Changelog
Every release, newest first — 187 in total.
0.7.113
A tidier, faster Roadmap window. Everything lines up in columns, it remembers what you had open, and big sections open much faster. Also: the Flatpak build can now run git and search tools on your computer, the status bar shows Claude Code's context use again, and several security hardening fixes.
0.7.112
/mcp) and it picks them up; the terminal and your other sessions keep running. The welcome window sets this up for you.0.7.111 — hotfix
Hotfix. Restores the Mageia package, which 0.7.110 could not build.
0.7.111 RC2 — Patron preview
Respin of v0.7.111-rc1 with cherry-picked fixes: a3134d7e
0.7.111 RC1 — Patron preview
This release shipped without written notes.
0.7.110
A performance report for contributors: tools/perf-report.sh runs every benchmark and compares each number against a saved baseline, flagging what got slower (ANTS-5133) Development tooling only; nothing ships in the user binary. How-to: docs/qa/perf-harness.md.
The roadmap repair step can now remove leftover detail lines at the end of an item's description (ANTS-4507)
roadmap_log op:"repair_trailers" takes strip_runs:true. It removes
the Kind / Source / Layman / Lanes lines a description stored by an
older version still ends in, but only where they repeat values the
item already stores. Where they disagree, nothing is removed and the
item is listed for review. Each removal is kept in the item's history.
apply_edits reports each edit's own replacement count in edit_replacements, and an optional per-edit expect_count skips an edit whose match count differs, before it touches the file. (ANTS-4838)
The Project Audit dialog's code is split into five source files; the dialog behaves exactly as before. (ANTS-1044) src/auditdialog.cpp keeps the dialog's core, and the check catalogue, Debt Sweep tab, AI triage and export each move to their own file. Tests read the class through one source list, and new checks keep the files in step with it.
MCP rate limiting remembers each caller folder's resolved path for 30 seconds, so repeat tool calls skip a filesystem lookup on the window's thread (ANTS-5090)
The etag option on MCP tools now states that a not-modified check needs the same arguments as the call that issued the etag (ANTS-4859)
On a store-backed project, a roadmap append that moves .roadmap-counter marks it counter_mirrored: true, since the store allocates the id and the file only mirrors it. (ANTS-4969)
Every MCP tool description now fits the 800-byte wire budget; the longer prose moved to each tool's detail, served by tool_info, and a test checks every registered tool. (ANTS-5152)
Switching back to a Claude tab resumes its task lists instead of re-reading the transcript (ANTS-5050) The task and background-task trackers keep each recently read transcript's parse position, so a tab switch back, or a second pane on the same session, carries on from where reading stopped.
workspace_search, cited_by and co_change_family now read ripgrep's output line by line instead of holding it whole (ANTS-5127) co_change_family also keeps only its strongest max_sites sites while it scans, instead of every candidate. Results and counts are unchanged.
Searching the terminal's history is more than twice as fast. (ANTS-2000) Each search reads history lines directly instead of one cell at a time. What a search finds is unchanged.
The test-audit window opens without the pause it used to have (ANTS-5126) Opening the window, and editing its dimensions field, ran a content scan over every test file before drawing the panel. The panel now asks only for the file grouping it displays, and the scan runs when an audit is actually dispatched.
The periodic session save now skips tabs that have not changed (ANTS-5030) Every thirty seconds each tab's whole history was re-serialised, compressed and written even when nothing had happened in it. A tab is now written only when something it stores has actually changed.
The diff viewer and clipboard guard build clean under the project's clang-tidy check set (ANTS-5129) Internal tidy-up with no change to behaviour: helper linkage, two byte-cap constants, one avoidable string copy, and the size of an internal enum.
The Independent Review and Cold-eyes dialogs search the project once per review round instead of twice (ANTS-5125)
CI's memory-checked test run skips the benchmark tests, as the local pre-push check already does (ANTS-5014) The benchmarks cost 285 s of a 1121 s serial sanitized run (measured 2026-09-11). Two correctness tests labelled perf for their large fixtures still run in CI's Release job.
The AI review windows keep one copy of the code that stops their requests on close (ANTS-5009) LlmDispatcher now only schedules jobs. ReviewDialogBase's runner creates the LlmClients and aborts them when the window closes. The dispatcher's own copy never ran, so nothing changes for the user.
Streaming AI answers are parsed without re-copying the buffer for every line (ANTS-5005)
Cold-eyes re-reviews now run cold and keep a round-by-round log (ANTS-2011) A re-review no longer tells the reviewer what was fixed, so a fix that did not hold shows up again. The results view lists each round: the lanes it checked and how many findings it produced.
Audit: the Contract-Doc ↔ Code Drift check now reports standards and specs as two separate lanes (ANTS-3849) One category held both, and the specs half outnumbered the standards half by more than fifteen to one — over 80% of the whole report. Nobody reads a category that size, so the real findings in it were invisible. Splitting suppresses nothing: both lanes still report every finding, and the small, readable half can now be acted on by itself.
The remote-control grab-image command saves a relative path inside the test artifact folder, and its reply names where the file really is (ANTS-5132)
Links and highlights stay correct on busy output when trigger rules are set (ANTS-5078) With a trigger rule that colours or links matching text, a burst of output could leave clickable links and highlight colours on the wrong place on lines printed later in the same burst. Those lines are now rechecked when the screen is next drawn.
Exporting from the terminal's right-click menu tells you when it fails (ANTS-5078) Export Scrollback as Text or HTML and Share Block could leave a half-written file on a full disk and say nothing. They now write to a temporary file that replaces the old one only when everything was written, and show a message in the status bar when an export fails.
Right-click command actions always act on the command you clicked (ANTS-5078) With shell integration, right-clicking a command offers Copy, Re-run, Fold and Share. If new output arrived while that menu or the Share dialog was open, the action could land on the neighbouring command, so Re-run could type a different one. Each action now remembers which command you clicked, and does nothing if that command is gone.
The verify-config trust prompt lists every gate and can remember a repo for good (ANTS-5082) When Claude asks to run a repo's own verify commands, the prompt now lists how many gates would run and their names, and "Trust this repo" has a checkbox (on by default) to ask again if the file changes. The trust file keeps the date each entry was first trusted instead of resetting it on every save, and Ants warns when that file can be read by other users.
Claude can no longer freeze the window by asking for a terminal's whole history (ANTS-5219) Claude's scrollback tool took any line count and read all of it on the window's thread. It now returns at most 10,000 lines, trims to the same size limit as reading a tab's text, and marks a reply that was cut so Claude knows it did not get everything.
The review dialogs no longer build every lane's brief at once (ANTS-5082) Starting a review built the full brief for every lane up front and held them all in memory while only a couple ran at a time. Each brief is now built when its lane actually starts, which keeps memory inside the review budget.
A background audit started by Claude no longer leaves its thread behind (ANTS-5080) An audit Claude starts in the background is run on its own thread. That thread was cleaned up only through the object that tracks such audits, so if that object went away mid-audit the thread was never freed. The thread now cleans itself up when the audit finishes.
Pasting a large screenshot no longer freezes the window while it saves (ANTS-5077) Pasting an image saved it as a PNG on the window's own thread, so a large screenshot froze every tab until the file was written. The image is now saved in the background, and its path is pasted once the file exists.
Long unbroken lines no longer leave tens of megabytes cached per tab (ANTS-5077) The drawing cache kept a fixed number of text runs, however long each run was. A line with no spaces is one run as wide as the row, so a few tabs of such output could hold tens of megabytes after it stopped. The cache now has a budget for the amount of text it keeps, and very long runs are drawn without being cached.
Closing a tab whose shell is stuck no longer risks freezing the whole window (ANTS-5077) When a tab's background reader was still busy two seconds after the tab closed, it was forcibly killed, which can freeze the window if it held a lock at that moment. It is now left to finish on its own and clean itself up.
A clickable link that wraps off the bottom row stays clickable on both lines (ANTS-5076) A link printed near the end of the last row wraps and scrolls the screen. Its recorded start row did not move with the scroll, so the link was dropped from both lines. It now keeps a span on each.
A large paste now reaches the shell whole, closing marker included (ANTS-5075) A paste bigger than the terminal's write queue used to lose its tail, and could lose the bracketed-paste end marker, leaving the shell stuck in paste mode. The paste is now fed to the shell in slices as it reads, and keys typed during a long paste arrive after it.
The audit drift checks no longer freeze the window, and audit_run keeps them under its time limit (ANTS-5067) The Audit dialog runs its spec, contract-doc and changelog drift checks on a background thread, so terminals keep updating and Cancel works while a check runs. A result that arrives after Cancel or a new Run is discarded. audit_run now counts these checks in its aggregate time cap; one still running at the deadline is reported timed_out and marks the run partial.
CI's Release job finishes again, and fails red instead of silently cancelling when it runs long. (ANTS-5188) cppcheck now runs in its own CI job, multi-threaded with a cache of its analysis. The Release job's build and test steps each have a time guard, its compile cache is saved even when a run stops early, and its tests run two at a time.
doc_citations no longer checks a bare line number against a file named in an earlier paragraph. (ANTS-4923)
A line reference like :120 on its own now borrows a file only from a
citation in the same paragraph. After a blank line or a heading it is
reported as unresolved, instead of a confident result about whatever
file the previous section happened to mention.
doc_citations no longer reports citations inside a review loop-log row as stale. (ANTS-4918) A citation on a Cold-eyes loop-log table row is still listed, marked loop_log_row:true and counted, but only:"stale" leaves it out, since a landed loop-log row is never edited. A line reference that follows a quoted path from another project is now foreign_path too, instead of a missing_file warning that could never be cleared.
Re-importing a roadmap no longer rewrites the wording of each item's plain-English summary (ANTS-4955) Summaries are now stored without their closing full stop and written back with one, whichever route saved them. Each roadmap file gains the stop on summaries that lacked it at its next render: a one-time diff.
A line such as ```json inside an open code block no longer ends the block, so the headings and text after it stay treated as code by the outline, spec log, feedback and changelog tools (ANTS-4987) Matches the CommonMark rule that a closing fence carries nothing after its backticks. Files with Windows line endings still close their blocks, and the code-block mask now shares the same rule as the other readers.
The MCP socket answers a malformed request line with a JSON-RPC parse error instead of closing silently after five seconds (ANTS-5089)
A remote-control client that stops reading a large reply is disconnected after 30 seconds instead of holding the socket open (ANTS-5093)
The Test Audit dialog drops a collected report when its chunk now covers different test files, so stale findings are not filed (ANTS-5102)
The status bar and tab dot follow a Claude session that was started before its first message, instead of showing it as idle for the whole session (ANTS-5156) The per-tab tracker now keeps looking for the session's transcript until Claude Code creates it, and no longer borrows another project's transcript in the meantime.
indie_review_partition says why it ignored a .indie-review/partition.json (map_rejected with a reason, such as a missing "version": 1) and no longer reports that file as the source it used. (ANTS-4846)
New tab shells no longer inherit a dead Claude Code session's identity variables, which switched transcript saving off in a claude started there; other CLAUDE_CODE_ settings still pass through. (ANTS-4541)
The per-tab thinking-level chip reads the last prompt a person typed instead of stopping at a tool reply, so it no longer shows Unknown on almost every live tab. (ANTS-1892)
invariant_check's scope note now says on every reply that specs are matched by path, so a spec citing the module only by symbol is known to be missing even when another spec matched. (ANTS-4972)
apply_edits flags a batch that applied some edits and skipped others with partial:true (would_be_partial:true on a dry run), instead of a bare ok:true. (ANTS-4856)
roadmap_log op:"set_body" accepts a replacement body up to 65536 characters, so long bodies can be written back whole; amend_body keeps its 4096-character fragment cap. (ANTS-4841)
co_change_family, docs_index, feedback_query and session_orient now declare the compact argument they already honoured, so a caller can pass compact:false to them. (ANTS-4657)
roadmap_query's tool detail now says when parseable_bullets and its warning appear, and that a store-served read never carries them, so their absence can be read correctly on either backend. (ANTS-4964)
roadmap_query answers a migrated project whose store holds no items from the store, with zero items, instead of falling back to the file and telling the caller to Read it. (ANTS-4861)
roadmap_query names the kind or source filter when it empties a result, instead of blaming the ID filter and calling the roadmap malformed. (ANTS-4971)
Roadmap writes leave a roadmap file untouched when its content has not changed, and list it under files_unchanged rather than files_written. (ANTS-5016)
workspace_search no longer warns about short terms when the whole regex is one word-bounded group, such as \b(TODO|FIXME|TBD|XXX)\b. (ANTS-5139)
file_outline lists gtest TEST, TEST_F and TEST_P blocks as Suite.Case, and no longer reports locals declared inside a test body as functions. (ANTS-5019)
mutation_probe refuses a baseline that collected no tests (pytest exit 4 or 5) as baseline_did_not_run, naming the test command, instead of reporting the suite as red. (ANTS-4852)
file_outline lists out-of-line C++ constructors and destructors, so read_region can fetch them by name, including a bare ~ClassName. (ANTS-5021)
mutation_probe, verify_changes and the test-results cache read ctest's all-passing summary ("100% tests passed out of N"), so a green run counts as green and require_green_baseline can pass. (ANTS-4996)
feedback_query flags a shipped id with no recorded ship date as possibly_stale_binary, marked stale_check:"no_shipped_date", instead of leaving the check silent. (ANTS-4843)
roadmap_query says when it applied a different body cap than the max_body_bytes asked for, with body_cap_clamped and the effective value, on an id or ids fetch. (ANTS-4981)
ants-terminal --remote waits for a slow first byte of the reply, up to its overall deadline, instead of printing "no response" while the server is still answering. (ANTS-5138)
workspace_search resolves a glob or exclude_glob containing / against lane when one is set, instead of returning no matches. (ANTS-5117)
A refused MCP call narrowed with fields no longer lists every requested field as unmatched, so the refusal's own code is the only cause shown. (ANTS-4979)
find_definition finds C functions whose return type, name and parameter list sit on separate lines, the style common in older C code such as DOOM. (ANTS-4828)
find_definition no longer lists a wrapped ternary arm (: sym(a);) or a stream insertion (out << sym(i);) as a declaration of the symbol. (ANTS-4924)
An offloaded reply's per-row preview now samples each object row's own JSON, so every head shows what its row is instead of the same {"v": prefix. (ANTS-5153)
tools/perf-report.sh -R <regex> --save-baseline now keeps every other benchmark's saved numbers instead of replacing the whole baseline. (ANTS-5137)
The git diff reader ends a file's hunks at a merge-conflict section instead of assigning its lines to the file before it. (ANTS-5111)
The debug log now rotates when it passes its size limit during a session, instead of growing until Ants restarts (ANTS-5110)
read_log refuses a path that is not a regular file, so a named pipe can no longer freeze it (ANTS-5110)
Review Changes no longer lists a repository's whole history on every refresh when it has no remote, and a crashed git command no longer refreshes the window twice (ANTS-5109)
A line break in a changelog entry's summary, id or dated headline stays on that line and can no longer create a fake heading (ANTS-5108)
project_query no longer refuses every query for the rest of the session after 64 slow queries (ANTS-5107) A query worker that was left running past its deadline and later finishes now frees its slot.
project.read refuses anything that is not a regular file, so a named pipe can no longer hang a query (ANTS-5107)
Two open Ants windows no longer corrupt each other's saved tabs (ANTS-5106) Each process now writes its own temporary file before replacing a saved session or the tab order.
A failed tab-order save (for example on a full disk) keeps the previous tab order instead of replacing it with a cut-off file (ANTS-5106)
read_region clips an oversized first line to its byte limit instead of returning a huge single-line file whole (ANTS-5103)
An automatically sized results page always includes at least one row, so following next_offset can no longer loop forever (ANTS-5103)
codebase_index reports a failed cache write instead of ignoring it (ANTS-5103)
The AI client reports an error for a successful reply it cannot read, instead of an empty answer that looks like success (ANTS-5105) A proxy login page or any other body that is neither a stream nor a chat completion used to show nothing in the AI chat and read as a clean review lane.
Build-error fix hints recognise GCC's curly quotes, so they work with this machine's compiler output (ANTS-5102)
verify_changes no longer loses ctest's failed-test list when one large chunk of output arrives at once (ANTS-5102)
An Independent Review lane now stops at its time limit even when the AI endpoint keeps trickling data, and a reply over 10 MiB is cut off (ANTS-5101)
The debt sweep's automatic fix no longer corrupts bytes that are not valid UTF-8 elsewhere in the file (ANTS-5101)
spec_conformance refuses a pattern that does not compile instead of passing every no-match row against it (ANTS-5100)
doc_lint's table-of-contents fix recognises code blocks the same way the check that reports the gap does (ANTS-5099) It used its own simpler rule, which could insert a contents row in the wrong place.
docs_index reports a failed cache write instead of ignoring it (ANTS-5099)
indie_review_fold_in's narrative mode honours dry_run and no longer writes ROADMAP.md on a preview (ANTS-5097)
verify_changes no longer serves a cached pass when git status fails or times out (ANTS-5097) A failed status used to hash as a clean tree; that snapshot is now marked uncacheable.
session_message returns at most 200 messages per inbox call and refuses a message_id that is not a positive whole number (ANTS-5098) A limit of 0 or less used to return the whole mailbox.
changelog_log add_batch accepts at most 200 entries per call (ANTS-5098)
A roadmap flip_batch or annotate_batch where no item resolves in the roadmap database now reports failure under its own name instead of success (ANTS-5095) The review-kind warning on append_batch also names the ids the batch assigned rather than the headlines.
mutation_probe writes the mutated file and the restored original atomically, so a full disk or crash cannot leave a source file cut short (ANTS-5096)
build_target_for validates its cmake_path argument, and file_outline accepts at most 100 paths per call (ANTS-5096)
Editing a roadmap item's plain-language summary, source or evidence through roadmap_log can no longer break a line in ROADMAP.md or split an evidence path (ANTS-5094) The single-field edit now cleans values the way adding an item does.
roadmap_query's report mode refuses a since or until date it cannot read instead of quietly reporting a different period (ANTS-5094)
The Claude permissions window says when its settings file cannot be read, and the transcript window no longer shows an empty User line for each tool result (ANTS-5092)
The Background Tasks window reads each task's output from the exact path it checked, and reads a finished task's output only once (ANTS-5092)
The roadmap window's kind filter covers every roadmap kind, and a roadmap file over 8 MiB is no longer cut off (ANTS-5088) Perf, security, feature and six other kinds can now be filtered, and a kind filter no longer hides them. The live roadmap is read up to 64 MiB.
Publishing the roadmap refuses a path that leaves the project before creating any folder, and never writes through a symlink that points outside it (ANTS-5087) A dry run no longer creates folders either.
On a pass-headings roadmap, a hand-added pass block is protected from being deleted by the next roadmap write (ANTS-5087) The write now refuses and names the block, as it already did for ordinary bullets the database has never imported.
Roadmap migration and export rebuild match item ids with non-ASCII letters the same way the roadmap database does (ANTS-5087) The database folds only A-Z when comparing ids, while the migration and the rebuild also lowercased letters such as Ä, so those ids could link to the wrong item or none.
An older Ants build meeting a roadmap database from a newer build now refuses at once instead of waiting for the write lock (ANTS-5086) The new roadmap database file is also created owner-only before any roadmap data is written into it.
Since-last-run audits no longer re-count header findings as new or grow the recorded finding list every run (ANTS-5085) A finding a tool reports in an unchanged file that the previous run also had is carried forward once, not added again.
Suppressing an audit finding no longer rewrites the whole rule-quality history file (ANTS-5085) The history is saved at the end of a run and when the audit window closes, and only when something was recorded.
Audit context filters skip very large referenced files, and the review false-positive ledger cache is thread-safe (ANTS-5085)
Audit auto-fix repairs files with Windows line endings, and the audit cache folder is always created private (ANTS-5085) Auto-fix used to treat every planned repair in a CRLF file as out of date; it now applies them and keeps each line's ending. The SARIF and HTML exports and the gitleaks config writer create the .audit_cache folder owner-only from the start.
Audit window: the Since baseline filter and changed-lines scope no longer hide every finding, and batch AI triage is throttled (ANTS-5084) A run with Since baseline on keeps the changed-line data that filter needs. When git cannot say what changed (a short history, a timeout, a failure), the status line says so and the changed-lines filter is skipped instead of hiding everything; a repository with fewer commits than the window counts every line as changed. Batch AI triage sends two batches at a time through one shared network client, refreshes the list once at the end, and stops reading a reply past 10 MiB.
Audit window: project detection no longer walks build trees, and exports, labels and the changed-files filter are correct (ANTS-5084) Opening the audit window lists only the top few folder levels when detecting Docker and Terraform files. The signal banner shows one percent sign. A finding containing an HTML comment opener can no longer blank the HTML report. A tool that failed to start is labelled a tool issue, not a timeout. A failed export says so. A changed oo.cpp no longer matches src/foo.cpp.
Audit auto-fix skips findings suppressed after the run, and the contract-drift lanes keep every finding (ANTS-5083) Auto-fix now checks suppressions when it runs, skips findings with no line before reading any file, and does not read very large files. The per-check findings cap no longer applies to lanes that are uncapped on purpose.
A suppression that could not be saved no longer hides the finding, and a new audit run re-reads source lines (ANTS-5083) The audit dialog now checks the suppression file write before hiding the finding and says so when the save fails. Each run clears the cached source lines, so an edited file is shown as it is now.
The themed stylesheet header comment now matches the rules it builds (ANTS-5082)
A mistyped colour in a custom theme now uses the default colour instead of turning black, and an oversized theme file is skipped (ANTS-5082)
The About dialog no longer describes a GPU renderer the terminal no longer has (ANTS-5082)
Clicking Re-review while a review is still running no longer wipes that round's failures, so a failed lane can't be mistaken for a clean one (ANTS-5082)
A full disk can no longer replace the verify trust list with a cut-off file, and a damaged trust file is kept aside instead of being overwritten (ANTS-5082)
A global hotkey that the desktop refuses to register now shows the 'Global hotkey unavailable' message instead of failing silently (ANTS-5081)
The global hotkey now works on desktops that start their shortcut service on demand (ANTS-5081)
A tab's tooltip keeps naming the tool Claude is using instead of being reset to a generic state on every redraw (ANTS-5081)
The Public/Private repository badge recovers when the gh tool is missing or stops responding, instead of never updating again (ANTS-5080)
Restoring the window position through KWin no longer leaks a helper process or leaves the loaded script behind when a step cannot start (ANTS-5081)
The Review Changes button no longer shows the previous tab's git state after a quick tab switch, and a stuck git check can no longer freeze it for the session (ANTS-5080)
Start-up cleanup of old MCP sockets now removes only real sockets owned by you (ANTS-5080)
Exporting scrollback now tells you when the file could not be written, instead of reporting success (ANTS-5079)
Moving or centring the window through KWin no longer leaves a stray process and temp file behind when the helper cannot start (ANTS-5079)
Connecting to an SSH host no longer risks a crash if the new tab closes straight away (ANTS-5079)
Copying a very large selection with formatting no longer builds hundreds of megabytes of styled text (ANTS-5078) Formatted copy now groups characters that share a style, and above a size limit copies plain text only.
Claude running in a split pane now lights its tab's status dot, and closing a pane or a split tab releases what it was tracking (ANTS-5079) Split panes were never registered with the Claude and background-task trackers, and closing one left its shell listed in them.
A search pattern that can match an empty spot no longer hides the real matches later on the same line (ANTS-5078)
Jumping to a search match or a bookmark now moves the scroll bar too (ANTS-5078)
Right-clicking with a very large selection no longer pauses while the whole selection is copied just to measure it (ANTS-5078)
A terminal recording no longer garbles a character whose bytes arrived in two pieces (ANTS-5078)
A background image keeps its shape instead of being stretched to fit the window (ANTS-5077)
A command's duration never shows as negative when the system clock steps backwards (ANTS-5077)
Drawing the terminal no longer builds a highlight cache entry for every line when no highlight rules are set (ANTS-5077)
In broadcast mode, Ctrl key combinations and accepted suggestions now reach every pane (ANTS-5077) Ctrl+C, Ctrl+arrows and accepting an autocomplete suggestion went only to the focused pane, and did not clear a highlighted selection.
A pasted screenshot's path works at the prompt even when the paste folder has spaces, and the saved file is private to you (ANTS-5077) The saved screenshot's path is now quoted like a copied image file's, the file is readable only by you, and nothing is announced when saving fails.
A tab whose shell failed to start no longer behaves as if a shell were running (ANTS-5077)
Turning on the underline debug log no longer slows output, and the log stops growing past a fixed size (ANTS-5076)
Resizing the window no longer glues together words that were split across two lines (ANTS-5076) When a line wrapped exactly at a space, resizing dropped that space and the words on either side ran together.
Clickable links disappear when the text under them is cleared or overwritten (ANTS-5076) A link stayed clickable over blank space after the screen or line was cleared, and redrawing the same link over and over kept piling up copies of it.
Sixel images decode faster, so a crafted image can no longer stall the window as long (ANTS-5076) Each pixel was set through a slow general-purpose call; the decoder now writes pixels directly and skips drawing past the image edge.
A garbled terminal control code no longer spills its leftover characters onto the screen (ANTS-5075) When a program sent a control code with a stray byte in it, the terminal gave up on the code and printed the rest of it as text. It now discards the whole code, the way other terminals do.
Clearing the screen now always clears a text selection, even when the clear arrives in two pieces (ANTS-5075) The terminal looked for the clear-screen code inside each chunk of output it read, so a code split across two chunks was missed and the old selection stayed highlighted.
A shell opened on a desktop session with a very large environment now receives every variable, including the one Claude Code uses to find the terminal (ANTS-5075) The terminal copied its environment into a fixed-size list and stopped when the list filled, so the variables added last were lost. The list is now sized to fit.
Opening a long Claude transcript no longer freezes the window while the whole file is decoded (ANTS-5089) The transcript viewer shows only the last 2000 entries, but it decoded every entry in the file first, on the window's own thread. It now decodes just the entries it shows and only counts the rest.
Claude's status shows up for a Claude session started from a subfolder or subshell (ANTS-5089) Ants found the Claude session's transcript by looking at the shell's current folder. If Claude was started after changing folders in a subshell, that folder belonged to a different project, so no transcript was found and the status stayed blank. Ants now looks at the folder Claude itself is running in first.
Reading terminal text over remote control no longer copies the part it throws away (ANTS-5093) When a get-text reply was over its size limit, Ants copied the whole dropped part just to count its lines. It now counts them from what it already holds.
Tabs without Claude running no longer make Ants read every process's details every two seconds (ANTS-5089) To spot Claude Code in a tab, Ants checks the shell's child processes every two seconds, and when none was Claude it also read the details of every process on the system. That full scan now runs once every five checks for a tab with no Claude running, and on every check while one is.
The Claude transcript viewer and session summaries no longer drop messages longer than 64 KiB (ANTS-5089) Ants read Claude Code's transcript one record at a time but stopped at 64 KiB, so a long tool result or a pasted file was split into pieces that could not be read and was silently left out. Records up to 16 MiB are now read whole, and a larger one is skipped cleanly instead of in fragments.
Saving a large answer with many rows to disk no longer slows down sharply as the row count grows (ANTS-5104) When a big MCP answer is saved aside, Ants builds a short per-row preview that has to fit a size budget. It rechecked the size by rebuilding the whole preview after every row, so the cost grew with the square of the row count. It now keeps a running total and builds each row once, with the same preview as before.
Ants no longer rewrites Claude Code's settings.json on every launch, and no longer replaces a hooks value it cannot read (ANTS-5104) When the Ants session hook is already in place, starting Ants now leaves ~/.claude/settings.json untouched instead of rewriting it. A "hooks" value that is not an object, or a "SessionStart" value that is not a list, is now refused with a warning instead of being silently replaced.
Calls to tool names that do not exist no longer grow the token usage counter, and each MCP call does less bookkeeping (ANTS-5104) Every made-up tool name used to add its own permanent entry to the per-session token counter. They now share one "(unknown tool)" entry, still counted as failures. The tokens-saved chip also reads its total directly instead of rebuilding and sorting the full report on every call.
The tokens-saved chip shows 1M instead of 1000K for counts just under a million (ANTS-5104) A count that rounds up to the next unit (999,950 and above) now moves to that unit, so it reads 1M rather than 1000K, and the same for M to B.
Reading a large saved MCP answer page by page no longer garbles characters at page edges (ANTS-5104) read_spill could end a page in the middle of a multi-byte character, which showed as a replacement character and threw the page length off. Pages now end on a whole character, and each page reads only its own part of the saved file instead of the whole file.
Suspending the computer or changing its clock no longer lets two audits run on one project (ANTS-5090) The short read cache, the "already running" guard for audit_run and indie_review_dispatch, and the async audit job list now measure elapsed time on a monotonic clock. A clock step backward could keep a cached read fresh, and a suspend could free a guard slot while its sweep was still running.
A "too busy, retry shortly" answer is no longer remembered and repeated for a cached MCP verb (ANTS-5090) When the dispatch queue was full, the refusal for a cached read verb was stored in the short-lived read cache, so the next call got the same refusal after the queue had drained. The refusal is no longer cached.
Large answers from background MCP verbs no longer stall the Ants window while they are prepared (ANTS-5072) The reply transforms (etag, field projection, compaction, hints, offload and the data wrap) now run on the dispatch worker for verbs that already run there. The window thread keeps only the cache insert, the usage record and the socket write. Measured before the change: a 4 MiB reply cost about 92-95 ms on the window thread.
changelog_query finds an id cited only in a dated topic's headline or prose (ANTS-5145) A dated topic heading written by changelog_log op:add_subsection is now an entry of its own: its text is the headline and its body the prose under it, so an id lookup reports an id cited there instead of calling it missing.
Remote-control and MCP tab tools act on the window you last used, not on a hidden first window (ANTS-5121) Delivered by the shared listener (ANTS-5144): each request goes to the most recently active visible window, whose tools act on that window.
Closing a New Window window no longer cuts MCP, Claude hooks and remote control off for every other window (ANTS-5144) Each window used to take the three socket paths from the window before it, and closing it removed the socket files. Now one listener per path serves the whole process: requests go to the most recently active visible window, and a hook event goes to the window whose tabs track its session. A second Ants process no longer takes the remote-control socket from the first.
The sanitizer build builds again (ANTS-5143) The perf benchmarks are no longer built when ANTS_SANITIZERS is on, so a full build-asan build and its test suite complete, and the pre-push hook's ASan leg runs again.
changelog_query returns every entry for an id lookup (ANTS-5147)
An id or ids lookup now ignores offset and limit, as its
design and roadmap_query do, so offset:10 no longer returns an empty
page for an id that is present.
A malformed extended-colour code no longer sets text attributes (ANTS-5136)
An SGR 38, 48 or 58 introducer whose selector is neither 2 nor 5 now
abandons the rest of the sequence. Before, ESC[38;1m applied bold
and ESC[48;3m italic.
changelog_query refuses an id or ids of the wrong JSON type (ANTS-5146)
A number, boolean or object passed as id or ids now refuses
bad_args instead of returning the whole changelog.
changelog_log add_subsection refuses a date it cannot parse (ANTS-5148)
A date that is not a valid yyyy-MM-dd date now refuses bad_args,
so every dated heading it writes is one changelog_query can read.
changelog_query reads entries under dated topic headings (ANTS-5071)
A ### <date> <Category> — <headline> heading, the form
changelog_log op:add_subsection writes, now sets the category named
after a valid date. Its bullets are returned, found by id lookups and
counted in version_index. ANTS-3533 § 3 and INV-10 were amended first.
plan_template's includes_tests:false now drops each task's test file line and test-first steps (ANTS-5068) The option was documented and parsed but never read, so false returned the same skeleton. The remaining steps are numbered from 1; the default output is unchanged.
The window no longer freezes while a --remote search, tree walk or git command runs, and the roadmap store is no longer used from two threads at once (ANTS-5073) Remote-control socket routes whose MCP twin already runs off the GUI thread now run on the same MCP dispatch worker, with the reply written later. ANTS-5051 closes with it. indie_review_dispatch still freezes the window for a review.
A synchronous audit_run no longer freezes the window for the whole sweep (ANTS-5035) It replies from a completion slot instead of joining its worker on the GUI thread. A second synchronous call for the same project while a sweep runs is now refused with already_running.
Closing a second Ants window no longer makes Claude commands in the remaining window refuse (ANTS-5142)
The window no longer stalls while Claude writes to its transcript. (ANTS-5050) The task and background-task trackers now read only the newly appended part of the conversation log instead of re-reading up to 16 MiB of it every time Claude writes. Measured on a 65 MiB transcript, one update went from 258 ms of frozen window to 0.09 ms. Whenever the log cannot be resumed safely — truncated, replaced or rewritten — the full read still happens, so nothing is missed.
Leftover text styling is cleared when a foreground program exits. (ANTS-5135) A program interrupted while printing dimmed text used to leave the dim attribute set, and on distributions whose shell prompt carries no colour reset of its own — openSUSE's default among them — every later prompt and everything typed stayed dim. Ants now clears the character attributes when a program hands the foreground back to the shell. Shell builtins keep their styling, since they never leave the shell's own process group.
A malformed colour escape no longer leaves the text after it dim (ANTS-5130) An extended-colour sequence too short to carry its colour had its own selector run as an attribute code, switching dim on for every cell that followed until the next reset. Such a sequence is now abandoned whole.
Closing Review Changes now cancels its background git probes before they are torn down (ANTS-5128) Closing the window while a probe was still running destroyed it mid-flight, which ran the probe's completion handler against the half-closed window. The probes are now stopped first, on every way the window can close — the Close button, the parent window, and quitting the app.
The Review Changes window no longer freezes on a huge diff: it shows the first 1 MiB with a note, skips redrawing when nothing changed, and ignores an older refresh that finishes after a newer one (ANTS-5059)
The roadmap window reads its "last touched" dates in the background and can no longer put a date on the wrong card when the roadmap changes mid-read (ANTS-5047)
The audit window's Debt Sweep tab scans in the background, so the window stays usable while it runs and after every fix or allow click (ANTS-5057)
The audit window no longer freezes after a run while it looks up who last changed each flagged line; it asks git once per file in the background and fills the names in when they arrive (ANTS-5040)
Saving roadmap changes no longer re-adds up the whole change history for every row it writes, and a history size that cannot be measured now refuses the write instead of switching the size limit off (ANTS-5046)
A push interrupted during its Qt 6.2 compile check now stops that check's container, and the next check does not trust the half-built tree (ANTS-5124)
The "Close tab?" question box now matches the Ants theme, can be resized and remembers its size (ANTS-5123)
The build no longer prints its GCC warnings, and the tests they pointed at still check what they read (ANTS-5114)
project_query refuses a file too large for its memory limit before reading it, and its Lua callbacks no longer leak memory when a query runs out of room (ANTS-5070)
Code searches stop reading ripgrep's output past a size limit and say the answer is partial, instead of holding all of it in memory (ANTS-5052)
The Claude status tick stops reading saved conversations once none of the rest can be the live one (ANTS-5048)
project_query stops building a result as soon as it is over its size limit, so a small Lua snippet can no longer exhaust memory (ANTS-5069)
Independent Review and Cold-eyes never walk into build and dependency folders (ANTS-5058)
The test-audit partition never walks into build folders, walks the project once for all patterns, and keeps a project that sits inside a folder named build (ANTS-5062)
doc_citations reads each quoted document once per call, inside its read budget and size limit (ANTS-5054)
The subsystem lane cache is locked, so the window and the MCP worker can read it at the same time safely (ANTS-5074)
The Compiler Warnings audit check can finish, and no longer leaves a build folder behind in /tmp (ANTS-5039)
A verify_changes gate that times out stops everything it started, with SIGTERM first so a build can stop cleanly (ANTS-5063)
A timed-out or cancelled audit check, and an audit_run tool past its time cap, stop every process the tool started (ANTS-5038)
The review engines share one locked file cache with a size limit, so the window and the MCP worker no longer write one cache at once (ANTS-5056)
recent_errors finds every missing-include hint in one project scan instead of one scan per symbol (ANTS-5053)
recent_errors reads a line the terminal wrapped as one line, so an error wider than the pane is reported whole (ANTS-5061)
co_change_family clips a long match in one pass, so a very long line no longer stalls it (ANTS-5066)
Closing a tab while its Background Tasks window is open no longer risks a crash; the window closes with the tab. (ANTS-5049)
The test-result tools now name the tests that failed, not just how many failed. (ANTS-5064)
The link checker no longer reports false broken links in a document too large for it to read; it skips that document instead of reading part of it. (ANTS-5055)
The test-audit summary returns one page of reports when no page size is given, instead of all of them. (ANTS-5065)
An audit of recent changes checks the whole project when git fails, instead of reporting a clean result. (ANTS-5043) This covers a repository with no branch called main, a mistyped tag, and a git call that times out.
Dialogs remember their size when you close them with the close button or Esc. (ANTS-5037)
Filtering or sorting audit results no longer lowers how serious findings look. (ANTS-5041) A line that two tools both flag is now ranked higher, as intended.
Scrolling back through a command's output stays smooth even when that output is huge. (ANTS-5027) The command name pinned at the top while you scroll is now built from only as much text as fits, instead of rereading the whole output on every redraw.
Closing a window now asks first when a program such as an editor is still running in any of its tabs, as closing one tab does. (ANTS-5120) The question offers the same "don't ask again" choice as the tab one, and turning it off applies to both.
A program that floods the terminal with output no longer holds up Ctrl+C and other keys, and a finished shell is reported once. (ANTS-5026) Ants stops reading as soon as it falls behind and resumes when it catches up, instead of queueing output without limit.
Terminal output can raise at most ten desktop notifications a minute per tab. (ANTS-5033) A program or file that prints notification codes in a loop no longer starts a notification process for each one.
Closing one of several windows closes its tabs and ends their programs, including the original window, which used to keep them running out of sight. (ANTS-5118) The last window still keeps its tabs for the next start. The tabs close a moment after their window, so a logout that closes every window still saves them all.
A second window (File → New Window) opens with one fresh tab instead of reopening every saved tab, and a restart brings back the tabs of every open window. (ANTS-5032) Saved tabs are restored once per run, into the first window. Each window now saves the whole run's tab list rather than only its own, so the last window to save no longer drops the others' tabs.
Changing View -> Opacity updates terminals that are already open (ANTS-5034)
audit_run reports a tool that ran out of time as timed out, not crashed, and ignores its half-finished output (ANTS-5044)
A saved tab that cannot be restored keeps its session file, restores all or nothing, and very large scroll histories save only what can be restored (ANTS-5031)
Re-running a command reads the right line after the scroll history has filled up (ANTS-5029)
Opening a link from the suspicious-link warning is safe while output keeps streaming (ANTS-5028)
Dialogs keep remembering their size after a second window closes (ANTS-5036)
Audit runs and project queries no longer leak memory, and the everyday checks now catch new leaks (ANTS-3847) Each audit tool's process kept itself alive and was never freed, and a refused project.read or project.list skipped its cleanup. Both are fixed. The sanitized test runs before a push, in tools/ci-parity.sh and on CI now check for leaks, so a new one fails the run instead of going unnoticed.
Closing Ants while a Claude tool call waits on the window no longer hangs (ANTS-5113) At shutdown, a tool call already waiting for the window thread used to wait forever while the window waited for it. Shutdown now releases that call without running it, so Ants exits.
The verify.json trust prompt opens on the main thread (ANTS-5025) verify_changes runs off the GUI thread, so the "Trust .ants/verify.json?" prompt used to be built and run on a background thread, which Qt does not allow. It now always opens on the main thread. If Ants is shutting down, no prompt is shown and the untrusted commands are not run.
indie_review_dispatch replies instead of hanging Ants (ANTS-5024) The verb's worker now receives the project root the provider has already resolved on the GUI thread. It used to resolve the root itself, which asked the GUI thread for every tab's cwd while the GUI thread was waiting for that worker, so every call froze Ants for good.
Ants's own git checks can no longer make your git commit fail with "index.lock: File exists" (ANTS-4999) Every read-only git check Ants runs now skips git's optional index lock. That covers the Review button's two-second check, the git MCP verbs, audit scoping and the Claude git-context hook. A hook that is already installed picks this up when reinstalled from Settings.
The AI review job pool reports a finished round exactly once, even when a job completes the moment it starts (ANTS-5000) Before, each nested step announced the same round again. Only test code reached this path; live AI requests always finish later.
Four end-to-end checks now fail when the typed command never runs (ANTS-5013) The ANSI colour, CJK, scrollback and resize cases looked for text the typed command already contained, so the terminal's echo satisfied them. Each now looks for text only running the command can produce.
AI answers from a provider that ignores streaming now appear, and a server's own error message now reaches you (ANTS-5007) LlmClient keeps the raw reply until an SSE line proves it is a stream, then reads a plain JSON answer or error body whatever the HTTP status.
Reaching the AI answer size cap now ends the download (ANTS-5008) The request is aborted at LlmClient::kMaxBytes and reported as a truncated answer, even while the server holds the connection open.
Closing an AI-backed window mid-request no longer crashes the app (ANTS-5002) An AI client destroyed with a request in flight now shuts down quietly. The audit dialog's AI triage could reach this when closed mid-answer.
Long AI answers that arrive all at once are delivered whole (ANTS-5015) An answer longer than one parsing pass, sent in a single burst by a fast or local AI server, used to come back cut short with no warning.
Review dialogs report a failed AI request as a failure, not a clean review (ANTS-5003) A lane whose request failed now shows the error in its tab and is named in the status line, and it is left out of the results. Test audit's resume now retries such a chunk instead of treating it as reviewed.
The review dialogs' Dispatch button is disabled while a round runs (ANTS-5004) A second click mid-round used to wipe the collected reports and pay for every lane again.
Dispatching a review with no lanes says so instead of reporting a finished review (ANTS-5006)
The sanitizer CI job's test step has room to finish again (ANTS-5011) Passing runs were finishing within a minute of the 16-minute limit, and three were cut off at it. The limit is now 22 minutes, re-measured from recent runs, and the job's overall cap rises to match.
The end-to-end test cases wait for the screen instead of sleeping a fixed second (ANTS-5012) A slow instance, such as the sanitizer build in CI, no longer fails them. Each case now polls for its expected text for up to about 15 seconds.
The blocked-roadmap-write message now describes the rule the code actually enforces (ANTS-4749) A roadmap write blocked for a missing plain-language summary told you the summary had to be the first line of your note. It never did — the key may lead any line of the note. Following the old wording cost nothing; reading it as exhaustive led you to conclude a fix would not work when it would.
Files that hold your project's content or your terminal's output are no longer written when Ants cannot make them readable by you alone; settings and other small files log a warning instead (ANTS-5151) Covers audit exports and caches, saved reports, the debug log, and session logging and recordings, which now say in the status bar why they did not start. The roadmap database warns rather than refusing, so every project keeps working.
Resuming a Claude session quotes the session id in the command it types (ANTS-5080)
The command that asks Claude to review an audit quotes its prompt, so an unusual results path cannot run shell code (ANTS-5079)
The optional key-press debug log no longer records what you type, so passwords typed at a prompt stay out of it (ANTS-5077)
A remote-control command name can no longer add fake lines to the debug log (ANTS-5093) The remote-control socket wrote the command name it received straight into the debug log, so a name containing a newline could add a line that looked like a real log entry, and an escape sequence could affect a terminal showing the log. Control characters are now written as visible escape codes.
OpenAI service-account and admin keys are now scrubbed before text is sent to an AI model (ANTS-5104) The secret scrubber recognised OpenAI project keys and the old key format but not the sk-svcacct- and sk-admin- keys OpenAI now issues. Both are now replaced with a [REDACTED] marker.
A large MCP answer saved to disk is discarded if Ants cannot make the file private (ANTS-5104) The saved copy holds the answer itself. If its permissions cannot be set to owner-only, the file is removed and the answer is returned directly instead.
Resetting the terminal no longer lets a program get around the clipboard-write and user-variable limits (ANTS-5122)
Re-run Last Command shows the command and asks before running it, unless prompt signing proves the command came from your shell (ANTS-5029)
Audits no longer take tool options from the audited project's .audit-config.json, so a downloaded project cannot use them to write files or run code (ANTS-5045)
Batch AI triage and indie_review_dispatch remove passwords and keys from source text before sending it to the AI endpoint (ANTS-5042)
SSH bookmark extra arguments are checked the way ssh reads them, including spaced, quoted and combined option forms, config files and library-loading options (ANTS-5060)
A keyless AI request sent over plain http to another machine now warns that it is not encrypted, and still sends (ANTS-5010) The AI chat, the review dialogs, the audit dialog's triage and the indie_review_dispatch reply each show the same warning naming the host. Localhost and https endpoints are unaffected.
The indie_review_dispatch MCP tool now runs the same AI send checks as the rest of the app (ANTS-5018) It refuses an API key over plain http to a remote host, private and link-local IP addresses, and credentials embedded in the URL, and it no longer follows redirects.
0.7.110 RC1 — Patron preview
This release shipped without written notes.
0.7.109
Roadmap items can now be searched by where they came from (ANTS-4985)
roadmap_query gained a source filter — a case-insensitive prefix, or a
list of them — so "how much review work is still open?" is one call. Each
item now reports its source as a field alongside its kind and lanes. The
provenance was always recorded and nothing could read it.
Filing a review's fix under the wrong label now says so (ANTS-4989)
Adding a roadmap item whose origin is a review but whose kind is a plain
fix returns an advisory suggesting review-fix or audit-fix. It never
refuses the write — the label may be deliberate — and it names which
origin it matched.
A roadmap publish that overwrites unrecoverable text now keeps a copy of the file and names it (ANTS-4947)
discarded_backup_paths[] rides the lost-text arm of a store-backed
roadmap write. A lost line is text the render reproduces in no
styling, so the file on disk was its only copy and the capped
twenty-line discarded_text echo was all a caller had to retype
from. The copies go under the shared data directory, never beside
the project, so they cannot dirty git status. Restyled and
repunctuated drift keeps nothing — that text survives in the render.
similar_code accepts query as an alias for shape (ANTS-4951)
shape wins when both are sent, as every other alias here does.
roadmap_migrate reports the note population it capped, and lets you raise the cap (ANTS-4559)
notes_summary maps each note code to its total count over all notes,
tallied before the row cap and never truncated, so a truncated notes[]
can still name what it dropped. max_notes moves the row bound, default
200 and clamped to [1, 2000]; the envelope echoes the effective value.
roadmap_query accepts q as an alias for query (ANTS-4927)
Unrecognised, q was dropped and the call answered with the whole list,
which reads as a search result. Follows the existing filter/status
and max_results/limit aliases; query wins when both are sent.
apply_edits accepts old_text/new_text and a batch-level path (ANTS-4936)
The spelling roadmap_log's op:"amend_body" uses now works here too,
alongside the native Edit tool's old_string/new_string. A path given
once at the top level supplies every edit that omits one; a per-edit
path still wins.
doc_lint can repair a drifted table of contents (ANTS-3669)
Pass fix:true and the one auto-fixable finding is corrected in place: a
missing Contents entry is inserted in document order, a duplicated one is
removed. Every other line is left exactly as it was — sub-entries,
free-text rows and entries pointing at nothing are all preserved, because
a rebuild-from-headings would quietly delete them. dry_run:true shows
what would change without touching the file. Everything else the verb
reports stays report-only.
doc_lint — every deterministic document check in one call (ANTS-3663)
Runs doc_integrity, doc_citations, doc_dedup, doc_symbols and spec_lint over one enumeration and one shared read, returning a single findings list in a total order so two runs diff cleanly. Narrow it with checks[]; an unknown name is refused rather than read as "all". spec_lint is applied only to documents under the project's specs directory, so a README is no longer flooded with findings about invariants it does not have. Reports only — nothing is written.
search the roadmap by kind of work (ANTS-4836)
Every item records what kind of work it is, and nothing could select on
it — so "which review fixes are still open?" meant fetching every open
item and sorting by hand. roadmap_query now takes a kind, composing
with status, section and keyword. An unrecognised kind is refused with the
list of real ones rather than quietly returning everything, which would
have read as "nothing of that kind exists".
read a long roadmap item's latest note in one call (ANTS-4904)
A long-lived item is an append-only progress log, so what a resuming
session wants is its last paragraph — and the body could only be kept from
the head, with a fixed 1 KiB tail. Asking for more raised the head with it
until the reply spilled and the preview dropped the body altogether: three
calls and a hand-picked byte offset to read one closing paragraph, and an
offset guessed too high missed the note silently. body_from_end on a
targeted id fetch keeps the END instead, sized by max_body_bytes, and
every truncated body now reports body_bytes — its true size — so what is
missing is a number rather than a guess.
project_settings op:"get" — the read, named like one (ANTS-4903)
Asking what a project has declared meant calling op:"detect", whose
name and documented purpose are to propose settings for a project that
may have none — so the answer arrived with an inert suggestion block
attached, and a session reading the op as write-shaped could skip the
check entirely. op:"get" returns the declaration and what is still
unset, and no proposal. Both are read-only and share one implementation,
so they cannot disagree about what is declared.
UnrealScript (.uc) files are outlined, indexed and searchable for definitions (ANTS-4902)
file_outline reported unknown with no symbols for .uc, so a project
whose whole runtime surface is UnrealScript had to read those files whole.
It now outlines as brace-family — classes, structs, consts and every
function spelling the language uses, including the modifiers no other
brace language has. codebase_index and find_definition admit the
extension in the same change, which the three lists' own in-step rule
requires. var members are the known gap.
spec_lint reports how many invariants it recognised, so a spec it read none of is visible (ANTS-4894)
A spec whose invariants are written in a form the format standard does
not define parses to zero, every per-invariant check then runs over an
empty set, and the envelope was byte-comparable with one that checked
thirteen. invariants_found is emitted on every run — zero is the alarm
— and the hint that says invariant_no_test always runs now points at
it rather than reading as evidence that any invariants were seen.
The roadmap format standard now requires every item to record where it came from (ANTS-4985)
Source: moves from optional to required (format v1.2), matching what the
writing tool has always enforced, and the standard now states that an
item's kind and its origin are separate questions — a review legitimately
produces work of any kind.
A narrowed response names the fields it does carry, not just the ones it doesn't (ANTS-4930)
fields_available now accompanies fields_unmatched, so a misspelled
field name is visible rather than inferred, and a field that is simply
not populated on this call can be told from one this backend never
carries.
The body-scrub warning names the text it removed (ANTS-4938)
body_scrubbed_tool_xml now carries removed_fragments, so a caller can
confirm the strip from the envelope instead of re-reading the stored
body. It reported a count alone, and the re-read it asked for usually
found nothing missing.
workspace_search states that caller_cwd picks any tree to search (ANTS-4928)
Pointing caller_cwd at another project, or at a path outside any repo,
searches that tree. That was documented only inside the lane
argument, so the verb read as project-scoped.
project_query names file enumeration where the tool is chosen (ANTS-4929)
Its one-line hint now covers listing files by name via project.list,
and project.list leads the API list. Previously the hint described
counting inside file contents, so the enumeration case was never
found.
a refused absolute path now names the way to read that file (ANTS-4899)
A file handed to a session from elsewhere on the machine was refused by
file_outline and read_region with nothing else to go on, so the
fallback was reading the whole thing raw — the one thing those tools exist
to avoid, and worst on the large documents where it matters. The boundary
has not moved: there was already a way to read such a file, and the
refusal now names it. Reads only, absolute paths only, and only for a file
that exists — a relative ../ is still a traversal and is still answered
with nothing.
workspace_search now says what its enclosing-symbol annotation managed (ANTS-4901)
Asking for the annotation over files whose language has no outline
returned every row unannotated and said nothing, which is
indistinguishable from the documented case of a match sitting above the
first symbol — except that here no row in the file can ever be annotated.
The envelope now carries enclosing_annotated and, when it applies,
enclosing_symbol_unavailable naming those files. Both ride the opt-in.
The spec standard fixes one review-log table shape, and the spec skeleton now carries it (ANTS-3682)
Review logs had drifted into many incompatible column layouts because
the spec skeleton handed authors a placeholder where a table should be.
The skeleton now ships the header row, and § 5.7 fixes the columns: one
count per review question, n/a where a genre does not ask one, and the
outcome cell last. Existing logs keep the shape they have.
roadmap_migrate's re-run refusals carry the project's registration as fields (ANTS-4893)
Its description names the dry run as the read-only way to ask "is this
project already migrated?", but both re-run guards fire before the
preview is built — so on a project whose stored export_slug was not
derived from its directory name, that question came back as a refusal
with the answer only in the message. The refusal now carries
project_id, store_backed, stored_export_slug and
stored_project_name. It still refuses: a preview that reached a
different verdict from the real call would be worse than an
unanswerable one.
Creating a level-2 roadmap section no longer swallows the level-3 sections below it (ANTS-4848)
roadmap_log op:"create_section" could adopt trailing subsections, producing
a section whose later appends landed somewhere unexpected.
Already in 0.7.108; recorded here late.
roadmap_log's description now names every op that refuses on a pass-headings roadmap (ANTS-4863)
It said only create_section refuses; amend_body refuses too, so a caller
planned around a limit the tool did not actually have.
Already in 0.7.108; recorded here late.
doc_integrity no longer reports false duplicate headings in a mirrored standard (ANTS-4878)
Heading numbers were counted across the MIRROR boundary, so every standard
carrying a project delta plus its upstream copy reported duplicates that
were not there.
Already in 0.7.108; recorded here late.
roadmap_query no longer reports one project's storage backend to another caller (ANTS-4885)
The backend witness was cached on the roadmap's path alone, so a second
caller could be told which backend a different project was using.
Already in 0.7.108; recorded here late.
spec_lint no longer blames a document for a check it did not run (ANTS-4889)
Its single-file path reported test coverage as unchecked and attributed that
to the spec, while the same spec passed the same check on the other path.
Already in 0.7.108; recorded here late.
roadmap_migrate now names the second format when a roadmap mixes two (ANTS-4492)
A mixed-format roadmap was classified by majority with nothing saying the
other format was present, so the items written in it were silently
reinterpreted.
Already in 0.7.108; recorded here late.
An unknown MCP tool now reports the same code in the envelope and in telemetry (ANTS-2173)
The refusal envelope said unknown_tool while telemetry recorded
tool_not_found, so the two could not be correlated.
Already in 0.7.108; recorded here late.
The changelog and test-audit parsers now use the shared CommonMark fence rules (ANTS-4404) Both hand-rolled their own "is this a code block?" test, and both got it wrong in ways that can swallow the rest of a document — a line quoting fence syntax opened a block that never closed. Measured before changing anything: neither fault currently bites on this project's own files, so this closes a latent defect of a class that has already caused two real bugs.
ants.notify() now shows a notification instead of doing nothing (ANTS-4273)
The plugin API's notification signal reached the plugin manager and stopped
there — the call was accepted, raised no error and had no effect. It now
goes to the desktop notification service, falling back to the status bar
when that does not take it. Plugin authors: the "currently a no-op" note
is gone from PLUGINS.md.
The Ctrl+Shift+V paste handler guards the clipboard pointer Qt documents as nullable (ANTS-3831)
All three paste branches dereferenced QClipboard::mimeData() without a
check. A null now makes the paste a no-op instead of undefined behaviour.
No crash was reproduced — the documented trigger does not apply to this
call site — so this closes a documented contract, not an observed bug.
A test run can no longer write to the developer's real git repository (ANTS-3841)
Both test-bundle entry points now scrub the git environment variables
that redirect which repository git acts on, and the shell tests that run
git do the same. Without it, a session with GIT_DIR exported had its
git-fixture tests operate on the real repo — one such run rewrote local
main onto fixture commits and left stray branches behind.
The pre-push gate now passes from a git worktree (ANTS-3842)
Git hands a worktree's hooks an absolute GIT_DIR, which reached the
test suite and pointed its git fixtures at the real repository, so the
gate failed for a reason that appeared nowhere in the diff. Fixed by the
same scrub as ANTS-3841.
The pre-push ASan gate now resolves a pending CMake regeneration and gates on the real edge count (ANTS-4536)
A changed CMakeLists.txt leaves a pending regen that hides every
pending build step behind it, and the gate stood the sanitizer leg down
whenever it saw one — so the leg went dark on exactly the pushes most
likely to need it. The hook now runs the regen (a CMake re-run, not a
build) and measures what it reveals.
The pre-push interrupt-marker skip now states how old the marker is (ANTS-4943) The marker that disables the sanitizer leg after a killed build never expires, and its skip message read identically whether it was written minutes or days earlier. It now reports the marker's age, so a long-dark gate is visible in a run that still ends "push allowed".
ANTS_PREPUSH_NO_ASAN can now be used for the push it documents (ANTS-4883) The hook's own test inherited the variable from the caller, so exporting the documented escape hatch made the suite the hook runs go red — the push was refused either way. The test now scrubs the hook's tunables and sets only what each case exercises.
Three MCP flags whose false is the answer now survive terse mode (ANTS-4956)
file_in_sync, store_backed and markdown_rewritten were emitted
correctly and then dropped as dead weight, because terse responses are on
by default. file_in_sync was therefore present exactly when the roadmap
was in sync and absent when it had drifted; markdown_rewritten is always
false and so never appeared at all.
roadmap_migrate op:"init" no longer writes a link that cannot resolve (ANTS-4962) The generated roadmap names the format standard in prose. A project bootstrapping one has no docs/standards/ to link to, so the link was dead in the file the project is told is its source of truth.
feedback_query's path help no longer names a folder holding no feedback file (ANTS-4980)
It points at derivation instead. The corpus root is configurable, so a
literal path in the help is wrong on any machine that moved it.
changelog_log op:add writes at the top of a mixed changelog instead of into its tail (ANTS-4563)
A ## [Unreleased] whose dated topics lead over a legacy flat tail fell
between two guards that answered its shape in opposite directions, and
the entry landed in the gap — appended into the tail, reported ok:true,
measured on one project at roughly ten thousand lines below the newest
entry. Both guards now read one classifier, and op:"add" routes such a
section to the dated-topic form at the top, saying so with
routed_to_subsection. A direct call to the flat writer refuses the new
mixed_section rather than burying anything.
spec_lint says whether a project ADOPTED the global spec-format standard or adopted none (ANTS-4926)
sections_source prefixing a hit with ~global/ said which standard
answered and nothing more. A conforming project keeps no local copy —
the standard forbids one and sends deltas to
docs/standards/spec-format-overrides.md — so adoption looked identical
to a project with no format standard at all, and a caller dropping
section findings on that prefix dropped real ones.
sections_source_reason now separates the states, with
sections_source_hint on the three arms a caller can misread.
Resolution is unchanged: the overrides file is reported, never read.
The MCP usage figures now count a call by what actually happened to it (ANTS-4457) Two errors cancelled each other out badly. A call that returned "nothing changed since last time" — the cheapest possible result, where Ants sends no data at all — was filed under failures, which is the column meant for wasted effort. And a call a tool rejected for bad input was filed under successes, because only rejections made by the dispatcher itself were counted. Both are now filed correctly, and the diagnostic trace names the rejection instead of reporting "ok".
Ants no longer re-reads its settings file about eight times a second (ANTS-4457) A background check that watches for the Claude "Switch model?" prompt re-opened and re-parsed the whole settings file on every one of its polls, and those polls run continuously by default whether or not Claude is even running. It now reuses the copy already held in memory, which is refreshed whenever the file changes on disk — so changing a setting still takes effect as before.
Claude status no longer gets stuck needing a tab switch to wake up (ANTS-4457) Ants finds the Claude session's log file to follow what Claude is doing. It looked exactly once, right after spotting Claude start — and Claude often has not written that file yet at that moment. When the first look came up empty, Ants never looked again for the rest of the session, so the status stayed blank and every update from Claude was ignored. Switching tabs and back was the only cure. It now keeps looking until it finds the file, and stops as soon as it does.
Permission prompts and failed tool calls now reach the status bar (ANTS-4457) Ants had working code to show when Claude is waiting for you to approve something, and when a tool call failed — but it never asked Claude to tell it about those two things, so neither ever happened. Both are now requested. If you installed the Claude hooks before this version, Settings will stop saying "Hooks installed" until you press the button again; that is accurate rather than a new fault, because the two new ones genuinely are not set up yet.
Resetting the terminal no longer throws away your theme and scrollback setting (ANTS-4456)
Some programs reset the terminal when they finish, and reset does
it on purpose. That used to snap the text colours back to the
built-in ones and shrink your scrollback back to the default —
while the cursor and selection colours stayed on your theme, so you
were left with a mismatch. A reset now clears the screen and history
as it should, and leaves your settings alone.
Selection highlight no longer runs past the end of Chinese, Japanese, Korean or emoji text (ANTS-4456) Characters that take up two columns were being counted as three when the highlight behind them was drawn. The error added up, so a selection containing several of them spilled further and further to the right of the text it belonged to.
Choosing a new bold or italic font now takes effect straight away (ANTS-4456) Text already on screen kept drawing in the previous font until something else forced a redraw — resizing the window, or changing the font size — so the setting looked like it had been ignored.
Opening a file by name from the terminal can no longer consult the wrong process (ANTS-4456) The terminal looks up your shell's current folder to resolve a short filename. The record of which shell to ask was being updated on one thread and read on another without coordination, so just after a shell exited the lookup could land on an unrelated program that had been given the same identifier.
A terminal whose shell has gone away no longer pins a CPU core (ANTS-4456) When a write to the terminal failed for good — the shell reaped, the connection torn down — the code that retries queued keystrokes kept being woken, kept failing, and never stopped, burning a whole CPU core silently for as long as the window stayed open. It now gives up on the first unrecoverable error. Keystrokes discarded that way are also reported in the log, as they already were on the two other paths that can drop them.
Sending a multi-line scratchpad now submits the text, and only the text (ANTS-4456) Sending from the scratchpad sent the Enter before the paste confirmation was answered. The shell received a bare Enter first, running whatever was already typed at the prompt, and the scratchpad text then arrived without being submitted; cancelling the confirmation did not take the Enter back. The Enter now travels with the paste, so it is sent when the paste is and not at all when it is cancelled. Multi-line sends always took this path, since a newline is what raises the confirmation.
Sixel decoding is now bounded by a whole-payload work budget (ANTS-4456) Displaying a file containing a crafted Sixel image could hang the terminal. The decoder now stops and reports an inline error when a payload's work exceeds a budget derived from the image it declares, alongside the existing dimension and image-size caps. Ordinary images, including multi-colour ones that redraw a row once per colour, are unaffected.
session_orient counts findings that carry no Proposed ID slot (ANTS-4941)
A feedback finding with no **Proposed ID:** line at all is a different
state from one whose slot is blank, and the session-start summary could
see only the second. Files whose sole pending input is slotless now list,
under their own suspected_untagged_count.
doc_citations caps an ambiguous quotation's candidate list (ANTS-4939)
The citation path has capped its candidates since ANTS-3636; the quote
path returned every same-named file in the tree. All three ambiguous
arms now share one emitter and report candidates_total and
truncated_candidates.
a checker that printed its help text is no longer reported as 92 findings (ANTS-3846) When a code checker refuses to run it prints its usage banner and exits successfully, and the audit read that banner as its output — so one report announced 92 findings whose contents were lines like "--help Display available options". A tool that analysed nothing looked exactly like a tool that found problems, which is worse than the tool being missing: missing reads as missing, this read as coverage. Such a run is now reported as an incomplete tool. The markers are matched only at the start of a line and only near the top of the output, so a genuine finding that mentions usage is still a finding.
one refusal code for a bad kind, whichever route reported it (ANTS-4750)
Setting an unrecognised kind of work was refused under one name through
roadmap_log's own argument and a different one through
op:"amend_field", so a caller that handles the first and re-prompts for
a valid kind silently mis-handled the second. Both now report bad_kind,
and the refusal lists the real kinds — taken from the same source the
check itself uses, so the two cannot drift.
a batch of changelog entries now reads in the order it was written (ANTS-4854) Each entry is inserted at the top of its category, so applying a batch front-to-back left the last entry on top and the release notes came out backwards from the list their author wrote. The entries are now applied in reverse, which puts them on the page in input order. This replaces an earlier contract of being byte-identical to the same entries added one call at a time — a property nothing checks, where the order on the page is one every reader sees. The reports back to the caller are unchanged and still follow the input list.
"where is this defined?" no longer lists local variables as definitions (ANTS-4880) A function-local variable whose value continues on the next line was reported with the same kind as the real definition, so a symbol appeared to have several homes and a reader could be sent to any of them. Such lines were already reported as declarations when they fitted on one line; they are now treated the same way when they wrap. They are still listed — knowing where a name is declared is the point — they just stop claiming to be the definition.
a feedback corpus in its own folder no longer causes a second, empty record (ANTS-4900) The guard that stops a derived feedback file landing where no maintainer sweep reads it only looked UP the tree. Move the corpus into a folder of its own — which is what happens once a machine carries two dozen of these files — and it sits BELOW instead, so the derivation read as "this is the first file in a new corpus": the read answered "nothing filed yet" and the write created an empty second record and reported success, while the real file went on sitting one directory down. The derivation root's children are now probed too, and the refusal names both the corpus it found and the setting that would point derivation at it.
cross-project id resolution survives the feedback corpus moving into its own folder (ANTS-4905)
feedback_query resolves a foreign-prefix id against the project whose
roadmap owns that prefix, and found that project by scanning the feedback
file's own directory for sibling projects. That held while the corpus WAS
the shared root; once it moved into a folder of its own, the scan found
no projects and every id fell back to foreign_repo — measured at 64 of
64 from a contributor project. It now searches the corpus directory and
then its parent, nearer first. Ownership is still decided the same way,
so a prefix nobody owns is still reported as foreign.
test bundles no longer read the developer's live settings (ANTS-4898)
Both bundle mains sandboxed XDG_DATA_HOME so a test could not open the
real roadmap store, and left XDG_CONFIG_HOME alone — so every test
process read ~/.config/ants-terminal/config.json. Setting one real
setting (the shared feedback corpus root) turned twelve tests red, and
their writes went to the real corpus rather than the temporary directory
each of them asserts against. Both mains now redirect the config
directory too, per process; a test that wants its own config still
overrides it.
session_orient's feedback backlog now follows the configured corpus root (ANTS-4896)
The feedback_pending block scanned the parent of the project root and
nothing else, while feedback_query had consulted
claude.mcp_feedback_root since ANTS-4471 — so moving the shared corpus
to a folder of its own left the session-start summary reporting nothing
waiting rather than saying it had looked somewhere else. It now scans
both roots, names the declared corpus as shared_root, and carries
searched — the directories actually read — so an empty result says
where it looked. The scan moved into its own builder, which is what let
it be tested at all: the bundle verb refuses without a window.
spec_lint checks a project against its own spec standard, not a mirrored copy of the global one (ANTS-4895)
Where a project keeps a verbatim mirror of the global spec standard
beside its own, the mirror was consulted first and the project's own
required-section list was never read — so specs were graded against a
rulebook the project had deliberately replaced. A project shipping no
standard of its own is unaffected.
changelog_log's schema no longer tells callers that summary is ignored under add_from_roadmap (ANTS-4888)
It has overridden the cited bullet's headline since ANTS-4360, with
category, id and the Layman body still inherited, but the schema still
described it as ignored. A session read that, took the roadmap headline
into its release notes, and hand-edited the file afterwards — the
round-trip the op exists to remove. Three descriptions corrected; the
behaviour was already right and is unchanged.
roadmap verbs run from a git worktree are served by the project's store instead of patching the worktree's copy (ANTS-4887)
The store keys a project on its main checkout, so a worktree was a path
with no entry: reads answered from the worktree's ROADMAP.md, which is
generated output, and a write patched that file, reported ok:true, and
was erased by the next render from the main checkout. A worktree now
resolves to its main checkout — read out of the .git file git already
writes, with no subprocess — provided that checkout really holds the
project's roadmap. Anything else is left exactly as it was.
project_conventions no longer states a convention whose source document is not in your project (ANTS-4460)
The rule table is curated and cites Ants Terminal's own documents, so on
another project it stated this project's conventions as that project's —
one report was told to write a tests/features/<name>/ conformance test
by a project that keeps unit tests elsewhere. A rule whose cited document
is not present under the caller's root is now dropped and listed in
conventions_dropped, with a warning naming the missing documents and
sources_resolved saying how many were found. The only previous signal
was a per-row exists:false, which compact:true drops by design.
spec_log op:append_loop refuses instead of writing an empty row and discarding the caller's cells (ANTS-4886)
The verb finds the log by its ## Cold-eyes loop log heading. Against a
file whose table sat under a different heading it fell through to the
bullet arm, which renders from label and body alone — so a call
passing only cells wrote the literal row - ** ** — , dropped every
cell, and returned ok:true. It then appended the heading it had failed
to find, below the existing table, so the next call wrote into that stub
and the file ended up with two logs. It now refuses in all three cases,
and still creates the section for a spec that genuinely has no loop log
yet.
roadmap verbs called from a project subdirectory are now served by the roadmap store (ANTS-4884)
The store-vs-markdown dispatch asked the store by caller_cwd, and a
project is keyed on its canonical root — so from a subdirectory the
lookup matched nothing and the verb fell back to reading and splicing
ROADMAP.md, which is an output of the store. Reads answered from a
file that could lag the store; writes reported ok:true and were
discarded by the next render. Both dispatchers, the render's
containment check and the project-row lookups now resolve the caller's
directory to the project root first.
roadmap_log's id allocator now applies its store floor when called from a project subdirectory (ANTS-4882)
The floor that stops a migrated project reissuing an id the store
already holds asked the store by caller_cwd. A project is keyed on
its canonical root, so from a subdirectory the lookup matched nothing
and the floor did not apply. Both op:"append" and
op:"append_batch" now ask by the directory the roadmap was resolved
under, which is the registered root in every layout the resolver
supports.
doc_citations counts each quotation once when the document contains a fenced code block (ANTS-4697)
A blank line after a closing code fence used to open a scan window of
its own, so every quotation in the paragraph below it was counted twice
— inflating quotes_checked and the quote_counts buckets that a
reviewer reads.
A fields=-narrowed MCP reply that matched nothing now still says it succeeded (ANTS-4877)
Asking a verb for a field it does not carry returned only the list of names it
could not find, with no ok — indistinguishable from a call that produced no
envelope at all. Such a reply now carries ok. A request where at least one
name matched is unchanged: the matched field already proves the call worked.
Mirrored standards no longer carry dead links (ANTS-4875)
A link a mirrored standard cannot carry — the owner's review history, a
foundation document, a skeleton — is now copied down as plain text instead of
as a link, so it no longer 404s for a reader on GitHub. The path is still
named. tools/check-standard-mirrors.sh does the de-link inside the same step
that feeds its drift check, so the copy stays derivable from its owner and no
mirror is hand-edited.
The published 0.7.107 RPM now builds on Fedora 44, via a backport carried in the spec. (ANTS-4870) Fedora's package repository was stuck one release behind because 0.7.107 could not build there, while openSUSE, Leap and Mageia were unaffected. The already-released 0.7.107 now carries the fix as a packaging patch, so Fedora users no longer have to wait for the next release. The patch applies only to 0.7.107 and older and is skipped from 0.7.108 onward, which carries the fix directly.
cut-rc.sh hotfix --continue resumes when the emergency release is already tagged. (ANTS-4872)
An emergency release interrupted between tagging and writing its notes back to the main branch could not be finished by re-running the command; it refused, and the last steps had to be done by hand. It now recognises its own tag — the one sitting on the work in progress — and completes the remaining steps. A tag of that name made by anything else is still refused.
cut-rc.sh cycle no longer skips its second phase after an interrupted release cut. (ANTS-4871)
The weekly cadence command would report success having done nothing, for the same reason new-rc used to refuse: it read the notes already moved into the version's own section as an empty changelog. It now looks for that moved work — but only where the version is not already released, so the ordinary run still skips cleanly instead of failing.
cut-rc.sh new-rc resumes after an interrupted run instead of refusing. (ANTS-4869)
Cutting a release candidate first moves the pending notes into the version's own section, then builds. If the build was interrupted, re-running said there was nothing to release — reading the move it had already made as an empty changelog. It now recognises its own completed work and carries on, and still refuses a genuinely empty cut.
cut-rc.sh now names the release phase it abandoned when a pre-commit hook refuses one of its commits. (ANTS-4865)
The run already stopped, but said nothing: the only text on screen was the hook's own refusal, which says nothing about the release, and a run piped through tail loses the exit status as well. It now reports which phase aborted, what did not happen (no tag, nothing pushed), and that the staged files are left as written so you can see what had been changed.
Harder to forge the boundary around tool results (ANTS-4457) Results Ants sends to Claude are wrapped in a marker so Claude can tell where the data ends. Text inside the data that looked like that end-marker was being blanked out — but only in its plainest form. A slightly dressed-up version slipped through, which a file or a commit message could use to make the rest of its content look like it came from Ants rather than from the file. Both forms are now blanked.
Project folder paths from Claude's session files are now checked (ANTS-4457) Ants reads Claude's own records to work out where each project lives on disk. One of the three places it reads was checked for paths that climb out of their folder, and the other two were not — including the one it consults first. All three are now checked, and a project whose location cannot be trusted is skipped rather than shown.
Session logs and session recordings are now readable only by you (ANTS-4456) Turning on session logging, or hitting Record Session, saves a file holding everything the terminal printed — which can include keys or tokens a command happened to display. Those files, and the folders holding them, were being created with the system default that lets any other account on the machine read them. Both are now locked to your account only. Files saved by an earlier version keep the permissions they were created with; delete them or tighten them by hand if that matters to you. The three "Export..." menu items are unchanged, since you choose where those go and may mean to share them.
0.7.109 RC1 — Patron preview
This release shipped without written notes.
0.7.108
Roadmaps written as numbered passes can now use the database (ANTS-4803) A project whose roadmap is organised as numbered pass headings could be copied into the database but not written back out, so it could use none of the database features. It can now be written back out, which makes the database usable for those projects.
Review corroboration now spots the same problem found in several different files (ANTS-4814) When reviewers each cover a different part of the codebase they rarely cite the same line, so finding one recurring problem in several places used to read as no agreement at all. Those are now reported, grouped by the function they sit in.
Cross-lane review corroboration now reports near misses, and can group them on request (ANTS-4817) Two reviewers who find the same problem but quote different lines of it were previously reported as no agreement at all, with nothing to say why the result was empty. Those groups are now listed as near misses, and an optional tolerance merges them into one finding naming the span. Exact matching is unchanged unless that tolerance is asked for.
changelog_log and changelog_query can name a nested component changelog (ANTS-4812) A project that ships a second changelog for a bundled, separately versioned component can now write and read it through the verbs, with an optional path. Previously only the root CHANGELOG.md was reachable, so the nested file had to be edited by hand.
Shell scripts can be outlined, so Claude can see their shape without reading them whole (ANTS-4826)
Asking for the outline of a .sh file used to return nothing at all. It now lists the functions and the settings at the top, and recognises a script with no file extension — a git hook or a launcher — from its first line. Those scripts also become visible to the code search and review tools, which had been skipping them.
Writing a note instead of a filename in a roadmap entry's Evidence line now warns you (ANTS-4527) That field is split on commas, so a sentence written there was silently stored as several fragments. The write still succeeds — it tells you rather than refusing.
Documented how doc_integrity really behaves (ANTS-3652) It answers a path that does not exist with a clean pass rather than an error, and silently stops after a fixed number of documents. The per-verb notes now say so.
Committed a linter configuration for the helper scripts (ANTS-4801)
ruff.toml pins the checks to the ones that catch code that misbehaves, leaving formatting alone. Two real findings in the roadmap survey script were fixed with it.
spec_lint checks a spec's Invariants against its Tests section (ANTS-4623)
Catches a spec that declares ten contracts and remembers to test eight —
the two sections restate one set in different words, so they share no
token to search for and could disagree indefinitely. Coverage written in
notation that cannot be expanded, such as INV-1..7, is reported as
unverifiable rather than as a pile of gaps.
find_definition resolves C++ data members (ANTS-3668)
Asking where a struct field or an m_-prefixed member is declared used
to come back empty, though the answer was sitting in a header. Measured
on this project's own docs, the unresolved data-member share fell from
26% to 7%, which also lifts doc_symbols' signal-to-noise.
doc_symbols gains an only= row filter (ANTS-3689)
only:"unresolved" (or "not_checked") drops the resolved rows nobody
reads, which is what let a single dense document spill past the response
cap. counts stays whole-document and symbols_filtered_out reports
what was withheld; an unrecognised value refuses rather than widening.
Shell scripts are checked automatically, in CI and locally (ANTS-4518) The project ships a lot of load-bearing shell — git hooks, build and release tooling, test drivers — and nothing checked any of it, so a committed hook had silently lost a suppression its installed copy carried with no gate to say so. The gate blocks rather than reporting, because the surface is clean at the severity it runs. It covers the two hooks that carry no .sh extension, which are the scripts the original finding was about.
Review lanes over a test tree are labelled as such (ANTS-4806) A computed partition groups by directory, so the tests become a lane like any other — and a code review's scope excludes them while a test review's is exactly them. The lane is labelled rather than dropped, so each consumer takes what it wants instead of the tool choosing.
Review lanes report how many lines they cover, not just how many files (ANTS-4804) A lane holding one very large file counted as one file and tripped no size threshold, so the biggest lanes were the ones least likely to be flagged. Six projects reported it independently. Each lane now carries a line count and an oversized signal alongside the existing file-count one, which keeps its meaning.
A way to rewrite a roadmap entry's text from scratch (ANTS-4808) Editing an entry's prose could only replace a matched piece of it, so text that could not be matched had no route back — one project carried a kilobyte of garbled prose in a shipped entry for that reason. roadmap_log op:"set_body" replaces the whole body. It records the previous text in history and reports how much it replaced, and it refuses an old_text rather than ignoring one, since a caller who sends it believes the write is bounded by it.
An .editorconfig recording the layout this project actually uses (ANTS-4790) Spaces, four wide, LF, UTF-8, trimmed trailing whitespace and a final newline, with two deliberate carve-outs: markdown keeps trailing whitespace because there it is a line break, and the roadmap round-trip vectors keep both rules off because they back a byte-identical comparison. It does not declare a shell style — no setting makes shfmt agree with these scripts, so arming that tool would mean reformatting every one of them.
A spell-check vocabulary file, so typos is usable on this tree (ANTS-4789) A whole-tree run was almost entirely false positives — identifier fragments, VT100 notation, git SHA prefixes and namespace aliases read as misspellings. The config names the project's vocabulary rather than silencing the rule, so genuine misspellings still report.
A committed clang-tidy check set, so the C++ analyser actually analyses (ANTS-4787) With no configuration clang-tidy selects an empty check set: it reads every file, reports nothing, and exits successfully, which is indistinguishable in a log from a clean tree. This project's own audit_run had been invoking it that way. It now runs against a declared set of checks.
Conformance test holding the committed review partition to every source file (ANTS-4794)
The build now fails when a file under src/ belongs to no lane in
.indie-review/partition.json, when a lane names a path that has been
renamed away, or when two lanes claim the same file. The reviewable-file
set comes from the review engine's own walk rather than a second copy of
its filters, so a change to what counts as reviewable moves the test with
it.
Every source file now belongs to a review lane (ANTS-4793)
A committed .indie-review/partition.json declares 32 lanes covering all 313 src/ files exactly once. The module map's prefix rule had left 189 files and 59,257 lines — mainwindow.cpp among them — in no lane at all, so a full-codebase review silently skipped a third of the tree.
Release gate: a CHANGELOG bullet that merely repeats its roadmap headline is now reported (ANTS-4759)
A defect item's roadmap headline states the problem, so a summary copied from it puts the bug in the release notes where the fix belongs. tools/check-shipped-coverage.sh now flags any [Unreleased] bullet whose bold summary is byte-identical to the item's stored headline — 14 of 56 were, across seven commits, and all 14 have been rewritten to say what shipped. Advisory like its sibling check, so it cannot block a release over wording.
spec_lint accepts fields= and compact (ANTS-4663)
A corpus run previously spilled and had to be parsed by hand.
roadmap_log op:"amend_headline" rewords a headline on a store-backed project (ANTS-4683) A headline was effectively immutable after creation while body notes accumulated a correction the headline went on contradicting.
roadmap_log op:"render" publishes the store's canonical ROADMAP.md (ANTS-4509) Normalising a file used to require inventing an unrelated write purely to trigger the render, which left a roadmap bullet nobody wanted.
A read-only way to ask whether a project is in the roadmap store and in sync (ANTS-4488) Previously only a dry-run answered it, as a side effect, at the cost of its whole note list.
Large MCP read bodies spill to a scratch file instead of flooding the reply (ANTS-2094) The response carries the head plus a pointer, so a big read no longer costs its full size in context.
roadmap_query mode:"headline_only" returns lean rows (ANTS-1881)
{id, status, headline_oneline, section_slug} per bullet, skipping the body — roughly a tenth the payload on a dense section.
feedback_log op:"set_format_version" corrects a file's format marker (ANTS-4707)
A feedback file declaring a version that was never defined could only be
fixed by the hand edit the assign-id pipeline exists to avoid. The marker
is a claim about the content, so the write happens only where the content
already matches — stamping 2 needs the inline-**Proposed ID:** model,
stamping 1 needs its absence, and either way round refuses with a hint
saying which. Refuses a version above what the build can read, and refuses
a file with no marker at all: building that shape is migrate_v2's job.
An already-correct marker reports no change, so a corpus sweep is
re-runnable; dry_run previews.
The pre-push docs-only set is checked against ci.yml's paths-ignore, and the skip decision is recorded (ANTS-4726)
The hook's docs-only path list was a hand-maintained twin of the
workflow's paths-ignore that nothing compared, so drift would let it
skip a gate CI then runs. A static test now asserts the two name the same
paths in both directions, and that the regex is anchored. The hook also
prints the ref range and the paths its skip was decided on — previously a
wrong skip and a right one printed the same line and exited 0. The
misclassification that prompted this is still unexplained and the item
stays open; the log is what a recurrence needs.
doc_citations resolves a backticked path::symbol to the symbol's line (ANTS-4748)
Only where the path resolves and the symbol is unique in that file; the
entry carries resolved_via:"symbol", since its start_line was looked
up rather than authored. An overload, a name that has moved, or an
unresolvable path resolves to nothing and is counted as unchecked — a
false ok is worse than the admitted silence ANTS-4743 added. The
unqualified suffix of a qualified definition is matched too, which is how
the corpus writes these; two definitions sharing a suffix are ambiguous
and refused.
invariant_check accepts paths as an alias for files (ANTS-4744)
invariant_check now accepts paths as an alias for files,
the spelling its sibling multi-path verbs use, so the argument name a
caller carries across works here too. files wins when both are sent,
and the refusal for neither names both spellings.
invariant_check says why a zero match can be wrong (ANTS-4742)
invariant_check returning zero matches now says why it can
be wrong. Matching is by file path only, so a spec naming the module by
symbol was reported as no spec at all — the reading that sends a
session to write against no contract. The zero case now carries
path_match_only and a hint naming the fallback search. Matching
itself is unchanged.
feedback_query does the stale-binary comparison itself (ANTS-4741)
feedback_query now does the stale-binary comparison instead of
telling you to make a second call for it. A shipped id whose ship date
is not older than the running server binary's build date comes back
with possibly_stale_binary, alongside the build date and commit. A
same-day ship sets it, because the ship date carries no time.
roadmap_migrate op:"init" bootstraps a project that has no ROADMAP.md (ANTS-4740)
roadmap_migrate op:"init" bootstraps a project that has no
roadmap file at all. It writes a conforming skeleton with one
## Backlog section and then registers the project the ordinary way,
replacing the hand-authoring step a greenfield project used to need —
where a mis-cased trailer label would migrate cleanly while silently
dropping fields. Refuses roadmap_exists rather than overwriting;
dry_run:true returns the skeleton without writing.
A document can opt out of spec_lint's required-sections check (ANTS-4739)
A document may carry spec-lint: no-required-sections
to opt out of the required-sections check -- for a mixed corpus holding
pre-standard specs, build plans and ledgers that can never conform. The
number of exempted documents comes back as sections_exempt_docs, so
an exemption is never mistaken for a clean pass.
spec_lint can match a required heading by prefix, so a descriptive suffix passes (ANTS-4738)
A format standard may write its marker as
required-sections: prefix, which matches a numbered heading on its
number and name and lets a descriptive suffix pass. Exact matching
stays the default.
roadmap_query can list WHICH items shipped in a date window, not just how many (ANTS-4715)
shipped_since / shipped_until narrow mode:"bullets" and
mode:"headline_only" to items the store dates inside a half-open window,
so the release-time question — what shipped this cycle that the changelog
does not mention — can be asked through the verb. It refuses rather than
ignores the arguments wherever the branch that would honour them is not
the one that runs, and reports shipped_undated when the store holds
shipped items with no date, since those can appear in no window at all.
apply_edits takes match_wrapped, applying an edit whose old differs from the file only in where its line breaks fall (ANTS-4723)
Opt-in and fallback-only — consulted just where the verbatim search found
nothing, so it can turn a refusal into an edit and never changes a call
that already succeeds. Uniqueness is enforced on it as on the verbatim
pass, and it declines a span whose continuation lines carry column
alignment or open a list item. When it fires the envelope says so, so a
re-flowed span is never silent. Same rule and spelling as
workspace_search's match_wrapped.
Release gate: work that shipped but reached no CHANGELOG entry is now reported (ANTS-4714)
tools/check-shipped-coverage.sh reads the roadmap store's shipped dates
and lists every item closed since the last public tag that no CHANGELOG
bullet cites. The existing gate only checked the opposite direction — that
entries claiming to have shipped really had — so work that shipped and was
never written down was invisible. The first run found 19, one closed the
same day as the release it was missing from.
amend_body's not-found refusal now shows the text it searched (ANTS-4807) The refusal echoes the stored body's tail and its length, and says plainly that this is the text being matched against and not what roadmap_query returns. A caller comparing their search string to that echoed tail gets in one call what three rounds of guesswork could not.
Right-clicking a tab now shows which colour that tab is currently set to. (ANTS-4862) The palette listed 26 colours and marked none of them, so there was no way to read the current setting with the menu covering the tab. The active entry is now ticked and bold; a colour picked from the custom dialog marks "Custom colour..." and shows itself as that entry's swatch.
Adding a test that reads a source file now rebuilds one file, not the whole bundle (ANTS-4797) Each test's file-path setting is attached to that test alone rather than to every file in its bundle. Measured on one bundle: adding a setting rebuilt 54 files before, and rebuilds 1 now.
cppcheck's advisory useStlAlgorithm class is suppressed, with its reason recorded (ANTS-4782) It asked for hand-written loops to become <algorithm> calls, printed in every run and was never acted on. Suppressed in both the workflow and the local parity script, each stating why and how to re-enable it.
The standards mirror gate now names the links it cannot check (ANTS-4825) A link out of a mirrored standard that points outside the mirror set is skipped by rule. The gate now names each such target and why, instead of reporting a bare "links resolve" over a tree where doc_integrity finds broken links. Both tools were right; only the reporting disagreed.
The document-integrity check accepts the same compact setting as its sibling tools (ANTS-4858)
It was the only one of that group that ignored the setting, so asking for a smaller reply there did nothing. Replies stay full-size unless you ask, because some of its empty fields carry meaning.
The project-setup check separates settings you can fill in from ones you cannot (ANTS-4815) A project that deliberately keeps no docs or specs folder was listed as missing them forever, with no way to clear the list short of creating folders it did not want.
CI now drops a superseded build on branches and PRs (ANTS-4799)
Pushing again to the same branch cancels the run the previous push started. Runs on main are never cancelled, so a cancelled build there still means something went wrong.
Symbol resolution answers many names in one tree walk, and skips lines that cannot match (ANTS-3680)
doc_symbols over the whole docs/specs + docs/standards corpus went
41.8 s to 3.4 s with identical output. A single find_definition went
410 ms to 81 ms on this tree; 200 names resolved together went 15.6 s to
0.35 s. New lib entry point SymbolQuery::findDefinitions, which returns
exactly what the per-name call returns.
session_orient offers project settings for any non-standard layout (ANTS-3353)
A project whose code sits where Ants expects but whose roadmap, changelog
or specs live somewhere unusual now hears that it can pin those paths —
previously only a project with misplaced SOURCE was told. The suggestion
carries every path it found, not just the source directories.
A gap in a spec's invariant numbering now reads as evidence, not a defect (ANTS-3684) A spec that continues a parent spec's numbering keeps its ids, and renumbering would break the citations — so the gap is real and the document is correct. The finding is now a candidate and names that cause, so it can be triaged without opening the parent spec.
CI jobs show readable names in the GitHub Actions UI (ANTS-4796) The three jobs carried no name, so Actions labelled each by its internal yaml key.
The cppcheck gate reads tests/ as well as src/ (ANTS-4788) Findings under tests/ were previously seen by no run. cppcheck cannot parse this project's test macros, so its syntax-error rule is suppressed under that directory; without that the widening reports a parse failure per affected file. The gate stays informational and does not fail a build. tools/ is not in scope because it carries no C or C++ source.
Two per-query tokenisers compile their search pattern once per process rather than once per call (ANTS-4776)
wrapmatch.cpp and findsources.cpp hold their constant QRegularExpression as static const.
The roadmap high-water scan is 41% faster on a large roadmap (ANTS-4426)
maxDeclaredId — the scan behind roadmap_query's file_ahead_of_store witness and the allocator's floor — ran a regular expression once per line of the file, 62,918 times on this project's 4.96 MB roadmap. It is now a character check with identical semantics: the scan drops from 22.6 ms to 13.3 ms, and a cache-miss roadmap read on a migrated project from ~26.5 ms to ~16.6 ms. Verified by verdict diff over 400 markdown files and 234,692 lines, with no verdict moved.
A roadmap note with no line breaks of its own is wrapped to match the corpus (ANTS-4532) It was written verbatim, so a caller who did not pre-wrap left one very long line in a file where every other body is wrapped. Prose carrying no newline has no structure to preserve, so it is wrapped at the width used around it; a note that brought its own line breaks is still written exactly as given. The wrap runs after the trailer-key guard, never before it — moving a word to the start of a line could otherwise manufacture a declaration the caller never wrote.
apply_edits says a transport timeout means UNKNOWN, not failed (ANTS-4733)
Measured: a batch reported a timeout, the files showed nothing written,
and the write then landed — so the obvious recovery of looking and
re-sending races a write still in flight. The description now says to
re-read before deciding, and why the unique-old form is the safe one.
Three source-scrape tests now anchor on the construct they mean, not on the first match (ANTS-4727) Each opened its search window on a string that occurs several times, and landed on the intended site only because of where the translation units sit in the build's source list. Two of the three would have kept passing from the wrong place. Each is now scoped to the function or branch under test.
Two unverified ANTS-1881 combinator claims are now checked against the code (ANTS-4725)
Both hold. The section etag keys on the section's bytes rather than on mode, so an etag taken under one mode is answered unchanged under another — the row now records that edge. And the list path's mode echo is gated, which is what § 2.4's byte-identical back-compat claim rests on.
A source-scrape test now anchors inside the branch it measures (ANTS-4724)
Inv2BypassesStatusAndPagination anchored on a string that also matches an earlier else if, so its byte window opened above the branch it names — a test that passes from the wrong position gives no signal when the thing it names moves. Re-anchored on a needle that occurs once, and proved by mutation in both directions.
Session preamble now reports when the installed hooks have drifted from the repo (ANTS-4516) Fixes to the helper hooks could be written, tested and green while the installed copies stayed a month stale, with nothing saying so. One preamble line now names the stale hooks and the command that installs them. It reports rather than writing to the hook directory at launch.
spec_log op:append_loop no longer corrupts a table-form review log (ANTS-3685) It appended a bullet after the last table row whatever shape the log was in. It now detects a table, takes the row's cells, checks them against the header's column count, and refuses rather than writing a row that breaks the table. Row order is inferred too, so a row lands at the end the log actually runs from. Shipped under ANTS-4364 and ANTS-4353.
Tab-colour menu now ticks the colour the tab is already on (ANTS-4864) The active entry gets a tick drawn into its swatch plus a checkmark on its label, so the current colour is findable in a 27-row menu. Qt's own checked state was rendering as a frame around the swatch, which looked like every other swatch.
Typing in the scrollback search box no longer freezes the window. (ANTS-2000) Each keystroke re-scanned the entire scrollback -- 50,000 lines by default -- so typing a word could lock the window for seconds. The search now waits for a brief pause in typing and scans once. Pressing the regex toggle still searches immediately.
Updating the compiler no longer leaves precompiled headers that can crash the build. (ANTS-2107) The build now records which compiler built its precompiled headers and discards them when that changes. Previously a compiler upgrade left large incompatible files in place, which warned and then crashed the compiler part-way through a build, and the only remedy was knowing to delete them by hand.
The test-audit dialog's "Per-finding" fold-in now works. (ANTS-4445) Choosing Per-finding reported a failure every time: nothing ever collected the findings it needed. The reviewer's reports are now read into findings and folded in with roadmap IDs allocated. When a report carries no findings in the expected form, the dialog says so and points at Narrative mode instead of reporting a failure.
Dismissing an audit finding as a false positive now actually hides it. (ANTS-4444) A dismissed finding was remembered, but it kept appearing in the results list, the summary counts, the HTML and SARIF exports and the roadmap fold-in, and was sent to the AI again on every triage. Every one of those places now checks the learned-false-positive list, and does so live, so a dismissal takes effect without re-running the audit.
A plugin that reaches its memory limit can no longer take the terminal down with it. (ANTS-4442) Sending an event to a plugin already at its 10 MB budget crashed the whole terminal, which is exactly what the plugin documentation promises cannot happen. The event data is now prepared inside a protected call, so the plugin is told it ran out of memory and the terminal carries on.
SECURITY.md now lists the link schemes the terminal actually accepts. (ANTS-4453) Hyperlinks are limited to http, https and mailto. The document promised ftp and file as well, which the code has never accepted at these sites.
Opening a file whose path contains a percent sign now passes the right path to the editor. (ANTS-4784)
Ctrl-clicking a file:line in terminal output built the editor argument
with a form that could substitute into the path itself. The same form was
corrected wherever an untrusted path, lane name or caller-supplied string
is placed into a message before another value.
Files that Claude Code changes are tracked again when the hooks are installed. (ANTS-4451) The hook path read the wrong field names from the event, so no file was ever recorded and nothing downstream fired. The transcript path was unaffected.
Audit dedup keys and learned false-positive fingerprints no longer collide on a file path containing a percent sign. (ANTS-4452)
Both keys are built from the file path first. The string-building form
used could read a %2 inside the path as a placeholder, letting two
unrelated findings hash to the same key -- so dismissing one could silence
the other.
roadmap_query now answers check_sync on every reply, including the section-inventory fallback. (ANTS-4860)
Asking whether the roadmap file matches the store could previously come
back with the field simply absent, which is also how the reply looks when
nobody asked. The fallback reply now says the check did not run and why.
Documentation now says where review lanes actually come from (ANTS-4785) The subsystem map described itself as the source of the review lanes, which stopped being true when every source file was given a lane in a committed override. Both files now state which is which, and which one to edit to change a lane.
A code fence inside one roadmap bullet now ends with that bullet (ANTS-4823) roadmap_log's walkers bind a fence to the body it was opened in, so a hand-written or legacy bullet that opens a fence and never closes it can no longer make every bullet below it unwritable. A fence opened in section prose still masks what follows, unchanged.
Projects whose item IDs start with a digit can name their own IDs again (ANTS-4849)
An ID like 3D_E-0682 was rejected by the one setting that exists to let you choose an ID, while the rest of the same tool accepted it. The only route left quietly depended on a counter file being correct. A plain number with no letters is still rejected, because nothing could tell it apart from an automatic ID.
A preview of a batch file edit no longer claims it wrote the files (ANTS-4834) Asking "what would this change?" replied as though the change had already happened, so a caller checking that reply could conclude the work was done and skip doing it. Previews now say what they would do. The matching tool for roadmap edits was fixed the same way earlier.
Asking for a roadmap section's items now tells you which sub-section each one is in (ANTS-4824) Every item came back labelled with the section you asked for, so a nested one looked like it lived in its parent. Filing a new item next to one of them would then have put it in the wrong place without saying so.
Cutting a release now reports the date it stamped on the heading (ANTS-4833) That date is what release-note tooling searches for, and it was the one thing the preview could not show you — left out, it quietly uses the machine's idea of today.
A roadmap note that quotes code-block markers is handled correctly again (ANTS-4450) The check that stops such a note running away and swallowing the entries below it had drifted apart from the reader that enforces the same rule. It could both miss a runaway note and insert a stray backslash into a note that was perfectly fine.
The roadmap tool's set_body operation is listed among its operations again (ANTS-4842)
It worked and was documented, but was missing from the list callers read to check what exists — so sessions concluded it was unavailable and left a garbled entry unrepaired.
The spell checker's settings now cover the noise they were written for (ANTS-4800) Findings dropped from 131 to 4, and the four that remain are documented as deliberate. Verified the checker still catches ordinary misspellings.
A roadmap entry can show its own format in a code block without that example becoming real data (ANTS-4526) Keys written inside a fenced block are now read as illustration, matching what already happened for keys quoted inline. Tilde-fenced blocks are covered too, which they were not before.
Roadmap entries below a certain point can be updated again (ANTS-4823) One entry contained an unfinished code block, which made the roadmap tool refuse every entry after it.
Both guides quote the README's real version line (ANTS-4540) They described it with wording the README stopped using, so a reader searching for it found nothing.
The contributor guide and bump recipe name the tools that still exist (ANTS-4530)
Both routed the version bump to a /bump command removed in August, and credited tagging to the wrong tool. They now name cut-release and packaging/cut-rc.sh.
The debt-sweep engine's design doc matches the shipped code again (ANTS-3510) Its detector list, scan order, apply-fix contract, dry-run path and perf budget had fallen behind several months of changes. The two places that listed detectors by name now point at the engine's own roster, so adding one no longer needs a doc edit.
find_definition tags a brace-initialised member declaration, so both spellings of one field agree (ANTS-4821)
A brace only opens a body when a parameter list, a lambda capture list
or a class-key precedes it; otherwise it initialises the declarator it
abuts. QTimer *m_t{nullptr}; now matches int m_x = 0;.
The Review Changes dialog refreshes when a file is saved in place (ANTS-2229) Closed by the watcher rewrite already shipped under ANTS-3509: edits to an existing file, and edits anywhere in a subdirectory, now refresh the diff without a manual Refresh.
A spec that quotes the test-clause marker parses whole (ANTS-3676)
The spec reader took the first occurrence of the marker wherever it
appeared, so an invariant naming it in backticks was cut in half at the
mention. It now skips a marker inside an inline code span. spec_lint
reads this parser, so its findings were least reliable on the specs that
discuss spec format.
The closer-length rule now reaches every reader that tracks fences itself (ANTS-4820) The docs index, file outline, spec lint, spec log and feedback-file scanners each kept their own fence loop and still ended a block on a short run. A document quoting code-block syntax no longer leaks its sample headings into the outline, the docs index, or a feedback file's untriaged delta.
Code fences now honour CommonMark's closer-length rule (ANTS-3678)
A short run of backticks or tildes no longer ends a block opened with a
longer one, so a document that quotes fence syntax keeps its sample text
— headings included — inside the block. Affects every reader built on
fenceMask, read_region's section resolution among them.
The bash veto no longer blocks commands that merely mention a git diff (ANTS-4517) The rule matched its phrase anywhere in the command, so an echo, a heredoc or a sed replacement carrying it was vetoed while running no git at all. The match now anchors to a command position; a later stage of a pipeline or a compound is still a real invocation and still vetoed.
roadmap_query section= now returns a heading's nested sections too (ANTS-4819)
A level-2 slug answers with its level-3 children's items, matching the
descendants mode:"section_index" already rolled into that slug's
counts. Store-backed projects previously got an empty read for any
parent heading while the index promised its children's bullets.
The citation scanner appears in the project's own code map (ANTS-3687) docs/subsystems.md listed all five of its siblings and not it, so the code map answered "no such thing" to a session asking where the citation checker lives.
Build and release workflows are visible to the review splitter (ANTS-4805) The walk never descended into .github at all, so a project's CI files appeared in no review lane and in no coverage report either — an absent lane and a clean lane look identical downstream. This session's own sweep found four real workflow defects, two security-relevant, on files no partition would have offered a reviewer.
A registered project is no longer told the store holds no record of it (ANTS-4802) Only one roadmap dialect is served from the store, while the migration reads three — so a project could be registered, hold hundreds of rows, and still be told there was no row for it, which sends the reader to re-run the migration, the one action that cannot help. The refusal now asks the store first, and the migration says what a non-served format costs rather than reporting a bare flag.
A spec id one tool hands out is now accepted by the tool that reads specs (ANTS-4810) invariant_check returns a spec's id as its filename stem, and on projects whose specs carry a topic suffix spec_query rejected that exact id — breaking the two-call lookup the coding workflow prescribes.
The roadmap summary view answers the in-sync question instead of ignoring it (ANTS-4818) Asking whether ROADMAP.md matches the store was accepted and silently dropped in report mode, while the documented rule says a missing answer means the question was never asked — so a caller who did ask read the silence as their own omission.
A roadmap entry's reply says which of its lines the tool wrote (ANTS-4813) An entry's text mixes an author's prose with lines composed from stored fields, and the two are edited by different operations. Nothing distinguished them, so passing a line back to be edited failed on text that had just been read. The reply now names the composed lines.
Lanes that count zero now say whether they are empty or unmeasured (ANTS-4816) A lane of packaging files, YAML or shell counted zero, because the count admits only file types the code index can read — and zero read exactly like an empty directory while being what the size warnings key on. Such a lane now reports how many files were there and not counted.
A project keeping all its code in one directory gets review lanes again (ANTS-4811) Every partitioner grouped by top-level directory and returned nothing unless grouping produced more than one group, so a project whose modules all live in one package got an empty partition that looked exactly like having no subsystems. Such a directory is now split by size, and the reply says which stage produced the lanes so "the map yielded nothing" is distinguishable from "there is no map".
Review lanes no longer cover files the repository ignores (ANTS-4809) The partition walk knew the conventional exclusions but never asked git about the project's own, so one project got fourteen of seventeen lanes over an ignored directory of scraped data. A lane is an instruction to read what it names, so this mattered more than the noise.
indie_review_partition now reports which files no review lane covers (ANTS-4786) It could only answer that for a partition it had computed itself. On the declared path — the normal one here and on every migrated project — the reply listed lanes and said nothing about what they left out. The envelope now carries the uncovered count, a per-suffix breakdown, a bounded sample of paths, and a reason saying whether the files were dropped by a filter or simply never named.
remotecontrol_state.cpp includes the Qt header it uses instead of relying on the module precompiled header (ANTS-4774)
The translation unit built only because the shared Qt PCH happened to supply QElapsedTimer; it now compiles on its own terms.
The command palette builds and positions itself on every reveal, not only when shown through a typed pointer (ANTS-4775)
It now overrides the virtual setVisible(bool) rather than shadowing the non-virtual QWidget::show(), which could otherwise yield an unbuilt, unpositioned and unfocused palette.
indie_review_partition now reports the source files its computed walk left out of every lane, so a review can no longer read as full coverage while skipping part of the tree (ANTS-4771) The computed partition admits only suffixes the codebase index can outline, which is narrower than "source" on purpose. On a mixed Python and shell project that silently dropped every shell file. The envelope now carries unassigned_count, unassigned_by_suffix and a hint whenever it hands back a computed partition.
The mirrored standards now include the C++, Qt and Python spellings that coding.md names as owning how this project's code is written (ANTS-4764) A reader with only the checkout could open coding.md, read that the language spellings live elsewhere, and have no way to follow it. What stays out of the mirror set is now a stated decision rather than an omission: a skeleton is a template to copy, not a rule to conform to, and docs/standards/README.md records the boundary.
The public copies of the shared standards now carry every sibling they link to, so their cross-references resolve for a reader who has only the checkout (ANTS-4761) The mirrored half of a standard is a verbatim copy, so a link in it pointing at a global file this repo did not carry simply dead-ended. The mirror set is now closed under its own links, and tools/check-standard-mirrors.sh fails on a link out of a mirrored half that this repo does not carry.
Two false-positive compiler warnings no longer appear on every build (ANTS-4544) GCC reported a possible null dereference on a line that checks for null, which it loses sight of when optimising. Scoped suppression, same as the three earlier instances of this class. Reproduced before being silenced and confirmed gone after.
The roadmap store's history budget counts the bytes it says it counts (ANTS-4503) Three units were in play — SQLite counted characters, the incoming size counted UTF-16 code units, and every name for the budget said bytes. They agree on plain ASCII and diverge on emoji status markers, em dashes and non-Latin text, so a real store could hold several times the history the cap names. All four measurement sites now count UTF-8 bytes. On such a store the effective budget is smaller, which is the documented intent, and reaching it still writes no history and never aborts the operation.
Asking for fields= no longer switches compaction on as a side effect (ANTS-4673)
The two shared one allowlist, so narrowing a reply also applied a transform nobody asked for — which folded away spec_lint's flag saying a check had never run.
An odd number of backticks no longer inverts code-span masking for the rest of a body (ANTS-4598) Everything after the stray backtick was treated as the opposite of what it was.
counter_advanced_past no longer names a position the allocator has passed (ANTS-4579) The one field a caller would predict the next id from was knowably false.
op:append_batch reports the counter fields op:append already emitted (ANTS-4551) Their absence read as evidence that the counter had not moved.
A dry run on a store-backed project reports the id it would allocate (ANTS-4592)
ANTS-4508's would_be_id never reached that path.
A roadmap_log dry run no longer looks like it reserved an id (ANTS-4508) It returned the next id under the same key the real write uses, so a caller reading that field took it for a reservation. Nothing is reserved, and the id is wrong if any other write intervenes.
roadmap_migrate's routine re-import is no longer thousands of tokens of noise (ANTS-4408) Every call returned about a thousand notes whether or not anything changed.
A rolled-back roadmap_migrate no longer describes writes that did not happen (ANTS-4144) It emitted its full note list regardless, spending thousands of tokens on a transaction that was undone.
A project that has allocated no ids is not reported as diverged (ANTS-4430) Its zero high-water mark was indistinguishable from the divergence that blocks writes.
A migrated project no longer reports a false divergence (ANTS-4410) roadmap_migrate left the store's high-water mark at zero, which reads as the state that means "do not write".
The store-divergence warning no longer fires forever (ANTS-4406) It compared two different quantities, so it warned on every project and both of its numbers were wrong.
roadmap_query names the backend that actually answered (ANTS-4143) It answered from the store while the envelope named ROADMAP.md, so a caller could not tell which document it had read.
Specs no longer cite the pre-split remotecontrol.cpp or the deleted isFieldProjectionTool (ANTS-4711)
Roadmap-query code is attributed to the translation unit that holds it,
and the stale line pins are dropped rather than re-derived — every one
landed on unrelated code or past the file's end. The file was corrected
only where the subject moved: dispatch() and the rc* helpers those
specs cite really did stay put. Instructions to enrol a verb in
isFieldProjectionTool now say there is nothing to enrol in, ANTS-4524
having removed the fields= allowlist; two statements asserting the
deleted symbol still exists are gone, which is what invited re-adding it.
Loop-log rows are untouched — they record what a review pass found on a
date. CLAUDE.md's own applicationDirPath citation is corrected with them.
changelog_query's bad_mode refusal names the version argument (ANTS-4754)
A caller after one version's entries reaches for mode:"unreleased",
reads an accepted-mode list that does not contain it, and concludes the
verb cannot filter by version — when version:"Unreleased" has done
exactly that all along and composes with every mode. The refusal now says
so and gives a value that works. No filter was added; the capability was
never missing, only unfindable from the reply that turned it down.
workspace_search explains a zero caused by its own hidden-path filter (ANTS-4753)
A glob naming a dot-directory (.github/workflows/*.yml) returned no
matches and no reason, because include_hidden defaults to false and
ripgrep prunes the hidden directory during the walk — indistinguishable
from the string genuinely being absent. Such a zero now carries
hidden_paths_skipped:true and a hint naming include_hidden:true, on
the match list and on the count_only, files_only and matches_only
replies alike. A call that returns rows is unchanged. A lane naming a
hidden directory is searched as before and does not flag it.
The Arch and Debian packages now declare ripgrep, so their test suites run instead of failing (ANTS-4717) The guard that catches this class was reading only the CI workflow, so the same missing declaration was still live in two package recipes. It now reads every recipe that runs the test suite, and matches a declaration rather than a mention.
doc_citations says whether a zero is a clean run or a silent one (ANTS-4743)
doc_citations returning zero now says whether that is clean
or silent. It recognises only path:line citations, and the spec-format
standard tells authors to cite the symbol instead — so a conforming spec
corpus scanned to zero and read as checked. Every reply now names the
forms it recognises, and a zero over a document holding citation-shaped
text reports how many it could not read, with a hint not to record the
run as a passing check.
spec_lint counts are taken over the whole scan, not after the cap (ANTS-4737)
spec_lint counts are now taken over the whole scan, and
max_findings trims only the returned list. Previously a capped run
under-reported the total, so two runs at different caps could not be
compared -- one measurement read 39 where the truth was 81. A new
findings_total gives the uncapped figure.
mutation_probe stops a batch before it outlives the transport timeout (ANTS-4736)
mutation_probe now stops a batch before it outlives the MCP
transport's read timeout. Previously the loop kept mutating the source
after the caller had been told the call timed out, so a session could
see an unexplained red suite and commit a file the probe still held.
The deadline is measured from the slowest run so far, not predicted
from timeout_sec (a worst-case cap), so batches that fit are
unaffected; skipped mutations report not_run and the reply names the
arithmetic. transport_budget_sec (default 60, 0 disables) tunes it.
read_region no longer returns a short Python function body while reporting it complete (ANTS-4735)
Symbol mode brace-matched every function, and { opens nothing in
Python — so a body containing a dict literal ended at that dict's closing
brace, dropping everything after it, with truncated:false saying the
read was whole. In py mode the extent now comes from the flat outline,
which is what indentation already encodes.
doc_citations resolves a bare basename naming a markdown document where the codebase index cannot (ANTS-4728)
A skill citing a standard as commits.md — a file living under
standards/ — matched none of the resolver's three routes in a tree with no
codebase index, so the quotation came back unverified and, in an
aggregate, looked like a pass. The caller now supplements the basename map
with a bounded markdown scan, and only where the index was absent or
truncated. A basename with two matches stays ambiguous rather than being
guessed.
roadmap_query's section_etag_match row now matches what the code does (ANTS-4732)
The row said the shortcut was ignored under mode:"headline_only"; in fact the section argument alone selects it, and it is honoured. The row now says so, and carries the edge that follows: the etag keys on the section's bytes, so varying mode between calls can be answered unchanged for a payload shape you did not ask for. No code change — the code was already right.
section_index now warns when it returns no sections, and says which of the two reasons applies (ANTS-4710) Asking a project for its roadmap sections used to give the same empty answer whether the file was unreadable or simply had none. The warning names the case, using whether any id-bearing bullet parsed to tell them apart.
check_sync emits sync_checked on both arms (ANTS-4730) The field the schema tells every caller to read first was emitted only where it would be false, so a caller doing the documented check read the healthy answer as "nobody looked" — defeating exactly the misread the field was added to prevent. Its absence now means only that check_sync was not requested. Two tests asserting the old shape were updated with the reason recorded.
discarded_external_edits no longer fires on the renderer's own format marker (ANTS-4729) The one flag documented as the only place a silently discarded hand-edit surfaces was firing once per not-yet-marked project, on a header the render itself adds — and a high-trust flag with a routine false positive is one sessions learn to skim. The boolean is now derived from the file's own classified lines; the total count is deliberately unchanged, since deciding which differences are cosmetic is not this check's judgement.
Removed a dead duplicate lambda that made every build of the Claude-integration library warn (ANTS-4687) A second, never-called copy of a helper sat in a different method from the live one. Dead since it was written.
roadmap_query's schema and spec now state that mode:"headline_only" with ids also carries input_index (ANTS-4712)
Both described a strict four-key shape while the ids path has always
emitted a fifth. The key is kept rather than removed: results come back
in document order, so dropping it would reintroduce the mis-pairing it
exists to prevent, for exactly the leanest caller. A test now locks the
ordering that produces it.
read_regions refuses a non-object region item by naming the shape it wanted, not the selector rule (ANTS-4721)
Passing a [start, end] pair per item — a natural reading of the
regions alias — was refused with "exactly one of {line range, symbol,
section} required", which diagnoses a choice the caller never made and
cannot act on: an array has nowhere to hold a selector. The refusal now
names the expected shape, what arrived, and the item's index, and stays
per-item so one bad item does not cost the batch its good ones.
file_outline no longer reads a # comment inside a fenced code block as a markdown heading (ANTS-4722)
A path comment labelling a fenced source listing was emitted as a
level-1 heading. With sizes:true that phantom took its extent to the
end of the file, so it claimed most of the document and every real
section below read as its child with its own size wrong —
turning the answer to "where is the natural seam?" into a comment.
max_heading_level was no escape, since the phantom was level 1.
A multi-line old that misses now reports where it nearly matched, instead of a bare not_found (ANTS-4723)
The near-miss report previously covered single-line misses only, so the
commonest multi-line cause — a hard wrap falling elsewhere — was
indistinguishable from the text being absent. It now reports the location
and names match_wrapped in its hint, so the fix is one re-send rather
than a re-read. This matters most mid-batch, where earlier edits have
already landed and the file has moved under the caller.
Releasing now updates the package repositories instead of only the recipe in git (ANTS-4716) Promoting a release pinned the new tag in the OBS recipe and pushed it, but nothing told OBS to re-fetch — so the repositories kept rebuilding the sources they already had, while the git side looked correct. That is how package users ended up two releases behind twice. The promote step now runs the submit itself, and says so loudly if it cannot.
The AI triage request is scrubbed for secrets before it is sent. (ANTS-4448) This path built and posted its own request rather than going through the shared client, so it never got that client's scrub. The prompt embeds the finding's source snippet -- and for a secrets or gitleaks finding, that snippet is the credential. The status line now says when something was redacted, so a scrubbed prompt is not silent.
Credentials in a git remote URL no longer reach SARIF exports or the wire. (ANTS-4448)
A remote like https://user:token@host/owner/repo.git was written
verbatim into SARIF provenance and read back out of it by
last_audit_summary. The userinfo is now stripped on both the write and
the read, so SARIF files captured before this change stop leaking too.
An scp-style remote (git@host:owner/repo.git) is left alone.
Roadmap and changelog lookups stay inside the project they were called from. (ANTS-4447) The search for a project's ROADMAP.md or CHANGELOG.md walks up from the calling directory. It now stops at the enclosing git repository before walking rather than after each step, so a call made from a directory that is not in a repository no longer reaches a different project's files.
The release workflow defaults to read-only and grants write on one job (ANTS-4795)
release.yml declared contents: write for the whole workflow; it now
defaults to contents: read and grants write on the single job that
creates the release and uploads the AppImage, mirroring ci.yml. Same
effect today, but a job added later inherits read rather than a write
grant nobody decided to make.
Release-workflow values now reach the shell through env:, never template expansion (ANTS-4792)
All thirteen remaining ${{ }} expansions into run: bodies in
release.yml are gone; the tag and RC flag are passed as environment
variables. ANTS-4772 had closed the direct inputs.tag path but the same
value was written to $GITHUB_OUTPUT and read back into the script body
at eight further sites. A new conformance test enumerates
.github/workflows/ and fails on any expansion reaching a shell body, so
a workflow added later is covered on the day it lands.
CI and release checkouts no longer leave the git credential in the runner workspace (ANTS-4773)
All four actions/checkout steps set persist-credentials: false. Neither workflow invokes git, so nothing depended on the persisted token, and it can no longer reach an uploaded artifact via .git/config.
The release workflow passes its tag input through env: instead of expanding it into the shell (ANTS-4772)
A workflow_dispatch tag is now read as "$INPUT_TAG" rather than substituted by the Actions template engine before the shell parses the line, so a crafted tag can no longer become shell source in a runner holding contents: write.
0.7.108 RC2 — Patron preview
Respin of v0.7.108-rc1 with cherry-picked fixes: 0086e07b
0.7.108 RC1 — Patron preview
This release shipped without written notes.
0.7.107
feedback_query reports a format version it does not recognise (ANTS-4702) A feedback file claiming a version that has never existed was accepted in silence. It is still read, and the envelope now says the marker is unrecognised.
read_region takes max_line_bytes to clip long lines (ANTS-4700) Reading part of a document whose weight sits in a few very long lines returned a truncated head and none of the rows you were reading toward. Clipping per line is applied before the size cap, so it makes room rather than competing with it.
A prose-heavy migration explains its item count (ANTS-4694) Migrating a large milestone document could report only a handful of items, which read as though most of the file had been skipped. Nothing is skipped — the rest is stored as prose and tables — and the report now says so.
feedback_log can record a whole triage in one write (ANTS-4671) The new assign_id_batch fills many findings' id slots at once, where before each one cost a separate read and write of the file.
The tab colour menu now offers 25 colours, including greys and black (ANTS-4689) Right-click a tab to pick from 25 presets instead of 14. New: Orchid, Indigo, Sand, Olive, Lime and Mint fill gaps between the existing shades, and a grey ramp -- Silver, Gray, Slate, Charcoal, Black -- is grouped at the bottom of the list. They use the same gradient as every other tab colour.
roadmap_log op:"amend_field" edits an item's Layman, Kind, Source, Lanes or Evidence (ANTS-4667) These could be set when an item was filed and never changed afterwards. Trying through amend_body failed with the text reported missing, because those lines are assembled when the roadmap is written out rather than stored. That error now points at the new command.
roadmap_log's envelope could report when its XML scrub actually removed something. (ANTS-4572) A safety net quietly cleans up a mistake, so nobody learns they made it.
read_region gives no positive signal that a read did NOT spill once fields + compact narrow the envelope away. (ANTS-4567) A deliberately cheap probe can come back saying only "ok", which is indistinguishable from the probe being wrong.
invariant_check cannot see a spec that cites the file by basename, so the wrong spec comes back looking like the answer. (ANTS-4566) Asking which design document governs a file returns a confidently wrong one when the right one names the file informally.
roadmap_migrate accepts fields= and compact, so its envelope can be narrowed (ANTS-4429)
A no-op re-sync returned ~25 KB with no way to ask for less — the schema
sets additionalProperties: false, so fields= was not merely absent, it
was refused. Both args are consumed by the dispatcher after the handler
returns, so they narrow the report and never the migration. Closes the
last of ANTS-4144, ANTS-4408 and ANTS-4429, the older half of which
ANTS-4649 had already settled.
feedback_log op:"set_title" — rename a feedback file's heading after a project rename (ANTS-4646)
The filename was always derived correctly; the H1 was writable by no op, so a rename left the title contradicting it and the only route was a hand edit, against the file's own "don't hand-edit" instruction. title is the project name, not the whole heading: the existing prefix is preserved verbatim, since this renames a project rather than normalising a corpus. Re-runnable — an already-correct title reports no change and writes nothing.
invariant_check states that the ROADMAP was not consulted (ANTS-4645) Every reply now carries roadmap_scanned:false and a scope_note naming roadmap_query and task_priors. The envelope was confident and complete-looking on a non-zero answer, and a project rebuilt a feature its own roadmap had already specified.
doc_citations quotes: widen attribution from the line to the paragraph — coverage goes from 26 to 111 of 249 checkable quotations. (ANTS-4638) The checker only links a quote to its source when both sit on one line, which hard-wrapped prose almost never does.
Feedback lookups can be pointed at a shared corpus directory (ANTS-4471)
When feedback_query could not find a project's feedback file it said
only "not found", even though it already knew how to list nearby files —
it was looking in one directory, and for a project whose feedback lives
on a different part of the disk that directory is empty. A new
claude.mcp_feedback_root setting names the shared folder, searched in
addition to the usual one, and a miss now reports which directories were
searched plus how to widen it. Empty by default, meaning the previous
behaviour is unchanged.
roadmap_query can now answer "is my roadmap file in sync with the database?" (ANTS-4462)
Pass check_sync:true to have the roadmap rendered from the database and
compared against the file on disk — the same comparison a write already
runs before it overwrites, so the two can never disagree. Returns
file_in_sync plus a breakdown of how many lines differ and which text
would be lost. The existing warning only fired when the file mentioned an
item the database had never seen, so a status changed by hand looked
perfectly healthy; this catches that. Opt-in and never cached: it costs a
full render (~204 ms on a 2,267-item project), which is why it is a
question you ask once rather than something every query pays for.
Every roadmap_log flip/annotate envelope names the path that wrote (ANTS-4464)
The store-backed and markdown paths declare different fields on
purpose — a store has no line numbers — but nothing said which had
run, so a caller reading note_line got nothing with no way to tell
why, and identical calls changed shape across a migration. A new
write_path ("render" or "patch") makes the absence a statement. The
store path also now emits the post_bullets echo that
return:"headline_only" promises and only the other path supplied.
roadmap_log gains op:"annotate_batch" (ANTS-4470)
Adds a note to N roadmap items in one read and one atomic commit,
where N separate annotates cost N reads of a large ROADMAP.md, N
commits, and raced the file watcher. Takes flip_batch's existing
locators[] unchanged, so there is no second shape to learn; a
to_status is refused rather than ignored, and a locator with no
note is skipped without costing the batch its other closures.
Build-time Qt-version guard: refuse a tree whose objects predate a Qt upgrade (ANTS-4625) A Qt point upgrade could silently poison an incremental build. RPM preserves upstream header mtimes, so newer headers can be older than your objects; ninja is mtime-based, rebuilds nothing, and links the new libraries against stale objects. Measured on 6.11.1 → 6.11.2: 859 of 3838 tests failing with heap corruption. The build now compares the installed Qt against the version the tree was configured with, and refuses with the clean-rebuild command. A reconfigure alone never fixed this — it regenerates build.ninja but invalidates no objects.
Cross-session mailbox: session_message, so Claude sessions on different projects can leave each other notes (ANTS-4622)
Mail is addressed to a PROJECT by its export_slug, never to a session:
a session is its shell PID, which is ephemeral and reused, so mail
addressed to one is undeliverable the moment that session ends. Three
ops — send, inbox and ack — with confirmation held as two states in one
nullable column, and unread mail surfaced by session_orient at the
start of the next session on that project. Complements the
*_Ants_MCP_Feedback.md corpus rather than replacing it: that channel
asks for work to be FILED and ends in a roadmap id, this one asks for a
person to KNOW something and ends in a receipt. Adds the store's tenth
table and moves the schema to version 3.
A gate that markdown tables, code fences and blockquotes survive the store round trip (ANTS-4616) Once a project is store-backed the whole-file re-render fires on any write, so every element type a roadmap body can carry has to survive it. Nested lists were measured; these three were assumed. All four new cases passed on the first run — no defect was found, and the assumption is now a gate.
roadmap_migrate op:"deregister" removes a project from the shared store (ANTS-4617)
Trying a risky operation on a throwaway copy is the careful thing to do,
and it permanently polluted the shared database — there was no inverse of
migrate. deregister removes a project and everything hanging off it in
one transaction. It refuses while the project's folder still exists
unless you pass confirm:true, because the database holds history the
roadmap file does not; a project whose files are gone needs no confirm.
roadmap_log op:"render" publishes the store to ROADMAP.md on demand (ANTS-4614)
After adopting the database, the tidy-up a migration computed could not
be written to the file until some unrelated edit happened to trigger it —
so the only route was to invent a roadmap bullet nobody wanted, and a
clean git status after migrating was indistinguishable from the
migration never having run. render publishes it directly: no locator,
no semantic change, idempotent, dry_run previewable, and it runs every
gate the ordinary write ops run.
roadmap_query mode:section_index can be narrowed by section name (ANTS-4610)
Resolving the one slug a roadmap_log write needs used to cost the whole
index — 40 section objects for one string, with neither fields nor
compact able to cut it. query now filters the index by section
headline or slug, and the reply carries sections_considered /
sections_filtered_out so an empty result is legible. whole_word and
regex compose, as on the bullets path.
A project can declare its own roadmap ID format in .ants/project.json (ANTS-3771)
Add an id_format object with a prefix, a pattern, or both, and the
roadmap reader stops guessing. prefix decides what a new ID is called and
refuses one written with the wrong prefix; pattern says which part of a
bold heading is actually the ID, so a bullet written
**AX1. Geometric occlusion** is reported as item AX1 with "Geometric
occlusion" as its headline, instead of the whole sentence being treated as
the ID. Set it with project_settings op:"set" id_format:{…}. A project
that declares nothing is completely unaffected, and a heading the pattern
does not match keeps exactly the ID it had before — nothing loses its
identity. Unblocks the one-time converter for checklist-format roadmaps.
roadmap_query says which ids it GUESSED from a bold prose lead-in (ANTS-4575)
A bullet now carries id_inferred: true when the reader adopted its id
from a bold phrase instead of reading an id the author wrote in brackets
— so a session planning work on a checklist-style roadmap can tell a real
id from a guessed one. Gated: absent when the id was declared, so a
roadmap whose ids are all bracketed sends exactly the same bytes as
before. It is not the existing synthetic flag, which marks an id
invented from a content hash when there was no text to take one from; the
two can never both fire on one bullet.
roadmap_query now says when a roadmap file mixes two bullet styles (ANTS-4604)
A roadmap carrying both checkbox and emoji bullets reported only the
first style it found, so it looked identical to a file written
entirely in that style. It now adds mixed with a count of each,
and only when the file really carries both.
roadmap_log op:"repair_trailers" — recover the trailer values migration cut short (ANTS-4585)
A one-off, per project, with a dry run. Values cut at a hard line wrap or
at the full stop inside e.g. / config.yaml are recovered by re-parsing
the prose that still holds them, since the legacy inline run was never
stripped. A value is written only where the stored one is a strict prefix
of the re-parse: measured across 16 projects, seven items hold a column
newer than its prose, and an unguarded pass would revert those to
superseded text. 824 layman values, 58 source and 54 lanes are repairable.
A daily audit of what is actually published (ANTS-4588) Checks the current release against reality rather than against the last build's exit code: that the release carries its AppImage, .zsync and permanent alias, that the permanent URL serves a real ELF binary (fetched as bytes, not inferred from a listing), and that every package repository publishes the release's version. Written because two channels broke on the same day while both reported success, and because a per-run check cannot catch a job that dies before reaching it. A fetch that could not run reports as unchecked rather than as drift.
A permanent direct-download link for the latest AppImage (ANTS-4586)
A stable release now also publishes the same binary under a version-less name, so github.com/milnet01/ants-terminal/releases/latest/download/Ants_Terminal-x86_64.AppImage downloads the current version and never needs updating anywhere. Previously the filename embedded the version, so no permanent direct URL could exist and every link a site or README could offer was a page to navigate. The versioned filename is unchanged and RC builds are excluded, since latest never resolves to a prerelease.
README.md's numeric claims are verified against the code on every push (ANTS-4584) tools/check-readme-claims.sh checks the version banner against CMakeLists.txt, the MCP tool count against claudeintegration.cpp, and the colour-theme count against themes.cpp — the tool count in particular had no verifier but a person remembering. It runs from the pre-push hook ahead of the docs-only skip, so a README-only push cannot exempt itself; every tenth push additionally raises the prose for a human read.
workspace_search gains match_wrapped:true — find a quotation that is hard-wrapped in the file. (ANTS-4547)
A run of whitespace in the pattern matches a run of whitespace and
markdown blockquote markers in the text, so a sentence spanning a line
break is found and line reports where the span starts. Prose here
wraps at ~70 columns, so a line-oriented search missed text that was
present and exact — and a review gate that dismisses a finding whose
quote it cannot locate then ships the real defect. Two-sided: paste the
quotation with its own newlines and > markers and it still matches
unmarked text. Literal patterns only; off by default.
The roadmap can now answer "when" — dates are recorded going forward, and the past is backfilled from git (ANTS-4501)
roadmap_query mode:"report" shipped with nothing to count: the item table's date columns existed and nothing ever wrote them, so every "closed this week" was computed over an empty set. Two changes fix that. Every roadmap write now stamps when an item was created, when it last changed, and — only on the move into shipped — when it closed; a later edit to a closed item no longer moves its closure date, and reopening one clears it. And roadmap_log op:"backfill_dates" walks the project's own git history once to date everything filed before today: 1525 revisions in 4 seconds on this project, dating all 2179 items with none left over. It never overwrites a date that is already there, so it is safe to re-run and a hand correction survives.
roadmap_query and roadmap_log now document that they spell one dialect two ways (ANTS-4606)
roadmap_query reports github-task-list where roadmap_log reports
gfm for the same format, so comparing the two envelopes read as a
disagreement that was not one. Both spellings stand — unifying them would
silently break a caller already keying on either, and the roadmap store is
machine-global, so that caller may be another project's session. Each
verb's detail now says so and tells you to map both spellings.
workspace_search's exclude_glob now names its use for scoping a cold review lane (ANTS-4568)
A subagent told which files it may read cannot honour that through an
unscoped search: one project-wide match returns context lines from the
documents it was forbidden. The description now says so, so a caller
reaches for the argument instead of trusting the instruction.
workspace_search's glob says how to name several paths at once (ANTS-4672)
It takes one pattern where its neighbour exclude_glob takes a list, and nothing said that several paths are written as docs/{coding,testing}.md. The description now says both.
Seven more MCP verbs now run off the GUI thread (ANTS-4682) The five test_audit_* verbs, caller_cwd_info and project_query no longer hold the window while they work. The other seven inline verbs stay on the GUI thread on purpose, and each now records why in the code.
fields= and response compaction are now separate per-verb answers, so spec_lint and feedback_query are no longer compacted unless asked (ANTS-4524)
A single predicate granted both, so a verb added to the list for fields=
silently began compacting its replies for every caller — which folded away
spec_lint's flag saying a check had never run. They are now two columns of
one table, and adding a verb makes you answer each. An explicit
compact:true is still honoured wherever the schema declares it; only the
unasked-for default is per-verb. Partial: making fields= universal is the
remaining half, and the blocker that half named is now gone.
feedback_query ignores fields=, so a duplicate check pays for the whole delta it was passed to avoid. (ANTS-4665) Asking the feedback reader for just a couple of small numbers returns the entire backlog anyway.
compact_resolved retires a legacy tracking heading once every id in it has shipped (ANTS-4646) A v1 "Tracked in ROADMAP" heading survived migration with no verb able to touch it — append_tracking refuses on a v2 file, and assign_id needs a finding those ids do not have. It is retired under the gate compact_resolved already applies, all-or-nothing per heading. It is deliberately KEPT on a fully condensed file carrying no inline proposed ids, where that line is the project's only record of what it reported.
roadmap_migrate collapses repeated notes instead of restating one fact hundreds of times (ANTS-4649) A 357-item migration returned 357 identical "field defaulted" objects — about 6 KB — beside two envelope fields that already said the same thing, and both the preview and the real call emitted it. Repeated notes are now merged by (code, detail, source), each merged row carrying a count and up to three sample lines. Measured on the 250-item fixture: 500 notes become 2 rows and 197 bytes, and the answer is now complete where before it was capped at 200 rows and still truncated.
a ripgrep start failure now names its cause instead of asking whether ripgrep is installed (ANTS-4650) "rg failed to start (is ripgrep installed?)" could only ever name one cause, and it has been emitted both for ripgrep genuinely absent and for ripgrep present and working — so in one of those cases it sent the reader to the wrong repair every time. The three rg-backed verbs now share one classifier that distinguishes a vanished search root, rg missing from PATH (naming the PATH searched), rg still starting when the wait expired (a loaded host — retry, do not fall back to grep), and rg refused by the OS.
A 5-second git-spawn budget in eight test fixtures reddens build-asan at random. (ANTS-4651) One shared, env-overridable budget replaces twelve hard 5-second literals across eight fixtures, and a timeout now says which command blew it and after how long — distinctly from a helper that could not start at all.
Large-result previews shrink each row's text sample before dropping it (ANTS-4474)
When a preview could not fit a sample of every row it discarded all of
them, leaving row sizes but no hint of what any row contained — usually
the only part that says what the row is. The sample width now steps down
(60 → 40 → 24 → 12) and the narrowest that still covers every row wins,
reported as rows_preview_head_chars. Only when no width fits is the
preview omitted.
A migrated project whose file was never rendered is write-locked, and the render that would free it is gate-refused. (ANTS-4628) The roadmap's "plain-English summary" rule now applies to the items an edit touches, instead of to the whole project. Editing one item no longer demands that every other item in the project already carry a summary — which had made some projects impossible to edit at all, including the one whose repair was itself blocked by the rule.
Split the roadmap-write translation unit, which was at its line cap (ANTS-4620) The file holding the roadmap write operations had reached the 6,000-line ceiling the project sets, leaving no room for the next change. It is now two files — one for single-item writes, one for batch and section writes — with the batch and section operations moved out. No behaviour changes; two tests that named the old file by name were widened so they still cover both halves.
Declare that roadmap_log op:"append" writes ants-v1 into any roadmap dialect (ANTS-4605)
Appending to a github-task-list roadmap has always written an ants-v1
emoji bullet, leaving the file mixed, and nothing said so — the
standard actively claimed the opposite ("projects that prefer GFM stay
on GFM"). Now declared in roadmap-format.md § 3.10.2 and in the verb's
own op:"append" reference. Behaviour is unchanged: the store re-renders
the file in one dialect, so the mixed state is transient. op:"flip"
does not do this and never did — it applies per the bullet's own
format.
discarded_edit_lines now comes with a breakdown and the lost text itself (ANTS-4615)
One count mixed harmless reformatting with sentences that were deleted,
so nobody could tell the difference — a measured report of 84 was 24
restyled bullets and one sentence that no longer existed. A drifted
write now also reports discarded_restyled_lines,
discarded_text_lines and discarded_text[] naming the lost lines. The
total is unchanged: this adds a breakdown rather than suppressing
anything.
roadmap_log's bad_section refusal now carries ranked near-miss candidates (ANTS-4556)
A mistyped slug threw the composed body away and left one route: a
section_index call answering every slug in the file. The refusal now
ranks the real slugs against what you typed (capped at 10), reusing
read_region's ranker rather than growing a second copy. It still
refuses — a near miss is a hint, never a resolution.
roadmap_query: body is the item's notes, no longer the whole rendered bullet (ANTS-4557)
Two projects reported opposite things about this field and both were right: the database column holds notes, while the tool rebuilt the bullet and handed back its title line and metadata block too. The field is now the notes alone — the title is already returned separately. The metadata lines stay, because nothing else carries their text.
roadmap_log: dry_run now says what a green preview is, and is not, evidence for (ANTS-4548)
The preview is not a simulation — it performs the store write inside a transaction and rolls it back, so every refusal the real call would hit reaches it too. It stops short of the commit and of the write to disk, and the tool description now says so instead of claiming the preview cannot differ from the result.
workspace_search: a lane that escapes the project root now names the route out (ANTS-4569)
The refusal was correct but terminal. Both the lane schema
description and the bad_path envelope now say that to search a
different tree you set caller_cwd to that tree's root and confine
with glob — so a refused call carries its own repair.
Roadmap card columns now line up across sections whatever the font (ANTS-4718) The shared column grid was not actually shared: a section whose rows carry a kind grew that column past its reserved width when the text did not fit, while a section without one kept the reserved width, so the two sat on different grids. It only showed where fonts render the kind text wider — never on a typical desktop. The grid is now measured across every card and applied uniformly, and no text is truncated to achieve it.
changelog_log add_subsection's docstring described a guard the code does not implement (ANTS-4562)
It claimed a refusal against any [Unreleased] carrying flat category
blocks, and told you to empty the section first. The guard actually tests
ORDER: a section led by a dated topic is accepted however many flat blocks
trail below. The remedy is to lead with a dated topic, and a legacy tail no
longer reads as a blocker.
roadmap_migrate no longer advises a re-run that a known open defect says destroys correct rows (ANTS-4564)
Its hint called re-ingesting "routine and idempotent", while ANTS-4507 has
established that parse(render(x)) is not identity — so items_updated can
be a re-parse artefact and applying it overwrites a good row. Both the hint
and the description now say so and point at updated_items[].
doc_citations resolves a quotation inside a blockquote (ANTS-4706)
A quote written as an indented > block was reported out of date even
when it matched its source exactly, and the reported text carried the
block markers so copying it into a search found nothing.
An offload envelope always reports how many rows exist (ANTS-4705) A summarised result could say its preview was shortened without ever reporting the total it was shortened against — the figure that decides where to read next.
doc_citations no longer blames a nearby document for a stale quote (ANTS-4696) A quotation credited to a document cited without its file extension was reported out of date against an unrelated file mentioned earlier in the same paragraph. It now says the target was not recognised, and names it.
Publishing the roadmap reports punctuation changes separately (ANTS-4695) A render that adjusted the punctuation of hand-written summary lines counted it as cosmetic restyling, so the field a caller checks before allowing an overwrite read zero while their sentences changed.
roadmap_migrate finds a roadmap declared in .ants/project.json (ANTS-4693) A project keeping its roadmap somewhere other than the usual filename could be read and written but not adopted into the database. The other verbs already honoured the declared path; this one now does too.
An offloaded result keeps the body's own top-level fields (ANTS-4692)
When a large roadmap_query answer was parked to disk, fields like
source and file_in_sync vanished with it — the safety information a
session is told to read before writing. They now survive, and anything
dropped for size is named in preserved_omitted.
A roadmap_log dry run no longer leaves a counter file behind (ANTS-4691)
A preview that refused still created .roadmap-counter, so pointing a dry
run at a repository left an untracked file in it. Both the single-append
and batch paths are fixed.
A narrowed reply keeps the note explaining how to read it (ANTS-4698) Asking a verb for only certain fields could drop the warning that says a zero means "this roadmap's format wasn't recognised" rather than "nothing left to do". That warning now always survives.
Asking the roadmap for a few rows using the search tool's word now works (ANTS-4701)
roadmap_query calls its row cap limit where workspace_search calls it max_results. Sending the wrong one used to be ignored and answered with every row; it is now accepted as an alias.
token_usage says so when it cannot read the figures, instead of reporting zero (ANTS-4684) Asked while the window is closing, it reported no savings rather than saying it could not look -- a wrong number where an error was owed.
A roadmap item's title can be corrected again (ANTS-4668) roadmap_log op:"amend_headline" refused outright on any project whose roadmap lives in the database, which is all of them -- so a title was fixed only by refiling the item under a new number and losing every reference to the old one. It now edits the title in place.
The window no longer freezes while Claude runs an MCP command (ANTS-2132) Ants answered every MCP request on the thread that draws the window, so the whole window was frozen for as long as the request took — and with several Claude sessions live, their requests queued behind each other on that one thread, which is why the freezes had no pattern. Eligible verbs now run on a background worker and the window keeps painting. Requests still run one at a time, so nothing that could not previously overlap now does. Measured: a 10 ms timer fired once across a 200 ms verb before, ~20 times after. Two verbs are NOT covered yet and still block for a whole sweep — audit_run and indie_review_dispatch (ANTS-4682).
fields= is documented universal but allowlisted, so every other verb drops it AND cannot report it. (ANTS-4524)
fields= is honoured on every verb now, and declared on every schema, so
a caller who narrows a response gets what they asked for instead of the
full payload in silence. Compaction stays per-verb on purpose: it has a
default, and a default nobody asked for is what folded away a flag saying
a check had never run.
changelog_log op:release echoes the whole closed section back with no way to suppress it, so it fails on a large [Unreleased]. (ANTS-4561) Closing a big changelog section returns the entire section as its reply, which is too large to receive.
spec_lint's surfaces hint states a FALSE cause: the check does have an input, and it is a hard-coded layout. (ANTS-4679) A tool tells you a check can never be switched on. It can -- the tool just only looks in one place for the thing that switches it on.
spec_lint names a check in skipped[] while returning that same check's findings, so the field cannot be read either way. (ANTS-4666) The spec checker reports a check as "did not run" in the same breath as reporting what that check found.
roadmap_query's sync_checked:false is folded away by default compaction -- ANTS-4673 in a second verb, already live. (ANTS-4677) A flag meaning nobody checked whether the file matches the database is quietly deleted from the reply, so it reads as checked and fine.
spec_lint's skip entries and their hint fields share no token, so a hinted skip reads as unexplained. (ANTS-4676) A tool says it skipped a check and does explain why, but the explanation is labelled so differently that readers cannot tell the two go together.
doc_citations quotes:true detects a quoted span per LINE while matching folds newlines, so a hard-wrapped quotation is never checked and lands in no bucket. (ANTS-4664) The quotation checker silently ignores any quote that runs onto a second line, and reports the document as clean.
Neither quotation checker survives a hard wrap, so a true quotation reads as unverifiable. (ANTS-4607) Two tools that check whether a quote is real fail when the quote is split across lines, which is how this project writes everything.
An unknown roadmap section now suggests near misses on database-backed projects too (ANTS-4591) The "did you mean?" hint had reached only the file-backed path, so every migrated project got a bare refusal and a full section listing as its only route. The schema now also says the hint exists — documenting it first was the original request, and probing the field before writing about it is what found it missing.
A refused preview no longer blames an item id that exists nowhere (ANTS-4593) A dry run creates its candidate row inside the transaction and rolls it back, so a render-gate refusal listed an id nobody could look up. The preview's own row is now reported separately, and when it is the only offender the message says the problem is the arguments rather than the roadmap.
A truncated first-call file digest now says so under a findable name (ANTS-4560)
On a project with no module map the digest is a flat file list, and the
only flag describing it was named for the lane digest — while the
obvious-looking files_truncated describes the index cap and is
correctly false. A reporter read those and concluded a large project had
no renderer. source_files_truncated now travels alongside.
doc_citations accepts the ~global sentinel its sibling verbs already took (ANTS-4565)
It was the one verb of the doc-checking set that refused it, so checking
a document under ~/.claude/ meant hand-rolling the citation pass. The
basename fallback is unavailable under that root and the description now
says so.
A narrowing argument sent to a verb that ignores it now says so (ANTS-4578)
fields, compact, offload and etag_match were exempt from the
unknown-argument advisory on every verb, while the dispatcher honours
them only for the verbs that support them — so asking an unsupporting
verb to narrow its reply did nothing and reported nothing. The exemption
is now decided per call.
A roadmap note naming a C++ symbol is no longer read as a trailer declaration (ANTS-4608)
A double colon is a scope operator, not a heading. Writing about a symbol
whose scope name ends in Source, Kind, Lanes, Layman or
Evidence was refused; the quieter half was a scoped symbol inside a
value silently truncating it at that point.
A spilled MCP reply now still names the argument it ignored (ANTS-4626) The unknown-argument advisory was attached to the response body and then discarded when a large body was spilled to a handle — so it reached the caller only when the reply was small, which is backwards: a filter that was silently ignored is what makes a reply large in the first place.
A repeated MCP refusal no longer comes back as ok:true (ANTS-4446)
The ETag short-circuit built its {ok:true, unchanged:true} reply before
parsing the response, so it never learned the body was a refusal — and
the same function attached an etag to refusals in the first place, which
is what put a replayable etag in the caller's hands. It now parses first
and returns a refusal verbatim.
a nested project's feedback file is no longer created where nothing reads it (ANTS-4647)
Omitting path derives the file into the parent of caller_cwd, which is the shared corpus only for a project sitting directly under it. A repo nested inside another workspace got a file in that workspace — invisible to every maintainer sweep, and starting an empty parallel record beside the workspace's real one — and was told it succeeded. It now refuses with the ancestor corpus's files as candidates. A configured claude.mcp_feedback_root redirects the write there instead, and a genuinely new corpus with no feedback file anywhere above still creates its first file.
project_settings op:detect omits its source counts when no walk ran, instead of reporting 0 (ANTS-4648) A project with a settings file short-circuits before the directory walk, so both counts were still their initialiser — and a ~3000-line project was reported as holding no source at all. They are omitted now, with counts_computed saying which shape the reply is.
invariant_check retries a zero with the path's suffixes, so a spec that cites a module by name is no longer invisible (ANTS-4644) A spec cites a module the way its author writes it, so the project-relative path the verb's own description prescribed was routinely the one form that matched nothing — and matched_count:0 beside specs_scanned:247 read as "nothing governs this file". On a zero that did look, the scan now retries with each path's suffixes, longest first; the bare basename is a separate later tier, used only when every fuller form has failed. fallback_match rides on every reply, and matched_as names the form that actually matched. Measured over this repo's 247 specs: 29 source files that returned a confident zero now return their governing specs.
doc_integrity broken_link: a ~/-prefixed target is resolved as a relative path and reported as a missing file. (ANTS-4643)
Links written with the home-directory shortcut are reported as pointing at nothing, with a message that sends you looking for a file that is there.
doc_integrity broken_link: skip a link whose enclosing span is a verbatim quotation of another document. (ANTS-4642) Quoting another document word for word drags its links along, and the checker reports them as broken.
doc_integrity heading_sequence: suppress a numbering gap the document itself accounts for, and say that it did. (ANTS-4641) A file that explains why it skips numbers 2, 3, 10 and 13 still gets flagged for skipping them.
doc_citations quotes: a table row's LAST backticked path wins attribution, which is the wrong one in an impact table. (ANTS-4640) In a table listing which documents a change affects, the quote gets checked against the affected file instead of the source.
doc_citations quotes: resolve a bare basename against the scanned document's own directory, and stop making non-document tokens targets. (ANTS-4639) A quote pointing at a sibling document in the same folder is reported as having no source at all.
doc_citations quotes: skip a quotation sitting in a loop-log table row — 271 of 520 in one corpus (52%) are historical by design. (ANTS-4637) The quote checker flags wording that a review deliberately replaced, so most of what it reports is not a problem.
roadmap_query: a freshly migrated project no longer warns that its file is ahead of its own store (ANTS-4636)
idHighWater() reads the id_prefix row, and migration does not write one — only an id-allocating append does. A migrated-but-never-appended project therefore reported a store high-water of 0 while holding every id in its file, so file_ahead_of_store was pinned on permanently and read as "the migration silently wrote nothing". The mark is now the higher of that counter and the ids the items actually hold. Reported by Album_Builder on a 357-item project.
Confirmed that file_ahead_of_store's baseline firing is a real defect, not a fixture artefact (ANTS-4633) Album_Builder reported the same flag on a real 357-item project whose ids were minted under a previous directory name, which is this item's own hypothesis confirmed in the wild: a high-water lookup keyed on a prefix re-derived from the path finds no rows and reports 0. The fix is tracked as ANTS-4636.
roadmap_log: op:"append_batch" now reconciles .roadmap-counter on the store path, as op:"append" already did (ANTS-4635) The cache stayed stale for a whole batch until some later single append's scan repaired it, so anything reading it to predict the next id got a wrong answer that looked right. Reported by Games_Hub, who measured three batches allocating GHUB-0113 through GHUB-0118 with the counter still reading 110. The reconciliation is now a shared helper both paths call, and the batch envelope reports counter_advanced_to / counter_advanced_past when the cache moves.
roadmap_log: a store-backed dry run reports would_be_id / would_be_ids, never id / ids (ANTS-4634)
ANTS-4508 established that a preview must not report the id under the real write's key — a caller reading one field takes it for a reservation, and nothing is reserved. That rename landed on the markdown branch only, so every migrated project, which takes the store branch, still got id back from a dry run. Reported by finbreak and reproduced directly. A test in roadmap_write_half had pinned the store branch to the old behaviour; it now asserts the new key and the absence of the old one.
roadmap_log reports the status it just committed, not the one it read (ANTS-4466)
On a store-backed project, op:"annotate" described the stale file it
read rather than the result the render had committed — so after a
git checkout -- reverted ROADMAP.md, the envelope said the item was
still planned while the write had correctly restored shipped. Nothing
was lost; the report was wrong, in exactly the case where a caller
confirming a write most needs it right. A new file_status names the
file's value when the two disagree, so a divergence is visible rather
than silently resolved.
Roadmap id allocation reads the store's id columns instead of scraping the rendered roadmap (ANTS-4631)
A migrated project's ROADMAP.md is a rendered output of the store, but the
allocator still floored its next id to a \bPFX-NNNN\b scan of that file's
whole text — so it re-read the store's own answer back through prose, which
cannot tell an allocated id from a documented example. One deliberately
absurd ANTS-9999 inside a body was read as the high water: the next append
issued ANTS-10000 and burned ~5,370 ids, and every roadmap_query reply
reported the file ahead of a store that was exactly in sync, pinning on the
one indicator that would reveal a real divergence. The store path now asks
two columns and reads no file. The two callers with no database to ask — an
un-migrated project, and the file-vs-store witness — count only lines that
declare an id, a top-level list item outside a fence.
roadmap_query elides a long body BEFORE spilling it, so the middle is unreachable by any argument. (ANTS-4630) A very long roadmap note can now be read right through. Before, its middle was cut out and the suggested fix could not reach it.
changelog_log op:"add" double-wraps a summary that already carries its own emphasis or id. (ANTS-4629) Writing a changelog entry whose text already carried bold or its own reference number produced a mangled line, and the reply gave no sign of it. That is now refused with an explanation, and every write reports the line it actually wrote.
render_gate_unmet now names a remedy that works — repair every offender in one flip_batch (ANTS-4434)
The render's Layman gate is per project and is evaluated after the
mutation, so repairing one item at a time is refused by whichever
offenders remain and rolled back — a single repair commits only when it
is the last one outstanding, and a project with two or more could not be
repaired one item at a time at all. The refusal previously named the
offenders and no way out. It now names the route that works: one
flip_batch carrying every id from gate_failures, each locator with a
note whose first line declares the Layman: trailer and to_status set
to the status those items already hold. No gate change, so ANTS-3758
INV-5 is untouched. Locked by a new RoadmapWriteHalf case that asserts
both halves — each single repair refused while the other is outstanding,
and one batch repairing both. ANTS-4434 stays open for the bulk case,
where the offender count is 585 and a batch of hand-written summaries is
not a remedy.
session_orient now honours fields=, narrowing the largest response in the session (ANTS-4624)
The verb joined the projection allowlist in 0.7.105 but never gained the
matching fields schema property, so the argument reached the dispatcher
as a string, failed its array check, and was skipped in silence — the
full bundle came back on every narrowed call. The property is now
declared, and the test that should have caught it derives its count from
the allowlist instead of hardcoding one.
roadmap_migrate now declares its op and confirm arguments (ANTS-4621)
op:"deregister" and confirm:true were documented in the verb's
description but declared in neither its schema properties nor any enum.
The schema sets additionalProperties:false, so a strictly validating
MCP client refused the call outright — and on the permissive path the
ignored_args advisory named op as ignored on the very call that
argument had just steered.
amend_body no longer collapses an aligned or nested block when a match spans it (ANTS-4612)
A wrapped match re-flows the lines it spanned into one, which is right
for a hard wrap and destroys a column-aligned table or a nested list —
and every repair attempt ran the same pass, so the damage accumulated.
Such a span now refuses body_match_wrapped_block and writes nothing.
Ordinary hard wraps are unaffected, including a bullet whose
continuation sits at a deeper hanging indent.
feedback_log no longer derives a stray feedback file from a dotted project leaf (ANTS-4613)
Omitting path from a project like ~/.claude created
<home>/.claude_Ants_MCP_Feedback.md and reported success — a
hidden-looking file that could never match the naming convention, so a
session quietly started a second feedback file nobody reads while the
real one went untouched. A dotted leaf now refuses with the sibling
files as candidates. Only the leaf counts: a project under a dotted
parent still derives normally.
roadmap_query's no_roadmap_loaded refusal now says where it looked (ANTS-4611)
Passing an explicit caller_cwd and getting back "no ROADMAP.md detected
for the active tab" sent callers to diagnose tab misrouting. The refusal
now reads "no ROADMAP.md under " and carries resolved_root; the
tab is named only where the tab supplied the path. The verdict is
unchanged, and the schema now says the code is an answer — this project
keeps no roadmap — rather than a failed call.
roadmap_log body scrub now strips leaked OPENING tags, not just closing ones (ANTS-4609)
Peeling a closing tag off the edge of a body exposed the opening half of
the same pair, alone on its line with its value, and the scrub stopped
there — so a bare <compact>true line reached ROADMAP.md under an
ok:true envelope. Only a lone scalar after the tag counts as leakage,
so markup followed by prose still survives.
a trailer value your own text got wrong is no longer reported as a database fault (ANTS-4577)
A body declaring an unrecognised Kind: refused under store_failed,
so a caller branching on the code read an engine fault for their own
text — and a retry is the wrong answer to both readings. It now
answers bad_kind, which is what the same value in an argument has
always answered.
find_definition now resolves C++ definitions whose return type wraps onto its own line (ANTS-4603) Every definition anchor matched a single line, so the common shape for a long return type resolved to nothing — which reads as "no such symbol" for a function that is plainly there. 30 such definitions in this tree, including 14 of RoadmapStore's.
roadmap_log's store-only ops now name the project they refused (ANTS-4602)
repair_trailers and backfill_dates reported the store holds no row for "" on every refusal, naming neither the project nor the
file — the shared lookup filled its out-params only when it
succeeded.
A throwaway temporary folder can no longer register itself as a real project (ANTS-4600) The roadmap database is shared by every project on the machine, so anything recorded in it as a project stays there. A session's temporary scratch folder had been recorded that way: 33 items that were copies of another project's, under a folder that was deleted minutes later. Every machine-wide count has been reporting one project too many ever since. Migrating from a folder under the system temporary directory is now refused, and says which folder to use instead. The already-recorded copy is removed separately.
roadmap_query: the include_body schema now says a trailer line can be composed from a column (ANTS-4599)
The store-backed record is built by rendering the item and re-parsing it,
so the body carries a trailer line for every column whose key the stored
prose does not declare. The field answers what the bullet says in
ROADMAP.md, not whether the stored body declares that field — and the
schema had promised the second. Behaviour unchanged; locked by INV-6.
A Lanes: list is no longer cut short at a dotted lane name (ANTS-4597)
The lanes capture stopped at the first full stop, which for this key is
usually a dot inside a filename, so every member after it was lost —
remotecontrol.cpp, AuditDialog, RoadmapDialog, … was read as the single
lane remotecontrol. The value now ends at its sentence stop instead,
where a dot inside a token does not count, and the list is split from the
finished text. Measured over the roadmap store: 15 bullets parse
differently, 5 gain whole members, nothing loses one.
The layman capture stops at the first period, not the trailing one. (ANTS-4596) A plain-English roadmap summary containing "e.g.", "etc.", a decimal like 41.5, or a filename like .deb was cut off at that dot when the roadmap was read. The capture now runs to the end of the line and removes only the sentence's own closing period, which is what the two neighbouring trailer keys already did.
bullet_ambiguous no longer advises narrowing by the locator that just failed (ANTS-4574)
When an id matched two bullets and neither carried an anchor, the
refusal said "narrow with anchor or id" — naming two routes that were
both closed. It now names the id as the ambiguous one and points at the
headline locator, or says outright that no locator can address them.
Roadmap render no longer doubles the full stop on a trailer value that already ends in one (ANTS-4582)
Kind:, Source: and Lanes: now get their terminating period only
when the value lacks one. The period belongs to the format, not the
value (roadmap-format.md § 3.5 names Evidence: as the one key rendered
without one), so a value supplied as a sentence was rendering with two.
Package-manager users were two releases behind (ANTS-4587)
All four repositories (openSUSE Tumbleweed and Leap 16.0, Fedora 44, Mageia 10) served 0.7.103 while the current release was 0.7.105. The _service file in git correctly pinned v0.7.105, but the OBS project's own copy still pinned v0.7.103 — the submit step was never run for 0.7.104 or 0.7.105, so OBS kept rebuilding the source archive it already had and reported every build as succeeded.
A bullet listing four lanes no longer stores three (ANTS-4553) Reported against finbreak, where a bullet's prose listed four lanes while the stored column held three. Same cause as ANTS-4542: the list was truncated at the line wrap rather than diverging from a second source. Values already truncated in existing stores are not repaired by this change and are tracked as ANTS-4585.
A hard-wrapped roadmap trailer value keeps the text below the wrap (ANTS-4542)
Every trailer pattern stopped its capture at the line break, so a value wrapped mid-phrase lost its tail: Lanes: MainWindow, stored one lane and dropped TerminalWidget. from the next line, and the render then re-emitted the short value terminated with a full stop, which reads as a correct declaration. A wrapped value now continues onto the following line, stopping at its own sentence-terminating full stop, a blank line, a line opening another declaration, or four lines. No-op on generated content, which is always emitted unwrapped.
A release can no longer be published with no downloadable file (ANTS-4583) v0.7.105 shipped with no AppImage: its build hit a 25-minute job timeout (GitHub reports that as "cancelled") inside the first step, so nothing was built or uploaded, while the release page existed because the tag created it. The ceiling is now 60 minutes — sized against the slowest observed run, 22m36s, not the fastest — and a new final step re-reads the release and fails the run unless both the AppImage and its .zsync are attached.
roadmap_query: an item's notes come back with their own indenting (ANTS-4558) Notes read back through the tool were flush-left even where the file had lists and indented commands, so text read out and written back lost its structure. The block is now dedented by its shared indent only.
Roadmap items keep their indenting: a nested sub-bullet is no longer re-attached to the bullet above it (ANTS-4554) Reading a bullet stripped the indent off every line of its notes, so writing the file back put them all at one level — which in Markdown makes an indented sub-point belong to the wrong parent. The shared indent is now removed and anything deeper is kept, so lists and command lines survive a write. Items imported before this change are already flat; their original indenting is only in git history.
roadmap_log: deleting a Kind: or Source: line from an item's notes no longer fails with a database error (ANTS-4576)
Removing a trailer declaration from a body used to clear the column, and four of the five columns cannot be empty — the caller got a raw NOT NULL constraint failed. What un-declaring means is now the column's own contract: the plain-English line clears, lanes and evidence empty to a list with nothing in it, and kind and source keep their value, which the roadmap file then states canonically. An amend_body new_text is guarded like a note, and a recognised kind is canonicalised rather than refused.
roadmap_query's duplicate_ids[] now reports an authored id shared by two bullets, whatever its shape. (ANTS-4546)
On a roadmap whose bullets lead with a bold span, two bullets sharing
that span both carry it as their id — and roadmap_log then refuses
bullet_ambiguous, so a silence on the read side was paid for at write
time. The detector had been keyed to the canonical [PROJ-NNNN] shape
to stop the reader's own content-hash nonces over-reporting, and that
key excluded the authored ids too; it now excludes only the nonces,
which is what it was for. Read the field as "this id addresses more
than one bullet". A roadmap whose ids are all canonical is unaffected.
A roadmap_log note is prose — a trailer keyword named mid-sentence no longer rewrites that column. (ANTS-4549)
op:"annotate", op:"flip" and op:"flip_batch" append the note to the
bullet's body, and the trailer columns are then re-derived from that body
— so a bare Kind: in a sentence was read as a declaration and its
following words written to the column. Caught on one report only because
the store's CHECK constraint refused an invalid kind; an accepted one
would have written silently, and the other four columns have no CHECK.
A note may now name a trailer key only with the label first on its line,
which is how the render writes a declaration; mid-line it refuses
body_shadowed, quoting the text and naming both remedies. Position
rather than a value check, so the one deliberate use — a Layman: note
filling that column on a migrated item, which the render gates on — keeps
working.
roadmap_log op:"amend_body" amends a phrase that straddles a hard-wrapped line break, in one call. (ANTS-4550)
old_text is now exact apart from its whitespace runs, which span a
line break. The wrapped pass runs only after the exact one finds
nothing, so no previously-succeeding call changed, and the
body_match_ambiguous guard applies to whichever pass matched. This
removes the multi-call hazard ANTS-4097 could only echo. Shares one
matching rule (src/wrapmatch.cpp) with workspace_search, so the two
verbs cannot answer one quotation differently.
ROADMAP.md's version banner no longer silently reverts after a release (ANTS-4529)
The bump recipe hand-edited a version number into ROADMAP.md, which is
generated from the roadmap store — so the next roadmap write of any kind
re-rendered the file and discarded the edit, putting the banner back to
the previous release and re-opening version drift long after the release
commit had looked clean. The banner now states no version at all: the
number duplicated CMakeLists.txt's project(VERSION), and the sentence
carrying it already links CHANGELOG.md, which lists every shipped version
with its date. ROADMAP.md leaves both the bump recipe and the version-drift
gate, so there is nothing left to keep in step.
The sanitizer CI job finishes again, and a timeout now reads as a failure (ANTS-4533) The memory-safety half of CI had been quietly giving up part-way through for two days — cancelled at its 30-minute cap on 36 of the last 40 runs. It did not look like a failure, because a job killed by a timeout is reported as cancelled rather than failed, so nothing went red and nothing notified. Three causes, all measured: the long steps are now wrapped so an overrun genuinely fails; the compiler cache is saved even when the job does not finish (a cancelled job skipped the save, which froze the cache and made every later build slower, which made the timeout likelier); and the sanitized test run is parallel instead of serial. Measured after the fix: 922s to 704s on the tests, and a first fully green run.
Re-importing the roadmap no longer piles metadata lines up inside entries (ANTS-4506)
The migration now drops the trailing run of trailer-only lines from what
it stores, because that block is the renderer's own output rather than
anything an author wrote. Measured before the fix: one re-import moved
599 entry bodies, 458 gaining a Kind: line and 97 a Source: line. A
trailer line sitting ABOVE authored prose is the author's and stays, and
a line is only dropped when it is the sole declaration of its key — the
condition that stops an entry rewriting its own metadata. The
indentation half of the round-trip report is ANTS-4528 and is still
open.
A roadmap entry's metadata can now be corrected by editing the file (ANTS-4505)
The renderer used to keep a trailer line (Kind:, Source:, Layman:,
Lanes:, Evidence:) out of its own output only when the entry's text
repeated that value exactly. Because the renderer appends its block at
the END of an entry and the reader takes the last declaration, a value
read wrong once won for ever: correcting the real line in ROADMAP.md was
silently ignored. Suppression is now presence at the start of a line —
whatever the value — so the file wins and the store follows on the next
import. A mid-sentence mention still declares nothing, and an
unrecognised Kind: value cannot displace a recognised one.
0.7.107 RC2 — Patron preview
Respin of v0.7.107-rc1 with cherry-picked fixes: 143fa73a (ANTS-4718 — roadmap card column grid is now genuinely shared across sections; the previous RC fails the openSUSE/Mageia package build).
0.7.107 RC1 — Patron preview
This release shipped without written notes.
0.7.106
ignored_args is suppressed on refusals, so a wrong-argument call never learns its args were wrong. (ANTS-4525) When a call fails, the tool stops telling you that you also used the wrong option names.
mutation_probe mutates every occurrence with no way to state how many were intended. (ANTS-4521) A test-quality check can quietly change more places than you meant, and still report success.
roadmap_query's bad_mode refusal names the id/ids[] route (ANTS-4511)
mode:"by_id" is the obvious guess for a single-item lookup, and none of
the accepted mode names answers it — item lookup is an argument, not a
mode. The hint fires on every bad mode, not just that spelling.
roadmap_migrate states that it never rewrites ROADMAP.md (ANTS-4482)
markdown_rewritten is on every response and is always false. The file being byte-identical after a migration is expected, not a sign it did not run.
roadmap_migrate says whether the project will be served from the store (ANTS-4490)
store_backed answers it on every response: only an ants-v1 roadmap is served from the store afterwards, so a github-task-list project migrating successfully now says so rather than leaving the caller to notice which fields a later query does not carry. sections_unchanged accompanies sections_written, so 0 written reads as an idempotent re-run rather than as a dead counter.
roadmap_migrate names which items it would update, and which fields (ANTS-4479)
updated_items lists the id and the changed column names for each item items_updated counted, capped at 200 with updated_items_truncated. Under dry_run that turns a bare count into a plan you can review before running it.
apply_edits says where the occurrences are when an edit matches more than once. (ANTS-4473)
An ambiguous skip was correct and silent about location, so recovering from it meant re-reading the file. It now carries candidates:[{line,text}] — the same field a not_found near-miss already returns — plus match_count for the true total and candidates_truncated when the ten-entry list is capped. Recover by extending old with surrounding context, by naming the occurrence with the start_line/end_line form, or with replace_all:true.
roadmap_log and changelog_log accept each other's word for "add an entry" (ANTS-4475)
One spells it append and the other add, and each already accepted the
other's word for the batch version. Both now accept both. Error messages
still name the canonical form.
section_index can return just the slugs, which no longer overflow on a large roadmap (ANTS-4467)
Pass slugs_only:true to get a flat list of section slugs. On roadmaps
big enough to overflow the reply budget the full form returned only the
first handful of sections — and slugs are exactly what roadmap_log
needs to file a bullet.
a refused cross-project feedback-file read now names the call that works (ANTS-4421) The refusal is correct — the file lives outside the project — but it was a dead end. It now names the folder to use as the working directory.
searching with HTML entities now warns instead of silently finding nothing (ANTS-4420)
A pattern written as <h1> rather than <h1> found nothing and read
as a confident answer; it now asks whether the literal characters were
meant.
a failed text edit now points at the line that nearly matched (ANTS-4418) When an edit misses only because of spacing, apply_edits reports the line number and that line's exact text, so the retry needs no re-reading and no guessing at the whitespace.
an empty code map now explains why it is empty (ANTS-4419) Instead of a bare "empty" flag, codebase_index says whether the project was never registered, whether its declared source folders hold no code, or whether there genuinely is none — and names the folders that do hold code.
Feedback v2 gains an "awaiting reporter" triage state (ANTS-3631)
assign_id takes a third disposition, awaiting, writing
_(awaiting reporter — <question>)_. It stays un-triaged on purpose so
the question reaches the reporting session's feedback_query delta, and
it is classified before the id and closure tests so a question quoting
an id is not mistaken for an assignment. feedback_query emits
awaiting[]; session_orient separates the maintainer's outbox from
their inbox.
Layman: When I triage another project's bug report and need to ask
them something, there is now a way to ask it that they will actually
see.
build_target_for — which build target owns a source file (ANTS-3745)
New read-only MCP verb. Parses CMakeLists.txt and answers with the
owning target, the cmake --build --target line, and the ctest -R
filter for the source's gtest suites. Replaces an awk over
CMakeLists.txt and a --gtest_list_tests launch per bundle; the
bundle mapping is not guessable from the path.
Layman: Ask "which test program do I rebuild after editing this
file?" and get the exact build and test commands back.
spec_lint falls back to the global spec-format standard (ANTS-4080)
A project carrying only a delta was linted against nothing, and the skip
read as a clean structural result. The four candidate paths are now
retried under ~/.claude/ through the same ~global re-rooting
doc_integrity uses; sections_source prefixes a global hit with
~global/. Projects shipping their own standard are unaffected.
Layman: The spec checker now falls back to the shared standard when a
project has no copy of its own, and says which one it used.
spec_lint: a document can declare a deliberate invariant-id floor (ANTS-3784)
A spec that continues a sibling's numbering can now say so with an
<!-- invariant-id-base: N --> line outside fenced code; missing numbers
below N are counted as suppressed instead of reported. Measured over
docs/specs/: 39 gap findings became 28, one document moved.
Layman: The spec checker kept flagging deliberately skipped numbers
as mistakes; a spec can now say the skip is on purpose.
roadmap_query now says which backend answered — source: "store" | "markdown" (ANTS-4402)
The store has been primary for roadmap reads since ANTS-3793, while ANTS-4141's
standing workaround has roadmap edits made by hand to ROADMAP.md — so a hand
edit lands in a file no reader reads, and the envelope gave no way to tell: it
named ROADMAP.md as its path, returned ok:true, and served the store.
file_ahead_of_store (with file_highest_id and store_high_water) fires when
the file holds ids the store has never seen. One-directional by design: it
proves staleness, never freshness, because a status flip or a body edit moves no
id. Reconciling the two backends remains ANTS-4141.
co_change_family — every edit site of one settings field, in one call (ANTS-3368)
Adding a setting by copying an existing one is a ten-file lockstep
change, and the easiest sites to miss are the ones that are not
spelled like the field: the text key in the config file
("claude.mcp_enabled") and the helper names built from it
(setClaudeMcpEnabled, m_claudeMcpEnabled, setMcpEnabled). Ask
for one example field and this lists them all, grouped by file, each
line tagged with what it is. Measured on this repo's own
claude.mcp_enabled family: it spans 11 files under src/, where
searching for the whole name finds 6 — and one of those 6 is found
only by a comment, while the two lines you actually have to copy in
that file match nothing at all.
changelog_log op:normalize now tidies stray prose, not just category order (ANTS-3381)
A paragraph left stranded between the category blocks of
[Unreleased] is folded into the entry above it, in the same
one-shot tidy that puts the categories in order. Every fold is
listed back to you as moves[], under dry_run as well as on a
real write, so you can see what would be absorbed before it is.
A line separated from its entry by a sub-heading is left alone
and still reported. Completes the half ANTS-3495 deferred.
cited_by — one call that says which document cites which anchor (ANTS-3716)
Give it the names a change touched and it reports which documents
mention which of them, with an occurrence count and a first line
per pair — plus the anchors nothing cites, which is the half a
caller cannot infer from an absence. One ripgrep run per anchor, so
each match belongs to its anchor by construction. Replaces a
per-anchor search loop a review pass otherwise runs by hand four
times a round. Scope defaults to the project's docs directory,
README.md and CLAUDE.md.
doc_integrity catches a skill that calls an MCP tool it never granted itself (ANTS-3719)
A Claude Code skill lists the tools it may use in its frontmatter and
names the ones it needs in its instructions. When the two drift the
skill cannot be run as written — and the skill that hit this twice was
a review skill, so the check meant to catch everything else was itself
broken. A new ungranted_tool finding reports any mcp__ants__ verb
the body calls that the allowed-tools list omits. It fires only on
files that actually carry that frontmatter key, so ordinary
documentation never sees it. doc_integrity also accepts the
~global / ~claude-config root that search and outline already
take, which is what lets it read the skill tree at all.
A spilled read_region emits a headless rows_preview that costs ~1k tokens and conveys nothing. (ANTS-4519) A big read sends back a long list of empty rows that tells you nothing but still costs money.
roadmap_migrate's description now states that it never writes ROADMAP.md, and which roadmaps are served from the store. (ANTS-4482)
A byte-identical file and a clean git status immediately after a migration are expected — the file is first re-rendered by the next roadmap_log write — and only an ants-v1 roadmap is served from the store afterwards, so a github-task-list project migrates successfully while roadmap_query keeps answering from markdown. The description also records that the store is machine-global rather than per-project. (ANTS-4490)
roadmap_migrate's selection hint says that re-running is the routine way to reconcile a drifted store. (ANTS-4480)
It read "Use ONCE per project", which two sessions read as one-shot-and-dangerous — one nearly skipped the re-run that reconciled its drift, the other hand-edited a 616 KB generated ROADMAP.md instead. The hint now says re-ingesting is routine and idempotent, names dry_run as the preview, and states that the verb never writes ROADMAP.md.
focused_test now points at mutation_probe (ANTS-4472)
Three sessions in a row hand-wrote the mutation loop that verb replaces,
because nothing surfaced it. The session-start list is full to within 10
bytes of its cap, so the pointer lives on focused_test instead — which
is where a session running tests will be looking anyway.
roadmap_query reads ROADMAP.md once per call, not once per branch (ANTS-4431)
A cold section= query opened and read the whole roadmap twice — the
heading index and the per-section ETag each built their own file
reader, in the same function, so neither could reuse the other's
memoised body. One reader is now created per call and every branch
shares it, folding five readers into one.
roadmap_log op:append no longer reads ROADMAP.md to check for near-duplicates (ANTS-4426)
On a migrated project the ANTS-2043 advisory took its records from the
whole 3.8 MiB file; it now takes them from the roadmap store, so the
append path reads no document body at all. No answer changes — a file
carrying a bullet the store has never imported is refused outright, and
that is the only input on which the two sources could differ.
Roadmap reads dispatch before loading the file, so a migrated project no longer pulls 3 MiB of ROADMAP.md to ask which backend serves it (ANTS-3863)
The read seam's markdown parameter became RoadmapSource::RoadmapText, a
move-only provider that reads lazily and only as far as the caller's path
needs: the dispatch spends the detector's window — at most 300 non-blank
lines or 1 MiB, whichever comes first — and the whole body is read only on
the unmigrated path that was always going to parse it. Thirty call sites
converted; no verb's response shape, refusal code or message changes.
the session startup tool list now spells out that read_regions batches (ANTS-4422) It read as a plural of read_region, so sessions made several calls where one would do.
the roadmap write refusal now says where backticks must go (ANTS-4424) "Wrap the key in backticks" was true only if they sit immediately around the key; a key inside a longer code span is still read as metadata, so a caller who already had backticks was refused twice with the same advice.
Roadmap dialog — the filter bar collapses to three controls (ANTS-4412)
Sixteen checkboxes across two rows became Status: all ▾,
Kind: all ▾ and Reset filters, with the density combo moved up to
the search row. The reset button is enabled exactly when something
narrows the list, so a collapsed control cannot hide why the list looks
short. Keyboard access is unchanged.
Layman: The cluttered row of options at the top of the roadmap
window is now three tidy buttons.
Audit: the Contract-Doc ↔ Code Drift lane no longer reports any head:value citation (ANTS-3849)
The 2026-08-13 guard covered a path-shaped head only (remotecontrol.cpp:2540);
it now covers any head, so a symbol:line citation (applyTheme:3118) and a
JSON field-and-value fragment (sections_checked:false) are skipped too.
Source spells the head and the literal separately, so the joined form can never
resolve. Measured before shipping: 128 of the 129 distinct heads still reaching
the guard resolve in source, so the finding carried no information about its
head. 175 of 1,175 findings, 14.9%; verdicts diffed old-vs-new over the whole
corpus with 0 newly firing.
A filename mentioned in an entry's Source line is no longer reported as a missing file. (ANTS-4502)
The import treated a token as a path if it carried EITHER a directory
separator OR a filename-shaped ending, and either half alone is a bad test.
The ending alone flags every bare filename and every §4.4-shaped fragment:
11 of 12 notes on one project, every file existing one directory down. The
separator alone flags every prose slash: measured over this project's 1,781
entries it accepts 97 tokens of which 2 are real paths, the rest being
slash-command names, id pairs and phrases like precision/convenience.
Requiring both gives 1 candidate, 1 real, 0 false reports. A path with no
extension — a bare directory — is no longer checked, which is the whole cost.
A roadmap entry that quotes the format no longer has its own example read as real data. (ANTS-4504)
The guard against a quoted trailer key was three fixed-length lookbehinds, so
it only saw a backtick one to three characters before the label. An entry
whose example wrapped across two lines had the second key parsed as a real
declaration — ANTS-3808 gained a Lanes: value it never wrote. Every key is
now matched against a masked copy of the body in which inline code spans are
blanked, with values taken from the original text so a real value keeps its
own backticks. Span boundaries come from the project's existing CommonMark
primitive rather than a fourth hand-rolled pairing.
roadmap_log dry_run returns the next id under the same key the real write uses, so it reads as a reservation. (ANTS-4508) The preview tells you which number your entry will get, but does not hold it — and the wording makes it look held.
read_region section= and file_outline cannot see a heading inside a blockquote. (ANTS-4520) A heading written inside a quoted block is invisible to the tool, so the one section you were told to read cannot be addressed by name.
read_regions refuses conflicting alias arrays instead of picking one (ANTS-4512)
Sending both paths and regions silently used one and then reported a
shape error against the other. It now refuses bad_op_combo naming both,
and a successful call echoes items_key.
read_log no longer reads as a git-log reader in the session bootstrap (ANTS-4510)
It sat between git_state and model_switch_stats described as a
"filtered log tail", and now says Ants debug-log tail (NOT git log).
session_orient honours fields= (ANTS-4523)
It had accepted and silently dropped the argument — ~16k tokens for a
three-integer answer, on the documented first call. Narrowing trims the
payload, not the work: the eager codebase_index refresh still runs.
Bash guard-hook now routes a --stat diff to git_state op:diff (ANTS-2169)
It had named get_git_status and roadmap_query, neither of which can
return changed lines, so obeying the block led to a dead end and then to
the bypass token. The status/log branch keeps get_git_status.
roadmap_log now reports the hand-edits its re-render overwrote (ANTS-4462, ANTS-4465)
On a migrated project every roadmap_log op rewrites the whole ROADMAP.md
from the store, so anything typed into the file by hand — the preamble
above the first heading especially, which no verb can write — was
silently put back. The write now renders the store as it stood before
the change, compares that against the file, and reports
discarded_external_edits with a line count (would_discard_* on a
dry run). It reports rather than refuses, so one hand-edit cannot block
every roadmap op on a project. Expect one true report per project on
the first write after a migration, where the render canonicalises
formatting the migration kept verbatim.
A roadmap migration no longer overwrites a stored field with a value it invented (ANTS-4498)
Where a roadmap file does not declare an item's Kind: or Source:,
the import supplies a default. Re-running a migration then wrote that
invented value over whatever the database already held, so re-importing
a project could quietly replace real information with placeholders — on
this project's own roadmap, 384 items were exposed to it. A defaulted
value is now withheld from any field that already holds one, and the
migration reports each withheld write. The effect is that a re-run can
only add information, which is what makes it safe to run without taking
a backup first.
The roadmap's id counter cache no longer drifts behind the database (ANTS-4141) Filing an item on a migrated project allocates its number from the database and left the on-disk counter untouched, so the counter fell further behind with every item — measured at 99 behind on this project. Anything still reading it, such as a fresh clone or a hand-written entry, would then pick a number already taken. It is now refreshed on each write, as a best-effort cache update that cannot fail the write itself.
A roadmap item that quotes an example above its own details no longer imports the example's values (ANTS-4497) Four of the five detail lines an item carries were read from the FIRST place they appeared in the item, so an item quoting a sample block — or whose title happened to contain one of the labels — was filed under the sample's values instead of its own. All five now take the last one written, which is the rule the import contract already specified and which only one of them followed.
The first roadmap item filed after a migration no longer reuses an id the database already holds (ANTS-4493)
When a project is migrated, items with no id of their own are given one.
Those ids live only in the database, and the code that hands out the next id
was reading only the roadmap file — so the next item filed took a number
already in use. It now checks the database too. counter_advanced_past is
reported alongside counter_advanced_to, so the response says what the
number was skipped over rather than only where it landed.
A long roadmap headline no longer splits its item in two when the file is rewritten (ANTS-4484) A headline that wrapped across two lines was stored with the line break inside it, and written back out with the second half flush against the left margin — where markdown reads it as a new bullet, so the item visibly split and the trailers below it were then read as belonging to the fragment. The headline is now joined into one line both when it is read into the store and when it is written back, so an existing roadmap with the damage in it is repaired by the next write.
roadmap_migrate's preview reports the real project id instead of 0 (ANTS-4478)
A dry run over an already-migrated project answered project_id: 0 beside counts that could only have been computed against that project's real rows. It now reports the store's id for that root on both paths; 0 means only that the root is not registered yet, so a preview answering project_id > 0 reads as "already migrated".
roadmap_migrate no longer reports a resolvable path as missing when the provenance is a sentence (ANTS-4481)
A Source: value is prose far more often than a bare path, and the whole
value was being resolved as one filename — so every item whose provenance
mentioned a real file was reported unresolved. The unit is now the
whitespace-delimited token, as the import contract always said. A token
that ends a sentence also gets its punctuation stripped on a retry, and
the note points at the line carrying the path rather than at the bullet's
first line.
codebase_index can see HTML pages, so a website project's code map is no longer near-empty. (ANTS-4425)
file_outline has outlined HTML since ANTS-4361 while the index gate did not admit it, so on a site of ~40 HTML/CSS/JS files the index counted 9 — the .js and .py — and every page was invisible, to codebase_index and to the indie_review computed partition alike. Both now key on one shared predicate. CSS stays out until it has an outline mode, because counting files the outline cannot read is the same drift in the other direction.
project_layout finds an AppStream metainfo file kept one level under packaging/ or pkg/. (ANTS-4439)
Every probe was single-level, so a project packaging for more than one distro — packaging/obs/ beside packaging/flatpak/ — reported no AppStream metadata at all, and a release path keyed on that field silently skipped its release-notes sync. The legacy *.appdata.xml spelling is now accepted too, probed after every *.metainfo.xml candidate so the current name always wins. Flat candidates are still tried first, so no project whose file already resolved resolves differently.
changelog_log op:"add_subsection" no longer refuses a dated changelog because of a legacy category tail at the bottom. (ANTS-4489)
The guard classified ## [Unreleased] by first match, so any Keep-a-Changelog category heading anywhere in the section made the whole section "flat" — refusing a section of ~500 dated topics on the strength of six legacy headings 10,000 lines below them. It now classifies by position: a dated topic above the first flat heading means a dated section with a legacy tail, and the insert (which lands at the top) is accepted. The refusal that remains reports what it found — MIXED, with both populations and their line ranges — instead of asserting the section is flat, and no longer advises the op that created the mixture.
file_outline says how many symbols a max_symbols cut left out (ANTS-4469)
Truncation reported only that it had happened, so there was no way to
tell one missing symbol from four hundred. It now emits
symbols_dropped, the same field the byte cap already used.
spec_query declares its mode argument, and rejects a misspelled one (ANTS-4468)
mode was accepted but never declared, so every call passing it was
reported as having an ignored argument. A misspelled mode also fell
through to the list behaviour and answered a drift question with a spec
list; it now refuses and names the valid values.
roadmap_log dry runs no longer report the write as done (ANTS-4463)
A preview returned files_written and note_appended:true for a write
that did not happen, so a caller checking either field read a dry run as
a completed one. Those names are now absent on a dry run and the values
arrive as would_write / note_would_append, matching what
changelog_log already documents.
SARIF export labels suppressions with status, the v2.1.0 field name (ANTS-4454)
The suppression object wrote state, an earlier-draft spelling that
no v2.1.0 validator accepts — strict consumers rejected the object and
lenient ones read no status at all. The feature spec and its
conformance test carried the same error and are corrected with it.
doc_symbols now reports its qualified unresolved-span count (ANTS-4443)
The classifier tested the scope-stripped needle for ::, which it can
never contain, so every qualified non-call span was filed as bare
and the counter was structurally always zero. It tests the span.
Lua: registering an event handler from inside a handler no longer corrupts memory (ANTS-4441)
LuaEngine::fireEvent iterated the live handler vector inside
m_handlers while lua_pcall ran a handler, so a plugin calling
ants.on() from a handler could reallocate that vector (same event)
or rehash the QHash (new event) out from under the loop. It now
iterates a copy taken before the dispatch begins — which also fixes
the contract: a handler registered during a dispatch runs from the
next dispatch, never the current one.
A duplicate id_fold refused a whole project as a bare SQL constraint, naming neither bullet. (ANTS-4428)
roadmap_migrate now refuses a duplicate id before the insert, naming
both path:line sites and the folded identity, instead of aborting the
whole project with a bare UNIQUE constraint failed: item.project_id, item.id_fold.
a section search that matches nothing now says the search emptied it (ANTS-4423) It previously reported that no item carried an ID and suggested two settings that cannot help. The same fix shipped for whole-file searches in July and had never been applied to single-section ones.
roadmap_query no longer returns a body fragment as a bullet's headline (ANTS-4417)
On the **ID** headline bullet shape (id bolded, headline plain), the
headline search ran past the bullet's first line and adopted whatever the
body bolded next — so a listing could read "already received" or "fixed
2026-08-02" for open items. Also stopped a latent migration fault: the same
offset is the boundary the roadmap import strips at, so importing such a
project would have dropped real text.
mutation_probe's green-baseline gate no longer passes when it cannot tell (ANTS-4401)
require_green_baseline tested for red and read everything else as
green, so a baseline whose output could not be parsed, or that ran no
tests at all, satisfied it. Both now refuse with baseline_unreadable,
naming which of the two it hit.
Layman: A safety check meaning "stop if the tests are already
broken" quietly did nothing when it couldn't tell whether they were.
spec_lint's skip hint no longer states a false cause on an empty walk (ANTS-4080) A run that matched no document reported "no required-sections block was found", which is a claim about the format standard that the run never tested. It now says no document was checked.
Roadmap migration no longer loses a quarter of the roadmap to a line of prose that quotes fence syntax (ANTS-4403)
The migration walk decided code-fence extents with its own
startsWith("```") test instead of the shared MarkdownScan rule, so a
multi-backtick inline span — the standard way to write *about* fences —
opened a fence nothing closed and masked the rest of the file. Bullets below
it were never recorded and headings never opened a section, with no note: the
plan was well-formed and simply short. On this project's own ROADMAP.md one
such line at 31,081 dropped the plan from 2,040 items to 1,559 and from 218
sections to 135 — 481 items, 24% of the roadmap. It was also the whole of the
446 "orphaned" store rows that had been read as a policy question about rows
the file no longer explains; the dry run now reports 0. The walk takes
MarkdownScan::fenceMask(), so the fence rule is stated once.
roadmap_log no longer deletes bullets the store has never imported (ANTS-4141)
The store-backed write path renders the whole roadmap from the store
on every op, which assumes the store holds everything the file does.
Where it does not, a bullet filed by hand between two verb calls was
simply absent from the render and the publish removed it with no
trace. RoadmapWrite::commitAndRender() now compares the dry
render's id set against the ids the files it would rewrite hold
today and refuses render_would_drop, naming them, instead of
publishing. It sits at the one seam all eight ops share, so none can
bypass it, and a dry_run preview reports the same refusal.
Re-importing a roadmap no longer aborts on an id the store and the file both claim (ANTS-4142)
A migration re-run refused the whole project with UNIQUE constraint failed: item.project_id, item.id_fold when a bullet's hand-filed id had
already been synthesised by an earlier run for a different, id-less
bullet. Two causes: matching walked the plan once in document order, so
the weaker id-less key could consume a row the id itself named (now two
passes, id-bearing first); and allocation took the store's high-water
instead of the source file's highest id rather than the higher of the
two, so it could re-issue an id the file was already using.
feedback_log op:compact_resolved no longer lists the same id twice (ANTS-3739)
An id named twice on one **Proposed ID:** line now reports once, matching op:assign_id.
feedback_query again reports mapped_ids for a fully-condensed feedback file (ANTS-3744)
A file tidied down to its ## Tracked in ROADMAP … pointer line returned no ids, so the reporting project could not see whether its own findings had shipped. Inline ids still win where a file has them.
Audit: the Contract-Doc ↔ Code Drift lane no longer reports path:line citations (ANTS-3849)
A remotecontrol.cpp:2540-shaped token can never resolve — the path index holds no line numbers — so half the category was noise by construction. 1,146 of 2,315 findings on this repo. Citation staleness is doc_citations' job.
0.7.106 RC1 — Patron preview
This release shipped without written notes.
0.7.105
The four shared standards now carry their full text in-repo, gated against drift (ANTS-4133)
docs/standards/{coding,documentation,testing,commits}.md keep their
project-specific delta and now mirror their global owner verbatim below it;
security.md joins them under the same markers. This repo is public, and a
pointer to ~/.claude/standards/… left the rule simply absent for anyone
reading on GitHub. tools/check-standard-mirrors.sh (--check / --write)
compares each mirrored region against its owner and tools/hooks/pre-commit
refuses a commit that has drifted; a checkout without the global tree skips
and passes.
spec_lint resolves the test surface a spec cites, instead of reading it (ANTS-4127)
A spec's Test: clause claims an invariant is locked by something real, and
nothing checked that the thing it names exists — three specs in this corpus
name test directories that never have, two of them claiming to have shipped.
spec_lint now resolves a cited tests/features/<name> against disk and
against CMakeLists.txt, adding test_surface_absent, test_surface_unresolved
and test_surface_unwired, plus surfaces_resolved / surfaces_checked on
the envelope. The bucket is chosen by the spec's own Status:, so only a
shipped one yields a finding and a draft's forward reference does not; a
wildcard (tests/features/audit_*) names a family and is not resolved. The
engine still opens no file — both filesystem facts are injected by the verb
layer, gathered once per run — and an empty set means skip, never fail.
Qt 6.2 floor guard now runs before every push (ANTS-4131) (ANTS-4131)
New tools/qt62-guard.sh compiles the whole project against the
oldest supported Qt inside a container, and the pre-push hook now runs
it whenever a push touches compilable source. Both the toolchain layer
and the build tree are cached, so a check that used to take about 25
minutes now takes 5-7 seconds — and rejects a too-new Qt call in about
one. This is the failure that broke CI three times on ANTS-4108 and
could not be caught on this machine at all.
spec_conformance — an MCP verb that RUNS the regex patterns a spec prescribes against the | input | expected | table beside them (ANTS-4108)
A cold read passes a wrong pattern; running it does not. A row whose
actual result differs from expected comes back as a finding, a
```regex fence with no table beside it as a candidate, and a per-case
timing as an observation — which is what catches a fixture that returns
before the pattern under test ever runs. Executes patterns only, never
fenced code: a pattern applied to a string cannot reach the filesystem
or the network whatever it contains. pcre2 only; any other engine tag
refuses per fence rather than substituting one. max_cases outside
[1,1000] refuses the call rather than clamping, so a partial run can
never read as a complete one. Nothing in this repository uses the
convention yet, so it reports nothing here today — ANTS-4128 (the
re.search call form) and ANTS-4127 (executing fixtures) are what make
it earn its keep.
indie_review_partition reports each lane's real size and flags lanes too big to review (ANTS-4100) A module map naming one directory produced a single lane covering a whole 21,000-line application, and nothing in the response said so. Each lane now carries a measured file count, and one over 30 files is marked too_coarse with a hint to split by cohesion. The threshold is set above this project's own largest lane so a good partition never trips it.
doc_integrity takes a single path, so a post-fix sweep cannot be scoped to the files a run actually edited. (ANTS-4106)
After fixing five files you can only re-check one folder or the whole tree, so unrelated old problems get blamed on your change.
Archive a released version from the roadmap store, and rename a section (ANTS-4070)
roadmap_log op:"rotate_minor" moves a closed minor's headings and
everything under them into docs/roadmap/<M>.<N>.md, and
op:"retitle_section" renames a heading. Both re-derive the section's
internal address the way a re-read of the file would, so the roadmap
still matches itself afterwards. Rotation refuses a minor that still
has unfinished work in it, and previews with dry_run.
The roadmap store records which roadmap dialect each project was migrated from, and reports it when the file later disagrees (ANTS-3815) The database now remembers whether a project's roadmap was written in Ants' own format, a GitHub task list, or pass headings, instead of working it out from the file every time. If the file is later rewritten in a different dialect behind the database's back, Ants says so plainly and points you at re-running the migration, rather than quietly guessing which of the two to believe. Projects migrated before this release carry no recorded dialect and behave exactly as they did.
This is also the store's first schema-version bump (1 → 2), so an existing store is upgraded in place on first open, keeping all its data. An older Ants will refuse to open a store this version has upgraded — relaunch the newer one.
roadmap_migrate — the MCP verb that loads a project into the roadmap store (ANTS-3855)
The migration engine has been complete and unreachable since ANTS-3756:
schema, read half, load half, archives, render, export, section
provenance, consumer cutover and write half all shipped, and every
invoker was a test. This is the production entry point. One project per
call, resolved from caller_cwd; optional project_name /
export_slug default from the leaf directory (slug derived, a supplied
one validated verbatim and never rewritten). dry_run:true reports the
real run's counts and rolls back — it does open the store, and creates an
empty schema if none exists, because those counts are a diff against
existing rows. Re-running an unchanged project is idempotent; re-running
with a different name or slug is refused rather than silently applied.
Refusals: no_project, no_roadmap, case_ambiguous, not_utf8,
format_mismatch, bad_args, slug_collision, store_failed,
migrate_failed (the last three codes are new to the taxonomy). The
handler is deliberately thin — it resolves the root, stamps the clock
once and names RoadmapStore::defaultPath() — with all store work in a
free RoadmapMigrateVerb::run(storePath, req) seam in its own
translation unit, so the feature test drives it against a temp store
instead of the developer's real one. 11 feature tests, INV-1..INV-10.
Roadmap write half — roadmap_log's eight ops mutate the store, then re-render. (ANTS-3809)
On a project migrated to the roadmap store, all eight roadmap_log
ops now mutate the store and re-render, instead of splicing markdown.
Unmigrated projects, and migrated ones whose roadmap is a GFM or
pass-headings dialect, are unaffected. Adds the render_gate_unmet,
render_failed, store_failed, locator_unsupported and
body_shadowed refusals.
Dropped 52 stale file references from the remote-control code, cutting build work (ANTS-4125) When this area was split into a dozen smaller files, the code that used each helper moved out but the line naming that helper stayed behind — so the main file still pulled in 48 helpers it never touched. Every candidate was removed and then re-compiled to prove it was genuinely unused; the five that turned out to be needed were put back. No behaviour change. (Also closes ANTS-4123, the single-file version of the same thing.)
test_audit_partition's dimensions:"auto" no longer includes the five test-STYLE dimensions. (ANTS-4111)
They are still accepted explicitly via csv:, and now appear in dimensions_skipped[] with a reason. dimensions_active[] is what /test-audit seeds subagents from, so returning all 18 briefed reviewers on exactly what the skill had removed.
Split the remote-control implementation into eleven source files (ANTS-3833) The MCP verb implementation was one 24,752-line file — slow to rebuild, and awkward for anything that reads it. It is now eleven files, one per verb family, with no change to what any verb does. A new conformance test keeps the split honest: it checks the files stay in cut order, none grows back past 6,000 lines, and nothing goes back to reading just one of them.
The audit window now recognises Qt projects that keep their code in the top folder (ANTS-4124)
Two separate pieces of code decided whether a project was a Qt one, and
they had already drifted apart once — that drift was the bug fixed in the
previous release. They are now one piece. The window inherits the better
version as a result: a Qt project whose source sits in the top folder
rather than a src/ subfolder is now detected, so it gets the correct
analysis settings instead of being treated as a plain C++ project.
The push safety check no longer starts a sanitizer build it cannot finish in time (ANTS-4118) Pushing from an automated session could be cut off part-way through the second half of the check, failing the push and leaving the sanitizer build tree half-done. The check now measures how much work is pending first and, if that is more than a couple of minutes' worth, says so and skips that half rather than starting it (the cloud build still covers it). If it does get interrupted, it marks the tree and the next run tells you the one command that repairs it. It also now tells you up front how to skip that half, instead of only mentioning it once you already had.
The file map now shows top-level data blocks in JavaScript and similar files (ANTS-4090) A quick map of a file listed its functions but skipped plain top-level values — which, in a file that keeps big blocks of text or styling in a named value, are the largest parts of it. One reported file showed a 250-line hole. Those entries now appear, labelled as data rather than code, with only their name shown so a 95-line block never lands in the reply. You can also pull one up by name now.
Opening one roadmap item by ID now returns its whole body, not the ends with the middle cut out (ANTS-4091) A long item was trimmed to its first and last part with the middle replaced by a marker — and the middle is where a resume plan lives. Asking for a specific item by ID now returns it in full (up to 16 KB), because naming the item has already kept the reply small; asking for many at once still trims, so a wide query cannot balloon. The trim marker also now says how to read the omitted part.
roadmap_query mode:"bundles" no longer labels a group with punctuation and filler (ANTS-4088)
The "what should I work on next" view opened with labels like - be bookmarked or 57 blocks duplication: — leftover quotes, stray hyphens, bare numbers and filenames scraped out of the item titles. Group labels are now cleaned the same way the grouping itself already cleaned them, so they read as words.
The bug scanner no longer goes blind on non-Qt C++ projects (ANTS-4094)
Ants told the scanner to expect Qt code, always — because Ants itself is
a Qt app. On a project that isn't, that setting makes the scanner treat
the word emit as special. Any file using emit as an ordinary variable
name then failed to read at all, so the scanner found nothing in it and
the report still looked complete. One real project lost every finding in
its largest file this way: 10,000 lines, zero results. It now reports 59.
Ants now checks whether a project actually uses Qt before turning that setting on, so Qt projects keep the Qt-aware scan and everyone else stops losing whole files.
The audit's changelog-coverage check now recognises tests linked by ticket number (ANTS-4099) On a project that ties each release note to its tests by ticket number, every entry was reported as untested even though all of them had tests. The check was comparing the wording of the release note against the title of each test contract — but a release note is written for users and a test title for developers, so the better the note read, the worse it matched. When an entry ends with its ticket number, that number is now looked for in the tests directly. Entries without one are matched by wording as before, and an entry whose ticket appears nowhere is still reported.
The audit's spec-drift check no longer flags a test contract for naming its own test file (ANTS-4098)
A per-feature contract that says "covered by test_min_size.py" was
reported as referring to something that does not exist, even with that
file sitting in the same folder. The check compared the name against
the contents of every file in the project, and a file's name is not
written inside itself. It only ever worked here by luck, because this
project's build file happens to list every test by name. The check now
also looks at the list of files that exist, so naming a real file is
enough. A name that matches nothing at all is still reported.
spec_log set_status now reports the value it replaced, and its schema no longer teaches one project's Status vocabulary to every project (ANTS-4114)
The verb writes the caller's status string verbatim and validates no
vocabulary — but its schema offered "accepted (2026-06-03)" as the
example, which is Ants' dated lifecycle. A project whose standard defines
its own permitted values got a spec its own lint gate then failed, minutes
after the verb reported success. set_status envelopes (and dry_run
previews) now carry previous_status, the replaced value joined across
its continuation lines, so a vocabulary mismatch is visible at the call;
the schema states that no vocabulary is imposed and points at that field.
No value is refused — the permitted sets live as prose in each project's
own standard, in shapes that share no grammar, so an extractor would
sometimes refuse a correct write.
Review lanes now cover hand-written shader code instead of the compiled arrays beside it (ANTS-4096) A shaders folder was partitioned into 22 generated byte-array headers and none of the 19 shader sources next to them: the generated-file filter only recognised name prefixes, and shader files were missing from the project's indexable-file list entirely (find_definition and file_outline had both supported them for months). The same fallback also suggested 38 nonsense lane merges built from its own boilerplate text; those are gone.
indie_review_corroborate now resolves citations that name a file without its directory (ANTS-4095)
Reviewers cite the filename a brief showed them, so a report saying d_main.c:1049 for a file at linuxdoom-1.10/d_main.c resolved to nothing and the whole cross-check silently returned "no agreement". A unique basename now resolves; an ambiguous one still does not, since guessing between two same-named files would invent agreement. The response also reports how many citations it saw versus resolved, so a parse failure can no longer pass for a clean result.
audit_falsepos_log's ledger suppresses nothing in audit_run — the writing verb and the reading sweep use different files. (ANTS-4105) You mark a false alarm as dismissed and the next scan reports it again, because the note was filed where nothing reads it.
changelog_log's interleaved-prose advisory fires on trailing HTML comments and on the continuation bodies its own body parameter writes. (ANTS-4103)
The changelog tool warns that your file is malformed — about text the same tool wrote, in the shape it was designed to write.
project_settings op:detect under-reports itself — it suggests more keys than documented, and says "no override needed" while five keys sit undeclared. (ANTS-4093) The setup tool tells you there's nothing to configure when five settings are still unset.
project_settings op:detect suggests gitignored directories as source_roots, including a 2 GB state directory. (ANTS-4092) The tool that is supposed to stop the guessing suggested pointing the code index at 2 GB of session transcripts.
roadmap_log op:amend_body is single-line by design, so two successful calls can leave a wrapped paragraph incoherent with nothing in the envelope to show it. (ANTS-4097) Fixing two lines of the same paragraph one at a time can produce a sentence that contradicts itself, and both edits report success.
roadmap_log op:append renders a non-conforming bullet on a pass-headings roadmap — Status: todo, unindented body, no separator. (ANTS-4117) On one project the roadmap tool writes entries in the wrong shape, so every append needs fixing by hand afterwards.
roadmap_log's id locator matches zero bullets on a roadmap whose ids are bold (**LOTTO-0019**), while roadmap_query resolves the same id fine. (ANTS-4109)
Marking items done silently does nothing on some roadmaps, and still reports success.
Test chunks are bounded by bytes as well as file count. (ANTS-4113)
A suite whose tests live in a few very large files returned one chunk no reviewer could read against the full dimension list. The applied budget is echoed as chunk_byte_budget.
The cpp_exit and system_shell_out pre-pass patterns only fire on C/C++ files. (ANTS-4112)
They were matching English prose inside a quoted string in a shell script, and Python's canonical sys.exit(main()) — half the pre-pass signal on one project.
doc_citations says which counts keys are overlays, so the status buckets can be summed. (ANTS-4087)
counts mixes statuses with overlays on the ok subset, so it never summed to count and a clean pass read as half-verified. counts_overlay_keys names them.
feedback_query reports "nothing filed here yet" as a result, not an error. (ANTS-4104)
A derived path with no file returns ok:true / found:false with the candidates and hint intact; an explicitly passed path that does not resolve still refuses.
spec_lint no longer reports an invariant id gap for a number a sibling spec owns. (ANTS-4110)
On a project that numbers invariants once across the corpus this fired on every neighbouring id. The scheme is detected from ownership, and suppressions are counted in id_gaps_suppressed rather than being silent.
Sub-lettered invariant ids (INV-3b) parse as invariants of their own. (ANTS-4107)
Their text used to be absorbed into the preceding invariant's body, and invariant_no_test could not see them at all — so an invariant split by a review could ship with no test surface and pass the lint.
spec_query / spec_lint find an Invariants section whose heading does not begin with the word. (ANTS-4115)
## 5. Correctness invariants matched nothing, so every invariant under it was invisible — a whole project's specs returned an empty contract. Also picks up four Ants specs (## 4. Contract — invariants, ### 2.5 Invariants).
doc_integrity no longer reports template placeholder links as broken (ANTS-4102)
A skeleton file's deliberately-unfilled <ID>-<topic>.md links are
placeholders, not paths, so the check reported the same non-problem
on every run of any project that keeps a template.
roadmap_log reports the roadmap file it actually wrote, in its real name (ANTS-4116)
On a project whose roadmap is lowercase roadmap.md, the reply
named "ROADMAP.md" — a file not in the repo — so there was no way
to confirm where the change landed.
file_outline reports a C++ namespace as a namespace, not as a class (ANTS-4101) Asking the file map for a file's classes no longer hands back the namespace they live in.
apply_edits accepts the native Edit tool's old_string/new_string spelling (ANTS-4089)
Every session reaches this verb holding the key names the built-in
Edit tool uses, so the first call was refused. Both spellings now
work; old/new still win if a call sends both.
Roadmap entries writing their work-type or owners in bold are read correctly (ANTS-4077)
An entry writing **Kind:** fix. or **Lanes:** core. in bold — a style
the summary line has always accepted — was read as declaring nothing, so
its work-type or owners were lost. Twenty-eight entries were affected.
An entry that merely quotes one of those labels while writing about the
roadmap format is still correctly ignored.
Roadmap items marked partial import as in-progress rather than not-started (ANTS-4071)
A pass-headings project marking a block partial means the work has
begun. The importer had no entry for the word and filed it as planned,
which says the opposite about work a reader would act on.
A roadmap headline quoting code no longer loses everything after the quote (ANTS-4066)
A headline that quotes a bold marker inside backticks — a C signature like
runMain(int argc, char **argv), or a bullet discussing the roadmap
format itself — was cut short at those two asterisks, and the rest of the
title was dropped without a word. Twelve entries across all projects were
affected. The reader now recognises backticked code as literal text.
Roadmap import no longer loses a declared Kind:, invents a Source:, or defaults a field in silence (ANTS-4086, ANTS-4063)
The markdown→store import read Kind: only at the start of a line, so
the 99 bullets in this project that write the trailer inline were read
as declaring nothing and quietly filed as implement. It also wrote
Source: planned. back into the file for every bullet that had never
declared one, turning "this said nothing" into a statement the next
import believed. Both are fixed: the label is matched anywhere in the
bullet (guarded against quoted mentions, case-sensitive now, last match
wins), a source the import supplied is no longer rendered, four
further Kind: spellings map to their canonical value, a path cited by
Source:/Evidence: that no longer exists is reported rather than
ignored, and every defaulted field raises a note with a per-field tally
on the migration report. The re-migration this unblocked is verified
over every bullet — 1,434 declare a kind, 0 mismatch the store.
The roadmap database can now be brought up to date instead of refusing to open (ANTS-3781) A roadmap database written by an older build had no route forward: it fell through to the create-tables step and died on "table already exists", a message naming neither its version nor the one the running build expects. There is now an upgrade ladder — one step per version, validated end to end before a single statement runs, applied inside one transaction so a failure leaves nothing half-done, and every refusal names both versions. The database's layout number and the backup file's record-format number were also one and the same, so any change to the database would have spoiled every backup ever written; they are now separate numbers that move independently. No file on disk changes.
Tests can no longer reach the user's real data directory (ANTS-3856)
Every test bundle now runs with XDG_DATA_HOME pointed at a
per-process temporary directory, so a test that omits its roadmap-store
path opens a throwaway store instead of the live one under
~/.local/share. A fixture project had leaked into the real store that
way; it has been removed. No production code changed.
The roadmap database now records an auto-numbered item ID as store-generated rather than author-supplied. (ANTS-3838) Every item carries a note saying where each of its values came from. Both places that add items were labelling the ID as "the author chose this" even when the system had just picked the next number — so an item nobody named looked identical to one somebody did. Only an ID you supply yourself is now labelled as yours.
doc_integrity truncated a Contents list at its first WRAPPED entry. (ANTS-3836) The documentation checker reported table-of-contents entries as missing when they were in fact listed, whenever an entry ran onto a second line — it stopped reading the contents list at that point and called everything below it absent. It now reads a wrapped entry as one entry. A genuinely missing entry is still reported.
0.7.105 RC1 — Patron preview
This release shipped without written notes.
0.7.104
Roadmap edits on a migrated project now change the database, and the file is rebuilt from it (ANTS-3809) All eight roadmap-editing commands — add an item, add several, mark one done, mark several, add a note, patch a note's wording, add a section, add a table row — now write to the roadmap database and regenerate ROADMAP.md from it. Before this they still edited the file's text by hand, so on a project that had been migrated the read came from the database, the write went to the file, and the next read saw a database that never heard about the edit.
Nothing is written unless the regenerated file is valid: the
database change and the rebuild either both happen or neither
does. A preview (dry_run) still commits nothing. New IDs are
allocated from the database rather than the .roadmap-counter
file, floored to the highest ID anywhere in the committed project
so a fresh clone can never hand out an ID that is already in use.
Two things a caller will notice. Locating a bullet by line range is now refused on a migrated project, per locator, so the rest of a batch still applies — a database has no line numbers, and the old comparison would have matched every bullet in the file. And a request that sets a field the note text would contradict is refused rather than silently stored, because the file would render one value and re-read as the other.
Roadmap read seam — consumers can take their bullets from the store instead of ROADMAP.md (ANTS-3793) The roadmap database has been the primary copy for a while, but nothing read it back — every part of the app still re-read and re-parsed the ROADMAP.md text file, so the database quietly drifted out of date the moment anyone edited that file. This adds the reader, and it produces exactly the same entries the text file would have, so nothing downstream can tell which one it was handed. A project that has not been imported into the database keeps using its text file, unchanged. A database that is present but broken now says so instead of silently pretending the project was never imported. Reading a project's whole roadmap out of the database takes about 10 ms; re-reading the text file takes 33 ms. Not yet switched on for the app's own screens — that is the next step.
file_outline and read_region understand GLSL shaders (ANTS-3800)
Shader sources (.comp, .frag, .vert, .glsl and the rest of the
stage extensions) are outlined and can be sliced by function name, using
the same extension set find_definition and find_caller already
accepted. Previously a shader reported an unknown language with no
symbols, and slicing one by symbol failed outright.
Roadmap render — ROADMAP.md can now be generated from the roadmap store at full fidelity (ANTS-3758)
The inverse of the migration: every item is written in full
roadmap-format.md § 3.5 bullet form, so the generated file is the file
that exists today, written by the store instead of by hand. Section order
comes from the stored document order, and each section is routed back to
the file it was read from, so a rotated archive cannot fold back into
ROADMAP.md. Publishing is gated on every open item having a plain-English
Layman: line, per project and all-or-nothing.
find_caller takes an optional lane to scope the scan to one subdirectory. (ANTS-3805)
Without it a short generic method name such as update returns hundreds of unnarrowable callers from every class that has one. The walk itself is scoped, so callers_count and truncated describe the scoped set, and the applied lane is echoed back.
Roadmap migration now reads rotated archives, and records which file each section came from (ANTS-3766, ANTS-3782)
The migration used to see only a project's live ROADMAP.md. Closed
minor versions that /bump had rotated out into docs/roadmap/0.6.md
were invisible to it — so once the roadmap file becomes machine-generated,
the first regeneration would have quietly deleted them. Discovery now
returns the live file plus every conforming archive beside it, detects
each one's format separately, and namespaces each archive's sections by
the file they came from so nothing in the live file shifts. The store
gained one column recording the source file per section, which is what
lets the archives later be written back where they belong instead of
being folded into the main file.
Roadmap migration, load half — a project's roadmap can now be written into the store, atomically and repeatably (ANTS-3765) The reader built in ANTS-3757 could turn a ROADMAP.md into a plan and had nowhere to put it. This writes the plan into the roadmap database: one project at a time, all-or-nothing, and safe to run again after the file changes — a second run updates what changed, keeps anything you edited in the app that the file cannot express, and never duplicates or renumbers what is already there. An item you delete from the file is kept, re-filed and reported rather than thrown away, because it is far more often a rename or an archive move than a deletion. Nothing is user-visible yet; the read side that surfaces it is still to come.
Roadmap migration, read half — a hand-written roadmap becomes a loadable plan (ANTS-3757)
New RoadmapMigrate::findRoadmap() + planFrom() in ants_core_lib
(Qt6::Core only). Discovery resolves a project's roadmap
case-insensitively and refuses rather than guesses on a missing,
case-ambiguous or non-UTF-8 file; the pure planFrom() turns the
markdown into sections, items, tables, narration, the status legend
and a report, over all three roadmap formats. Nothing is written yet
— the load half is ANTS-3765. Every non-blank source line lands in
exactly one carrier, so nothing is silently dropped.
Two new debt_sweep_scan detectors — dead lint suppressions and disagreeing dependency pins (ANTS-3743)
dead_suppression flags a # noqa: CODE whose code no ruff selector
enables — a directive that reads as a reviewed decision and suppresses
nothing. It stands down entirely unless the project states an explicit
select list, so it never guesses. dep_pin_mismatch diffs every
pkg==version in build scripts, CI and container files against the pin
in requirements*.txt (or against the first hardcoded copy, for a
package with no manifest entry) and reports only where two pins
disagree — one of them is then definitely wrong.
cold_eyes_brief returns a per-doc section_index (ANTS-3740)
Each lane doc now arrives with a map of its own headings —
{heading, slug, level, start_line, end_line} — so a reviewer cites a
section anchor instead of a line number (an anchor survives an edit
above it) and fetches any section with read_region section=<slug>.
The slug is that verb's own key, so every listed section is fetchable.
Lane docs only; capped per doc, with truncated:true when a doc has
more headings than the cap.
audit_run takes exclude_paths — skip code that is not yours (ANTS-3710)
Point it at a bundled dependency folder and the audit stops reporting that
folder's problems, without narrowing the sweep away from the rest of your
project. The result tells you which exclusions were applied, and names any
tool that could not honour them, so a partly-filtered result never reads as
a fully-filtered one. Requested by the DOOM Ants session, where bundled
Windows build dependencies supplied over half of every sweep's most serious
findings.
cold_eyes_brief returns an input_hash (ANTS-3718)
A fingerprint of everything a review lane reads — the brief, its docs, the small cross-reference contracts and every cited code file. A review loop can skip a lane whose fingerprint has not changed since it last came back clean, without having to read the lane to work that out. Requested by the claude-config session.
Mageia 10 packages, and one less thing that can break a build (ANTS-3731) Ants Terminal now builds for Mageia 10 on the openSUSE Build Service. The version number stamped into the package is now worked out when the recipe is submitted rather than by a helper running inside the build machine — that helper was unavailable on Mageia and blocked it entirely. Removing it takes a dependency away from every distribution we build for, not just Mageia.
openSUSE Leap 16.0 packages, and a distro-portable RPM spec (ANTS-3727) The OBS repository now publishes for openSUSE Leap 16.0 alongside Tumbleweed, so Leap users can install and update through zypper like everyone else. The spec grew %if arms for every package name that differs between distributions — Ninja, Lua, the AppStream validator and the man-page suffix — which is what made a second distribution possible at all. Fedora 44 is wired up and resolves; it turns green from the next release, which carries the Lua header fix below.
Install Ants Terminal on openSUSE with zypper (ANTS-3726) Ants Terminal is now built and hosted on openSUSE's own build servers, so on Tumbleweed you can add one repository and install it like any other program — and get updates automatically instead of downloading a file each time:
sudo zypper ar https://download.opensuse.org/repositories/home:/milnet:/ants-terminal/openSUSE_Tumbleweed/ ants-terminal sudo zypper in ants-terminal
Getting there turned up six problems in the packaging recipe that had gone unnoticed because nobody had ever built it on a clean machine — including one that would have left out the shell-integration scripts entirely. All six are fixed.
Roadmap consumers read the store — roadmap_query, changelog_log, feedback_query/feedback_log, roadmap_branch_drift and the Roadmap dialog (ANTS-3793)
Until now nothing read the roadmap database: every tool parsed
ROADMAP.md, so the database was a write-only copy. Each of those tools
now asks the database first, and falls back to reading the file for any
project that has not been migrated — which today is all of them, so
nothing changes yet for anyone. A project served by the database also
gets its own status legend in the Roadmap dialog; every other project
keeps the built-in one, unchanged. When the database exists but cannot
be read, the tools say so instead of quietly reading the file behind
its back.
Roadmap reader records what a headless migration needs (ANTS-3764)
RoadmapParse::BulletRecord gains five fields: the verbatim
- **Status**: value, the Source: value, the leading-slot id token as
written, the pass designator, and a 1-based inclusive line span. Additive
— no existing caller reads a field that did not exist, so the roadmap
dialog, roadmap_query and roadmap_log are unchanged.
debt_sweep_scan's scope_note derives its per-category detector counts from the live roster (ANTS-3743)
debt_sweep_scan now shows how thinly each category is checked (ANTS-3707)
A zero next to a category used to read as "this area is clean". It really
meant "the checks we have did not fire" — and some categories rest on a
single check while others rest on seven. The result now lists the checks
behind every count, so a zero can be weighed properly. Reported
independently by the DOOM Ants and Fin Break sessions, both of which read a
zero as an all-clear and then found real problems by hand.
The render has no table renderer — a migrated table re-emits as raw JSON. (ANTS-3832)
A roadmap table stored in the database now comes back out as a
table. It was being written into ROADMAP.md as raw JSON, because
the file generator replayed every non-item block verbatim while
the migration stores a table as structured data. Cells keep a
literal | in the store and gain their \| spelling only in the
file, so the round-trip survives it.
Migrated roadmap bullets no longer render their headline and every trailer key twice (ANTS-3808)
The migration stored the whole bullet — headline line included — as the
item's body, while the render emitted its own headline, then that body,
then the trailer again from the database columns. The migration now stores
only the leftover text, and the render emits a Kind: / Source: /
Lanes: / Layman: / Evidence: line only when the body does not already
say the same thing. Values are compared rather than mere presence, so a
stale line in the body can never silently replace the canonical one.
Pasting a copied image file now inserts its path, not a file:// URI (ANTS-3828)
Ctrl+Shift+V branched on hasImage() only, so a screenshot on the
clipboard (a raster) worked while a copied image file — which
carries text/uri-list and no raster — fell through to the plain-text
paste and wrote file:///home/…/shot.png verbatim. Claude Code can't
resolve a file:// URI as a path, so the attachment never formed.
A hasUrls() branch now inserts the bare local path; nothing is
re-saved, since the file already exists. Paths outside the shell-safe
character set are passed through shellQuote(), which also fixes
filenames containing a space. Non-image files keep the existing
text-paste behaviour.
Roadmap migration: an archive's own preamble does survive (ANTS-3806) A reported migration limitation — that every source file's content above its first heading collapsed into one row per project, so an old archive's own header was lost — turned out not to exist. Verified by running a two-source project end to end rather than by reading the code: each file keeps its own preamble and gets it back when the roadmap is written out again. Added a regression test that will notice if that ever stops being true, and corrected the spec and comments that had recorded the wrong explanation.
feedback_log op:compact_resolved now finds shipped items in other projects' roadmaps (ANTS-3802)
Cross-session feedback files are written from one project about
another's tooling, so their item ids belong to that other project. The
compaction step looked them up only in the calling project's own roadmap
and therefore almost never found them, while the query step on the same
file found them all. Both now share one lookup, a missing local roadmap
is no longer fatal, and the refusal says which roadmaps were searched.
changelog_log op:add_batch now reports created_category, per entry and as a rollup. (ANTS-3804)
The batch path emitted the field nowhere while the single-op path always did, so callers read the absent boolean as false and could conclude no heading had been created when one had.
changelog_log op:add now refuses a feature-grouped [Unreleased] written in the format op:add_subsection itself produces. (ANTS-3803)
The detector required a flush-left bold run and missed list-item ones, so the entry was appended at the end of the section instead of being refused.
doc_citations no longer reports a fabricated out_of_range against a file the document never named. (ANTS-3801)
A citation that failed to resolve left the inherited-path antecedent pointing at the last one that did, so a following bare :NNNN was range-checked against an unrelated file. Such a continuation now reports unresolved instead.
spec_lint no longer reports a false invariant_no_test for every bullet-form invariant in a spec that also contains an | INV-N | table row. (ANTS-3799)
The parser skipped bullet invariants entirely once any table row matched, so one withdrawn-invariants summary table cost a conforming spec eleven fabricated findings. Both forms are now parsed and merged in document order.
docsindex carries a third copy of the Status-field rule, with the same truncation bug. (ANTS-3786) A third piece of code reads a document's status line and, like the other two, stops at the first line — so a status written across two lines is cut short in the docs index.
Spec header fields may wrap: spec_log no longer corrupts a multi-line **Status:**, and spec_query no longer truncates one (ANTS-3785, ANTS-3672)
A spec's **Status:** value routinely runs onto a second line — 49 of the
172 specs carrying one do, because the format standard tells authors to
append review progress inline and the corpus wraps at ~80 columns. Both
sides of the toolchain assumed a single line: spec_query returned the
value truncated mid-sentence, and spec_log op:set_status replaced only
the first line, leaving the rest behind as an orphaned paragraph while
reporting success. Neither failure was visible — the file still parsed and
the stray text read like a botched hand-edit. Both now share one rule
(SpecParse::headerField) for where a header field ends. Also fixed: the
writer could rewrite a **Status:** line quoted inside a fenced code
example and report success; the field search is now bounded to the header
block.
Re-importing a roadmap no longer duplicates entries that share a heading (ANTS-3765) When two entries in the same section had identical wording and no ID of their own, each re-import added a fresh copy of both and set the previous pair aside. Repeated often enough that grows without limit. They are now matched in order, so importing the same file twice changes nothing.
Seven roadmap entries shared one malformed ID and now have their own (ANTS-3769)
Seven separate fixes had been written down with the same broken ID (a
stray &), so none of them could be looked up and the roadmap could not
be imported. Each now has its own number; no wording changed.
A bold label at the start of a checkbox-style roadmap line is no longer mistaken for an ID (ANTS-3773)
In roadmaps written with - [x] checkboxes, a bold phrase opening a
line was always read as that item's ID. That is right for a project
writing **MT1**, and wrong for one writing **Phase 9E-2:** before
some prose — and when two lines shared such a label they became one
item, which blocked that project from importing at all. A label ending
in a colon is now never an ID.
Roadmap store: lanes, evidence and extras are now writable (ANTS-3767)
The roadmap database had a column for each of an item's lanes, evidence
files and extra fields, but nothing could actually put a value in one —
every write reported success and left the column empty. All three can now
be written, both when an item is created and when one is edited, and a
badly-shaped value is refused rather than stored.
dependencies.md cites build directives instead of line numbers (ANTS-3648)
The authoritative version-location map told you symbol names are the
stable anchor and then cited every pin by line — and every one of those
line hints had already drifted, pointing a dependency sweep at the wrong
places.
roadmap_query mode:"bundles" now flags plain-prose gates (ANTS-3389)
A to-do whose body says "Wait for X to close" in ordinary prose is now marked blocked, like one using a formal blocked by marker. Reported by the MAME Curator session, whose gated items were reading as ready to start.
doc_integrity, doc_symbols and spec_lint returned a false ETag 304 after documents changed (ANTS-3737)
These three report findings rather than content, and the ETag is a hash of the response. Editing documents without changing any finding left the response identical, so etag_match answered "unchanged" and the caller skipped the re-check — precisely when it was looking for a new problem the edits had introduced, and precisely in the common zero-findings case. Each verb now emits docs_digest, a fingerprint of the checked set, which makes the ETag track the documents themselves. (Fin Break feedback.)
roadmap_query truncated long bullet bodies from the head, dropping the current status (ANTS-3736)
A roadmap body is an append-only progress log: the head says what the item is, the tail says where it stands. Head-only truncation served the oldest text as the answer to "what is the state of this?" — one fetch reported work as pending that had shipped three days earlier. A truncated body now keeps both ends, joined by an explicit … [body elided — tail follows] … marker. Fixed at the cache truncation site too, so a body past the 16 KiB store cap — previously unreachable at any max_body_bytes setting — also surfaces its tail. (DOOM Ants feedback.)
file_outline lost every symbol after a declaration whose ; was followed by a comment (ANTS-3735)
The scanner decided whether a matched definition opens a body by testing the raw line's last character, so extern "C" int f(char* c); // note did not look like it ended in ;. It was read as a definition awaiting its {, and the next brace it met — an anonymous namespace { — became that phantom function's body, suppressing every function symbol until the namespace closed. On DOOM's r_vulkan.cpp that swallowed 5,742 lines: detected definition sites went 29 → 130 with the fix, 98 of them inside the namespace. workspace_search enclosing_symbol was collateral — with the outline blind, it attributed every match in that span to the last symbol it had seen, the struct VulkanState. (DOOM Ants feedback.)
Project search is much faster on the first search after a build (ANTS-3732)
The MCP workspace_search tool ran ripgrep with a single worker thread, which switched off its parallel directory walk and made every file read wait its turn. On a spinning disk that is the slowest possible order. A wide search over 31,500 files took 24.6 seconds; it now takes a fraction of a second, and ordinary searches no longer time out when a build has just filled the disk cache.
Lua headers are found on distributions that do not use openSUSE's layout (ANTS-3727) luaengine.cpp included <lua5.4/lua.hpp>, which names a directory only openSUSE creates; Fedora installs the same headers directly in /usr/include. The build failed there even though CMake had located Lua and enabled the plugin system. The include is now unqualified and the directory comes from CMake, which was already supplying it. No effect on openSUSE, where the compile line has carried -I/usr/include/lua5.4 all along.
openSUSE package ships a stripped binary, with debug symbols in their own subpackages (ANTS-3729)
The OBS project now enables debuginfo, so /usr/bin/ants-terminal is
stripped and ants-terminal-debuginfo / ants-terminal-debugsource are
published alongside it — a crash report from a user can now be resolved
to source, where previously the symbols were discarded during the build.
The rpmlint report for the package is clean (0 errors, 0 warnings); the
one remaining error was a false positive in rpmlint's own branding check,
which misreads hicolor-icon-theme as a themed branding package, and is
filtered with the analysis recorded next to the filter.
0.7.104 RC1 — Patron preview
This release shipped without written notes.
0.7.103 — hotfix
0.7.102
apply_edits can now edit by line range, not just by matching text (ANTS-3711) Deleting a big block used to mean pasting the whole block back in just to say which one you meant. Now an edit can name a line range instead — "lines 200 to 276" — which is far cheaper for a large chunk. You also hand it the exact text of the first and last line as a safety check: if the file has changed underneath you and those lines no longer match, the edit is refused instead of quietly rewriting whatever moved into those positions.
doc_integrity catches numbered sections that run out of order (ANTS-3700)
Nothing used to notice a document whose sections go 5.6, 5.8, 5.7 — the eye reads them in the intended order and every link still works. The checker now flags out-of-order, skipped and repeated section numbers. Tuned to stay quiet: 2 findings across this project's 275 documents, and both are real.
indie_review_partition works out review lanes from the file tree when a project has no module map (ANTS-3709)
Previously a project that describes its layout in prose got back an
empty list, even though the server already knew every source file, and
the caller had to build the review plan by hand. It now groups the files
by folder (splitting big folders into numbered chunks) and hands that
back marked as worked-out rather than declared, so you can adjust it and
save it as your project's own partition file.
audit_run says WHY a file could not be parsed (ANTS-3706)
New parse_failures_detail[] gives {file, tool, reason} — the
check-id and first diagnostic — next to the existing bare-path
parse_failures[]. A missing include path (a config problem you can
fix) now reads differently from a construct the analyser cannot handle
(something to route around), without re-running the tool by hand.
project_settings op:detect shows what the project already declares (ANTS-3705)
detect now returns declared (the six layout keys as stored) and
declared_missing[] naming any declared path that no longer exists, so
checking the layout no longer means opening .ants/project.json by
hand. A path that has been moved or deleted is silently ignored by the
reader — previously indistinguishable from never having been declared.
indie_review_corroborate accepts lane reports outside the project (ANTS-3713)
Pass allow_outside_project:true and reports_dir may be an absolute
path — the session scratchpad, /tmp — so a review sweep no longer has
to write scratch files into the working tree just to use the cheap
server-side read. Same opt-in name as test_audit_synthesis_prompt's.
Path checks are unchanged apart from the root anchor; the default
project-relative path still refuses an absolute directory.
workspace_search gained exclude_glob — search everywhere except a given tree (ANTS-3704)
Until now you could narrow a search to one folder or one file pattern, but there was no way to say "everywhere except the docs". On a project with a lot of written documentation that meant the results were mostly prose, and the only way round it was to drop back to the raw command-line search tool. Pass exclude_glob (one pattern or a list) and those paths are left out. Write the pattern plainly — docs/**, not !docs/**.
doc_dedup MCP verb — find the same passage written twice across a doc set (ANTS-3660)
Splits each markdown file into paragraphs, compares them as word
3-grams, and reports every pair that overlaps enough to be the same
fact written twice. Groups the pairs into clusters, because five
documents sharing one stanza are one thing to fix rather than ten
findings. Skips fenced code (two identical samples are not duplicated
prose), pointer-only lines, and generated files whose whole purpose is
to repeat a template. Report-only — it never decides which copy to
keep, because that is a judgement about where a fact belongs.
spec_lint MCP verb — the greppable half of the spec-format contract (ANTS-3662)
Checks the structural defects /cold-eyes § 1e otherwise greps for by hand on every review pass: an invariant with no test-surface clause, a gap in a spec's own id sequence, a cold-eyes loop-log row with no recorded outcome, and a test clause that is a command stating nothing it should return (a candidate — never auto-fixable, and no subprocess is ever run). Tombstoned invariants are exempt. missing_section runs only when the project's format standard carries a machine-readable block, so it ships skipped with sections_checked:false rather than firing against a guessed list. Size is reported in line_count, never as a finding.
doc_symbols MCP verb — resolve the identifiers a doc claims exist (ANTS-3661)
Checks that every function or class name a document mentions actually
exists in the code. Harvests inline code spans shaped like Foo::bar()
and looks each up with find_definition's resolver, skipping fenced
samples, paths, language keywords, MCP verb names and doc-examples
regions. It only ever reports: deciding whether a missing name is rot
or something the document is about to create stays with the reader, so
there is no severity and nothing auto-fixable. A name the run never got
to is reported as not_checked — never as one that does not exist.
doc-examples regions: mark a passage's file references as illustrations so doc_citations stops reading them as rot (ANTS-3659)
Documents that explain citation syntax have to write references that
point nowhere. The checker read all 58 of them in one spec as broken,
against a single real citation. An author can now wrap such a passage in
<!-- doc-examples: begin --> / <!-- doc-examples: end --> and the
scan skips it entirely. Two new response fields, unterminated_examples
and examples_suppressed, keep a runaway or over-broad region visible
rather than silently blanking real citations.
doc_citations now checks whether a cited line still mentions the symbol named beside it (ANTS-3654)
When a document writes a symbol in backticks just before the place it
points at, the verb now looks for that symbol in the lines it cites and
reports anchor_found. only:"stale" returns those drifted citations
alongside the ones that failed to resolve, so a docs sweep surfaces
"this line moved" as well as "this file is gone".
New doc_citations tool — check that a document's "see file, line 12" pointers still point where they claim (ANTS-3636)
Design docs are full of references like src/window.cpp:120, and they
go stale silently as the code moves. Reviewing one used to mean opening
every file by hand. This reads the document, finds every reference, and
shows you the line each one points at today — plus which ones now point
at a missing file, past the end of a file, or at a name that matches
several files.
Citation scan — the path:line grammar and scrape rule as a standalone contract, split out of ANTS-3636. (ANTS-3653)
The part that just finds file.cpp:42 references in a document, specified on its own so it can be built and tested without touching the filesystem.
changelog_query accepts section as an alias for version. (ANTS-3618)
Callers arriving from roadmap_query reach for its vocabulary and passed section:"Unreleased", which was silently ignored — so every entry came back, reading as a filter that matched everything rather than one never applied. Passing both with different values now refuses bad_args.
Review Changes dialog shows each touched file's total line count. (ANTS-3632)
The Status list now renders a dimmed (N lines) after each entry, so file size is visible before clicking in. Renames report the new path; deleted, binary, unreadable and over-64 MiB files show no suffix rather than a misleading number. Counted in 64 KiB chunks, so memory stays flat on multi-MiB files.
audit_dismiss — let Claude tell the audit tool "that finding isn't a real bug". (ANTS-1713)
Until now only the Audit pop-up could mark a warning as a false positive so later audits stop reporting it. Claude can now record the same verdict directly. It is remembered by content rather than by line number, so it survives edits that move the code around.
New doc_integrity MCP verb + engine — deterministic dead-anchor / broken-link / TOC-coverage checks for markdown docs (ANTS-3601)
Finds the internal-consistency rot in long contract docs that no test caught and only a careful human reader spotted: a [text](#heading) link whose heading was renamed, a [text](../file) whose target moved, or a hand-maintained Table of Contents that drifted from the sections. GitHub-compatible heading slugs, fence-aware (fenced examples ignored). Runs as the doc_integrity verb (scope a file/dir, filter by kind) and feeds the cold-eyes review pre-pass automatically.
New contract_doc_drift audit lane — flags back-ticked literals in docs/standards/docs/specs that no longer exist in the code (ANTS-3600)
Catches the doc-vs-code drift where a contract doc quotes a file path, symbol, or config key that a rename left behind. Fence-aware (fenced examples are skipped) and path-aware (slash-paths resolve). Human-triage severity, never auto-fixed; a project-local .ants_doc_drift_allow.txt suppresses intentionally-illustrative literals. Runs in the GUI Audit dialog.
audit_run: richer partiality + zero-coverage surface, raised time ceilings (ANTS-3585) (ANTS-3585)
A big C/C++ sweep no longer misleads a caller: incomplete_tools_detail[] says WHY each tool is incomplete ({tool, status, elapsed_ms, truncated} — truncated ⇒ cut off at the cap, false ⇒ crashed) and parse_failures[] names source files cppcheck could not parse (a C++23 TU its frontend chokes on gets zero coverage instead of being silently absent). The per-tool wall-clock cap ceiling is raised 60→300 s and the aggregate 240→900 s (default stays 30 s; high caps are meant for the async path, which survives the transport timeout). Both surfaces ship on the sync and async-poll envelopes.
changelog_log: add an opt-in feature-grouped-subsection insert mode for ### <DATE> <Category> — <headline> CHANGELOGs. (ANTS-3584)
Some projects group their changelog by dated feature blocks rather than flat Added/Fixed headings; give changelog_log a native way to insert those instead of forcing a hand-edit.
read_regions: accept an optional top-level path used as the default for any item that omits its own path (per-item path still wins). (ANTS-3589)
When you want several slices of ONE file, read_regions makes you repeat the same filename on every slice; let it take one filename at the top like its sibling read_region does.
invariant_check answers in a few hundred bytes by default (ANTS-3699)
The "does any spec already cover these files?" check used to return every matched spec's full invariant text. One real call over three files came back at 267,070 characters — too big to read, so people learned to skip the step. It now returns just the list of matching specs and how many rules each has; ask for mode:"full" when you actually want the text.
Review tools no longer tell you not to use them (ANTS-1581) Every cold-eyes and indie-review tool used to carry a line saying the matching review command doesn't call it, so reach for it only when building your own pipeline. That advice is now backwards — those commands prefer these tools. The line is gone. One warning stays, on the tool that runs the review on your own local AI model instead of Claude: it now says plainly that this is a weaker reviewer, not a cheaper way to get the same review, and the tool picker no longer advertises it as the way to start a review.
apply_edits documents the per-call batch ceiling separately from the 4 MiB file cap (ANTS-3712)
A several-KB batch can fail in the client's transport before the server
sees it, with an error blaming escaping; the 4 MiB figure applies to the
file being edited, not the request. The description now says to split the
batch rather than audit quoting.
roadmap_log reports bytes_written as the byte delta on every op, plus file_bytes (ANTS-3702)
Whole-file-rewrite ops reported the size of the file, so six short notes
came back as a 459 KB write. The full size is still available.
audit_falsepos_log now accepts debt-sweep as a review kind (ANTS-3701)
Dismissing a false alarm found during a debt sweep had to be filed under "audit" with a hand-written note, so the record of where it came from was wrong. It can now be filed under its own name.
doc_symbols no longer reports bare lowercase words it cannot resolve (ANTS-3692)
A backticked word like counts, dry_run or truncated is almost always
a JSON field name rather than a symbol anyone declared, and the checker
cannot tell those from a genuine stale reference. It now stays quiet about
them unless the lookup actually finds the thing, so the list you get back
is the one worth reading. Names that do resolve are still reported, which
keeps shell and Python functions — they are lowercase by convention.
Measured over the project's own docs: 45% fewer entries, and slightly more
real symbols found, because the lookup budget is now spent on the promising
names first.
Source-scrape tests can strip comments before scanning (ANTS-3662)
A test asserting that code does not do something fired on a comment that merely named the thing — the same false-positive class as the fixed-byte scrape windows ANTS-3681 replaced. tests/_support/srcgrep.h gains stripComments, which removes comments while leaving string and character literals intact.
Hoisted the inline-code-span scanner into MarkdownScan::codeSpans, and taught fenceMask to report an unterminated fence's opener line (ANTS-3649)
A shared text-scanning helper moved into the shared module so the tools
that need it all use one copy. Nothing changes for you today; it is the
groundwork for the citation checker.
Dropped the source-scrape size ceiling from the roadmap-format test (ANTS-3633) The ceiling had been raised eleven times and never once caught the runaway it was meant to catch — it only taxed each edit to the roadmap-query code with a second round trip. The check that actually guards against reading the wrong function is kept, and its failure message now names the real cause.
find_sources is now labelled C/C++-only before you call it, not after. (ANTS-3619)
On a Python or JS project it returns files_count:0, which is indistinguishable from a genuine "nothing calls this" — the worst possible wrong answer when checking a change's blast radius. The limit was only explained in the response. Both pre-call surfaces (the session hook menu and the tool catalog hint) now state it, and the adjacent find_definition line names its languages so the asymmetry is visible rather than inferred.
roadmap_query's bad_mode refusal now lists the accepted modes. (ANTS-3617)
A caller who guessed a mode name wrong was told only which value was rejected, not which were valid, so finding the right spelling meant going to read the schema. Matches the accepted array its changelog_query sibling and its own bad_status already emit.
The /test-audit command now uses the fast server-side path, like /audit already does. (ANTS-3625)
There was a quick way for the test-audit command to run its detection and chunking work inside Ants rather than the slow way, but the shortcut had never been switched on in the command's instructions. It is now — so a test-suite sweep no longer has to load the command's full manual to get started.
Refreshed five stale code comments in the audit runner (missing tool, dead status value, incomplete scope list, missing caveat). (ANTS-3610)
Headless audit_run MCP verb does not dispatch the in-process (inProcessRunner) audit lanes. (ANTS-3605)
The headless audit_run MCP verb now runs the in-process doc-vs-code drift lanes (spec↔code, contract-doc, changelog↔test) that previously fired only in the GUI Audit dialog, so automated Claude/CI audits get the same coverage. A tool-less host no longer refuses an auto-detect sweep — the lanes need no external binary.
Hoisted the duplicated CommonMark fence-scanner into a shared MarkdownScan module (ANTS-3603)
feedbackfile.cpp and speclog.cpp had verbatim copies of the fence-opener regex + opener-char helper; both now call the single src/markdownscan.{h,cpp} copy (Qt6::Core-only), which also exposes a fenceMask primitive for the doc-integrity arc. No behaviour change.
project_settings op:detect: also propose test_roots/changelog/docs_dir when the conventional dirs/files exist on disk, not just source_roots. (ANTS-3588)
The auto-detect for project layout only fills in the source folder even when it can plainly see a tests/ dir and a CHANGELOG.md — make it offer those too so setup is one step, not three.
cold_eyes_brief: caution reviewers that ROADMAP/CHANGELOG are append-only HISTORY — a ✅/Resolved bullet is PAST state, a 📋 bullet is already-tracked work. (ANTS-3586) Cold-eyes reviewers keep re-reporting already-fixed bugs because a 'Resolved' roadmap line reads to them like an open problem; add one sentence to the brief telling them to verify against current source first.
test_audit_partition's inert quick field, which was accepted and forwarded but never read. (ANTS-3630)
quick:true ran an identical full walk and pre-pass. --quick is caller-side (the grep pre-pass is the audit; only the subagent phase is skipped), so the field promised a fast path that did not exist. Callers passing it now get a schema refusal rather than a silent no-op.A "no window" safety check that handed control to code needing one (ANTS-3725) An internal guard for the no-window case passed the problem straight to a function that assumed a window existed, and threw away the caller's own answer on the way. Harmless in the running app, but it crashed any test driving these tools directly — which is why several were testing their own source text instead of their behaviour.
spec_log reports how much it wrote, not how big the file is (ANTS-3724)
Editing a spec reported the size of the whole file as "bytes written". It now reports what the edit actually added or removed, with the full size alongside — the same two numbers the roadmap and changelog tools already give.
changelog_log now reports how much it added, not the size of the whole file (ANTS-3723)
Adding a two-line entry used to come back as "1,150,003 bytes written",
which is the entire CHANGELOG. It now reports the bytes it actually
added, alongside the full file size as a separate number — the same way
the roadmap tool already did, so the two can be read side by side.
Filing feedback whose example contains a shell comment no longer swallows the finding (ANTS-3695)
A # at the start of a line in a pasted command example was being read
as a document heading, which cut the finding short and hid it from the
maintainer's review list — silently. Examples are now indented so that
can't happen, and files already written the old way are read correctly
again.
test_audit_partition no longer refuses a project that told it where the tests are (ANTS-3708)
If a project declares its test folders in .ants/project.json, or you
name the test files outright, the tool now just gets on with it instead
of insisting it can't find a recognised test framework. Hand-rolled test
harnesses — the kind with no framework to detect — work again. When it
genuinely can't tell, the refusal now lists everything it looked for and
the two ways to answer it.
A roadmap bullet quoting Lanes: in its prose no longer acquires a bogus lane (ANTS-3722)
Writing about the trailer keys made the reader treat them as that item's
real fields — so the docs describing the roadmap format were the bullets
most likely to be mis-parsed. Inline Kind: x. Lanes: y. trailers still
parse as before.
Feedback-file standard covers a project whose directory begins with a dot (ANTS-3714) The derived filename would start with a dot and never match the glob the standard calls authoritative, so that project's input would be invisible.
audit_run SARIF now carries level on every result (ANTS-3694)
Asking severity_floor:"error" always returned nothing, because the
headless writer emitted no level and the line-based parser recorded every
cppcheck finding as UNKNOWN — the real severity survived only as the
message's text prefix. Both are fixed, so real errors are findable again.
Debt sweep counts a test that names its invariant in the function name (ANTS-3693)
test_INV8a_... cannot carry the hyphen, so the commonest citation shape
read as uncovered. The same fix closes a false negative in the other
direction: a test citing INV-80 was accepted as covering INV-8.
spec_query no longer runs a bullet-form invariant through a following heading (ANTS-3697)
The last invariant before a subheading absorbed that heading and its prose,
and test_surface inherited the over-run.
roadmap_log strips a stray XML tag at the start or end of a note (ANTS-3703)
A malformed call could write a literal </note> into ROADMAP.md — a
permanent line in an append-only file. Markup mid-sentence is prose and is
left alone.
roadmap_query honours filter as an alias for status (ANTS-3698)
The response echoes the applied lifecycle as filter, so callers sent that
name back; it was silently ignored and answered with the full set, handing
planning sessions already-shipped items.
roadmap_log resolution notes land at the end of a multi-paragraph bullet (ANTS-3696)
A note used to be inserted after the body's first paragraph, leaving the
rest of the problem statement and the Kind/Source trailer below it — so a
well-documented item read as resolved and then went on arguing for itself.
spec_query could swallow an invariant from a malformed table (ANTS-3683)
parseSpecBody's table-row pattern separated cells with a whitespace class that matches newlines, so on an invariants table missing its test column one match ran across two rows — returning the first invariant with the second row's text as its test surface, and dropping the second invariant entirely. Well-formed three-column tables were unaffected. Found by spec_lint, in the parser it consumes.
spec_query now returns test_surface for bullet-form invariants (ANTS-3665)
The spec standard promises that an invariant's *Test:* clause comes back as
its own test_surface field, but only the table form delivered it — so nearly
every spec in the project returned the clause buried in the invariant body.
Asking a spec "which of your invariants have no test?" now has an answer.
Test suite no longer goes red at random on a busy machine (ANTS-3658) One audit test reads the whole source tree, so it takes about 2 seconds when the files are already in memory and nearly 10 when they are not — against a 10-second limit. It now gets 60 seconds, so a failure there means a real hang rather than a busy computer.
MarkdownScan::fenceOpenerChar misreads a 4-backtick inline span as a fence opener. (ANTS-3655)
A line that shows backticks as example text can be mistaken for the start of a code block, which makes the scanner skip the rest of the document.
The doc link-checker now recognises code blocks indented inside a bullet list (ANTS-3638) A code block nested under a bullet was invisible to the scanner, so its sample text was read as prose and any link-shaped line in it was reported as broken. Code blocks are now recognised relative to the bullet they sit under, exactly as Markdown defines them, while an indented block outside a list is still left alone.
roadmap_log no longer writes a body or note that opens a code fence it never closes (ANTS-3640) A roadmap note quoting a bare fence opener used to be written verbatim, which silently turned every bullet below it into one big code block — the next edit down there was refused with a message pointing at the innocent bullet. Unclosed openers are now escaped on the way in (balanced ones, which leave the file well-formed, are left alone), and the refusal names the line that opened the fence.
Review-tool descriptions no longer deny a call their own skill instructs (ANTS-3639) The "your slash-command skill does not call this tool" note was pasted onto every tool whose name started with cold_eyes_, indie_review_ or test_audit_. But /test-audit drives the whole test_audit_ family, and /indie-review instructs indie_review_partition and indie_review_corroborate by name — so sessions were told not to call verbs their skill mandates two lines later. Those seven are now exempt.
doc_integrity: a code span that crosses a newline is now masked end to end (ANTS-3635)
The link checker masked inline code one line at a time, so the tail of a two-line … span was still harvested as a link — a C++ lambda split across lines read as a broken link. Masking now runs over the whole document. An unpaired backtick still masks nothing, so one stray tick cannot silently stop the checker on the lines below it.
Link checker no longer reports every section of a spec as "missing from TOC" when that spec's contents list is plain text. (ANTS-3634)
(ANTS-3634) doc_integrity's TOC-coverage check matches sections by
anchor slug, but it only required that a contents region exist — not
that the region actually contained any #anchor links. A spec whose
contents list is written as plain text (- §1 Problem) therefore
offered nothing to match against, and every one of its sections was
reported missing. Three specs (ANTS-1870, ANTS-2023, ANTS-2141)
accounted for 23 of the 26 findings over this project's own docs — all
false alarms; the documents were fine. The check now stands down when a
contents region yields no anchors. A contents list that does use
anchors keeps full coverage, including the existing rule that a single
linkless bullet is skipped without ending the list.
doc_integrity no longer reports example links written inside backticks as broken. (ANTS-3623)
Fenced code blocks were skipped but inline `[text](target)` spans were not, so 21 of the 22 broken-link findings over this project's docs were prose examples — 12 of them in doc_integrity's own spec, which has to contain sample links to explain what the checker does. Real links whose text is code, [a/b.md](a/b.md), are still checked.
A SARIF report could silently drop findings while reporting findings_truncated: false. (ANTS-3629)
The document is capped at 10 000 results across all tools, but only the per-tool caps set the flag. Several tools each individually under the cap could sum past it, so results vanished from the report a reviewer treats as the complete finding set, with nothing to say so. The document-wide shed is now reported.
A test_audit_* contract test that asserted the opposite of the shipped behaviour and passed anyway. (ANTS-3628)
It checked that the verbs were NOT caller_cwd-Required — the inverse of the truth — and passed only because the source window it searched had drifted off the lines it meant to inspect, so its assertions were vacuous and a real regression would have gone unnoticed. Rewritten to call the contract function directly, covering all five verbs; verified to fail when a verb is flipped.
Two stale audit_run schema descriptions that under-reported what a sweep does. (ANTS-3622)
tools claimed auto-detect skipped clang-tidy (it skips only mypy — 9 tools run by default, clang-tidy among them), and paths claimed only cppcheck and clang-tidy honoured it (eight tools do; only the repo-global gitleaks and trivy ignore it). The second mattered most: believing narrowing didn't reach ruff/bandit/semgrep, a caller would skip paths and pay for a full sweep.
feedback_log op:compact_resolved accepts v2-or-later feedback files, not exactly v2. (ANTS-3621)
The gate was != 2 while the parser gates on >= 2, so a future v3+ file was read with v2 semantics but refused as not_v2 and pointed at migrate_v2 — a no-op on an already-migrated file, leaving no way forward.
audit_run refuses an unrecognised formats value instead of silently producing no report. (ANTS-3626)
An unrecognised entry such as ["json"] was non-empty enough to suppress the ["sarif"] default yet matched neither output branch, so a sweep that could run for 15 minutes finished with ok:true, no artifact, and no signal. Now refused as bad_args naming the accepted set, before any tool starts. An empty array still means ["sarif"].
The audit's "hide these known-OK warnings" list now works outside the app window too. (ANTS-3615)
The .audit_allowlist.json filter only ever ran in the Audit pop-up, so an entry that hid a finding there did nothing when Claude or CI ran the same audit. It is now shared by both. The suppressions setting, which had quietly done nothing, is honoured too — and an unrecognised value now says so instead of being ignored.
audit_run cleans up the temporary report files it leaves when a project folder is read-only. (ANTS-3614)
When the audit can't write into your project it drops its report in a temp area, and those were never cleaned up. They are now swept after a week, on the next audit that needs the fallback.
audit_run no longer treats a long-but-healthy sweep as a dead worker. (ANTS-3611)
The safety timer that stops two audits running on the same project at once still used an old 4.5-minute limit, but audits can legitimately take up to 15 minutes. A long sweep could have its slot freed underneath it, letting a duplicate start. The timer is now derived from the real limit instead of hardcoded.
Aggregate concurrency cap on audit_run — a third simultaneous sweep now refuses server_busy instead of allocating. (ANTS-3612)
The audit engine was meant to run through a small pool of worker lanes to cap memory, but that pooling was never built — each audit just started its own worker. Up to 16 could run at once (~30 GiB of tool memory), enough to run a 32 GiB machine out of RAM. runAudit now takes one of two process-wide slots before doing any work, so audits across different projects are bounded too.
docs_index no longer indexes headings or links inside fenced code blocks (ANTS-3604)
A # Heading or [link](x) written inside a ``` example block was being treated as a real document heading/link. The scanner is now fence-aware, so illustrative examples in code fences no longer pollute the documentation map.
indie_review_orchestrate: don't fire summary-similarity suggested_merges for lanes built by the file-list directory-grouping fallback (boilerplate summaries → false merges). (ANTS-3591)
When indie-review groups a file-list project by folder, every lane's description is near-identical boilerplate, so the 'you could merge these lanes' hint wrongly suggests merging tests with src.
audit_run/last_audit_summary: filter zero-content SARIF results (empty ruleId AND message AND artifact uri AND startLine 0) out of total_raw/total_actionable and top_findings. (ANTS-3590) A completely clean repo reported '1 actionable finding' — a blank MAJOR row with no file, rule, or message — because trivy emitted one empty placeholder result. A close-phase/audit gate could misread that as a real blocker.
session_orient: keep top-level ok:true when the only problem is absent-but-optional artifacts (no ROADMAP/specs/empty index) — surface them via warnings[]/notices[]. (ANTS-3587)
A brand-new project with no roadmap yet gets a scary 'ok:false' from the first orientation call, as if something broke — but 'no roadmap yet' is normal.
The cross-session feedback-file exception is now bounded by directory, not just filename. (ANTS-3616)
One shared coordination file is deliberately allowed to sit outside the project root. That exception keyed on the filename suffix alone with no location check, so any path ending _Ants_MCP_Feedback.md anywhere on the filesystem qualified — and since apply_edits uses the same check, that made the naming convention a broad write permission. It now also requires the file to sit in a directory above the project. The apply_edits description, which stated the root-escape refusal with no mention of the exception, has been corrected — that omission is what led a contributor to report it as a suspected security hole.
audit_run's compile-DB escape validator no longer skips every non-default build tree. (ANTS-3624)
It probed only build/ while clazy and clang-tidy resolve their compile database from seven build-dir names. On a tree whose DB lives in build-fast/ — this project's documented iteration workflow — the validator found no file, concluded there was nothing to validate, and the tools then consumed that DB with its include paths never checked for escapes. Both now go through the same resolver.
test_audit_partition now root-anchors a caller-supplied scope, closing a path-traversal read. (ANTS-3627)
A scope:"path:../.." (or a files: entry with ../) was concatenated onto the project root with no anchor check, so the walk enumerated files outside the project and returned their paths in the envelope. Both branches are now checked against the canonical root and refuse bad_path; a files: list refuses as a whole rather than silently dropping the bad entry.
0.7.102 RC1 — Patron preview
This release shipped without written notes.
0.7.101
recent_errors / build_status — attach likely_fix add_include for "'X' has not been declared" diagnostics. (ANTS-3374) recent_errors and build_status now attach a likely_fix:{add_include, defines, at} hint on undeclared-symbol diagnostics ('X' has not been declared / was not declared in this scope / unknown type name 'X' / use of undeclared identifier 'X'), resolving the declaring header so the missing #include no longer needs a second lookup.
indie_review partition deriver: optional fallback to group a file-list ## Module map into lanes by top-level directory. (ANTS-3507)
Right now the auto-reviewer only works when a project's module map lists named subsystems; a project that lists plain file paths (common) gets refused and must hand-write a partition file. Let it fall back to grouping those files into lanes by their top folder.
indie_review_partition now emits a sparse_partition_hint on an empty/≤1-lane partition (ANTS-3567)
Symmetry with cold_eyes_partition (ANTS-1634a): when the module-map
deriver yields ≤1 lane (no CLAUDE.md ## Module map, no
docs/subsystems.md), the response now carries sparse_partition:true plus a
hint pointing at indie_review_brief's source_paths[] ad-hoc mode
(ANTS-3375) and the .indie-review/partition.json override — instead of a
bare empty partition the caller can't act on.
changelog_query read verb — mirror roadmap_query so drift-checks stop reading the whole CHANGELOG. (ANTS-3533)
Add a quick lookup for the changelog like the one the roadmap already has, so reviews don't read the whole 276K-token file.
Surface MCP tokens-saved as a live status-bar chip + persisted month / YTD / all-time aggregate. (ANTS-3572) A little pill on the status bar shows how many tokens the Ants MCP has saved you — live this session, plus this-month / this-year / all-time totals that survive restarts.
spec_query silently under-reports invariants declared inline (` (ANTS-3569) When a spec writes a rule as a sentence instead of a list or table, the quick spec-reader reports zero rules, so a caller trusting the count thinks the spec has no invariants.
indie_review_brief gains a source_paths[] ad-hoc-lane mode — a lightweight single-reviewer broker for a cold review of a code / dependency diff
When lane is not in the auto-partition, pass source_paths[]
(the changed files of a dependency bump / code diff) and the brief
is synthesised over just those files — the code-review analogue of
cold_eyes_brief's doc_paths[]. Paths are traversal-guarded;
rejected entries surface in source_paths_rejected, and an unknown
lane with no usable override now lists known_lanes for recovery.
(ANTS-3375, ANTS-3493)
feedback_query — resolve foreign-prefix mapped ids (ANTS-*) against the canonical sibling roadmap under the shared root. (ANTS-3519) feedback_query now resolves foreign-prefix mapped ids (e.g. ANTS-* cited in a consumer project's feedback file) against the sibling roadmap that owns the prefix — the mapped_id_status entry carries the real status, a resolved_from field naming the owning project, and shipped_date when shipped, instead of a bare "foreign_repo". Cost-gated on a foreign id being present and RAM-bounded (one roadmap parse per distinct foreign prefix, discarded after).
find_definition is blind to GLSL/shader symbols — add a glsl language family. (ANTS-3558)
In projects built around shaders (like the DOOM renderer), the "where is this defined?" tool can't see any shader code, so it wrongly reports symbols as missing.
roadmap_query mode:bundles clusters on headline tokens only — same kind+lane+stem items return as all-singletons, indistinguishable from no-clusters. (ANTS-3388) roadmap_query mode:bundles now groups per-module " /spec.md"-style to-dos that share the same Kind, a lane, and a template shape (previously returned as indistinguishable singletons), and always reports no_clusters_found so a caller can tell "grouped into singletons" from a real grouping.
The status-bar "tokens saved" badge is now scoped to the current project instead of a global all-projects total. (ANTS-3579) The pill shows the savings for the project of the tab you're looking at — its live-session number on the face, this-month/year/all-time in the tooltip — and switching tabs re-scopes it. A separate "All projects" tooltip line preserves the previous global total. Savings are attributed per Claude Code call's project folder; the per-project store is bounded (64 projects, LRU). Switching to a tab for a project you have history for (but haven't used this session) shows that project's all-time.
changelog_query entries[] auto-downshifts to the lean headline_only shape when a page would truncate, keeping the whole tail (mirrors roadmap_query). (ANTS-3576) When a full-detail entries page overflows the soft cap, the server projects every entry to {version, category, ids, text_oneline} and re-pages so nothing is silently dropped; the response then carries downshifted:true. Opt out with an explicit limit or include_body.
Auto-downshift a would-be-truncated roadmap_query / workspace_search list to its titles-only form instead of cutting off the tail. (ANTS-3543)
When a result list is too big to send whole and you didn't ask for a
specific page size, the server now keeps every item but drops the long
descriptions (showing just the titles / file:line) rather than silently
cutting off the end — so a scan for "is X tracked anywhere?" no longer
misses a hidden remainder. It emits downshifted:true, and only pages
if even the titles-only list overflows.
debt_sweep_scan self-describes its scope (detectors_run + scope_note) so total_findings:0 is not misread as "no debt". (ANTS-3564) The debt-sweep tool only looks for a narrow set of markers; when it finds nothing it should say "I only checked these things", not silently imply the project is clean.
Default max_match_bytes ceiling on workspace_search — clip pathological long lines by default (token-saver default-ON). (ANTS-3548)
One giant machine-generated line shouldn't be able to flood a search result; trim over-long lines automatically unless asked not to.
roadmap_log flip_batch: line_range locator has no ants-v1 fallback in a mixed-format roadmap (ANTS-3565 residual). (ANTS-3570) In a to-do file that mixes two bullet styles, flipping a newer emoji-style item by its line number still fails — though you can now flip it by its ID or title instead, which is the usual way.
feedback_log op:assign_id now honours its documented note param (ANTS-3571)
assign_id{ids|closure, note} silently dropped the note — it was never
read from the request. It now renders as a single - **Note:** <note>
bullet under the finding's Proposed ID line (newlines folded to spaces),
replaced in place on re-assign so the write stays idempotent; the
envelope echoes note_written:true.
roadmap_query: prose/ID-less roadmaps now warn on every status filter, not just status:all (ANTS-3583)
A status-filtered query (planned/in-progress/considered) against a
roadmap with zero [PROJ-NNNN]-tagged bullets returned a bare count:0
with no warning — dangerously readable as "no work left". The empty
path now scans the whole file for any id-bearing bullet and, when none
exist, emits a machine-detectable parseable_bullets:0 plus a "format
not recognised" warning identically on every filter.
roadmap_query mode:"headline_only" now measures the lean row for the auto-truncate cap, so more title-only bullets fit per page. (ANTS-3577) Restores ANTS-1881 INV-6: the projection to the 4-key headline shape now runs before pagination (it had drifted to projecting the page after measuring the full-size row, dropping more rows than needed).
read_regions requests/paths/regions aliases are stripped by the inputSchema (additionalProperties:false) before reaching the ANTS-3500 fallback. (ANTS-3568)
The batch file-reader has handy synonym keys that were coded in but the tool's input rules silently throw them away, so they never work — you always hit an error unless you spell it exactly items.
roadmap_log flip/annotate ID-locator suggestions now rank by sibling-id prefix, not headline overlap (ANTS-3566)
A bullet_not_found on an id like 3D_E-0031 no longer suggests every bullet whose title merely starts with "3"; it offers same-prefix sibling ids (or none when nothing is related).
roadmap_log op:amend_body and op:flip_batch locate ants-v1 emoji bullets in a mixed-format roadmap (ANTS-3565)
Extends the ANTS-3561 fix to two more verbs: in a GFM-majority ROADMAP.md with appended - 📋 [ID] emoji bullets, amend_body and flip_batch can now patch/flip those bullets by id or headline (flip_batch resolves each locator against either format and can flip GFM and emoji bullets in one commit).
roadmap_log flip/annotate: digit-led-id bullets have no working locator — headline fallback also fails. (ANTS-3561)
roadmap_log op:flip / op:annotate can now locate an ants-v1 emoji bullet (- 📋 [ID] **…**) inside a GFM-majority (mixed-format) roadmap. Previously the ants-v1 walker only ran when the file had ZERO GFM bullets, so a mixed file's appended emoji bullets were unaddressable by id or headline — bullet_not_found — even though roadmap_query read them fine.
Bash guard-hook no longer vetoes a grep -v ROADMAP.md / --exclude=ROADMAP.md exclusion as if it were a read of the roadmap (ANTS-2169).
read_region/read_regions symbol mode should resolve namespace/class-qualified names, not only the bare identifier. (ANTS-3513) When asking to read a function by name, typing its full name (like ClassName::method) fails and you must strip the prefix — the tool should accept both.
Malformed ids filter fails silently to the full list instead of refusing — waste + wrong result. (ANTS-3541)
If a tool call asks for specific roadmap items but formats the request slightly wrong, it quietly dumps the whole list instead of saying 'that was malformed' — wasting tokens and hiding the mistake.
e2e run leaks a throwaway config path into the real ~/.claude/settings.json SessionStart hook. (ANTS-3562) Running the built-in end-to-end test suite quietly edits your real Claude settings to point a startup script at a temporary folder; once that folder is deleted, every later Claude session shows a "file not found" hook error until the app is relaunched and rewrites it.
roadmap_query zero keyword-match emits a misleading "every entry has no [PROJ-NNNN] id" warning. (ANTS-3560) When you search the roadmap and nothing matches, the tool wrongly says "none of your items have IDs" and points you at the wrong fix — when the truth is simply that your search text matched nothing.
0.7.101 RC1 — Patron preview
This release shipped without written notes.
0.7.100
User-level e2e feature case suite (ANTS-2050)
A lane-partitioned QA suite that drives the app as a user through the
ANTS-2049 harness: tools/e2e/cases.sh automates the terminal-basics,
scrollback, resize and theme-restyle-liveness lanes (run all or one lane),
wired into ctest -L e2e; docs/qa/e2e/cases.md enumerates every feature as
drive→expected cases marked auto/manual/pending-hook.
End-to-end test harness (ANTS-2049): a --e2e launch flag + socket-only inject verbs let the app be driven as a user for testing.
New --e2e flag forces the RemoteControl socket open and unlocks four
socket-only inject verbs (inject-key, inject-click, resize-window,
grab-image) behind an m_e2eMode gate; a generic --remote-json client
passthrough carries their argument shapes. A shell driver
(tools/e2e/run.sh) + 8-case smoke suite (tools/e2e/smoke.sh) launch a
throwaway isolated instance, drive it, and reap it, wired as an opt-in
e2e ctest label. The inject verbs are NOT MCP tools, so they are
unreachable against a normal instance (they refuse e2e_disabled).
read_spill row-paging — page a parked list's rows cheaply instead of re-downloading the whole body (ANTS-3545)
When a large list reply is parked (offloaded) and its preview shows an array of rows, read_spill now accepts row_offset/row_count to fetch the NEXT batch of rows, already parsed — instead of byte-paging (which lands mid-row) or re-fetching everything. Returns {mode:"rows", key, rows, row_offset, total_rows, truncated}; page on via row_offset + rows.size(). Completes the ANTS-3538 preview. Pages the same dominant array the preview named (shared detection). Bounded by a stat-before-read 1 MiB parse cap (refuses too_large without loading an over-cap body); non-array bodies refuse not_array (byte-page them). Byte-mode read_spill is unchanged.
workspace_search offset cursor — continue a truncated search instead of re-running it wider (ANTS-3547)
offset:N returns the window [offset, offset+max_results); when more matches remain the reply carries next_offset, which you pass back as offset to page forward (mirrors roadmap_query) — instead of re-running the search wider, which re-scans from scratch and re-emits every match already seen. Pairs with count_only (count first, then page). Dedup applies within each page.
workspace_search files_only mode — return the deduplicated matched-file set with per-file hit counts, no match rows (ANTS-3549)
The rows-eliminated answer to "which files reference X?": files_only:true returns {files:[{file,count}], files_count, count} and omits every match line. Much smaller than headline_only when a symbol recurs many times in one file. File set and per-file counts are complete (uncapped by max_results); count is the true total. Complements count_only (counts only) and headline_only (one line per match).
Structured preview on offloaded Ants-MCP replies — first rows + total, no second fetch (ANTS-3538) When a large array-shaped reply is parked to a spill file, the preview now includes the first few whole results plus the total count (head_rows / row_count), so most callers answer without a read_spill round-trip. The preview is capped and always smaller than the original (INV-9); it falls back to the byte-prefix head for non-array, tabular, over-1-MiB, or marginal bodies.
workspace_search count_only mode — return {count, files_count} with no match rows (ANTS-3537)
A rows-eliminated existence/frequency check ("is X referenced?", "how many call-sites?") that skips serialising match bodies. count is the true total (uncapped by max_results); truncated flags only a cut-off scan. Complements headline_only and max_match_bytes.
cold_eyes_brief: resolve cited-code REGIONS (symbol / file:line) so cold-eyes reviewers read regions + outline, not whole files. (ANTS-3522)
cold_eyes_brief now emits cited_code_regions[] ({path, lines[]}) — the exact lines from each path:line citation, grouped per file — so a cold-eyes reviewer reads a window around each cited line (outline + read_region) instead of the whole file. Additive alongside cited_code_paths; structural completeness is preserved (the reviewer still outlines the file). Cuts review-loop token cost without losing accuracy.
changelog_log op:normalize — one-call tidy for the [Unreleased] section (ANTS-3495) Reorders the ### category blocks under ## [Unreleased] into canonical Keep-a-Changelog order (Added/Changed/Deprecated/Removed/Fixed/Security) without moving any content — the companion fix the malformed-section advisory previously lacked. Non-destructive (bullets stay under their heading; duplicate/custom headings keep their place), with a dry_run preview and a no-write no-op when the section is already canonical.
session_orient now actively flags a stale MCP-server binary (ANTS-3499)
When the running server's build commit is behind the project's HEAD, session_orient emits a server_build_stale block (behind_commits, built, head, hint) so a Claude Code session on an old binary is told to relaunch before trusting shipped status — instead of silently re-reporting already-fixed bugs. Advisory only; never blocks; skipped when the build commit is unknown to the repo or caller_cwd is not a git checkout.
More tab colours, plus a "Custom colour…" picker for any colour you like (ANTS-1374). Right-click a tab → the colour list now offers 14 preset colours (up from 7), and a new "Custom colour…" entry opens a full colour picker so you can tag a tab with any colour at all. Custom colours are remembered across restarts and stay put when you drag tabs around, exactly like the presets.
Prompt-jump chips in the scrollback. (ANTS-1330) Two small floating buttons (↑ / ↓) now appear above the "back to bottom" chip when you scroll up: click ↑ to jump to your previous command, ↓ for the next — the mouse companion to the existing Ctrl+Shift+Up/Down shortcuts. They only show once you've scrolled up and your shell has command markers (OSC 133) enabled.
Stamp the ship-date (+ short commit) into shipped feedback-file findings so a reporting session can self-check its server_build before re-reporting. (ANTS-3504) When we mark a bug fixed in another project's feedback file, also record WHICH build it shipped in. Then that project's Claude session can compare against its own running terminal and know "my copy predates this fix — restart before assuming it's still broken," instead of re-filing an already-fixed bug.
session_orient codebase_index lane digest is empty for repos with no ## Module map — add a capped file-path fallback digest. (ANTS-3503)
The one-call session opener's code map is blank for projects that don't declare a module map, so those projects still get no 'where is the code' hint on the first call — even though the tool already knows the file list.
read_regions should accept requests/paths/regions as aliases for the items batch key. (ANTS-3500)
The batch file-reader only accepts one exact spelling for its list of things to read; a natural guess gets rejected. Accept the obvious synonyms too.
feedback_query — render each assigned id's live ROADMAP status (the deferred ANTS-3448 follow-up). (ANTS-3478) Make the feedback reader show, at a glance, whether each finding's assigned task is planned / in-progress / shipped — pulled live from the roadmap — instead of just listing the IDs.
feedback_log op:append_tracking should refuse (or warn) on a :2 file, pointing at assign_id. (ANTS-3477) The old way of recording IDs (a tracking table) is retired; if someone tries to use it on a new-format file, it should politely refuse and point at the new way.
Add id_prefix override param to the three fold-in verbs (test_audit/cold_eyes/indie_review_fold_in) — parity with roadmap_log op:append. (ANTS-3498) Let a caller force a specific ID prefix on the review fold-in tools, the way the roadmap-append tool already allows.
feedback_log op:migrate_v2 gains backfill_from_tracking — carry inline Proposed-IDs over from the v1 tracking tables (ANTS-3474)
Confidence-gated token match (heading↔tracking-item, per-id overlap-coefficient with a margin) replaces each finding's blank Proposed-ID placeholder with its already-assigned ANTS-NNNN; ambiguous/folded matches stay blank (never a wrong id). Reports backfilled[] {heading, line, id, confidence_pct}; dry_run previews. Unblocks per-finding compact_resolved on migrated files.
git_state op:diff hunk-header mode for clean commit splits (ANTS-3377)
hunks=true returns per-file @@ hunk headers {path, hunks:[{header, old_start, old_count, new_start, new_count, lines?}]} instead of --numstat counts; staged=true diffs the index vs HEAD; include_lines attaches hunk bodies; context sets the unified-context width. Backed by the pure GitWrap::parseDiffHunks helper.
verify_changes orphaned-source lint (ANTS-3373)
Warns when a source file added in the working tree (.cpp/.cc/.cxx/.c++/.c) is referenced by no CMakeLists.txt / *.cmake — it compiles in isolation but is silently never built. Surfaced as an advisory orphaned_sources[] array (does not fail the build gate; emitted only when non-empty). Build/vendor dirs are pruned so a generated build-tree CMakeLists never masks a real orphan.
session_orient's embedded codebase_index sub-object is still counts-only — extend the ANTS-3438 symbol/lane/file selectors (or a compact digest) to the first-call orientation surface. (ANTS-3468) session_orient's embedded codebase map now carries a compact per-lane source_files digest (each subsystem's non-test source paths, sorted, capped) so the first-call "query before grep" map is navigable — a session can jump straight to file_outline/read_region instead of grepping. Opt-in on the standalone codebase_index too via lane_files:true; deterministic, so it keeps the bundle's ETag-304 stable.
MCP read replies now teach each advisory hint once per session, then hush (ANTS-3550) The next_call_hint / leaner_call_hint nudges on read replies are educational the first time and pure recurring overhead after; a process-scoped "already-taught" latch emits each (tool, hint-kind) pair once, then omits it. Keyed per verb so a distinct verb's own tip is never hidden. On by default (token-saver); flip the claude.mcp_hint_latch config key off to get the tips on every call.
Cold-eyes: stop full-reading large append-only cross-ref logs (ROADMAP/CHANGELOG) — consult by search/region. (ANTS-3526) Right now every review helper is told to read the entire 25,000-line roadmap and 19,000-line changelog top to bottom — the single biggest review cost. Have them search those two big logs for the exact thing they're checking instead.
feedback_query mapped_id_status marks cross-repo ids foreign_repo instead of unknown. (ANTS-3518) A mapped ANTS-id whose prefix is absent from the caller project's own roadmap now resolves to "foreign_repo" (with a mapped_id_status_note) rather than the ambiguous "unknown", so a consumer session no longer misreads a shipped suggestion as never-shipped. Fuller cross-repo status resolution is tracked in ANTS-3519.
Faster Roadmap dialog on large histories — the bullet parser is memoized across debounced renders, so typing in the search box no longer re-parses the whole (up to 64 MiB) archive on every keystroke. (ANTS-2119)
Lower idle CPU — a blinking cursor (and hover/expose repaints) no longer redraw the whole terminal. paintEvent now honours the damage rectangle (ANTS-3454): it redraws only the screen rows the update actually touches. The cursor blink invalidates just its own cell, so instead of re-shaping all 40 rows every 550 ms it now repaints a single row. Full-screen updates (heavy output, resize, theme change) are unchanged, so there is no risk to normal rendering.
Faster text rendering under heavy output — the terminal no longer re-computes the shape of every letter on every redraw. A new shaped-run cache (ANTS-3453) remembers the glyph layout of each run of text keyed by its content and style, so an unchanged run is redrawn without re-running the (expensive) text-shaping step. Measured 1.6-1.9x on the shape-and-draw path at a ~100% hit rate; it eases the per-frame cost behind the typing freeze during heavy Claude Code output. Cache is bounded (~8 MB worst case) and cleared on font change.
Skip the shell-history autosuggestion scan when the input line is unchanged, cutting per-batch work during output. (ANTS-3455)
Preserve URL/highlight span caches across streaming frames — only the changed scrollback band is invalidated instead of a full wipe on every push. (ANTS-3452)
Frame-pace output-driven repaints to ~60 fps so heavy terminal output no longer starves keystrokes (typing-freeze relief). (ANTS-3451)
Roadmap ID counter (.roadmap-counter) is now a derived, gitignored cache instead of committed state (ANTS-3450)
The next roadmap ID is now floored to the true high-water mark computed from the committed corpus (ROADMAP.md + CHANGELOG.md + docs/roadmap/ archives), so a stale, wiped, or fresh-clone-absent counter can never reissue a live or migrated ID. This removes the recurring "counter bump left out of the commit" drift entirely — git can't drift a file it no longer tracks.
find_sources no longer stalls on the first call after a restart (ANTS-3444) session_orient now warms the find_sources file set into the OS cache on a background thread, so the first find_sources of a session reads fast instead of paying a one-time cold-disk penalty. Completes ANTS-3444 alongside the earlier CPU cut and the 60s bridge-timeout guard.
MCP server-side latency: find_sources and spec_query list mode are dramatically faster (ANTS-3444)
find_sources now ASCII-folds file bytes in place and searches with
QByteArray::indexOf instead of decoding every file to a lowercased
UTF-16 string — ~5× faster warm (106 ms → 21 ms on the full tree),
identical results. spec_query list mode reads an 8 KiB header per
spec instead of the whole body (title/status always sit in the
first ~124 bytes), skipping the wasted full-body invariant parse.
Crash when changing the colour theme from the menu (ANTS-3556) Picking a theme from View → Theme could crash the terminal on Wayland. The restyle now runs a beat after the menu closes instead of while it is still closing, so the earlier safety guard is no longer bypassed. Locked with a regression test.
Large-result offload never hands back a preview envelope larger than the original reply (ANTS-3540) At a head≈threshold config the fixed envelope overhead plus the head could exceed a body that only just cleared the head guard, so offload could produce a net loss. offloadBody now measures the finished envelope and fails open (returns the untrimmed body) when it would not be strictly smaller — honoring the INV-9 net-saving guarantee at every config.
last_audit_summary no longer mislabels a since-tag / since-last-run / branch-diff / files audit as a single-file rerun. (ANTS-3517)
The narrow_run_warning suppression (ANTS-3512) only trusted the full scope; every explicit multi-file changeset scope now suppresses the false "a broader recent file may exist" alarm, since the sidecar-recorded requested scope is authoritative.
Audit-cache retention no longer deletes a report a kept run still references — a same-second/same-commit filename collision could reap a SARIF/HTML file that a surviving history entry pointed at. (ANTS-2119)
Output triggers now fire once per matching line — a non-instant trigger (notify/command/run_script) fires on every completed line that matches, instead of once per output batch and only when the batch happened to end on a newline. (ANTS-2119)
ANTS-2119 hardening tail (partial): 11 correctness/robustness fixes across the PTY, audit, MCP-dispatch, and Claude-integration lanes. Small, low-risk repairs the June indie-review flagged: a stale-fd guard and a documented env-buffer reserve in the PTY reader; PTY write-loss is now logged instead of silent; the audit review-report handed to Claude gets explicit owner-only file permissions; the review dispatcher enforces its reply size limit in bytes (not characters) and rejects lane names that could escape the reports directory; the drift-check helper no longer throws away the script's error output; project discovery and the brief-file cache are bounded so a corrupt file or long session can't grow memory without limit; and a dead field was removed.
last_audit_summary mislabels a genuine full-tree audit as a single_file/narrow rerun when findings land in one file. (ANTS-3512) The audit summary sometimes warns 'this looks like a one-file re-run, a bigger run may exist' even after a full project scan — because it guesses the scope from how many files had problems, not from what was actually asked for.
workflow_state arg-validation messages — name all missing required args and distinguish absent skill from malformed. (ANTS-3511) When a Claude session calls the workflow_state tool without all its required inputs, the error now tells it everything that's missing in one go, instead of making it guess twice.
Debt-sweep shipped_without_commit now honours the review window. (ANTS-3342)
The detector re-checked every ✅ item in the whole ROADMAP against git log, re-surfacing ~433 pre-convention shipped items each sweep. It now only considers items flipped to ✅ within since..HEAD (diff ROADMAP.md over the window, added-but-not-removed ✅ lines) (ANTS-3342).
Debt-sweep code-quality detectors no longer flag deliberate test-fixture data. (ANTS-3344) orphan_q_unused / obsolete_qstring_idiom / dead_branch_after_return now skip the tests/features/ and tests/audit_fixtures/ subtrees, whose files embed those exact patterns as conformance inputs (ANTS-3344).
Review Changes dialog now refreshes live on file edits, not just commits (ANTS-3509)
The dialog advertised "live — auto-refresh on git changes" but went stale until
you pressed Refresh, because its QFileSystemWatcher could not see a content edit
of a file inside a watched directory (Qt only surfaces add/remove/rename). It is
now driven by a new raw-inotify DirTreeWatcher that watches the gitignore-aware
set of working-tree directories (edits included) plus the .git metadata — one
inotify fd, directories-only (≈4× fewer watches than one per tracked file), and
every watch is released when the dialog closes. Git dir is resolved via
git rev-parse (worktree / sub-dir cwd safe), and all probes run with
GIT_OPTIONAL_LOCKS=0 so a read-only status can't rewrite the index and self-loop.
Silenced 4 more harmless compiler warnings from the diff/review viewer (ANTS-3505) The same false-alarm "-Wnull-dereference" warning quieted in the dialog helper (ANTS-3358) also fired 4 times in the diff/review viewer's "back to top" button positioning. Quieted the same way (a tightly-scoped, GCC-only pragma), keeping the build warning-clean. No behaviour change.
Silenced 33 harmless compiler warnings from the dialog-frame helper (ANTS-3358) A full optimised build printed ~33 false-alarm "-Wnull-dereference" warnings from the dialog centring/resize helper — the compiler couldn't prove a Qt widget pointer was set, even though the code already checks it. Quieted with a tightly-scoped, GCC-only pragma (the same approach as the earlier ANTS-1554 fix), so real warnings no longer get lost in the noise. No behaviour change.
Release version bumps now reliably update the version shown in the README and roadmap (ANTS-2163). The automatic version updater was searching for old wording that no longer matched those two files, so it quietly skipped them and the "Version X" line could show a stale number after a release. Realigned the search text, removed a dead rule for an AppImage filename that is now a wildcard, and taught the drift-check to actually validate both banners so they can't silently fall behind again.
indie_review_orchestrate returns no_lanes when a ## Module map is present-but-unparseable (misleading "heading absent" error). (ANTS-3481) The auto-reviewer says "you have no module map" even when the project DOES have one — it just can't read that particular list format. The error should say so instead of claiming it's missing.
FeedbackFile::skeleton() still births new feedback files as v1 (marker :1 + append_tracking banner). (ANTS-3476) When a brand-new project logs its first feedback, we still create the file in the OLD format with old instructions — new files should be born in the new format.
find_sources returns files_scanned:0 (early-bail) for a valid multi-word topic that leads with a symbol name. (ANTS-3489) The who-uses-this search sometimes scans nothing and returns no results for a reasonable query, when a plain search finds the answer instantly.
debt_sweep_defer renders [ANTS-] hardcoded + un-padded — misses the ANTS-3473/3480 fold-in ID treatment. (ANTS-3497) One of the auto-ID tools (debt-sweep) still writes the fixed 'ANTS' prefix without leading zeros, unlike its siblings which were fixed.
roadmap_query id/ids cannot locate a bullet whose ID prefix is digit-leading (e.g. [3D_E-NNNN]) — the whole project ID scheme is invisible to id-lookup. (ANTS-3492) For a project whose IDs start with a number, the roadmap lookup by ID finds nothing at all — none of its items can be fetched or flipped by ID.
Fold-in verbs emit un-zero-padded IDs (FIBR-82 not FIBR-0082) + inconsistent allocated_ids type — ANTS-3473 residual. (ANTS-3480) The auto-ID feature I added today gets the project's ID prefix right but drops the leading zeros (writes FIBR-82 where every other ID reads FIBR-0082), so the IDs sort and match wrong.
fold-in verbs stamp the project's own roadmap ID prefix instead of a hardcoded ANTS- (ANTS-3473) test_audit_fold_in / cold_eyes_fold_in / indie_review_fold_in / plan_template now sniff the dominant [PREFIX-NNNN] prefix from the project's ROADMAP.md (new RoadmapFoldIn::sniffIdPrefix), so a fold-in into e.g. a FIBR-NNNN roadmap allocates FIBR- and no longer collides with the Ants roadmap's own IDs.
audit_run no longer counts mypy note: (annotation-unchecked) lines as actionable findings (ANTS-3472)
A deps-less mypy run over untyped helpers emitted informational note: lines that inflated total_actionable, misleading /close-phase triage into a phantom fix-pass on a tree the gated mypy reports clean. Note lines are now dropped before the finding count.
codebase_index now covers repo-root source files on non-standard layouts (ANTS-3390) (ANTS-3390) project_settings op:detect now suggests source_roots:["."] when a project's source sits loose at the repo root (RetroArch's retroarch.c, configuration.c, runloop.c … ~9k LoC), instead of a subdirs-only suggestion that silently dropped those files from the index. The whole-root walk also had its keys normalised — a "." source_root previously stored "./retroarch.c", so codebase_index(file_path:"retroarch.c") returned found:false; it now keys the bare repo-relative path, which also repairs flat-root ["."] projects (ANTS-3393) whose file_path lookups and tests/ role-detection were silently broken.
focused_test returns an MCP -32000 transport timeout when a broad changed-file (e.g. remotecontrol.cpp) maps to a large ctest set that outruns the transport deadline. (ANTS-3449) focused_test now runs ctest in parallel (-j, capped at 4 / host cores), so a broad change that falls back to the full suite finishes in ~26s instead of ~78s and no longer trips the MCP transport timeout (spurious -32000) while the tests were still running fine underneath.
feedback_query/feedback_log path auto-derivation doubles a trailing '_Ants' in the project leaf. (ANTS-3426) feedback_query/feedback_log no longer double a trailing "_Ants" in the auto-derived filename (fork checkouts like DOOM_Ants now resolve DOOM_Ants_MCP_Feedback.md instead of the never-existing DOOM_Ants_Ants_MCP_Feedback.md), so a path-omitted call reads the real file and cannot fork history into a wrong-named duplicate.
feedback_log op:append_tracking does not pipe-escape | in item/notes cells — a pipe in the text corrupts the markdown table row. (ANTS-3469)
When the triage tool writes a tracking row containing a pipe character in its text, the table row breaks, and a later cleanup step can't read the row's status.
roadmap_log op:amend_body now reaches body text on blank-line-separated nested sub-bullets, and its body_match_not_found refusal hints when old_text sits outside the bullet or spans a hard-wrapped line break. (ANTS-3467)
workspace_search now hints 'did you mean regex:true?' when a regex:false pattern carrying regex metacharacters (e.g. A|B|C) returns zero matches, so an empty result isn't mistaken for 'symbol absent'. (ANTS-3466)
find_definition now resolves C++ type definitions (struct / class / union / enum) — a pure type with no same-named constructor previously returned zero definitions. (ANTS-3465)
roadmap_log note append is now retry-idempotent (ANTS-3440) A roadmap_log op:flip/annotate that committed its ROADMAP.md write but surfaced to the caller as a transport timeout used to duplicate the resolution note when retried. appendBodyNote now dedups a byte-identical trailing note, skipping the second append and reporting note_already_present:true. Applies to flip (GFM + ants-v1), annotate, and flip_batch (shared primitive).
Kitty graphics images are now strict-base64 decoded — a corrupt image payload is rejected outright (matching the OSC 52 / iTerm2 paths) instead of silently feeding a garbage-prefixed stream to the image loader. (ANTS-2119)
OSC 8 hyperlink phishing warning now covers the common case — a benign-looking link label ("Download", "click here") pointing at a punycode/IDN-homograph or bare-IP destination now prompts before opening, not only when the label itself looks like a mismatched hostname. (ANTS-2119)
Audit window AI-review buttons now enforce the same network safety checks as the main AI client (ANTS-2121). The audit window's "AI triage" buttons talk to the AI server over their own connection. They were missing three protections the main AI client already has: refusing internal/cloud-metadata addresses (SSRF), refusing sneaky redirects, and refusing credentials hidden in the URL. All four egress checks now live in one shared validator both paths call, so the audit buttons are guarded identically.
0.7.100 RC1 — Patron preview
This release shipped without written notes.
0.7.99
feedback_log op:"assign_id" — v2 inline triage write (ANTS-3447)
Fills one ### finding's **Proposed ID:** line in a shared *_Ants_MCP_Feedback.md file with the assigned ANTS-NNNN id(s) or an n/a — <reason> closure — the inline replacement for the v1 op:append_tracking table. Locates the finding over the fence-aware ### enumerator (target_not_found / target_ambiguous+candidates), replaces the placeholder/existing line or inserts one when absent, idempotent, atomic, dry_run-previewable. No roadmap read.
feedback_log op:"migrate_v2" — one-shot mechanical v1→v2 feedback-file migration. (ANTS-3446)
Bumps the version marker to <!-- ants-mcp-feedback: 2 --> and stamps a
blank - **Proposed ID:** _(maintainer to assign)_ line on every
finding-shaped, below-watermark ### block that lacks one. Leaves the v1
tracking tables in place (no move/collapse), so the v1 watermark and the
shipped un-triaged delta are preserved. Reports stamped/orphans/
unclassified; already_v2 no-op on a v2 file; dry_run previews;
idempotent, atomic. This is the converter that lets compact_resolved
(ANTS-3443) act on the all-v1 corpus.
feedback_log op:compact_resolved — collapse a shipped v2 finding's write-up to a roadmap-driven stub (ANTS-3443)
Maintainer, v2 files only. Auto-discovers every ### finding and collapses each whose inline **Proposed ID:** id(s) are all ✅ in the live ROADMAP.md to a one-line → shipped ✅ (write-up compacted, ANTS-3443) stub that keeps the heading + the id line — never deletes (the write-up survives in git + under its ANTS-NNNN). Per-finding, idempotent, atomic, dry_run byte report; per-finding skips (no_shippable_id / already_compacted / roadmap_unresolved_ids / has_open_id) in skipped[]. Refuses not_v2 on a v1 file (run migrate_v2 first) and roadmap_unavailable when no ROADMAP.md resolves. Introduces the shared fence-aware ### -block enumerator the pending v2 feedback_query/migrate_v2 verbs will reuse.
find_sources emits a redirect hint on an empty result (ANTS-3435) An empty files[] now points at workspace_search / find_definition / find_caller so a symbol-led query is not mistaken for genuine absence.
feedback_log op:"prune_tracking" — dedup superseded maintainer tracking rows (ANTS-3442)
Removes superseded duplicate maintainer tracking-table rows from the cross-session *_Ants_MCP_Feedback.md files, keeping each id's authoritative last-per-id row plus every heading/header/separator. Two-stage pass preserves mappedIds (a notes-cell-only id pins its row); idempotent, atomic, dry_run preview, optional scope_ids. The deduplication complement to compact_shipped's write-up collapse.
*feedback_log: compact-shipped op to trim confirmed-implemented findings in _Ants_MCP_Feedback.md (keep a stub, not a delete). (ANTS-3421) The cross-session feedback files keep growing; once a suggestion is built and the other session confirms it works, its long write-up should shrink to a one-line stub so the file stays small.
feedback_query / feedback_pending honour the v2 marker (marker-aware un-triaged delta) (ANTS-3448)
On a <!-- ants-mcp-feedback: 2 --> (or higher) file the un-triaged delta is now the findings whose inline **Proposed ID:** is still unfilled, not the region after the last v1 tracking table. feedback_query gains format_version + suspected_untagged[]; session_orient feedback_pending picks up the v2 count with no code change. The v1 rule is retained verbatim for un-migrated files.
changelog_log refuses a feature-grouped [Unreleased] (dated ### topics + **Bold** category runs) with feature_grouped_section instead of inserting a mis-ordered flat category. (ANTS-3416)
On changelogs that group notes by feature (newest-first), the writer now stops and asks you to hand-edit, instead of adding an entry at the wrong place that you then have to move.
prune_tracking wrongly drops distinct id-less closure rows sharing a non-ANTS id-column token (e.g. (schema fix)) — confirmed data loss. (ANTS-3445)
The tool that de-duplicates feedback-file bookkeeping rows can delete a real row by mistake when two unrelated rows are labelled the same non-ID tag. Fix it so only real ANTS-IDs count for de-duplication.
feedback path derivation tolerates checkout-leaf vs package-name case/separator mismatch (ANTS-3439) A Fin_Break checkout resolves finbreak_Ants_MCP_Feedback.md via a normalized compare; a normalized-equal sibling is floated to candidates[0] without a false all_other_projects.
roadmap_branch_drift scans legacy no-ID ants-v1 roadmaps (ANTS-3437) No more false scanned_bullets:0 all-clear; id-less bullets are identified by headline slug and the envelope flags roadmap_format.
spec_query / spec_log accept the numeric NN-topic ids that spec list mode emits (ANTS-3436) e.g. id=17-emission-model no longer refuses bad_id; the read surface accepts the identifiers it hands out.
read_region bare-name symbol-mode confirmed for two-word-return-type members (ANTS-3434) Verified as a downstream effect of the file_outline drop; a bare method name resolves the qualified member via the existing suffix fallback once it is kept in the outline.
file_outline no longer drops out-of-line C++ members with two-word return types (ANTS-3433) unsigned int / long long / unsigned char / const T& return types on a Class::method definition are outlined again (and so resolvable via read_region symbol-mode).
roadmap_log op:bundle_row now works end-to-end (ANTS-3432)
The bundle_row op (ANTS-1691) advertised cells/header/position/sort_col but never declared them as schema properties, so with additionalProperties:false the MCP client stripped the args and every call refused missing_field. Declared the four properties; the existing handler is now reachable. Regression-locked with a schema source-grep invariant.
Bulk debt-sweep defer now warns before dumping a whole scan into the roadmap. (ANTS-3346)
The Project Audit "Defer all to ROADMAP" button (and the debt_sweep_defer MCP verb) used to fold every finding of a scan — real, false-positive, and mechanical alike — straight in as tracked items; that is how 1,106 false positives once landed there. Deferring more than 25 findings at once now asks for confirmation in the dialog (pointing you at "Triage with AI" first) and is refused by the MCP verb unless it passes triaged:true. The guard keys on how many findings you defer, not on whether they look auto-fixable — a scan that is mostly auto-fixable can still be mostly false positives.
debt_sweep_scan produces unbounded output — transport timeout + token overflow. (ANTS-3345) Running the scan on a real release-sized diff either times out or returns a 130k+ character blob the assistant can't read.
General read/write verbs must reach *_Ants_MCP_Feedback.md outside the project root (supersede ANTS-3419 hint-only decision). (ANTS-3430)
The shared feedback files live one folder ABOVE every project, so the normal read/edit tools refuse to touch them. Let those tools reach feedback files directly instead of only pointing at the special feedback commands.
roadmap_query by-id: max_body_bytes (ANTS-3402) is inert — a single-bullet fetch truncates the body at the 2000 default regardless of the requested cap. (ANTS-3425) Asking the roadmap tool for one item's full text still cuts it off at ~2000 characters even when you ask for more, so the important tail (like a decision paragraph) is lost and you have to read the whole 2 MB file instead.
audit_run no longer auto-runs mypy in its deps-less sweep (it emitted 28-31 spurious import-not-found/import-untyped warnings on every full run); mypy stays available on explicit tools:["mypy"]. (ANTS-3418)
The built-in bug-scanner stopped falsely reporting dozens of 'missing library' problems on Python projects every time it ran; CI and pre-commit still run the real, fully-installed type check.
0.7.99 RC1 — Patron preview
This release shipped without written notes.
0.7.98
subsystem op:map accepts an optional name filter (ANTS-3414)
Pass name to narrow the returned lane list to subsystems whose name contains that substring (case-insensitive); omit it for the full map. Previously name was ignored on op:map.
RoadmapDialog surfaces a duplicate-ID warning banner (ANTS-1694) When the roadmap accidentally carries the same [PROJ-NNNN] id on more than one bullet — which roadmap_query already flags — the cards view now shows a warning banner at the top, using the same canonical-ID predicate as the MCP detector.
roadmap_log op:"amend_body" — patch a bullet's body prose in place (ANTS-3406) Locate a bullet by id/anchor/headline, then replace an exact single-line substring of its continuation body (old_text→new_text) with a unique-match guard, dry_run preview, and leaked-XML scrubbing. Works on GFM and ants-v1 roadmaps. Fixes the cross-session gap where correcting a stale phrase in a roadmap entry meant a raw text edit.
roadmap_query max_body_bytes raises the body cap for a targeted id/ids fetch (ANTS-3402)
A single-bullet / id-set read can now return a large multi-phase epic body (clamp [2000, 16384]); list and section queries stay at the 2000 cap.
roadmap_query accepts the granular lifecycle filters planned / in-progress / considered (ANTS-3400)
The read verb now takes roadmap_log's own status vocabulary (planned=📋, in-progress=🚧, considered=💭) alongside active/shipped/all, and echoes an accepted list on an unknown value. section_index collapses the granular names to their aggregate.
Opt-in async mode for the audit_run MCP verb (ANTS-3396). (ANTS-3396)
Pass async:true to start a slow sweep detached and get a job handle back instantly ({job_id, status:"running", poll_with:"audit_poll"}) instead of blocking on the ~60s MCP transport cap; poll the new audit_poll {caller_cwd, job_id} verb for completion (running/done/error/expired). The sync path is unchanged (default async:false). The job registry is bounded (16 entries, 270s reap) and root-scoped so a poll only sees its own project's jobs. Results are written to .audit_cache regardless of poll.
roadmap_log append accepts optional evidence:[paths], echoed by roadmap_query (ANTS-3382) op:append / append_batch take an evidence file-path array, rendered as an Evidence: line (dots in paths preserved) and echoed back by roadmap_query as an evidence array — so a bug logged from screenshots/logs records where the files are for a later session.
feedback_log / feedback_query: derive the feedback-file path when path is omitted (ANTS-3376)
Omit path and the verb derives _Ants_MCP_Feedback.md
at the shared root, so a first-time log no longer needs a filesystem hunt;
the reply carries path_derived:true. On a not-found, the candidate list now
floats your own project's file first, or flags that every candidate belongs
to another project. The canonical per-project filename rule is documented in
docs/standards/mcp-feedback-files.md (ANTS-3384).
feedback_query can now report which of a session's past suggestions shipped. (ANTS-3371) Pass include_tracking:true and the reply carries a compact tracking list (item, roadmap IDs, status, notes) pulled from the maintainer tables — so a contributor session sees what's done without re-reading the whole 100KB+ feedback file.
The session-start summary now flags the audit-warning count as stale when the saved audit predates your latest commit. (ANTS-3370) current_state / session_orient now carry open_audit_findings_count_stale, mirroring last_audit_summary's staleness signal — so a session doesn't chase warnings that were already fixed after the audit was cached.
Donate menu in the menu bar with GitHub Sponsors and Patreon links A new Donate menu (just left of Help) opens the project's GitHub Sponsors and Patreon pages in your browser, so supporters can find them without leaving the terminal.
dry_run preview on the last two mutating Ants-MCP verbs — test_audit_fold_in and debt_sweep_apply_fix (ANTS-2227)
Completes dry_run parity across all 9 mutating verbs. test_audit_fold_in previews the would-be ROADMAP IDs + block without bumping the counter or writing; debt_sweep_apply_fix runs every guard and computes the patched source but skips the write, reporting would_apply. Pass dry_run:true to any of them to see the change before it happens.
dry_run:true preview extended to the ROADMAP fold-in verbs (ANTS-2227)
indie_review_fold_in, cold_eyes_fold_in and debt_sweep_defer now accept
dry_run:true, returning the would-be allocated IDs and rendered block
without bumping .roadmap-counter or inserting into ROADMAP.md — so Claude
can preview a fold-in before committing it. Backed by a new non-mutating
RoadmapFoldIn::peekIds primitive that mirrors allocateIds' ID math.
dry_run:true preview on four more mutating Ants MCP verbs (ANTS-2227)
apply_edits, project_settings, feedback_log and audit_falsepos_log now
accept dry_run:true, returning the would-be result (carrying
dry_run:true) without writing to disk — so Claude can pre-flight a
multi-file edit (catching a no-match before any partial write) or any of
these appends before committing. Joins roadmap_log / changelog_log /
spec_log, which already supported it. Each preview shares the verb's real
write code so it can't drift from the actual result.
Screen-reader support for the terminal area (AT-SPI / Orca). (ANTS-1078) The terminal's content is now readable by assistive technology. A QAccessibleInterface adapter (role Terminal) with a text interface exposes the visible viewport as plain text, reports the caret, and fires a throttled change event as output arrives — so a blind user's screen reader can read the terminal, not just the window chrome. Costs nothing when no screen reader is running. Core slice of the H9 accessibility bundle; richer text attributes, command-boundary batching, and selection write-back are tracked separately (ANTS-3363). Also resolves the terminal-a11y gap in ANTS-2205 (4).
Tabular (columnar) encoding for homogeneous-array MCP replies (ANTS-2090).
Opt-in encoding:"tabular" on the list-shaped read verbs (roadmap_query,
workspace_search, file_outline, find_sources, find_caller, codebase_index,
docs_index) packs each top-level array-of-objects into a columnar
{cols,rows} form — the per-row keys are dropped, so big list
replies are 30–60% smaller. Self-guarding per array; never made larger
than plain JSON; refusals/304s untouched. Decode by zipping cols with
each rows row (a null cell = key absent on that element).
Ants MCP: project_settings op:detect no longer ranks a bundled-dependency tree (e.g. mingw-deps/, third_party/) as a project's source root — vendored dirs are discounted from the suggestion and the file total. (ANTS-3357)
Ants MCP: spec_query with no id/path now lists the specs directory ({mode:"list", specs, count}) instead of erroring — spec discovery without a shell ls. (ANTS-3360)
Ants MCP: spec_query / spec_log accept any <PREFIX>-NNNN spec id (e.g. DOOM-0009), not just ANTS-NNNN, and resolve topic-suffixed spec files (DOOM-0009-path-tracer.md) from the bare id. (ANTS-3356)
Generalised the spec id routing the same way roadmap_log was (ANTS-2076); path stays the explicit override for non-standard layouts.
MCP dispatch silently ignores unknown / misspelled args — echo an ignored_args advisory. (ANTS-2175)
If Claude passes a setting name a tool doesn't recognise (a typo, or an option that doesn't exist), the tool quietly ignores it and can return wrong-looking results. It should name the inputs it ignored so the mistake is caught on the first call instead of after several.
Implement INV-13 scroll-position persistence (roadmapdialog). (ANTS-1264) the Roadmap dialog is supposed to remember where
read_regions MCP verb — batched multi-selector read (read-side mirror of apply_edits). (ANTS-2219)
Let Claude fetch several code snippets in one request instead of one call each — fewer round-trips, fewer tokens.
Surface feedback_query / feedback_log to contributor CC sessions (self-advertising banner). (ANTS-2226)
Tell the other projects' Claude sessions they can read and add to these feedback files through Ants directly, instead of editing them by hand.
read_region markdown section selector — read a heading's body by slug. (ANTS-2221)
Let Claude ask for one section of a doc by name instead of working out its line numbers first.
workspace_search enclosing_symbol — annotate each match with the function it lives in (ANTS-2220).
Opt-in enclosing_symbol:true adds enclosing:"Foo::bar" to each match (the nearest-preceding symbol from the file outline), folding the usual "which function is this in?" follow-up into the search. One outline scan per distinct matched file, cached; off by default. (DOOM_Ants feedback S2.)
file_outline multi-path — outline several related files in one call (ANTS-2223).
Pass paths:[...] instead of path to outline a header + its impl + a consumer in a single MCP round-trip; returns a files:[{path, symbols, etag, …}] array, each entry resolved and byte-capped independently. An optional etags map ({relPath: priorEtag}) 304s any unchanged file to a compact stub, so re-outlining a set after editing one file re-sends only the changed bodies. Single-path callers are unaffected. (DOOM_Ants feedback S5.)
workspace_search now emits a regex_advisory when a regex:true alternation contains very short un-anchored terms (e.g. tan) that substring-match inside longer words, flooding results. (ANTS-2181)
Claude can now ask Ants to compute an answer across your project's files and get back just the result — not the file text — via a new sandboxed project_query MCP verb. (ANTS-2093)
The "code-execution" token-saver: Claude sends a small read-only Lua snippet (with project.read/list/root + string/table/math/utf8), Ants runs it safely over the project and returns only what it computes — a count, a filtered list — instead of reading every file into the conversation. Fully sandboxed: read-only, confined to the project directory, no network/exec/write, and memory/time/output-capped on a worker thread so a runaway snippet can never freeze the app. On by default (Settings → General → "Let Ants run read-only project queries for Claude"); compiled out in builds without the Lua subsystem.
Master on/off switch for the Ants MCP integration (ANTS-1901) (ANTS-1901)
Settings → General → "Enable Ants MCP integration" (on by default) is a
single master switch for the whole Ants ↔ Claude Code helper. Turn it
off and Ants stops the integration: the MCP socket isn't bound at next
launch, the start-of-session cheat-sheet is removed, the auto model
switcher stands down, and every MCP tool call is refused with a
mcp_disabled message. Turning it back on takes effect on the next Ants
launch; turning it off is honoured immediately. Every dependent toggle
(auto-switcher, cheat-sheet, lean replies) greys out when the master is
off. Plain terms: one switch to fully turn the Ants–Claude integration
on or off.
Added an out-of-cadence hotfix path, cut-rc.sh hotfix. (ANTS-2165)
Two phases around a /bump that ship the latest public release plus a cherry-picked fix as the next public patch (and roll any in-flight RC up one number), for an urgent bug that can't wait for the next Wednesday.
Roadmap dialog: items now lay out in aligned columns instead of a cramped inline run. (ANTS-3392) Each roadmap item's status, type, summary and ID line up in neat columns (the dialog renders each section as a table) instead of running together into a wall of text, and section-heading counts no longer fuse into the section title.
roadmap_query MCP provider now forwards args verbatim (rcDelegate), retiring the hand-maintained per-arg allowlist that silently dropped new options. (ANTS-3422) The mainwindow provider that bridges the roadmap_query MCP verb to its handler used to copy each option across the boundary by hand; forgetting a line silently disabled that option (it happened five times). It now forwards the whole request, so every current and future option reaches the handler automatically.
Read/write MCP verbs point you at the right tool for a feedback-file path (ANTS-3419)
A bad_path refusal on a *_Ants_MCP_Feedback.md path (which lives above the project root) now carries a hint redirecting to feedback_query / feedback_log — the verbs that do serve it — instead of a bare error.
find_sources accepts symbol as an alias for topic (ANTS-3415)
The 'who uses X' verb now takes symbol (filled in only when topic is absent), so copying an identifier between find_sources and find_caller no longer fails the first time.
tools/ci-parity.sh now mirrors every locally-reproducible CI gate (ANTS-3410) Added --lints (cppcheck / appstream / desktop / groff / version-drift / bash+zsh+fish completions), --asan (Debug ASan/UBSan build + sanitized ctest + --version/--help smoke) and --full phases. Gates accumulate and report all failures at the end; a gate whose tool is absent is skipped with a loud warning so a local "green" never masks a gate CI still runs. The qt62-baseline job is intentionally not mirrored (needs a different distro + Qt). Motivated by ANTS-3391's wire-budget overrun staying CI-red for three commits before ANTS-3409 caught it.
Document the Evidence: bullet field in roadmap-format.md § 3.5 (ANTS-3382 follow-up). (ANTS-3386) The new Evidence line on roadmap items works and is documented in the live tool schema, but the human-facing roadmap-format standard doesn't list it yet.
roadmap_log documents that flip/annotate resolve top-level bullets only (ANTS-3403)
Nested phase sub-bullets inside a narrator bullet's body are out of scope for the locators and need a direct Edit; the op descriptor now states this.
changelog_log malformed-section advisory now names the alternative insert path (ANTS-3401)
When a feature-grouped [Unreleased] layout is detected, the advisory suggests op:add_from_roadmap into a named subsection or a hand-edit.
roadmap_log counter strategy auto-creates .roadmap-counter at 0 on a greenfield roadmap instead of refusing (ANTS-3397)
A counter-strategy op:append / op:append_batch on a roadmap that has no .roadmap-counter AND no existing bullet ids now auto-initialises the counter at 0 and proceeds (first id allocates as -0001), rather than refusing with counter_missing and forcing a shell echo 0 > .roadmap-counter side-step that broke the all-MCP workflow. A roadmap that already carries ids but lost its counter file still refuses (counter_missing) — re-allocating from 0 would mint duplicate ids. Reported by the Contact_List CC session.
project_settings op:detect now suggests all first-party source dirs on a non-standard layout, always explains itself, and echoes the roots/dirs it resolved (ANTS-3369). Cross-session feedback from four projects: the layout auto-detector gave up silently on projects that don't use a plain src/ folder (it required one dominant subdir to cover ~90% of the code, so a spread layout or a small entry-point dir like app/ got nothing). It now suggests every first-party source subdir on a miss, always returns a non-empty reason, and adds would_use_roots (the roots already in effect) + excluded (the vendored/build dirs it skipped) so the output is never an opaque set of zeros.
read_region / read_regions / file_outline tool descriptions now warn that an MCP read does not satisfy the native Edit read-precondition (ANTS-3383). Cross-session feedback: reading a file via these verbs and then editing it fails the harness's "File has not been read yet" check, so a native Read is still required first. Documented inline in the tool help so sessions stop paying for a doubled read.
*feedback_query / feedback_log not_found returns sibling _Ants_MCP_Feedback.md candidates. (ANTS-3366) When Ants can't find a feedback file at the guessed name, it now lists the real ones nearby so Claude doesn't have to shell out to find them.
git_state defaults op to "status" when omitted. (ANTS-3365) Asking Ants for git status in one call now just works without naming the sub-command.
De-duplicate the KWin scripting chain in moveViaKWin / centerWindow into a shared runKWinScript helper (ANTS-2205 7b) (ANTS-2205)
read_regions discoverability — surface the batch reader in the SessionStart prelude + cross-link it from read_region. (ANTS-3359)
When Claude needs several chunks of a big file, the cheaper "read them all in one go" tool wasn't visible up-front, so it fell back to the slower built-in reader. Make that batch reader easy to find.
Review-dialog and clipboard/AI size caps: honest unit naming + de-duplicated truncation tail (ANTS-2205)
Renamed the clipboard and AI-command length caps from …MaxBytes
to …MaxChars so the name matches what they measure (QString
length units, not bytes — they never changed behaviour). Pulled
the review dialogs' shared "clip-to-cap then close any open code
fence" step into one BriefDispatch::clipToCapClosingFence helper
and lifted the 200 KiB prompt cap to the shared base, removing
three copies. Internal tidy-up; no user-visible change.
Extract ClaudeTranscriptWalker (claude-trackers). (ANTS-1261)
the two task-list trackers have ended up as
cppcheck cleanup: default member initializers on 32 struct fields + 2 const-ref getters. (ANTS-3341) An /audit run flagged a batch of small code-quality warnings: several data structures had number/flag fields with no default value (a latent source of "uninitialised value" bugs if one were ever used before being filled), and two small accessor functions copied a string instead of handing back a reference. Gave every flagged field a sensible default (0 / false / the documented default) and switched the two accessors to return by reference. Verified with cppcheck that the warnings are gone and all 2314 tests still pass.
read_region section selector: forgiving match for a heading's short title when it carries a trailing parenthetical (ANTS-2221 follow-up). (ANTS-2234)
When a Claude session wants to read one section of a spec, it naturally types the short heading title (e.g. "7. Build order"). But many headings have a descriptive bit in brackets after them ("7. Build order (cheapest-first; ...)"), and right now the short title doesn't match — the session has to type the whole thing or fall back to line numbers. This makes the section reader accept the short title as long as it unambiguously points at one heading.
Link with the mold linker when available (-fuse-ld=mold, auto-detected, opt-out). (ANTS-2233)
Switched the build's final "linking" step to a much faster linker called mold (it was already installed but unused). Linking is the slowest, most memory-hungry part of the build, and mold is both faster and lighter on memory — so the build finishes sooner without risking the machine hanging. Falls back to the old linker automatically if mold isn't installed.
Parallelise the ctest presets (-j4 default/fast, -j2 workstation; debug/perf stay serial). (ANTS-2231)
Running the whole test suite used to take about 78 seconds because the tests ran one after another. They now run 4 at a time, finishing in about 19 seconds — roughly 4× faster — capped at 4 so it won't bog the machine down.
Regression test for the per-language lineIsCode lexer (ANTS-1270 follow-up). (ANTS-2210)
Add an automated test proving the audit pipeline correctly treats # comments (Python/shell) and Lua --/[[ ]] as comments, so the ANTS-1270 fix can't silently regress.
read_region symbol-mode returns a struct/class/union's full body (ANTS-2222).
Asking for an aggregate by name used to return only its declaration line (the flat outline ends it at the first member); it now brace-matches to the closing brace and returns the whole body. A namespace (also tagged kind class) still resolves to its head — use line mode. From DOOM_Ants cross-session MCP feedback.
Roadmap dialog (Cozy density, the default) — the status / kind / section-count labels are now 12px (was 10–11px), keeping a readable Compact (11px) < Cozy (12px) < Comfortable (13px) ladder. (ANTS-2211)
Roadmap dialog (Compact density) — the smallest text (status / kind / section-count labels) is now 11px instead of 9px, for readability. (ANTS-2211)
Status bar stops polling while the Ants window is in the background — saves battery and scheduler wakeups on laptops. (ANTS-1363)
computeConfidence moved to AuditEngine so non-GUI consumers score findings without Qt6::Widgets. (ANTS-1262)
The confidence formula was a pure data-transform stranded in the GUI dialog. It now lives in AuditEngine::computeConfidence (single source of truth); AuditDialog::computeConfidence forwards, so call-sites and the public static surface are unchanged.
Codebase/docs index refresh now scans the project tree once per call instead of two-to-three times. (ANTS-2198)
Roadmap pagination cut-point search is now O(n) instead of O(n log n) (ANTS-2200) Each list element is serialized once and accumulated, rather than re-serializing a growing prefix on every binary-search probe. Same output, less work on long lists.
MCP read responses now surface the etag re-use hint (next_call_hint) regardless of body size — repeated small read_region / file_outline slices can now 304 their re-reads. The 4 KiB gate now applies only to the leaner_call_hint. (ANTS-2180)
Ants MCP now offloads oversized read replies to a short preview + pointer by default Large MCP read results (roadmap_query, workspace_search, get_scrollback, …) are saved to a scratch file and returned as a short preview plus a handle, so a Claude session spends a few tokens instead of thousands and fetches the rest (via read_spill) only if it needs it. Shipped switched off in the prior release; now on by default per the "token-savers default ON" rule, and it fails open — if the scratch file can't be written, the full reply is sent as normal. A new Settings → General toggle ("Offload huge Ants MCP replies to a preview + pointer") turns it off. (ANTS-2094 fast-follow)
Hardened the weekly RC→release→new-RC cadence into one guarded cut-rc.sh cycle. (ANTS-2164)
new-rc refuses an empty RC and auto-rolls [Unreleased]; promote refuses an empty/placeholder or stale (>14-day) RC and auto-date-stamps the CHANGELOG/metainfo/debian release notes; version-drift is now a hard gate. Stops the recurring empty-RC / half-finished-release class of incident.
Deleted the decayed claude_statusbar_extraction source-grep test (ANTS-1381).
It froze the 0.7.74 ClaudeStatusBarController extraction with byte-for-byte signature/connect-site/LoC greps; the controller's behaviour is covered at runtime by the model_near_miss_ledger and model_auto_switch_outcome_fillin tests. Investigation found the other five "extraction" tests in the same roadmap item are NOT decayed (two carry live runtime unit tests; three are the sole guard for their behaviour) and were kept.
Deleted the dead IndieReviewEngine::assembleBrief zombie — a latent prompt-injection trap (ANTS-2187).
The unfenced v1 brief assembler had no production caller (only doc
comments + one feature test) yet inlined raw source into an LLM prompt
with no fence-hardening, and its name read like the main entry point —
so a future rewire would have reopened a prompt-injection hole. The live
path is assembleBriefForDispatch (fence-hardened via
BriefDispatch::fenceBody) / assembleBriefManifest, both already covered
by their own tests. Removed the function, its declaration, the stale
header contract comments, and the orphaned Inv3 shape test.
roadmap_log / changelog_log output is now pre-commit-clean (ANTS-3417)
Emitted continuation lines are right-stripped, so the ubiquitous trim-trailing-whitespace pre-commit hook no longer rewrites the file and aborts the commit after a roadmap/changelog append.
file_outline no longer drops methods with std::function<void()> params or inline const accessors (ANTS-3412)
The C++ scanner's argument matcher stopped at the first ), so a parameter whose type carried inner parentheses (e.g. std::function<void()>) truncated the match and the whole method vanished from the outline; and a trailing const/noexcept on a one-line accessor (int workerCount() const { ... }) also hid it. Both are now recognised.
roadmap_query's query keyword filter now actually filters (ANTS-3420)
The keyword search shipped in a prior release but was silently inert: the MCP dispatch layer never forwarded the query argument to the handler, so a search returned the whole roadmap instead of the matching items. The same audit fixed three sibling arguments that were also being dropped over MCP — max_body_bytes (large-body id fetch) and the per-section ETag args (include_section_etags, section_etag_match).
file_outline now detects extern "C" function definitions (ANTS-3351) (ANTS-3351)
The C++ outliner skipped any function with an extern "C" linkage prefix, so its interior local variables were mislabelled as top-level functions. Reported against DOOM's r_vulkan.cpp (the RB_Vulkan_* entry points were missing and devs/exts locals showed up as functions).
RoadmapDialog section chips now roll up descendant counts, matching the MCP (ANTS-1693) Parent (level-2) section chips previously showed only their direct bullets, so they could disagree with roadmap_query's section_index active_count for the same slug. RoadmapIndex::rollupCounts is now a shared template that both the dialog and the MCP use, so a parent chip sums self + descendants.
roadmap_query MCP one-line description back under its 800 B wire budget (ANTS-3409)
Trimmed the redundant bundles mode gloss so the reconstructed
wire description is 790 B (was 814), restoring the
mcp_tool_detail_field INV-5 budget. Full per-op detail is
unchanged — still served via tool_info {name:"roadmap_query"}.
roadmap_query granular status filters now work without a section arg (ANTS-3408)
status="planned" / "in-progress" / "considered" silently returned 0 on
the full-file (no-section) query path — ANTS-3400 added those arms to the
section= branch only. The full-file predicate now mirrors them. Root
cause was section-path vs full-file-path divergence, not the roadmap
format. Reported by the Contact_List CC session.
read_region call_sequence — suppress false-positive callees from comments and type constructors. (ANTS-3379) The 'list the steps in this function' feature sometimes lists words from comments or type names as if they were function calls.
similar_code — down-weight trivial one-line/one-token signatures in ranking. (ANTS-3380) The 'find code like this' search sometimes puts a tiny irrelevant one-liner at the top.
roadmap_query id/ids + roadmap_log flip/annotate return a silent miss (not bad_id_format) for a format-nonconforming bracket token that exists in the file. (ANTS-3387) When a project uses an ID style Ants doesn't recognise, looking it up says 'not found' with no hint that the ID shape is the problem — so it looks like the item vanished.
token_usage now credits the read/search verbs for the tokens they save (ANTS-3361) file_outline, read_region(s), workspace_search, codebase_index, find_definition/sources/caller and build_status previously reported ~0 tokens saved despite replacing a full-file Read or a grep. They now carry conservative per-call baselines (deliberately under-estimated and floored at 0 so the headline never over-claims).
workspace_search returns a soft rg_failed envelope instead of a raw transport timeout on huge match streams (ANTS-3405)
The JSON parse loop is now bounded by the wall budget (not just the rg process), so a whole-repo query over large data blobs no longer blows the client's transport timer with an un-catchable -32000.
file_outline now emits Python class methods as Class.method (ANTS-3404)
The Python outliner tracks enclosing classes by indentation, so read_region symbol mode can address a method (e.g. PlayQueue.next); top-level defs/classes stay bare.
file_outline no longer glues a namespace-qualified C++ return type onto the symbol name (ANTS-3399)
A method like JPH::BodyID PhysicsWorld::createStaticBody(...) now outlines as PhysicsWorld::createStaticBody (was the whole return-type-plus-name), so read_region symbol mode resolves it. read_region also gained an unambiguous qualified-suffix fallback so a bare method name resolves its qualified entry.
audit_run no longer turns a tool's progress bar or fatal error into a fake finding (ANTS-3395) JSON-emitting tools (ruff, bandit, semgrep, gitleaks, trivy, shellcheck) are now parsed strictly as JSON, so bandit's progress bar and trivy's FATAL log line stop being counted as findings. A trivy scan that aborts is reported as a crashed/incomplete tool instead of a single misleading finding.
audit_run scope:"full" now skips build-output, vendored, and cache dirs (ANTS-3394) The default exclusion set (logs, data, database, node_modules, dist, build, _build, pycache, .venv, venv, env) is now passed to each whole-tree audit tool (ruff, bandit, semgrep, trivy, mypy), so a full-project sweep no longer drowns real findings in build output or downloaded media. Scoped sweeps (an explicit file list) are unaffected.
codebase_index + project_settings no longer index a project's committed virtualenv (ANTS-3393) On a flat-layout project with a committed venv/, codebase_index (under source_roots=["."]) and project_settings op:detect both treated the vendored virtualenv as the project's source — indexing/suggesting thousands of library files instead of the ~10 real ones. The shared isNoiseDir skip-list now covers bare venv/env/pycache (.venv was already skipped), and codebase_index's walk applies the same skip-list, so vendored / build / virtualenv trees are pruned even when "." is an explicit source root.
roadmap_log op:flip locates GFM bullets by the canonical headline roadmap_query reports (ANTS-3378) Flipping a GFM task-list bullet by headline now matches the de-marked-up bold span / post-em-dash prose / bold-ID label (the token roadmap_query echoes), not just the raw stored head — so the exact title copied from roadmap_query no longer returns bullet_not_found. On a genuine miss, suggestions rank by token overlap against the canonical headline and carry the bullet line.
Audit summaries no longer rank a scanner's progress-bar text as the top finding. (ANTS-3372) Junk lines scraped from tool output (e.g. a "Working... 100%" progress bar) are now filtered out of every audit-summary parser before findings are ranked, so a real issue lands at the top instead of terminal noise.
Ants MCP read verbs no longer return doctored bytes for frame-sensitive source (ANTS-2218)
read_region / read_regions / workspace_search now accept raw:true,
returning file content byte-for-byte inside an unforgeable nonce frame
instead of the default lossy scrub that neutralises literal
</ants_mcp_data> close-tags and <!--/--> comment markers. Without
it, an agent reading a file that itself contains those tokens (the MCP
source, a spec, HTML/markdown with comments) and building an
Edit/apply_edits from the output would corrupt the file. The default
(scrub) framing is unchanged, so the ANTS-1294/1670/1996 breakout
protections are untouched.
MCP audit_run description documents the client transport-cap + last_audit_summary fallback for long sweeps (ANTS-2183) (ANTS-2183)
A full-tree sweep can exceed the MCP client's ~60 s request timer (Claude Code's), surfacing transport: timed out outside the response even though the run completes server-side and writes its SARIF. The verb description now states this and points the caller at last_audit_summary (or the envelope's cache_path) to read the result, mirroring the two-tier-timeout note verify_changes already carries.
MCP audit_run now resolves the project's compile_commands.json for the C/C++ tools (ANTS-2182) (ANTS-2182)
cppcheck full runs drive off --project=<db> (plus the missing-include suppress set as a no-DB fallback), and clazy/clang-tidy resolve their build dir via the new shared AuditEngine::resolveCompileCommands probe (build/, build-fast/, build-asan/, …) instead of a hardcoded build/ path. Fixes the ~2330-finding missingIncludeSystem/namespace-as-C noise flood and clazy's 0-findings result when the compile DB lived in a non-build/ tree — the MCP audit now matches the in-app AuditDialog / ants-audit CLI signal.
Fix -Wshadow in test_roadmap_viewer_archive.cpp — inner QTemporaryDir tmp shadows outer. (ANTS-2176)
A test file declares the same temporary-folder variable name twice (one inside the other), which the compiler warns about. Harmless today, but warnings should be clean — rename the inner one.
cppcheck uninitMemberVarNoCtor residue in tests/ — 10 test-only helper-struct members lack default-member-initialisers. (ANTS-3354)
Added {} default-member-initialisers to 10 cppcheck-flagged members across 6 test-helper structs (BenchResult, Fixture, Entry, PositiveCase, NegativeCase, Cand, Site). Completes the uninitMemberVarNoCtor cleanup started in ANTS-2184 — our code (src/ + tests/) is now fully clean of the warning; only third-party googletest remains. Test-only, no behaviour change.
cppcheck uninitMemberVarNoCtor brace-init hygiene sweep — ~52 helper-struct members in src/ lack a default-member-initializer (e.g. Hit::score, InlineImage::row/col, UrlSpan/PaintTextRun fields, RunOut::started/timedOut/crashed). (ANTS-2184)
Added {} default-member-initialisers to 11 cppcheck-flagged scalar/pointer members across 8 helper structs in 7 src/ files (Marker, Hit, SpecToken, ChangelogBullet, PendingTypedef, Head, Frame, StatusHit). Forecloses the uninitMemberVarNoCtor warning class in src/; no behaviour change (all were assigned before use). Verified: 0 findings post-fix, clean build, 242 touched-subsystem tests pass.
Ants MCP, from DOOM cross-session feedback: read_region/read_regions symbol-mode now returns a symbol's definition rather than its forward declaration; roadmap_log/changelog_log resolve ROADMAP/CHANGELOG from a project subdirectory (walk up to the git repo root) like the read verbs; session_orient's project_settings_suggestion gained a next_step nudge. ANTS-3349 / ANTS-3350 / ANTS-3352. A residual file_outline most-vexing-parse mis-tag (ANTS-3351) remains deferred — its root cause is an undetected enclosing function, same class as ANTS-2159.
Debt-sweep detector false positives: brace-less if/else/for bodies no longer flagged as dead branches; Q_UNUSED of Qt-typedef params and in comments no longer flagged as orphans; shell (test_*.sh/.bash) tests now count toward invariant coverage.
Also removed two duplicate #includes in mainwindow.cpp. Deeper detector-scope issues (shipped_without_commit windowing, missing_inv_test behavioural matching, fixture-dir exclusion, scan output capping) deferred to ANTS-3342..3347.
C-style // and /* comment introducers aren't excluded from lineHasCode's code-detection (ANTS-1270 left them; the #/-- analogue). (ANTS-2230)
The audit code that decides "is this line real code or just a comment?" was taught (in an earlier fix) that # (Python/shell) and -- (Lua) start comments, so findings inside those comments get dropped. But for C/C++ it was never taught the same about // and /* — so a secret or TODO written inside a // comment in a C++ file is still reported as a real finding. This makes the audit cry wolf on commented-out lines.
file_outline registers typedef struct TAG_s { … } ALIAS_t; aggregates (the dominant C struct idiom). (ANTS-2228)
Let Claude pull one C struct definition by name instead of reading the whole header to find it.
roadmap_query section=<slug> misses #### Pass N.M bullets nested under a prose-shaped ###. (ANTS-2225)
When Claude asks for the to-do items in one section, don't silently skip the ones tucked under a mostly-text heading.
find_definition include_body / read_region symbol-mode — brace-match function bodies so a missing outline boundary can't over-read. (ANTS-2224)
Stop Claude's 'show me this function' from accidentally grabbing the next function too.
SessionStart MCP catalog now tells you to call each verb by its full mcp__ants__<name> — the bare names it listed weren't directly callable. (ANTS-2213)
file_outline / read_region now capture C/C++ functions whose parameter list wraps across multiple lines (id-Software / K&R prototypes) — previously dropped. (ANTS-2212)
Roadmap dialog now has a deterministic keyboard tab order through its filters and controls. (ANTS-1350)
Audit summary reports an honest suppressed-finding count for semgrep runs instead of always zero. (ANTS-2123)
Debug Mode category selection is remembered across restarts instead of resetting to off every launch. (ANTS-1863)
Subsystem recent-changes no longer freezes the UI on slow git — the per-file git calls are now one batched invocation. (ANTS-1340)
Per-tab image memory budget no longer double-counts shared image buffers. (ANTS-1265)
A Kitty T stores an image in m_kittyImages and pushes a COW copy into m_inlineImages; recomputeImageBudget summed both, over-counting real RAM ~2x and rejecting legitimate later transfers. It now dedups by QImage::constBits() so a shared buffer counts once.
Audit path rules now match findings reported with absolute paths. (ANTS-1271) applyPathRules matched project-relative globs against Finding::file, so absolute-path findings (clang-tidy/semgrep/mypy) silently no-op'd the rules. Finding::file is now normalised to project-relative before matching.
Audit pipeline: comment/string detection is now language-aware (#, Lua --/[[ ]], Python triple-quotes). (ANTS-1270)
lineIsCode applied a C-style lexer to every file, so # TODO: "10.0.0.1" in a Python/shell file scanned as code and its IP/secret/TODO finding survived. The lexer now dispatches on file extension and the bare comment-introducer no longer reads as code.
Roadmap dialog: the document preamble (intro before the first section) now shows on the Full preset. (ANTS-1275)
renderCardsHtml gated preamble prose on sectionExpanded, which is still false before the first ## heading, silently dropping the document intro. The collapse gate now applies only to section-intro prose.
Settings reports Claude hooks as installed only when all five hook events are present (previously it could show green with two missing); failures to lock down the Claude settings-file permissions are now logged instead of ignored. (ANTS-2205)
With two Claude tabs open, a permission prompt arriving during the brief tab-switch window is no longer attributed to the wrong tab. (ANTS-2190)
Audit history file is now guarded against two Ants/Claude sessions auditing the same project at once (no lost history or orphaned reports). (ANTS-2189)
Wedged project-query workers are now bounded and logged instead of leaking silently (ANTS-2194) A hung query snippet that can't be killed parks its worker thread; past a hard ceiling Ants refuses new background queries (and logs the count) so a repeating hang can't slowly exhaust memory unnoticed.
Shell launch resolves HOME before forking for signal-safety (ANTS-2204) The child's working-directory fallback no longer calls getenv() in the fork/exec window; HOME is captured pre-fork like the other launch parameters.
Remote-control socket gained the same re-entrancy guard as the MCP socket (ANTS-2202) A latch + consumed-line clear prevent a command from being dispatched twice should a future verb pump a nested event loop. Also corrected a misleading code comment.
Fold-in verbs no longer burn roadmap ID numbers when there's nothing to write under (ANTS-2201) indie_review_fold_in / debt_sweep_defer / cold_eyes_fold_in now check for an active release heading before reserving IDs, so a missing heading returns a clean no_release_heading refusal instead of consuming counter values into a permanent gap.
Model-switch banner detection hardened against a quoted banner in scrollback (ANTS-2197) directModelSwitchVisible now requires the tier title and the banner's 'saved as your default' tail together, so a bare quoted fragment no longer trips a continuation.
Model-switch ledger now hard-bounded even when outcomes never settle (ANTS-2196) A secondary ceiling (4x the soft cap) drops the oldest record regardless of pending state, so a run of never-settling pending outcomes can't grow the file unbounded.
Model auto-switcher: a /model typed while idle no longer starts an unrequested billable turn (ANTS-2186) The direct-switch continuation now gates on an active turn, matching every sibling path. Pre-picking a model at idle is not consent to resume work.
Refuse truncated DCS/APC image payloads; correct the AI command-preview byte warning. A DCS (Sixel) or APC (Kitty/iTerm2) payload that overflowed the parser's 10 MiB accumulator is now refused outright instead of being decoded from a corrupt prefix (ANTS-1663 was previously honoured only for OSC 8/52). The AI command-confirmation dialog's "N additional bytes will be executed" warning is now accurate (and never negative) for commands containing HTML-special characters.
roadmap_log no longer reissues a live roadmap ID when .roadmap-counter lags the file. Both op:append and op:append_batch now reconcile the counter against the highest existing [PREFIX-NNNN] id, skip past it, self-heal the counter, and surface counter_advanced_to; an id_hint that collides with a live id is refused id_taken instead of overwriting a bullet. (ANTS-2179)
Flatpak installs now reach Ants MCP from Claude sessions (ANTS-1900) (ANTS-1900)
Inside a Flatpak sandbox the user's shell runs on the host via
flatpak-spawn --host, which does not inherit the caller's
environment — so the ANTS_MCP_SOCKET path that lets a Claude Code
session find Ants's MCP helper never crossed the boundary. The
SessionStart cheat-sheet (and every socket-gated MCP affordance) was
therefore silent in Flatpak installs. src/ptyhandler.cpp now
forwards --env=ANTS_MCP_SOCKET=<path> on the flatpak-spawn argv,
gated on a non-empty value (omitted when MCP is disabled). Plain
terms: install Ants Terminal as a Flatpak and Claude can now use the
cheaper Ants tools just like a native install.
concurrent_writer_lock predictable /tmp/ants-cwl-<pid>-<time>.dat + symlink-attack exposure. (ANTS-1380)
concurrent_writer_lock test now roots its lock file in a private QTemporaryDir (0700, random name) instead of a predictable /tmp/ants-cwl--.dat, with O_NOFOLLOW on the lock open — closes a symlink-attack window on shared /tmp. Test-only; production ConfigWriteLock (user-private config dir) was never exposed.
--remote client now has an overall read deadline (slow-loris guard) (ANTS-1671).
The per-iteration 2 s timeout on the --remote client read loop reset every iteration, so a hostile same-machine peer that won $ANTS_REMOTE_SOCKET could dribble bytes under the 1 MiB cap and hang the client forever. A 10 s cumulative deadline now bounds the total wait.
Untrusted project path from a Claude transcript is now validated before use (ANTS-1670).
The cwd read from a (attacker-influenceable) .jsonl transcript is gated through a new absolute-path / no-..-traversal / no-NUL check before it becomes the working directory used to launch or resume a claude process; a bad value falls back to the safe path decoder.
MCP data-wrapper now neutralises a spoofed open tag, not just the close tag (ANTS-1670).
wrapMcpData already scrubbed a literal </ants_mcp_data> (and case/whitespace variants) from payloads; it now scrubs the symmetric <ants_mcp_data …> open form too, so attacker-controlled scrollback / commit text can no longer spoof a nested wrapper-open and desync the frame a consuming model sees.
Plugin loader refuses a second plugin that resolves to an already-loaded name, preventing a silent overwrite via symlinked directories. (ANTS-1370)
Flatpak shell launch now runs flatpak-spawn with the same pre-fork sanitised environment as the direct path (consistent terminal-variable handling across the sandbox boundary). (ANTS-1994)
workspace_search rejects a leading ! in glob (negation that resurrects ignored trees). (ANTS-1274)
ripgrep treats a !-prefixed --glob as a negation whose precedence is above .gitignore, so !.git/!node_modules/** could un-exclude an ignored directory. The glob validator now rejects a leading ! (the only resurrection vector) alongside the existing ../256-byte guards; legitimate brace/charclass globs are unaffected.
The "which Claude session is current" check no longer trusts a file's modified-time when a live Claude process is known, closing a same-user spoofing gap. (ANTS-2191)
project_query's file listing no longer discloses names of files symlinked outside the project (ANTS-2203) project.list now re-validates each entry against the project root, so a symlink whose target escapes is skipped entirely — matching project.read, which already refuses to read it.
debt_sweep now rejects a malformed since revision instead of passing it to git (ANTS-2199)
The user-supplied git ref is shape-checked before it reaches the git diff argv, giving a clean error instead of a confusing mid-scan git failure.
Model auto-switcher: code-level parked guard so a config bug can't re-arm keystroke injection (ANTS-2195) The auto-switcher's only actuator is keystroke injection and is deliberately parked as unsafe. A new compile-time guard (kAutoSwitchActuatorParked) short-circuits the firing site before the /model send, so the parked decision can no longer be undone by config alone.
Terminal query-responses are now rate-limited to bound a \e[6n-flood amplification (ANTS-2192).
DA/CPR/DSR/DECRQSS/colour/Kitty query replies fired an unconditional
PTY write per request, so a program streaming a query in a tight loop
forced unbounded response writes. All replies now funnel through
TerminalGrid::sendQueryResponse(), a rolling 1 s cap (256/sec) mirroring
the OSC 52 quota — a flood is bounded while a legitimate synchronous
query is never throttled. Response content was already fixed/derived;
this bounds volume only. Regression-locked by
query_response_ratelimit (INV-1/2/3).
audit_run scoped positionals are now guarded against argv option-injection (ANTS-2185).
A scoped file path beginning with - (e.g. a hostile-clone file named
-rf.cpp) was appended bare to each tool's argv and parsed as a flag.
toolArgv now ./-prefixes dash-leading relative paths at a single
chokepoint so they are unambiguously paths; absolute and ordinary
relative paths are unchanged (the since-last-run delta keeps matching
them). -- was rejected because ruff/bandit/shellcheck/mypy append
flags after the path list. Regression-locked by mcp_audit_run INV-18
(behavioural + wiring).
audit_run now secret-scrubs raw tool output before it reaches disk or the LLM (ANTS-2188).
trivy's --scanners secret surfaces literal secret values; unlike
gitleaks (which runs --redact) trivy's output was embedded verbatim
into the .audit_cache/*.sarif notification text and returned over MCP
via top_findings. Each tool's output is now run through
SecretRedact::scrub at the single capture point in the runner, before
it feeds either sink — closing both the on-disk artifact and the LLM
envelope (OWASP LLM06). Regression-locked by mcp_audit_run INV-17.
Harden several input boundaries (indie-review #8 sweep).
Secret-redaction now covers GitLab personal-access tokens (glpat-) and Slack app-level tokens (xapp-) before context is sent to an LLM. The repo-visibility gh repo view call now passes a -- end-of-options sentinel so a hostile clone's --prefixed origin slug can't be parsed as a flag, and the roadmap-card HTML escaper now escapes " so a quote in a ROADMAP bold-ID can't break out of an HTML attribute (CWE-79).
0.7.98 RC1 — Patron preview
0.7.98 RC1 — Patron preview (public ship target: 2026-07-09)
A large Ants-MCP / Claude-integration release. Highlights:
project_query — Claude runs read-only Lua sandboxed over your project, returning just the result (ANTS-2093)dry_run preview across all 9 mutating MCP verbs (ANTS-2227)audit_run async:true + audit_poll for slow sweeps past the transport timeout (ANTS-3396)0.7.97 — MCP integration briefs, project auto-detect, work-bundles
A large MCP / Claude-integration milestone. Ants now answers "where do I hook into this pipeline?" (read_region call_sequence) and "show me the canonical idiom" (similar_code include_bodies) in a single call, auto-detects projects whose code isn't under src/ and offers to write their settings file, groups related to-dos into thematic work-bundles, and keeps its most-used tools instantly callable so long sessions don't drift back to raw grep.
Highlights
read_region call_sequence:true returns the ordered steps inside a function with line anchors + the accessors a new step needs; similar_code include_bodies:true returns full matching definitions so you copy a canonical idiom in one call. (ANTS-2157 / ANTS-2156)src/ and offers to write its .ants/project.json; new project_settings verb creates/updates it. (ANTS-2160 / ANTS-2161)roadmap_query mode:"bundles" groups active to-dos by theme, flagging blocked items and shipped siblings. (ANTS-1922)file_outline stops mislabelling locals and case labels as functions and now finds old-C-style definitions; plus a sweep of related fixes. (ANTS-2159 et al.)0.7.97 RC1 — Patron preview
Theme: Rolling Patron preview of the next release. Fixes and features land in this section as they ship during the freeze window; the section gets its public date when 0.7.97 is promoted to a stable release.
0.7.96 — MCP token-frugality, teardown-crash sweep, security hardening
Theme: Ants-MCP token-frugality and roadmap / read-verb expansion, a sweep of teardown / use-after-free crash fixes, and indie-review security hardening.
Added
read_region, apply_edits, project-wide codebase_index (ANTS-2021 / 2022 / 1637)compact:true (ANTS-2085 / 2091)roadmap_log dry-run / custom stable IDs / near-duplicate advisory; changelog_log batch op; server build-identity in session_orient (ANTS-2077 / 2078 / 2043 / 2044 / 2073)Fixed — crash & stability sweep
Security — indie-review hardening
See CHANGELOG.md for the full list (~85 entries). The x86_64 AppImage attaches automatically once the release workflow finishes building.
0.7.95
Settings-dialog & menu usability fixes — readable fields at any window size, no surprise dialog closes, and menus that stay open while you toggle checkboxes (ANTS-1980/1981/1982).
0.7.94
Auto-switcher refinement + MCP feedback/spec tooling — chooseable home tier, smarter safe-downgrade signals (ANTS-1974 + ANTS-1970..1973 + ANTS-1961/1962/1963).
Settings: user-selectable default/home model tier — auto-switcher returns to it and only downgrades on clear evidence. (ANTS-1974) Let the user pick their home model (e.g. Opus) in Settings. The auto-switcher then treats that as the baseline — staying there for normal work and only stepping down to Sonnet/Haiku when the work is clearly mechanical/cheap.
Yield the composer_not_empty veto during a long-ToolUse foreground wait (the 67%-dominant near-miss blocker). (ANTS-1973) During a long build/test wait the text box often holds stale queued text, not active typing — yet that blocks the cheaper-model switch 67% of the time. When a foreground command has been running a while and the box hasn't changed, treat the text as abandoned and allow the downgrade.
Conditionally surface feedback_log / feedback_query in SessionStart orientation when a *_Ants_MCP_Feedback.md exists. (ANTS-1971)
When a project has a cross-session feedback file, the startup hint should mention the two tools for reading/writing it — otherwise contributors fall back to hand-editing the file (which risks format drift).
Regret signal over-counts: any non-auto-authored /model is scored as a user-override regret, even when it's an independe… (ANTS-1957) The "regret rate" that decides whether to trust the auto-switcher counts every manual model switch you make after it acts as the switcher being wrong — even when you switched for your own reasons (testing, or picking a model for a specific task). So the feature looks worse than it is, especially during sessions where you're switching models a lot by hand.
spec_log MCP write verb — append cold-eyes loop entries / flip Status / append INVs on docs/specs/*.md. (ANTS-1963)
Lets Claude update a spec file's status and review-log through Ants MCP instead of expensive hand-edits.
feedback_log MCP verb — write to a *_Ants_MCP_Feedback.md file (contributor append + maintainer tracking block). (ANTS-1962)
Lets any CC session add feedback to the shared report files, and lets the maintainer session stamp a tracking block with roadmap IDs, without manually editing markdown.
feedback_query MCP verb — read the un-triaged tail of a *_Ants_MCP_Feedback.md file. (ANTS-1961)
Lets the Ants maintainer session pull just the new feedback from a cross-session report file instead of re-reading the whole thing every week.
Auto-switcher task-shape signals round 2 — foreground-subprocess-wait (safe-downgrade) + post-clarification suppression … (ANTS-1959) Make the model-switcher smarter about WHEN it's safe to drop to a cheaper model: it's safe while waiting on a long build/test, but risky right after you answer a design question (hard work usually follows). Two cheap signals that catch the most common real sessions.
model_switch_stats surfacing/scoring polish — suppress calibration regret_rate, credit clean idle-end downgrades, do… (ANTS-1960)
Three small honesty fixes to the model-switcher's stats: don't show a scary "50% regret" when it's based on one event; give credit when it correctly wanted a cheaper model right at a clean session end; and note that two of its counters only match when read at the same instant.
Auto-switcher task-shape signals round 2 — foreground-subprocess-wait (safe-downgrade) + post-clarification suppression … (ANTS-1959) Make the model-switcher smarter about WHEN it's safe to drop to a cheaper model: it's safe while waiting on a long build/test, but risky right after you answer a design question (hard work usually follows). Two cheap signals that catch the most common real sessions.
Near-miss telemetry needs the ANTS-1941 epoch boundary too — dominant_blocker stays contaminated by stale-binary records… (ANTS-1954) When you rebuild Ants, the auto-switcher's "why didn't it fire?" stats keep showing yesterday's reasons for about a day, because they don't reset at the rebuild boundary like the other stats do. That's exactly why we can't yet judge whether the switcher's remaining block is real. This makes those stats reset cleanly at each rebuild.
User-typed /model: inject a continuation prompt after auto-confirming the switch dialog. (ANTS-1953) When you type /model sonnet yourself, Ants now auto-clicks the confirmation dialog — but it still leaves you at a blank prompt instead of typing "please continue" to resume your work. This would add that last step.
Surface the MCP server's build identity (git SHA + build time) so callers can detect a ship-vs-live binary gap. (ANTS-1952) When a fix is committed but the running Ants build is older, the MCP tools report stale data and we waste time re-investigating. Expose the build's git SHA and build timestamp so a Claude session can instantly tell the server predates a fix.
Auto-confirm the 'Switch model?' prompt for user-typed /model commands too. (ANTS-1951) When you manually type /model to switch models, you still get a "are you sure?" popup — even though you just typed the command. This would make that popup automatic too.
find_definition could hint when a query matches a file stem but no symbol. (ANTS-1950)
If you ask 'where is X defined' and X is actually a filename (not a function), the tool returns nothing — a one-line 'did you mean the file X.cpp?' hint would save a follow-up.
roadmap_log op:flip should accept common status synonyms ("done"→shipped, "wip"→in-progress, "todo"→planned). (ANTS-1932)
When marking a roadmap item finished, the tool only accepts the word "shipped", not the more obvious "done" — so the first try fails and has to be redone. Teach it to accept the common everyday words too.
Trust-signal needs a fix-epoch boundary, not just an age window — pre-fix records poison regret_rate even when only days… (ANTS-1941) The auto-switcher scorecard still reads ~70% regret, but that is leftover noise from the buggy May-29 sessions, not the fixed behaviour. The 30-day window can't filter it out because the bad records are only a few days old. We need to mark a 'fixed from here' line so the scorecard only counts switches made by the current, fixed logic.
Auto-switcher calibration robustness — regret-driven conservatism + task-shape downgrade signal. (ANTS-1940) Now that the auto-switcher actually fires, the first real data shows it guesses wrong about half the time (on a tiny sample). Make it more cautious while it's still learning, and make it notice fiddly mechanical work (safe to use a cheaper model) versus deep reasoning (don't).
Extend the ANTS-1908 human-idle soft-veto to the focused_state_not_idle gate (now the dominant near-miss blocker). (ANTS-1939)
The auto-switcher won't change models while a Claude session looks "busy" — but during a long hands-off task, "busy" means the agent is grinding away, which is exactly when a cheaper model would save the most. Let it switch when YOU haven't typed in a while, even if the agent is still working.
model_switch_stats firing/regret aggregation needs a recency window — stale pre-fix records poison the trust signal in… (ANTS-1936)
The model-switch report mixes in old data from before the May-29 bug fixes, so it keeps showing a high "regret" number even though the one switch since the fix went fine. Let the report focus on recent switches so it reflects how the feature behaves NOW.
model_switch_stats should split upgrades/downgrades by trigger (auto vs manual) — current totals can't answer "is the … (ANTS-1934)
The model-switch report mixes together the switches YOU make by hand and the ones the app makes automatically
Auto-switcher tracks pending-switch intent when blocked by composer_not_empty (ANTS-1919) When all guards pass but the composer has recent text, the target tier\nis stored in m_autoSwitchPendingTier so the switch fires on the next\n2 s tick once the composer clears. Status-bar visual annotation TBD.
session_orient now includes top-20 active roadmap bullets inline (ANTS-1922)
Adds an active_bullets key (headline_only, status:active, limit:20)\nto the session_orient response envelope. Sessions can now identify the\nnext work bundle in one call without a follow-up roadmap_query.\nToken budget bumped to 17 000; does not affect allOk.
workspace_search adds max_match_bytes: + headline_only:true knobs — first-attempt tax on research sweeps with long… (ANTS-1904)
When searching for threading keywords across a big roadmap, the response is dominated by 2-5 KB bullets each. Two new knobs (clip per-match bytes, or skip context entirely) would let the search return manageable results on the first try instead of forcing a regex-narrowing retry.
SessionStart orientation cheat-sheet: add roadmap_query + workspace_search + the verb-modes / ETag-and-fields footno… (ANTS-1910)
The 9-tool cheat-sheet at session start is helpful but misses three things I reach for every session. Add roadmap_query (the read-side partner to roadmap_log), workspace_search (the cheap-grep), and a one-line footnote noting that most write verbs have multiple op: modes and most read verbs accept etag_match + fields=.
spec_query accepts project-configurable doc-path layouts (e.g. docs/phases/phase_<NN>_<topic>_design.md). (ANTS-1906)
spec_query only finds specs at docs/specs/ANTS-NNNN.md — but Vestige's design docs live at docs/phases/phase_NN_topic_design.md. Either re-use the project_layout probe-set discovery, or accept path_prefix + id_pattern args.
roadmap_log op:append auto-detect stable-string-ID project shape — empty .roadmap-counter should default to `id_stra… (ANTS-1905)
Some projects use stable string IDs (Ts20-FL1) instead of numeric ones (ANTS-1234). On those projects the append tool tries to allocate from an empty counter and produces wrong IDs. Auto-detect the project's existing style and default to the stable-prefix escape hatch.
roadmap_query per-section etag — let section=X short-circuit when only other sections changed. (ANTS-1907)
Today reading the same section twice always re-parses if any byte of the file changed. A per-section ETag would let unchanged sections return 'not modified' independently — saves token cost across /cold-eyes and /test-audit loops that touch many sections.
Auto-switcher: enrich model_switch_stats headline with calibration progress + near-miss dominant-blocker. (ANTS-1909)
When the auto-switcher hasn't fired yet, the headline says 'no switches yet' — which sounds broken. Rewrite the headline to also surface 'N near-misses blocked by composer_not_empty' so it reads as 'evaluating but blocked' instead of 'doing nothing'.
Auto-switcher: composer_not_empty soft-veto with stale-text detection — unblock long autonomous sessions where the dom… (ANTS-1908)
The auto-switcher won't fire while there's any text in the input box — but in long autonomous sessions there's always leftover text. Make the rule smarter: if the text has been sitting untouched for 5+ minutes, treat it as stale and allow the switch.
Direct /model <tier> (explicit arg, no dialog) doesn't resume work — ANTS-1969 only covers the menu/dialog path. (ANTS-1975)
When you type /model sonnet (with the model name) Ants doesn't resume the task — it goes idle. The fix from before only handles the pop-up menu version of /model, not the direct one.
Doc note: firings near_misses.total_24h excludes idle-end-suppressed; +idle_end_suppressed_24h reconciles to near_miss… (ANTS-1972)
Two of the switcher's counters look like they disagree (18 vs 21) but don't — one excludes idle-end-suppressed near-misses. A one-line doc note stops the next session re-filing it as a bug.
User-typed /model does not continue work after auto-confirm (ANTS-1958 revisit). (ANTS-1969) When you manually type /model X and Ants auto-confirms the dialog, Claude Code goes idle rather than continuing work — you have to send a new message. When auto mode is on, a continuation prompt should fire just like after an auto-switch.
Pending-switch chip annotation (→ Tier) reads as a recommendation rather than an automatic queued action. (ANTS-1956)
The model chip shows "Haiku (→ Sonnet)" when a switch is queued, but users interpret it as "Ants is suggesting you switch" rather than "Ants will switch automatically once you send your message." Rewording or a tooltip would make the intent clear.
model_switch_stats: document by_blocked_by gate-hit counting and /model ledger semantics (ANTS-1947 + ANTS-1949) Two clarifications in the MCP tool descriptor. (1) by_blocked_by\n values can sum to more than window_24h.total — a single near-miss\n blocked by several gates contributes one count per gate. (2)\n by_trigger.manual is structurally 0: only the autonomous switcher\n writes ledger records; user /model commands and chip-clicks do not.
Auto-switcher pending-switch visual indicator — model chip annotation when switch queued on composer_not_empty. (ANTS-1926) When the auto-switcher wants to change model but you have text in the input box, show a subtle 'pending → haiku' label on the model chip so you know a switch is waiting.
ModelRecommender::score() mtime cache — avoid re-parsing 512 KB transcript every 2 s tick. (ANTS-1927) The auto-switcher re-reads up to 512 KB of conversation history every 2 seconds even when nothing has changed
Colored cell backgrounds (diff green/red highlights) appear more opaque than the base when terminal opacity < 1. (ANTS-1864) When you make the terminal see-through, the green/red highlight bars (like Claude's diff view) stay nearly solid instead of going see-through to match the rest of the window. Make them respect the same transparency level.
Drop deprecated cwd field from session_memory schema in 0.7.93. (ANTS-1420)
clean up the placeholder schema entry left behind
mcp-review-engines H1 — Ledger load per brief assembly is re-parsed N times for N lanes. (ANTS-1672)
mcp-review-engines M3 — extractCitedCodePaths slurps every cited doc body unbounded; ROADMAP.md (~600 KB today) is rea… (ANTS-1674)
--preset=fast is functionally identical to --preset=default — give it a real differentiator. (ANTS-1556)
The "fast build" mode doesn't actually build faster than the normal mode — they have the same settings. Need to make it really faster, or drop it.
--preset=fast differentiator: bump link_pool from 1 → 2 for parallel test-bundle linking. (ANTS-1558)
Let the "fast build" mode link two programs at the same time instead of one — should make full rebuilds a few minutes faster.
project_layout roadmap-format sniffer still returns unknown on Vestige despite the ANTS-1632 ship claim — re-investi… (ANTS-1903)
Vestige reports that the format-detector for ROADMAP files is still returning 'unknown' on their project despite the previous fix being marked shipped. Either the fix isn't reaching their build, or it doesn't cover their specific roadmap shape. Needs a trace field + CI fixture.
SessionStart orientation prelude exceeds its 1200-byte cap (test red). (ANTS-1970) The MCP cheat-sheet that prints when a Claude session starts grew past its size limit as new tools were added, so a test that guards the limit is failing. Either trim the cheat-sheet or raise the limit.
/model slash command double-posts in the transcript (command + confirmation echoed twice from one invocation). (ANTS-1958)
When you type /model to switch models, the command and its "model set" confirmation sometimes show up twice from a single press. Harmless but messy — this de-duplicates it.
Auto-confirm user-typed /model dialog now catches it within ~120 ms instead of up to 2 s (ANTS-1955) The 2-s status tick was the sole trigger, so a dialog rendering just after a tick fire was missed for nearly 2 full seconds. Added a short polling burst (mirrors the auto-switch handshake path) that fires every 120 ms for up to 1.9 s after each tick, ensuring the confirm lands almost immediately after the dialog appears.
Near-miss telemetry no longer records idle steady-state as a blocked switch (ANTS-1946) When the auto-switcher's recommended model already matches the current\n model, it never wanted to switch — so that tick is no longer logged as\n a "blocked" event. Previously these idle records dominated the\n "what's blocking it?" counter, making composer-typing and cool-down\n look less significant than they are. The fix removes ~36% ledger noise\n and clears the path for accurate dominant-blocker reporting.
Auto-switcher thrash at context-compression boundary — fires /model to the model already set when the transcript's /mode… (ANTS-1944) After the conversation fills up and gets summarised, the auto-switcher loses track of which model was already set and keeps issuing the same switch command over and over (8 times in one observed session).
model_switch_stats MCP verb relies on StatsConfig struct-default for windowDays while the controller sets it explicitly … (ANTS-1942) Two parts of the auto-switcher read the same scorecard but configure the 'recent window' differently — they agree today only by luck. If the default ever changes, the human-facing number and the switcher's own caution dial would silently disagree.
Model-switch actuator: confirm by watching PTY output, not a blind 250 ms timer. (ANTS-1920) The auto-switcher sends "/model opus" then blindly waits a quarter-second and presses "1" to confirm. If the confirmation box is slow, or you have a message queued, the "1" lands in the wrong place and the switch gets stuck (observed by the user 2026-05-29 — model stayed on Sonnet, "/model opus" stranded in the composer). Watch for the actual "Switch model?" prompt before answering it.
model_switch_stats 24h near-miss count + dominant-blocker disagree between firings and near_misses modes. (ANTS-1938)
The model-switcher's "what's blocking it today" number is different depending on which report you ask for — one says 2, the other says 22, and they even name different top blockers. Make them agree.
roadmap_log op:flip/flip_batch can't resolve composite IDs containing comma+space (read-path roadmap_query id= can… (ANTS-1937)
When a roadmap item's ID has a comma in it, the "mark it done" tool can't find it — even though the search tool finds it fine. Make the two agree.
Inline graphics: back-to-back Sixel/Kitty/OSC sequences no longer lose a byte (ANTS-1777) (ANTS-1777) Terminals that stream pictures in back-to-back chunks (Sixel, Kitty graphics) dropped a byte between two chunks, rendering the second as garbage. The VT parser now re-arms a new string sequence when the byte after a string-terminating ESC is a string introducer (] P _ X ^), while still discarding the dangerous ESC-finals (c=RIS, etc.) and the CSI introducer — so the 0.7.53 hardening against hostile terminal-reset injection is fully preserved.
changelog_log op:add_from_roadmap drops the Layman line from the CHANGELOG body — regression, failing test. (ANTS-1933)
The shortcut that copies a roadmap item into the changelog is currently leaving out the plain-English summary line — so changelog entries created that way are missing their friendly description
Auto-switcher stable-counter reset-hysteresis reduces false ticks_target_stable_insufficient blocks (ANTS-1925) A single score-boundary noise tick no longer wipes the stability counter.\nRequires kStableResetTicks (=2) consecutive target==current ticks before\nresetting, matching the same 4 s window as the forward stability gate.
ANTS-1924: model-switch PTY handshake — ESC + ENTER confirmation sequence + continuation prompt. (ANTS-1924) When Ants switches Claude models, it now sends the right keystrokes to confirm the switch dialog (ESC then Enter), then automatically sends "please continue" so Claude picks up where it left off — no manual typing needed.
Review Changes button doesn't appear when Claude only writes new files (no insertions/deletions on tracked files). (ANTS-1874) When Claude creates a brand-new file, the bottom-bar "Review Changes" button stays hidden — it only lights up if Claude edits an existing file. So new-file changes have no quick way to be reviewed.
Strip the "ccache + PCH" framing from --preset=fast docs. (ANTS-1557)
Update the "fast build" mode's description to match what it actually does (just an isolated build folder).
/model actuator sent \n not \r — text landed in composer, never submitted (auto-switch dead loop). (ANTS-1912)
When the auto-switcher fired /model sonnet, the text appeared in the Claude composer waiting for the user to press Enter — it never actually submitted. Single-char fix: send CR (0x0D, what the Enter key produces per terminalwidget.cpp:1930) instead of LF (0x0A).
Cross-tab Claude-state mix-up — focused tab's status bar shows ANOTHER tab's Claude: state. (ANTS-1911)
When you have two Claude tabs open (e.g. Ants Terminal + Vestige), the status bar at the bottom sometimes shows the WRONG tab's state. Screenshots 2026-05-28: the focused Ants Terminal tab was actively running cold-eyes loops + 3 background agents but the status bar said 'Claude: idle' and the tab dot was idle; switching to the Vestige tab (which was actually idle) showed 'Claude: thinking' — the state of the OTHER tab leaked into this one's status surface.
0.7.94 RC1 — Patron preview
This release shipped without written notes.
0.7.93
Theme: Auto-switcher visibility — the silent /model injection
finally announces itself + the operator-trust ledger
(ANTS-1893 surfacing + ANTS-1894 near-miss telemetry +
ANTS-1891 honest headline + ANTS-1897 MCP cheat-sheet).
0.7.93 RC3 — Patron preview
Theme: Auto-switcher visibility — the silent /model injection
finally announces itself + the operator-trust ledger
(ANTS-1893 surfacing + ANTS-1894 near-miss telemetry +
ANTS-1891 honest headline + ANTS-1897 MCP cheat-sheet).
0.7.93 RC2 — Patron preview
Respin of v0.7.93-rc1 with cherry-picked fixes: c6cc905
0.7.93 RC1 — Patron preview
Theme: Auto-switcher visibility — the silent /model injection
finally announces itself + the operator-trust ledger
(ANTS-1893 surfacing + ANTS-1894 near-miss telemetry +
ANTS-1891 honest headline + ANTS-1897 MCP cheat-sheet).
0.7.92
Theme: MCP token-saver depth + frozen-RC release pipeline. Wires
the 0.7.92 milestone (pulls 34–42 MCP/test-audit/indie-review fold-ins)
and bootstraps the weekly-Wednesday + Patron-RC cadence. Public ship
target 2026-05-27; the inaugural v0.7.92-rc1 is cut from these bits.
0.7.92 RC1 — Patron preview
Theme: MCP token-saver depth + frozen-RC release pipeline. Wires
the 0.7.92 milestone (pulls 34–42 MCP/test-audit/indie-review fold-ins)
and bootstraps the weekly-Wednesday + Patron-RC cadence. Public ship
target 2026-05-27; the inaugural v0.7.92-rc1 is cut from these bits.
0.7.91 — indie-review fold-in (3 CRIT + 18 HIGH inline; 30 follow-ups roadmapped)
Theme: indie-review fold-in. The 2026-05-13 multi-agent /audit
vtparser — 8-bit C1 controls now routed. Raw 0x80-0x9F bytes
in the PTY stream were being silently mis-decoded as malformed
UTF-8 → U+FFFD, which made the ST (0x9C) terminator branches in
OscString / DcsString / ApcString / IgnoreString dead code.
ECMA-48 § 5.3 / Williams VT500 state machine treat C1 as
first-class controls. Fixed at vtparser.cpp:feedByte — bytes in
[0x80, 0x9F] route through processChar(byte) directly, plus a
Ground-state dispatch table for 0x9B (CSI), 0x9D (OSC), 0x90 (DCS),
0x9F (APC), 0x98/0x9E (Ignore). Other C1 emit Execute.
lua-plugins — heap cap silently bypassed. Lua 5.4 manual:
when ptr == NULL, the allocator's osize parameter encodes the
object type (0..8 small int), not byte count. The custom
allocator treated it as a byte count, drifting m_luaMemUsage
downward unboundedly on every fresh allocation — a plugin could
silently exceed the documented 10 MB cap. Fixed at
luaengine.cpp:luaAlloc (zero osize when ptr == nullptr).
roadmapdialog — HTML injection (CWE-79). rec.kind was
emitted unescaped into the rendered card HTML; rxKind admits
<, >, &, ", so a hostile ROADMAP.md bullet like
Kind: <img src=x onerror=…> rendered as raw HTML in
QTextBrowser. Fixed at roadmapdialog.cpp:1337 —
htmlEscape(rec.kind) before emission.
claudeintegration — cold-start hook gate tightened. During the
1–3 s window after a tab switch (while m_transcriptPath was
empty), foreign-tab PreToolUse / PostToolUse / Stop events
passed the focus gate and mutated singleton state — exactly the
ANTS-1161 symptom returning transiently. Now only SessionStart
PermissionRequest land while the path is empty; other
state-mutating hooks drop with a diagnostic log
(DebugLog::Claude).ptyhandler — child signal mask + dispositions reset before
exec. Qt / glib / dbus install handlers for SIGCHLD / SIGPIPE /
SIGUSR2 etc. POSIX § 2.4 keeps SIGPIPE=SIG_IGN across exec, so
child shells inherited the parent's mask and pipelines silently
hung. ~Pty also now sends SIGCONT alongside SIGTERM so a
stopped child doesn't burn the full 500 ms escalation budget.
remotecontrol — cwd validation on cmdLaunch + cmdNewTab
(CWE-22 parity). Other path-bearing verbs (workspace-search,
file-outline, git_state) validated their path parameter
against control-bytes / backslashes; launch and new-tab
did not. Same-UID model holds today, but the verb is the seam
Claude Code (and any future MCP) crosses with user-supplied
input — validate at the boundary, not the kernel.
remotecontrol — getsockopt(SO_PEERCRED) len check. On
getsockopt failure with a zero-initialised cred, the
disconnect log read "peer UID 0" — a phantom "root tried to
connect" alarm. Now checks len == sizeof(cred) before reading
cred.uid.
antshelper — --repo-root traversal hardening. Rejects
.. substrings + NUL bytes; canonicalises before script lookup
so a planted packaging/check-version-drift.sh at an attacker-
controlled path can't pose as a project root.
lua-plugins — init.lua symlink rejection. Plugin directory
was already canonicalised; the init.lua file inside wasn't
symlink-checked. A symlink to /etc/passwd or ~/.ssh/id_rsa
would surface the file's first line in lua_tostring(err) when
the parser failed. Rejected at pluginmanager.cpp:202 with a
qWarning.
terminalgrid — Kitty chunk buffer flush on alt-screen + resize.
An attacker m=1 chunked image transfer + alt-screen flip → stale
bytes prepended to the next legitimate APC m=0 from a different
sender → poisoned image attributed to new sender. Fixed at the
alt-enter site + at the top of TerminalGrid::resize.
terminalgrid — OSC 8 id= length cap (256 bytes). URI was
capped at MAX_OSC8_URI_BYTES = 2048; the id was uncapped. With
the parser's 10 MiB per-OSC cap, a hostile id=<10MB> sequence
scaled to ~800 MiB across 80 rows of scrollback.
terminalgrid — Sixel first-pass payload cap (4 MiB). The vt parser caps DCS bodies at 10 MiB, but Sixel's first-pass iterator walks every byte to compute width/height with no cycle budget. A 10 MiB payload of valid sixel bytes pinned the parse thread for ~10 M iterations.
claudebgtasks::poll() — watch-loss recovery mirror of
claudetasklist::poll(). Re-adds the QFileSystemWatcher when
the transcript path (re-)appears + mtime-shortcircuits to skip
reparse when nothing changed. Closes the documented CLAUDE.md
claim that both trackers had poll() parity (only foreground
did pre-0.7.91). Wired into
ClaudeStatusBarController::refreshBgTasksButton.
claudebgtasks::parseTranscript — isSidechain +
isCompactSummary filters. Foreground tracker already filtered
these (ANTS-1158 sidechain rationale + ANTS-1224 compact-summary
state reset); bg-tasks did not, so subagent-launched background
tasks silently inflated the parent session's running count, and
post-/compact resume left pre-compact bg entries lingering.
m_errorHideTimer owned member on ClaudeStatusBarController
— replaces the prior QTimer::singleShot pattern in setError.
Re-entry now cancels the prior auto-hide instead of inheriting it
(rapid back-to-back errors no longer flash and disappear). Sticky
errors via autoHideMs <= 0.
Accessibility — setAccessibleName on three previously-silent
status-bar chips. m_bgTasksBtn, m_tasksBtn, m_errorLabel
now expose Orca-readable labels per docs/standards/documentation.md § 7. The remaining a11y gaps (keyboard-nav to expand cards,
aria-expanded on collapse anchors) are tracked as ANTS-1277.
vtparser — kSafeMinSigned / kSafeMaxSigned named constants
for the SSE2 safe-ASCII scan boundary (-96 / -2 were magic
numbers).
Q_ASSERT in handleAsciiPrintRun verifying the
[0x20, 0x7E] precondition the SIMD scanner establishes.
CLAUDE.md — three doc-drift fixes. terminalwidget is
QWidget (has been since 0.7.4), not QOpenGLWidget.
featurecoverage lists 2 in-process lanes + 1 shell-based
(test_health runs via QProcess, not inProcessRunner).
claudetasklist / claudebgtasks bullet now reflects ANTS-1246
done/total chip semantics + the new bg-tasks poll().
SARIF partialFingerprints key bumped to
primaryLocationLineHash/v1 (SARIF v2.1.0 § 3.27.13 versioning
convention).
computeDedup chained single-arg .arg() — closes a
%n-placeholder collision vector if a file path contains a
literal "%3" / "%4" substring (legal on filesystems).
terminalwidget — dead #include <QSurfaceFormat> removed
(no QSurfaceFormat reference left in the file).
terminalwidget — dead duplicate loop in invalidateSpanCaches
removed (lines 3472-3478 re-erased the exact key set the
previous loop had just erased; was a no-op).
terminalwidget — unused m_perfLastPaintUs member + the
overlay line that read it dropped. The member was never written;
the overlay always showed Paint: 0 us.
terminalwidget — isCellSearchMatch early-returns on empty
m_searchMatches. Called twice per cell per frame; ~16,000
wasted comparator invocations/frame at 60 fps when no search
active.
luaengine::lua_ants_on — per-event handler cap of 64. A
plugin calling ants.on(event, fn) in a loop would otherwise
grow the handler list until the heap cap fired.
CI — build-asan job timeout 20 → 30 min. Mirrors the
earlier 421e32d build-test 15 → 25 bump shape; ASan is
2-3× slower than the release build, and the 0.7.91 sweep
pushed it past the prior cap.
ROADMAP.md — 30 follow-up cards allocated ANTS-1260…
ANTS-1317 across five themes: indie-review deferred items
(1260…1277), MCP renderer bug (1278), MCP orchestration
consolidation (1279…1288), skill displacement + context
discipline + security hardening + namespacing (1289…1298),
general Claude Code workflow MCPs (1299…1312), and visibility /
Anthropic-discoverability (1313…1317). .roadmap-counter:
1259 → 1317.
antshelper — silent false-clean on hung script.
proc.waitForFinished(60000) return value was ignored;
Qt's exitStatus() returns NormalExit + exitCode() returns
0 by default on a still-running process. A hung drift script
was reported as {ok:true, clean:true} exit 0. Now branches
on the timeout: kills the process and returns
{ok:false, error:"drift script timed out after 60s", code:"script_timeout"} with exit 1.
antshelper — isatty(0) short-circuit on readStdin.
Interactive runs no longer block until Ctrl-D when stdin is
a TTY.
claudeintegration — m_planModeByPid PID-reuse leak. Cache
was only pruned on explicit closeTab. Linux PID-reuse is fast;
an abnormally-exiting shell left a stale entry that poisoned the
next tab assigned the same PID. Now also pruned on the
NotRunning transition in pollClaudeProcess.
claudetasklist + claudebgtasks — m_lastRescanMtimeMs reset on
path change. A re-bind to the same path shortly after a clear
could short-circuit poll() on a matched mtime, skipping the
legitimate rescan.
claudestatuswidgets — m_promptActive now cleared in
resetForTabSwitch (was the only render-state flag left out).
claudestatuswidgets — empty initial label on m_tasksBtn. The
prior "☰ 0/0" literal contradicted the ANTS-1246 hide predicate
(total <= 0 || done >= total). The chip is hidden until first
refresh anyway, but the latent inconsistency is gone.
claudeintegration — qWarning on the 100 MiB transcript cap.
Previously returned an empty QJsonArray silently; "transcript
empty / status frozen" reports now have a logging breadcrumb.
tests — remote_control_launch test window 3000 → 3500 chars
to admit the new cwd validation block in cmdLaunch.
0.7.90 — ANTS-1113 v1: /debt-sweep fold (engine + 4 MCP tools)
Theme: ANTS-1113 v1 — fold /debt-sweep mechanical scan into
Ants. The four canonical post-feature drift categories (code drift,
test coverage, doc drift, packaging drift) lift out of the
file-reading subagent into 4 new MCP tools backed by a
DebtSweepEngine namespace; per-finding triage continues to live in
optional Claude calls. Estimated saving per /debt-sweep run: ~15-40 K
orchestrator tokens (subagent never opens 95% of the project files;
findings come back as pre-classified JSON instead of free-form prose).
Qt "Debt Sweep" tab deferred to ANTS-1259 v2.
ANTS-1113 — DebtSweepEngine helper (src/debtsweepengine.{h,cpp},
Qt::Core only). Per-detector pure functions:
detectStaleTypeComments(projectPath, opt) — code drift (a):
walks git diff <since>..HEAD --name-only for *.cpp/h/py/js/ts/tsx,
extracts comments, flags leading-cap CamelCase tokens (≥4 chars,
\b([A-Z][A-Za-z0-9_]{3,})\b) absent from the project source blob.detectAddedTodos(projectPath, opt) — code drift (c): unified=0
diff parser flags TODO / FIXME / XXX / HACK markers added in scope.detectOrphanQUnused(projectPath, opt) — code drift (d): per-file
pass over git ls-files '*.cpp' '*.h' finds Q_UNUSED(x) /
(void)x; markers wrapping a variable not declared anywhere in
the same file. The only v1 detector that sets autoFixable=true.detectMissingInvariantTests(projectPath, opt) — test coverage:
parses \bINV-([0-9][0-9a-zA-Z]*)\b from each
tests/features/*/spec.md (covers INV-7 and INV-8b); reports
invariants not mentioned in any sibling test_*.{cpp,py,js,go,rs}.detectRoadmapShippedWithoutCommit(projectPath, opt) — doc drift
(a): reads ROADMAP, runs git log --all --format=%s once, flags
✅ items whose stable ID is unmentioned.detectChangelogStaleBullets(projectPath, opt) — doc drift (b):
parses CHANGELOG [Unreleased] block, flags bullets citing files
not in git diff <since>..HEAD.runPackagingDrift(projectPath, opt) — packaging drift: wraps
the existing packaging/check-version-drift.sh; parses stdout
into Finding structs.scanAll(projectPath, opt) — convenience: runs every enabled
detector in canonical category order.applyMechanicalFix(projectPath, finding) — applies one mechanical
edit; returns ApplyVerdict {applied, errorCode, errorMessage} so
MCP handlers can disambiguate file_changed / not_fixable /
io_error no-ops without inspecting QFile errno state.templateDebtSweepFoldInBlock(deferred, ids, dateIso) —
### 🧹 Debt-sweep fold-in (DATE) block per
roadmap-format.md § 3.8 + § 3.5.3. <dateIso> byte-identical
between heading and per-bullet Source: line (locked by INV-10).triagePrompt(llmShaped) — pure string templating of the LLM
triage prompt for the judgment-required subset.
Locked by tests/features/debt_sweep_engine/ (13 tests).ANTS-1113 — 4 new MCP tools registered via the consolidated
registerToolProvider registry from ANTS-1253:
debt_sweep_scan — runs the four-category scan; returns
{findings, total_findings, by_category, since_resolved} (input:
optional since, optional categories subset).debt_sweep_apply_fix — applies one mechanical fix in-place;
ok=true even on recognised no-ops (file_changed /
not_fixable); ok=false only on io_error. Re-grep guard in
applyMechanicalFix § 3.9 makes the operation idempotent against
re-application (locked by INV-13a).debt_sweep_defer — allocates IDs via RoadmapFoldIn::allocateIds,
renders the fold-in block, and atomically inserts it into
ROADMAP.md when findActiveReleaseHeading succeeds. Eager ID
allocation (no rollback on written:false) — same trade-off as
ANTS-1111 § 2.5; envelope returns the allocated IDs so the caller
can splice manually.debt_sweep_triage_prompt — emits the LLM triage prompt for a
caller-filtered subset of findings.
All 4 follow the existing UID-scoped 0700-perms IPC trust model;
schemas use additionalProperties: false. Locked by
tests/features/mcp_debt_sweep_tools/ (5 tests).ANTS-1259 (new ROADMAP item) — ANTS-1113 v2: AuditDialog
"Debt Sweep" tab + per-finding Fix / Defer / Allow buttons +
Triage with AI button (aidialog dispatch). Adds README CLI flag
drift detector (needs binary execution). Spec § 1.1 of
docs/specs/ANTS-1113.md documents the v1/v2 split rationale.
featurecoverage.{h,cpp} — three internal helpers lifted to
FeatureCoverage:: public surface so DebtSweepEngine can reuse
them without copying the extension list, skip-dir set, or
containment-check fallback chain:buildProjectSourceBlob(projectPath) — concatenates every source/
config/doc file in the project tree (per the canonical extension
list) into one UTF-8 blob; skips build dirs + spec.md files.existsInSource(blob, token) — substring containment +
:: and . tail-fallbacks (with the same identifier-shape
guards runSpecDriftCheck had inline).specStopwords() — public accessor for the existing internal
kSpecStopwords set.
runSpecDriftCheck is rewired to call them; behaviour unchanged.tests/features/debt_sweep_engine/ (13 tests) — engine pure-fn
invariants (INV-3, INV-4, INV-9, INV-10, INV-11, INV-13a, INV-13btests/features/mcp_debt_sweep_tools/ (5 tests) — source-grep
verification of the wiring layers (tools/list, registerToolProvider,
remotecontrol.h declarations, remotecontrol.cpp definitions, schema
additionalProperties:false).tests/features/mcp_indie_review_tools/test_mcp_indie_review_tools.cpp —
AllSchemasUseAdditionalPropertiesFalse region boundary tightened
so it doesn't drift when new tool blocks land after it.0.7.89 — ANTS-1112 v1: /indie-review fold (engine + 5 MCP tools)
Theme: ANTS-1112 v1 — fold /indie-review orchestration into
Ants. The mechanical halves of the multi-agent independent review
(partition, brief assembly, cross-lane corroboration, synthesis-prompt
templating, ROADMAP fold-in) lift out of orchestrator context as 5
new MCP tools backed by a IndieReviewEngine namespace; per-lane
review judgment + dispatch continue to live in Claude subagent
calls. Estimated saving per /indie-review run: ~20-50 K orchestrator
tokens. Qt dialog deferred to ANTS-1258 v2.
ANTS-1112 — IndieReviewEngine helper (src/indiereviewengine.{h,cpp},
Qt::Core only). Six pure functions:
derivePartition(projectPath) — reads CLAUDE.md ## Module map (src/) via the existing SubsystemMap::cachedLanes helper, walks
src/ to compute per-lane source-file lists; honours
<projectPath>/.indie-review/partition.json override when present.assembleBrief(projectPath, lane) — verbatim brief text for one
lane: header + source bodies + ROADMAP slice + standards links.
Pure file IO bounded to projectPath.extractFileLineCitations(projectPath, report) — regex pass over
a single review report; rejects paths that escape projectPath
(defense against fabricated cites).corroboratedFindings(projectPath, reports, minLanes=2) —
cross-lane corroboration; (file, -1) (file-level) and
(file, 42) (line-level) are distinct keys (intentional).synthesisPrompt(reports, threatModelExtras) — pure string
templating of the optional cross-cutting synthesis prompt.templateIndieReviewFoldInBlock(actionable, ids, dateIso) —
### 🔍 Indie-review fold-in (DATE) block per
roadmap-format.md § 3.8 + § 3.5.3.assembleThreatModelExtras(projectPath) — MCP-handler helper
that concatenates CLAUDE.md / SECURITY.md / .semgrep.yml under
=== <header> === markers.
Locked by tests/features/indie_review_engine/ (13 tests).ANTS-1112 — 5 new MCP tools registered via the consolidated
registerToolProvider registry from ANTS-1253:
indie_review_partition — returns lane list (no input).indie_review_brief — returns the assembled brief for one lane
(input: lane).indie_review_corroborate — returns cross-lane corroborated
findings (input: reports map, optional min_lanes).indie_review_synthesis_prompt — returns the rendered synthesis
prompt (input: reports, optional include_threat_model_extras).indie_review_fold_in — allocates IDs via
RoadmapFoldIn::allocateIds, renders the
### 🔍 Indie-review fold-in (DATE) block, and atomically
inserts it into ROADMAP.md when findActiveReleaseHeading
succeeds (input: actionable array, optional date_iso,
optional release_block_heading).
All 5 follow the existing UID-scoped 0700-perms IPC trust model;
schemas use additionalProperties: false. Locked by
tests/features/mcp_indie_review_tools/ (5 tests).ANTS-1258 (new ROADMAP item) — ANTS-1112 v2: Qt
IndieReviewDialog that wraps the v1 engine for users who want
in-app indie review without Claude orchestration. Spec § 1.1 of
docs/specs/ANTS-1112.md documents the v1/v2 split rationale.
0.7.88 — ANTS-1111 v1: /audit triage fold (engine layer)
Theme: ANTS-1111 v1 — fold /audit triage into the Project
Audit tool's engine layer. Five mechanical pieces lift out of the
LLM round-trip: cross-tool corroboration severity-tier shift,
framework auto-detect, ROADMAP fold-in helper, the
// audit: drop[=rule] inline-suppress alias, and the
RuleQualityTracker::noisyRuleIds accessor that feeds the shift.
UI affordances (Fold-into-ROADMAP button, per-finding Allow
button, Since-baseline pill) deferred to ANTS-1257 v2.
Cumulative payload: ~750 LoC engine + ~33 new test cases, all
green. Foundation ships ahead of ANTS-1112 / 1113 (which reuse
RoadmapFoldIn).
ANTS-1111 — RoadmapFoldIn helper (src/roadmapfoldin.{h,cpp},
Qt::Core only). Three operations: allocateIds(projectPath, n)
reserves N consecutive IDs from .roadmap-counter under
::flock(LOCK_EX|LOCK_NB) (5 s budget; adopts the
configbackup.h pattern); insertBlock(projectPath, heading, block) performs an atomic insert immediately after a named
## heading via QSaveFile, preserving original file
permissions; findActiveReleaseHeading(projectPath) returns the
first (target: …)-marked heading, falling back to the first
shipped release block. Returns false on heading-not-found —
caller is responsible for creating the heading first.
Locked by tests/features/roadmap_fold_in/ (12 tests).
ANTS-1111 — AuditEngine::applyCorroborationShift
(src/auditengine.cpp). Severity-tier promotion when ≥ 2
distinct CheckIds cite the same (file, line) (clamped to
Blocker); demotion when a single-tool finding's checkId is in
noisyRules (clamped to Info). Wired into
AuditDialog::renderResults after the enrichment pass; the
noisyRules set comes from the new
RuleQualityTracker::noisyRuleIds(fpThreshold=50, minSamples=5)
accessor (src/auditrulequality.{h,cpp}). Locked by
tests/features/audit_corroboration_shift/ (8 tests).
ANTS-1111 — AuditEngine::templateRoadmapFoldInBlock
(src/auditengine.cpp). Pure-string templating of a
### 🔍 Audit fold-in (DATE) subsection per
roadmap-format.md § 3.8 (subsection shape) + § 3.5
(per-bullet fields). Caller pre-allocates IDs via
RoadmapFoldIn::allocateIds. Empty input → empty string.
Covered by tests/features/roadmap_fold_in/Inv5TemplateShape.
ANTS-1111 — AuditHygiene::detectProjectFrameworks
(src/audithygiene.{h,cpp}). Probes the project root for 7
framework markers (flask, django, react, vue, qt6, rust, go) by
reading requirements.txt / pyproject.toml / package.json /
CMakeLists.txt / Cargo.toml / go.mod / manage.py.
Companion semgrepRulePacks(QStringList) maps recognised
frameworks to {"--config", "p/<fw>"} argv pairs. Pure-IO,
bounded to projectPath. Wiring into the live runNextCheck
semgrep invocation deferred to ANTS-1257. Locked by
tests/features/audit_framework_detect/ (10 tests).
ANTS-1111 — // audit: drop[=rule] inline-suppress alias.
Shorter ergonomic form of the existing // ants-audit: disable
token. Same parser code (auditdialog.cpp:2055); same
semantics (bare form suppresses every rule on the line; =rule
suffix targets one). Both forms coexist indefinitely. The
generic audit: prefix is confined by the verb constraint
(\s*drop) plus the existing rule-list parser
(auditdialog.cpp:2083-2107); collision-prevention covered in
spec § 2.6 + INV-12. Locked by
tests/features/audit_drop_alias/ (2 tests).
ANTS-1111 — Widened-allowlist documentation. The pre-existing
.audit_allowlist.json filter (AuditDialog::allowlisted(),
auditdialog.cpp:3991) was already cross-detector — the call
site lives in the main per-finding loop and matches by exact
checkId equality, so any detector's rule (clazy-X,
cppcheck-Y, etc.) can be allowlisted. The doc-comment at
auditdialog.h:101-110 now documents this widened scope (was
previously labelled grep-rule only). No behaviour change.
ADR-0003
(docs/decisions/0003-cc-fold-relax-gate-and-draw-boundary.md)
— relaxes the ADR-0002 dec 8 gate (ANTS-1120 measurement no
longer pre-requisite for the CC-fold bullets), draws the
ANTS-1108 ↔ ANTS-1111 / 1113 per-surface boundary, and
enumerates pre-existing scaffolding so the subsequent ANTS-1112
/ 1113 specs don't re-invent it.
ANTS-1257 (new ROADMAP item) — ANTS-1111 v2: UI affordances
(Fold-into-ROADMAP button, per-finding Allow button,
Since-baseline pill, semgrep wiring). Spec § 12 of
docs/specs/ANTS-1111.md documents the v1/v2 split rationale.
0.7.87 — MCP token-reduction pack
Theme: MCP token-reduction pack — five new MCP tools
(workspace_search, file_outline, git_state, subsystem,
last_audit_summary) collapse common Bash/Read patterns into
structured envelopes, plus a token-saving hook pack and a
provider-registry consolidation that makes the next tool a one-line
add. Cumulative session saving on a typical /indie-review +
/audit workflow: ~50-100 K tokens.
setXProvider/m_xProvider setter+member pairs
in ClaudeIntegration (each added by ANTS-1244 / 1247-1251) with
a single registerToolProvider(QString name, ToolHandler handler)
surface backed by std::map<QString, ToolHandler> m_toolProviders,
where ToolHandler = std::function<QString(const QJsonObject&)>.
The 92-line tools/call else if (toolName == "X" && m_XProvider)
chain in claudeintegration.cpp collapses to one inline branch
for get_session_info (the documented carve-out — it reads
ClaudeIntegration's own state, not an external delegate) plus a
single m_toolProviders.find(toolName) lookup. MainWindow::setupClaudeMcpProviders
now makes 12 registerToolProvider("name", lambda) calls; each
lambda absorbs the dispatcher-side argument extraction + result
formatting that previously lived in the per-tool dispatch case.
No behaviour change. Source delta: −157 LoC across
claudeintegration.{h,cpp} + mainwindow.cpp. Spec:
docs/specs/ANTS-1253.md (5-loop cold-eyes pass, ship-ready).
New regression test: tests/features/mcp_provider_registry/
(10 invariants source-greped, pre-fix red verified — 9/10 fail
against the unrefactored tree; INV-10 carve-out preservation
passes both sides).ANTS-1254 — last_audit_summary MCP tool. New read-only tool
that opens the latest .audit_cache/audit-*.sarif and returns a
compact summary: counts (error/warning/note/suppressed) plus
top_findings[] sorted by SARIF level desc → confidence desc →
file asc → line asc. Default top_n=5, severity_floor="warning";
server-clamps top_n to [0, 50]. Saves ~5-15 K tokens per
audit consultation vs reading the HTML report (which today's flow
uses). Backed by the new AuditEngine::summariseSarif(path, topN, levelFloor) pure parser + RemoteControl::cmdLastAuditSummary
with a single-entry mtime-keyed cache ((path, mtime, topN, floor)
4-tuple). Latest-SARIF discovery uses lex-max filename
(audit-YYYYMMDD-HHmmss.sarif is sortable at second granularity);
html_path derives via extension swap, falling back to lex-max
audit-*.html within ±60 s of the SARIF timestamp (the SARIF and
HTML export buttons each call QDateTime::currentDateTime()
independently). Severity resolution: rule-index lookup
(runs[0].tool.driver.rules[].properties.severity) with fallback
for foreign SARIF (error→CRITICAL, warning→MAJOR, note→INFO).
Per-tab gate inherits from resolveRootCanonical(MainWindow*).
Lands on the post-1253 registry (one registerToolProvider call
in setupClaudeMcpProviders, no per-tool setter). Spec:
docs/specs/ANTS-1254.md (5-loop cold-eyes pass, ship-ready).
New regression test: tests/features/mcp_last_audit_summary/
(10 invariants — 6 parser-side against committed
fixture_min.sarif + fixture_empty.sarif; 4 wiring-side via
source-grep). Pre-fix red verified — summariseSarif symbol
absent from stashed engine fails build.
ANTS-1252 — Token-saving hook pack. Five bash hooks plus
tools/install-hooks.sh that nudge Claude Code toward cheaper
MCP tool calls. SessionStart preamble emits a ≤ 500 B
branch/ahead/last-commit summary (cap enforced via head -c).
PreToolUse(Bash) veto blocks grep -r src/, git status,
cat ROADMAP.md | grep etc. with reasons capped at 200 bytes,
redirecting to mcp__ants__workspace_search /
mcp__ants__get_git_status / mcp__ants__roadmap_query. Per-
command escape hatch: trailing # ants-bypass comment, stripped
before pattern match (INV-12 — never appears in the emitted
reason text). PreToolUse(Read) veto blocks full reads of
ROADMAP.md > 50 KiB; bypassed cleanly via offset/limit
args. Stop hook backgrounds an ants-helper drift-check under
flock (INV-11) with a sane-toplevel marker write (INV-9).
PreCompact hook walks the transcript JSONL and writes the most
recent TodoWrite snapshot to
~/.cache/ants-terminal/precompact_<sessionId>.json only if
sessionId matches ^[a-zA-Z0-9_-]{1,64}$ (INV-2). Per-project
gate via committed .ants-project marker — non-ants sessions
exit silently in sub-millisecond stat ascent (no git rev-parse
fork). Install-hooks hardening: lstat symlink abort (INV-5),
cp --no-dereference backup, sentinel-key fence
ants_hooks_pack_v1 rather than text-fence comments (INV-6 —
jq strips JSON comments), tmpfile + jq empty validate-
before-rename (INV-8), idempotent re-install, --dry-run and
--uninstall flags. Token saving: ~30-100 K/week depending on
which siblings shipped first (calibrated in spec § 5).
Conformance harness at tests/features/hook_pack/test_hooks.sh
(shell-driven per audit_self_test.sh pattern, no C++ link)
covers INV-1/2/3/4/7/9/10/12 + bash-veto behaviour + read-
roadmap-veto behaviour + install round-trip + symlink abort —
36 assertions, all green; INV-6/8/11 deferred to manual smoke
with documented rationale. Spec: docs/specs/ANTS-1252.md.
Cold-eyes review on tests/features/hook_pack/spec.md returned
7 findings (1 HIGH on source-vs-runtime semantics for INV-12,
2 MEDIUM on doc-code drift, 4 LOW); all fixed inline.
ANTS-1251 — subsystem MCP tool (consolidated; map / files /
recent_changes via op discriminator). Pre-parses the project's
CLAUDE.md ## Module map (src/) H2 into a lanes[] array and
serves per-lane chunks so /indie-review reviewers don't each
re-read the file. Three ops: map returns [{name, summary}, …],
files returns the lexicographically-sorted src/<lane>* glob,
recent_changes returns merged-by-sha git log entries across
every file in the lane. Cache is mtime-only on CLAUDE.md (no
wall-clock TTL — concurrent reviewers share warm cache). Defensive
parser drops bullets that don't match the `name` — summary
shape and splits multi-name bullets like `a` / `b` into one
Lane each. Hardening: lane membership check precedes any
filesystem call (closes cold-eyes S1251-1 path-traversal vector);
per-result canonical-startsWith re-check on resolved files
(defence in depth against malicious symlinks inside src/).
Composes cmdGitState({op:"log", path:<file>}) per lane file for
the recent_changes op — no duplicated git plumbing. Distinct
error codes bad_op (input enum) and unknown_lane (with
lanes:[…] echoed in the response so the caller can recover).
Single setter / member / provider-lambda triple on
ClaudeIntegration + MainWindow. Locked by feature test
mcp_subsystem/ (12 invariants — decl, INV anchors, IPC dispatch,
MCP tools/list schema with op enum + op in required[],
MCP tools/call dispatch, header surface, mainwindow lambda,
op-switch literals, error-code surfacing, cmdGitState composition,
parser surface, CMake wiring, and the ≥ 15-lane CLAUDE.md parser
floor). Token math: ~24 K saved per /indie-review run (6
reviewers × ~3.5 K → ~250 each); permanent schema cost ~115
tokens. Side-effect: mcp_workspace_search test INV-3c now uses
a word-boundary regex (\bsystem\() instead of substring match
so cmdSubsystem( is not false-flagged as a shell escape.
ANTS-1250 — git_state MCP tool (consolidated; status / log /
diff via op discriminator). Replaces multiple Bash calls to
git status, git log, git diff with a single structured tool.
Cold-eyes pass 2 collapsed three originally-proposed verbs into
one to save ~240 permanent schema tokens per session start; per-
call savings are ~14–300 tokens depending on op. Status returns
{branch, upstream, ahead, behind, files:[{path,index,worktree}], untracked[]} parsed from --porcelain=v1 -b. Log returns
{commits:[{sha,subject,date,body?}], truncated} parsed from
unit-separator-framed --pretty=format:, with n+1 probe to
detect truncation and per-body 1 KiB cap when body:true. Diff
returns {files:[{path,added,removed}], totals} parsed from
--numstat, with binary files surfaced as null added/removed.
Hardening: shell-less QProcess::start("git", QStringList...)
via the new gitwrap.{h,cpp} synchronous helper (5 s terminate
kill, 4 KiB stderr cap); strict regex on
range excludes leading - from any rev-component (closes
cold-eyes S1250-1 flag-injection); -- argv separator before
every user-derived positional arg, with ./ prefix on
--leading paths; canonical-startsWith path-escape guard for
path mirroring ANTS-1248's lane check; distinct error codes
bad_op, bad_range, bad_path, git_failed, git_missing,
not_git_repo. Locked by feature test mcp_git_state/ (13
invariants spanning decl, INV anchors, shell-lessness in
gitwrap.cpp, IPC route, tools/list schema with op enum +
required:["op"], tools/call dispatch, header decl + member,
provider lambda, op-dispatch chain, the stricter regex literal,
the 2-tier kill constants, and CMake wiring). Spec:
docs/specs/ANTS-1250.md.ANTS-1249 — file_outline MCP tool. Returns a compact
{header_doc, symbols:[{line, kind, name, signature}], total_lines, total_bytes} envelope for a single file instead
of a full Read. ~13-39× compression on typical C++ source
(e.g. auditdialog.cpp 67 K tokens → ~2 K). New
src/fileoutline.{h,cpp} translation unit hosts a 6-regex
scanner (C++ member / type / free-func / Qt-marker + Python +
Markdown headings). Each regex is a static const QRegularExpression with .optimize() invoked at first use,
so the JIT compiles once per process. Possessive quantifiers
path argument is canonicalised + NFC-normalised + checked
against the project root via startsWith (same security
posture as ANTS-1248's lane). Locked by feature test
mcp_file_outline/ (10 invariants: wiring + runtime floor of
≥ 8 symbols against the in-tree auditdialog.cpp to catch
regex-set regressions, plus not-found path). Spec:
docs/specs/ANTS-1249.md.ANTS-1248 — workspace_search MCP tool (ripgrep wrapper).
Replaces typical Bash grep -r ... src/ patterns with a
structured {ok, matches:[{file,line,text}], truncated, elapsed_ms} envelope. Shell-less argv (QProcess::start("rg", QStringList...) — never bash -c, never a single-string
overload). Server clamps max_results at 500 (default 50);
lane and glob NFC-normalised and rejected if they contain
control chars / backslash / .. segments / parent-traversal
past the project root (canonical-startsWith check via
QFileInfo::canonicalFilePath()); 4 KiB stderr cap surfaced
only on ok:false; 2-tier hard kill (2 s terminate() then
200 ms grace → kill()) so a catastrophic regex can't outlive
the wall budget. Estimated saving ~6–15 K tokens per typical
bug-investigation session at ~150 permanent schema tokens.
Locked by feature test mcp_workspace_search/ whose 10
invariants cover decl, INV anchors, shell-lessness, IPC route,
tools/list schema (with required: ["pattern"]), tools/call
dispatch, header decl + member, provider lambda, the literal
ripgrep flags (--json, --no-heading, --line-number,
--max-columns, --threads), and the 2-tier kill wiring.
Spec: docs/specs/ANTS-1248.md.
0.7.86 — MCP integration end-to-end + inputSchema compliance
Theme: MCP integration end-to-end — Claude Code sessions inside
Ants tabs can now call the in-process MCP server. Three internal
remote-control verbs (roadmap_query, tab_list, get_text) are
promoted to MCP tools; a Python stdio bridge (tools/mcp-bridge.py)
connects Claude Code's stdio MCP client to the Ants QLocalServer
Unix socket; and a Zod-compliance fix repairs the silent
zero-tools failure that the bridge surfaced on its first probe.
Plus a Tasks chip semantics tidy-up and an optional status filter
on roadmap_query that cuts typical "what's next" queries by ~7×.
ANTS-1244 — roadmap_query / tab_list / get_text as MCP
tools. Wires three existing RemoteControl IPC verbs as MCP
tools so Claude Code sessions inside Ants tabs can query terminal
state via tool-call rather than Bash / Read. The handlers
(cmdRoadmapQuery, cmdTabList, cmdGetText) are promoted to
public on RemoteControl; MainWindow lambdas in
setupClaudeMcpProviders delegate directly, sharing the existing
roadmap-query mtime cache. Token saving per
session-with-roadmap-need: ~110 K tokens (a 120 K Read of the
482 KiB ROADMAP.md → ~13 K of structured JSON for 397
status-emoji bullets at ~133 B/bullet measured). Per-call saving
on tab_list / get_text: 30–100 tokens of bash-glue Claude no
longer composes. Locked by 7-invariant source-grep test
mcp_extra_tools/. Spec converged across 7 cold-eyes loops.
ANTS-1247 — status filter on roadmap_query MCP tool /
IPC verb. Optional status argument ("all" default,
"active" = 📋+🚧, "shipped" = ✅; case-insensitive) on the
roadmap_query MCP tool. On this repo at ship time: "active"
returns 57 of 399 bullets (~1.75 K tokens vs ~12 K), a ~7× saving
per "what's next" query. Provider lambda widened to thread the
filter through; cache continues to hold the full unfiltered array
(filter runs post-cache). Pack-mate with ANTS-1248..1252
(spec-only) and ANTS-1253/1254 (planned follow-ups). Cold-eyes
pass 2 converged across 4 lanes (performance / token reduction /
security / optimisation) over 4 loops — final pass clean.
ANTS-1255 — MCP stdio bridge (tools/mcp-bridge.py). Unblocks
the entire MCP pack on the consumer side. The Ants in-process MCP
server exposes JSON-RPC over a QLocalServer Unix socket; Claude
Code's MCP client speaks stdio. Until now, every server-side tool
registered since ANTS-1244 was unreachable from a Claude Code
session. ~100-line Python 3 stdlib script reads line-delimited
JSON-RPC from stdin, opens one AF_UNIX connection per request
(server is one-shot by design), forwards the request, writes the
reply to stdout. Notifications (no id) are forwarded without
awaiting a response. Socket selection: $ANTS_MCP_SOCKET override
or newest-mtime /tmp/ants-terminal-mcp-*. Register with
claude mcp add ants -- /path/to/tools/mcp-bridge.py once per
machine. End-to-end verified: initialize handshake → 9 tools
listed → roadmap_query status="active" → bad_status error
path → clean exit on stdin close.
ANTS-1246 — Tasks chip progress semantics + Mode B batch reset.
Two coordinated fixes for the same user-visible symptom (the
bottom-right Tasks chip showing wrong or no info during an active
Claude task list). (1) Chip now reads ☰ <completed>/<total> and
stays visible iff 0 < done < total, closing the visibility hole
left by ANTS-1221's pending-only predicate — visible end-to-end
through every active run, hides cleanly at 100 %. (2) The
TaskCreate path in ClaudeTaskListTracker::parseTranscript
(Mode B) piled up completed tasks forever; now, when a new
TaskCreate arrives and every task in out is completed, the
prior batch is cleared before appending. Mode A (TodoWrite
snapshot) was already correct. Partial batches (any
pending/in_progress) are preserved. Locked by 4-INV
tasks_chip_done_over_total/ and 3 new behavioural cases in
claude_task_list/. Spec: docs/specs/ANTS-1246.md (2 cold-eyes
loops, clean pass).
ANTS-1256 — MCP tools/list inputSchema compliance. The six
zero-arg MCP tools (get_cwd, get_session_info,
get_last_command, get_git_status, get_environment,
tab_list) were emitted without an inputSchema field. Claude
Code's MCP client validates tools/list with Zod and rejects the
entire response when any entry omits the field — the
connection reports "Connected" but registers zero tools. Surfaced
2026-05-12 on the first probe through the ANTS-1255 bridge: the
log read tools[1].inputSchema expected object, received undefined for indices 2/3/4/5/7 as well. Fix: a shared
QJsonObject emptySchema; emptySchema["type"] = "object";
sentinel, assigned to each zero-arg tool's inputSchema. Locked
by feature test mcp_tools_list_schema/ whose third invariant
pins tools.append() == ["inputSchema"] = count parity inside
the tools/list block — so any future tool addition that omits
the field breaks CI before it reaches a user.
0.7.85 — Roadmap density toggle + Qt::Popup blocker fix
Theme: Roadmap dialog density toggle (compact / cozy /
comfortable) — Material 3 / Linear / GitHub-style density
selector that persists across sessions. Plus a regression fix
that surfaces while building it: Qt::Popup widgets (combo
dropdowns, menus, color pickers) opened from inside a dialog
were being eaten by the ANTS-1051 pseudo-modal blocker. The
fix is one short-circuit in dialogfocus::shouldSuppressEvent ForDialog and benefits every dialog that hosts a popup, not
just the new density combo.
QComboBox lives at the trailing edge of the filterRow (after
the existing addStretch(1)) with three options: compact
(9/10/11/14 px text tier, proportional padding), cozy (the
pre-1238 baseline — 10/11/12/13/16 px), and comfortable
(12/13/14/15/18 px). Selection drives a CSS class on
renderCardsHtml's root container; per-tier sizes/padding are
defined in STYLE_COMPACT / STYLE_COZY / STYLE_COMFORTABLE.
Persists via a new roadmap_density Config key, with graceful
fallback to "cozy" on missing/invalid value and silent
best-effort on persistence-write failure (matches the existing
Config write convention). The default is cozy, byte-equal to
the pre-1238 baseline render — INV-1 in the spec's test bundle
locks that. Regression-locked by ANTS-1238-INV-1..9 in
tests/features/roadmap_density/spec.md (tier-unique sentinels
9/16/18 px scoped to renderCardsHtml; combo present and wired;
Config getter/setter round-trip; persistence-failure contract).
No keyboard shortcut in v1 — the ?-cheatsheet (ANTS-1236)
doesn't gain a row. Spec: docs/specs/ANTS-1238.md
(cold-eyes-clean after 4 loops + ~50 verified findings fixed;
full audit trail at ROADMAP.md
### 📝 Cold-eyes 2026-05-12 (ANTS-1238 spec)).Qt::Popup events were swallowed by the
pseudo-modal blocker. Surfaced by the ANTS-1238 density combo:
clicking an option in the dropdown popup did nothing — the
popup didn't even close. Root cause: a Qt::Popup window
(QComboBox dropdown, QMenu context menu, color-picker popup, …)
is its own top-level widget, and QWidget::isAncestorOf stops
walking at window boundaries — so popup descendants opened FROM
inside a dialog read as "outside the dialog's tree" and the
pseudo-modal blocker (ANTS-1051) ate every mouse event on them.
Fix is a one-block short-circuit at the top of
dialogfocus::shouldSuppressEventForDialog: if
QApplication::activePopupWidget() is non-null and target is
either that popup or a descendant of it, return false
immediately. Regression-locked by a new INV-2i in
tests/features/dialog_pseudo_modal/spec.md —
source-grep-asserted rather than behaviourally driven because
the offscreen QPA platform doesn't promote Qt::Popup widgets
to activePopupWidget(). Affects every dialog that hosts a
combo / menu / color picker, not just the Roadmap density combo
— Settings, AuditDialog, etc. all benefit transparently.0.7.84 — RoadmapDialog v2 finishing pass
Theme: RoadmapDialog v2 finishing pass — keyboard ergonomics
(/-focus, Esc-clear, ?-cheatsheet), accessibility (status and
theme glyph labels for screen readers), and a freshness signal
("Updated N days ago" on in-progress cards). Plus theme-aware
frameless title bars + dark-theme link fixes across every
dialog.
ANTS-1237 — "Updated N days ago" line on 🚧 cards in the
Roadmap dialog. Surfaces stall signal at-a-glance: every 🚧
in-progress card now shows when its bullet block was last
touched, mirroring GitHub Projects / Linear / Jira's default
"last activity" affordance. Derived from one
git blame --line-porcelain call against ROADMAP.md (cached
by mtime via the same pattern as parseShippedDates), then
MAX(author-time) over each - 🚧 [ANTS-NNNN] bullet's line +
its 2-space-indented continuation lines. Renders as
<span class="rm-date">· Updated <X></span> inline with the
ID chip, where <X> is one of: today / yesterday /
Nd ago (2-13d) / Nw ago (2-8w) / Nmo ago (2-12mo) /
Ny ago (≥1y). Reuses the existing .rm-date CSS class —
no new style rule. Status-gated (✅ already shows shipped
date; 📋/💭 don't need it). Graceful degradation on non-git
checkouts: git blame failure → empty hash → no "Updated"
line, mirroring the ✅-card path when shipped date is absent.
ID mentions outside the bullet block (e.g. audit-trail prose
in another card's body) do NOT contribute to MAX — the walker
only enters block-mode at lines matching
^- 🚧 \[ANTS-NNNN\]. Cost: ~175 ms blame call on first
dialog open per session against the current 8723-line
ROADMAP, cached thereafter; ~4 MB transient peak for the
porcelain output (verified by wc -c); ~1 KiB persistent
per dialog at realistic 5-15 🚧 bullet counts. Regression-
locked by ANTS-1237-INV-1..8 in
tests/features/roadmap_inprogress_age/spec.md (renderer-
layer + parser-layer test split; parser tests early-skip with
GTEST_SKIP() when git is not on PATH). Spec:
docs/specs/ANTS-1237.md (cold-eyes-clean after 4 loops + ~40
verified findings fixed; full audit trail at ROADMAP.md
### 📝 Cold-eyes 2026-05-11 (ANTS-1237 spec)).
ANTS-1234 — Roadmap dialog /-focus + Esc-clear + body-only
auto-expand. Three keyboard-first ergonomics additions to the
Roadmap dialog's existing substring search box. (1) Pressing /
anywhere in the dialog focuses the search box and select-alls
any existing predicate so the next keystroke replaces it
(Linear / GitHub / Notion / VS Code / Slack convention).
Layout-robust via event->text() == "/"; gated on
!m_searchBox->hasFocus() so the user can still type / into
the predicate (URLs, path tokens). (2) Pressing Esc while the
search box has focus clears the predicate and removes focus —
one press undoes a search, a second press closes the dialog as
before. Implemented via installEventFilter(this) on the
search box + a dialog-side eventFilter override that consumes
the Escape KeyPress before it can reach QDialog::reject. For
every non-Escape key, the filter falls through so the QLineEdit
receives PgUp / F5 / typing characters normally. (3) Cards
whose match lives only in the body continuation prose (not in
ID / headline / Layman) auto-expand for that render so the
matched substring is visible in context. Render-time only;
m_expandedItems is never mutated, so clearing the search
reverts every auto-expanded card to its prior user-driven
state. The id:NNNN jump shortcut also auto-expands the
matched card. The cheatsheet (ANTS-1236) gains a / row in
lockstep — kRoadmapShortcuts[] bumps 9 → 10 rows and the
static_assert + test assertions all move 9 → 10 in the same
commit. Cost: +4 case-insensitive QString::contains calls per
card per render (~0.3 ms on 200 cards, well below the 120 ms
search debounce). Regression-locked by ANTS-1234-INV-1..9 in
tests/features/roadmap_search_keybinds/spec.md. Spec:
docs/specs/ANTS-1234.md (cold-eyes-clean after 7 loops + ~25
findings fixed).
ANTS-1236 — Keyboard-shortcut cheatsheet in the Roadmap dialog.
Press ? inside the Roadmap dialog to open a sub-dialog listing
every keyboard shortcut it ships today. Mechanism: a file-scope
kRoadmapShortcuts[] data table in src/roadmapdialog.cpp is the
single source of truth; roadmapShortcutRows() exports it to
RoadmapShortcutsDialog (src/roadmapshortcutsdialog.{h,cpp}),
which renders a two-column QTableWidget (Shortcut / Action).
10 shortcuts at ship after ANTS-1234 added /: ?, /, Esc,
F5, Ctrl+C, Ctrl+A, ↑ ↓, PgUp PgDn, Home End,
Tab Shift+Tab. The trigger uses event->text() == "?"
(layout-robust — AltGr / dead-key paths on non-US layouts still
hit it) and gates on !m_searchBox->hasFocus() so the user can
still type ? into the substring filter. The overlay is lazy +
reused (one instance via QPointer), inherits the active
terminal theme through DialogChrome, and announces via
setWindowTitle(tr("Roadmap Keyboard Shortcuts")). The
static_assert(std::size(kRoadmapShortcuts) == 10, …) guard
locks the row count to the test's exact-10 assertion so adding a
future shortcut must bump the test in lockstep. Regression-locked
by ANTS-1236-INV-1..8 in
tests/features/roadmap_shortcuts_cheatsheet/spec.md. Spec:
docs/specs/ANTS-1236.md (cold-eyes-clean after 7 loops + ~27
findings fixed).
ANTS-1235 — Accessible status / theme glyph labels in the
Roadmap dialog. Screen readers (Orca / NVDA / VoiceOver)
announce "✅" as "white heavy check mark" by default — useless
as a scan cue on a several-hundred-bullet roadmap. The fix
emits a short, lowercase text label inline alongside each
status emoji on every card: <span class="rm-state">✅</span> <span class="rm-state-label">shipped</span> and friends.
Section-header count chips gain trailing words too —
✅ 47 shipped · 🚧 2 in progress · 📋 3 planned · 💭 8 considered
in place of the prior bare ✅ 47 🚧 2 …. Filter checkboxes
get setAccessibleName() setters ("Show shipped items"
etc.) so Orca speaks the verb-led form instead of the visible
label. The visible ✅ Done checkbox label is renamed
✅ Shipped for vocabulary consistency with the rest of the
roadmap-format standards (§3.3 accepts both terms). Theme
glyphs (🎨/⚡/🔌/🖥/🔒/🧰/📚/📦/🐛/🔍/🧹) stay unlabelled
because the heading text they prefix already labels them.
Verified mechanism: HTML aria-label is dead on arrival in
Qt 6 QTextBrowser (parser strips unknown attributes;
QAccessibleTextInterface only reads toPlainText()), so
any a11y solution must put the label in the rendered text.
docs/standards/documentation.md gains a new Accessibility
section documenting the Qt 6 a11y path so future contributors
don't repeat the aria-label investigation. Regression-locked
by INV-18..24 in tests/features/roadmap_dialog_cards/spec.md
(302/302 tests pass). Spec: docs/specs/ANTS-1235.md. Cold-
eyes audit trail: ROADMAP.md ### 📝 Cold-eyes 2026-05-11.
QPalette. The main terminal window already worked around this
by going frameless (Qt::FramelessWindowHint) and drawing its
own TitleBar widget; dialogs did not. Fixed by introducing a
shared DialogChrome helper (src/dialogchrome.{h,cpp}) that
any dialog ctor can invoke with one line:
auto chrome = DialogChrome::install(this, themeName); — it
sets the frameless flag, prepends a themed TitleBar widget,
wires close / minimize / maximize signals, and returns a
content QWidget to use as the layout parent for the rest of
the ctor. Applied to all ten QDialog subclasses:
RoadmapDialog, AuditDialog, SettingsDialog, SshDialog,
AiDialog, ClaudeAllowlistDialog, ClaudeTranscriptDialog,
ClaudeProjectsDialog, ClaudeBgTasksDialog, and
ClaudeTaskListDialog. MainWindow::applyTheme now also calls
DialogChrome::setActiveTheme(name) so dialogs that don't
receive a theme name still pick up the current one. Live-theme
refresh is preserved on the dialogs that previously supported
it (Roadmap, bg tasks, task list) — the helper's
applyTheme(dlg, bar, name) re-styles both the dialog palette
and the title bar.#NNNN on summary row +
larger ID font. Third-pass user feedback after ANTS-1240
(2026-05-11): the card's #NNNN (and shipped date) were
emitted on a separate <div class="rm-meta"> row below the
summary line, which (a) hit the same Qt nested-block
QPalette::Base frame issue that ANTS-1240 fixed on the
expanded body — painting a darker band under the ID on dark
themes — and (b) was visually noisy. Fixed by emitting the
hashed ID and shipped date as inline <span> children of the
card div, on the summary row immediately before the
rm-toggle anchor. CSS .rm-id font bumped from the old
10 px meta size to 12 px so the number is scannable at a
glance. The .rm-meta CSS rule is gone. Regression-locked by
spec INV-17 in tests/features/roadmap_dialog_cards/spec.md.QTextBrowser's QPalette::Link / LinkVisited were
themed in ANTS-1239; the QDialog window, the QListWidget
TOC sidebar, and the viewer's Base / Text roles all
fell through to Qt's default dark palette. So the dialog
looked greyer than the rest of the app (e.g. against the
Tokyo Night / Dracula / One Dark navy backgrounds). Fixed
by applying bgPrimary to QPalette::Window / Base
on all three widgets, textPrimary to
WindowText/Text, and accent to the TOC's
Highlight so the selection indicator stays visible.<div>
block elements with their own QPalette::Base background
frame — it does not visually inherit from the parent
<div>'s background CSS. So the renderer's
<div class="rm-body"> wrapper inside each
<div class="rm-card"> painted bgPrimary over the
card's bgSecondary background, creating a visible
colour break at the divider. Fixed by emitting body
<p> paragraphs directly as children of the card (no
wrapping div) — <p> doesn't create a separate
background frame, so each paragraph paints over the
card's bgSecondary and the visual continuity is
restored. First body paragraph carries
class="rm-body-first" (dotted divider + extra
padding-top); subsequent lines carry
class="rm-body-line" (indent only). Regression-locked
by spec INV-15 + INV-16 in
tests/features/roadmap_dialog_cards/spec.md.<a> text on dark themes. Two bugs surfaced by user
test-drive of 0.7.83's card renderer:### Performance h3s (under 0.7.0,
0.8.0, and Beyond 1.0) all slugged to performance, so
expanding one expanded all three, and bySection["performance"]
pooled bullets from every Performance section into a single
bucket. Fixed by tracking a seen set across each walk and
appending -2, -3, … to repeat slugs (uniqueSlug helper).
Both parseBullets and renderCardsHtml walk the same
sourceText in the same order, so their slug sequences agree
and bySection[slug] keys match the URLs emitted into the
click anchors. Same fix covers other duplicate h3s in the
file (Platform, Cross-cutting themes, Security,
Tier 2 — hardening sweep, …).<a> foreground using the widget's QPalette::Link
role, ignoring the inline <style> block's a{color:…} rule
and not propagating through color:inherit. On Qt 6's default
palette that role renders as near-black on most dark themes,
so the chevron + heading text in each section header
disappeared. Fixed by setting QPalette::Link /
QPalette::LinkVisited on the QTextBrowser to the active
theme's textPrimary, plus emitting explicit
color:<textPrimary> on .rm-section-toggle /
.rm-section-title (replacing the prior color:inherit) as
belt-and-braces. Regression-locked by spec INV-13/14 in
tests/features/roadmap_dialog_cards/spec.md.Theme: RoadmapDialog v2 ships — card-style rendering for glance-friendly scanning of what's done, in-progress, and outstanding. Live-iterated visual polish (status counts lead section titles, parent-h2 breadcrumb on h3 sections, Table-of-Contents h2 suppressed inside the content pane, section headings + chevron both clickable) plus a Wayland Copy-from-roadmap fix.
⚙ implement, 🐛 fix, etc.), a one-sentence summary, and a meta row carrying the
ID and the shipped date — instead of a wall of prose. Sections
start collapsed with status-count chips leading the title (✅ 4 🚧 1 Performance) so a partially-sighted scan reads the counts
before parsing the name. Click the chevron or the heading
text itself to expand/collapse. Non-Full tabs (History /
Current / Next / Far Future) now strip prose intros and empty
sections so you only see cards relevant to that tab. h3
sub-sections show their parent h2 as a breadcrumb (Performance · 0.7.0 — shell integration) to disambiguate orphan headings that
recur under multiple releases. The Table of Contents h2 is
suppressed inside the content pane — the QListWidget sidebar
already serves that role. Per-item + per-section expand state
persist via four new Config::roadmap* keys. New optional
Layman: line in any roadmap bullet body — when present, shown
on the card face instead of the bold headline. Two new footer
buttons: Refresh (F5) and Reset View (two-click confirm).
Shipped scope is presentation-only; the original "tagged-text
format v2" portion of the charter was deferred (see §1156
sub-(2) in ROADMAP for the spin-out). All existing
roadmap_viewer* / roadmap_kind_facets /
remote_control_roadmap_query tests stay green — renderHtml /
parseBullets were extended additively, not replaced; the new
card renderer is a sibling helper named renderCardsHtml. New
feature test roadmap_dialog_cards wired into the existing
test_dialogs bundle (no new add_executable per ANTS-1217).
Source: user-2026-05-02 (original) + user-2026-05-11 (reframe).
([ANTS-1154])docs/standards/roadmap-format.md §3.5 now documents the
optional Layman: <one-sentence>. line. Additive on v1 — older
parsers ignore it. No format-version pragma bump. Authors add
Layman: lines opportunistically as they touch bullets; the
bold headline remains the fallback when absent. ([ANTS-1154])QDialogButtonBox::Close in 0.7.49
— the auto-popup attaches to a surface the compositor never
grants input to. Switched to Qt::CustomContextMenu with a
dialog-parented QMenu so the popup attaches to a stable xdg
surface; Copy + Select All now work and outside-click dismisses.
([ANTS-1154])Theme: live-repro fix bundle from the 2026-05-10 session — two
real bugs surfaced when the user /compact'd a Claude Code session,
/exit'd, then "Continue previous coding session"'d back in. The
Tasks chip kept showing five stale pre-compact tasks (ANTS-1224, a
parser miss on the isCompactSummary checkpoint) and the bottom-bar
Claude: status chip stayed hidden until a tab-switch (ANTS-1225, a
PID-replacement gate too narrow to catch a Claude exit + respawn
inside the 2-second poll window). Plus the session-isolation wiring
lock-in (ANTS-1219) — source-grep contract that catches future drift
in the resolver → tracker rebind path before the user does.
isCompactSummary
checkpoint, so pre-compact tasks survived into post-claude --resume
state. Layman: after you /compact'd a Claude session, exited,
and chose "Continue previous coding session", the Tasks chip kept
showing the todos from before the compact — a phantom list of work
Claude already considered closed. Root cause: parseTranscript
walked every JSONL event but never observed the isCompactSummary: true checkpoint, so any TaskCreate with no terminal TaskUpdate
from the pre-compact phase counted forever. Fix: inside the per-line
loop right after the sidechain filter, treat isCompactSummary:true
as a state-reset — clear out, idxByToolUseId, and sawTodoWrite
before the type-dispatch, so the checkpoint event itself contributes
zero entries. Multiple checkpoints converge on "state after the
final one" by induction. Zero new persistent state; all four new
test cases wired into the existing test_claude bundle (no fresh
exe, mindful of the ANTS-1217 build-OOM guardrails). Source:
user-2026-05-10 live reproduction. ([ANTS-1224])/compact + /exit + "Continue previous coding session" until a
tab-switch nudged it. Layman: the bottom-bar Claude: <state>
chip disappeared whenever a new Claude process replaced the one
that just exited inside the 2-second poll window, and only came
back when you switched tabs and back. Root cause:
pollClaudeProcess rebound the transcript only when
m_claudePid == 0, which catches first-launch but not live PID
replacement — between two polls a Claude could exit and a new one
spawn, leaving m_claudePid holding the stale dead PID and the
rebind branch never firing. Tab-switching worked by accident
because setShellPid zeroes m_claudePid as a side-effect. Fix:
generalise the gate to m_claudePid != foundPid, which catches
both initial detection and live replacement uniformly. Diagnostic
signature in the debug log (procStartMs=0 for every
sessionPathForCwd/result: line on the stale PID) explains why
the tasks chip kept working while the status indicator was
broken — the freshness resolver degrades to recency-only mode and
still finds the right JSONL, but m_claudePid correctness is what
the status indicator depends on. Zero new persistent state. Source:
user-2026-05-10 live reproduction, same session as ANTS-1224.
([ANTS-1225])mainwindow.cpp /
claudestatuswidgets.cpp surfaces as a red build rather than a
user-visible regression. New
tests/features/claude_task_list_session_isolation/ (source-grep,
no GUI, wired into the existing test_claude bundle) anchors six
// ANTS-1219-INV-N comments to load-bearing call sites — same
shape as the ANTS-1161 per-tab gate test that locked in the
session-id discriminant. Source: user-2026-05-10. ([ANTS-1219])0.7.81 — tasks-chip semantics polish + About-dialog build context
Theme: small follow-ups to the 0.7.80 user-feedback batch — a two-fix bundle that finishes the Tasks-chip semantics arc started in ANTS-1216, plus a long-requested About-dialog enhancement that surfaces enough build context to triage bug reports without a "what build are you on?" round-trip.
Build: line with the build date, build type (Release/Debug/RelWithDebInfo), short git commit (or unknown for tarball builds), and the compiler used (GCC or Clang with version). Implementation: a CMake configure_file() block at the top of CMakeLists.txt writes build/generated/build_info.h at configure time; the compiler ID is detected at compile time inside aboutdialogs.cpp so it reflects the actual toolchain (env-var overrides, distcc, distro alternative chains) rather than what CMake found. Reconfigure (cmake build/) to refresh the SHA between commits.☰ X/Y chip used to mean "X tasks left out of Y" (counted down) and stayed lit while a single in-flight Claude task remained. It now means "X done out of Y" (counts up, like every other progress display in the app) and disappears as soon as nothing is waiting on you. Two bundled changes: (1) ClaudeTaskListTracker::unfinishedCount() now counts only pending tasks — a task Claude is actively working on is Claude's problem, not yours, so it no longer keeps the chip visible. (2) The chip's numerator is now total - unfinished instead of unfinished, so X/Y reads like a GitHub-PR-checks progress bar. Pre-existing unfinished <= 0 hide branch (from ANTS-1216) still fires — combined effect: chip cleanly hides at 100%, monotonically increases as tasks complete, and never surfaces a single in-flight task as actionable.CI is now building the AppImage and other release artifacts via release.yml; they'll attach to this release page once the workflow completes (typically a few minutes on the public-repo Linux runners).
Theme: post-0.7.79 follow-ups + small high-signal user-feedback
fixes plus a build-OOM cure. Three notification/UI false-positives
that made the desktop feel noisy in normal use; a per-tab Claude: <state> widget bug that mis-attributed state across tabs; two
refactors carried over from the 0.7.78 indie-review #2 sweep that
finally land here (setupMenus per-menu carve-out and the Tier 4
Qt6 idiom polish); and ANTS-1217 — src/*.cpp refactored into
STATIC libraries, build-time guardrails (ccache /
CMAKE_OPTIMIZE_DEPENDENCIES / JOB_POOLS cap / Debug
-gsplit-dwarf) so a cmake --build build -j$(nproc) cannot OOM a
32 GiB host any more. Full ANTS_TESTS=ON build now peaks at
526 MiB RSS (was >32 GiB OOM). 144/144 ctests green; 1/1 new
claude_status_bar_per_tab feature test green.
onTabChanged; the
first call at startup saw shellPid() == 0 and shellCwd() == ""
(the shell hadn't started yet), the widgets hid, and nothing
re-fired until a tab switch. Same shape as the
refreshBgTasksButton fix from 0.7.49. Fix: wire both
refreshRoadmapButton and refreshRepoVisibility to the 2 s
m_statusTimer alongside the existing review/bg-tasks/tasks
refreshes (mainwindow.cpp:711-712), so a stale-hidden state
recovers within one tick of shellCwd() resolving. New feature
test tests/features/roadmap_status_bar_refresh/ locks both
widgets in via source-grep on the timer connections. This is
ANTS-1160 Phase 2 (status-bar bug fixes); the broader Roadmap
dialog redesign remains in progress (see ROADMAP § Roadmap
dialog redesign + format spec v2).Claude: <state> status-bar widget showed
another tab's state. Layman: with two Claude Code tabs open,
the bottom-of-window status text could read "Claude: bash" on
a tab whose Claude was actually thinking, because a sibling tab
running a Bash tool spammed its hook events into the singleton's
state. The per-tab dot indicator was already correct (it derives
state from each shell's own transcript). Root cause: the hook
server is one Unix socket shared across every Claude under any
tab, and ClaudeIntegration::processHookEvent mutated
m_state / m_currentTool and emitted stateChanged for every
inbound event regardless of session_id. Fix: gate every
state-mutating branch on the event's session_id matching the
basename of the focused tab's transcript path
(m_transcriptPath). PermissionRequest stays ungated — its
slot routes per-tab via lastHookSessionId(). Pre-poll tolerance
built into the gate so the bootstrap SessionStart (which fires
before pollClaudeProcess resolves the transcript) still wires
m_activeSessionId. New 4-invariant feature test
claude_status_bar_per_tab (verified fails I1 against pre-fix
code via the gate-neutralization repro).notify-send repeating from
a system-resource watchdog tab, breaking focus during normal
work. Root cause: terminalwidget.cpp:checkIdleNotification()
fired for any burst-then-idle pattern regardless of exit
status. Fix: gate on m_grid->lastExitCode() != 0 (OSC 133 D
exit-code). Successful commands and shells without OSC 133
integration stay silent; non-zero exits raise a
dialog-warning-iconed "Command Failed" notification with the
exit code in the body..directory (KDE Dolphin
metadata). Root cause: mainwindow.cpp:refreshReviewButton()
treated any non-header git status --porcelain=v1 -b line as
"dirty," including untracked (??) lines. Untracked files
don't appear in git diff, so clicking the button opened an
empty diff viewer. Fix: skip ?? lines in the dirty count.
Common false positives now silenced: .directory, IDE caches,
swap files, build artefacts that aren't gitignored. Once the
user git adds an untracked file it becomes A and counts
again.ClaudeTaskListTracker::unfinishedCount() used
status != "completed", so a deleted task counted as
unfinished — the chip showed "1" when the dialog showed zero.
Header-stated contract was "pending + in_progress";
implementation drifted. Fix: count only pending +
in_progress. (2) refreshTasksButton only hid the chip when
total <= 0, so a fully-done task list kept the chip visible
at "☰ 0/N", reading as actionable chrome. Fix: also hide when
unfinished <= 0.setupMenus 947-LoC monolith carved into seven
per-menu helpers + About dialogs moved to their own TU. Phase
A (in-place): setupMenus() is now an orchestrator of
setupFileMenu, setupEditMenu, setupViewMenu,
setupSplitMenu, setupToolsMenu, setupSettingsMenu,
setupHelpMenu. Phase B (partial): About-Ants and About-Qt
carved to src/aboutdialogs.{cpp,h} as namespaced free
functions (AboutDialogs::showAboutAnts(parent) /
showAboutQt(parent)) sharing an internal makeAboutDialog
helper. The Wayland modal-exec workaround comment (QTBUG-79126
/ QTBUG-90005) survives in the new TU verbatim. Deferred
(re-evaluate when surrounding pressure increases):
showSnippetsDialog and checkForUpdates stay on MainWindow
— both have tight coupling that would force threading
Config& + terminal-getter callbacks through free-function
signatures for marginal win.pluginmanager.cpp setRecentOutput / setCwd:
m_engines.values() → std::as_const(m_engines) direct
value-iteration; snapshot sites in unloadAll/fireEvent
intentionally keep .values() per ANTS-1173 UAF defense. (b)
mainwindow.cpp:4915,4955:
currentDateTime().toMSecsSinceEpoch() →
currentMSecsSinceEpoch(). (c) featurecoverage.cpp:392:
unused QDir projectDir deleted. (d) titlebar.h:19:
QColor("#e74856") → QColor::fromRgb(0xe74856). (e) Chained
.arg(int).arg(int) left as-is — idiomatic for ints.tests/features/ had
grown to 141 standalone add_executable(test_*) blocks, each
carrying its own MOC autogen and its own copy of the src/*.cpp
files it linked. src/mainwindow.cpp was compiled 42 times
per build. Default cmake --build build -j$(nproc) peaked above
RAM+swap on a 32 GiB host and tripped earlyoom — losing the
linker mid-write. Phase 0: flipped ANTS_TESTS default to OFF,
added Qt aggregate-header PCH (<QtCore/QtCore>, <QtGui/QtGui>,
<QtWidgets/QtWidgets>) on the main exe. Phase 1: refactored
src/*.cpp into 6 mandatory + 2 conditional STATIC libraries
(ants_core_lib / ants_vt_lib / ants_chrome_lib /
ants_claude_lib / ants_audit_lib / ants_dialogs_lib +
ants_lua_lib if LUA_FOUND + ants_helper_lib if
ANTS_ENABLE_HELPER_CLI); each src/*.cpp now compiles once
for the whole build. Layered three belt-and-suspenders
optimizations: (i) ccache as CMAKE_CXX_COMPILER_LAUNCHER
(cache hits skip cc1plus entirely); (ii)
CMAKE_OPTIMIZE_DEPENDENCIES=ON (Ninja skips relinking dependents
whose lib interface didn't change); (iii) JOB_POOLS cap
(compile_pool=nproc-2, link_pool=2) so a future regression
cannot OOM the host even on a naked -j$(nproc). Plus
-gsplit-dwarf on Debug builds — linker no longer relocates
hundreds of MiB of debug strings. Phase 1 verified on the
32 GiB host: full ANTS_TESTS=ON build at 526 MiB peak RSS in
1m36s wall (was >32 GiB OOM); 144/144 ctests green. Spec at
docs/specs/ANTS-1217.md; Phases 2–7 (test-bundle consolidation
via GoogleTest + workstation preset + tools/safe-build.sh)
pending.Theme: two bug-fix passes back-to-back. Pass 1 (morning,
user-driven) chased a cluster of user-visible scrollback /
output-pile / spinner-duplication / app-exit-crash reports —
8 ANTS items (1187..1194), one of which (ANTS-1194) turned out
to be the structural root cause of three of the others.
Pass 2 (afternoon, /indie-review #3) ran a scoped 4-lane
review on the 8.2k-LoC terminalgrid.cpp + terminalwidget.cpp
surface — the densest VT/render code in the repo — and shipped
17/17 calibrated findings (ANTS-1195..1213) in one session:
4 calibrated CRITICAL, 7 HIGH, 6 MEDIUM, INFO swarm. Headline
patterns: doc-rot at scale (4 of 4 lanes flagged comments that
survived a refactor and now lie about current code), and a
recurring "missing soft-reset / partial parse" shape — ANTS-1194
(scroll region didn't grow on resize), ANTS-1196 (DECSTR soft-
reset handler missing entirely), ANTS-1197 (OSC 9;4 short-payload
misclassified as a notification with body "4"). 4 new feature
tests added (decstr_soft_reset 8 INVs, osc9_progress_disambiguator
6 INVs, styled_font_kerning_off 3 INVs, wide_char_resize
2 INVs) plus scroll_region_growth_on_resize 7 INVs from Pass 1
and contract catch-up updates to debuglog_perms (post-ANTS-1190)
and pty_dtor_off_main_thread (post-ANTS-1189). 141/141 ctests
green.
scrollRegionTop() / scrollRegionBottom() accessors so a
user with a misbehaving TUI has a one-click recovery without
needing to know about \e[r._int_free_chunk SIGABRT on every app exit.
Layman: closing Ants Terminal would print free(): invalid pointer and dump core every single time, in front of three
separate users with identical fingerprints today. Root cause:
Pty::~Pty ran a detached std::thread for the SIGTERM →
SIGKILL escalation. The 0.7.0 VtStream refactor moved ~Pty
itself onto the parse-worker thread, making the 0.7.33 detach-
for-GUI-responsiveness redundant — but the detached worker
outlived the destructor and raced main-thread Qt teardown via
glibc malloc arenas. Escalation now runs synchronously on the
parse worker (1.5 s budget, within m_parseThread->wait(2000)).
The post-SIGKILL reap is now also bounded — pre-fix
waitpid(_, _, 0) could hang on uninterruptible-sleep children
and trip Qt's "destroying a running QThread aborts" assertion,
worse than the detach race. Test rewritten with 12 invariants
on the synchronous-on-worker contract; verified fails 7/12
against pre-fix code.DebugLog::clear() truncated the file but didn't reopen the
category file handles, so subsequent appends went to a deleted
inode. Fix reopens the file post-clear. The debuglog_perms
feature test had its I2 invariant updated to lock the new
contract (post-fix slipped through yesterday's ship without the
full features sweep — caught and fixed under this section)./home/me/my_project and the BG-tasks chip in
the status bar would silently disappear; switch to a project
with no underscores in the path and it returned. Closes the
ANTS-1052 regression. Root cause: encodeProjectPath (which
derives the ~/.claude/projects/<encoded-cwd>/ lookup key)
collapsed every other separator but left underscores intact,
so the on-disk transcript directory and the in-memory key
diverged. Now collapses underscores too, matching Claude Code's
own canonical encoding.TerminalGrid::resize() only
clamped the scroll region; it never widened it on grow.
Tab opens at 24 rows → maximise to 60 → m_scrollBottom stays
at 23 (clamp(23, 0, 59) = 23) → new output piles into rows
0-23, bottom 36 stay empty. ANTS-1130's clamp was correct for
shrink (preserves explicit DECSTBM from tmux/mc/less) but
silently wrong for grow when no DECSTBM was ever set. Fix
captures liveWasFullScreen / savedWasFullScreen before
m_rows updates, then conditionally widens on full-screen and
clamps on explicit partial — matches xterm semantics (default
region tracks the screen, explicit DECSTBM is preserved). New
scroll_region_growth_on_resize feature test, 7 invariants
(default-grow, default-shrink, explicit-preserved-on-grow,
explicit-clamped-on-shrink-below-bottom, alt-default-grows,
alt-explicit-preserved, sequential grow/shrink); verified fails
5/7 against pre-fix. Closes ANTS-1187, ANTS-1193 (pipeline was
innocent), and likely ANTS-1188.CSI ! p) soft-reset handler was
missing. Layman: vim, neovim, mc, htop and other conformant
TUIs emit DECSTR on startup as a defensive recovery from
hostile prior shell state. Pre-fix Ants silently dropped the
sequence — same shape as ANTS-1187 (a stale scroll region
traps the user) but for the broader contract. Per xterm
ctlseqs / VT220 reference, DECSTR is a subset of RIS: it
resets scroll region (full screen), origin mode, autowrap,
cursor visibility, SGR and saved-cursor — but does NOT wipe
screen, scrollback, hyperlinks, OSC 133 prompts, integration
callbacks, the OSC 133 HMAC key, the kitty keyboard stack, or
alt-screen state. (ESC c / RIS keeps doing the hard reset.)
New decstr_soft_reset feature test, 8 INVs on the spec-
defined reset surface plus three "DECSTR is soft, NOT hard"
preservation invariants; fails 6/8 against pre-fix. Added
originMode() and autoWrap() accessors needed by the test.OSC 9;4 "remove
progress" sequence (the minimal form, payload literally 9;4)
used to pop a desktop notification with body "4" instead of
clearing the progress indicator. Pre-fix disambiguator
required payload.size() >= semi + 3 AND payload[semi+2] == ';', so the 3-character minimal form fell through to the
notification path. New predicate accepts 9;4 (state=0,
percent=0 defaults) AND keeps rejecting 9;42 / 9;4Hello.
6 INVs in osc9_progress_disambiguator; fails 3/6 pre-fix.
Open question logged in spec: a notification body literally
starting with 4; (e.g. 9;4;30 PM meeting) still
misclassifies as progress — fundamental wire-format ambiguity,
flagged for future stricter range-check.setBoldFontFamily /
setItalicFontFamily / setBoldItalicFontFamily rebuilt the
per-style QFont but bypassed updateFontMetrics, inheriting
Qt's default kerning-enabled state. Three-line fix
(setKerning(false) on each); styled_font_kerning_off
source-grep test, 3 INVs.recalcGridSize cleared
the snapshot but waited for the next onVtBatch to re-capture.
Now re-captures immediately when m_syncOutputActive. Spec
ANTS-1148 §H1 had named this exact gap.m_syncOutputActive flag and cleared the snapshot — but the
grid's m_synchronizedOutput latch stayed set, so the next
batch's pre-scan re-armed the flag and restarted the timer.
New TerminalGrid::forceClearSynchronizedOutput() is now
called from the safety timer after clearScreenSnapshot();
malformed BSU now escapes on first timer fire.c / p / s / q / 0-7 per xterm) was parsed and
discarded; every clipboard write landed on the system
clipboard. Now plumbed through the existing OSC52:
sentinel envelope as OSC52:<selChar>:<bytes>. Selection
collapse: c → system clipboard, p / s → primary; empty
/ unknown defaults to c (modern xterm.js / kitty behaviour
— primary's middle-click paste is surprising when not
explicitly requested).id= parser used naive
find("id=") which matched embedded id= inside a value
(x=id=trick produced id="trick") — switched to a proper
per-key scan; (c) URI scheme allowlist was applied on open
only, defense-in-depth gap if any future path mutates
m_hyperlinkUri outside the open validator. Refactor
extracted pushHyperlinkSpansForActive() so close and
reopen-close paths share the multi-row emit + scheme re-check.Cell::isWideChar / isWideCont when splitting logical
lines. Fix: if the chunk would split a wide-char pair AND
the chunk doesn't end at total, decrement chunk by 1 to push
the lead into the next iteration with its continuation. New
wide_char_resize test, 2 INVs targeting the dangerous
mid-pair boundary; fails 4/4 pre-fix.m_cols * m_rows to m_cols * 2 — a malicious payload
\e[16000b\e[16000b... could pin the parser thread on a
single byte for a 16,000× CPU multiplier; real REP use
cases (tput rep, less border drawing) need at most
m_cols, so 2 * m_cols is generous without giving an
attacker a pin. (b) OSC 133 marker validated before HMAC
verification: with m_osc133Key set, every received
OSC 133;X;ahmac=<hex> used to fire a SHA-256 round even
for unknown markers — only A/B/C/D are spec'd, drop early.
(c) inputMethodQuery::ImCursorRectangle now uses
effectiveCursorRow/Col instead of the live grid cursor,
so during a sync-output BSU block the IME panel doesn't
jump anywhere mid-redraw.ECH (CSI X) honours BCE and clears
non-bg attrs. Layman: erase-character used to leave bold
/ italic / underline / inverse active on the resulting
space, and could leave a stranded wide-char marker. Per
xterm, ECH should clear all non-bg/fg attrs and set bg via
the Background-Color Erase state — same shape as clearRow.
Now zeros CellAttrs first, then assigns default fg +
eraseBg(), and resets isWideChar / isWideCont.refreshTasksButton diagnostic logging.
Added ANTS_LOG(Network, …) instrumentation around the
Claude-tasks status-bar button refresh path so the next
user-reported "the chip should be visible but isn't" report
has a transcript to triage from. No user-visible behaviour
change. Pairs with the post-mortem of ANTS-1192.m_fallbackFont / m_nerdFallbackFont
removed. Indie-review #3 Lane D Critical: the emoji/CJK
and Nerd Font fallback fonts were probed at construction,
resized on font-size change, but never consulted by any
render path (zero non-declaration grep hits). Today's
emoji/CJK/Powerline rendering relies entirely on Qt's
implicit FontConfig substitution. Decision: delete (vs
wire) — FontConfig substitution works on Linux/macOS and
the flag pattern (m_hasFallbackFont / m_hasNerdFallback)
was the wrong shape for any future explicit per-codepoint
route. Net delta: −34 lines of code that never affected any
render frame. Comment in the constructor names the deletionliveWasFullScreen / savedWasFullScreen
rename + m_altScrollBottom default init. Two Lane A
findings bundled. (a) The primaryWasFullScreen /
altWasFullScreen names introduced by ANTS-1194 read
backwards during alt-screen mode (the grid's
m_scrollTop/Bottom are the LIVE values for whichever
buffer is active; m_altScroll* are the SAVED values for
the OTHER buffer). Behaviorally correct but misled the
next reader — renamed to live / saved with explanatory
comment. (b) m_altScrollBottom defaulted to 0 in the
header until first alt-screen entry; symptom-free today
but asymmetric with the constructor's primary init and a
foot-gun for any future caller. Now initialised to
m_rows - 1 in the ctor body alongside the primary slot.m_paintLayout.clearFormats() discipline.
The mutable QTextLayout member is reused across runs and
paints; setText() resets glyph state but not additional-
format ranges installed via setFormats(). No caller sets
formats today, so this is a no-op in practice — but any
future addition that calls setFormats({...}) once would
leak that formatting into every subsequent run for the
lifetime of the widget. Bug shape is invisible to tests
and brutal to track down; one-line defensive
clearFormats() at the top of the runs loop with a
comment naming the contract.terminalwidget.cpp. Two
stale comments cited the retired QOpenGLWidget+FBO
architecture as the reason for current code behaviour
(paintEvent background fill rationale at line 629; resize
base-class call rationale at line 1934). Updated to describe
the current QWidget backing-store + translucent-compositor
reasoning. The history comments at lines 80 / 89 (which
document the 0.7.4 refactor itself) stay — those are
deliberate institutional memory, not stale rationale.QFonts as
members — pre-fix paintEvent constructed 4 fresh fonts
per frame (badge × 4pt-bold; cmd-mark labels at 2 sites;
quick-select labels; perf overlay), every one triggering a
FontConfig family lookup; now built once in
updateFontMetrics() alongside the styled variants and
reused (m_badgeFont, m_smallFont, m_quickSelectFont,
m_perfFont, all inheriting setKerning(false)); (b)
ASCII fast-path for QString::fromUcs4 in the hot run
loop — >90% of TUI content is <0x80, where the
UTF-32→UTF-16 transcode is pure overhead; pre-allocated
QString + QChar copy is roughly 2-3× faster on ASCII
runs, falls through to fromUcs4 only when a non-ASCII
codepoint is detected; (c) m_freeCellBuffers.clear()
in TerminalGrid::resize() now guarded by cols != oldCols — pre-fix cleared unconditionally, so a pure
rows-resize lost cached cell allocations for no reason.m_grid->processAction
loop: if any installed callback synchronously pumped the Qt
event loop, a queued repaint could fire mid-batch and the
snapshot fallback (cellAtGlobal → live grid) would tear.
Audit walked all 8 installed callbacks (response, bell,
notify, progress, commandFinished, userVar, osc133Forgery,
lineCompletion) — none synchronously pump. Closes as INFO
with no code change beyond a defensive contract comment at
the top of onVtBatch: future callbacks that want
synchronous GUI dispatch (e.g. modal dialogs) MUST queue
via QMetaObject::invokeMethod with Qt::QueuedConnection.setLineWidth fudge replaced + setFontFamily
order hazard documented. (a) tline.setLineWidth in the
run loop used runText.length() * m_cellWidth * 2. UTF-16
length conflated with glyph cells (CJK is 1 unit / 2 cells;
emoji surrogate is 2 units / 1 cell), and the * 2 was a
fudge to mask both directions. Replaced with
qreal(INT_MAX) — the correct semantic for monospace
cell-grid rendering ("do not break this run on width")
regardless of run length. (b) setFontFamily clobbers
per-style family overrides set via setBoldFontFamily /
setItalicFontFamily / setBoldItalicFontFamily.
Documented as an order hazard (callers that want both basesetFontFamily FIRST,
then the per-style setters) rather than fixing the
behaviour — the propagate-to-all default matches the
"I picked one font for everything" user mental model.handleCsi claimed "DECRQSS (DCS $q) is not implemented
to prevent echoing attacker-controlled data" while
handleDcs IS implemented and replies to DECSTBM r / SGR m
/ DECSCA " q / DECSCUSR SP q. Threat-model calibration
confirmed the implementation is safe (replies are terminal-
controlled — 4 fixed format strings, no attacker bytes
echoed back, so the CVE-2003-0063 echo-back class doesn't
apply); the comment was misleading, not the implementation.
Comment updated to match reality. The CSI 20t/21t
XTWINOPS portion of the original note stays — that one IS
still absent and the CVE rationale still holds.Theme: independent-review sweep #2 — pre-0.8.0 audit + multi-
agent code review, briefed cold against external standards
(ECMA-48, xterm ctlseqs, OWASP LLM Top 10, POSIX forkpty(3),
SARIF v2.1.0, Lua 5.4 sandbox, RFC 8259, freedesktop). The
review surfaced 4 calibrated CRITICAL + 22 HIGH after threat-
model calibration; this release ships 22 of 23 (ANTS-1181 — a
1500-LoC mainwindow.cpp carve-out — deferred to its own
session). Headline pattern: ANTS-1163's just-fixed dialog-
staleness bug recurred in 5+ more cached dialogs, exactly the
"self-graded homework" trap that motivated the sweep — bundled
into ANTS-1168 + a structural lint guard. Plus the rolled-over
ANTS-1163 fix from 2026-05-07. 137/137 ctests, gitleaks 0,
semgrep 0.
Pty::write logged a 60-byte percent-encoded
preview when ANTS_DEBUG=pty was active, with NO redaction.
Short tokens (ghp_…, AKIA…, sub-60-char passwords pasted
at sudo) landed verbatim in ~/.local/share/ants-terminal/ debug.log. Fix: route the slice through SecretRedact:: scrub() before percent-encoding. Pairs with ANTS-1170.a=d,d=a wiped Sixel + iTerm2 images
cross-protocol. Kitty's "delete all images" cleared the
shared m_inlineImages vector, erasing every Sixel and
iTerm2 image the terminal had displayed. Exploitable from
untrusted PTY output as cross-protocol visual-context
redaction. New InlineImage::Origin enum; delete-all now
removes only Kitty-tagged entries via std::remove_if +
recomputeImageBudget.tailFile canonicalizes against
/tmp/claude-$UID/ + lstat S_ISREG (LLM01-reachable);
6 path setters reject >4096 chars or NUL-embedded;
setBackgroundImage peeks via QImageReader::size() against
MAX_IMAGE_DIM=4096 (was vulnerable to a 50000×50000 PNG →
~10 GB alloc); Kitty APC c/r clamped to MAX_IMAGE_DIM;
OSC 8 trigger URI capped at 4096 chars; parseTranscriptTail
falls back to last-complete-record on the 4 MiB cap instead
of returning empty; runClient 1 MiB receive cap mirrors
server; --remote send-text 1 MiB stdin cap. MAX_IMAGE_DIM
promoted to a public static so non-grid intake paths share it.ANTS_DEBUG opt-in gate + log rotation.
Honoring ANTS_DEBUG now prints a one-line stderr banner
("Ants: debug log active (ANTS_DEBUG=…) → /path") so a
forgotten env var inherited from a CI image / .envrc is
visible at startup. debug.log rotates to debug.log.1
when it exceeds 10 MiB, preventing unbounded secret-capture
across multi-day forgotten sessions.LUA_MASKCOUNT-only doesn't fire inside pure-C
Lua calls (string.gsub, table.sort); a plugin author
feeding PTY output into data:gsub("(.-)+", …) could freeze
the UI. New startPcallBudget() sets a 1500 ms wall-clock
deadline; the hook now combines LUA_MASKLINE | LUA_MASKCOUNT
@ 100K ops with a deadline check on every callback. Closes
the only documented Lua sandbox-escape path.QRegularExpression::isValid() + AuditEngine:: isCatastrophicRegex() at save (catastrophic patterns silently
dropped). AI key field gains Qt::ImhSensitiveData | ImhHiddenText | ImhNoAutoUppercase | ImhNoPredictiveText
so virtual-keyboard / IME predictive caches don't capture
the value.AIza… (Google API keys, 39 chars) and
ya29.… (Google OAuth tokens). Closes a known gap in the
AI-context redaction pass.ClaudeIntegration::sessionPathForCwd picked the
newest .jsonl in ~/.claude/projects/<encoded-cwd>/ by
mtime alone. After a reboot → claude sequence, the
prior session's transcript outranked the new (empty)
transcript until the new session appended its first event —
surfacing the prior plan's TodoWrite tasks in the chip and
dialog. /clear was a no-op because the bug lived in file
selection, not in tracker state. Fix is two-layered: (a)
process-anchored identity filter — when the focused tab's
m_claudePid is known, candidates whose effective last-event
ms predates the Claude Code process start by more than 5 s
are dropped; (b) 24 h liveness floor — independent of (a),
candidates whose effective last-event ms is older than 24 h
are dropped. New static helpers processStartTimeMs(pid) +
lastEventTimestampMs(path); new overload
sessionPathForCwd(cwd, minLastEventMs, nowMs) with default
args preserving legacy newest-by-mtime behaviour. New feature
test claude_session_freshness (16 INVs).Ctrl+Shift+Up/Down configured bookmark
shortcut silently dead. Default next_bookmark /
prev_bookmark chord collided with the OSC 133 prompt-
navigation chord that TerminalWidget::keyPressEvent
intercepts before Qt dispatches QShortcuts; the bookmark
shortcut never fired whenever the terminal had focus. Default
changed to Ctrl+Alt+Up/Down.forkpty F_SETFL silent fall-through. If
F_GETFL or F_SETFL failed, Pty::start swallowed the
error and continued with a still-blocking master fd; a
spurious QSocketNotifier wakeup would freeze the GUI thread
inside read(). The failure path now logs and aborts
start().SshDialog form fields (added
clearForm() called on setBookmarks); AiDialog last-
error / in-flight reply (new resetTransient());
ClaudeAllowlistDialog rule input + validation hint;
ClaudeProjectsDialog first-show-not-refreshed (now
unconditional); ClaudeTranscriptDialog showing global
newest-by-mtime regardless of focused tab (new
setProjectFilter + recentSessionsForCwd);
pollClaudeProcess second site of unscoped global newest-
by-mtime (now scoped via /proc/<shellPid>/cwd →
sessionPathForCwd); setShellPid not clearing
m_changedFiles on PID change.handleCheckOutput reality (filter-then-cap; capping first
would surface junk findings while real ones get pushed out
behind).PluginManager::unloadAll snapshot before
iteration. Pre-fix code didn't snapshot m_engines.values(),
unlike fireEvent. An unload handler that re-entered the
event loop (status signal → palette repaint → keypress →
fireEvent) could dereference an already-deleteLater'd
engine — deterministic UAF window in dev/hot-reload mode.std::make_shared<QMetaObject::Connection> self-disconnect
patterns with Qt::SingleShotConnection (Qt 6.0+). Added a
transient m_paletteProxyHolder that gets replaced on every
rebuildCommandPalette() so proxy QActions no longer
accumulate as orphans on the MainWindow's child list across
plugin reloads / config refreshes.kEnvpCap=512 now emits
a qWarning (was the silent root of the user-reported "ants
sometimes opens a shell with no PATH" symptom on
500+-entry desktop sessions). New childUnreapedAtEof
signal disambiguates "child closed PTY but waitpid hasn't
reaped" from "child exited with -1, reaped"; legacy
finished(-1) continues to fire on the same code path."drift script killed by signal" error message;
QProcess::exitStatus's exit code is unspecified on
CrashExit per Qt 6 docs.setRoadmapActivePreset and
setRoadmapKindFilters now reject unknowns at write
(mirroring the getter's known-set filter), so future preset
renames don't leave zombie values stranded on disk. Failed
user-theme JSON files are rotated to <name>.json.corrupt. <epoch> so the user has a forensic copy on disk and the
next launch doesn't keep silently skipping it.ANTS_LOG(Network, "rc dispatch cmd=%s tab=%d text_bytes=%d", …) on every dispatch. Same-UID-attack
post-mortem now has a record. Payload bodies deliberately
excluded.fillRect coalescing. Contiguous
same-bg cells now produce one fillRect per run, mirroring
the existing TextRun aggregation. Fully-styled vim status
lines previously paid cols × per-frame fillRect calls;
hot-path frames now scale with run count, not cell count.findChildren<TerminalWidget*>() walks
collapsed. Replaced 12 sites in mainwindow.cpp with
liveTerminals() over a maintained QList<QPointer< TerminalWidget>>. The hottest path was the broadcast-mode
keystroke lambda — under 20 tabs × split panes that previously
triggered a tree walk per keystroke.config.json schema versioning. New
_schema: 1 stamp + migrate(int from, int to) placeholder
so future on-disk renames have a canonical home. Pre-stamp
configs (anything written by ≤ 0.7.77) read as v0 and
migrate-up on first load."Claude: <state>" (idle / thinking / tool use / bash /
planning / auditing / compacting / awaiting input) so Orca
/ Windows Narrator can announce state alongside the tab
title. The visual dot was previously invisible to assistive
surfaces.Theme: ANTS-1150 Phase 1 — UI / chrome state persistence. User
ask 2026-05-01: "Everything that makes sense to persist across
sessions, please make that persist." Six persisted UI choices that
previously reset on every launch — SettingsDialog last-active tab,
RoadmapDialog active preset + Kind facet checkbox set + 5 status-
emoji checkboxes, AuditDialog severity-filter pills + show-new-only
toggle. Six new Config getter/setter pairs follow the existing
storeIfChanged + atomic-write hygiene; round-trip identity tested
via a sandboxed-XDG_CONFIG_HOME QTemporaryDir lane. 15 ANTS-
1150-INV-N invariants in tests/features/ui_state_persistence/.
Cold-eyes spec review folded 2 CRITICAL + 5 HIGH + 6 MEDIUM + 4 LOW
before code; indie-review of the implementation folded 1 MEDIUM + 2
LOW (Custom-preset silent rewrite to Full + dead-defense
QSignalBlockers on tab/show-new restore — re-ordered connect-
before-restore so blockers actually defend). 132/132 ctests pass.
ANTS-1150 Phase 1 — six pieces of UI state now survive a restart. Phase-1 cohort, all driven by user ask 2026-05-01:
settings_dialog_last_tab).
Open Settings on whichever of the 9 tabs (General / Appearance /
Terminal / AI / Highlights / Triggers / Keybindings / Profiles /
Plugins) you closed it on, not always General. Index-keyed,
clamped at the call site against the dialog's tab count.roadmap_active_preset,
string: "full" / "history" / "current" / "next" /
"far_future" / "custom"). Re-opens to your last tab choice;
invalid stored values fall back to "full".roadmap_kind_filters,
sorted ASCII-codepoint array). Reverses the explicit ANTS-1106
"Kind filter resets on each open" decision per 2026-05-01 ask.
Setter sorts on write so two saves of the same set produce
byte-identical JSON (diff-friendly). Unknown Kind strings drop
silently on read — defends a future KindEntry rename.roadmap_status_filters, 5-key boolean object —
done/planned/in_progress/considered/current).
Restored only when the persisted preset is "custom" (named
presets ride applyPreset's canonical mask anyway). Empty
object on first read defaults each checkbox to true (the
.toBool(true) Qt contract on missing keys).audit_severity_filters,
5-key boolean object — blocker/critical/major/minor/info).
The MIN+INF-off post-noise-sweep state survives a re-run.audit_show_new_only). Restored at dialog ctor only when
m_hasBaseline is true (no point checking the box without a
baseline to compare against). Lazy-invalidate stickiness:
persisted bool stays through baseline-deletion gaps; next
saveBaseline re-honours the preference.Implementation hygiene: two-write-site invariant for
m_activePreset — applyPreset (named-preset path) and
onCheckboxToggled (Custom-divergence path) both call the new
RoadmapDialog::persistActivePreset helper, which uses a
switch on Preset for compiler -Wswitch-enum coverage of
any future enum addition. Lifted KindEntry table to file
scope as kKinds[] so the build loop and the persisted-Kind
restore iterate the same source-of-truth; new
m_kindCheckboxes (QHash<QString, QCheckBox *>) hash for
O(1) lookup during restore. AuditDialog ctor signature
changed to require Config * as a third arg — no
= nullptr default per the cold-eyes review (single call site
at mainwindow.cpp:1351 always has &m_config available; a
default would be dead scaffolding).
Config::setRoadmapKindFilters now sorts ASCII-codepoint-wise
on write (not at the call site). Centralizes the canonicalization
so future call sites can't accidentally write an unsorted list
and break diffability.0.7.75 — DEC mode 2026 sync output atomicity fix (ANTS-1148)
Theme: ANTS-1148 — DEC private mode 2026 (Synchronized Output /
BSU/ESU) atomicity fix. Pre-fix, onVtBatch suppressed only its own
update() during BSU but left the live grid mutated by
processAction; any other paint trigger (blinkCursor,
focusIn/Out, hover, selection, visual-bell flash) would call
paintEvent which read the live half-applied state and leak mid-
sync content to screen. The fix unifies sync output onto the
existing 0.6.33 frozen-screen snapshot machinery: pre-scan a
VtBatch for CSI ?2026h, capture the snapshot before the
processAction loop, route paintEvent's cursor reads through new
effectiveCursorRow/Col() accessors, and clear on ESU or via the
500 ms safety timer. Single TU touched (terminalwidget.{cpp,h}),
behavioural-correctness only — no LoC delta on mainwindow.cpp.
132/132 ctests pass.
ANTS-1148 — Synchronized Output (DEC mode 2026) is now
atomic against non-batch paint triggers. Pre-fix, BSU
(CSI ?2026h) suppressed onVtBatch's trailing update() but
did nothing about the live grid mutation underneath: any
paintEvent driven by blinkCursor (550 ms cadence),
focusInEvent / focusOutEvent, mouse-hover OSC 8 hyperlink
rollover, selection drag, or visual-bell flash would re-render
from the half-applied grid mid-BSU and leak partial state —
exactly the tearing the protocol exists to prevent. Fix folds
sync output onto the existing 0.6.33 frozen-screen snapshot
path: a new file-scope helper batchEntersSyncOutput pre-scans
the VtBatch for CSI ?2026h (CsiDispatch + intermediate ? +
finalChar h + 2026 in params), and onVtBatch captures the
snapshot under
(m_syncOutputActive || batchEntersSyncOutput(*batch)) && m_frozenScreenRows.empty() before processAction runs. The
left disjunct handles eviction-recovery — resize / RIS / alt-
screen toggle clear the snapshot mid-sync, and we need to
re-capture for the rest of the block. Cleanup runs at end-of-
batch when !m_syncOutputActive && m_scrollOffset == 0 && !m_frozenScreenRows.empty() (drops the prior wasSync && gate
that stranded same-batch BSU+ESU snapshots — cold-eyes C2); the
500 ms safety timer also clears on force-end. paintEvent's
three cursor-render sites (cursor draw, under-cursor glyph,
autocomplete ghost) route through new
effectiveCursorRow/Col() inline accessors that read
m_frozenCursorRow/Col when m_frozenScreenRows is non-empty
and live otherwise; the eight typing / IME / semantic-output
cursor reads (keyPressEvent, inputMethodQuery,
clickToMoveCursor, findMatchingBracket, toggleFoldAtCursor,
toggleBookmark, updateSuggestion, lastCommandOutput)
intentionally keep direct m_grid->cursorRow/Col() reads —
those paths want the real cursor, not the rendered one. Spec at
docs/specs/ANTS-1148.md cold-eyes-reviewed before code (10
findings folded — 2 CRITICAL, 3 HIGH, 3 MEDIUM, 2 LOW); new
feature test at tests/features/sync_output_snapshot/ locks
9 INVs by source-grep. Out-of-snapshot scope (intentional,
matches xterm/foot): cursor visibility / shape / blink
(DECSCUSR, DECTCEM) and inline images (Sixel / Kitty / iTerm2)
read live mid-BSU.
ANTS-1148 — updateScrollBar's frozen-snapshot predicate
extended to cover sync output. Indie-review HIGH on commit
9674e5a caught a missed call site: updateScrollBar at
terminalwidget.cpp:2793 still used the pre-1148 predicate
wantFrozen = (m_scrollOffset > 0). Closes spec transition
row 7 — scrolled-back-during-sync, user scrolls back to bottom
while BSU is still active: without the sync clause, the
snapshot is evicted and the next paint reads the half-applied
grid. Fix extends to (m_scrollOffset > 0) || m_syncOutputActive,
matching the unified-predicate pattern already in use at the
other three call sites (onVtBatch pre-scan, end-of-loop
cleanup, safety-timer slot). Test gains INV-4b locking the
disjunction at this third site (74395a6).
blinkCursor partial-update rect routes through
effectiveCursorRow/Col(). Pre-fix the blink invalidation
rect was computed from m_grid->cursorRow/Col() directly; under
sync the live cursor position can have advanced past the
frozen-rendered cell, so the partial-update rect wouldn't cover
the cell paintEvent actually draws and the cursor would
visually stall mid-BSU. Routing through the centralised
accessor locks the invalidation rect to the rendered cell — a
small consistency win that drops out of the snapshot
unification rather than a separate fix.Theme: two user-driven items landed back-to-back. ANTS-1158 —
status-bar surface for the focused tab's Claude Code task list,
replayed from the session JSONL transcript (per-tab, live-updated
on QFileSystemWatcher + 2 s coalesce, hidden when there's no
plan). ANTS-1159 — crash-safe session persistence: pre-fix code
called saveAllSessions() only from closeEvent, so any
abnormal termination (SIGSEGV / OOM-kill / power loss) discarded
every write since the last graceful shutdown. New 30 s
m_sessionSaveTimer + tab-create / tab-close / tab-reorder hooks
into a cheap new saveTabOrderOnly() slot bound the loss window
to ≤ event-loop latency for the tab list and ≤ 30 s for
scrollback. Three follow-up regression fixes (Task List dialog
refresh under atomic-rewrite; RoadMap button + GitHub repo-type
badge cold-launch visibility) ride along under ANTS-1160 P2,
plus a new docs/standards/status-bar.md contract to prevent
recurrence. 13 INVs for ANTS-1158, 9 INVs for ANTS-1159, 5 INVs
for ANTS-1160 P2; 136/136 ctests pass.
<unfinished>/<total> (in_progress + pending out of all). Click
opens ClaudeTaskListDialog with one row per task (☐ pending /
◐ in_progress / ✓ completed) plus a header
"<N> tasks — <Y> running, <Z> outstanding, <X> done". Hidden
when no plan exists for the focused tab. New
claudetasklist.{cpp,h} (ClaudeTaskListTracker + static
parseTranscript) and claudetasklistdialog.{cpp,h}; wired
through claudestatuswidgets (button + tooltip + retarget on
setTranscriptPath) and MainWindow::onTabChanged (per-tab
retarget mirrors m_bgTasks). Replay handles two transcript
shapes: Mode A (snapshot) — most-recent TodoWrite wins,
found by 64 KiB EOF block-walk so we don't scan the whole file
to grab a near-tail event; Mode B (incremental) — forward
walk over TaskCreate (paired with tool_result for the
allocated ID, mirroring claudebgtasks.cpp:222–243) +
TaskUpdate flips, 16 MiB tail cap. Filters out
isSidechain == true events (subagent tool calls inlined in
the parent transcript) and Task / Agent tool_uses with
subagent_type (sub-agent dispatch ≠ plan row). Dialog is
non-modal with plain QPushButton close — no
setModal(true), no QDialogButtonBox (per the
debug_wayland_modal_dialog memory: QTBUG-79126 drops button
clicks on KDE+KWin+Qt 6.11+frameless+translucent parent).
Cold-eyes spec review folded before code. 13
ANTS-1158-INV-N invariants in
tests/features/claude_task_list/.ANTS-1159 — session state now persists on a 30 s timer and on
every tab create / close / reorder, not only on closeEvent.
User reported 2026-05-02 that after a SIGSEGV, restart restored
zero tabs. Pre-fix, MainWindow::saveAllSessions() was wired
only to closeEvent (mainwindow.cpp:3798), so any abnormal
termination — SEGV, OOM-kill, power loss, kernel panic —
bypassed it and discarded every write since the last graceful
shutdown. ANTS-1141 had already plugged a related leak (the
destructive read on tab_order.txt); this closes the bigger
one. Fix shape (option B, user-approved): new
QTimer *m_sessionSaveTimer on MainWindow, 30 000 ms
autoreloading, timeout → existing saveAllSessions(). Started
in the ctor after restoreSessions() so the first save
doesn't race the restore; stopped in closeEvent before the
explicit save so it can't re-enter mid-shutdown. Three new tab
hooks call a cheap new saveTabOrderOnly() slot (just walks
m_tabWidget, builds the tabOrder QStringList + active
index, calls SessionManager::saveTabOrder — no scrollback
serialisation): from newTab() (post-addTab), from
performTabClose() (post-removeTab), and from
m_coloredTabBar's tabMoved signal. The same
sessionPersistence()-off and 5-s uptime-floor guards as
saveAllSessions apply (no on-disk overwrite during
half-restored startup). currentChanged is intentionally
not hooked — tab switches fire constantly and only update
activeIndex; the periodic timer covers it. Worst-case loss
window: ≤ event-loop latency for the tab list, ≤ 30 s for
scrollback. 9 ANTS-1159-INV-N invariants in
tests/features/crash_safe_session_persist/. Stale-binary
triage for the originating SEGV (/proc/<pid>/exe reported
(deleted) with binary mtime 29 min before crash; restart
from the fresh binary did not reproduce) is documented in the
spec body — not a code bug, but the persistence-loss it
exposed is.
Claude Code Task List dialog now refreshes when the
transcript file is rewritten (ANTS-1158 follow-up).
QFileSystemWatcher silently drops its inotify watch when the
watched file is replaced via atomic-rewrite — Claude Code's
transcript update pattern — and never re-arms, so the dialog
stayed pinned to whatever state was current when it opened.
Added ClaudeTaskListTracker::poll() with mtime-gated rescan
(QFileInfo::lastModified() compared against
m_lastRescanMtimeMs), called from
claudestatuswidgets::refreshTasksButton() which already ticks
on the 2 s status-bar timer. Same fix shape as
ClaudeBgTaskTracker::poll(). The existing 13
ANTS-1158-INV-N source-grep invariants still hold; no new test
needed.
RoadMap status-bar button + GitHub repo-type badge now
appear on cold launch instead of waiting for the first tab
switch (ANTS-1160 P2). Both widgets read shellCwd(), which
reads /proc/<shellPid()>/cwd; shellPid() is set by
startShell() AFTER newTab()'s setCurrentIndex() triggers
onTabChanged(0) for the first time. The previous
onTabChanged-only wiring saw shellCwd() == "" on first
fire, hid the widget, and nothing rescheduled. Fix is
additive: both refreshRoadmapButton and
refreshRepoVisibility are now also invoked from the 2 s
m_statusTimer periodic refresh AND the startup
QTimer::singleShot(0, ...) lambda. Same three-connection
pattern that refreshBgTasksButton already follows correctly.
Same regression class as the v0.6.29 review-button regression.
New docs/standards/status-bar.md documents the
State-category-widget refresh contract (shellCwd() readers
must register on all three connection points) plus a checklist
for adding new widgets, so this regression class doesn't
recur. 5 ANTS-1160-INV-N invariants in
tests/features/roadmap_status_bar_refresh/.
0.7.74 — Bundle G Tier 3 closeout (mainwindow.cpp -785 LoC)
Theme: Bundle G Tier 3 carve-out — second and third of three
L6 LoC decompositions of src/mainwindow.cpp. Ships the
ClaudeStatusBarController extraction (ANTS-1146) and the
themedstylesheet helper extraction (ANTS-1147), together cutting
mainwindow.cpp from 6249 → 5464 LoC (-785, -12.6 %) — the
largest single-bundle reduction in the file's recent history.
Adds three roadmap items (ANTS-1155 / 1156 / 1157) covering the
in-app self-update gap, a roadmap-system audit, and the Project
Audit tool flesh-out. Closes with a post-bundle debt sweep
(5 trivials + 3 behaviorals). 131/131 ctests pass.
ANTS-1146 — ClaudeStatusBarController extracted from
mainwindow.cpp. L6 LoC2 carve-out. Three contiguous
Claude-status sections — applyClaudeStatusLabel,
updateClaudeThemeColors, setupClaudeIntegration (~680 LoC
combined, including the ~120-LoC permission-button factory) —
moved to a top-level ClaudeStatusBarController : QObject in
new translation unit src/claudestatuswidgets.{cpp,h}. State
booleans (m_claudePromptActive, m_claudePlanMode,
m_claudeAuditing, m_claudeLastState, m_claudeLastDetail)
and the six widgets they drive (m_claudeStatusLabel,
m_claudeContextBar, m_claudeReviewBtn, m_claudeErrorLabel,
m_claudeBgTasks, m_claudeBgTasksBtn) all migrate; service
objects (m_claudeIntegration, m_claudeTabTracker) stay on
MainWindow and are observed via attach(). Coupling shape:
six signals out (reviewClicked, bgTasksClicked,
allowlistRequested, reviewButtonShouldRefresh,
statusMessageRequested, statusMessageCleared) plus four
std::function providers in (current/focused/at-tab terminal
setupClaudeIntegration for historical convenience but aren't
Claude chrome (Roadmap button, update-available QAction, 5 s
startup update-check singleShot) stay on MainWindow under
the renamed setupStatusBarChrome host. mainwindow.cpp:
6249 → 5656 LoC (-593). Spec at docs/specs/ANTS-1146.md
cold-eyes-reviewed before implementation; new feature test at
tests/features/claude_statusbar_extraction/ locks 9 INVs;
three pre-existing tests (claude_bg_tasks_button,
claude_state_dot_palette, allowlist_add) re-pointed at
claudestatuswidgets.cpp per the spec's INV-9 table.ANTS-1147 — themedstylesheet helpers extracted from
mainwindow.cpp + cache-and-compare branch chip. L6 LoC3
carve-out. applyTheme's ~178 LoC inline QSS body and four
per-widget restyle templates moved to pure-function builders
in new translation unit src/themedstylesheet.{cpp,h}: six
public helpers (buildAppStylesheet,
buildMenuBarStylesheet, buildStatusMessageStylesheet,
buildStatusProcessStylesheet, buildGitSeparatorStylesheet,
buildChipStylesheet). The chip QSS template that was
inlined three times (applyTheme, updateStatusBar,
refreshRepoVisibility) collapses onto one parameterised
helper; margin asymmetry preserved byte-for-byte (unit-free
0 vs <N>px). mainwindow.cpp: 5656 → 5464 LoC (-192);
themedstylesheet.cpp: 272 LoC. New feature test at
tests/features/themedstylesheet_extraction/ locks 8 INVs;
three pre-existing tests (menubar_hover_stylesheet,
tab_close_button_visible, review_changes_clickable)
re-pointed at themedstylesheet.cpp.
updateStatusBar runs every 2 s on the status timer; pre-fix
code rebuilt and re-applied the branch-chip QSS unconditionally
on every tick even when nothing changed. Two new private
members on MainWindow (m_lastBranchChipQss +
m_lastBranchChipValid) cache the last-applied stylesheet;
the tick now computes the new QSS via
themedstylesheet::buildChipStylesheet and only calls
setStyleSheet when it differs. applyTheme invalidates the
cache on theme change so the next tick re-applies. User-visible
only as a small CPU reduction in the steady-state status-tick
path; the optimisation was the roadmap entry's stated
motivation alongside the extraction.ANTS-1155 added to 0.8.0 — true in-app self-update (no
AppImageUpdate dep). Today's "Update" click in
handleUpdateClicked is in-place auto-update only when the
user already has AppImageUpdate (GUI) or
appimageupdatetool (CLI) on $PATH AND is running the
AppImage build; every other path falls through to
QDesktopServices::openUrl — a glorified browser link. User
feedback 2026-05-02: "the terminal must download and apply the
update directly, no link." Plan: download full AppImage via
QNetworkAccessManager, SHA-256 verify against the *.sha256
artefact already published by release.yml, atomic
rename(2) over $APPIMAGE, restart-with-session-preservation.
Three-bucket distribution-channel contract (ANTS_BUILD_CHANNEL
= appimage / distro / source): distro-packaged builds
suppress the notifier structurally — Flatpak / RPM / .deb /
Arch users update via their package manager, not via the app.
Placed in 📦 Distribution readiness between H7 (website) and
H13 (outreach launch) — auto-update lands before outreach so
the wave of new users actually stays on latest.
ANTS-1156 added to 0.8.0 — Roadmap-system audit (split /
tag / integrate / display / number / write). User ask
2026-05-02: "We need to iron out how the roadmap is going to
work" — six concrete questions. Three already covered (defer
to ANTS-1154 tagging, ANTS-1139 summary-table renderer, and
docs/standards/roadmap-format.md § 3.5.1 numbering); three
genuinely open (file-splitting strategy as ROADMAP.md
approaches the § 3.9 archive-rotation threshold; Ants Terminal
↔ roadmap integration via Help-menu access + cross-roadmap
navigation + two-pane reading mode; Claude Code ↔ roadmap
integration via auto-load, fold-in of /audit and /debt-sweep
findings, and a roadmap MCP capability). Six child
deliverables (1156-A through 1156-F) enumerated to spin out
once decisions land. Sequence dependency: ANTS-1154 (format v2
tagging) lands first since 1156's child items consume v2
tags. Placed at the lead of 🧰 Dev experience.
ANTS-1157 added to 0.8.0 — Project Audit tool flesh-out
(cross-skill / cross-project history). User ask 2026-05-02:
flesh out the Project Audit concept by analysing every
/audit, /indie-review, and /debt-sweep run across all
projects. Six capabilities sequenced smallest-first: (1) run-
history persistence as dated SARIF v2.1.0 files in
~/.local/share/ants-terminal/audit-history/<slug>/; (2)
per-finding fingerprint + recurrence detection via SARIF
partialFingerprints (§ 3.36); (3) MTTR / open-finding-age
tracking + a new History tab in AuditDialog; (4) cross-skill
correlation lane for findings flagged by both static analysis
and cold-eyes review; (5) Projects tab with per-project
summary metrics modeled on Codacy/SonarQube; (6) roadmap-
fold-in audit trail (raw → ANTS-N → status flip → CHANGELOG)
integrating with ANTS-1117 IPC verbs and ANTS-1154 tagged-text.
Online research synthesised from SARIF 2.1.0 (OASIS),
DefectDojo v2.55.1, Microsoft sarif-tools, SARIF Visualizer,
Codacy, and SonarQube. Placed at the lead of 🧰 Dev experience
above ANTS-1156.
Post-bundle debt sweep — 5 trivials + 3 behaviorals.
Trivials: stale comment in mainwindow.h:4 referencing the
pre-1146 applyClaudeStatusLabel rewritten to point at the
controller; refreshStatusBarForActiveTab doc-comment at
mainwindow.h:404 updated similarly; unused <QApplication>
and <QDir> includes dropped from claudestatuswidgets.cpp;
tests/features/claude_bg_tasks_button/spec.md and
tests/features/claude_state_dot_palette/spec.md INV prose
updated to match the post-1146 controller-routed call shapes.
Behaviorals (user sign-off 2026-05-02): deleted write-only
m_lastBranchChipPrimary + m_lastBranchChipTheme cache
members from ANTS-1147 (the QSS string itself encodes the
theme × primary × margin triple, so per-flag fields were YAGNI
redundancy); added "verified externally via git diff --stat;
this in-process test asserts only the floor on the new TU"
disclaimer to the two-sided LoC anchors in both
docs/specs/ANTS-1146.md (INV-8) and docs/specs/ANTS-1147.md
(INV-7) so spec text matches what the source-grep harnesses
can actually enforce.
Theme: Bundle G Tier 3 carve-out — first of three L6 LoC
decompositions of src/mainwindow.cpp. Ships the
DiffViewerDialog extraction (ANTS-1145) and a roadmap-hygiene
pass that flips three stale 🚧 statuses (ANTS-1118 / 1133 /
1139) to ✅ where the work was already shipped but the emoji
hadn't been updated. Adds ANTS-1153 to the 1.0 milestone:
fresh-eyes audit of the feature-test corpus.
MainWindow::showDiffViewer (~430 LoC of dialog constructionQFileSystemWatcher +
scroll-preservation + Copy-Diff payload) moved to a free
function diffviewer::show(QWidget *parent, const QString &cwd, const QString &themeName) in new translation unit
src/diffviewer.{cpp,h}. mainwindow.cpp shrank from 6636 →
6213 LoC; the post-extraction showDiffViewer slot is ~25
meaningful LoC (status message → focused-terminal lookup →
button disable → call into diffviewer::show → connect
dialog destroyed for re-enable + refreshReviewButton).
Spec at docs/specs/ANTS-1145.md was cold-eyes-reviewed
before the implementation: 1 CRITICAL + 2 HIGH + 3 MEDIUM
findings folded in (per-test re-pointing table, exact-
signature INV, user-visible-string preservation INV,
surface-design justification, theme-capture-at-open + QProcess-
parent-change semantics named explicitly). New feature test
at tests/features/diffviewer_extraction/ locks 7 INVs;
three of the four pre-existing review_changes_* tests
re-pointed at diffviewer.cpp per the spec table; the
fourth (clickable) was about button-gating policy and
didn't need re-pointing. Full ctest 129/129 green.Roadmap hygiene — three stale 🚧 statuses flipped to ✅. Survey of Bundle G surfaced three items whose work shipped in earlier 0.7.x releases but whose emoji status was never updated: ANTS-1118 (smooth-scroll snapshot race — shipped 2026-05-01 in 0.7.65 per CHANGELOG); ANTS-1133 (VT parser CSI verbs — REP/CHT/CBT/HPA shipped 2026-05-01 in 0.7.69 + 0.7.70, edge cases explicitly deferred); ANTS-1139 (RoadmapDialog markdown subset — 3 of 5 sub-fixes shipped 2026-05-01 in 0.7.70, remaining 3 explicitly deferred). All three flipped with deferred-sub-items called out inline so future readers know what to re-open as separate IDs if the edge cases ever bite a user.
ANTS-1153 added to 1.0 milestone — fresh-eyes audit of
feature-test corpus. ~190 feature tests live in
tests/features/; many were written alongside their fix and
source-grep the patch body rather than the behavioural
contract. As the codebase matures, two failure modes leak
in: tests that pin implementation details and break on
cosmetic refactors without catching real regressions, and
tests whose INVs no longer match their spec.md because the
spec evolved but the test didn't. Plan: walk every
tests/features/<dir>/ against its spec.md, mark each as
keep / rewrite / retire, fold rewrites into a dedicated
bundle. Pairs naturally with ANTS-1085 (perf-regression
suite) as the "what does CI actually catch?" audit.
ANTS-1154 added to 0.8.0 — tagged-text roadmap format v2 +
RoadmapDialog tab filtering on tags. User report
2026-05-02: the RoadmapDialog tabs (Full / History / Current
/ Next / Far Future / Custom) over-show because they infer
scope from prose shape rather than reading explicit tags. v2
promotes the H-bundle summary-table format (| ID | Description | Status | Target release |) to a first-class
section type and adds machine-readable tags on every line
via HTML comments (<!-- header:section-intro -->,
<!-- kind:fix -->, <!-- prose --> etc.) so per-tab
filtering is exact instead of heuristic. RoadmapDialog
parseBullets + renderHtml learn the new tags; v1 docs
degrade gracefully. One-shot migration of ROADMAP.md tags
every existing item in place.
Theme: Bundle G perf optimisations from the post-revalidation review — closes ANTS-1140 (RoadmapDialog Kind extraction double-walk fold) and ANTS-1149 (paintEvent QTextLayout reuse across runs). Both were tagged "pure perf optimisation, not a bug" and originally deferred from the Tier 2 sweep; user explicit direction 2026-05-02 to ship them rather than defer.
ANTS-1140 — Kind extraction folded into single pre-walk +
cache. RoadmapDialog::renderHtml previously did a per-bullet
peek-ahead when kindFilter was non-empty: for each bullet,
walk forward through continuation lines, append into a
bodyFull QString, regex-match the Kind: line. With ~270
bullets and 8-10 renders/sec while the user types into the
search box (debounced 120 ms) with a Kind filter active,
this was the dominant render cost. Replaced with a single
pre-walk at the top of renderHtml that builds a
QHash<int, QString> kindByLine keyed on bullet line index,
cached thread-locally across calls (mirrors the
reverseTopLevelSections cache from 0.7.70). The bullet
block now does an O(1) lookup. All 7 INVs of
tests/features/roadmap_kind_facets/ pass against the
refactor.
ANTS-1149 — paintEvent QTextLayout reuse. Pre-fix code
constructed a fresh QTextLayout layout(runText, *drawFont)
inside the inner paint loop on every text run. Claude Code's
heavily-styled output produces dozens of runs per row × N
visible rows × 60 fps, so the impl-alloc cost was a real hot
spot. New mutable QTextLayout m_paintLayout member is
reused across all runs in a paint via setText (and
setFont only when the run's font role changes — bold /
italic / both / regular). HarfBuzz shape pass still runs per
unique text — load-bearing for ligature support. A future
row-content-fingerprint cache could skip even the shape, but
is left for later.
tests/perf/bench_vt_throughput baseline comparison (0.7.71 →
0.7.72): minor parser-side gains (ASCII print 29.96→30.10
MB/s, newline 6.04→6.30, ANSI SGR 18.02→18.32, UTF-8 CJK
8.34→8.80) consistent with compiler optimisations on the
recent code. The QTextLayout reuse impact is in paintEvent,
not vt_throughput; a bench_paint_throughput harness (per
ANTS-1115) is still pending — without it, the win is
qualitative rather than quantified.
Theme: Bundle G post-revalidation closures. The 2026-05-01
revalidation pass (/audit re-run + scoped /indie-review on
the changed surface) accepted all 13 Tier 1+2 closures shipped
across 0.7.65–0.7.70 but surfaced 4 net-new findings worth
closing before the codebase is called "audit-clean". 128/128
ctests pass.
ANTS-1151 — Hook + MCP socket SO_PEERCRED + idle timeout.
Symmetric extension of ANTS-1132's remote-control hardening
to the two QLocalServers that ClaudeIntegration owns:
onHookConnection + onMcpConnection. The L3 indie-review
reviewer flagged that the file-side guarantees (0700 perms +
UserAccessOption + safeToUnlinkLocalSocket) hold, but the
peer-side trust-model reasoning that motivated SO_PEERCRED on
remote-control applies symmetrically — a same-UID-but-
different-process attacker (e.g. a malicious browser plugin)
could otherwise inject hook events shaped like
processHookEvent consumes, or call MCP verbs (filesystem
reads, git status, environment). 5-second per-connection
idle timeout closes the slow-loris foot-gun.
ANTS-1152 — GlobalShortcutsPortal permanent-failure latch.
ANTS-1142 H-1's 0.7.67 fix (clear m_pending + m_sessionHandle
on BindShortcuts rejection) was correct on first failure
but didn't enforce terminality across re-binds — a future
config-reload / "Reload Plugins" path that re-bound shortcuts
would re-enter the same backend failure (the loop on GNOME
the original finding flagged). Lane 6 reviewer caught this
during revalidation. New m_permanentlyFailed member set in
every sessionFailed-emit path; bindShortcut early-returns
when latched. Today's only caller fires once at startup so
the bug was dormant; latch makes the contract enforced
rather than aspirational.
ANTS-1136 (cancel path) — RuleQualityTracker flush on
cancel. The 0.7.68 cycle-completion flush in runNextCheck
closed the kill-9-loses-fires class on clean completion;
this closes the symmetric hole on cancel. cancelAudit now
also calls m_qualityTracker->save() before renderResults().
Lane 4 reviewer caught the asymmetry during revalidation.
ANTS-1134 (modifier list extension) — CapsLock, NumLock,
Super, Hyper added to the modifier-only keypress guard.
Lane 2 reviewer flagged the original list missed Qt::Key_CapsLock,
Qt::Key_NumLock, Qt::Key_ScrollLock, Qt::Key_Super_L/R,
Qt::Key_Hyper_L/R, Qt::Key_Mode_switch. Same reflexive-
modifier class — pressing CapsLock while scrolled-up
shouldn't kill the scroll position either.
Static-analysis closures. cppcheck identicalConditionAfterEarlyExit
at auditdialog.cpp (m_cancelled check after early return —
always false; redundant check removed). cppcheck passedByValue
at globalshortcutsportal.cpp (Qt SLOT signature requires
QString-by-value for old-style connect string-match;
suppressed inline with rationale comment). shellcheck SC1087
×2 in packaging/rotate-roadmap.sh (array-expansion brace
hygiene).
Theme: Bundle G Tier 2 closeout — heavy items. ANTS-1135 (post-fork setenv signal-safety, the headline correctness fix in this release), ANTS-1133 partial (3 more CSI verbs), 1140 partial (reverseSections cache), 1139 partial (3 of 5 markdown subset gaps closed). 128/128 tests pass.
ANTS-1135 — post-fork setenv async-signal-safety. Pre-fix
code called ::setenv() ×5 in the child after forkpty,
but setenv is NOT on POSIX's async-signal-safe list
(§ 2.4.3) — it can call realloc() to grow environ.
CLAUDE.md and the flatpak path's pre-fork-allocation
comment block both claimed "child only does execvp"; the
non-flatpak path was silently breaking the discipline. Fix:
pre-fork build of envp (parent thread, signal-safe context):
walk environ, skip the 5 keys we override, append our 5
overrides as static-storage / stack-allocated entries.
Child uses execle(shellCStr, argv0, nullptr, envp)
instead of execlp + setenv loop. Matches the flatpak-path
discipline. flatpak_host_shell test INV-5 updated to
recognize the execle invocation.
ANTS-1133 (4/5) — CSI verbs. Three more on top of REP from 0.7.69: CSI Pn I (CHT, cursor horizontal forward tabulation), CSI Pn Z (CBT, cursor backward tabulation), CSI Pn ` (HPA, horizontal position absolute — same as CHA but spelled with backtick). All three are common in ncurses, less, and bash tab-completion. Combining-after- wrap edge case + wide-cont rewrap orphan remain deferred as discrete rare-edge items.
ANTS-1139 (3/5) — RoadmapDialog markdown subset. Two
fixes: (a) **bold** recognized in body prose via a new
pass in applyInline (pre-fix code rendered the literal
** characters in body text — every bullet's "bold
headline" was visually wrong); (b) markdown tables
render as <table> not <pre> — every |-row becomes
<tr>, first row is <th>, separator row (|---|---|)
auto-skipped, body rows are <td>. Per-cell applyInline
so backticks + bold work inside cells. Sub-bullet
rendering (heavy refactor) + Kind regex multi-value +
bold regex anchor hardening remain deferred.
ANTS-1140 (1/2) — RoadmapDialog perf. reverseTopLevelSections
gains a function-local thread-local cache keyed on the
input markdown. History-mode renders + search keystrokes
DescendingChronological sort) now short-circuit on
identical input. With archive markdown attached
(ANTS-1125), the input can reach 64 MiB; pre-fix code
re-walked the entire document on every render. Kind
double-walk fold (the harder optimisation) remains
deferred — perf only, not a bug.Theme: Bundle G Tier 2 final closeouts. ANTS-1134, 1138, 1142, 1144 all flip ✅ (sub-fixes from 0.7.67/0.7.68 closed out); ANTS-1133 partial — CSI REP shipped, other CSI-spec gaps remain deferred. 128/128 tests pass.
ANTS-1134 (2/2) — span cache invalidation across
scrollback push. New m_lastScrollbackPushed member
tracks the monotonic counter; invalidateSpanCaches clears
both URL + highlight span caches when the count changes
since the last paint. Pre-fix code only invalidated entries
for screen rows the grid flagged dirty — but scrollback
push shifts every cached scrollback-line entry to point at
a different historical line. Symptom: URL highlights
occasionally painted at the wrong column ranges on
scrolled-back lines after high-throughput output.
ANTS-1138 (2/2) — auto-profile-rules pattern cache
invalidation. New Config::autoProfileRulesGeneration()
counter bumped on every successful setAutoProfileRules
write; MainWindow::checkAutoProfileRules compares to a
function-local s_lastRulesGen and clears s_patternCache
s_warnedInvalid on change. Pre-fix code retired
patterns lingered forever, and a fixed-then-rebroken
pattern wouldn't re-warn after the third edit.ANTS-1142 (4/4) — KDE-presence guard lifted into shared
helper. kwinPresent() now lives as a free function in
kwinpositiontracker.h (header-only, env-var check
inlined). KWinPositionTracker::setPosition,
MainWindow::moveViaKWin, and MainWindow::centerWindow
all share the guard. Pre-fix code: moveViaKWin and
centerWindow unconditionally fired the kwin-script +
dbus-send chain on GNOME / Sway / Hyprland / etc.,
orphaning /tmp/kwin_*_ants_*.js files and triggering a
dbus-send to a non-existent service every time. Existing
kwin_position_tracker feature test updated to source-grep
the header for the env-var literals + assert
setPosition() calls kwinPresent() before any
QTemporaryFile constructor.
ANTS-1144 (3/3) — Claude transcript render cap.
ClaudeTranscriptDialog::loadTranscript now caps rendered
entries to the last 2000 with a "showing last N of M"
header. Pre-fix code built full HTML for transcripts up to
the integration-layer 100 MiB cap and handed it to
QTextEdit::setHtml, freezing the UI for tens of seconds
on long Claude Code sessions. 2000 is empirically generous
— even busy sessions stay well under that.
ANTS-1133 (1/4) — CSI Pn b (REP) handler. Repeats the
preceding graphic character N times. Common in less,
ncurses (tput rep), and zsh's reset-prompt path as a
bandwidth optimisation for runs of ─ borders, spaces,
etc. Pre-fix code dropped this silently — visible as
missing border chars in less redraws + blanks in TUI
separator lines on high-latency PTYs. Other CSI-spec gaps
(Z CBT, I CHT, ` HPA, combining-after-wrap, wide-cont
rewrap) remain deferred as discrete items.
Theme: Bundle G Tier 2 closeouts — knocking off the deferred sub-fixes annotated in 0.7.67's ROADMAP entries. ANTS-1136, 1141, 1142, 1144 each move from 🚧 partial to substantively closed. 128/128 ctests pass.
ANTS-1136 (5/5) — RuleQualityTracker durability.
Final sub-fix: runNextCheck calls
m_qualityTracker->save() when the audit cycle completes
(m_currentCheck >= m_checks.size()). Pre-fix code only
flushed at destructor (RAII), so a SIGSEGV / SIGKILL /
power loss between recordFire calls lost up to 30 minutes
of fire records.
ANTS-1141 (5/5) — parent-dir fsync after rename. New
secureio::fsyncParentDir(path) helper (Postgres /
SQLite / Git pattern). Called after every successful
std::rename in Config::save and
SessionManager::saveSession / saveTabOrder. ext4 with
data=ordered could otherwise lose the rename's
directory-entry update on a crash between rename(2)
returning and the journal commit.
**ANTS-1142 (4/4) — MainWindow sessionFailed listener
GlobalShortcutsPortal::sessionFailed with a qWarningshowStatusMessage fallback. Pre-fix code emitted
the signal to no listener — on GNOME (where
CreateSession succeeds but BindShortcuts fails) the
user just saw "F12 doesn't work" with no diagnostic. The
KDE-presence guard lift sub-fix (for moveViaKWin /
centerWindow) deferred — small but distinct refactor.ANTS-1144 (3/3) — BgTasks dialog ANSI strip. tailFile
in claudebgtasksdialog.cpp now strips CSI/SGR escape
sequences before HTML-escape via a regex pass. Build
logs with ncurses-style progress bars (make -j$(nproc),
claude-code's status bar, etc.) no longer show literal
^[[2J / ^[[31m etc. in the dialog. Per CLAUDE.md
rule 3 (reuse-before-rewriting), the ideal would route
through VtParser/TerminalGrid — but the regex pass
covers ~99% of in-the-wild noise without pulling in the
parser link footprint. Same regex also strips OSC
sequences (window-title updates). Transcript chunked
render still deferred (touches claudetranscript.cpp's
setHtml flow more invasively).
Theme: Bundle G Tier 2 progress sweep. 8 ANTS IDs touched (ANTS-1134, 1136, 1137, 1138, 1141, 1142, 1143, 1144) — 3 fully shipped (1137, 1143 doc-only, AI partial-stream fail-closed in 1144), 5 partial (sub-fixes shipped, remaining sub-fixes annotated in ROADMAP for 0.7.68). 128/128 ctests pass throughout.
ANTS-1134 (partial) — modifier-only keypress guard.
TerminalWidget::keyPressEvent no longer resets
m_scrollOffset on bare modifier presses (Shift / Ctrl /
Alt / Meta / AltGr / Key_unknown). Pre-fix code killed the
scrolled-up workflow when the user reflexively pressed Ctrl
in preparation for Ctrl+Shift+C. Span-cache invalidation
on scrollback push deferred to 0.7.68.
ANTS-1136 (4/5) — audit pipeline doc + correctness.
CLAUDE.md pipeline-order line corrected to match
handleCheckOutput reality (.audit_suppress marks not
drops; dedup runs after drop steps); Confidence formula
description spelled out (floor +10, severity×15, cross-tool
+20, AST +10, grep-and-short −5, test path −20, AI verdict
caps ≤30 / ≥80). consolidateMypyStubHints dedup-key
width 16→24 hex via AuditEngine::computeDedup. cancelAudit
sets m_snapshotPersisted = true before renderResults()
so cancelled-run partial picture doesn't pollute
trend.json. static QString sourceForCheck trampoline
dropped — one call site uses fully-qualified
AuditEngine::sourceForCheck. refactor_set_permissions_pair
rule removed from audit_rules.json (duplicate of
hardcoded setPermissions_pair_no_helper). RuleQualityTracker
durability deferred to 0.7.68.
ANTS-1137 — MainWindow chrome perf hotspots.
refreshRoadmapButton replaces unbounded
QDir::entryInfoList (which enumerated the entire CWD on
every 2 s status tick) with three explicit
QFileInfo::exists calls for the case-variant ROADMAP.md
filenames (O(1) instead of O(N) per tick). New
m_repoVisibilityProbeInFlight map mirrors
m_reviewProbeInFlight to drop redundant gh repo view
QProcesses on fast tab-switches.
ANTS-1138 (partial) — applyTheme early-return.
MainWindow::applyTheme(name) early-returns when
name == m_currentTheme && !m_currentTheme.isEmpty(). Closes
the auto-profile-rules re-entrancy path (updateStatusBar
tick → checkAutoProfileRules → applyTheme → setTheme →
onConfigFileChanged → applyTheme) without depending on
the m_inConfigReload latency. Pattern-cache
invalidation deferred to 0.7.68.
ANTS-1141 (4/5) — Config + persistence hardening.
Sessions directory tightened to 0700 after mkpath
(matches the config dir which was already 0700 by accident
of an earlier run; no longer leaks filenames + mtimes +
sizes via directory listing). Config::setKeybinding /
setPluginGrants / setPluginSetting / setRawData
short-circuit when m_loadFailed is true (no more
fictional in-memory state on a parse-failure recovery
session). cleanupOldSessions extended to sweep orphan
*.tmp files older than 1 day (catches rename-failure
leftovers). SessionManager::loadTabOrder no longer
destructively removes tab_order.txt on read (the
atomic-write on save overwrites it; the destructive read
lost the order on a 0-5 s crash-window before
saveAllSessions could fire). Parent-dir fsync deferred
to 0.7.68.
ANTS-1142 (2/4) — Wayland portal + debug log.
GlobalShortcutsPortal::onBindShortcutsResponse drains
m_pending and clears m_sessionHandle on
BindShortcuts failure — sessionFailed is now
terminal-per-process, no more failure-loop pathology on
GNOME (where CreateSession succeeds but BindShortcuts
fails). Debug-log file creation wraps s_file.open() in
umask(0077) save/restore so the kernel applies 0600 at
create time, eliminating the same-UID race window between
open and the post-open setOwnerOnlyPerms call. Two
sub-fixes deferred to 0.7.68: KDE-presence guard lift
(for moveViaKWin / centerWindow), MainWindow
listening to sessionFailed with qWarning fallback.
ANTS-1143 — PLUGINS.md spec drift. string.dump
added to "libraries removed" list (the runtime nils it
via LuaEngine::stripDangerousGlobals for CVE-2014-5461
defence; doc was missing it). os.date() example
replaced with sandbox-safe ants._version so copy-paste
doesn't crash. C-call timeout caveat documented (string
regex / table.sort comparator chains bypass the
instruction-count hook). Optional Lua wall-clock watchdog
stays explicitly out-of-scope per the original spec.
ANTS-1144 (1/3) — AI dialog partial-stream Insert
fail-closed. m_insertBtn->setEnabled(!hadError) —
pre-fix code enabled Insert on truncated responses (e.g.
network drop mid-fenced-block leaving the prefix intact),
inviting the user to confirm a command missing its tail.
Now Insert is only reachable on cleanly-completed
responses. Two sub-fixes deferred to 0.7.68: transcript
incremental render, BgTasks dialog ANSI strip via
VtParser reuse.
Theme: Bundle G Tier 1 remainder. Two HIGH fixes from the 2026-05-01 indie-review sweep — ANTS-1131 (PTY child-process lifecycle) and ANTS-1132 (IPC socket trust-model gaps). Tier 2
ANTS-1131 — PTY child-process lifecycle. Two related
bugs in src/ptyhandler.cpp + src/claudeintegration.{cpp,h}
src/mainwindow.cpp:Pty::onReadReady orphans m_childPid on
EOF-before-reap (indie-review L3 HIGH-2). Pre-fix code
unconditionally cleared m_childPid = -1 even when
waitpid(WNOHANG) returned 0 (child alive, just not
reaped) — defeating the destructor's SIGTERM/SIGKILL
escalation block. Fix: gate the clear on w > 0. When
waitpid returns 0, the destructor's escalation thread
handles cleanup as designed.m_planModeByPid PID-reuse poisoning (indie-review
L3 HIGH-3). The per-PID plan-mode cache grew
monotonically over a session — never pruned on tab
close. Linux PID reuse meant a freshly-launched shell
could inherit a stale plan-mode flag from a closed
Claude tab. Fix: new ClaudeIntegration::forgetShell(pid)
called from MainWindow::closeTab alongside the
existing m_claudeTabTracker->untrackShell(pid).ANTS-1132 — IPC socket trust-model gaps. Three related hardenings (indie-review L5 HIGH-1 + L3 HIGH-4):
SO_PEERCRED UID match on remote-control accept.
Trust-model comment at remotecontrol.cpp:31-36 claimed
UID-scope but the code only enforced file-side perms
(0700). Now getsockopt(SO_PEERCRED) confirms the peer
UID matches getuid() before any verb dispatches —
defence-in-depth for the case where the socket path is
ever moved (e.g. ANTS_REMOTE_SOCKET env override) and
file ACLs stop being load-bearing.UserAccessOption + safeToUnlinkLocalSocket
precheck on Claude hook + MCP servers. Pre-fix code
called removeServer() against a /tmp/ants-claude- hooks-<pid> path with no symlink guard — a hostile
same-UID process could pre-create a symlink there
pointing at e.g. ~/.ssh/known_hosts and removeServer
would unlink the target. Now both paths gate the
unlink behind the lstat-checked S_ISSOCK + UID match
guard (lifted into secureio.h from its file-scope
home in remotecontrol.cpp so all three sockets share
one helper).QTimer::singleShot(5000, socket, &QLocalSocket::abort) per accepted connection. A peer
that opens an idle connection without sending bytes
can no longer hold the socket indefinitely.Theme: Bundle G Tier 1 fix-pass. Two CRITICAL fixes from the 2026-05-01 multi-agent indie-review sweep — ANTS-1118 (smooth-scroll snapshot race during streaming, root cause identified by the terminal-widget-paint reviewer) and ANTS-1130 (VT alt-screen + scroll-region invariants — DECSET 47/1047/1049 conflation, resize() destroying DECSTBM, alt scroll-region OOB on shrink-resize). Two more Tier 1 items (ANTS-1131 PTY child lifecycle + ANTS-1132 IPC socket trust gaps) defer to 0.7.66 to keep this release's blast radius small. Tier 2 + Tier 3 follow per the fold-in plan.
ANTS-1118 — smooth-scroll snapshot race. User report
2026-04-30 + downgrade 2026-05-01 ("scrolling up during a
Claude Code stream briefly shows overwritten text"). Root
cause: wheelEvent set m_smoothScrollTarget and started
the timer, but the first 16 ms tick had intStep = int(0.9) = 0, so updateScrollBar() didn't fire and the snapshot
stayed uncaptured. During the 16–32 ms race window
onVtBatch saw m_scrollOffset == 0 and let the grid
mutate the rows the user was scrolling past — once the
snapshot finally captured, it captured the post-mutation
grid. Fix: trigger captureScreenSnapshot() +
m_grid->setScrollbackInsertPaused(true) on scroll intent
(positive m_smoothScrollTarget from offset 0 with no
existing snapshot) rather than on committed offset
transition. Plus smoothScrollStep timer-stop branch calls
updateScrollBar() so any intent-captured-but-never-
committed snapshot is dropped. Spec at
docs/specs/ANTS-1118.md; feature test at
tests/features/scroll_snapshot_intent/ drives 4 INVs
(source-grep on the wheelEvent intent branch +
smoothScrollStep cleanup + onVtBatch regression guard).
ANTS-1130 — VT alt-screen + scroll-region invariants.
Three findings from indie-review L1 (VT parser + grid),
all in terminalgrid.cpp, collapsed into one fix-pass:
m_altScreenMode
member tracks which mode entered alt-screen; on exit
only the 1049 path restores DECSC. Programs sometimes
enter with 1049 and exit with a different code — xterm
uses entry-mode for the decision regardless of exit
code, and we now match.resize() preserves DECSTBM. Pre-fix code did
m_scrollBottom = m_rows - 1; m_scrollTop = 0; —
destroying any TUI's scroll region on every window
resize (tmux split panes, less with status line, mc,
all lost their regions). New behaviour: clamp the
existing top/bottom to the new row range. If the user
shrinks below the previous bottom, bottom is clamped;
top is preserved.m_altScrollTop /
m_altScrollBottom at all — after a shrink-resize
while on alt screen with a non-default scroll region,
regionSize = m_scrollBottom - m_scrollTop + 1 could
exceed m_screenLines.size() and the next scrollUp's
std::rotate would read past the end of
m_screenLines. Latent UB → crash class. Same clamp
as item 2, applied symmetrically to the alt scroll
region.Theme: Companion partial closeouts pre-Bundle-G. ANTS-1106
ships the Roadmap viewer's Kind-faceted secondary filter — the
last half of the "mandatory Kind: field + viewer faceted
categorisation" item, with the field-required half having
landed in 0.7.59 alongside the standards doc update. Three
status hygiene flips: ANTS-1116 v1 (ants-helper drift-check)
and ANTS-1117 v1 (roadmap-query + tab-list IPC verbs) both
shipped in 0.7.59 but the status emoji never flipped — done
now, with v2 of each explicitly noted as deferred per their
specs. ANTS-1118 reframed from HIGH to MEDIUM after the user
confirmed the bug downgraded post-0.7.63 (still a temporary
visual artifact, but the scrollback buffer itself is intact).
RoadmapDialog underneath the existing status-
checkbox row, with one toggle per Kind value (12 total — the
10 spec-defined implement / fix / audit-fix /
review-fix / doc / doc-fix / refactor / test /
chore / release, plus the de-facto research and ux
values that occur in the live ROADMAP). Each checkbox label
uses an emoji prefix as a visual cue (✨ implement,
🐛 fix, 🔍 audit-fix, etc.). Empty filter set = no
narrowing (current behaviour preserved); non-empty filter =
OR-include semantics across the active Kinds; bullets
without a Kind: line are excluded under non-empty filters
(the user filtering by Kind doesn't want unclassified
bullets sneaking through). Implementation: extended the
static RoadmapDialog::renderHtml with a defaulted
const QSet<QString> &kindFilter parameter; bullet walk
peeks ahead for the Kind: line via the same regex
parseBullets already uses; m_kindFilter set member +
per-Kind QPointer<QCheckBox> widgets with stable
roadmap-filter-kind-<value> objectNames for QSS targetingdocs/specs/ANTS-1106.md;
feature test at tests/features/roadmap_kind_facets/
drives the helper through 7 INVs (pure-helper drive across
empty / single-Kind / multi-Kind / missing-Kind shapes,
source-grep on the row + objectNames). Standards doc
(docs/standards/roadmap-format.md) was already updated to
mark Kind: as Required as of v1.1 in 0.7.59 — no
doc change needed in this release.ants-helper drift-check subcommand on
the optional CLI binary (-DANTS_ENABLE_HELPER_CLI=ON,
default OFF). ANTS-1117 v1 = the roadmap-query +
tab-list IPC verbs on the existing remote-control Unix
socket. Both have full feature-test coverage from when they
shipped — the status flip is documentation hygiene only. v2
of each (MCP server wrapper + CLAUDE.md §8 bias rule for
1116; audit-run IPC verb for 1117) explicitly stays
deferred per their respective specs — both are gated on
ANTS-1120 measurement validation./audit + /indie-review
sweep) as a candidate finding — multi-agent fresh-eyes
pass over terminalwidget.cpp paint code is the next step.Theme: Outstanding-queue cleanup pre-Bundle-G. ANTS-1014
clipboard-write redaction lands as the one real code fix
(centralises 15 raw setText(...) sites behind a single
hygiene funnel — OSC 52 was the headline exfil vector flagged
by the 7th-audit memory). Two regressions get explicit
"awaiting repro" status flips so the upcoming /audit + /indie-
review pass doesn't re-flag them as untreated bugs. Six big
items get explicit deferral annotations with re-evaluation
triggers so orphan 📋 bullets read as "deferred-by-design,
not forgotten."
src/clipboardguard.{h,cpp} exposes a pure
sanitize(text, source) helper splitting on a Source enum
(Trusted / UntrustedPty / UntrustedPlugin) plus a thin
writeText(text, source, mode = Clipboard) wrapper. All
sources strip embedded QChar(0) (NUL bytes that round-trip
through clipboard managers in unpredictable ways).
UntrustedPty (OSC 52 from the shell) and UntrustedPlugin
(ants.clipboard.write from Lua plugins) cap at 1 MiB —
defence-in-depth against accidental DoS via
cat /dev/urandom | base64. Trusted (user-initiated UI
copy actions) passes large inputs through unchanged so a
10 MiB context-menu copy of a file dump still works. All 15
raw QApplication::clipboard()->setText(...) sites (13 in
terminalwidget.cpp, 2 in mainwindow.cpp) converted with
appropriate source classifications. We deliberately do NOT
apply pattern-based secret detection — false-positive
redaction is worse than no redaction for this user surface.
Spec at docs/specs/ANTS-1014.md; feature test at
tests/features/clipboard_redaction/ drives the helper
through 8 INVs (pure-helper drive across the three sources
for NUL-strip + cap behaviour, source-grep that no raw
setText( survives, source-grep that the two untrusted
classifications appear).1abf768) waiting for the user
to capture one repro session each. Status change so the
/indie-review run knows the bugs are mid-investigation.bench_paint_throughput.cpp / bench_search_throughput.cpp
scaffolding first means the sweep ships measurable wins
rather than vibe-driven refactors.Theme: Bundle B of the pre-0.7.61 outstanding-item triage — two small UX wins that close ROADMAP carryover. The update-available indicator moves out of the status bar onto the menu bar so the one-shot "you have a new version" call-to-action reads as visually loud chrome instead of competing with the steady-state widgets, and the git-branch chip on the status bar gains a colour cue (green for main/master/trunk, amber for everything else) that pairs it visually with the existing "Public/Private" repo-visibility pill next to it.
ANTS-1124 — update-available indicator → menu bar. The
GitHub release-check probe used to surface a clickable QLabel
on the right edge of the status bar (m_updateAvailableLabel,
added in 0.7.45). 0.7.62 promotes it to a top-level menu-bar
QAction (m_updateAvailableAction), inserted to the right
of &Help by call order. The one-shot CTA framing now reads
louder, the status bar reclaims its rightmost slot for
steady-state telemetry, and a future setCornerWidget move
is unblocked. URL is stashed via QAction::setData; the
triggered lambda replays it through the existing
handleUpdateClicked slot — same in-place AppImageUpdate
flow, same Update / Skip / Postpone dialog, no probe-logic
changes. tests/features/update_available_menubar/ locks the
migration via 8 source-grep INVs (no leftover
m_updateAvailableLabel token, action wired via
m_menuBar->addAction or insertAction, visibility-toggle
on the action, whitespace-tolerant connect grep, action
starts hidden). Spec at docs/specs/ANTS-1124.md carries the
cold-eyes review fold-in (F1 / F2 / F3 / F4 / F5 / F7 / F15).
ANTS-1109 — git-branch chip colour cue. User screenshot
2026-04-30: the main branch chip read inconsistent with the
framed Public repo-visibility pill next to it. The chip's
shape (border-radius, padding, font) was already
aligned; what was missing was a branch-derived colour. New
pure helper branchchip::isPrimaryBranch (in
src/branchchip.h) returns true for main/master/trunk,
false for every other ref. Both branch-chip styling sites
(MainWindow::applyTheme for theme switches +
MainWindow::updateStatusBar for per-poll branch changes)
consult the helper and pick theme.ansi[2] (green, same role
as Public) or theme.ansi[3] (amber, same role as
Private). Visually pairs the chip with the pill;
functionally tells the user at a glance whether they're on
the project's primary branch. tests/features/status_bar_branch_chip/
drives the helper over six branch-name inputs (positive +
negative + empty + case-sensitive edge case) and source-greps
every chip-setStyleSheet site for the helper consult and the
palette-index wiring (cold-eyes F1 fold-in: pin every site,
not just the first match). Margin asymmetry between the chip
and the pill is preserved on purpose — chip is the leftmost
status-bar widget, pill sits to its right.
ANTS-1100 + ANTS-1119 — ROADMAP status flips (Bundle A).
Both shipped in 0.7.59 but the status emoji never flipped:
ANTS-1100 (RoadmapDialog faceted-tabs redesign — five
preset filters + search + persisted geometry) and ANTS-1119
(AuditEngine extraction — Qt6::Core-only module that
unblocks non-GUI audit consumers like ants-helper v2 and
the future MCP server). 📋 → ✅, no code touched. ANTS-1049
(audit-pipeline populateChecks-as-data-table) body trimmed
to a "subsumed by ANTS-1044" note — same scope as the
structural-tier auditdialog.cpp decomposition, kept under
one ID per roadmap-format.md § 3.5.1.
tests/features/github_status_bar/ — INV-1, INV-2, INV-11
reflect the ANTS-1124 migration. Previously locked the old
QLabel *m_updateAvailableLabel shape (member declaration,
addPermanentWidget, &QLabel::linkActivated,
setOpenExternalLinks(false)). Now asserts the new
QAction *m_updateAvailableAction shape (member, addAction,
&QAction::triggered) plus a negative regression check that
no m_updateAvailableLabel identifier survives anywhere in
mainwindow.{h,cpp}. Stale <regex> include also removed.Theme: dialog-and-roadmap polish bundle — fixes three user-visible chrome bugs (theme propagation to dialogs, focus-return on dialog close, pseudo-modal click-blocking on the parent), splits the now- 260 KiB ROADMAP.md by minor version so tooling can keep up, retires two long-deprecated standards docs, and folds in a batch of indie-review Tier-3 closures (KWin position tracker rename + leak fix, post-fork stack-only argv).
docs/roadmap/<MAJOR>.<MINOR>.md. The viewer pulls
archives in only on demand (History preset OR non-empty search)
via three new pure-static helpers
(archiveDirFor, loadMarkdown, shouldLoadHistory) that the
feature test drives without instantiating a dialog. Numeric
(major, minor) descending sort defeats the lexical-sort trap at
minor 10 (0.10.md before 0.9.md). Per-file 8 MiB cap + total
64 MiB cap with truncation sentinel. Thematic-break + HTML-
comment sentinel separators between archives. Symlink-safe
canonical-path resolution. Initial archives at docs/roadmap/0.5.md
and 0.6.md; 0.7 stays in the live ROADMAP.md until 0.8.0 cuts.
/bump-driven rotation via packaging/rotate-roadmap.sh (atomic
via mktemp+mv, idempotent, no-clobber on existing archives,
regex-escapes the closed-minor's ., handles EOF sections,
walks closed-minor sub-headings together).
docs/standards/roadmap-format.md § 3.9 documents the rotation
convention. Spec at
tests/features/roadmap_viewer_archive/spec.md — 18 INVs across
three rounds of cold-eyes review (4+8+5+5 → 0+3+4+5 → 0+0+3 →
clean).ANTS-1128 — theme stylesheet now reaches top-level dialogs.
User reports 2026-04-30 (menubar dropdown background + Review
Changes dialog footer) traced to MainWindow::applyTheme calling
setStyleSheet(ss) on the MainWindow instance — Qt's stylesheet
engine only propagates through a widget's render subtree, NOT
through QObject parent-child relationships, so top-level QDialogs
never inherited the theme. Switched to qApp->setStyleSheet(ss),
which fans out to every widget including not-yet-constructed
dialogs. All 9 dialog classes plus 7 ad-hoc new QDialog(this)
instances (about, paste preview, snippet editor, Review Changes,
settings-restore, etc.) now pick up the active theme.
ANTS-1050 — auto-return focus to terminal when any dialog
closes. User request 2026-04-28: "once any dialog box is
closed, automatically shift focus back to the terminal prompt."
The existing qApp eventFilter gained a Close-event branch backed
by a pure-logic helper dialogfocus::shouldRefocusOnDialogClose
in src/dialogfocus.h. Helper returns true iff the event is a
QEvent::Close, the watched object is a QDialog subclass, AND
no other QDialog is still visible (stacked-dialog suppression).
Refocus is deferred via QTimer::singleShot(0, ...) so the
dialog teardown completes before MainWindow grabs focus.
Null-guarded against early-startup dialogs closing before any
terminal exists.
ANTS-1051 — pseudo-modal click-blocking on the parent. User
request 2026-04-28: "when a dialog box is open, only the dialog
box is interactive, anything behind the dialog box should not
be interactive." 0.7.50 made all dialogs non-modal to dodge
QTBUG-79126 (KDE+KWin+Qt 6.11+frameless+translucent: setModal(true)
drops button clicks on Wayland), losing the click-blocking
semantics. Restored manually via the qApp eventFilter and a new
helper dialogfocus::shouldSuppressEventForDialog. Mouse/key/
wheel events landing outside any visible dialog's tree are
swallowed; events on the dialog itself (or its descendants) pass
through. Stacked dialogs both allowed; the strict-ancestor edge
case (a widget is not its own ancestor in Qt) is handled
explicitly so clicks on the dialog's own frame work.
KeyRelease paired with KeyPress to prevent modifier-state
desync.
ANTS-1058 — menubar dropdown flicker on mouse motion. User
confirmation 2026-04-30 that the 0.7.5 NoAnimStyle plus the
0.7.5+1 HoverMove suppression were fully sufficient on the
user's stack; the KWin-side experimentation hypothesis didn't
materialise. Closed without further fix-pass.
ANTS-1054 — mystery flashing dialog in centre of terminal.
User confirmation 2026-04-30 that another Claude Code session
resolved the flashing-dialog popups by adding a YML configuration
file (file local to the user's environment; Ants source has no
related change). The ANTS_TRACE_DIALOGS=1 trace tooling
introduced in 0.7.57/0.7.58 stays in place for future incidents.
ANTS-1045 — XcbPositionTracker rename + non-KWin bail +
temp-file leak. Indie-review-2026-04-27 Tier-3. Class +
files renamed to KWinPositionTracker (the class doesn't use
XCB at all; it talks to KWin's scripting D-Bus interface).
KWin-presence guard checks KDE_FULL_SESSION=true OR
XDG_CURRENT_DESKTOP containing KDE and bails before any
temp-file write — non-KWin desktops no longer accumulate
kwin_pos_ants_*.js files at ~10 per Quake-mode toggle.
QScopeGuard-style cleanup runs QFile::remove(scriptPath) on
every synchronous failure path; the async dbus chain dismisses
the guard and owns cleanup explicitly via the QProcess
finished / errorOccurred lambdas. The 0.7.12 TOCTOU
fix (QTemporaryFile + setAutoRemove(false)) is preserved.
ANTS-1046 — post-fork heap allocations in flatpak detect
path retired. Indie-review-2026-04-27 Tier-3. The previous
std::string + std::vector<const char *> argv build between
forkpty and execvp("flatpak-spawn", ...) relied on the
glibc malloc fork-handler — usually safe, not strictly POSIX
async-signal-safe in a multithreaded program. Detection +
argv string buffers now build pre-fork in the parent via
snprintf (POSIX § 2.4.3 async-signal-safe) into stack
buffers; the child only assembles a stack-allocated
const char *argv[12] table and calls execvp. Zero
std::string / std::vector references between forkpty
and execvp. <string> and <vector> includes dropped from
ptyhandler.cpp.
ANTS-1012 — RoadmapDialog::rebuild unbounded read. The
remaining open gap from indie-review-2026-04-27's "unbounded
reads" cross-cutting theme. f.readAll() now caps at 8 MiB per
the loadMarkdown helper's per-file budget — defends against
symlinked /dev/zero or accidental binary content.
ANTS-1013 — refreshReviewButton git-status QProcess
dedup. The remaining open gap from indie-review-2026-04-27's
"2 s status-timer redundant work" cross-cutting theme. New
m_reviewProbeInFlight member skips the git status --porcelain=v1 -b spawn when a previous probe is still alive;
cleared by both the finished and errorOccurred handlers.
STANDARDS.md and
RULES.md retired (user authorised). Both predated the
docs/standards/ bundle (2026-04-13/14) and duplicated content
living canonically at docs/standards/coding.md and
docs/standards/commits.md. Live references updated in
CONTRIBUTING.md, PLUGINS.md, README.md,
docs/RECOMMENDED_ROUTINES.md, and src/auditdialog.cpp
(the "Review with Claude" header now prepends the four
docs/standards/ files instead).ANTS-1121 — documentation drift fold-in reconciliation. The
9-theme finding list from doc-cold-eyes-2026-04-30 was largely
addressed in flight during the 0.7.50–0.7.59 cycle; this
release closes the bullet with a per-theme reconciliation
pass. T1 (GPU references in retired docs) folds into ANTS-1105
above; T6 (44-bullet Source: field backfill) deferred to
ANTS-1129 as a focused follow-up.
ANTS-1122 — audit fold-in fixes flipped to ✅; the four range-loop-detach / constVariablePointer / returnByReference fixes have been in code since the 0.7.59 / 0.7.60 cycle. This release just refreshes the roadmap narrative.
ANTS-1126 — three-pass cold-eyes review fold-in for the ANTS-1125 archive spec. 4 CRITICAL + 8 HIGH + 5 MEDIUM + 5 LOW on first pass; 0 CRITICAL + 3 HIGH on second pass; PASS on third pass with 3 LOW polish folded inline. Plus an INV-14 cluster review on the rotation script: 0 CRITICAL + 3 HIGH + 4 MEDIUM + 4 LOW, all folded.
ANTS-1007 → ANTS-1011 — indie-review cross-cutting themes
closed. Atomic-write drift (closed by 1016+1017),
frameless+translucent exec() (closed by 1015), argv --
separator + quote tokens (closed by 1024+1028+1030),
Lua + ssh permission allow-lists (closed by 1022+1024),
WCAG 1.4.1 colour-only state (closed by 1034+1035+1041) —
each cross-cutting bullet's narrative refreshed with the
constituent fix references.
Theme: ANTS-1123 indie-review Tier 2 + Tier 3 fold-in — the
remaining hardening + polish bullets from the 0.7.59 4-agent
independent review of the Claude Code companion features. Cache
contract on the roadmap-query IPC verb honours the spec
"≤ 100 ms" budget literally; AntsHelper rejects malformed JSON
input shapes; AuditEngine's parseFindings no longer emits bogus
SARIF locators for version-string tokens; RoadmapDialog gracefully
handles a corrupt persisted geometry blob and preserves user
state on Custom-tab clicks. CLAUDE.md streamlined (glrenderer
removal acknowledged, module map extended for the new modules
that landed in 0.7.59).
ants-helper now rejects non-object JSON request bodies ([],
42, "foo") with a usage error envelope instead of silently
treating them as {}. M1: parseFindings reJustFile regex
tightened to require an alphabetic-led extension — bare
version-string locators like cargo/1.75 no longer create
bogus SARIF physicalLocation entries. F1: roadmap-query
cache now honours the spec INV-10 "≤ 100 ms" wall-clock
bound on top of the existing mtime check, so an in-place edit
within the same mtime tick is still picked up. F9: ANTS-1117
spec § Error-response shape rewritten to the flat-envelope
convention ({ok:true, bullets:[...]}) the codebase actually
ships — earlier nested-on-data wording was the outlier.
LOW-1: corrupt persisted Roadmap-dialog geometry blob is now
cleared instead of masquerading as valid forever. LOW-3:
clicking the "Custom" tab on the Roadmap dialog now preserves
the user's checkbox tuning and sort order instead of resetting
to document order. LOW-2: tab-order array gained a
static_assert against the Preset enum size to catch future
enum drift at compile time. F4 / F7 / F8: helper docstring
usage-order corrected, stdin-open failure surfaces a usage
error instead of degrading to empty input, and Linux-only
platform note + TOCTOU note documented in antshelper.h. L1:
short-circuit regex chain in parseFindings (was matching
three regexes per line; now one). M2 / L2: comments added on
the applyFilter empty-line drop and the highConfidence
field's live cross-tool corroboration role.Theme: Claude Code companion features — faceted tabs +
search on the Roadmap viewer (ANTS-1100), two read-only IPC
verbs Claude can drive (tab-list, roadmap-query; ANTS-1117
v1), an optional standalone CLI helper (ants-helper drift-check;
ANTS-1116 v1), an audit-engine extraction that unblocks future
non-GUI audit consumers (ANTS-1119 v1), and a measurement gate
that grounds further companion work in real token counts
(ANTS-1120). Bundled with three rounds of fold-in fixes:
documentation cold-eyes review (ANTS-1121, T1-T9 across PLUGINS
/ STANDARDS / RULES / SECURITY / archival relocation), audit
findings (ANTS-1122), and indie-review findings (ANTS-1123, Tier 1
shipped in this cut). Strategic context in the new ADR-0002
(cold-eyes scope cleanup of the broader companion-feature plan).
Roadmap dialog: faceted preset tabs + search + persisted
geometry. Six tabs above the existing checkbox row —
Full / History / Current / Next / Far Future / Custom —
each maps to a (filter, sort) preset evaluated in pure
C++ over the already-parsed bullet model (zero LLM
round-trips). Search box accepts substring or id:NNNN
shorthand to jump to a specific bullet. Default size is now
1200×800 (was 900×700) with the user's resize persisted via
the new Config::roadmapDialogGeometry
saveGeometry/restoreGeometry round-trip. New static
helpers (RoadmapDialog::filterFor / sortFor /
presetMatching) plus a non-history Document sort that
preserves authored section order.
Locked by tests/features/roadmap_viewer_tabs/ (13 INVs).
(ANTS-1100)
Two new remote-control IPC verbs (read-only):
roadmap-query + tab-list. Claude Code can now interrogate
the running Ants Terminal for parsed ROADMAP bullets (id,
status, headline, kind, lanes per top-level
status-emoji bullet) and per-tab state (index, title,
cwd, shell_pid, claude_running, color) without
reading the underlying files or running ps. Both verbs
inherit the existing RemoteControl transport's UID-scope
trust (lstat + S_ISSOCK + UID match) and only run when
remote_control_enabled is on. New helpers:
RoadmapDialog::parseBullets (pure markdown → bullet
records), MainWindow::tabsAsJson (rich per-tab snapshot
alongside the existing slim tabListForRemote used by ls).
Cache on the roadmap-query path keyed on file mtime so
repeated calls don't re-parse the 4500-line ROADMAP.md.
Locked by tests/features/remote_control_roadmap_query/
(9 INVs) + tests/features/remote_control_tab_list/ (7 INVs).
(ANTS-1117 v1)
ants-helper CLI binary (optional, default OFF). New
standalone binary built when CMake is invoked with
-DANTS_ENABLE_HELPER_CLI=ON. v1 ships a single subcommand:
ants-helper drift-check wraps packaging/check-version-drift.sh
and emits a unified JSON envelope ({"ok": true, "data": {"clean": true|false, "violations":[…], "raw":"…", "exit_code": N}}) Claude can parse without re-running
the script's stdout through interpretation. Exit codes
per docs/specs/ANTS-1116.md § INV-8: 0=clean, 3=drift,
1=handler error, 2=usage error. Builds against Qt6::Core
only. v2 (audit-run, id-allocate, test-runner)
contingent on ANTS-1120 measurement validating per-call
token savings.
Locked by tests/features/local_subagent_framework/ (8 INVs).
(ANTS-1116 v1)
AuditEngine GUI-free module. Extracted the data-shape
types (CheckType, Severity, OutputFilter, Finding,
CheckResult, AuditCheck) and the three pure parsing
functions (applyFilter, parseFindings, capFindings)
from auditdialog.cpp (~5800 lines, mixed engine +
presentation) into src/auditengine.{h,cpp} depending on
Qt6::Core only. The dialog re-exports the data types
via a transitive include of auditengine.h; the dialog's
static helpers isCatastrophicRegex / hardenUserRegex
now forward to the engine (closes the divergence vector
the indie-review C-cluster surfaced — pre-fix the engine's
local copies had drifted from the dialog originals on
LIMIT_MATCH value, double-prefix guard, and shape
detector). sourceForCheck + computeDedup likewise
promoted to the engine's public surface. Pure refactor;
audit pipeline output is byte-identical pre/post.
Unblocks ANTS-1116 v2 / ANTS-1117 v2 audit-run (no
Qt6::Widgets in the linker path).
Locked by tests/features/audit_engine_extraction/ (9 INVs).
(ANTS-1119 v1)
Companion-instrumentation harness (ANTS-1120). New
scripts/measure-companion-tokens.sh skeleton + journal
placeholder at docs/journal/ANTS-1120-measurement.md. Per
ADR-0002 § Decision 5, before any further companion bullet
beyond ANTS-1117 v1 + ANTS-1116 v1 ships, this harness runs
one Claude task in two configurations (baseline vs stubbed
helpers) N≥3 times each, captures per-prompt token counts
from the API usage field, and produces a per-bullet
keep / iterate / drop / inconclusive verdict. v1 ships the
framework; the actual measurement run is gated on stub
capture + a representative scripted task. (ANTS-1120)
auditdialog.cpp slimmed by the engine extraction. The
three moved functions (applyFilter / parseFindings /
capFindings) now live in src/auditengine.cpp; the
dialog's orchestration path (onCheckFinished) calls
through AuditEngine:: directly. isCatastrophicRegex /
hardenUserRegex / sourceForCheck are unified — single
definition in the engine, dialog-side static helpers are
thin forwarders. (ANTS-1119 v1, ANTS-1123 indie-review
C-cluster + H1 fix-pass.)
ants-helper JSON envelope on usage errors. Both the
unknown subcommand and invalid JSON paths now emit a
{"ok": false, "error": "…", "code": "usage_error"} envelope
on stdout in addition to the human-readable stderr line.
Keeps the contract consistent — Claude consumers parsing
stdout no longer need to special-case usage errors.
(ANTS-1123 indie-review F1.)
ants-helper error-string match to spec INV-5.
bash-unavailable case now returns {"error": "bash unavailable", "code": "missing_bash"} (was: "could not start bash"). (ANTS-1123 indie-review F2.)
Kind: field now required on every actionable
ROADMAP.md bullet (previously optional, inferred from
section context). Backfilled all 100 unannotated bullets;
Source: added where non-default. Spec updated at
docs/standards/roadmap-format.md § 3.5 and synced to the
user-level /start-app template. Partial implementation of
ANTS-1106 — viewer faceted categorisation by Kind remains
📋. (ANTS-1106)
Documentation tree refresh (ANTS-1121, full-app cold-eyes review):
STANDARDS.md,
RULES.md, PLUGINS.md — glrenderer.cpp was removed
in 0.7.44; only the QPainter+QTextLayout path remains.PLUGINS.md (v0.6.9 → v0.7.58+),
SECURITY.md Supported Versions table (0.6.x → 0.7.x;
last-updated stamp), packaging/README.md (recipe-version
placeholder rather than a baked 0.6.20), README.md
install-snippet comment, CONTRIBUTING.md example commit
subject (<ID>: shape per docs/standards/commits.md).DISCOVERY.md, AUDIT.md,
FIXPLAN.md, project_audit_updates.md) relocated from
project root to docs/journal/ with "historical snapshot,
not current" headers and date prefixes; cross-refs in
EXPERIMENTAL.md and the
audit_2026_04_13_v2 agent-memory entry updated.docs/AUTOMATED_AUDIT_REPORT.md (17-day-stale
un-dated copy) retired; the dated archives at
docs/AUTOMATED_AUDIT_REPORT_2026-04-{11,13}_*.md remain.
docs/AUDIT_TRIAGE_2026-04-16.md gained a "historical
snapshot" header so readers don't mistake its 4 verified
findings for current state.## 0.7.0 / 0.7.7 / 0.7.12 / 0.7.50 headings re-tensed from "(target: 2026-NN)"
to "shipped 2026-MM-DD" (or the in-flight equivalent for
0.7.50–0.7.59).Audit-engine fold-in fixes (ANTS-1122, scoped /audit
pass): clazy range-loop-detach fixed in antshelper.cpp
drift-violation parser and in roadmapdialog.cpp::parseBullets
(bind split() results to a const QStringList local
before iteration); cppcheck constVariablePointer fixed in
auditengine.cpp::applyFilter (the cached fileLines slot
is read-through-pointer, written-through-reference);
cppcheck returnByReference fixed on
MainWindow::roadmapPathForRemote(). Strengthened
tests/features/roadmap_viewer_tabs/ test (debt-sweep
finding 2.x): added INV-9b non-history-sort negative case,
parsed-numeric INV-12 dimension check (previously a
static-string match), and INV-13 covering
sortFor(Custom) == Document.
AscendingDocument sort enum value retired as YAGNI
(debt-sweep finding 1.1) — it was an alias declared
"reserved for future use" with zero call sites; removed
rather than carried forward.
docs/standards/roadmap-format.md section reference
drift in newly-authored specs — internal cross-refs in
docs/specs/ANTS-1117.md, docs/specs/ANTS-1120.md, and
docs/decisions/0002-cold-eyes-companion-cleanup.md cited
§ 1.2 / § 3.1 / § 3.5 for facts that actually live
in § 3.3 (status emojis) and § 3.5.1 (stable-ID
immutability). Corrected during the cold-eyes review's
full-doc cross-check pass.isCatastrophicRegex / hardenUserRegex between
auditengine.cpp and auditdialog.cpp: differing
LIMIT_MATCH values (200000 vs 100000), no double-prefix
guard on the engine's version, and shape detectors that
caught different sets of catastrophic-backtracking patterns.
A pattern that triggered the bound on one path could slip
through on the other — the divergence-correctness vector
was the actual bug. Unified to a single definition exported
from the engine; both call sites delegate.
(ANTS-1123 indie-review C1+C2+C3+H1.)docs/decisions/0002-cold-eyes-companion-cleanup.md —
ADR for the scope reset of the Claude-Code companion
bundle (reorder ANTS-1116/1117 priority, retire ANTS-1110
catalogue + ANTS-1056 + ANTS-1114 wishlist, drop ANTS-1113
memory-drift category, register ANTS-1119 + ANTS-1120).docs/specs/ANTS-1116.md, docs/specs/ANTS-1117.md,
docs/specs/ANTS-1119.md, docs/specs/ANTS-1120.md —
per-bullet contracts (acceptance criteria, INVs, deliverables).Theme: runtime toggle for the ANTS-1054 dialog tracer — the 0.7.57 env-var-only path required restarting Ants Terminal to diagnose the user-reported mystery flashing dialogs. New menu action lets the user enable / disable the tracer mid-session without losing tab state.
DialogShowTracer the
ANTS_TRACE_DIALOGS=1 env var path installs. Output goes to
stderr (always) and to the events debug-log category when
enabled. Status-bar transient confirms the toggle. (ANTS-1054
follow-up)DialogShowTracer extracted from src/main.cpp to its own
src/dialogshowtracer.{h,cpp} translation unit so both the
startup env-var path and the runtime menu toggle call one
setActive(bool) entry point. Process-global instance is
parented to qApp; idempotent install / uninstall.Theme: indie-review backlog cleanup — surgical Tier 3 fixes
(shellQuote whitelist, claudeChildrenOf extraction), a
diagnostic for the mystery flashing-dialog report, a cppcheck
returnByReference perf sweep, and a ROADMAP audit that flipped
~30 prior indie-review items from 📋 to ✅ once verified shipped
in 0.7.52–0.7.55.
ANTS_TRACE_DIALOGS=1 env var → DialogShowTracer event filter
installed on the QApplication. Logs every top-level
QWidget / QDialog Show event with className, objectName,
windowTitle, parent class+objectName, and geometry to stderr
(and the events debug-log category when enabled). Diagnostic
step (a) of ANTS-1054 — the user reported a small popup
flashing periodically in the centre of the terminal; the trace
attributes the next occurrence to either an Ants spawn site
(logged) or a child process (silent trace). Zero cost when the
env var is unset. (ANTS-1054 step a)
tests/features/shellutils_whitelist/ — feature test for the
ANTS-1047 whitelist fix. Verifies the four invariants from
spec.md: empty input maps to '', the whitelist
[A-Za-z0-9_\-./:@%+,] passes through unchanged, anything
outside it forces single-quote wrapping (specifically the
glob/redirection/bracket characters the old denylist missed),
and embedded single quotes round-trip via the POSIX '\'' form.
Pure C++ harness, links Qt6::Core only.
ClaudeIntegration::findClaudeChildPid(pid_t shellPid) —
extracted from the two near-identical /proc/<pid>/task/<pid>/ children walks in pollClaudeProcess and
ClaudeTabTracker::detectClaudeChild. Both call sites now
share one helper; detectClaudeChild also gains the /proc-
scan fallback that previously only pollClaudeProcess had
(resilience on kernels / containers that don't expose the
task/<pid>/children file). Static method on
ClaudeIntegration, no global state, safe to call from any
thread. (ANTS-1048)
Cppcheck returnByReference perf sweep across seven hot
getters: Config::loadFailureBackupPath,
ClaudeBgTaskTracker::transcriptPath,
GhostLineEdit::ghostSuffix, TerminalGrid::windowTitle,
TerminalWidget::shellTitle, TerminalWidget::imagePasteDir,
PluginManager::pluginDir. Each was returning a QString by
value — now returns const QString &. Hot read on the 2 s
status-timer path (transcriptPath) is the most consequential,
but each saves a heap allocation per call. Same class as the
0.7.55 ClaudeBgTaskTracker::tasks fix.
shellutils.h::shellQuote switched from a denylist regex
[\s'"\\$!#&|;(){}]to a whitelist regex[A-Za-z0-9_-./:@%+,]+. The denylist missed glob wildcards (*, ?), redirection (<, >), and bracket-expansion specials ([, ]); a path containing any of those reached the shell unquoted and underwent globbing. Whitelist closes the class — anything outside the safe set gets single-quote wrapped. Affects every caller of shellQuote—mainwindow.cpp claude-launch builders (cd %1 && claude
--continueetc.) andsshdialog.cpp` ssh-argv construction.
(ANTS-1047)[ANTS-1015..1042, 1047, 1048]
bullets from 📋 to ✅ after verifying each item shipped in the
0.7.52–0.7.55 cycle (or in this 0.7.57 bundle). Prior to the
audit, the roadmap presented these items as outstanding even
though the source files carried the corresponding fix
comments. Each ✅ bullet now cites the shipped version + the
exact source-comment line so the trail is reviewable. ANTS-1054
flipped from 📋 to 🚧 with the diagnostic-shipped note. Tier 3
refactors ANTS-1043, 1044, 1045, 1046, 1049 remain 📋 (large-
scope items deferred — file decompositions and Wayland-
specific work need their own focused passes).0.7.56 — App-Build suite alignment
Theme: App-Build suite alignment — confirm-on-close for tabs
running non-shell processes, four-doc shareable standards bundle
synced to the user-level /start-app template, CI-blocking
metainfo escape fix, and the strategic roadmap for incorporating
the App-Build workflow natively into Ants Terminal so users can
run as much of it as possible without spending Claude tokens.
Config::confirmCloseWithProcesses to false for subsequent
closes. Default on. Settings UI in the Terminal tab.
Default safe-shell allowlist: bash, zsh, fish, sh, ksh,
dash, ash, tcsh, csh, mksh, yash. Probe walks
/proc/<pid>/task/<pid>/children transitively (cap 256
visited PIDs) — Linux-only, mirrors the existing
ClaudeIntegration descendant probe. Locked by
tests/features/confirm_close_with_processes/ (11
invariants).docs/standards/
(coding · documentation · testing · commits) plus an index
README, with the detailed ROADMAP and CHANGELOG format spec
carried as a sub-spec at docs/standards/roadmap-format.md.
Commits standard mandates <ID>: <description> subjects so
every commit links back to a ROADMAP item; testing standard
mandates TDD by default (ANTS-1055, ANTS-1104).docs/decisions/ with
0001-record-architecture-decisions.md (Michael Nygard's
format) plus per-folder README.docs/specs/ and docs/journal/ placeholder folders for
per-feature spec drafts and per-phase outcomes.Project standards section in CLAUDE.md pointing at
the five standards files and the /start-app,
/app-workflow, /close-phase skills (ANTS-1104)./audit, ctest,
drift checks, /debt-sweep, /bump, /release, fold-in
templating, atomic CHANGELOG/ROADMAP edits) runs in C++
with zero LLM round-trips. Three supporting roadmap items:
ANTS-1106 (mandatory Kind: + viewer faceted
categorisation), ANTS-1107 (adopt App-Build documentation
folder structure: glossary.md, known-issues.md,
audit-allowlist.md, ideas.md, design.md,
.claude/workflow.md), ANTS-1109 (status-bar git-branch
chip restyle to match the Public/Private repo pill).docs/standards/roadmap-format.md for the format spec (was
docs/ROADMAP_FORMAT.md originally; briefly lived inline in
documentation.md § 3 before being extracted as a sub-spec
for token efficiency under ANTS-1104).docs/standards/ now byte-identical to the user-level
/start-app template at
~/.claude/skills/app-workflow/templates/docs/standards/,
with per-language idiom examples and push-policy details
delegated to the global ~/.claude/CLAUDE.md rather than
duplicated in the project standards (ANTS-1104).docs/standards/,
docs/decisions/, docs/specs/, docs/journal/; flags
STANDARDS.md / RULES.md as deprecated (retire under
ANTS-1105 with explicit user confirmation).[Unreleased] block
that the Roadmap dialog reads for current-work signaling
(roadmap-format.md § 4.1 mandates one always, even empty).& in the 0.7.55 metainfo <release> body broke
appstreamcli validate and turned CI red on every commit
since the release. Now &. (ANTS-1099)0.7.55 — VT parser correctness + audit hardening
Theme: VT parser correctness + audit-dialog hardening from the 2026-04-27 indie-review. Six fixes — multi-row OSC 8 hyperlink span emission, ITU/ECMA-48 colon-RGB form parsing, audit comment-suppress hyphen handling, trend-snapshot dedup, bg-tasks liveness sweep split from full reparse, and a hot-path returnByReference fix.
38:2::r:g:b dropped a channel. Parser now detects the 4-slot colon group (colorspace + R + G + B) and reads R/G/B from the right offsets. Legacy 38;2;r;g;b and 3-slot 38:2:r:g:b continue unchanged.// nosemgrep: bash-c-non-literal matched only bash because the terminator class included -. Now [)\]]|$.renderResults which fires on every severity-pill toggle. Now gated by m_snapshotPersisted, reset in runAudit().sweepLiveness() walks m_tasks via stat()+mtime check; the 2 s status-timer drives it instead of the 16 MiB transcript reparse. File watcher continues to drive full rescan() on transcript-changed.ClaudeBgTaskTracker::tasks() returned by value on hot path. cppcheck returnByReference. Now const &.0.7.54 — A11y + UX bundle (indie-review fold-in)
Theme: A11y + UX bundle from the 2026-04-27 indie-review. Six fixes — accessibility plumbing on the status bar + ToggleSwitch, plan-mode persistence across tab switches, an SSE iteration cap to prevent UI freeze on misbehaving AI endpoints, and IPv6 support in the SSH Quick Connect parser.
setAccessibleName. Branch chip, repo visibility chip, foreground-process label, status-message slot, Claude session label, Claude context bar, Review Changes button. Screen readers (Orca / NVDA / VoiceOver) now announce each chrome widget with semantic text instead of raw tab content or Powerline glyph codepoints. The Claude session label additionally updates accessibleDescription on every state transition so the screen reader announces the current state.ToggleSwitch accessibility plumbing. Default accessibleName + QAccessibleStateChangeEvent on every check-state flip (checkStateSet / nextCheckState). Keyboard users now get immediate AT-SPI / UIAutomation confirmation when their toggle lands.m_planModeByPid and restores it on tab switch back, so the indicator no longer flickers off→on when the transcript-tail window doesn't include the latched permission-mode event.singleShot(0). Caps at 256 lines per tick (covers ~8 seconds of legitimate streaming output before yielding); pathological bursts yield to the event loop. Prevents UI freeze on misbehaving endpoints.[2001:db8::1]:2222 no longer mis-parses as host=[2001 port=0. Matches RFC 3986 §3.2.2 / OpenSSH command-line convention.0.7.53 — Tier-1 remainders + VT/paste/plugin hardening
Theme: Tier-1 remainders + VT/paste/plugin hardening from the 2026-04-27 indie-review. Six security-class fixes — three completing the Tier-1 ship-this-week list, three from Tier-2 paste/image hardening.
ESC c triggered RIS (full terminal reset); ESC D triggered IND; ESC 7 / ESC 8 triggered DECSC/DECRC. RCE-adjacent because OSC payloads can arrive from a hostile remote shell, trigger-rule expansion, or pasted content. Fixed at the C0 pre-handler — string-state ESC no longer transitions to Escape; new OscStringEsc / DcsStringEsc / ApcStringEsc / IgnoreStringEsc peek-states consume the trailing byte and return to Ground. Locked by new tests/features/vt_osc_esc_discard/ (5 invariants × 11 sub-cases).col + 32 produced bytes ≥ 0xE0 when col exceeded 223; apps reading the emit stream mis-framed subsequent click bytes as UTF-8 continuations. Both wheel-event + mousePressEvent SGR fallback now clamp col/row to 223 in X10 mode.knownPermissions() (clipboard.write, settings); prompt-result is intersected with requested set before reaching the engine.m_imagePasteDir canonicalised + required under $HOME; UUID4 suffix on filename to prevent millisecond-collision clobbers.\r terminator spoofed dialog. Preview now normalises \r\n and bare \r to \n for display; actual paste still writes original bytes verbatim.\x9B[200~ not stripped. Now strips 7-bit (\x1B[), raw 8-bit (\x9B), and UTF-8-encoded 8-bit (\xC2\x9B) forms.tests/features/vt_osc_esc_discard/ — 5 invariants × 11 sub-cases.0.7.52 — indie-review tier-1 sweep (data-loss + Wayland + secret-leak)
Theme: First Tier-1 sweep of the 2026-04-27 indie-review findings. Eight fixes covering CRITICAL data-loss / security regressions plus HIGH security-class hardening — SessionManager silent scrollback loss, the same Wayland modal-grab pattern from 0.7.50 lurking in the update-confirmation dialog, SARIF/HTML reports leaking secrets at 0644, and five smaller HIGH items. The user-reported GitHub repo-type chip regression is also resolved (verified visible).
SessionManager silent data loss. saveSession + saveTabOrder used QFile::rename, which on POSIX refuses to overwrite an existing destination. Every session save AFTER the first silently failed: the .dat file held the original snapshot, .dat.tmp accumulated each new write, user scrollback never updated past the first save. Switched both to std::rename (POSIX rename(2) — atomic replace), mirroring the 0.7.12 Config fix. Errno logged on failure; orphaned .tmp removed.QMessageBox box(this); box.exec() — exactly the QTBUG-79126 / QTBUG-90005 click-drop pattern. Converted to the proven non-modal QDialog + plain QPushButton + clicked()→close() shape.gitleaks / secrets_scan rules. Switched to QSaveFile + commit() + setOwnerOnlyPerms (0600).new-tab / launch IPC commands bypassed the send-text C0 filter. Both now route through filterControlChars by default; raw: true opt-out preserved.file:// and ftp:// schemes removed from the hyperlink allowlist. xdg-open file:///foo.desktop autoexecutes via the desktop's .desktop handler — RCE-adjacent. Allowlist now covers only http, https, mailto.extraArgs quote-bypass on the dangerous--o allowlist. Switched to QProcess::splitCommand for proper POSIX shell quoting.extractCwdFromTranscript unbounded readLine. Caps at 64 KiB.file:// / gopher:// / bare-host schemes up-front.openFileAtPath argv-injection via attacker-controlled paths starting with -. Now prepends ./ for combined-path branches and inserts -- for separate-path branches.tests/features/persistence_post_rename_chmod/ INV-2 revised to accept either QFile::rename or std::rename + rc==0 gating.tests/features/github_status_bar/ INV-17 revised to lock the Wayland-correct dialog shape (forbid QMessageBox box(this), require new QDialog(this) + &QPushButton::clicked + &QDialog::close).tests/features/remote_control_new_tab/ INV-3c/3d added.tests/features/remote_control_launch/ INV-4b/4c added.0.7.51 — config-reload inotify-loop fix
Theme: Hot-reload doesn't loop any more. Same-day follow-up to 0.7.50 — three reported symptoms (status bar permanently sticking at "Config reloaded from disk", Help → Check for Updates appearing to do nothing, Settings → Preferences not opening) all traced to a single root cause: MainWindow::onConfigFileChanged re-entering itself in an infinite inotify loop. The 0.7.31 attempt at fixing the same loop (m_configWatcher->blockSignals(true/false) bracketing) never worked because Qt reads inotify events from the event loop after the slot returns — outliving the blockSignals window. Fix is at the source: make the setters idempotent so the slot has nothing to re-trigger.
Config reload re-entered itself in an inotify loop. Config::setTheme unconditionally called save() even when the value matched, so any applyTheme(m_config.theme()) from inside onConfigFileChanged rewrote the watched file → kernel inotify event → next event-loop tick re-emits fileChanged → slot re-enters → loop. setTheme now early-returns when the value matches, breaking the loop at its source. onConfigFileChanged additionally skips the no-op applyTheme call when the theme didn't change, and carries an m_inConfigReload re-entrancy flag (cleared via QTimer::singleShot(0, ...) so a save inside the slot is dropped but a subsequent genuine external edit is honored). The failed 0.7.31 blockSignals calls are removed.
User-visible cascade now resolves:
Help → Check for Updates produces a visible "Checking for updates…" then "Up to date — running v0.7.51 (latest)" toast (or the actual update notice when behind).Settings → Preferences opens the dialog instead of having it deleted-on-reload faster than the user can click.Locked down by tests/features/config_reload_loop_safety/ — INV-1 idempotent setter (call shape + functional mtime check), INV-2 no-op skip at the call site, INV-3 re-entrancy guard, INV-4 failed blockSignals attempt removed.
0.7.50 — Wayland-correct dialog dismissal
Theme: Round four on the same dialog-button bug — and the matching Roadmap-dialog Close button. Three prior fix attempts (0.7.22, 0.7.35, 0.7.49) each diagnosed a downstream symptom rather than the actual root cause; this release identifies it from upstream Qt bug reports and applies the fix that the bg-tasks dialog has been quietly using all along.
Both Help → About OK buttons (still) didn't dismiss the dialog in 0.7.49, and the Roadmap dialog's Close button hadn't worked since 0.7.43 either. User report 2026-04-28: "OK buttons still do nothing on the 2 About dialogs… Close button does nothing on the Roadmap either."
Real root cause: both upstream Qt bugs QTBUG-79126 ("Dialogs behavior on Wayland is wrong") and QTBUG-90005 ("global modality on Wayland") document that Wayland's xdg-shell protocol has no equivalent of Qt::ApplicationModal. Calling setModal(true) (which 0.7.49 added) is a no-op on Wayland except for an aggravating side effect: KWin/Wayland routes click events into the modal-grab handler which then drops them, instead of delivering them to the dialog's button. The role-based dispatch path inside QDialogButtonBox is a known second aggravator on the same bug.
The Background Tasks dialog has been working all along on the same parent-window flags — by being non-modal, with a plain QPushButton whose clicked() is wired directly to QDialog::close. Both About dialogs and the Roadmap dialog now follow that proven shape: non-modal, plain QPushButton, direct clicked() → close(). No more setModal(true), no more QDialogButtonBox in these handlers. The OK / Close click reaches its slot and the dialog dismisses on the first try.
Spec coverage extended: tests/features/help_about_menu/ INV-4 now asserts a plain QPushButton (negative grep on new QDialogButtonBox); INV-7f flips from "must call setModal(true)" to "must NOT call dlg->setModal(". tests/features/roadmap_viewer/ INV-14 mirrors the same. The regression history in help_about_menu/spec.md records the Wayland-modal root cause so the next contributor doesn't repeat the cycle.
🤖 AppImage builds via the existing release workflow. Wait for the workflow to attach the AppImage to this release.
Theme: Three follow-up fixes to user reports against 0.7.48 — the About menu's OK buttons silently no-op'd again, the GitHub Public/Private badge wanted to live next to the git branch on the left, and the Background Tasks chip showed phantom counts of tasks that had long since exited.
Both Help → About dialogs now dismiss on the OK button. The
0.7.35 attempt at this bug only swapped the dialog type (from
QMessageBox to a custom QDialog); both dialogs still went
through exec(). On KDE/KWin + Qt 6.11 with our frameless +
WA_TranslucentBackground MainWindow, the nested QEventLoop
exec() opens does not surface the dialog as the active input
window — the OK button never receives focus, the click silently
no-ops, and the user has to dismiss via the WM close button (the
same symptom shape as the 2026-04-25 report).
Both About handlers now use the heap + Qt::WA_DeleteOnClose +
show() + raise() + activateWindow() pattern that the
Background Tasks, Settings, and Claude Transcript dialogs already
use successfully under the same parent-window flags. The OK
button's clicked() signal is also wired directly to
QDialog::accept in addition to the existing
QDialogButtonBox::accepted connection — belt-and-braces in case
any platform plumbing interferes with the button box's internal
accepted-emission. The "About Qt" entry dropped
QMessageBox::aboutQt (whose internal exec() is the same
regression shape) for a custom QDialog with the same heap+show
treatment, surfacing the Qt runtime + build versions and a link
to qt.io/licensing. Locked by tests/features/help_about_menu/
spec INV-7 (heap+show pattern, with negative grep on
QDialog dlg(this) + dlg.exec() and on
QMessageBox::aboutQt).
Background Tasks chip now reflects only tasks that are
genuinely running. The transcript-only completion detection
used through 0.7.48 only marked tasks finished when Claude Code
recorded a matching KillShell event or a BashOutput
tool_result with status=completed/killed/failed. Background
tasks that were spawned and never polled — the assistant moved
on without checking their output — stayed finished == false
indefinitely, producing a phantom running-count chip (12 tasks
showing on a session with zero genuinely-running tasks per the
user report).
parseTranscript now performs a liveness sweep at the end: each
unfinished task whose outputPath either no longer exists OR
whose mtime is older than a 60 s staleness window is flipped to
finished. The 60 s window is generous on purpose — a slow CMake
configure or link step can have 30+ s of silence between progress
prints. ClaudeBgTaskTracker::rescan was promoted from a private
slot to public slots: and the 2 s status timer drives
MainWindow::refreshBgTasksButton, which calls rescan()
directly when the resolved transcript path is unchanged
(setTranscriptPath short-circuits on same-path so it can't be
the entry point for the periodic sweep). Result: the chip falls
to 0 within ~60 s of the last task going idle. Locked by
tests/features/claude_bg_tasks_button/ spec INV-12.
addPermanentWidget, shipped 0.7.45) where it sat alongside
the Claude Code chrome; the user asked 2026-04-27 to move it next
to the branch where the two badges read as a "branch ·
visibility" pair. Restyled to match the branch chip — same
rounded bgSecondary fill + border-1px frame — while keeping
the green-for-public / red-for-private foreground colour from
0.7.45 for the at-a-glance visibility cue. Locked by
tests/features/github_status_bar/ spec INV-2 (negative grep on
addPermanentWidget(m_repoVisibilityLabel) — must be on the
left via addWidget).Theme: Two tab-bar fixes the user spotted after 0.7.47 shipped. First, the per-tab Claude Code activity dot was only painting on one tab — the most recently focused one — even when multiple tabs each had their own running Claude Code session. Second, the dot sat too close to the tab text. The first one was a real correctness bug, the second is comfort.
Per-tab Claude state dot now renders on every tab with a live
Claude Code session, not just the most recently focused one.
ClaudeTabTracker::detectClaudeChild was assigning each shell
the system-wide newest *.jsonl as its transcript path on
first detection — so all N tabs running Claude ended up watching
the same file, and the inverse path→pid map (m_pathToShell)
was last-write-wins, fanning the QFileSystemWatcher::fileChanged
signal to a single shell. Result: only the last-tracked tab's
dot ever updated; the others stayed on their initial-detection
state.
The 0.7.44 fix to ClaudeIntegration::activeSessionPath (which
walks up projectCwd, encodes each ancestor, and probes
~/.claude/projects/<encoded>/ for the deepest match) already
had the right shape — it just wasn't reused by the tracker.
This release extracts that walk-up as a static helper
ClaudeIntegration::sessionPathForCwd(cwd) and calls it from
detectClaudeChild. The cwd comes from /proc/<claudePid>/cwd
(the directory Claude itself encodes into its project dir on
launch), with /proc/<shellPid>/cwd as a fallback. Two shells
in two distinct project trees now end up with two distinct
transcript paths and two independent state streams.
Pinned by INV-8 in
tests/features/claude_tab_status_indicator/spec.md: a
source-grep half asserts the tracker calls sessionPathForCwd
and reads /proc/<pid>/cwd; a round-trip half builds a
synthetic ~/.claude/projects/ tree with two encoded subdirs
and verifies each cwd resolves to its own subdir's .jsonl.
QTabBar::tab stylesheet → padding: 6px 16px 6px 22px). The
Claude state dot's center moves from r.left() + 8 to
r.left() + 11 so it stays gutter-centered, leaving ~7 px
between the dot's right edge and the first character of the
tab text instead of the previous 4 px. User feedback 2026-04-27:
"add a little space between the tab label and the Claude Code
status dot." Tabs without a Claude indicator pick up the same
widened leading padding; the visual change is subtle and the
bar still looks balanced.Theme: Smaller-footprint update checks + a pre-update warning dialog. Two pieces of user feedback after 0.7.46 shipped, both addressed in this release: (1) "An hourly check I think is a bit much. Let's do the check when the terminal is opened and when the user clicked on Help > Check for Updates." (2) "Before an update is processed, it should warn the user that it will be restarting the terminal. Any Claude Code sessions currently running will need to be reconnected." Cadence dropped to startup + manual; the update click now opens a confirmation dialog explaining the quit-and-relaunch consequence before kicking the updater.
Help → Check for Updates menu action (objectName
helpCheckForUpdatesAction). Click → 2-second Checking for updates… status-bar message + a userInitiated=true call to
checkForUpdates. The result lands as a status-bar message:
Up to date — running v0.7.47 (latest) if no newer release,
Update check failed: <error> on network error, or the existing
badge-show flow when a newer release exists.
Pre-update confirmation dialog in
MainWindow::handleUpdateClicked. Before the QProcess:: startDetached(updater, ...) call, a QMessageBox asks
"Download and install the new version now?" with informative
text explaining (a) the new version is written alongside in the
background, (b) the user must quit and re-launch to start using
it, (c) active Claude Code sessions will be disconnected and
need to be reconnected after the restart. The Cancel branch
short-circuits the spawn and surfaces "Update cancelled." in
the status bar so the click registers a visible
acknowledgement.
MainWindow::checkForUpdates(bool userInitiated = false).
New userInitiated parameter (default false). Background
startup probe stays silent on negative results
(userInitiated=false); manual triggers from the Help menu
surface their result so the user sees something happen.m_updateCheckTimer member + its 1-hour setInterval +
start. The hourly background poll is gone; only startup +
manual remain.tests/features/github_status_bar/ — INV-9 revised + new
INV-17 added (16 → 17 invariants):m_updateCheckTimer must be absent from
the header; the 5 s singleShot startup probe still required;
helpCheckForUpdatesAction objectName must exist in the
Help menu setup; action's connect must call
checkForUpdates(/*userInitiated=*/true).handleUpdateClicked must construct a
QMessageBox before the QProcess::startDetached call
(Cancel must short-circuit). The dialog text must mention
"quit and re-launch" AND "Claude Code" AND "reconnected" so
the warning is load-bearing rather than a generic confirm
prompt. Cancel must surface a "Update cancelled." status
message.api.github.com/.../releases/latest per session (instead of
per session + once per hour).Theme: Phase B of the auto-update story — actual in-place binary updates via AppImageUpdate / zsync. The 0.7.45 status bar notifier knew about new releases but only opened a download page; this release closes the loop. Click "↗ Update vX.Y.Z available" and the AppImage updates itself. v0.7.46 is the first release whose binary can be updated in place — pre-0.7.46 binaries shipped without the embedded zsync metadata and continue to be manual-download only (no way to retroactively add the ELF note).
UPDATE_INFORMATION ELF note + .zsync sidecar.
.github/workflows/release.yml now passes
UPDATE_INFORMATION="gh-releases-zsync|milnet01|ants-terminal|latest|Ants_Terminal-*-x86_64.AppImage.zsync"
to the linuxdeploy invocation. linuxdeploy embeds the string into
the AppImage's update-info ELF note AND produces a
<output>.zsync sidecar file. The upload step uploads BOTH the
AppImage and the .zsync sidecar to each release. The wildcard
pattern resolves against latest so the embedded URL is
version-stable — every shipped binary auto-updates to whatever's
tagged latest at the time the user runs the updater.
MainWindow::handleUpdateClicked slot wired to
m_updateAvailableLabel's linkActivated signal (replaces
the prior unconditional setOpenExternalLinks(true)). Probe
order:
AppImageUpdate (GUI) — preferred when present, gives the
user a progress window they can dismiss.appimageupdatetool (CLI) — silent fallback.QDesktopServices::openUrl — only when neither updater is
installed or $APPIMAGE is unset (binary isn't running as
an AppImage).When a tool is found AND $APPIMAGE is set, it's launched via
QProcess::startDetached so the updater outlives the parent
process. A status-bar message acknowledges the launch.
setOpenExternalLinks(false). Was true in
0.7.45 — clicks now route through handleUpdateClicked instead
of letting Qt auto-open the URL.tests/features/github_status_bar/ extended from 12 → 16
invariants:linkActivated to
MainWindow::handleUpdateClicked, no longer sets
setOpenExternalLinks(true).UPDATE_INFORMATION with the
gh-releases-zsync|milnet01|ants-terminal|latest|... schema.${OUTPUT}.zsync sidecar.handleUpdateClicked probes AppImageUpdate
(GUI) and appimageupdatetool (CLI); reads $APPIMAGE
env var; falls back to QDesktopServices::openUrl.QProcess::startDetached..zs-old suffix on the previous version) — the running
process keeps the old contents in memory until the user
restarts. Quit + relaunch picks up the new version.Theme: GitHub-aware status bar — two new badges that surface context the user previously had to alt-tab to a browser to see. A Public/Private repo visibility badge for the active tab's project, and an "↗ Update vX.Y.Z available" notifier that checks the GitHub releases API hourly and lets the user click through to the release page when a newer version exists.
m_repoVisibilityLabel — Public/Private repo badge. Small
QLabel (objectName "repoVisibilityLabel") added next to the
Roadmap button on the status bar. Theme-derived foreground:
th.ansi[2] green for public repos, th.ansi[3] amber for
private. Tooltip <owner>/<repo> on GitHub. Per-tab via
refreshStatusBarForActiveTab → new refreshRepoVisibility():
walks the active tab's shellCwd() up looking for a .git
ancestor, parses [remote "origin"] url from .git/config
(handles both https://github.com/owner/repo and
git@github.com:owner/repo URL forms), then resolves visibility
via gh repo view <owner>/<repo> --json visibility -q .visibility.
Result is cached by repo root with a 10-minute TTL. Failure
modes (no .git, non-GitHub origin, gh missing,
unauthenticated, network error) all hide the label.m_updateAvailableLabel — clickable update notifier.
Cyan-coloured QLabel (objectName "updateAvailableLabel",
setOpenExternalLinks(true), RichText) shown when a newer
GitHub release exists. New checkForUpdates() slot hits
https://api.github.com/repos/milnet01/ants-terminal/releases/latest
via QNetworkAccessManager, parses tag_name, and compares
against ANTS_VERSION via the new pure helper compareSemver.
An hourly m_updateCheckTimer plus a 5-second singleShot on
startup keep the badge fresh. Click opens the release page in
the user's browser. Sets a non-empty User-Agent header
(GitHub 403s requests without one).mainwindow.cpp (anonymous namespace):
findGitRepoRoot(start) walks up looking for a .git;
parseGithubOriginSlug(repoRoot) extracts owner/repo from
the origin URL; compareSemver(a, b) compares two X.Y.Z
triples component-wise as integers (so 0.10.0 correctly
outranks 0.9.0).tests/features/github_status_bar/ — 12 invariants on the
bundle:addPermanentWidget-ed, and start hidden.findGitRepoRoot,
parseGithubOriginSlug, compareSemver) exist.parseGithubOriginSlug recognises both URL forms.refreshStatusBarForActiveTab calls the new
refreshRepoVisibility.refreshRepoVisibility hides the label on every
failure branch (≥ 4 hide() call sites).gh invoked with the minimal --json visibility -q .visibility slug.singleShot first run.checkForUpdates hits releases/latest with a
User-Agent header.setOpenExternalLinks(true).compareSemver splits on . and uses toInt so
component compares are numeric (not lexicographic)..zsync sidecar AND embed the gh-releases-zsync|...
update-information string in the linuxdeploy command. Cannot
retroactively update binaries that didn't ship with the metadata.
This release only notifies; the user clicks through to download
manually.Theme: Two themed cleanups in one bundle. Retiring the dormant GPU-accelerated glyph-atlas renderer (compiled-but-unreachable since 0.7.4) and its no-op user-facing chrome (Settings checkbox + View menu action) — ~900 LoC removed, retires ROADMAP § "Renderer subsystem decision". Plus scoping the Background Tasks dialog to the active tab's project tree so sessions running in another project's window don't leak into this window's button.
ClaudeIntegration::activeSessionPath() walked
~/.claude/projects/ system-wide and returned the most-recently-
modified .jsonl, which meant a busy session in one project's
window would surface its bg-tasks count on a sibling window
pointed at a different project. The method now accepts a
projectCwd argument; the active tab's shellCwd() is encoded
via encodeProjectPath and the helper walks up the cwd
(cdUp) probing each ancestor's ~/.claude/projects/<encoded>/
subdir, returning the deepest match's newest .jsonl. Empty
projectCwd falls back to system-wide newest (kept for callers
that genuinely want it; nothing in tree currently does).src/glrenderer.{h,cpp} (~900 LoC). The optional GPU-
accelerated glyph-atlas renderer with GLSL 3.3 shaders +
instanced quads + a per-cell vertex buffer. Disabled in 0.7.4
when TerminalWidget switched from QOpenGLWidget to plain
QWidget; sat compiled-but-unreachable for 39 releases. ROADMAP
framed the choice as "revive via createWindowContainer or
delete" — picked delete. The QPainter+QTextLayout path remains
the only paint path and is fast enough for every shipped corpus.gpu_rendering Config key + Settings checkbox + View menu.
Config::gpuRendering() / setGpuRendering(), the Settings →
Appearance "GPU rendering (glyph atlas + GLSL shaders)"
checkbox, and the View → "GPU Rendering" menu action all wrote
to the dormant bool with no observable effect since 0.7.4. All
three are removed in this release. Existing config files with
"gpu_rendering": true are silently ignored — Config doesn't
validate unknown JSON keys.TerminalWidget::setGpuRendering(bool) / gpuRendering() +
the m_glRenderer / m_gpuRendering members. The
setGpuRendering setter was a no-op stub; both are gone.m_glRenderer cleanup in the TerminalWidget destructor and
paintEvent chain.tests/features/claude_bg_tasks_button/test_claude_bg_tasks_button.cpp
extended from 10 → 11 invariants:
activeSessionPath(const QString &projectCwd),
(b) implementation references encodeProjectPath AND cdUp
(the walk-up logic — no shortcut to a single string match),
(c) MainWindow::refreshBgTasksButton reads
focusedTerminal()->shellCwd() and passes it through to
activeSessionPath(cwd).Spec updated with the matching contract entry (linking the user-feedback quote 2026-04-27).
tests/features/settings_restore_defaults/spec.md updated to
drop the no-longer-applicable gpuRendering off clause from
Invariant 2's reset-default catalogue.
createWindowContainer + QOpenGLWindow). If that comes
back, it'll be a fresh design with current Qt 6 idioms, not a
resurrection of the 0.7.4-era code.Theme: Roadmap viewer polish — the dialog the previous release introduced gets a working Close button and a dynamic section-navigation sidebar that lists every heading in the active ROADMAP.md and lets the user click to jump. The TOC is built per project from whatever headings the file happens to have (no hardcoded section names) so the same dialog adapts to every repo that opens with a roadmap.
QSplitter(Qt::Horizontal)
inside RoadmapDialog puts a QListWidget (objectName "roadmap-toc") on the left and the rendered viewer on the right.
The list is rebuilt from the markdown on every refresh — including
live-update events on ROADMAP.md change — and entries are
indented two spaces per level above 1 with level-1 headings shown
bold. Click or activate an entry → the viewer scrolls to the
matching heading.RoadmapDialog::renderHtml now
prepends <a name="roadmap-toc-N"></a> before each <h1>/<h2>/
<h3>/<h4> it emits. The N-th anchor matches the N-th entry
returned by the new pure helper RoadmapDialog::extractToc, which
returns a QVector<TocEntry> of {level, text, anchor} triples.
Both helpers are pure and static so they're driven by the
feature-conformance test without spinning a Qt widget tree.QTextBrowser in place of QTextEdit. Required for
scrollToAnchor support. Links are pinned non-navigable
(setOpenLinks(false) / setOpenExternalLinks(false)) so a stray
markdown link can't replace the document.QDialogButtonBox::rejected signal
alone; under some Qt 6 builds the standard Close button does not
trigger that signal, leaving the button looking active but doing
nothing. The Close button is now wired both via rejected()
(kept as belt-and-braces) and via a direct clicked connection on
the underlying QPushButton retrieved with
button(QDialogButtonBox::Close). The direct connect is the
authoritative fix.tests/features/roadmap_viewer/test_roadmap_viewer.cpp extended
from 10 → 14 invariants:extractToc returns a 5-entry list with matching
levels, raw text, and roadmap-toc-N anchors against a known
multi-level input.renderHtml emits an <a name="roadmap-toc-0">
immediately before the first <h2> so scrollToAnchor can
position the viewer at any TOC entry.roadmapdialog.cpp confirms a
QSplitter, the roadmap-toc objectName, and a
scrollToAnchor invocation in the TOC handler.QAbstractButton::clicked, guarding against the
Qt 6 rejected()-only quirk.QListWidget)
rather than tree-shaped (QTreeWidget) — flat-list with leading
whitespace renders the hierarchy in O(1) widget cost without
per-node expand/collapse state, which would invite sync bugs on
live-update refreshes.0.7.42 — Command Palette ghost-completion (Tab to commit /slash-style suggestions)
Theme: Inline ghost-text completion in the Command Palette —
Claude Code's /slash-command UX, ported to Ctrl+Shift+P. As you
type, the unmatched suffix of the top fuzzy-match is rendered in a
dimmed colour right after the cursor; Tab commits the suggestion
into the input. Retires ROADMAP § "Command Palette ghost-completion
(near-term, small scope)".
GhostLineEdit subclass of QLineEdit. Lives in
src/commandpalette.h. Carries a m_ghost string set via
setGhostSuffix(); paintEvent calls
QLineEdit::paintEvent(event) first, then opens a fresh QPainter
and draws the ghost at cursorRect().right() + 1 using
palette().color(QPalette::Text) with setAlphaF(0.45). No
layout changes — the suffix overlays the existing line-edit content
area, so palette geometry is unchanged.
CommandPalette::updateGhostCompletion(filter). Invoked from
populateList after the result list is rebuilt. Looks at
m_list->item(0), recovers the underlying QAction, strips &
accelerators, and:
setGhostSuffix("");setGhostSuffix(name.mid(filter.length())), which
preserves the action name's original casing in the ghost.contains()-only matches (where the filter appears mid-string in
the top match) get an empty ghost — the visual contract is that
the suffix appears flush after the user's typed input, which only
makes sense for prefix matches.
Tab-key commit handler. eventFilter adds
Qt::Key_Tab → commitGhost(), which appends the ghost suffix to
the input via setText. The follow-up
textChanged → filterActions → populateList → updateGhostCompletion
cycle clears the ghost in the same dispatch since the new filter
equals the action name's tail exactly. The post-commit text equals
the visible composition (user-typed prefix + ghost suffix), so
user-typed casing is preserved — same as shell-completion
semantics. Tab is always consumed by the palette (even when
the ghost is empty) so focus cannot leak out of the input while
the palette is open. Tab does not also execute the action — the
user presses Enter to run, matching Claude Code's
/slash-completion contract.
tests/features/command_palette_ghost_completion/ — ten
invariants spanning the contract: I1 the input is a GhostLineEdit
findable by objectName; I2 empty filter → empty ghost; I3
prefix "ind" → ghost "ex Review" (first item in setActions
order is "Index Review"); I4 uppercase "INDEX" →
ghost " Review" (case-insensitive prefix, casing preserved
from the action name); I5 contains-only match → empty ghost;
I6 no-match → empty ghost; I7 Tab commit → text =
filter + ghost ("index Review") and ghost cleared; I8 Tab
with empty ghost is consumed (text unchanged, focus retained on
input); I9 source-grep on commandpalette.cpp (exactly one
setAlphaF(0.45 literal, ≥1 cursorRect( reference); I10 Esc
still hides the palette and emits closed() exactly once
(regression guard on existing dismiss behaviour). Drives
CommandPalette directly under QT_QPA_PLATFORM=offscreen —
no MainWindow link, no PTY. Pre-fix verification: with
commandpalette.{h,cpp} stashed, the test fails to even compile
(missing GhostLineEdit symbol); restored, test passes.💭 with deferral past 1.0. It
would touch terminalgrid.cpp paint path and vtparser.cpp
rather than commandpalette.cpp.0.7.41 — A11y for chrome (accessible names + AT-SPI introspection lane)
Theme: Accessibility — explicit setAccessibleName /
setAccessibleDescription on every glyph-only chrome control so
Orca, Speakup, and other AT-SPI screen readers announce them by
purpose ("Close window") instead of by codepoint name ("eight-spoked
asterisk", "multiplication sign x"). Two ROADMAP items retired in
one bundle: § "Accessibility pass on chrome" (mainwindow / palette /
title-bar control labels) and § "AT-SPI introspection lane" (an
automated check that every user-visible chrome control carries a
name).
Accessible names on TitleBar window controls.
centerBtn ("Center window" / "Center this window on the active
screen"), minimizeBtn ("Minimize window"), maximizeBtn
("Maximize window" / "Maximize or restore this window"), and
closeBtn ("Close window") each carry an explicit
setAccessibleName and setAccessibleDescription set immediately
after setText in titlebar.cpp:54-97. Each button also gains a
stable objectName (the previously-set closeBtn is joined by
the matching three siblings) so the introspection test can find
them deterministically. Visual rendering is unchanged — the glyphs
(✥ – □ ✕) still drive the on-screen pixels; only the AT-SPI tree
changed.
Accessible names on the Command Palette. m_input
(objectName commandPaletteInput) gains Command palette search
/ Type to filter actions; Tab to commit; Esc to dismiss.
m_list (objectName commandPaletteList) gains
Command palette results / Available actions matching the current filter. Placeholder text and visible behaviour are
unchanged. The placeholder is a typing hint for sighted users; the
accessible name is what screen readers announce on focus —
separating the two unblocks the keyboard-only workflow.
tests/features/a11y_chrome_names/ — eight invariants spanning
the contract: T1–T4 each TitleBar button has the right
accessibleName + non-empty accessibleDescription; T5/T6 the
CommandPalette input + list have theirs; T7 the AT-SPI
introspection lane (every reachable QAbstractButton carries
either a non-empty accessibleName() or a non-empty text(),
every reachable QLineEdit has an explicit accessibleName),
T8/T9 source-grep counts on titlebar.cpp and commandpalette.cpp
so a future refactor that drops the setAccessibleName calls
fails the build, not just AT-SPI usage at runtime. The test
constructs the TitleBar and CommandPalette directly under
QT_QPA_PLATFORM=offscreen so it runs in CI without a display.tr() translation hooks for the new strings are deferred to the
0.9.0 H10 i18n bundle so the .qm files cover both UI text and
accessibility strings in one pass — pulling them in now would mean
re-touching every setAccessibleName(...) call later.QAccessibleInterface adapter for TerminalWidget
itself (the text-changed event lane behind the H9 a11y bundle in
0.9.0) is a separate, larger design cycle and is out of scope
for this bundle.0.7.40 — Perf bundle — scroll-region rotate + VtBatch zero-copy
Theme: Performance — scroll-region rotate + VtBatch zero-copy
across the worker→GUI thread hop. Two Tier 3 perf items from
ROADMAP.md § 0.7.x ⚡ / 🏗 Tier 3 — structural retired in one
bundle. Both are observably equivalent at the API level — the
existing tests/features/scroll_region_rotate/ and
tests/features/threaded_parse_equivalence/ regressions pass
unchanged — so the win is throughput / cost, not visible behaviour.
TerminalGrid::scrollUp / scrollDown use std::rotate.
Replaced the per-iteration m_screenLines.erase + insert loop
with a single rotate over [scrollTop, scrollBottom], gated by a
pool-salvage / scrollback-push pre-pass and a fresh-blank
post-pass. The CSI 2J doubling-guard window check now runs once
per batch (elapsed time across a batch is microseconds, so the
per-iteration window-extend was redundant). m_scrollback cap
enforcement collapses from N pop_front passes to one. On CSI 100 S against an 80-row screen, the row-shift cost goes from
O(count × rows) memmoves (8000) to O(rows) (80).
VtStream::batchReady carries VtBatchPtr instead of
const VtBatch &. using VtBatchPtr = std::shared_ptr<const VtBatch>; is the new alias in vtstream.h. Qt's queued-
connection plumbing must own the parameters it dispatches, so
the prior const T & signal forced a deep copy of the entire
batch (the actions vector + rawBytes QByteArray) on every
worker→GUI hop, regardless of any move-from-pending shaping the
emitter performed. The shared_ptr wrap reduces the cross-thread
payload to a 16-byte atomic refcount bump; the underlying
VtBatch lives on the heap and is not duplicated. Both emit
sites (flushBatch, onPtyFinished) build via
std::make_shared<VtBatch>(). TerminalWidget::onVtBatch
signature changed to void onVtBatch(VtBatchPtr batch); field
access is now via batch->… instead of batch.….
tests/features/vtbatch_zero_copy/ — 5 source-grep invariants
locking the cross-thread signal shape so a future refactor
can't silently revert to the deep-copy form. Behavioural
equivalence on the action stream is already covered by
tests/features/threaded_parse_equivalence/.
tests/features/scroll_region_rotate/ — pre-existing 8
invariants on rotation correctness (I1–I8 in spec.md) still
pass. The spec was written algorithm-agnostic, anticipating this
swap.
0.7.39 — Claude state-dot palette + status-bar Roadmap viewer
Theme: Claude Code UX bundle + status-bar Roadmap viewer. Two user requests landed in quick succession.
User request 1 (Claude state-dot palette):
"Let's have a round dot on each tab that has a Claude Code session running (no icons or anything else other than the tab label). The dot will change colour with the various states that Claude Code is in. Each state has its own colour (grey for idle). Then extend those colours to the status bar Claude Code status too."
User request 2 (Roadmap viewer):
"Add a button on the status bar to view the roadmap. So, it brings up a dialog showing the roadmap. It should have filters as well to show what is outstanding and what is completed. If at all possible, it should also highlight what item is being done currently. The roadmap button should only show if there is roadmap documentation — let's simplify that to requiring a roadmap.md file only." Plus a clarification adding a fourth emoji toggle and elevating "Currently being tackled" to a peer filter.
Unified Claude state-dot palette across tabs and status bar.
New static helper ClaudeTabIndicator::color(Glyph) in
coloredtabbar.h is the single source of truth for an
eight-state palette: Idle grey #888888, Thinking blue
#5BA0E5, ToolUse yellow #E5C24A, Bash green #6FCF50,
Planning cyan #5DCFCF, Auditing magenta #C76DC7, Compacting
violet #A87FE0, AwaitingInput orange #F08A4B. Red is
intentionally absent — AwaitingInput is a normal interaction
state, not an error. ColoredTabBar::paintEvent calls the
helper for fill colour and uses a single kDotRadius = 4 for
every state; the prior AwaitingInput "outline + radius 5"
treatment is removed (per "no icons or anything else"). The
bottom status-bar Claude Code label was rewired to map current
state → Glyph → helper colour, replacing the prior
Theme::ansi[N] mappings that drifted from the tab dot's
colour and varied across themes.
Auditing now lights the per-tab dot. Previously surfaced
only on the active-tab status bar (m_claudeAuditing). Plumbed
into ClaudeTabTracker::ShellState::auditing (mirrored from
the existing ClaudeTranscriptSnapshot.auditing field). Tab
provider routes it to Glyph::Auditing magenta; hover tooltip
reads "Claude: auditing". Precedence chain unchanged across
both surfaces:
AwaitingInput → Planning → Auditing → state-derived.
Status-bar Roadmap viewer button (roadmapdialog.{h,cpp},
new files). Visible iff the active tab's shellCwd()
contains a ROADMAP.md (case-insensitive — accepts
Roadmap.md, roadmap.md too). MainWindow:: refreshRoadmapButton is called from the central
refreshStatusBarForActiveTab tick. Click opens
RoadmapDialog: a non-modal QDialog with a top row of five
peer category checkboxes (✅ Done · 📋 Planned · 🚧 In progress ·
💭 Considered · Currently being tackled), all default-checked.
A bullet renders iff any of its enabled category memberships
matches; plain narration bullets (no status emoji) always
render. The static RoadmapDialog::renderHtml helper is pure
so tests can drive it without spinning a Qt widget.
"Currently being tackled" highlight. A bullet matches the
current-work signal if its first-80-character payload (status-
emoji-stripped, normalised — lowercase, hyphens-and-underscores
as spaces, punctuation removed) contains a phrase from a signal
set built from (a) the local CHANGELOG.md [Unreleased]
block bullets, (b) the last 5 non-release/non-merge/non-revert
git commit subjects. Matched bullets get a yellow left-border
CSS highlight (border-left: 4px solid #E5C24A — the same
ToolUse yellow from the dot palette; the consistency is
intentional). When the signal set is empty, no highlight
renders — the feature is silent when there's nothing to point
at.
Live-tail dialog mechanics. QFileSystemWatcher on
ROADMAP.md and CHANGELOG.md, 200 ms debounce timer, and
the same scroll-preservation triple shipped with 0.7.37 /
0.7.38: m_lastHtml shared_ptr cache + skip-identical-HTML
guard, capture vbar before setHtml, restore with
qMin(saved, vbar->maximum()) clamp, was-at-bottom pin so
live appends stay visible.
ClaudeTabIndicator::Glyph enum gains Auditing. Inserted
between Planning and Compacting so the palette ordering
reads Done-side → Action-side → InFlight-side → BlockingUser.
ClaudeTabTracker::ShellState gains a bool auditing field
and a new comparison branch in maybeEmit so changes to the
audit latch trigger shellStateChanged like every other
field.
MainWindow::applyClaudeStatusLabel no longer reads
Theme::ansi[] for Claude-label colours. All state →
colour mappings now route through the unified helper. The
Theme &th = Themes::byName(...) local was removed (unused).
radius = 5 and
white outline). Colour alone is the differentiator now, per
the "no icons or anything else" user spec. The
outline.alpha() branch was removed from paintEvent.tests/features/claude_state_dot_palette/ — 8 source-grep
invariants asserting helper signature, full palette,
paintEvent helper-call, uniform geometry, mainwindow
status-bar wiring, Auditing plumbing through ShellState,
and double-use of the Auditing glyph (provider closure +
status applier).
tests/features/roadmap_viewer/ — 10 hybrid invariants
(source-grep + behavioral). Links the dialog source so the
static renderHtml helper can be driven against synthetic
markdown to verify five-bit filter semantics, the highlight
CSS marker on signal match, the marker's absence on empty
signal sets, the case-insensitive cwd probe, the wire-up
shape (m_roadmapBtn construction + click connection), and
the refreshRoadmapButton call from
refreshStatusBarForActiveTab.
/tmp/*.js TOCTOU and the
no-auth local-IPC chain — both shipped in the 0.7.12 Tier 1
batch but the umbrella narrative entries weren't updated at
the time).0.7.35 — About dialog OK button fix — QDialog + connected QDialogButtonBox
Theme: UX bug fix. Single user-reported regression in the Help → About Ants Terminal dialog (the GUI version indicator added in 0.7.22).
About Ants Terminal dialog OK button silently no-op'd
(mainwindow.cpp). Was: Help → About Ants Terminal… opened
the version dialog correctly, but clicking its OK button did
nothing — the dialog could only be dismissed via the
window-manager close (X) button. The 0.7.22 implementation
used QMessageBox::Ok with Qt::TextBrowserInteraction (which
pulls in Qt::TextSelectableByMouse); under the combination of
our frameless MainWindow, Qt::WA_TranslucentBackground, KDE
Plasma + KWin, and Qt 6.11 this caused the OK click to be
silently dropped. Now: the About handler builds a custom
QDialog with a QDialogButtonBox(QDialogButtonBox::Ok) whose
accepted signal is explicitly connected to QDialog::accept,
giving us a click path that's standard, testable, and
independent of QMessageBox's internal standard-button dispatch.
About dialog GitHub link was a visual-only no-op
(mainwindow.cpp). Was: the link in the About body was
rendered as clickable (cursor changed on hover) because
Qt::TextBrowserInteraction enables Qt::LinksAccessibleByMouse,
but setOpenExternalLinks(true) was never called on the label
— so clicking the link emitted linkActivated() into the void.
Now: the body QLabel has setOpenExternalLinks(true), so
clicking the GitHub URL opens it in the user's browser.
tests/features/help_about_menu/spec.md + test (8
invariants). Spec rewritten to lock the QDialog +
QDialogButtonBox shape, document the 2026-04-25 regression in
the History section, and assert via source-grep that the
QMessageBox::Ok + Qt::TextBrowserInteraction pattern
cannot reappear without the test failing. Pre-fix source
fails 6 of the 8 new I4 invariants; post-fix all 8 pass.0.7.38 — Background-tasks status-bar button — Claude Code live tail
Theme: Feature. Background-tasks status-bar surface for Claude Code. User request 2026-04-25:
"a button on the status bar when there are background tasks being run. We then click the button to view what Claude Code shows for the background tasks. The button opens a dialog showing the live update info on the background tasks."
Claude Code can spawn long-running tasks via Bash or Task with
run_in_background: true. The TUI shows a sidebar listing them with
their tail output; users dropping into a different tab to keep
working had no way to see those tasks from Ants Terminal — they had
to flip back to the TUI and parse the sidebar by hand. The new
status-bar button surfaces the live count and the dialog mirrors the
0.7.37 Review Changes update model (debounced live tail with
scroll-preservation).
Background-tasks tracker (new claudebgtasks.{h,cpp}). A pure
parser walks the active session's transcript JSONL, recognizing
tool_use blocks whose input.run_in_background == true (start)
and tool_result blocks whose toolUseResult.backgroundTaskId is
set (confirmation, with the on-disk output path extracted from
the result body). Completion / kill state is derived from
subsequent BashOutput results carrying status: "completed" | "killed" | "failed" and from KillShell tool calls. The
ClaudeBgTaskTracker class wraps the parser with a
QFileSystemWatcher on the transcript and emits tasksChanged()
when the running-count or shape changes. Static
parseTranscript(path) is testable without a watcher.
Status-bar button (mainwindow.{h,cpp}). Sibling to the
Review Changes button; same fixed sizePolicy contract. Hidden by
default; visible only while runningCount() > 0 for the active
tab's session. Label re-renders as "Background Tasks (N)";
tooltip discloses running + total. Re-targeted on tab switch via
refreshBgTasksButton() (called from
refreshStatusBarForActiveTab), so each tab's session drives its
own count independently.
Live-tail dialog (new claudebgtasksdialog.{h,cpp}). Mirrors
the Review Changes live-update model that 0.7.37 stabilized.
QFileSystemWatcher covers each task's .output file plus the
transcript path, so new starts and completions appear without
manual refresh. A 200 ms debounce timer collapses bursts of
fileChanged signals (common when the backgrounded process is
noisy) into one render. Skip-identical-HTML guard via a per-
dialog m_lastHtml (std::shared_ptr<QString>) preserves
selection and scroll position when the render is unchanged.
Capture-vbar/hbar before setHtml, restore after with qMin(..., maximum()) clamp keeps absolute scroll position when content
changes — same shape as the 0.7.37 Review Changes fix. Bonus
"was at bottom" detection: when the user is tailing the bottom,
pin them to the bottom across appends so live output stays
visible without manual scrolling.
tests/features/claude_bg_tasks_button/spec.md + test
(10 invariants). New regression test pinning the parser shape
(run_in_background and backgroundTaskId keys), the tracker's
QFileSystemWatcher + tasksChanged() signal, the button's
hide-when-empty contract, the connect to showBgTasksDialog,
the dialog's reuse of the 0.7.37 scroll-preservation pattern,
the outputPath watch enumeration in rewatch(), the debounce
timer interval (≤ 500 ms, single-shot), and the CMake source
list. Source-grep harness, no Qt link.0.7.37 — Review Changes dialog — scroll preserved across live refreshes
Theme: UX bug fix. Live-update regression in the Review Changes dialog (the QFileSystemWatcher + 300ms debounce wiring added in 0.7.32). User report 2026-04-25:
"When using the Review Changes dialog, the constant resetting of the text means that if I scroll, it resets to the beginning every refresh. That means I can't scroll basically."
The 0.7.32 finalize lambda called viewerGuard->setHtml(html)
unconditionally on every probe completion — every git change, every
debounce tic, every Refresh click. QTextEdit::setHtml re-parses
the document and snaps the vertical scroll bar back to the top, also
discarding selection and cursor position. On a long diff with active
live updates the dialog became unscrollable.
mainwindow.cpp). Two-layer guard around the
setHtml call inside MainWindow::showDiffViewer's finalize
lambda. (1) An auto lastHtml = std::make_shared<QString>() cache
threaded through runProbes and finalize lets finalize early-
return when the new render is byte-identical to the previous one
— the common case during idle live-update tics, where branch
metadata refreshes don't change anything visible. Selection,
cursor, and scroll are byte-perfectly preserved. (2) When content
does change, vertical and horizontal scroll-bar values are
captured before setHtml and restored after, clamped to
bar->maximum() so a shorter render after a commit doesn't over-
scroll. First-render carve-out (isFirstRender = lastHtml->isEmpty()) keeps the initial paint at the top.tests/features/review_changes_scroll_preserve/spec.md + test
(6 invariants). New regression test pinning the lastHtml
cache, the skip-identical guard, the
capture-scroll-before-setHtml + restore-after pattern, and the
first-render carve-out. Source-grep harness scoped to the
MainWindow::showDiffViewer body so other QScrollBar / setHtml
call-sites in mainwindow.cpp don't cause false positives.0.7.36 — Tab bar + status bar opaque under translucent parent
Theme: UX bug fix. Same translucent-parent failure mode that bit the menubar in 0.7.25/0.7.26 caught in two more places (user report, post-0.7.32):
"The tabs themselves are fine but the rest of the tab bar across the window is transparent as well as the background for the status bar."
The 0.7.32 close-button SVG drew the user's eye to the empty-area fill that had been mis-painted all along; the bug was not introduced by 0.7.32 but became visible because the new tab look re-balanced what the user noticed.
Tab bar empty area paints opaque under
WA_TranslucentBackground (coloredtabbar.{h,cpp}). Was: the
strip to the right of the last tab rendered the desktop wallpaper
through, with the QSS QTabBar { background-color: ... } rule
silently dropped by Qt's stylesheet engine on KWin + Breeze + Qt 6
once WA_OpaquePaintEvent was set on the widget. Same root cause
documented in opaquemenubar.h. Now: ColoredTabBar exposes
setBackgroundFill(QColor) and prepends a
CompositionMode_Source fillRect to its existing paintEvent,
before the base class draws tabs and before the colour-group
gradient overlay. applyTheme() pushes theme.bgSecondary into
the override.
Status bar background paints opaque under
WA_TranslucentBackground (new opaquestatusbar.h,
mainwindow.{h,cpp}). Was: the entire status bar strip rendered
the desktop wallpaper through — same root cause as the tab bar
above. Now: a header-only OpaqueStatusBar mirrors
OpaqueMenuBar (paintEvent fillRect → delegate). MainWindow
constructs an OpaqueStatusBar, installs it via setStatusBar(...)
before the first statusBar() call (Qt's lazy-creation would
otherwise install a plain QStatusBar that paints transparent),
and pushes theme.bgSecondary via setBackgroundFill from
applyTheme().
tests/features/tabbar_statusbar_opaque/spec.md + test (5
invariants). New regression test pinning the setBackgroundFillfillRect mechanism on both bars, the setStatusBar install
order, the WA_OpaquePaintEvent attribute on the tab bar, and the
paint order inside ColoredTabBar::paintEvent (fill before the
base class draws tabs). Source-grep harness, no display required.Theme: Terminal correctness. One ROADMAP item retiring the last
known DECOM (origin mode) bug — tmux/screen save-restore
round-trips were silently dropping the origin-mode flag and
ignoring scroll-region-relative coordinates on absolute cursor
moves.
CUP / HVP / VPA translate to scroll-region origin under DECOM
(terminalgrid.cpp). Was: with origin mode (DECOM, CSI ?6h)
active and a scroll region set via DECSTBM (CSI top;bottom r),
CSI 1;1H jumped to absolute (0, 0) and CSI 99;1H clamped to
the bottom of the entire screen instead of the bottom of the
scroll region. Programs that depend on origin-mode-relative
positioning (vim split-window cursor restoration, screen's
caption line, tmux status redraws) ended up writing to the
wrong row. Now: the 'H', 'f', and 'd' cases inside
processCSI add m_scrollTop to the requested row when
m_originMode is set, then clamp to [m_scrollTop, m_scrollBottom].
Behaviour matches xterm.
DECSC saves DECOM + DECAWM, DECRC restores them
(terminalgrid.cpp, terminalgrid.h). Was: saveCursor()
stored only (row, col, attrs), so a TUI that flipped origin
mode or auto-wrap, did CSI s / ESC 7, ran code that
re-toggled the flag, then CSI u / ESC 8'd ended up with the
flag in whatever state the inner code left it in — silent
coordinate-space corruption. Now: m_savedOriginMode and
m_savedAutoWrap members capture the flags on save and restore
them on restore, matching the VT420 spec for DECSC / DECRC.
DECSTBM home position respects DECOM (terminalgrid.cpp).
Was: setScrollRegion() always homed the cursor to absolute
(0, 0) after setting the region. Now: when m_originMode is
on at the time DECSTBM lands, the cursor moves to
(m_scrollTop, 0) — the top-left of the origin-mode
coordinate space — as xterm does.
Locked by tests/features/origin_mode_correctness/ — 12
behavioural invariants (CUP/HVP/VPA translation, scroll-region
clamping, DECSTBM home, DECSC/DECRC save-restore in both CSI s/u
and ESC 7/8 forms, DECAWM round-trip via the wrap-or-not at
last-column probe) plus 4 source-grep checks anchoring the fix's
shape (std::clamp(row, m_scrollTop, m_scrollBottom) near case 'H', m_savedOriginMode = m_originMode in saveCursor, the
mirror line in restoreCursor, and the m_originMode ? m_scrollTop : 0 ternary in setScrollRegion).
Theme: Lifecycle / cleanup. Three ROADMAP items addressing GUI-thread blocking on shutdown, an xdg-desktop-portal session leak, and two latent issues (OOM surface + symlink escape) in the Lua plugin loader.
PTY destructor escalation runs off the main thread
(Pty::~Pty). Was: SIGTERM-then-busy-wait-then-SIGKILL ran on
the GUI thread; N split panes closing together blocked the
window close N × 500 ms. KWin would throw a "window not
responding" hint at four splits. Now: the destructor still
does the cheap pre-escalation reap (SIGHUP + close master fd +
waitpid(WNOHANG)) inline — most shells exit on SIGHUP in
microseconds. If the cheap reap doesn't take, the
SIGTERM/SIGKILL escalation moves to a detached std::thread
capturing the pid by value (no this reference can outlive
the destructor). The thread creation is wrapped in a
try { ... }.detach() } catch (const std::system_error &)
that falls back to the synchronous escalation when thread
creation fails (rare — ulimit -u pressure at exit). Locked
by tests/features/pty_dtor_off_main_thread/ — 11 invariants
on <thread> include, lambda capture list (rejects [this]
and [&]), .detach() call, fallback retention, and the
pre-escalation cheap reap remaining inline.
GlobalShortcutsPortal closes its session on destruction
(GlobalShortcutsPortal::~GlobalShortcutsPortal). Was: no
destructor at all — the session handle returned by
CreateSession leaked for the lifetime of the D-Bus client.
xdg-desktop-portal accumulated one orphan session per Ants
invocation that crashed or was SIGKILLed before the QObject
parent-tree cleanup could implicitly close the bus connection;
visible via busctl --user introspect org.freedesktop.portal.Desktop .... Now: the destructor
issues an asynchronous
org.freedesktop.portal.Session.Close call against
m_sessionHandle when non-empty (early-returns on empty
handle so X11 / GNOME / no-portal paths don't crash Qt's
D-Bus marshaller). New kSessionIface constant in the
anonymous namespace alongside the existing
service/path/interface constants. Locked by
tests/features/portal_session_close/ — 8 invariants on
header dtor declaration with override, kSessionIface
constant, empty-handle early return, and the
asyncCall(createMethodCall(..., kSessionIface, "Close"))
dispatch.
Plugin manifest cap + canonical plugin path
(PluginManager::scanAndLoad). Two latent issues addressed
together. Manifest cap: f.readAll() was unbounded — a
multi-GB manifest.json (corrupt disk, malicious tarball)
would allocate that much RAM before QJsonDocument::fromJson
got a chance to reject it. Now: f.read(kMaxManifestBytes)
with kMaxManifestBytes = 1024 * 1024 (1 MiB ≈ 250 plugins
worth of permission/description text — real manifests are
<10 KiB, so the cap never bites legitimate content). Files
larger than the cap log a warning and skip without parsing.
**Canonical plugin path:** the scan now anchors on
QFileInfo(m_pluginDir).canonicalFilePath(), passes
QDir::NoSymLinks to entryList (cheap first-pass filter),
and per-entry verifies the resolved path is anchored inside
the canonical root via
startsWith(canonicalRoot + "/"). Closes the symlink-escape
shape where a hostile plugin tarball containing
evil -> /etc/cron.daily could trick the loader into
attempting init.lua from outside the user's plugin tree.
Locked by tests/features/plugin_manifest_safety/ — 12
invariants on the cap (named constant + value + bounded read
Theme: Dialog UX. Three bundle items from the Settings dialog ROADMAP list (dependency-UI gating, Cancel rollback for Profiles, Restore Defaults per-tab) plus one user feedback item from 2026-04-25 — the Review Changes dialog only surfaced current- branch state and missed work living on other branches.
Restore Defaults button per primary Settings tab
(SettingsDialog::setupGeneralTab, setupAppearanceTab,
setupTerminalTab, setupAiTab). Was: only Keybindings had
a defaults-reset button. A user who tweaked the dialog and
wanted to start over had to either remember every default or
delete ~/.config/ants-terminal/config.json (which lost
unrelated settings: highlight rules, profiles, plugin grants).
Now: each substantive tab has its own button with stable
objectName (restoreDefaultsGeneral, restoreDefaultsAppearance,
restoreDefaultsTerminal, restoreDefaultsAi). Reset slots
mutate widget state only — applySettings still commits, Cancel
still rolls back. Schema defaults match the second argument of
each Config::xxx() getter so the dialog and config layer can't
drift. Locked by tests/features/settings_restore_defaults/
— 22 invariants on objectNames, reset-value coverage per tab,
and the "no m_config-> writes from within reset slots" rule.
Review Changes dialog: live updates via QFileSystemWatcher +
manual Refresh button (MainWindow::showDiffViewer). User
feedback 2026-04-25: dialog should show live or near-live
updates. Was: probes ran once on dialog open and never again —
the user had to close and re-open to see changes from a commit
/ fetch / branch operation done in the terminal underneath. Now:
a QFileSystemWatcher watches cwd, .git, .git/HEAD,
.git/index, .git/refs/heads, .git/refs/remotes,
.git/logs/HEAD. fileChanged and directoryChanged signals
feed a 300 ms single-shot QTimer (debounce — git pull /
git fetch fire fileChanged O(refs) times in milliseconds; the
debounce coalesces them into one re-probe). The probe-spawning
logic was refactored into a runProbes lambda that constructs a
fresh ProbeState per call, so an in-flight probe whose finalize
outlives the next refresh can't decrement the new pending counter
and render a half-populated mix. Atomic-rename safe: the
fs-event handler re-adds paths that exist but are no longer
watched (Qt loses the watch on rename(2), which git uses for
HEAD/index/logs/HEAD updates). A live-status label
(reviewLiveStatus objectName) shows "● refreshing…" /
"● live — auto-refresh on git changes" so the user can confirm
the watcher is wired. A manual Refresh button
(reviewRefreshBtn objectName) bypasses the debounce for cases
where state changed outside the watched paths (a build script
in another shell, a different terminal). Locked by
tests/features/review_changes_branches/ (extended) — now 33
invariants total covering ProbeState fields, runAsync targets,
finalizer rendering, copy-handler payload, empty-state guard,
runProbes lambda, watcher armament, debounce timing, atomic-
rename re-watch, refresh-button bypass, and live-status label
states.
Review Changes dialog: per-branch summary + cross-branch
unpushed commits (MainWindow::showDiffViewer,
MainWindow::ProbeState). Was: three async git probes
(status, diff HEAD, log @{u}..HEAD) all scoped to the
current branch's working tree and HEAD lineage. A user with
five feature branches each holding unpushed work saw "no
unpushed" if they happened to be on a clean branch. Branches
without upstreams or with diverged ahead/behind state were
invisible. User feedback 2026-04-25: "the Review Changes
dialog doesn't consider changes in other branches of the
project." Now: two additional probes drop in alongside the
existing three —
git for-each-ref refs/heads --format='%(refname:short) \t%(upstream:short)\t%(upstream:track)\t%(subject)' for the
per-branch summary (with ahead/behind/gone/no-upstream colour
cues), and git log --branches --not --remotes --oneline --decorate for the cross-branch unpushed log (every commit
reachable from any local branch but not from any remote-
tracking branch). Both are O(refs) and finish in milliseconds.
Copy Diff includes both new sections. Locked by
tests/features/review_changes_branches/ — 15 invariants on
ProbeState fields, runAsync targets, finalizer rendering,
copy-handler payload, and empty-state guards.
Tab close button (×) is always visible, not hover-only
(MainWindow::applyTheme stylesheet —
QTabBar::close-button). User feedback 2026-04-25: "the
tabs still don't have a visible marker per tab that shows
where to click to close the tab. The mouseover works but we
need to also see it when onmouseout." The 0.6.27 fix removed
image: none to let Qt fall back to the platform's standard
close icon — that worked on Breeze/Adwaita but failed on
Fusion / qt6ct / certain Plasma colour schemes where the
platform style still rendered the × hover-only. Now: explicit
data-URI SVG image: url("data:image/svg+xml;...") rules in
both the default and :hover QTabBar::close-button variants.
Glyph re-tints with the active theme via textSecondary
(default) / textPrimary (hover); hover keeps the ansi-red
background-color will-click cue. URL-encoded %23 is
spliced into the arg list via QStringLiteral("%23") + theme.<color>.name().mid(1) rather than the format string —
prevents Qt's CSS parser from truncating the data URI at the
fragment delimiter and also avoids the QString::arg()
placeholder-numbering collision. Locked by
tests/features/tab_close_button_visible/ — 11 invariants on
data-URI presence, two-line × shape, arg-side splice, and
image-rule presence in BOTH state variants.
Dependency-UI enable gating (SettingsDialog::setupAppearanceTab,
setupTerminalTab, setupAiTab). Was: master checkboxes
(m_aiEnabled, m_autoColorScheme, m_quakeMode) gated logic
but not UI — typing an API key into a feature-disabled AI tab,
or selecting "Solarized" as the light-mode theme while
auto-switch was off, both produced silent no-ops. Now:
QCheckBox::toggled is wired to setEnabled on every
dependent sibling, with a one-shot sync call at construction
so the initial state matches the loaded config without
relying on setChecked() always emitting toggled (it only
emits when state actually changes). Disabled controls keep
their current values, so toggling the master back on restores
the user's prior selection rather than zeroing it out. Locked
by tests/features/settings_dependency_gating/ — 16
invariants on the three sync lambdas, dependent setEnabled
calls, toggled-connect wiring, and initial-sync call sites.
Profiles tab honors Cancel/OK semantics
(SettingsDialog::setupProfilesTab, loadSettings,
applySettings, m_pendingProfiles, m_pendingActiveProfile).
Was: profile Save/Delete/Load buttons mutated m_config
immediately via setProfiles() / setActiveProfile(). Cancel
could not roll those edits back — they had already been
persisted to config.json before the user's intent was known.
Every other Settings tab follows the standard "stage in
widgets, commit on applySettings, discard on reject" pattern;
Profiles broke that contract. Now: the three buttons mutate a
pending-state pair (m_pendingProfiles + m_pendingActiveProfile),
loadSettings re-initializes the pair from m_config, and
applySettings is the single commit point that calls
setProfiles / setActiveProfile. Cancel skips applySettings,
so dialog close leaves m_config unchanged. Locked by
tests/features/settings_profile_cancel_rollback/ — 11
invariants including a global "exactly one m_config->setProfiles
call site" check that catches a regression where a button
callback starts writing to m_config directly again.
Theme: Persistence integrity (cross-file). Four items from the
post-0.7.30 grouping plan, addressing the silent-data-loss /
permission-drift / concurrent-writer surfaces that span Config,
ClaudeAllowlist, SessionManager, and SettingsDialog. Splits the
growing secureio.h into secureio.h (perms) + configbackup.h
(rotation + cooperative write lock) before a third helper landed.
ConfigWriteLock cooperative write lock for shared config files
(src/configbackup.h). Was: two simultaneously-running Ants
instances saving the same ~/.config/ants-terminal/config.json
raced over <path>.tmp + rename(2) — both writers truncated the
shared tmp, partial bytes interleaved on the same inode, and
last-rename-wins silently dropped one process's keystrokes /
settings / profile / AI key. Same shape for ~/.claude/settings.json
where Allowlist + Install-hooks + git-context-installer all
read-modify-write with no serialization. Now: an RAII guard wraps
POSIX flock(2) on a sibling <path>.lock file with a 5-second
poll deadline (100 × 50 ms), advisory so cooperating callers
serialize while non-cooperating editors (vim, jq) bypass by
design. Config::save, ClaudeAllowlistDialog::saveSettings,
SettingsDialog::installClaudeHooks, and
SettingsDialog::installClaudeGitContextHook now construct
ConfigWriteLock writeLock(path) with an acquired() guard
before the write block; failure to acquire logs and returns
rather than risking the data race. Locked by
tests/features/concurrent_writer_lock/ (5 runtime invariants
via fork(2) so flock semantics are honestly tested across
processes, plus 7 source-grep invariants on each save site).
Pre-fix source fails 7; post-fix all 12 pass.
Belt-and-suspenders post-rename setOwnerOnlyPerms at every
persistence site (Config::save, SessionManager::saveSession,
SessionManager::saveTabOrder, SettingsDialog::installClaudeHooks,
SettingsDialog::installClaudeGitContextHook). Was: each site
set 0600 on the temp fd before write but relied on rename(2)
preserving perms across the swap. ext4/xfs/btrfs honor that, but
FAT/exFAT on removable media (no POSIX bits at all), some
SMB/NFS servers (server-side rename applies server umask), and
Qt's copy+unlink fallback (cross-device rename, exotic mounts —
copy creates the destination with the process umask) do not.
Files containing ai_api_key (config.json), Claude Code bearer
tokens (settings.json), or paste-buffer scrollback content
(session_*.dat) could land 0644 on those filesystems and leak to
every UID on the host. Now: each site re-chmods the final
inode after rename/commit() returns success — idempotent on
POSIX filesystems, essential elsewhere. Locked by
tests/features/persistence_post_rename_chmod/ (10 source-grep
invariants spanning config.cpp / sessionmanager.cpp /
settingsdialog.cpp). Pre-fix source fails 6; post-fix all 10
pass.
secureio.h / configbackup.h split. secureio.h is now
perms-only — setOwnerOnlyPerms(QFileDevice&) and
setOwnerOnlyPerms(const QString&), the original 0600-bitmask
helpers that ~12 callers reach for. configbackup.h is the new
home for rotateCorruptFileAside (silent-data-loss recovery,
added 0.7.12) and the new ConfigWriteLock (concurrent-writer
guard, added this release). The split preempts a third-helper
cliff: the file was straddling perms + recovery + lock concerns
and a fourth would have made the include cost-of-business across
the codebase. Each downstream caller now picks the include it
actually needs (config.cpp, claudeallowlist.cpp,
settingsdialog.cpp pull both; sites that only set perms keep
their secureio.h include). Locked by
tests/features/secureio_configbackup_split/ (13 invariants:
file-content boundaries, non-copyable lock, every caller's
include set). Pre-fix source fails 4; post-fix all 13 pass.
Config::load, ClaudeAllowlistDialog::saveSettings,
and SettingsDialog::install{ClaudeHooks,ClaudeGitContextHook} —
but only Config had a regression test
(tests/features/config_parse_failure_guard/). The other two
were locked only by adjacent grep-style asserts inside other
feature tests. Now they have their own dedicated test:
tests/features/settings_parse_failure_mirror/ (8 invariants:
rotation call site, return-false-after-rotation gating,
open-failure branch distinct from parse-failure branch, comment
anchors explaining the clobber-risk reasoning). Closes the
ROADMAP "Allowlist + settings-dialog feature-test analogs" item.
Pre-fix source already passes — these tests lock previously
shipped behaviour against future regressions, not catch a new
bug.Theme: Session-file integrity. Three ROADMAP § 0.7.12 Tier 2 items
shipped together from the post-0.7.27 grouping plan, all in
sessionmanager.cpp: a SHA-256 envelope around the qCompress payload,
a pre-flight on qCompress's 4-byte uncompressed-length prefix, and
per-cell QDataStream::status() checks inside the decode loops.
V4 SHA-256 envelope around session-file payload
(SessionManager::serialize, SessionManager::restore,
SessionManager::ENVELOPE_MAGIC,
SessionManager::ENVELOPE_VERSION). Was: session files at
$XDG_DATA_HOME/ants-terminal/sessions/session_<tabId>.dat were
raw qCompress output with no payload integrity. Anyone with
write access to that directory (the user's own UID, a compromised
local process, a runaway pip install post-exec, an npm
dependency) could plant a crafted session that fed arbitrary
codepoints, fg/bg colors, and attribute flags into the grid on
next restore — a render surface, not a sandbox. Now: serialize
wraps the qCompress output in a V4 envelope
[SHEC magic (0x53484543)][envelope version=1][SHA-256 of payload (32 bytes)][payload length (uint32)][compressed payload]; restore
peeks the magic, verifies the hash, and refuses to restore on
version mismatch, length disagreement, or hash mismatch. Inner
QDataStream format unchanged (still V3) — the integrity layer is
framing-only. Legacy V1-V3 files (no envelope) continue to load
via the magic-peek fall-through; their next save writes them out
as V4 organically. Regression test
tests/features/session_sha256_checksum locks four invariants —
serialize emits the envelope (INV-1), restore peeks the magic and
verifies the hash with a return-false on mismatch (INV-2),
ENVELOPE_MAGIC/ENVELOPE_VERSION declared on SessionManager with
the agreed-upon magic value (INV-3), envelope version remains 1
at this milestone (INV-4). Pre-fix source fails eight invariant
assertions; post-fix all four pass. ROADMAP § 0.7.12 Tier 2 entry
retired.
qUncompress length-prefix pre-flight
(SessionManager::restore, SessionManager::MAX_UNCOMPRESSED).
Was: a crafted file claiming 500 MB uncompressed in qCompress's
4-byte big-endian prefix triggered a 500 MB allocation that the
post-hoc raw.size() > 500MB cap could only catch after the
damage was done — and the on-disk payload could be tiny, so the
allocation pressure showed up with no concomitant disk-space
anomaly. Now: restore reads the first 4 bytes of the compressed
payload, reconstructs the big-endian uint32, and rejects any
claim above MAX_UNCOMPRESSED (500 MB) BEFORE qUncompress runs —
constant-time, no allocator pressure. Short-payload guard
(compressed.size() < 4) keeps the same path safe against
truncated inputs that can't carry a length prefix at all. The
post-decompression cap remains as a defense-in-depth backstop
against payloads that under-claim and over-deliver. Regression
test tests/features/session_qcompress_length_guard locks four
invariants — pre-flight reconstruction precedes qUncompress
(INV-1, INV-3), MAX_UNCOMPRESSED is the named constant (INV-2),
short-payload guard exists (INV-4). Pre-fix source fails four
invariant assertions; post-fix all four pass. ROADMAP § 0.7.12
Tier 2 entry retired.
QDataStream::status() checks inside cell-decode loops
(SessionManager::restore readCell / readCombining). Was:
the readCell lambda was void; a stream truncated mid-cell still
flowed default-constructed QRgb and uint8_t values through
QColor::fromRgba etc. into the cell, silently writing
uninitialized fg/bg/flags into the grid. The surrounding loops
didn't check per-iteration status either, so the grid kept
accepting cells from a stream already at ReadPastEnd. Now:
readCell returns bool and short-circuits on in.status() != QDataStream::Ok; every call site (scrollback cells, screen cells
in range, screen cells skipped on width shrink, screen cells
skipped on height shrink) is guarded by if (!readCell(...)) return false. The combining-character helper checks status after
each codepoint read, so a stream truncated mid-codepoint can't
push default-constructed 0 into the combining map either. Pre-
fix, a partial save (kernel crash mid-fsync, disk-full mid-write,
hostile sender truncating the envelope payload) could materialize
as garbage cells in the next restore — not a crash, but a
corrupted scrollback. Regression test
tests/features/session_cell_loop_stream_status locks three
invariants — readCell returns bool with status check (INV-1),
every call site uses if (!readCell(...)) and at least four
sites exist (INV-2), readCombining inner codepoint loop checks
status (INV-3). Pre-fix source fails three invariant assertions;
post-fix all three pass. ROADMAP § 0.7.12 Tier 2 entry retired.Theme: Audit pipeline II — output quality. Three ROADMAP § 0.7.12
items shipped together from the post-0.7.27 grouping plan, all in
auditdialog.cpp: SARIF v2.1.0 result.suppressions[] array, a
regex-DoS watchdog on user-supplied patterns, and a widened 96-bit
dedup key.
SARIF v2.1.0 result.suppressions[] array
(AuditDialog::exportSarif, AuditDialog::loadSuppressions,
AuditDialog::saveSuppression, m_suppressionReasons map,
Finding::suppressed). Was: SARIF export silently dropped
suppressed findings (those whose dedup key matched
~/.audit_suppress), producing a falsely-clean report that
defeated the suppression-trend telemetry already computed for
the in-app dashboard. Now: a parallel
QHash<QString, QString> m_suppressionReasons map populated
alongside m_suppressedKeys carries the user's free-text
reason from the JSONL file into memory; the parse pipeline
marks suppressed findings with Finding::suppressed = true
instead of dropping them; render paths (UI, HTML, plain-text)
continue to filter via isSuppressed; the SARIF export iterates
ALL findings and attaches a suppressions[] block (kind:
external, state: accepted, justification: the user's reason)
per SARIF v2.1.0 §3.34. GitHub Code Scanning / SonarQube /
VSCode SARIF Viewer now see the full audit picture and can
compute fires-vs-suppressions ratios across export boundaries.
Regression test tests/features/audit_sarif_suppressions locks
five invariants — map declaration (INV-1), loadSuppressions
populates + clears (INV-2), saveSuppression mirrors (INV-3),
exportSarif emits the suppressions JSON property with
external/accepted fields (INV-4), exportSarif no longer
drops on isSuppressed (INV-5). Pre-fix source fails seven
invariant assertions; post-fix all five pass. ROADMAP § 0.7.12
Tier 2 entry retired.
Regex-DoS watchdog on user-supplied audit patterns
(AuditDialog::isCatastrophicRegex,
AuditDialog::hardenUserRegex, applyFilter,
loadAllowlist). Was: user patterns from
audit_rules.json (OutputFilter::dropIfMatches) and
.audit_allowlist.json (AllowlistEntry::lineRegex) flowed
straight into QRegularExpression with no shape check and no
match-time bound. A pathological pattern committed in either
file could pin the GUI thread for seconds with classic
catastrophic backtracking on adversarial scanner output —
(.+)+, (\w*)*, (.*)+ against long lines. Now: a static
isCatastrophicRegex heuristic rejects nested-quantifier
shapes (a quantified group whose body itself contains a
quantifier) at compile time with a qWarning naming the
offending pattern; the rule continues to run without the
filter rather than refusing to load. Patterns that pass the
shape check are wrapped in PCRE2's (*LIMIT_MATCH=100000)
inline option via hardenUserRegex so even catastrophic
shapes that slip past the heuristic have a bounded match-step
budget — PCRE2 returns "no match" on overrun (fail-safe). 100
k steps handles every sane pattern (typical match completes
in < 1 k) and aborts adversarial patterns within
milliseconds. Regression test
tests/features/audit_regex_dos_watchdog locks four invariants
— the helper exists and is invoked at both user-pattern entry
points (INV-1), it recognizes nested-quantifier shapes
(INV-2), the LIMIT_MATCH=N cap is in the [1k, 1M] sane
range (INV-3), loadAllowlist emits a qWarning on rejection
(INV-4). Pre-fix source fails four invariants; post-fix all
four pass. ROADMAP § 0.7.12 Tier 2 entry retired.
computeDedup from 64 to 96 bits
(computeDedup, AuditDialog::isSuppressed,
Finding::dedupKey). Was: SHA-256 truncated to 16 hex chars
(64 bits) — same key serves as the SARIF partialFingerprint,
the suppression-JSONL key, the rule-quality bucket, and the
in-app "Suppress" anchor URL. The 64-bit collision threshold
(~2³² entries at 50 % collision probability) was comfortable
but tight given the multi-role usage; a single false-collision
cost a wrong-finding suppression. Now: .left(24) (96 bits)
raises the birthday threshold to ~2⁴⁸ for 8 extra bytes per
stored key — well past any plausible project's lifetime
collection. A new bool AuditDialog::isSuppressed(const QString &dedupKey) const helper encapsulates a backward-compat lookup:
match either the full 24-char key OR the leading 16-char prefix,
so existing pre-0.7.29 user ~/.audit_suppress entries
continue to suppress new 24-char findings without forcing a
migration. Six render-pipeline call sites that previously read
m_suppressedKeys.contains(f.dedupKey) now route through the
helper. Regression test tests/features/audit_dedup_96bit
locks four invariants — width ≥ 24 hex chars (INV-1),
isSuppressed exists with .left(16) legacy path (INV-2),
zero raw m_suppressedKeys.contains(f.dedupKey) sites
remain (INV-3), saveSuppression mirrors into
m_suppressionReasons (INV-4). Pre-fix source fails five
invariant assertions; post-fix all four pass.Theme: Audit pipeline I — process-side robustness. Three ROADMAP § 0.7.12 Tier 2 items shipped together as bundle 0.7.28 from the post-0.7.27 grouping plan: per-tool timeout overrides, incremental QProcess output drain with a 64 MiB cap, and a distinct tool-crash warning that no longer hides as "0 findings."
AuditCheck::timeoutMs field — per-check QProcess timeout
(AuditCheck, AuditDialog::runNextCheck, ctor lambda). Was:
m_timeout->start(30000) hard-coded inside runNextCheck, with
the timeout-handler lambda printing a literal "Timed out (30s)"
warning. Slow tools (cppcheck on a 500k-line tree, semgrep with
rule-pack compile, osv-scanner rate-limited by OSV.dev,
trufflehog over the full git history, clang-tidy / clazy on
Qt-heavy code) routinely exceeded the 30 s cap and got demoted
to tool-health warnings instead of producing findings. Now: a new
int timeoutMs = 30000; trailing field on the AuditCheck
aggregate (default preserves the pre-fix global so positional
call sites stay correct), m_timeout->start(check.timeoutMs) at
the use site, and the timeout-warning string formatted from the
actual cap. populateChecks ends with a calibration loop that
bumps known-slow tool IDs to 60 s (cppcheck, cppcheck_unused,
clang_tidy, clazy), 90 s (semgrep), or 120 s (osv_scanner,
trufflehog). Regression test tests/features/audit_per_tool_timeout
locks four invariants — the field declaration (INV-1), the
per-check use site with no hard-coded 30000 (INV-2), the
parameterised warning message (INV-3), at least one calibration
override above 30 s (INV-4). Pre-fix source fails all four;
post-fix all four pass. ROADMAP § 0.7.12 Tier 2 entry retired.
Incremental QProcess output drain with 64 MiB cap
(AuditDialog::onCheckOutputReady, AuditDialog::onCheckErrorReady,
AuditDialog::onCheckFinished). Was: m_process->readAllStandardOutput()
called exactly once inside onCheckFinished. Until that moment,
QProcess accumulated internal buffers without bound — a runaway
semgrep against generated code or a buggy plugin emitting a
tight printf loop could buffer hundreds of megabytes before the
timeout fired, with the audit dialog showing a frozen progress
bar the whole time. Now: the constructor wires
readyReadStandardOutput and readyReadStandardError to new
onCheckOutputReady / onCheckErrorReady slots that append
incrementally to m_currentOutput / m_currentError; if the
combined size exceeds MAX_TOOL_OUTPUT_BYTES = 64 * 1024 * 1024,
the process is killed and m_outputOverflowed is flagged. On
finished() the runner drains any tail data (Qt may buffer
between the last readyRead and finished()), reads from the
member buffers instead of the live process, and surfaces a
distinct "Output exceeded N MiB cap" warning when overflow
occurred. Buffers reset before each check so output never
concatenates across checks. A small connectProcessSignals()
helper centralises the three connections so the timeout-kill /
cancel-kill / reconnect cycles never lose a drain slot.
Regression test tests/features/audit_incremental_output_drain
locks six invariants — both readyRead* connections (INV-1),
buffer members declared (INV-2), MAX_TOOL_OUTPUT_BYTES cap in
the 4 MiB ≤ cap < 1 GiB defensible range (INV-3), per-check
reset (INV-4), onCheckFinished reads buffers not the live
process (INV-5), overflow path kills the process (INV-6).
Pre-fix source fails on the connect-grep alone; post-fix all six
pass. ROADMAP § 0.7.12 Tier 2 entry retired.
AuditDialog::onCheckFinished). Was: signature
void AuditDialog::onCheckFinished(int /*exitCode*/, QProcess::ExitStatus /*status*/) — both parameters discarded
via comments. A tool that segfaulted with empty stdout was
indistinguishable from a clean run with zero findings, silently
hiding both a real bug in the tool AND any findings the tool
would have reported. Same went for non-zero-exit-with-stderr-only
patterns (clang-tidy missing compile_commands.json, semgrep
failing to parse a rule). Now: parameters named (exitCode,
exitStatus); function branches on QProcess::CrashExit to
emit a "Tool crashed (signal exit)" warning; on exitCode != 0 && stdout empty && stderr non-empty to emit a "Tool exited N
with no findings on stdout" warning. Both warnings demoted to
Severity::Info so they don't sort to the top of the report
next to real findings. The four exit modes (timeout, overflow,
crash, non-zero-with-stderr-only) all funnel through a single
small file-scope makeToolHealthWarning() helper that
centralises the row shape (Info severity, warning flag, distinct
message prefix). Regression test
tests/features/audit_tool_crash_distinct locks four invariants
— named parameters (INV-1), CrashExit branch (INV-2), warning
emission with crash/exit-message (INV-3), Severity::Info
demotion (INV-4). Pre-fix source fails all four; post-fix all
four pass. ROADMAP § 0.7.12 Tier 2 entry retired.Theme: PTY-handler hardening sweep. Two ROADMAP § 0.7.12 Tier 2
items shipped together — child-side FD closure no longer relies on
a hard-coded fd<1024 cap, and master-side writes no longer drop
data on EAGAIN.
close_range(2) with RLIMIT_NOFILE-bounded
fallback (Pty::start). Was: for (int fd = 3; fd < 1024; ++fd) ::close(fd) — silently leaked any inherited FD with index
≥ 1024 into the user's shell on systemd-service / container /
hardened-server profiles where rlim_cur sits above the
hard-coded ceiling. Qt's display socket, D-Bus session
connection, plugin HTTP sockets, Lua VM eventfds, and the
remote-control IPC socket are all valid leak candidates; a
leaked AI HTTP socket is a credentials-exfiltration vector,
a leaked D-Bus socket lets the shell impersonate the user's
desktop session, a leaked remote-control socket is a UID-scope
RCE-by-proxy vector. Now: post-fork child branch issues
::syscall(SYS_close_range, 3, ~0U, 0) first — single
signal-safe syscall on Linux 5.9+, atomic over the whole
range, ignores the soft cap. If the syscall returns non-zero
(kernel < 5.9, missing build-time SYS_close_range), the
fallback path consults getrlimit(RLIMIT_NOFILE) and iterates
up to the runtime soft cap, capped at 65536 to bound the
worst-case syscall storm on profiles where rlim_cur is in
the hundreds of thousands. Two new headers
(<sys/resource.h>, <sys/syscall.h>) added to
ptyhandler.cpp. Regression test
tests/features/pty_closefrom locks five invariants —
SYS_close_range referenced inside the post-fork child
branch (INV-1), the hard-coded fd<1024 loop is gone (INV-2),
fallback consults RLIMIT_NOFILE (INV-3), required headers
included (INV-4), fallback bound is capped to 65536 to avoid
the unbounded-loop pathology (INV-5). Pre-fix source fails 4
of 5; post-fix all 5 pass. CWE-403 reference; ROADMAP § 0.7.12
Tier 2 entry retired.Pty::write queues on EAGAIN instead of dropping bytes
(Pty::write, new Pty::onWriteReady). Was: the master FD
is non-blocking, so a slow consumer on the slave side could fill
the kernel PTY buffer; on EAGAIN the write loop's else-clause
broke out with the comment // EAGAIN or fatal error, silently
dropping the unwritten remainder. Behaviourally invisible during
normal interactive use (the kernel buffer drains within
microseconds) but provoked by realistic bursts — large pastes,
AI-dialog command insertions, plugin-driven keystroke floods,
or any write into a slave whose reader is suspended. Now: a new
m_pendingWrite byte buffer and a QSocketNotifier( QSocketNotifier::Write) on m_masterFd (initially disabled
because PTY masters are writable nearly continuously and a
hot notifier would burn CPU). On EAGAIN the unwritten remainder
is moved to the queue and the notifier is enabled; when the
kernel signals writability, onWriteReady drains the queue and
disables the notifier on completion. Fresh write() calls
arriving while the queue is non-empty append rather than
bypass, preserving FIFO ordering so newer keystrokes never race
ahead of older ones. Queue capped at 4 MiB
(MAX_PENDING_WRITE_BYTES) — large enough for realistic bursts,
small enough that a permanently-stuck slave cannot OOM the GUI
process. Regression test tests/features/pty_write_eagain_queue
locks seven invariants — write-side notifier creation (INV-1),
queue member declared (INV-2), notifier pointer member declared
(INV-3), onWriteReady slot declared and connected (INV-4),
EAGAIN handled distinctly inside Pty::write (INV-5), 4 MiB
cap present (INV-6), direct-write path checks queue first for
FIFO ordering (INV-7). Pre-fix source fails 5 of 7; post-fix
all 7 pass. ROADMAP § 0.7.12 Tier 2 entry retired.Theme: menubar opacity — the actual root-cause fix. 0.7.25's palette + widget-local-QSS belt-and-suspenders did not in fact suspend any belt: the user reported the desktop wallpaper still showing through the menubar strip on KWin + Breeze + Qt 6.
WA_TranslucentBackground (OpaqueMenuBar, MainWindow::applyTheme).
User report 2026-04-25: "I can clearly see my desktop background
behind it." Root cause: under WA_TranslucentBackground parent +
WA_OpaquePaintEvent on the menubar, none of the conventional
opaque-paint paths runs reliably on every WM/style stack. Specifically:
autoFillBackground is suppressed by WA_OpaquePaintEvent (the
contract is "the widget paints all pixels"), QPalette::Window is
only consulted by autoFillBackground and inherits the suppression,
and QSS QMenuBar { background-color: … } — which is supposed to
draw via QStyleSheetStyle::drawControl(CE_MenuBarEmptyArea) —
is silently skipped on KWin + Breeze + Qt 6 when
WA_OpaquePaintEvent is set, because the QSS engine assumes the
widget owns those pixels. Net effect: every safeguard 0.7.25 added
was correctly installed and not painting anything; the menubar
surface stayed cleared-to-transparent and the compositor showed the
wallpaper through. New file src/opaquemenubar.h defines
OpaqueMenuBar, a QMenuBar subclass whose paintEvent runs
QPainter::fillRect(rect(), m_bg) with CompositionMode_Source
before delegating to QMenuBar::paintEvent. That is the only
path that actually keeps the WA_OpaquePaintEvent contract honest
under translucent parents. m_menuBar is now an OpaqueMenuBar and
applyTheme sets the fill colour via setBackgroundFill(theme.bgSecondary).
The 0.7.25 palette + widget-local QSS calls are kept, no longer load-
bearing for the strip's opacity but useful for child-widget theme
propagation and for scoping the ::item :hover/:selected/:pressed
rules on the menubar itself rather than relying on the top-level
cascade. Regression test tests/features/menubar_hover_stylesheet
extended with INV-8 — three new assertions that pin the
OpaqueMenuBar construction site, the setBackgroundFill call in
applyTheme, and the presence of paintEvent + fillRect inside
src/opaquemenubar.h. The full spec is rewritten with the
per-iteration history (0.6.42 → 0.7.26) and an explicit "manual
verification" recipe (bright wallpaper + dark bgSecondary +
~0.85 opacity on KWin) so any future drift is reproducible by hand.Theme: memory-DoS hardening on OSC 8 hyperlinks and Kitty APC graphics chunks, plus a menubar-opacity fix reported mid-session.
TerminalGrid). Two attacker-controlled accumulators used to be
bounded only by the VT parser's per-envelope 10 MB ceiling, with no
downstream ceiling of their own. A hostile program could emit one
\x1b]8;;<10MB-URI>\x07 sequence per scrollback line and wedge
tens of GB of URI bytes into per-row HyperlinkSpan copies +
scrollback; or keep sending \x1b_G...,m=1,...\x07 frames without
ever closing with m=0, growing m_kittyChunkBuffer without
bound. src/terminalgrid.h now declares
MAX_OSC8_URI_BYTES = 2048 (real URLs are < 2 KiB; abuse is not)
and MAX_KITTY_CHUNK_BYTES = 32 MiB (larger than any realistic
chunked image upload — Kitty itself transmits in ~4 KiB frames).
The OSC 8 open branch rejects oversized URIs down the same drop
path as the existing invalid-scheme rejection — following text
prints unlinked. The APC chunk path clears and shrink_to_fits
the staging buffer on cap breach, so a subsequent m=0 sees an
empty buffer rather than attacker-prepended garbage. Regression
test tests/features/osc8_apc_memory_caps locks five invariants —
OSC 8 happy path, oversized-drop, at-cap-accepted, APC single
frame, APC overflow-dropped. Pre-fix source fails INV-OSC8-B and
INV-APC-B; post-fix all five pass.WA_TranslucentBackground
(MainWindow::applyTheme). User report 2026-04-24: "the
background of the menubar is transparent." Root cause: the QSS
cascade from the QMainWindow-level stylesheet can race with the
compositor damage rect under WA_TranslucentBackground — on frames
where the compositor invalidates the chrome region before QSS polish
has produced a paint, the menubar shows through to whatever is
behind the window. applyTheme now installs a belt-and-suspenders
opaque fill on m_menuBar: QPalette::Window = theme.bgSecondary
(so autoFillBackground lays down an opaque fill before QSS
paints) plus a widget-local setStyleSheet(QMenuBar ...) block (so
the menubar's own style engine polishes the QSS independently of
the top-level cascade). Mirrors the same pattern already used on
the custom title bar and status bar. tests/features/menubar_hover_stylesheet
extended with INV-7 locking both the palette and widget-local-QSS
calls.Theme: wide-char (CJK / wide emoji) overwrite no longer strands its mate. 0.7.12 Tier 2 hardening item.
TerminalGrid). A wide character occupies two adjacent cells —
a first half (isWideChar=true, codepoint=CP) and a continuation
(isWideCont=true, codepoint=0). Before this release, three write
paths in src/terminalgrid.cpp left half-pairs in an inconsistent
state when a new write landed on only one half: (a) handlePrint
narrow write over a continuation left the mate at col-1 still
claiming isWideChar=true with no neighbor (rendered with a gap);
(b) handlePrint narrow write over a first half left the old
continuation at col+1 claiming isWideCont=true with no mate
(blocked selection / copy); (c) handlePrint wide write shifted
by one cell from an existing wide pair left the old continuation
at col+2 orphaned; (d) the SIMD-coalesced fast path
handleAsciiPrintRun had the same left/right edge issues on its
write span. Consolidated the policy into a single
breakWidePairsAround(row, startCol, endCol) helper that runs
before every write site: it blanks the stranded first half on the
left edge and clears the stranded continuation on the right edge.
Regression test tests/features/wide_char_overwrite_mate locks
five subcases — pre-fix source fails four of them (INV-1, INV-2,
INV-3, INV-4), post-fix all five pass.Theme: xterm-compatible Background Color Erase (BCE) across every scroll and erase path. 0.7.12 Tier 2 hardening item.
TerminalGrid). Apps
that paint with a non-default SGR background (\e[44m etc.) and
then scroll, insert lines, delete lines, delete chars, or insert
blanks expect the newly-exposed cells to inherit the current bg —
this is the xterm convention vim/less/tmux/mc/htop rely on for
full-screen painted backgrounds. Previously takeBlankedCellsRow()
hardcoded m_defaultBg, so CSI L / CSI M / CSI S / CSI T
/ LF-past-scroll-bottom all produced default-bg gaps through any
painted region. deleteChars / insertBlanks used raw
m_currentAttrs.bg without the .isValid() fallback that
clearRow already had. Consolidated the policy into one
TerminalGrid::eraseBg() helper; every erase/scroll callsite
now reads from it. Regression test:
tests/features/bce_scroll_erase/spec.md — 10 behavioral
subcases (IL, DL, SU, SD, DCH, ICH, ED2, EL0, LF-scroll,
SGR-reset-before-erase). Pre-fix source fails 5 subcases
(IL/DL/SU/SD/LF-scroll); post-fix all 10 pass.Theme: user-visible About menu (user ask 2026-04-24 — "How do I see what version of Ants Terminal I am running? Can you please add a GUI version?").
Help → About Ants Terminal… menu. New rightmost menu on the menubar (matching Linux desktop HIG: Help is always last). Two actions:
ANTS_VERSION (read directly from CMake's
project-wide single source of truth, no hardcoded literal),
the Qt runtime version (via qVersion()), the Lua engine
version (when compiled with ANTS_LUA_PLUGINS), a one-line
summary, and a clickable GitHub homepage link
(TextBrowserInteraction enabled so URLs in the dialog open
in the user's browser).QMessageBox::aboutQt dialog.
Inherits future Qt-version bumps automatically.Pre-fix, ants-terminal --version on the command line was the
only path to read the running version. Regression test:
tests/features/help_about_menu/spec.md — 6 invariants
(Help-menu-last, About-action present, ANTS_VERSION
referenced not hardcoded, Qt::RichText +
Qt::TextBrowserInteraction set, About-Qt action routed to
QMessageBox::aboutQt, no "0.7." literal inside the
handler body).
Theme: Lua sandbox hardening trio from the 0.7.12 /indie-review.
Three small defense-in-depth fixes in LuaEngine, behavioral +
source-grep regression test locked to all three.
string.dump removed from the plugin sandbox.
string.dump(f) returns the bytecode serialization of a Lua
function. Lua 5.4 has no bytecode verifier — the loader parses
any byte sequence beginning with \x1b as a binary chunk, and
crafted bytecode can corrupt Lua's internal state and escape
the sandbox. load/loadstring/loadfile are already nilled
at init, so there is no supported round-trip from
string.dump back to executing bytecode, but a future C API
added to ants.* that wraps luaL_loadbuffer with plugin-
supplied data would reopen the attack surface. Closing the
primitive at the sandbox layer — lua_setfield(m_state, -2, "dump") scoped to the string table, not a blanket
lua_setglobal — is cheaper than auditing every future C API
for the same rule.LuaEngine::loadScript forces "t" (text-only) load mode.
The pre-fix path luaL_dofile → luaL_loadfile →
luaL_loadfilex(L, path, nullptr) accepted both text and
binary chunks at the loader level. The 0x1b-first-byte peek in
loadScript was the first gate; the loader call is now the
second gate. A future refactor that drops the peek still gets
rejection at the Lua level.lua_close in
shutdown. lua_close runs every pending __gc metamethod
in dependency order. Metamethods can execute arbitrary Lua
code which the count hook observes. If the hook fires mid-
close and walks back into registry data or the dying engine
pointer via __ants_engine, we get a UAF window. Clearing
the hook first with lua_sethook(m_state, nullptr, 0, 0)
removes that window; the C-side cleanup proceeds without any
Lua-level observer.Regression test: tests/features/lua_sandbox_hardening/spec.md —
6 invariants, behavioral (string.dump is nil, valid-text loads,
0x1b-first-byte rejected) plus source-grep on the three fix
tokens (luaL_loadfilex(..., "t"), lua_sethook(m_state, nullptr, 0, 0) before lua_close, lua_setfield(m_state, -2, "dump")
scoped to the string table). Verified to fail against pre-fix
source via git stash — 5 of 9 invariants flip red on the
regression, all green after the fix lands.
Theme: Tier 2 hardening sweep — three open 📋 items from the
0.7.12 /indie-review landed as one release. Each has a behavioral
regression test, source-grep on the fix's load-bearing tokens, and
was verified to fail against pre-fix source via git stash before
locking.
debug.log lands 0600 regardless of umask.
~/.local/share/ants-terminal/debug.log was being created with
the process umask (typically 0644 under 0022). The log can
include PTY keystrokes (via the input / pty categories), AI
endpoint request+response bodies (network), OSC 133 HMAC digest
material (shell), and Claude transcript parse state (claude) —
every one of those is material that must not be world-readable.
DebugLog::setActive now calls setOwnerOnlyPerms twice after
the file opens: once on the QFileDevice to cover the just-opened
fd, and once on the path string to narrow any pre-existing 0644
file that append reused from a prior (pre-fix) run. Fix uses the
project-standard secureio.h helper, not a raw QFile::setPermissions
bitmask. Regression test: tests/features/debuglog_perms/spec.md
(4 invariants: fresh-open perms, clear-then-reopen, pre-existing
0644 narrowed to 0600, source uses the helper).file field. User-
supplied audit rules, audit_rules.json in a cloned project, and
external scanner regex outputs can all produce findings whose
file field is e.g. ../../etc/passwd. Pre-fix, six call sites
across AuditDialog naively concatenated m_projectPath + "/" + f.file and passed the result to readSnippet / lineIsCode /
comment scans / AI-triage POST bodies — a textbook CWE-22 +
OWASP LLM06 (sensitive-information disclosure via LLM) chain, since
the AI-triage surface exfiltrates snippet contents to the configured
/v1/chat/completions endpoint. New AuditDialog::resolveProjectPath
helper canonicalizes the candidate path (resolves .. and
dereferences symlinks in a single QFileInfo::canonicalFilePath
call), requires the result to be anchored under the canonical
project root (with a trailing-slash sentinel so /proj-foo can't
escape from /proj), and returns QString() on any rejection.
All six call sites migrated: dropFindingsInCommentsOrStrings,
inlineSuppressed, the enrichment pass, single-finding AI-triage
snippet fallback, batch AI-triage snippet fallback, and the
dropIfContextContains regex-captured-relPath read. Regression
test: tests/features/audit_path_traversal/spec.md — 5 invariants
behavioral (via a byte-faithful reference reimpl — production
helper is private on the heavy QDialog subclass) plus source-grep
on the production code to confirm migration + helper structure
(canonicalFilePath + anchored startsWith).MainWindow caches the SettingsDialog across Preferences...
opens; the dialog was constructed with &m_config and populated
its widgets from the then-current values at construction. When
QFileSystemWatcher fired onConfigFileChanged on an external
edit, m_config = Config() reloaded from disk but the cached
dialog still held pre-reload widget state. Next Preferences...
open would show stale values, and clicking OK would replay them
over the fresh Config — silently undoing the external edit.
onConfigFileChanged now closes the dialog if visible, calls
deleteLater(), and nulls the pointer, so the next open rebuilds
from the freshly reloaded Config. Regression test:
tests/features/settings_dialog_config_reload/spec.md —
4 source-grep invariants (cache nulled, visible-close gate,
deleteLater-not-delete, invalidation scoped to
onConfigFileChanged).Theme: CI un-break + tab-rename persistence. CI had been red since
0.7.17 on the AppStream metainfo validation step — the 0.7.17 release
description embedded git clone https://ghp_…@github.com/… as an
example of what the new secret-redactor scrubs, and appstreamcli
correctly rejects plaintext URLs in <description> bodies. Rewrote
the example to use an inline non-URL form; local
appstreamcli validate --explain now exits 0. Separately, user asked
whether manual tab renames (right-click → "Rename Tab…") survive
Ants restart. They didn't — the pin map lived only in MainWindow
memory and SessionManager didn't serialize it. Fixed by bumping
the session-file schema to V3 with a trailing pinnedTitle field;
V2 files still load with the out-param defaulting to empty.
m_tabTitlePins)
is now written to and read from each per-tab session file, so a
tab renamed to "Deploy" or "Prod DB" or "Claude #3" keeps that
label after the app exits and relaunches. SessionManager
schema bumped to V3: a trailing QString pinnedTitle field is
appended after the V2 cwd. V2 files continue to load via the
existing in.atEnd() gate — Ants 0.7.18 and earlier can't read
V3 files, but since session files are a per-user cache (not an
interchange format) that's by design. MainWindow::saveAllSessions
threads m_tabTitlePins.value(w) (keyed by the outer tab widget,
which may be a QSplitter for split tabs) into the save;
restoreSessions pulls the pin back, populates the in-memory
m_tabTitlePins map, and sets the tab label directly (pin takes
precedence over the shell-derived window title from the saved
grid's windowTitle()). Contract locked by
tests/features/tab_rename_persist/spec.md — 20 invariants
covering the V3 round-trip (three pin lengths including empty),
V2 backward compat (hand-crafted V2 stream → restore leaves
pinnedTitle out-param empty), and MainWindow source-grep that
both save-side and restore-side wiring remain threaded. Verified
to fail against pre-fix source before locking.appstreamcli validate --explain failed on the 0.7.17 / 0.7.18 release
descriptions because the 0.7.17 block contained
git clone https://ghp_…@github.com/… as a scrubber example,
and AppStream's description-has-plaintext-url rule rejects
raw URLs in description bodies. CI's "Validate AppStream
metainfo" step exited 3 on every push since 0.7.17. Rewrote the
example as <code>git clone</code> with an embedded <code>ghp_</code> token in the URL. CI build-test step should
be green again on this release.SessionManager::restore initializes optional out-params
before reading. Previously, callers that passed a
pre-populated QString *cwd or QString *pinnedTitle would
see their sentinel survive an older-format load (V2 files
leaving pinnedTitle untouched, V1 files leaving cwd
untouched). Fixed by clearing both out-params at function
entry regardless of the stream's version. The version-gated
read blocks still populate them only when the on-disk format
actually has the field; the clear-first guarantees they read
as empty rather than as whatever the caller happened to
pre-fill. Locked by tab_rename_persist I3.Theme: post-release documentation + tech-debt sweep. Three parallel
audit agents (doc drift, source debt, spec-vs-code drift) inventoried
every doc-vs-code mismatch, stale TODO, dead config key, and spec line
number across the tree. The shell_command config key turned out to
be the only actual functional bug: Settings → General's Shell picker
persisted the user's choice to config.json but no code path ever
read it — every tab opened $SHELL regardless of what the user picked.
Two other deferred-but-flagged items from the 0.7.12 ROADMAP also
landed: Qt::WA_OpaquePaintEvent on the menubar (closes the
mouse-move-over-menubar dropdown flicker on KWin that the 0.7.4
QOpenGLWidget→QWidget refactor didn't fully eliminate), and
background_alpha config key removed as redundant after user
confirmed the opacity knob already does what they want.
background_alpha config key + Config::backgroundAlpha() +
Config::setBackgroundAlpha() + TerminalWidget::setBackgroundAlpha()TerminalWidget::backgroundAlpha() + m_backgroundAlpha member.**
The key had no paint-site consumer (the fillRect at
terminalwidget.cpp:605 always used m_windowOpacity, which is
driven by the opacity config key) and no UI widget. User
confirmed (2026-04-24) they use opacity ~0.9-0.95 to make the
terminal area translucent while the chrome stays solid — which is
exactly what the existing paint path does. Removing the key
collapses two overlapping knobs into the one that works.
Backwards-compat note: stale "background_alpha" entries in
existing config.json files are harmlessly ignored on load (Qt's
JSON parser tolerates unknown keys). README config-defaults table
and example JSON both pruned; CLAUDE.md "Per-pixel bg alpha is
independent of window opacity" bullet rewritten to describe the
actual single-knob behaviour. m_windowOpacity member kept (name
is historical — it drives per-pixel terminal-area fillRect alpha,
not Qt's whole-window setWindowOpacity); header + paint-site
comments clarify the misnomer.Config::shellCommand() actually launches the user's shell.
Settings → General's Shell picker (Default / bash / zsh / fish /
Custom…) has persisted the user's choice to shell_command in
config.json since the original implementation, and the Settings
dialog loads/re-displays it correctly. But TerminalWidget::startShell
took only a working directory and invoked VtStream::start with a
hardcoded empty shell arg — Pty::start's empty-shell fallback
always fired ($SHELL → pw_shell → /bin/bash), silently
dropping the user's choice. Fix: startShell gains a shell
parameter, the QMetaObject::invokeMethod call forwards it
instead of QString(), and every production call site in
MainWindow (newTab, newTabForRemote, splitCurrentPane,
the restoredTabs loop in session restore) passes
m_config.shellCommand(). Contract pinned by
tests/features/shell_command_wiring/spec.md — 7 invariants
covering the signature, the Q_ARG forward, the four wired call
sites, a zero-orphan guard, and a Config round-trip. Verified to
fail against pre-fix source (5 invariants fire on missing
signature + hardcoded QString() + four orphan call sites)
before locking.Qt::WA_OpaquePaintEvent on m_menuBar). The menubar's
autoFillBackground + stylesheet fully cover every pixel on every
paint, so marking it opaque-on-paint stops Qt from invalidating
the translucent parent's compositor region under it when it
repaints. Without this, each mouse-move over the menubar item
owning an open dropdown damages the popup above the menubar on
KWin. MainWindow ctor now calls
m_menuBar->setAttribute(Qt::WA_OpaquePaintEvent, true); the
QSS comment at the :hover rule always promised this was set at
the construction site, but the call was missing. Flagged in 0.7.12
Tier 3 hardening sweep as "Missing per
tests/features/menubar_hover_stylesheet/spec.md INV-3b". Locked
by the reinstated INV-3b assertion in
test_menubar_hover_stylesheet.cpp (was previously punted to
terminal_partial_update_mode, but that test covers an orthogonal
fix — both now assert their own attribute).README.md config-defaults table corrected. session_persistence
default shown as false in both the example JSON block and the
key-description table; actual default in config.cpp:311 is true.
font_size range shown as 8–32; actual range in config.cpp:170
and settingsdialog.cpp:206 is 4–48. Both corrected — the
authoritative code defaults win.PLUGINS.md Lua resource-limit docs. Previously claimed limits
are enforced "per plugin per event invocation" and that each event
gets a fresh 10 M-instruction budget. The implementation sets
lua_sethook(LUA_MASKCOUNT, 10000000) once at engine init; the
counter accumulates across all handlers in the same VM until the
hook fires, at which point it resets. Doc rewritten to describe
the real contract: "≥10 M instructions anywhere in the VM aborts
the current pcall; plugin is not unloaded, next event starts from
wherever the hook last fired."tests/features/terminal_partial_update_mode/spec.md rewritten
to match the live test. The .cpp test was updated in 0.7.4 to
enforce the post-refactor invariant (TerminalWidget is a plain
QWidget, no QOpenGLWidget:: calls, no makeCurrent()), but
the adjacent spec.md still documented the pre-0.7.4
setUpdateBehavior(PartialUpdate) fix that didn't work. Spec now
describes the real fix (base-class change), lists the 4
invariants the test actually enforces, and preserves the
PartialUpdate detour in a History section for archaeology.
Directory name kept for CMake stability — rename would touch
add_executable / add_test / label wiring.osc133_hmac_verification
(1297 → 1722), session_persistence_default (config.cpp:218 → 310),
review_changes_click (mainwindow.cpp:74 → 148, ~411 → ~595),
ai_context_redaction (aidialog.cpp:114-136 → 115-137). Line-
number drift is inherent to spec.md files; this sweep resets the
ground state, no workflow change.Theme: hot-path parser optimization + stale-TODO sweep. VtParser
printable-ASCII runs now coalesce into a single VtAction carrying a
pointer+length slice of the feed buffer; TerminalGrid grows a
handleAsciiPrintRun fast path that batches per-row cell writes so
markScreenDirty, the combining-char erase, and the cell() clamp
all fire once per row instead of once per byte. Rebench on 8 MiB
corpora: ascii_print 24.4 → 30.7 MB/s (+26%), ansi_sgr
27.5 → 31.4 MB/s (+14%). Two Tier 2 TODOs also closed: SIGPIPE
was already shipped in 0.6.28 (bc97485) but never had a regression
test — now grep-locked; RIS (ESC c) preserved only 2 of 8 integration
callbacks, silently breaking notifications, progress, command-finished,
user-vars, and the OSC 133 HMAC forgery alarm after any tput reset —
now preserves all 8 plus the HMAC key.
UserPromptSubmit git-context hook (user ask
2026-04-24). New Settings-dialog button "Install git-context
hook" writes ~/.config/ants-terminal/hooks/claude-git-context.sh
and merges one entry into ~/.claude/settings.json under
hooks.UserPromptSubmit. The script prints a compact
<git-context> block — branch, upstream, ahead/behind counts,
staged/unstaged/untracked file counts — on every user prompt, so
Claude Code sees repo state up front without spending tokens on
a Bash(git status) round-trip. Global scope (the user's explicit
ask "can this be available to all projects?"): installed in
~/.claude/settings.json so it fires on every Claude Code session
regardless of which project. No-op outside a git repo, no-op when
git isn't on PATH, no-op when CLAUDE_PROJECT_DIR is unset and
$PWD isn't a repo — silent exit-0 so non-repo sessions see zero
behaviour change. Independent of the existing status-bar hook
installer (different direction of data flow: this one is
Claude→git, the status-bar one is Ants→Claude); users can opt
into one and not the other. Installer is idempotent and preserves
pre-existing user-added UserPromptSubmit entries (ripgrep
cheatsheet injectors, TODO listers, etc.) via the same
"append-only-if-not-already-present" loop used by the status-bar
installer. Parse-error-refuse on a corrupt settings file matches
the 0.7.12 /indie-review cross-cutting fix shape. Contract pinned
by tests/features/claude_git_context_hook/spec.md + 10-invariant
installer source-grep test + 5-scenario behavioral script test.src/secretredact.h — OWASP LLM06 defense layer. Header-only
module exposing SecretRedact::scrub(QString) → {text, redactedCount}.
Regex-scrubs 14 well-known secret shapes out of any string before it
leaves the process: AWS access keys (AKIA/ASIA), GitHub classic /
OAuth / app / fine-grained PATs, Anthropic sk-ant-, OpenAI
sk-proj- + legacy sk-, Slack xox[abpros]-, Stripe sk_live_ /
rk_live_, JWTs, Bearer <token> headers, generic
api_key=/token=/password=/secret= assignments, and
multi-line PEM -----BEGIN … PRIVATE KEY----- blocks. Replacements
are [REDACTED:<kind>] so the LLM still sees the surrounding
command/variable shape. Priority-ordered rule set guarantees
sk-ant-… is labelled anthropic, not mislabelled as legacy
openai; overlap resolution drops lower-priority matches in the
same range. Pure function, thread-safe, static-initialised once per
process.AiDialog::sendRequest now scrubs both outbound strings. The
terminal scrollback context (m_terminalContext) and the user's
own question (userMessage) both pass through SecretRedact::scrub
before either is interpolated into the system prompt or the user
message body. Pre-fix, a single cat .env, aws configure show,
or git clone https://ghp_…@… line in recent scrollback exfiltrated
the credential on the next AI request. When the combined
redactedCount > 0, the dialog appends a System chat-history
note — "Note: N secret(s) redacted from outbound request (OWASP
LLM06)" — so the user knows the payload differs from what they
saw/typed. The chat history's "You:" display intentionally shows
pre-redaction text (redaction is a network-boundary concern, not a
UX one).ai_context_redaction — 16 positive cases (one
per secret shape) + 6 negative controls (UUID, commit SHA, plain
URLs, prose mentioning "api token" without a value, empty string,
ls -la). Asserts no raw secret substring survives, the expected
[REDACTED:<kind>] label appears, surrounding text is preserved,
cumulative redactedCount is accurate across multiple secrets,
precedence orders sk-ant- above legacy sk-, and (source-grep)
AiDialog::sendRequest wires SecretRedact::scrub to both
inbound strings and gates the user-notice on totalRedacted > 0.
Verified to fail against pre-fix source (5 grep invariants) before
locking.tab_rename_pin — source-grep regression guard
that pairs the right-click rename handler with the pin-map
consumer side. Asserts the rename lambda calls
setTabTitleForRemote(…), never calls m_tabWidget->setTabText
directly, does not guard the call behind !newName.isEmpty()
(empty-string clear is deliberate UX), and that the consumer-
side m_tabTitlePins.contains(...) guards on both the
titleChanged handler and updateTabTitles still exist — if
either is deleted "to clean up," the fix silently regresses.vtparser_print_run_coalesce — locks grid-state
equivalence between the bulk-feed (SIMD + coalesce) path and the
byte-by-byte scalar path across 8 invariants (ASCII, CSI at every
lane boundary, wrap at right edge, cursor mid-row, mixed UTF-8,
empty input, combining-mark clearing).sigpipe_ignore — source-grep regression test that
pins std::signal(SIGPIPE, SIG_IGN) in main.cpp before the
QApplication constructor. Installing it after QApplication leaves
any write path triggered by Qt resource loading exposed to
termination.ris_preserves_callbacks — exercises all 8
TerminalGrid integration callbacks plus the m_osc133Key HMAC
config before and after an ESC c reset and asserts every one
survives. Pre-0.7.17 only responseCallback + bellCallback were
preserved; the other 6 silently wiped.VtAction carries an optional coalesced Print run. Two new
fields: const char *printRun + int printRunLen. When set, they
point into the caller's feed buffer (valid for the duration of the
ActionCallback call only). TerminalWidget's direct-callback path
uses this to skip wcwidth, the combining-char check, the wide-char
branch, and std::clamp inside cell() on every printable byte.VtStream::onPtyData callback expands runs into per-byte Prints.
Its batch outlives the feed buffer, so it MUST materialize the run
into per-byte Print actions before queuing for the GUI thread.
Lifetime invariant documented in vtparser.h alongside the field.tests/features/vtparser_simd_scan + threaded_parse_equivalence
canonicalize runs before comparing. Both tests establish action-
stream equivalence across feed strategies; the expand-runs-to-per-
byte canonicalization preserves the contract after coalescing.settings.local.json perms survive rename
fallback (belt-and-suspenders). ClaudeAllowlistDialog::saveSettings
previously set 0600 only on the QSaveFile temp fd, relying on
rename to carry the perms across to the final path. Most local
filesystems (ext4/xfs/btrfs) preserve fd perms across rename(2),
but FAT/exFAT on removable media don't carry POSIX perm bits at
all, some SMB/NFS servers apply umask to the rename destination,
and Qt's QSaveFile::commit falls back to copy+unlink on
cross-device renames — in all three cases the final file lands
at the process umask (typically 0644), leaving it readable to
every other UID on the host. Since settings.local.json can
hold Claude Code bearer tokens in merged env/model/custom-hook
keys, the widened perms are a credential-leak surface. Fix: after
file.commit() returns true, call setOwnerOnlyPerms(m_settingsPath)
on the final path too. The pre-commit fd-level chmod is retained
(covers the open-to-commit window on filesystems that do
preserve fd perms — neither call subsumes the other). Flagged by
the 2026-04-23 /indie-review sweep. Locked by
tests/features/allowlist_perms_postcommit/spec.md — runtime
stat check under umask 0022 (the POSIX default) plus source-grep
that both chmod calls remain and the post-commit call is
sequenced after the if (!file.commit()) return false; gate so
a failed commit can't chmod a path that may not exist.cat .env, env | grep, aws configure show,
git clone https://ghp_…@github.com/…, or freshly-pasted SSH key
landed in the provider's request logs or training-data pipeline.
The user's own pasted questions were equally exposed. Fix: both
outbound strings now pass through a 14-shape regex scrubber before
reaching the JSON body. Complements the 0.7.12 LLM01/LLM02 fix
(ai_insert_command_sanitize) — that handled untrusted input from
the AI to the PTY; this handles untrusted output from the PTY to
the AI. Contract pinned by
tests/features/ai_context_redaction/spec.md (9 invariants).mainwindow.cpp:4284 wrote directly to m_tabWidget->setTabText
without populating m_tabTitlePins; the per-shell
titleChanged signal handler and the 2 s updateTabTitles tick
both consult the pin map to decide whether to relabel, so any
tab whose shell writes OSC 0/2 (Claude Code writes one every
prompt iteration, ~3–5 s) had its manual name wiped within
seconds. Rename now routes through the existing
setTabTitleForRemote (the rc_protocol set-title path) so
non-empty names pin, and empty names clear the pin and restore
the format-driven / shell-driven label — gives the user an
in-UI "un-rename" path that didn't exist before. Locked by
tests/features/tab_rename_pin (4 invariants).ESC c) now preserves all 8 integration callbacks +
m_osc133Key. Previously notifyCallback, progressCallback,
lineCompletionCallback, commandFinishedCallback, userVarCallback,
and osc133ForgeryCallback silently became empty functions after any
tput reset / reset(1) / stty sane. Desktop notifications,
progress bars, command_finished plugin events, user-var theming
hooks (jj, starship), and the OSC 133 HMAC forgery alarm all died
after a single RIS. Security impact for the forgery alarm: a
well-timed tput reset silenced it permanently, enabling OSC 133
forgery to proceed without user notification. m_osc133Key (read
from $ANTS_OSC133_KEY at process start) also now survives RIS.scanSafeAsciiRun on SSE2/NEON) now emits one VtAction::Print
per run instead of one per byte. For a 10 KB TUI-repaint PTY read,
the allocation count drops from ~10 000 to 1 and the dispatch count
drops by the same factor. Bench deltas on 8 MiB corpora:
ascii_print 327 → 260 ms (−20%), ansi_sgr 291 → 255 ms (−12%).
The newline_stream and utf8_cjk corpora are bound by
newLine/scroll and UTF-8 multibyte decoding respectively — no
safe-ASCII runs to coalesce — so their numbers are flat.TerminalGrid::handleAsciiPrintRun batches per-row writes.
Skips the per-byte wcwidth, combining-char branch, wide-char
branch, and std::clamp inside cell(). Fires markScreenDirty
once per row instead of once per byte.Theme: per-tab Claude Code activity indicator — flagship response to
the 2026-04-23 user request for at-a-glance multi-tab Claude visibility.
Users running Claude Code in several tabs now see per-tab state glyphs
(idle / thinking / tool-use / Bash / planning / compacting /
awaiting-input) so they can tell which tab is blocked on them
without cycling through tabs. Permission prompts route by session_id
(not active-tab) so a prompt from any tab lights up that specific tab.
Supporting refactor: ClaudeIntegration::parseTranscriptForState is now
a thin wrapper over a new static parseTranscriptTail helper so the
new per-tab tracker can share the exact parsing logic with zero drift.
session_id
(matched against the <session-uuid>.jsonl transcript filename) so
a prompt emitted by Claude in tab 3 lights up tab 3's glyph, not
the currently-focused tab. Hover tooltip on each tab shows
"Claude: thinking…" / "Claude: Bash" / etc. so glyph colours can be
disambiguated without switching tabs. Settings dialog → General
adds a checkbox ("Show per-tab Claude activity dot …") that flips
claude_tab_status_indicator (default on); the change takes effect
on the next paint via the live config reload path, no restart
needed. New src/claudetabtracker.{h,cpp} keys state by shell PID
so tab reorder / close doesn't shift entries, and reuses the
existing transcript parser via a new static
ClaudeIntegration::parseTranscriptTail helper — no duplicated
parsing logic. Contract + 11-invariant regression test at
tests/features/claude_tab_status_indicator/. Roadmap item:
2026-04-23 user request.ClaudeIntegration::parseTranscriptForState refactored to a thin
wrapper around the new static parseTranscriptTail helper. Pure
function, no side effects — safe to call from per-shell trackers.
Behavior unchanged; claude_status_bar and
claude_plan_mode_detection regression tests still green against
the same transcripts.
Hook events stash session_id in
ClaudeIntegration::lastHookSessionId(). Handlers of signals
emitted by processHookEvent (e.g. permissionRequested) can now
read the id to route the event per-session rather than treating
every hook as "belongs to the active tab". Required for the
per-tab activity indicator to flag the correct tab on
PermissionRequest.
Theme: fold-in of the 2026-04-23 re-review follow-up list. Six of
nine items had already shipped in 0.7.12 but were never crossed off the
ROADMAP (code was present, bullet was not flipped); the remaining three
are real new work — a /tmp/kwin_*.js orphan sweep, a behavioral
replacement for the fragile source-grep gate test, and a signal-count
assertion on the plan-mode tail-preservation path. Two new regression
tests lock existing invariants.
/tmp/kwin_*_ants_*.js script files.
The position-tracker + move/center helpers write KWin scripts via
QTemporaryFile(autoRemove=false) and rely on a chained dbus-send
callback to remove them after KWin unloads. A crash, SIGKILL, or
dbus hang between write and unload orphans the file — no functional
harm (KWin already loaded its copy), but the files accumulate in
/tmp. MainWindow ctor now runs a once-per-process sweep of
kwin_{pos,move,center}_ants_*.js older than one hour. The
one-hour floor avoids racing an in-flight script another Ants
instance just wrote.remoteControlEnabled() within 400 chars
of m_remoteControl->start(). A refactor that renamed the getter or
hoisted the check into a helper would have silently defeated the
gate while still passing (or failing) the grep. Replaced with a
Config API round-trip assertion: default is false, setter persists
across instances, value is readable post-reload. A getter rename
now fails at compile-time in the test. The structural "start() must
live inside a conditional" check is retained as a plain-substring
lookback to catch a refactor that removes the gate entirely —
without the catastrophic-backtracking risk that killed the initial
regex attempt.runToggleFreeTailPreservesState confirms
planModeChanged fires exactly once across both parse phases
(seed toggles NotSet → plan = one emission; tail sees zero
permission-mode events and must not re-fire the same state).
Guards against a regression that re-emits on every parse.config_parse_failure_guard adds retention-cap invariant (INV-5).
Plants 8 stale .corrupt-* siblings with decreasing mtimes, trips a
fresh parse-failure, asserts only the newest 5 (fresh + 4 old)
remain. Verified to fail against a neutered prune loop (got 9,
expected 5). Sets mtimes explicitly via QFile::setFileTime since
QDir::Time ranks on mtime, not on the filename-embedded timestamp.remote_control_opt_in adds behavioral Config round-trip section.
Sandboxed via XDG_CONFIG_HOME (not setTestModeEnabled — that
routes to a Qt-test-specific path that has collided with real user
configs in practice). Four invariants: fresh-default false, in-
memory round-trip, cross-instance persistence, setter-back-to-
false round-trip.Theme: three Claude Code integration robustness fixes from the 0.7.12+1 re-review checkpoint. All three failed silently — no errors, no degraded-mode logs — so the symptoms were "a transcript- driven UI state that just stopped updating" or "a project that appeared under the wrong name in the picker". All three now have targeted fixture-based regression tests.
parseTranscriptForState read the final 32 KB of the JSONL
transcript and skipped the first line after the seek as "likely
truncated". When the final event was a user tool_result carrying
inline file contents (routine for Read/Grep tool results on large
files), the 32 KB window landed inside that one event, the
firstLine-skip ate it, and the parser returned with zero events —
state frozen at whatever it was before the turn. Replaced with a
doubling-grow window (32 KB → 4 MiB cap) that only trims a
potentially-partial prefix line when the buffer contains at least
two newlines, guaranteeing real content remains for the parser.thinking content
blocks. Real assistant events on extended-thinking-enabled
models lead with one or more thinking blocks followed by the
final text response. ClaudeTranscriptDialog::formatEntry only
rendered text, tool_use, and tool_result block types;
thinking fell into an implicit else that discarded the block.
Now rendered as italicized, dimmed Thinking: paragraphs so
readers can distinguish chain-of-thought from the visible reply.decodeProjectPath mangles project names with embedded
hyphens. Claude Code encodes absolute paths by replacing every
/ with - — lossy, since my-project (leaf name) and
my/project (two segments) produce the same encoded string. The
decoder replaced every - with / unconditionally, turning
~/my-project/sub-dir into ~/my/project/sub/dir. Rewrote as a
greedy left-to-right filesystem-probing walker: at each hyphen,
prefer the / form if that directory exists, fall back to -
(folded into the segment name) if that exists instead, default
to / otherwise. Legacy hyphen-free paths remain unchanged.
Preferred source of truth is still extractCwdFromTranscript
(used by discoverProjects); the probe-based decoder is the
last-resort fallback.claude_transcript_robustness (6 invariants) — tail window
grows past a 40 KB trailing event (INV-1), tail-growth is bounded
at the 4 MiB cap on pathological inputs (INV-2), leaf-name hyphens
are preserved when the directory exists (INV-3), intermediate-
segment hyphens are preserved when the full path exists (INV-4),
missing filesystem hints fall back to the legacy all-slashes
behavior (INV-5), and the canonical repo-style path still decodes
correctly (INV-6). Verified to fail against the pre-fix code on
INV-1/3/4, pass on INV-2/5/6 (which test the safety-cap and legacy
paths).Theme: scope-tightening pass on the audit rule-pack trust boundary
shipped in 0.7.12. The global audit_trust_command_rules bool
over-granted after one opt-in — trusting Ants's own repo implicitly
trusted every cloned repo the user later opened. 0.7.13 replaces the
global flag with a per-project hash-bound trust store so trust is both
narrow and self-invalidating.
audit_trust_command_rules bool is retired. Trust is now keyed on
(canonical projectPath, sha256(audit_rules.json bytes)) via three
new Config methods: isAuditRulePackTrusted, trustAuditRulePack,
untrustAuditRulePack. Consequences:projectPath is canonicalized via QFileInfo::canonicalFilePath,
so symlinks / trailing slashes resolve to a single record.ANTS_AUDIT_TRUST_UNSAFE=1 remains an escape hatch for CI that
can't round-trip through the persisted store.
Upgrade behavior: users who had the legacy bool set to true are
re-prompted (fresh trust decisions per project). This is the safe
default for a security feature.qWarning,
which was invisible outside a terminal-launched invocation. Stderr
copy retained for headless / CI cases.CLAUDE.md streamlined. Compressed from 325 → 147 lines
(19.5 KB → 6.5 KB, ~67% reduction) by dropping the full project-
structure tree (derivable from ls src/), collapsing the verbose
audit pipeline prose, retiring the per-dep install-command list
(each probes which <tool> and self-disables), and dropping the
Config-keys table (derivable from config.cpp). Retained: non-
obvious gotchas, rationale for design decisions, and the release-
file triad rule.tests/features/audit_command_rule_trust/. Seven-invariant
regression test against the new trust API: default-untrusted,
round-trip, hash-bound invalidation, cross-project isolation (with
byte-identical packs to rule out hash coincidence), path
canonicalization through symlinks + trailing slashes, idempotent
untrust, and cross-Config-instance persistence through
config.json. Spec-first: contract written before the code, every
invariant mapped to an assertion with a diagnostic label.Theme: independent-review sweep. Multi-agent /indie-review run
(14 subsystems, 60 HIGH findings), followed by a re-review of the
shipped Tier 1 batch that surfaced three more HIGH findings of the
same shape. Every fix carries a regression test anchored to an
external signal (published spec, CVE, real on-disk data) rather
than author reasoning — closes the self-graded-homework loop.
Final polish pass (this tag) handles the six mechanical follow-ups
from the re-review.
remote_control_enabled (default false, matches Kitty's
allow_remote_control=no). send-text payloads pass through a C0
control-char filter stripping {0x00..0x08, 0x0B..0x1F, 0x7F}
while preserving HT/LF/CR and UTF-8 multi-byte, unless the request
carries "raw": true. Closes the UID-scope keystroke-injection
chain — a same-UID attacker can no longer inject
bracketed-paste-disable + newline payloads. Gate now snapshots
once per process (static bool) so a second MainWindow reading
the config after a runtime toggle doesn't try to bind a socket
the first window isn't listening on — matches the "requires
restart" documentation.root stays empty → next save() writes defaults over the
user's bytes) affected src/config.cpp, src/claudeallowlist.cpp,
and src/settingsdialog.cpp (Install Claude hooks path, which
shares the settings.json file with the Allowlist dialog). All
three now rotate the corrupt file aside via the shared
rotateCorruptFileAside() helper in src/secureio.h
(ms-precision timestamp + collision-retry up to 10 suffixes),
log the backup path, and refuse to write further until the user
intervenes. Settings dialog shows the backup path in a
QMessageBox so the user has an actionable recovery.
rotateCorruptFileAside() now also prunes older
<path>.corrupt-* siblings beyond the newest five — keeps
recovery material available without letting repeated launches
against a broken file accumulate unbounded.std::rename return checked in Config::save. Failure logs
errno via DebugLog::Config and removes the orphan tmp file
instead of silently stranding it on disk.extraArgs sanitizer. SshBookmark::sanitizeExtraArgs
now rejects -oProxyCommand=, -oLocalCommand=,
-oPermitLocalCommand=yes, -oMatch=exec (OpenSSH
Match exec runs /bin/sh -c on CLI-supplied commands), and
-oKnownHostsCommand= (OpenSSH 8.5+, same RCE surface), in both
single-token and space-separated forms, case-insensitive per
OpenSSH's own option parser. Closes the CVE-2017-1000117-adjacent
RCE-via-bookmark-field surface.Insert Cmd sanitize + confirm. New
AiDialog::extractAndSanitizeCommand extracts the fenced code
block and strips C0 controls (minus HT/LF/CR), DEL, C1 controls
(0x80..0x9F incl. NEL U+0085), line/paragraph separators
(U+2028, U+2029), bidi overrides (U+202A..E, U+2066..9 —
Trojan Source, CVE-2021-42574), and zero-width codepoints
(U+200B..D, U+FEFF). Capped at 4 KiB. The click handler now
requires explicit confirmation via QMessageBox::question showing
the literal bytes plus a "N bytes filtered" footer when non-zero;
when the 500-char preview doesn't cover the full command, the
dialog surfaces a truncation warning so an attacker can't hide a
payload past the preview window. OWASP LLM01 + LLM02 mitigation.
Language-hint heuristic now also requires the first line be a
single unbroken token (no space / tab) before treating it as a
language identifier — foo bar\nls no longer has its first line
eaten.command fields gated. audit_rules.json
entries that carry a command: field are now skipped unless
Config::auditTrustCommandRules() is true (default false)
or ANTS_AUDIT_TRUST_UNSAFE=1 is set. A cloned-but-untrusted
repo can no longer run arbitrary shell via the Audit dialog.
Skipped count surfaces via qWarning./tmp/kwin_*.js migrated to QTemporaryFile. The three
callsites in xcbpositiontracker.cpp (window-position poll) and
mainwindow.cpp (Quake-mode move + centre) used predictable
filenames in /tmp, enabling a same-name symlink-swap TOCTOU and
a collision when two Ants instances ran concurrently. All three
now use QTemporaryFile with a XXXXXX template and
setAutoRemove(false) so the async KWin finished callback can
still read the script.claudeintegration.cpp was reading
value("mode") — a field that never appears in Claude Code v2.1.87
JSONL transcripts. Real field is permissionMode. Fixed, and the
toggle-free tail scan now preserves the latched plan-mode state
instead of resetting to false (otherwise a 32 KB window that
scrolls past the last explicit toggle silently loses the flag).qWarning with the QJsonParseError offset when a file
is malformed, so users can find their own typos instead of
silently dropping the plugin or theme from the list.rotateCorruptFileAside helper shared across three subsystems.
Before the re-review, each site had its own ~30-line backup-rotation
block; the three reviewers independently flagged the duplication.
Now a single inline helper in src/secureio.h with retention cap
and ms-collision retry.Config::remoteControlEnabled() into a static local at first
MainWindow construction and reuses across subsequent windows, so
a runtime toggle of the config key during a session can't partially
enable rc on a second window.filterControlChars header comment corrected. Previously said
"C0 / C1 control bytes"; only C0 is stripped at this layer —
C1 codepoints are UTF-8 continuation bytes and can't be
byte-filtered without mangling multi-byte characters. Comment now
names aidialog.cpp as the layer that does strip C1 (operating
on QChar, not raw bytes).tests/features/:claude_plan_mode_detection/ — 8 scenarios against the JSONL
transcript schema; each asserts behaviour against live on-disk
bytes rather than a mock.remote_control_opt_in/ — 17 assertions covering the gate,
filterControlChars strip set, "raw": true bypass, and the
stripped response field.config_parse_failure_guard/ — 12 assertions covering the
three load states (missing / valid / corrupt), backup rotation,
save suppression, and std::rename failure path.ssh_extra_args_sanitize/ — 22 assertions covering all five
dangerous ssh_config directives in both -oKey=val and
-o Key=val forms plus the toSshCommand() end-to-end
invariant.ai_insert_command_sanitize/ — 21 assertions covering the full
strip set (C0 / DEL / C1 / NEL / line separator / bidi overrides
/ zero-width), UTF-8 preservation, 4 KiB length cap, and the
language-hint heuristic regression (whitespace on first line
is not a lang ID).Theme: housekeeping pass — two long-deferred 💭 items from the 0.8+ queue
and a ROADMAP correction. The perf item rewrites the combining-char map
shift in deleteChars / insertBlanks to use node-handle extract() +
key mutate + insert() instead of rebuilding the map. The security item
is a defence-in-depth stat-guard around the remote-control socket's
/tmp fallback cleanup. The correction promotes the 0.7.9-era text-run
QString accumulator work from 💭 to ✅ (it was shipped, the ROADMAP
entry just missed the update).
src/terminalgrid.cpp,
deleteChars / insertBlanks). Old impl built a fresh
std::unordered_map<int, std::vector<uint32_t>>, moved each surviving
entry into it (re-hashing every key), then move-assigned it back.
Rewrote to collect affected keys in two small local vectors
(toErase / toShift), erase the deleted ones, then for each shift
extract the node handle, mutate its integer key, and reinsert —
zero bucket reallocation, zero re-hashing of the value vectors.
deleteChars processes shifts in ascending order (leftward shift
can't collide); insertBlanks in descending order (rightward shift
can't collide). Empty-map fast-path skips the block entirely, so
steady-state rows with no combining content (the vast majority) pay
nothing. Tests combining_on_resize and osc8_insert_delete_lines
still pass./tmp fallback stat-guard (src/remotecontrol.cpp).
RemoteControl::start() used to call QLocalServer::removeServer(path)
unconditionally when listen() failed on a stale socket. The
XDG_RUNTIME_DIR path (0700-owned) is safe, but the /tmp/ants-terminal -<uid>.sock fallback lives in shared-world-writable space where a
symlink or a foreign file of the same name could trick us into
unlinking something unrelated. Added a safeToUnlinkLocalSocket()
helper that lstat()s the path and refuses removal unless it's an
actual S_ISSOCK owned by getuid(). ENOENT passes through (nothing
to remove anyway). On refusal the remote-control layer disables itself
with a clear log message — no silent unlink of unknown files.QString construction
in the text-run accumulator" was actually shipped in 0.7.9 (both
TerminalWidget::paintEvent and GlRenderer::render accumulate
codepoints into a reusable std::vector<char32_t> on the Run struct
and call QString::fromUcs4(data, size) once at run-push time). The
stale entry is now promoted to ✅ with the same-style retrospective
write-up the free-list entry got.Theme: follow-through on the 0.7.9 null-result. The std::rotate scroll
refactor was rejected because it paid for save/restore overhead without
touching the real bottleneck — the per-scroll vector<Cell>(m_cols) heap
allocation for the new bottom row. This release addresses that bottleneck
directly with a small free-list of m_cols-sized cell buffers fed by any
path that discards a row (scrollback eviction, scrollDown, insertLines,
deleteLines) and drained by paths that need a fresh blank row. Steady-state
scroll work on a full scrollback runs allocator-free.
TerminalGrid — cell-buffer free-list for scroll paths. New
m_freeCellBuffers pool (cap 4 entries) + takeBlankedCellsRow() /
returnCellsRow() helpers in src/terminalgrid.{h,cpp}. scrollUp,
scrollDown, insertLines, deleteLines now recycle cell buffers
instead of calling makeRow(m_cols, …) on every scroll. When
scrollback hits m_maxScrollback and a pop_front evicts the oldest
line, its cells vector is salvaged into the pool rather than freed;
the next scroll's new bottom row pulls from the pool (blanked in
place), skipping the allocator entirely. resize() clears the pool
(stale-size entries would be wrong-width). Pool stays at 0–2 entries
in normal use — memory footprint is bounded and negligible. All 49
feature tests still pass; scroll_region_rotate (shipped 0.7.9 to
lock the contract) verified the invariants held across the refactor.bench_vt_throughput newline_stream: 5.26 → 6.09 MB/s (+15.8 %).
Median of ten-run pairs before / after. Other corpora are flat to
mildly improved — ascii_print ~23 MB/s (baseline 23), ansi_sgr
~16.5 MB/s (baseline 16.5), utf8_cjk 8.9 MB/s (baseline 7.9,
+12 %; the CJK payload hits scrollUp too). This is the first perf
result to validate the 0.7.9 investigation's thesis that the
container shuffle was a red herring and the allocator was the actual
hot spot.Theme: planned-work pass — three ROADMAP items from the 0.7.7 "deferred
to 0.8+" queue investigated and shipped or rejected with findings on record.
The investigations that didn't ship still left value behind (a feature-test
that locks scroll-region semantics for any future refactor), and the one
that did ship consolidates 13 path strings through one header and halves the
per-frame QString::fromUcs4 call count in both render paths.
scroll_region_rotate feature-test. Locks the observable behavior
of TerminalGrid::scrollUp / scrollDown — row motion inside
[scrollTop, scrollBottom], outside-region invariance, main-screen
scrollback push, alt-screen no push, hyperlink tracking, and
count > regionHeight clamp. Six scenarios, eight invariants. Pins
the contract so the 0.8 ring-buffer-screen work can refactor safely.
See tests/features/scroll_region_rotate/spec.md.src/configpaths.h — single source of truth for home-relative
paths. Inline namespace consolidating ~/.claude/projects,
~/.claude/sessions, ~/.claude/settings.json,
~/.claude/projects/<enc>/memory/MEMORY.md, and
~/.config/ants-terminal/hooks/claude-forward.sh. 13 duplicated
path-concat sites across claudeintegration.cpp / settingsdialog.cpp
collapsed to one call each. Future XDG-respect pass now has a single
update surface.QString::fromUcs4. Both the QPainter
(terminalwidget.cpp::paintEvent) and GL (glrenderer.cpp::render)
code paths now accumulate codepoints into a std::vector<char32_t>
per run and build the QString once at flush/draw time, replacing
the previous runText += QString::fromUcs4(&cp, 1) pattern that
allocated one small QString per non-space cell and repeatedly
reallocated the run string on append. Saves N per-frame small-QString
constructions in the common ASCII TUI case. Behaviour identical;
ligature shaping via QTextLayout still receives the run text at
the same point.std::rotate scroll refactor. Measured on
bench_vt_throughput newline_stream: the rotate path was 12–15%
slower (5.2 → 4.6 MB/s) than the erase/insert loop it replaces,
because the newline-stream case is always count == 1 where rotate
pays for a save/restore cycle that vector::erase + push_back
avoids. Real bottleneck is the per-scroll makeRow(m_cols, ...)
allocation + heap handoff to scrollback — not the container
shuffle. Reverted. The scroll_region_rotate test stays so the next
attempt can refactor safely. ROADMAP §0.7.7 updated with the
scrollback-push cell-reuse approach as the real 0.8 candidate.settingsdialog and claudeallowlist
actually use QDialogButtonBox with Ok/Apply/Cancel; four other
dialogs listed in the ROADMAP have bespoke button layouts. Would
save ~10 lines across 2 call sites — premature abstraction.auditdialog.cpp uses one shared QProcess
member that runs checks serially (not six copies as the ROADMAP
claimed). The timeout-and-cleanup dance already lives at one call
site. No duplication to extract.Theme: audit fold-in — every regression guard the 0.7.6 / 0.7.7 hardening pass left behind is now a repo-wide detector. Three new rules in the Project Audit dialog turn the one-off sweeps into persistent guards, so the next contributor that re-introduces the pattern gets caught at audit-time.
ssh_argv_dash_host audit rule (Security, Major). Flags any
<< shellQuote(...host...) argv construction without a preceding
args << "--" argv terminator — the CVE-2017-1000117 class. Runtime
filter drops findings when the guard appears within ±5 lines
(spans the if (!user.isEmpty()) / else split in sshdialog.cpp:67-71).
tests/audit_fixtures/ssh_argv_dash_host/ provides the bad/good pair.qimage_load_without_peek audit rule (Qt, Minor). Flags
.loadFromData( calls that aren't either tagged // image-peek-ok
or preceded by a QImageReader size-peek within ±5 lines — the
image-bomb DoS vector that the 0.7.6 hardening sweep fixed. Both
shipped sites in terminalgrid.cpp (OSC 1337 PNG + Kitty graphics)
carry the explicit reviewer-sign-off tag and drop cleanly.setPermissions_pair_no_helper audit rule (Qt, Info). Hygiene
nudge toward setOwnerOnlyPerms() from src/secureio.h when a raw
setPermissions(QFileDevice::ReadOwner | QFileDevice::WriteOwner)
bitmask appears. Pattern deliberately excludes the 0755 hook-script
case (extra | flags after WriteOwner) so the one legitimate
literal-bitmask site in settingsdialog.cpp doesn't false-fire. The
helper itself is suppressed via // ants-audit: disable-file in
secureio.h — it is the definition, not a call site.tests/audit_self_test.sh — three new run_rule lines, coverage
cross-check now sees the new rule ids automatically. 55/55 pass
(up from 52/52).Theme: the 0.7.7 planned-work pass from the hardening four-dimensional review — every item marked 📋 in ROADMAP.md's 0.7.7 section is now shipped. Two de-duplication refactors, three perf hoists, four regression feature-tests locking the 0.7.6 → 0.7.7 fixes.
tests/features/osc8_insert_delete_lines/ (CSI L / CSI M sync
invariant), tests/features/hyperlink_resize_clamp/ (shrink-resize
clamp on the active OSC 8 start coordinates),
tests/features/ssh_dash_host_rejected/ (the -- argv terminator
against CVE-2017-1000117 class hosts), and
tests/features/image_bomb_png_header_peek/ (static source-inspection
guard ensuring every loadFromData is peeked first). Each test is
designed to fail on pre-fix code and pass now. 49 feature/fast tests
total.setOwnerOnlyPerms helper (src/secureio.h) replaces the 11
copies of file.setPermissions(QFileDevice::ReadOwner | QFileDevice::WriteOwner) that were scattered across config.cpp,
sessionmanager.cpp (×2), claudeallowlist.cpp, auditdialog.cpp
(×4), claudeintegration.cpp (×2), remotecontrol.cpp, and
auditrulequality.cpp. One typo (ReadOther, ReadGroup) away
from silently widening access to files that may hold an API key.
Two overloads: one for open QFileDevice& handles, one for existing
path strings. The single 0755 case in settingsdialog.cpp (hooks
shell script — needs executable bit) stays literal by design.shellQuote helper (src/shellutils.h) replaces the previous
static function in sshdialog.cpp and the re-defined lambda in
mainwindow.cpp::openClaudeProjectsDialog. The two earlier impls
differed on empty-string and no-special-chars handling; both call
sites now share one inline implementation.QFontMetrics is no longer constructed per
underlined cell; updateFontMetrics() caches underlinePos and
lineWidth into members refreshed on font change. Selection bounds
are now normalised once per paintEvent into local scalars with an
inline cellInSelection lambda, rather than running std::swap +
std::min/std::max per cell. Measurable on search-highlighted +
selection-active frames.isCellSearchMatch — std::lower_bound instead of linear scan.
m_searchMatches is sorted by globalLine; the binary search jumps
to the first match at or after the target line and iterates only the
matches on that line. The previous implementation self-described as
"binary search" but walked the whole vector from index 0 on every
cell. Expected win: 5–20% of paint time on search-highlighted
content in deep scrollback.Theme: close the 0.7.x remote-control arc and close the audit-signal
gap the 0.7.5 hygiene work opened. Seven rc_protocol commands now cover
the "drive the terminal from a script" story end-to-end (ls, send-text,
new-tab, select-window, set-title, get-text, launch), and the
audit gains two feature-coverage lanes that catch shipped features
without locking tests and spec prose that drifted past a rename.
Audit — feature-coverage lanes (src/featurecoverage.{cpp,h}). Two
new in-process audit checks that close the signal-loss frontier left
by the 2026-04-21 RetroDB audit-hygiene work: scanner calibration was
tightened, so the remaining noise was upstream of the scanners —
features shipping without any locking test, spec prose referring to
symbols that had since been renamed. The lanes are Info/Minor severity;
the value is raising awareness, not gating releases.
Lane 1 — spec↔code drift. For each tests/features/*/spec.md,
extracts backtick-fenced identifier-shaped tokens (CamelCase,
snake_case, kebab-case, scoped::names, dotted.ids) and reports the
ones that no longer appear anywhere under src/. Scoped names fall
back to their tail component (Class::method → method) so rename
refactors that keep the leaf don't false-positive; a curated stopword
list drops ubiquitous Qt / language keywords that always exist.
Lane 2 — CHANGELOG↔feature-test coverage. Parses the top
## [x.y.z] section of CHANGELOG.md, walks the Added/Fixed bullets,
and fuzz-matches each against the set of feature-spec titles.
Backtick-token match wins; ≥2 significant-word overlap falls back
(stopword-filtered, ≥4 chars). Bullets with neither surface as
coverage gaps — release-note claims that never got a locking test.
Both lanes use the AuditCheck::inProcessRunner hook so the pure
parsers plug into the same parseFindings → suppress → render
pipeline as the shelled-out scanners, sidestepping QProcess. Pinned
by tests/features/feature_coverage/ (17 invariants across token
extraction shapes, stopword filtering, top-section isolation,
bullet-section tagging, and both fuzzy-match paths).
Remote-control — launch command. Seventh and final command in
the initial rc_protocol surface. Convenience wrapper that spawns a
command in a new tab — sugar for the idx=$(... new-tab) && ... send-text --remote-tab $idx --remote-text $cmd$'\n' pattern,
collapsed into one round-trip.
{"cmd":"launch","cwd":"<path optional>","command":"<string required>"}
-> {"ok":true,"index":<int>}
command is required (empty / missing yields a documented error
pointing at new-tab for the bare-shell case). Auto-appends \n
to the command if not already present — the convenience
differentiator from new-tab, which stays byte-faithful for
scripts that compose multi-line input or send control sequences.
ants-terminal --remote launch --remote-command 'htop'
ants-terminal --remote launch --remote-cwd /tmp --remote-command 'tail -f log'
Pinned by tests/features/remote_control_launch/.
Remote-control — get-text command. Sixth rc_protocol command.
Returns the trailing N lines of (scrollback + screen) joined with
\n — so scripts can capture output, dispatch on visible state,
or grab the last command's result.
{"cmd":"get-text","tab":<int optional>,"lines":<int optional>}
-> {"ok":true,"text":"<string>","lines":<int>,"bytes":<int>}
Default 100 lines; capped server-side at 10 000 so a runaway
--remote-lines 1000000 against a million-line scrollback can't
blow up the JSON envelope. Reuses
TerminalWidget::recentOutput(int) — already the AI dialog's
context-capture accessor, so format stays consistent across both
consumers. Client CLI adds --remote-lines <n> (validated via
toInt(&ok); non-numeric or negative values reject before the
socket call).
ants-terminal --remote get-text # default 100
ants-terminal --remote get-text --remote-lines 24 # screen-ish
ants-terminal --remote get-text --remote-tab 0 --remote-lines 500 # specific tab
Pinned by tests/features/remote_control_get_text/ — 6 invariants
including the 10 000-line cap, the response field set
(text/lines/bytes), and the client-side --remote-lines
validation.
Remote-control — set-title command. Fifth rc_protocol command.
Pins a tab label that survives both the per-shell titleChanged
signal (OSC 0/2 from the inferior) and the 2 s updateTabTitles
refresh — exactly the behaviour you want for "label this tab prod
while I work on it" workflows.
{"cmd":"set-title","tab":<int optional>,"title":"<string>"}
-> {"ok":true,"index":<int>}
Empty title clears the pin and the auto-title path resumes:
under tabTitleFormat == "title" (default) the most recent
shell-provided title is restored from TerminalWidget::shellTitle()
immediately rather than waiting for the next OSC 0/2; under
cwd / process formats updateTabTitles() rebuilds from current
state. Pin storage is keyed by the tab's QWidget* and freed
alongside m_tabSessionIds at tab-close time so the
QHash<QWidget*, QString> doesn't accumulate dangling keys.
Client CLI adds --remote-title <str> (isSet()-gated, so an
empty string legitimately clears the pin):
ants-terminal --remote set-title --remote-title '★ prod'
ants-terminal --remote set-title --remote-tab 2 --remote-title 'logs'
ants-terminal --remote set-title --remote-title '' # clears pin
Pinned by tests/features/remote_control_set_title/ — 8 invariants
including the dual signal + timer guards, the tab-close cleanup,
and the empty-title-restores-via-shellTitle behaviour.
Remote-control — select-window command. Fourth rc_protocol
command. Switches the active tab to a given 0-based index.
{"cmd":"select-window","tab":<int required>}
-> {"ok":true,"index":<int>}
tab is required — no implicit default-to-active, the caller
always spells out which tab they mean. Validated via
QJsonValue::isDouble() so a string-typed tab field doesn't
silently become tab 0. After the switch, focus is restored to the
new active terminal via QWidget::setFocus() so follow-up
send-text calls without an explicit tab field land on the
expected pane. Client CLI reuses the existing --remote-tab option:
ants-terminal --remote select-window --remote-tab 2
Pinned by tests/features/remote_control_select_window/ — 6
invariants including the required-tab contract, the isDouble
check, and the setFocus-on-switch behaviour.
Remote-control — new-tab command. Third rc_protocol command.
Opens a fresh tab and returns its 0-based index, so callers can chain
into send-text --remote-tab <i> without guessing.
{"cmd":"new-tab","cwd":"<path optional>","command":"<string optional>"}
-> {"ok":true,"index":<int>}
cwd omitted = inherit from focused/current terminal (same default
as the menu-driven newTab() slot). command written to the new
shell after a 200 ms settle via TerminalWidget::writeCommand
(matches onSshConnect timing — shells drop keystrokes written
before their init settles). Caller owns the trailing newline,
matching send-text semantics for one consistent rule across every
rc_protocol write-path. Client CLI adds --remote-cwd <path> and
--remote-command <str>:
ants-terminal --remote new-tab
ants-terminal --remote new-tab --remote-cwd /tmp
ants-terminal --remote new-tab --remote-cwd /tmp --remote-command 'pwd'
Idiomatic chaining:
idx=$(ants-terminal --remote new-tab --remote-cwd /tmp | jq .index)
ants-terminal --remote send-text --remote-tab "$idx" \
--remote-text $'ls -la\n'
Pinned by tests/features/remote_control_new_tab/ — 7 invariants
including the 200 ms settle, the QPointer<TerminalWidget> guard
on the deferred lambda (tab-close between enqueue and fire), and
the non-const int newTabForRemote(...) signature the dispatch
layer depends on.
Remote-control — send-text command. Second rc_protocol command
on top of the scaffold shipped in the previous Unreleased entry.
Writes a UTF-8 string byte-for-byte to a tab's PTY master — control
chars and escape sequences pass through unchanged, matching
Kitty's rc_protocol send-text.
Request shape: {"cmd":"send-text","tab":<int optional>,"text":"<string>"};
tab omitted targets the active tab. Response on success:
{"ok":true,"bytes":<int>}. Client CLI adds --remote-tab <i> and
--remote-text "<string>"; when --remote-text is absent, the
client reads stdin until EOF — enables shell piping without
inline-arg quoting pain:
ants-terminal --remote send-text --remote-text 'echo hi\n'
ants-terminal --remote send-text --remote-tab 0 --remote-text ''
printf 'ls\n' | ants-terminal --remote send-text
Does not auto-append a newline (matches Kitty; auto-newline would
surprise binary-stream callers). tab is disambiguated via JSON
isDouble() rather than toInt() == 0 so --remote-tab 0 stays
meaningful. Locked by new source-grep test
tests/features/remote_control_send_text/ (7 invariants including
the no-auto-newline negative guard and the stdin-read ergonomic).
Remote-control protocol — first slice (src/remotecontrol.{h,cpp}).
Kitty-style JSON-over-Unix-socket RPC. A QLocalServer listens on
$ANTS_REMOTE_SOCKET or, by default, $XDG_RUNTIME_DIR/ants-terminal.sock
(falling back to /tmp/ants-terminal-<uid>.sock when XDG runtime is
unset). Each connection handles one request / one response, LF-terminated.
The only command in this slice is ls, which returns
{"ok": true, "tabs": [{"index", "title", "cwd", "active"}, ...]}.
Unknown commands return {"ok": false, "error": ...} with client exit
code 2. The same binary doubles as the client via
--remote <cmd> (--remote-socket <path> overrides the socket).
Clean-foundation release — the remaining commands (send-text,
set-title, new-tab, select-window, get-text, launch) and the
X25519 auth layer land in follow-up commits against this scaffold.
Locked against future drift by tests/features/remote_control_ls/
(8 source-grep invariants including field-name stability and the
--remote-before-MainWindow-construction ordering).
Main-thread stall detector (Debug Mode → Perf category). A 200 ms
heartbeat QTimer in MainWindow measures the gap between its own
firings; when the wall-clock gap exceeds the scheduled interval by
more than 100 ms the event loop was blocked and the log records
the drift, cumulative count, and worst-case observed so far. Gated
by ANTS_DEBUG=perf (or all) so the detector is only armed when
debugging — zero overhead when off, zero log output when no stalls
are happening. Addresses the 2026-04-20 follow-up user report —
"slow down experienced at various times, when tab has been clear
or with lots of text, not one specific scenario." Intermittent
slowdowns that span both empty and full tabs point at a periodic
background handler rather than the PTY hot path, and this detector
will fingerprint the exact blockage when the user next reproduces
the slowdown.
VT throughput benchmark harness at tests/perf/bench_vt_throughput.cpp.
Drives four fixed corpora (ascii_print, newline_stream, ansi_sgr,
utf8_cjk) through VtParser → TerminalGrid::processAction at -O2,
emits CSV per corpus (bytes, actions, wall-ms, MB/s, actions/s). First
step of the 0.8.0 "Terminal throughput slowdowns" perf investigation
(user report 2026-04-20). Registered under ctest label perf so it
runs only on explicit ctest -L perf; default fast suite stays under
two seconds. Baseline on dev laptop: ascii_print 23 MB/s,
newline_stream 5 MB/s (4× slower — scroll/scrollback is the hotspot
to profile next), ansi_sgr 17 MB/s, utf8_cjk 8 MB/s.
QEvent::HoverMove / HoverEnter / HoverLeave in
MainWindow::eventFilter. The 0.7.5 NoAnimStyle fix only partially
reduced the user-reported flicker (mouse-movement-triggered);
Qt's style engine consults WA_Hover tracking — a separate channel
from QMouseEvent — to re-evaluate :hover on every cursor tick.
Without suppressing HoverMove the menubar was still repainting on
every pixel of intra-item mouse motion. Cross-item motion
(File → Edit) still passes through so menu switching works.
Residual flicker is still visible (confirmed by the user after
shipping this change) — pushed to the 0.8.0 roadmap's
"Carried over from 0.7.x" section for a different-angle fix.new-tab honours its documented "caller owns newline" contract.
The first ship of new-tab used TerminalWidget::writeCommand for
the deferred command write, which always auto-appends \n —
contradicting the documented send-text-style "caller owns the
trailing newline" semantics. MainWindow::newTabForRemote now
uses sendToPty (raw bytes) instead. The convenience case lives
in the new launch command (which auto-appends explicitly), so
callers who wanted "open + run" can switch with no behaviour
change. Pinned by an INV-5 (neg) rule in
tests/features/remote_control_launch/ against any future
re-introduction of writeCommand in newTabForRemote.Theme: same-day follow-up to 0.7.4 that resolves the dropdown-menu
flicker the 0.7.4 "Known Issue" section called out. Root cause finally
identified via the Debug Mode instrumentation that shipped in 0.7.4:
Fusion's QWidgetAnimator was creating a 60 Hz
QPropertyAnimation(target=QWidget, prop=geometry) cycle on the idle
window (1439 animation completions in a 23 s idle log). Each cycle
drove a full LayoutRequest → UpdateRequest → Paint cascade across
every widget, which the user saw as dropdown flicker on mouse hover.
Neither QApplication::setEffectEnabled(UI_AnimateMenu, false) nor
QMainWindow::setAnimated(false) covers the style-hint-driven cycle
— only overriding QStyle::SH_Widget_Animation_Duration does.
NoAnimStyle
QProxyStyle in src/main.cpp that zeroes
SH_Widget_Animation_Duration and SH_Widget_Animate. Wrapped
around the Fusion style at application startup
(app.setStyle(new NoAnimStyle(QStyleFactory::create("Fusion")))).
Idle-log result: QPropertyAnimation events per second:
0 (was ≈62 Hz in 0.7.4). Total event volume dropped >99 %
(29 096 → 62 log lines in 4 s idle). Pinned by the new
no_anim_style feature test (tests/features/no_anim_style/).QEvent::ChildAdded to QEvent::ChildPolished. At ChildAdded
time the child's vtable still points at QObject (the derived
constructor hasn't run), so child->metaObject()->className()
returns "QObject" and the QPropertyAnimation filter never fires.
ChildPolished fires after full construction. The old hook was
silently useless on the 0.7.4 investigation — this one actually
logs creation sites.Theme: a session's worth of user-facing UI fixes, the introduction
of a comprehensive debug-logging system, and the debt-cleanup that
came out of both. The centerpiece is the new Debug Mode (Tools
→ Debug Mode), which was what let us pin a ~54 Hz full-window
repaint cascade that had been driving the menubar / dropdown /
paint-dialog flicker everyone had been chasing for multiple
releases. Fixes to the Command Palette, Paste dialog, TitleBar,
QMainWindow animator, Qt UI effects, QMenuBar's hover rule, and
several smaller items all landed here.
src/debuglog.{h,cpp} — ~200 lines). A single
ANTS_LOG(category, ...) macro plus 14 independent category
toggles (Paint, Events, Input, PTY, VT, Render, Plugins, Network,
Config, Audit, Claude, Signals, Shell, Session). Categories can
be enabled narrowly (one) or broadly (all) via either
ANTS_DEBUG=<list> at launch (comma-separated names, or all,
or 1 for legacy paint-only) or Tools → Debug Mode →
[checkable submenu] at runtime. Output is timestamped and
written to ~/.local/share/ants-terminal/debug.log (append mode,
one header per process start with PID + active category mask);
stderr mirror only when the env var was used, so 2>file.log
redirection still works. Menu also offers Open Log File (xdg-
opens the log in the system viewer) and Clear Log File.
Hot-path cost when disabled is a single bit-test on a static
quint32. Replaces the ad-hoc ANTS_PAINT_LOG diagnostic hack
that was briefly landed mid-session.TerminalGrid::processAction's OSC 133 A handler. When a shell
emits ESC ] 133;A ST at the start of a new prompt, the terminal
now nudges the cursor to column 0 of the next line first if the
previous output didn't end with a newline (e.g.
cat file.json whose last byte is }). Fixes the long-standing
"prompt glued onto the end of the previous command's output"
UX papercut. Equivalent to zsh's PROMPT_SP shell-side option,
but works for bash / fish / any shell that emits OSC 133 A.
Requires the shell-integration scripts at
packaging/shell-integration/ants-osc133.{bash,zsh} to be
sourced; without OSC 133, the terminal has no way to know a
prompt is starting.QMenuBar::item base stylesheet rule. Two remaining leaks:
(a) the QMenuBar inherits Qt::WA_TranslucentBackground from the
MainWindow, which disables auto-fill for the widget tree, so each
hover repaint briefly rendered over cleared-to-transparent pixels;
(b) Qt's QStyleSheetStyle treats QMenuBar::item:hover as a
distinct pseudo-state from :selected, and the Breeze / Fusion
paths can hover-flash for one frame before selection engages —
on that frame, only :hover styling applied, and :hover was
never defined. Fix sets autoFillBackground(true) +
Qt::WA_StyledBackground on the QMenuBar so the stylesheet
background-color paints reliably, mirrors the :selected
highlight into :hover, and makes setNativeMenuBar(false)
explicit so DE global-menu integrations never hide the menubar
in our frameless window.&Paste Alt-mnemonic
path worked. Root cause under a frameless + translucent main
window on KWin: any dialog that goes through QDialog::exec()
inherits Qt's Qt::ApplicationModal transition, which doesn't
compose well with our QApplication::focusChanged redirect on
the frameless parent — same incompatibility the Review-Changes
dialog hit at 0.6.29. Rewrote pasteToTerminal() to use the
Review-Changes shape exactly: heap-allocated QDialog with
Qt::WA_DeleteOnClose, own QHBoxLayout of QPushButtons
(Cancel default, Paste setAutoDefault(false)), clicked
lambda calls performPaste() and closes. No blocking; caller
returns immediately. Previous attempts via QMessageBox::exec(),
custom QDialog::exec(), and show() + QEventLoop all hit the
same modal-vs-frameless regression.QListWidget inside CommandPalette was running its
internal QAbstractItemView machinery (layout scheduler,
selection animation, view-update timer) from MainWindow
construction onward, even though the palette itself was hidden.
Debug-log measurement showed the timer firing at ~50 Hz,
cascading LayoutRequest → UpdateRequest → full widget-tree
paint. Fix: lazy-create the list in CommandPalette::show()
so the view doesn't exist until the user first presses
Ctrl+Shift+P. Zero QAbstractItemView timers at idle.QWidgetAnimator was enabled by
default (setAnimated(true)). It exists to animate dock-widget
rearrangements; Ants has no dock widgets, but the animator
still ran at idle. setAnimated(false) kills the continuous
QPropertyAnimation(target=QWidget, prop=geometry) cycle that
was cascading a LayoutRequest through the whole widget tree on
every frame.Qt::UI_AnimateMenu /
UI_FadeMenu / UI_AnimateCombo / UI_AnimateTooltip /
UI_FadeTooltip / UI_AnimateToolBox) disabled globally at
startup via QApplication::setEffectEnabled(…, false). Menu
show/hide, combobox dropdown, tooltip fade, and toolbox expand
all went through QPropertyAnimation(geometry) at 60 Hz — each
animation frame triggered a LayoutRequest cascade. We're a
precision terminal, not a presentation app; these animations
were pure cost with no user benefit.QTipLabel show/hide cycle on every frame while the cursor was
near a titlebar button, animating the tooltip's geometry at
~60 Hz, cascading LayoutRequest through the whole widget tree.
Diagnosed via ANTS_DEBUG=paint,events — the log pinpointed
cls=QTipLabel name=qtooltip_label parent=QToolButton:closeBtn
as the repaint source. The button glyphs (✥ / ✕ / ⬜ / ⟩) + the
hover-colour feedback give enough affordance without tooltip
text.TerminalWidget switched from QOpenGLWidget to plain
QWidget. The GL widget composites its FBO through the
parent's backing store and the composition path always
invalidates the top-level window — we chased this at length via
PartialUpdate, WA_OpaquePaintEvent, swapInterval(0),
setAlphaBufferSize opt-out, etc. None fixed it. Switching the
base class eliminates the GL composition path entirely. The
default QPainter path — already ligature-aware via QTextLayout
/ HarfBuzz — is unaffected. The optional GL glyph-atlas
renderer (gpu_rendering config) is dormant in 0.7.4 and
returns in a future refactor as an embedded QOpenGLWindow via
createWindowContainer (the only shape that cleanly decouples
GL composition from the parent's backing store).tests/features/menubar_hover_stylesheet/ — pins the
QMenuBar::item:hover rule, autoFillBackground(true),
Qt::WA_StyledBackground, the QMenuBar::item base rule, and
setNativeMenuBar(false) for the menubar hover no-flash contract.tests/features/paste_dialog_custom/ — pins the async
pasteToTerminal() pattern: new QDialog(this) on the heap,
Qt::WA_DeleteOnClose, QDialogButtonBox::Ok+Cancel, Cancel
setDefault(true), Paste setAutoDefault(false), the
show() + raise() + activateWindow() activation dance, explicit
setFocus on a button, QPointer<TerminalWidget> tab-close
guard, performPaste() helper, no QEventLoop, no
QDialog::exec(). Forbids the old synchronous
bool confirmDangerousPaste(...) API from reappearing.tests/features/terminal_partial_update_mode/ — pins
TerminalWidget : public QWidget (not QOpenGLWidget) and the
absence of QOpenGLWidget::* / makeCurrent calls in the .cpp.A residual dropdown-flicker cycle driven by a
QPropertyAnimation(target=QWidget, prop=geometry) still runs in
some idle states on KWin + Qt 6. The per-iteration fixes listed
above resolved identifiable sources (CommandPalette QListWidget,
TitleBar tooltip, QMainWindow animator, Qt UI effects); what
remains is a Qt-internal animation we have not yet pinpointed. The
new Debug Mode (Tools → Debug Mode) is the instrument to hunt it
with — enable paint + events and grep the log for
QPropertyAnimation CREATED lines.
Theme: H6.1 of the 📦 distribution-adoption plan — Lua plugins
inside Flatpak. The 0.7.2 manifest shipped with plugin support
disabled because org.kde.Sdk//6.7 does not carry lua54-devel
and flathub/shared-modules has no Lua 5.4 entry today. 0.7.3
closes that gap with an in-manifest Lua archive module so a
Flatpak build loads plugins the same way the openSUSE, Arch, and
Debian builds do.
packaging/flatpak/org.ants.Terminal.yml). A dedicated lua
module appears before ants-terminal so the CMake configure
step sees /app/include/lua.h + /app/lib/liblua.a and takes
the LUA_FOUND=TRUE branch. Built from
https://www.lua.org/ftp/lua-5.4.7.tar.gz with a pinned
sha256 and the linux-noreadline target — the terminal only
statically links liblua.a, so pulling readline into the
sandbox would be pure bloat. MYCFLAGS="-fPIC" keeps the
static library PIE-safe for linking into the ants-terminal
executable. Installed to /app via
make install INSTALL_TOP=/app, which is exactly where
CMake's FindLua searches by default when
CMAKE_INSTALL_PREFIX=/app.flatpak-external-data-checker. The Lua module carries an
x-checker-data: stanza pointing at
https://www.lua.org/ftp/ with
version-pattern: lua-(5\.4\.\d+)\.tar\.gz and
url-template: https://www.lua.org/ftp/lua-$version.tar.gz.
Flathub CI opens a PR against the Flathub repo with a
refreshed url + sha256 on each Lua 5.4.x point release —
no manual hash churn per bump, and 5.5.x majors are correctly
excluded (they would break the in-source find_package(Lua 5.4)
floor).tests/features/flatpak_lua_module/. Source-grep
regression against the manifest YAML pinning six invariants:
Lua module precedes ants-terminal (order matters —
flatpak-builder evaluates modules in sequence, and CMake's
FindLua runs during ants-terminal configure), type: archive
with a pinned sha256:, MYCFLAGS="-fPIC" in the build
commands, make install INSTALL_TOP=/app, the linux-noreadline
build target (guarding against the default make linux that
aliases to linux-readline), and the x-checker-data stanza
with the 5.4.x version-pattern. A regression that strips any
of these fails at ctest time.packaging/flatpak/README.md: the "Lua plugins — limitation"
section is replaced with a "Lua plugins" section documenting the
enabled path (archive-module build, linux-noreadline rationale,
x-checker-data automation, success criterion of PluginManager
loading a plugin inside the sandbox).ROADMAP.md H6.1 flipped from 📋 to ✅ (shipped in 0.7.3).Theme: H6 of the 📦 distribution-adoption plan — Flatpak packaging.
Ships the org.ants.Terminal.yml manifest and the PTY wiring Flatpak
needs so a second tab's shell actually sees the host's $PATH and
filesystem. One artifact that runs on every distro unblocks the
"packaged everywhere" gating item without per-distro maintenance.
packaging/flatpak/org.ants.Terminal.yml).
Targets org.kde.Platform//6.7 — the KDE SDK brings cmake, ninja,
and every Qt6 component the build needs (Core, Gui, Widgets,
Network, OpenGL, OpenGLWidgets, DBus) so the manifest's modules
block is a single buildsystem: cmake-ninja entry. finish-args
cover Wayland + fallback-X11 + DRI, --share=network (AI endpointorg.freedesktop.portal.* covers
global shortcuts + file dialogs), desktop notifications, and
xdg-config/ants-terminal:create + xdg-data/ants-terminal:create
so config and data land under standard XDG paths. The manifest is
ready to re-point from type: dir / path: ../.. to
type: git / url / tag for Flathub submission — local flatpak- builder --install --user builds end-to-end against the in-tree
source today. See packaging/flatpak/README.md.ptyhandler.cpp. The forked
child now probes FLATPAK_ID (set by the flatpak launcher) and
/.flatpak-info (present in every sandbox regardless of launch
path); either signal triggers a branch that exec's the user's
shell via flatpak-spawn --host instead of direct execlp. The
sandbox doesn't inherit env or cwd across the host boundary, so
TERM, COLORTERM, TERM_PROGRAM, TERM_PROGRAM_VERSION (pulled
from the ANTS_VERSION compile definition, not a literal), and
COLORFGBG are forwarded explicitly via --env=KEY=VALUE; the
requested working directory crosses via --directory=<workDir>
gated on !workDir.isEmpty(). _exit(127) after execvp
matches the direct-exec fallback's failure shape for a missing
shell. This is the same PTY model Ghostty's Flathub build uses —
the terminal emulator stays sandboxed (VT parser, renderer,
plugin VM all inside the confined runtime); only the child shell
escapes so the user's $PATH, $HOME, and tooling are
reachable. Outside Flatpak the existing execlp(shellCStr, argv0, nullptr) path is hit byte-for-byte unchanged. Feature test
tests/features/flatpak_host_shell/ locks the branch shape:
detection probes both signals in an OR, --host and --
separators appear in argv, every TERM var is present as
--env=..., workDir gating is correct, the direct-exec call site
is preserved, _exit(127) is the post-exec fall-through.packaging/flatpak/README.md. Local build instructions
(flatpak-builder --install --user), host-shell rationale,
Flathub submission workflow (the tag-based manifest body differs
from the in-tree type: dir shape), and the documented Lua-
plugins-disabled limitation for the initial manifest — plugin
support returns via a shared-modules Lua entry once the tarball-
sha256 refresh cadence is wired up.packaging/README.md Flatpak section. Layout tree adds
flatpak/; new "Flatpak — org.ants.Terminal.yml" section
mirrors the openSUSE / Arch / Debian sections; shared-concerns
header updated from "three recipes" to "four recipes".ROADMAP.md H6 flipped from 📋 to ✅ (manifest shipped; Flathub
submission is the residual 📋); distribution-adoption table
entry updated to reflect the split shipped/remaining state.Theme: first of the 0.8.0 🎨 multiplexing items — SSH ControlMaster
auto-integration from the bookmark dialog. A second tab to the same
host now piggybacks on the first tab's session, skipping the full
auth handshake. Matches the precedent set by kitty's ssh kitten,
Warp's SSH blocks, and iTerm2's SSH profiles.
-o ControlMaster=auto,
-o ControlPath=$HOME/.ssh/cm-%r@%h:%p, and
-o ControlPersist=10m when the new ssh_control_master config
key is true (default). Opt-out via
~/.config/ants-terminal/config.json for sites that forbid
lingering multiplex sockets by policy. $HOME is resolved in
C++ via QDir::homePath() so the ControlPath works under dash /
POSIX sh (neither does tilde expansion on command-arg
foo=~/…); %r@%h:%p are OpenSSH ControlPath substitution
tokens, not shell metacharacters, so the existing shellQuote()
helper leaves them intact. The three -o options precede the
[user@]host destination for parity with OpenSSH's own
documentation examples. Feature test
tests/features/ssh_control_master/ pins six invariants:
default + explicit-false produce zero Control* tokens,
explicit-true emits all three, $HOME resolution replaces any
literal tilde, %r@%h:%p survives shell quoting, the flags
precede the destination arg, and they coexist with legacy
-p/-i/extraArgs without interference. See
ROADMAP.md §0.8.0
— the item is carried forward from the 0.7.0 "SSH
ControlMaster" roadmap bullet and lands ahead of the larger
mux-server / remote-control work.Theme: shell integration + triggers. Consolidates the 0.6.1 →
0.6.45 arc into the release the 0.7 ROADMAP plotted out: command
blocks as first-class UI, .cast recording, semantic history,
HMAC-verified OSC 133 markers, the full iTerm2-parity trigger set,
SIMD VT parsing, a dedicated PTY read + VT parse worker thread,
Wayland-native Quake mode (layer-shell + GlobalShortcuts portal),
and the H1–H4 distribution slice (SECURITY.md, AppStream metainfo,
man page, shell completions) that takes Ants from "side project" to
"distro-ready." 0.7.0 is the minor-release rollup of 30 patch
releases; the individual per-feature entries below 0.7.0 remain
authoritative for the SHAs they shipped under.
.cast). Shipped across
0.6.x; consolidated in §0.6.10..cast v2 recording
(spec).
Full-session via Ctrl+Shift+R; per-block export via the
command-block context menu. §0.6.10.path:line:col capture in
scrollback opens the file at the cited position. CWD resolution
via /proc/<pid>/cwd, so relative paths Just Work without shell
cooperation. Editor support broadened beyond VS Code + Kate to
VS Code family (code-insiders, codium), vi-family (nvim,
vim), nano, Sublime / Helix / Micro, and JetBrains IDEs.
§0.6.12.$ANTS_OSC133_KEY is set, every OSC 133 marker must carry an
ahmac= param computed as HMAC-SHA256(key, <marker>|<promptId> [|<exitCode>]). Forged markers are dropped and a status-bar
counter surfaces the count with a 5-second cooldown. Bash + zsh
integration scripts under packaging/shell-integration/. §0.6.31.instant flag:
bell, inject, run_script, notify, sound, command
shipped in §0.6.9; highlight_line, highlight_text, and
make_hyperlink shipped in §0.6.13 via a new TerminalGrid
line-completion callback so matches map to exact column ranges
on a real row before the row scrolls into scrollback.user_var_changed event. §0.6.9.
Disambiguated from inline images by the byte after 1337;.
NAME ≤ 128 chars; decoded value capped at 4 KiB.ants.palette.register({title, action, hotkey}). Always-on (no
permission gate); optional global QShortcut. §0.6.9.command_finished (exit + duration),
pane_focused, theme_changed, window_config_reloaded,
user_var_changed, palette_action. §0.6.9.Print actions for the
safe run. A signed-compare trick (XOR 0x80 → two cmpgt_epi8
against pre-computed bounds) flags any interesting byte with a
single movemask. Regression guard:
tests/features/vtparser_simd_scan/ asserts byte-identical
action streams across whole-buffer, byte-by-byte, and
pseudo-random-chunk feeds over a 38-case corpus. §0.6.23.VtStream QThread; parsed VtAction batches
cross to the GUI over Qt::QueuedConnection and apply on the
main thread (paint stays where it was). Back-pressure: at most
8 batches (≈128 KB) in flight before the worker disables its
QSocketNotifier and kernel flow control takes over; GUI
re-enables on drain. Resize goes over
Qt::BlockingQueuedConnection so winsize is current before the
next paint. ANTS_SINGLE_THREADED=1 kill-switch was retired
once the new path baked out. §0.6.34 / §0.6.37.cells.resize() with default-attr
padding or trailing-blank trim, skipping the allocation-heavy
joinLogical / rewrap round-trip. Multi-line soft-wrap
sequences still go through the full logic so correctness is
preserved. §0.6.15.find_package(LayerShellQt CONFIG QUIET) wires LayerShellQt::Interface when the
layer-shell-qt6-devel package is installed;
MainWindow::setupQuakeMode() promotes the window to a
zwlr_layer_surface_v1 at LayerTop, anchored top/left/right
with exclusive-zone 0. XCB path preserved for X11. §0.6.38.GlobalShortcutsPortal client wraps the
org.freedesktop.portal.GlobalShortcuts handshake
(CreateSession → BindShortcuts → Activated) behind a single Qt
signal. Wires the toggle-quake id on KDE Plasma 6 and
xdg-desktop-portal-hyprland / -wlr; the in-app QShortcut
fallback stays active on GNOME Shell. 500 ms debounce prevents
focused double-fire. §0.6.39.config.plugin_grants[<name>] and take effect at
next plugin reload. §0.6.11.TerminalGrid::ImageBudget tracks
total decoded image bytes across the inline-display vector +
the Kitty cache; cap is 256 MB per terminal. Sixel rejects
up front from declared raster size; Kitty PNG / iTerm2 OSC 1337
reject post-decode. Inline red error text surfaces the
rejection. The per-image MAX_IMAGE_DIM = 4096 dimension cap
remains in place. §0.6.11.SECURITY.md + CODE_OF_CONDUCT.md.
Coordinated-disclosure policy with supported-versions table,
reporting channel (GitHub Security Advisory + encrypted email),
disclosure timeline, severity rubric, in/out of scope lists.
Contributor Covenant 2.1 verbatim with maintainer email + the
private GitHub Security Advisory listed as conduct reporting
channels. Clears the Debian / Fedora / Ubuntu security-team
review gate. §0.6.16.packaging/linux/org.ants.Terminal.metainfo.xml (AppStream 1.0
with summary / description / releases / categories / keywords /
OARS content rating / supports / provides / launchable) and
packaging/linux/org.ants.Terminal.desktop (reverse-DNS id,
tightened Keywords, StartupWMClass, Desktop Actions for
NewWindow + QuakeMode). CMake install rules via
GNUInstallDirs cover desktop / metainfo / six hicolor icons;
CI runs appstreamcli validate --explain +
desktop-file-validate on every push. §0.6.17.packaging/linux/ants-terminal.1 in
groff -man covering synopsis, every CLI flag, environment
variables, files, exit status, bugs, authors, and see-also.
CMake installs to ${CMAKE_INSTALL_MANDIR}/man1/; CI lints the
source with groff -man -Tutf8 -wall. §0.6.18.bash-completion/completions/, zsh/site-functions/,
fish/vendor_completions.d/); all three are auto-discovered on
system-wide installs. CI lints each file with the matching
shell's parse-only flag. Closes the H1–H4 distribution slice;
remaining packaging work (H5 distro recipes landed in §0.6.20;
H6 Flatpak, H7 docs site, H13 distro outreach) lives in 0.8.0.
§0.6.19.unbounded_callback_payloads, §0.6.8).QNetworkReply::connect without context object
(qnetworkreply_no_abort, §0.6.8).catch(...) (silent_catch,
§0.6.7).addGrepCheck
(audit_fixture_coverage, §0.6.6), CI-enforced via
tests/audit_self_test.sh.missing_build_flags, §0.6.7).no_ci, §0.6.7).CONTRIBUTING.md — derived from STANDARDS.md; covers
build modes, test layout, adding an audit rule, version-bump
checklist. §0.6.7.actions/checkout v4.2.2 → v5.0.1. Runs on Node 24,
pre-empting GitHub's 2026-06-02 Node 20 deprecation. Both CI
pin sites SHA-pinned with # v5.0.1 humans-readable comment.ants.* schema break.Theme: Deferred VT-standards items from the 10th-audit research report, plus atomic-write hardening for long-lived audit state. 11th audit (this release) returned 0/~103 actionable — baseline remains clean.
termenv-based Go TUIs) probe the
terminal for default fg / bg / cursor colour with \e]10;?\e\\ etc.
to auto-detect dark vs light themes without relying on the
unreliable COLORFGBG env var. Response is the xterm 16-bit-
per-channel form \e]<n>;rgb:RRRR/GGGG/BBBB\e\\. Only the query
form (? payload) is honoured — the set form is deliberately
silently dropped to keep default fg/bg theme-driven and prevent
in-terminal theme injection.$q requests for
DECSTBM (scroll-region, r), SGR (current attributes, m),
and DECSCUSR (cursor shape, q) now reply with DCS 1 $ r …
/ DCS 0 $ r … per the xterm spec. tmux, neovim, and kitty's
kitten use these to save-and-restore state around their own
output. Unknown settings reply "invalid" rather than dropping
silently so callers fall back to defaults. Sixel DCS payloads
continue to route to the image handler — the DECRQSS branch
gates on the $q prefix before Sixel's q search.QSaveFile for four long-lived files:audit_rule_quality.json — per-fire/per-suppression history
(written frequently; corruption loses every rule score)..audit_cache/trend.json — cumulative per-run severity totals..audit_cache/baseline.json — baseline fingerprints anchoring
the "new findings only" filter..audit_suppress v1 → v2 migration rewrite (the simple append
path stays non-atomic on purpose — one-line writes; loader
already skips malformed lines).
Torn writes from crash / kill -9 now can't corrupt any of these.
config.json and session snapshots already had a tighter
fsync-then-rename pattern; those were left in place.osc_color_query — locks OSC 10/11/12 response shape + the
query-only gate (set form MUST NOT produce a response, preventing
theme-injection via in-terminal programs).decrqss — locks the DECSTBM / SGR / DECSCUSR replies plus
the invalid-reply bytes and a Sixel-routing regression guard so
the new DECRQSS branch can't eat image traffic.Theme: Project Audit dialog — self-review improvements sourced from the 10th audit run (2026-04-19). That audit returned 0/55 actionable findings, and the nature of the 55 pointed at three bits of friction the dialog was creating for itself: the dialog was auditing its own output, every "is this noise?" verdict needed an individual click, and users had no summary-level readout for "how much of this run was signal?". All three are closed.
renderResults()
now computes an "X% actionable" figure — signal / (signal + noise)
where noise = suppressions + AI-verdict FALSE_POSITIVE — and
prints it below the severity pills with colour-coded context
(green ≥60%, amber 20–59%, red <20%). A 30-day rolling average
pulled from RuleQualityTracker sits next to it as a peer
comparison point ("30d avg X% actionable, n=Y fires"). Makes the
"should I bother scrutinising this run?" question answerable at
a glance instead of requiring the user to count suppressions
against the severity rows.isGeneratedFile(). Three
new built-in skip rules, no project configuration needed:docs/AUTOMATED_AUDIT_REPORT_*.{json,md,html,txt} — the
previous-run artifacts contain SHA-256 dedup keys that gitleaks
consistently flagged as high-entropy secrets (19 of 23 findings
in the last run)..audit_cache/ and audit_rule_quality.json — the dialog's
own baseline + self-learning ledger; auditing them is circular.tests/audit_fixtures/** — scaffold data for the regex-rule
self-test; bad.* files intentionally embed strings that
third-party scanners misread as real findings (gitleaks picked
up 4 of these last run).🧠 Triage visible (N) button sits next to
the confidence-sort toggle in the filter bar. Click it and every
visible, not-yet-triaged finding goes to the configured LLM in
one JSON request (hard cap 20 per batch; above that it slices
and dispatches sequentially). Each batch's response is an array
of {key, verdict, confidence, reasoning} — verdicts are spliced
back onto the right findings by dedup key so a reordering model
can't misalign results. Prompt-injection hardening inherited from
the single-finding path (4-backtick snippet fencing, literal
```` scrub). Confirms the exact count before the POST;
hidden entirely when AI isn't configured. The button's label
re-calculates on every render so it always shows the current
visible-untriaged count.audit_rule_quality.json and
audit_rules.json are read-compatible with prior versions.fires as the denominator rather
than run count, because run count isn't directly recorded —
fires are a good enough proxy and the label makes the
denominator explicit (n=<count> fires).Theme: Two user-reported regressions closed, plus a quality-of-life default flip. The 0.6.42 focus-redirect guards didn't cure the menubar hover flash — the real culprit was a stylesheet gap, not focus churn. And users coming back to the app expected their tabs to be waiting for them; session persistence now ships default-on, matching what every modern terminal (iTerm2, Kitty, WezTerm, Konsole) does.
QMenuBar::item:selected without a
matching base QMenuBar::item rule, so Qt fell back to the native
style for the non-selected state and composited native item drawing
under the :selected overlay. On hover transitions the native and
stylesheet layers raced — visible as the flash. Added an explicit
transparent-background QMenuBar::item base rule (with matching
padding and border-radius) plus a :pressed rule so every menubar
item state is owned entirely by the stylesheet. The 0.6.42 focus
guards remain as defense-in-depth.Config::sessionPersistence() defaulted to false, so users who
never toggled Settings → General → "Save and restore sessions"
lost their tab set on every quit. Flipped the default to true;
users with session_persistence: false explicit in config.json
still get their opt-out honored (only the absent-key path reads the
fallback). Matches iTerm2 / Kitty / WezTerm / Konsole default
behavior. The 5-second uptime floor in saveAllSessions() still
protects against crash-on-launch wiping real state.Config::sessionPersistence() default changed from false to
true. First-run users now get tab restore out of the box.
Only affects users with no session_persistence key in their
config; explicit values are respected unchanged.tests/features/session_persistence_default/ — three-invariant
feature test (spec.md + test_session_persistence_default.cpp) that
locks in the truth table: missing key → true; explicit true → true;
explicit false → false. Linked against src/config.cpp only (no
GUI dependency). Verified against the pre-fix default by temporary
revert — test fails red on (default false), passes green on
(default true). Brings the feature-test count to 25, and the
total ctest count to 26.-Wshadow warning at mainwindow.cpp:3921 — the
Review-Changes finalizer lambda's const Theme &th shadowed the
outer scope's th. Renamed the lambda-local to lth, silencing
the build-time note without functional change.<cstdlib> include from the new session-
persistence test (clangd lint flagged it).Theme: Focus-redirect guard rails. The 0.6.26 auto-return-to-terminal
behavior was firing through two narrow windows where it shouldn't —
hover-across-menubar produced a visibly flashing highlight, and the
paste-confirmation dialog's "Paste" button silently swallowed mouse
clicks (only the &Paste keyboard shortcut worked). Both user-visible
regressions traced back to the same QApplication::focusChanged
handler. New guards short-circuit the redirect when a popup is open,
when the cursor is over a menu/menubar, or when any top-level
QDialog is visible — the last one covers the QMessageBox::exec()
modality-handshake window where activeModalWidget() can briefly be
null.
focusChanged cycles during
menubar hover; now could momentarily park on a chrome widget not
covered by the parent-chain exempt list, which triggered the
auto-return-to-terminal redirect and wiped the menubar highlight.
The focus-redirect lambda now early-returns when
QApplication::activePopupWidget() is non-null (covers QMenu /
QComboBox popup / tooltip) and when
QApplication::widgetAt(QCursor::pos()) has a QMenu or
QMenuBar ancestor (covers the menubar, which is NOT a popup).
Both guards are mirrored in the deferred QTimer::singleShot(0,…)
fire-time block so a menu that opens between queue and fire is
still honored.sudo, | sh, control
characters) triggered the "Confirm paste" QMessageBox, but
clicking the Paste button did nothing — only pressing P (the
&Paste keyboard mnemonic) worked. Root cause:
QMessageBox::exec() enters modal state inside a brief show()
handshake during which QApplication::activeModalWidget() can
return null for a tick; if a focusChanged event fired in that
window, the redirect stole the button's focus mid-click and
cancelled clicked(). The redirect now walks
QApplication::topLevelWidgets() at queue time and early-returns
on any visible QDialog, mirroring the check that already existed
at fire time.tests/features/focus_redirect_menu_guard/ — new source-grep
feature test pinning six invariants on the focusChanged lambda in
src/mainwindow.cpp: the popup and menu-hover guards (queue +
fire time), the visible-dialog walk at queue time, and the five
existing guards that must not regress (activeModalWidget,
QAbstractButton, mouseButtons, QDialog parent-chain,
CommandPalette). Uses the balanced-brace scanner pattern
(linear-time, constant stack) introduced in 0.6.41's
command_mark_gutter test to avoid std::regex blowing the stack
on a 4000+ line mainwindow.cpp.Theme: Quake-mode + OSC 133 UX polish. Surfaces the out-of-focus
portal binding state so users can tell at a glance whether their hotkey
escapes focus, and adds a scrollbar-adjacent tick-mark gutter that
shows OSC 133 command boundaries across the full scrollback — jump
targets for Ctrl+Shift+Up/Down that were previously invisible.
Command-mark gutter (show_command_marks, default on). A 4-px
vertical strip just left of the scrollbar draws one tick per
PromptRegion in the current scrollback. Each tick is color-coded
by last exit status: green (success), red (non-zero exit), gray
(in-progress, OSC 133 D not yet emitted). Ticks map linearly from
PromptRegion.startLine / (scrollbackSize + rows) to y — so a
50,000-line scrollback compresses into the widget height with every
boundary still visible. No-op when promptRegions().empty(), so
users without shell integration see nothing change.
Settings → Terminal toggle. New "Show command markers in
scrollbar gutter" checkbox (default on) under Scrollback, with a
tooltip explaining the color states. Propagates to all live
terminals via applyConfigToTerminal on apply — no restart needed.
Portal-binding status label in Settings → Quake groupbox. A one-line QLabel under the hotkey field reports whether out-of-focus hotkeys are available on this session:
xdg-desktop-portal-* backends that provide the service.Static check based on GlobalShortcutsPortal::isAvailable() — the
D-Bus service registration is a per-session property that doesn't
change while the app runs, so a snapshot at dialog construction is
accurate.
tests/features/command_mark_gutter/ used a [\s\S]*?^\} pattern
with std::regex::multiline to extract paintEvent's body. On a
4500-line terminalwidget.cpp, libstdc++'s backtracking executor
recursed 100,000+ frames into _M_dfs and segfaulted. Replaced
with a linear-time balanced-brace scanner
(extractFunctionBody(src, signature)) — same semantics, constant
stack. Left as a guiding pattern for future source-grep feature
tests against large files.Config::showCommandMarks() persists to show_command_marks with
default true.paintEvent gates gutter drawing on m_showCommandMarks AND
non-empty promptRegions().m_scrollBar->width() subtraction so
ticks never overlap the scrollbar thumb.applyConfigToTerminal(t) propagates the toggle so settings-apply
updates all live terminals.GlobalShortcutsPortal::isAvailable() with the exact strings
"Portal binding active" and "Portal unavailable" so the test can
lock them.Locked by tests/features/command_mark_gutter (CTest labels
features;fast, part of the release gate).
Theme: OSC 133 shell-integration polish. Adds two keyboard-driven
"last completed command" actions that complement the existing
block-at-cursor right-click menu (Copy Command / Copy Output / Re-run
Command / Fold / Share as .cast) with no-selection-needed top-level
equivalents. The iTerm2 ⇧⌘O / WezTerm CopyLastOutput convention,
surfaced through the View menu, the command palette, and configurable
keybindings.
TerminalWidget::copyLastCommandOutput() slot. Walks
promptRegions() backwards for the most recent region with
commandEndMs > 0 && hasOutput, delegates to outputTextAt(idx)
(unbounded), and places the result on the clipboard. Returns the
number of characters copied, or -1 when no completed command exists.
The backwards walk is load-bearing: if the tail region is still
running (user typed a command but it hasn't emitted OSC 133 D yet),
the method skips it and uses the previous completed region — so
Ctrl+Alt+O during sleep 30 copies the previous command's output,
never the partial in-progress one.TerminalWidget::rerunLastCommand() slot. Same backwards-walk
skip-in-progress rule, then delegates to rerunCommandAt(idx) which
writes the command text + \r to the PTY. Returns the re-run region
index or -1. Bailing on in-progress tails prevents splicing text
into whatever the user is typing.Ctrl+Alt+O, config key copy_last_output) and
"Re-run Last Command" (default Ctrl+Alt+R, config key
rerun_last_command) live under the View menu next to the existing
Previous/Next Prompt entries. Both surface in the command palette
automatically (it collects all menu actions). The Ctrl+Alt+*
modifier combo avoids the already-taken Ctrl+Shift+O (split pane)
and Ctrl+Shift+R (record session).lastCommandOutput()The existing lastCommandOutput() helper caps output at 100 lines
because it feeds commandFailed triggers where huge stderr noise is
harmful. A user-initiated copy expects the full output or a clear
"nothing to copy" signal, not a silently truncated string. The new
action uses outputTextAt(idx) (unbounded) instead, and a feature
test enforces this choice.
promptRegions() backwards and gate on
commandEndMs > 0.copyLastCommandOutput must call outputTextAt, never
lastCommandOutput (100-line cap).config.keybinding(...) so
users can rebind them via config.json.Ctrl+Alt+O / Ctrl+Alt+R must appear
exactly once in mainwindow.cpp.showStatusMessage(...)
toast so the operation isn't silent.Locked by tests/features/osc133_last_command (CTest labels
features;fast, part of the release gate).
Theme: Wayland-native Quake-mode, part 2 of 2. Completes the 0.7.0
ROADMAP item: out-of-focus global hotkey registration via the
Freedesktop Portal org.freedesktop.portal.GlobalShortcuts D-Bus API.
Replaces the originally-planned KGlobalAccel-on-KDE + GNOME-dbus
branching with one compositor-agnostic portal API that upstream has
converged on since 2023 (KDE Plasma 6, xdg-desktop-portal-hyprland,
xdg-desktop-portal-wlr).
GlobalShortcutsPortal client class. New
src/globalshortcutsportal.{h,cpp} wraps the three-step handshake
(CreateSession → Response with session_handle → BindShortcuts →
Activated) behind a simple activated(QString) Qt signal. The class
handles the portal's request-path prediction convention (subscribing
to the Response signal before dispatching the method call so the
portal's reply can't land in a window where no match rule is
installed), Qt D-Bus custom meta-type registration for the
a(sa{sv}) shortcuts array, and session lifecycle.MainWindow Quake setup. When
GlobalShortcutsPortal::isAvailable() returns true, MainWindow
instantiates the portal, binds toggle-quake with
config.quake_hotkey as the preferred trigger, and routes Activated
signals through toggleQuakeVisibility(). The in-app QShortcut
from 0.6.38 stays wired unconditionally as a defence-in-depth
fallback — for GNOME Shell (portal service present but no
GlobalShortcuts backend yet), for the window between CreateSession
and BindShortcuts response, and for any future revocation / re-bind
path.QShortcut lambda
and the portal activated lambda stamp m_lastQuakeToggleMs with
QDateTime::currentMSecsSinceEpoch() and reject any toggle if the
previous stamp is less than 500 ms old. Without the debounce, a
focused system where both paths deliver the same key press would
hide-then-show the window in one frame (visible flicker).qtKeySequenceToPortalTrigger()
converts "Ctrl+Shift+\"/"Meta+F12" into the freedesktop shortcut-syntax equivalents ("CTRL+SHIFT+grave", "LOGO+F12"). Deliberately minimal — modifiers plus a handful of common punctuation → xkb-keysyms; full keysym coverage is xkbcommon's job. Unrecognised keys pass through unchanged, and the user adjusts in System Settings if needed (the portal prompt treats preferred_trigger` as advisory anyway).tests/features/global_shortcuts_portal/ pins: (1)
isAvailable() gate on portal construction, (2) canonical
org.freedesktop.portal.Desktop / /org/freedesktop/portal/desktop
/ org.freedesktop.portal.GlobalShortcuts / .Request literals in
the impl, (3) bound id "toggle-quake" matches the Activated
handler filter, (4) in-app QShortcut fallback stays wired, (5) both
paths debounce via m_lastQuakeToggleMs (regex match count ≥ 2), (6)
CMake lists the source + keeps Qt6::DBus linked, (7) header
surfaces the expected public API (isAvailable, bindShortcut,
activated).Theme: Wayland-native Quake-mode, part 1 of 2. The 0.7.0 ROADMAP
item is split into (a) layer-shell anchoring + wiring the
long-dead quake_hotkey config key — shipped here — and (b)
Freedesktop Portal GlobalShortcuts for out-of-focus hotkey
registration, shipping in 0.6.39. After this release, running
Ants Terminal in Quake mode on Wayland no longer depends on the
compositor's goodwill for positioning/stacking.
LayerShellQt integration for Wayland Quake-mode. When the
layer-shell-qt6-devel package is available at build time,
find_package(LayerShellQt CONFIG QUIET) wires
LayerShellQt::Interface in and defines ANTS_WAYLAND_LAYER_SHELL.
At runtime on Wayland, MainWindow::setupQuakeMode() promotes the
QWindow to a zwlr_layer_surface_v1 at LayerTop, anchored
top/left/right with exclusive-zone 0 and
KeyboardInteractivityOnDemand. This is the Wayland equivalent of
X11's _NET_WM_STATE_ABOVE + client-side move() — without it the
compositor could place the dropdown anywhere and stack it below
other surfaces.LayerShellQt not found — Quake-mode on Wayland falls back to the Qt toplevel path and the binary still
runs. When the binary is run on X11, the XCB path (pre-0.6.38
behaviour) is taken — no regression.quake_hotkey config key was saved but never read. The
QLineEdit in Settings → "Dropdown/Quake Mode" persisted the user's
chosen sequence to config.json, but nothing consumed it — the
value was inert. MainWindow now wires a QShortcut with
Qt::ApplicationShortcut context to toggleQuakeVisibility()
when Quake mode is active. The shortcut fires only while Ants has
focus; true out-of-focus global hotkey registration goes through
the Freedesktop Portal GlobalShortcuts API coming in 0.6.39.QPropertyAnimation(this, "pos") path in toggleQuakeVisibility()
is a no-op on Wayland (the compositor owns position) and the
animation's end-value move() would visibly snap. The Wayland
branch now takes a plain show()/hide() toggle; the XCB slide
animation is unchanged.BuildRequires: cmake(LayerShellQt) >= 6.0.layer-shell-qt to makedepends and
listed it under optdepends for runtime discoverability.liblayershellqt6-dev to
Build-Depends. Older Debian releases packaged the devel headers
as liblayershellqtinterface-dev; the CMake QUIET flag lets the
build proceed either way.tests/features/wayland_quake_mode/ — source-grep regression
test pinning the 0.6.38 contract: (1) the <LayerShellQt/Window>
include is guarded by ANTS_WAYLAND_LAYER_SHELL; (2)
setupQuakeMode() does a runtime QGuiApplication::platformName()
dispatch and preserves the X11 flags; (3) toggleQuakeVisibility()
returns before the QPropertyAnimation setup on Wayland; (4) the
constructor wires quake_hotkey to a QShortcut connected to
toggleQuakeVisibility; (5) CMakeLists.txt uses
find_package(LayerShellQt CONFIG QUIET) and links
LayerShellQt::Interface when found.Theme: Phase 3 of the 0.7.0 threading refactor — retire the
ANTS_SINGLE_THREADED=1 kill-switch and delete the legacy
single-threaded read/parse code paths from TerminalWidget. Bake
period (0.6.34 → 0.6.37) proved the worker path; the 0.6.35 hotfix
proved the kill-switch worked. Shipping on the worker is now the only
path.
ANTS_SINGLE_THREADED=1 kill-switch. The env-var fork in
TerminalWidget::startShell() and the legacy branch it guarded
are gone. Every launch now goes through VtStream on
m_parseThread.m_pty / m_parser members and onPtyData slot. The
widget no longer owns a Pty or a VtParser directly; both live
on the worker inside VtStream. hasPty() and ptyChildPid()
simplified to worker-only checks; no behaviour change at the
call sites.#include "vtparser.h" and #include "ptyhandler.h" from
terminalwidget.h. Transitively pulled in via vtstream.h for
callers that still need VtAction / Pty types.Theme: Tenth periodic audit (10th in the series, post-0.6.35). Two actionable findings across ~110 raw — a 2% signal ratio consistent with the codebase's audit-calibration anchor. Both fixed in a single pass.
RuleQualityTracker::prune() dead code. Declared in
src/auditrulequality.h:133 and defined in .cpp:250; never
called. The retention-cutoff logic was inlined into save()
(const-method) when that method needed to filter records without
mutating. Two copies of the same logic would drift over time.
Removed the dead function — save() is the single source of
truth for the 90-day retention window. Flagged by cppcheck's
unusedPrivateFunction rule.pipeRun regex recompiled on every paste.
src/terminalwidget.cpp:2129
constructed a fresh QRegularExpression with the
CaseInsensitiveOption on every call to confirmDangerousPaste()
— a hot path that fires on every user paste. Hoisted to a
function-local static const so the pattern compiles once per
process and is reused thereafter. Flagged by clazy's
use-static-qregularexpression check. Safe because
QRegularExpression::match is reentrant and const-safe since
Qt 5.4.Theme: Hotfix for the 0.6.34 threaded-parse refactor. Keystrokes typed into the terminal were silently dropped — the echo never appeared on screen. Shipped 0.6.34, caught in minutes of dogfood, fix is small and the regression test makes it impossible to come back quietly.
TerminalWidget (keystrokes, Ctrl modifiers, arrow
keys, clipboard paste, focus reporting, mouse middle-click paste,
input method commit, scratchpad, context-menu paste and clear,
writeCommand, rerunCommandAt) still carried pre-0.6.34
if (m_pty) ptyWrite(...) or if (cond && m_pty) { ptyWrite(...) }
guards. Under the threaded parse path m_pty is null (the Pty
lives on the worker thread inside VtStream), so the guard
silently swallowed every write. Replaced each guard with either
hasPty() (path-agnostic boolean — true for either the worker or
the legacy path) or an unconditional ptyWrite(...) (the helper
is already null-safe). The legacy path's m_pty lifecycle code
in startShell / recalcGridSize / ~TerminalWidget is
unchanged.shellPid() / shellCwd() / foregroundProcess() were also
gated on the bare m_pty pointer, returning empty on the worker
path. Replaced with a new ptyChildPid() helper that reads from
whichever path is active (VtStream::childPid() on the worker;
Pty::childPid() on the legacy path). The PID is written once by
forkpty() during startShell's blocking-queued invocation, so
the cross-thread read is synchronised and mutex-free.tests/features/threaded_ptywrite_gating/ regression test.
Source-grep inspection: no if (m_pty) / && m_pty) / !m_pty
gate may appear outside the four allowlisted functions
(ptyWrite, ptyChildPid, startShell, recalcGridSize,
~TerminalWidget). Ground-truth verified — re-injecting the
regression makes the test fail at the offending line; restoring
the fix makes it pass.Theme: Decouple PTY read + VT parse from the GUI thread. Ships the first planned 0.7.0 performance item early since the architectural change landed clean with three feature-conformance tests locking the invariants.
Previously, every byte of PTY output — from a fast prompt echo to a
find / firehose — traversed the Qt GUI thread through
QSocketNotifier → Pty::dataReceived → VtParser::feed → TerminalGrid::processAction and then on to paint. Heavy streams
blocked scroll and repaint. As of 0.6.34, read + parse run on a
dedicated worker QThread (VtStream); only batched VtAction
streams cross back to GUI for grid application and paint.
VtStream). New worker-thread
wrapper around Pty + VtParser (src/vtstream.{h,cpp}). The
worker reads the PTY master via its own QSocketNotifier, feeds
bytes through VtParser, accumulates emitted VtActions into a
VtBatch, and ships the batch to the GUI over
Qt::QueuedConnection. GUI applies the batch to TerminalGrid and
repaints — the grid and paint path do not move.VtStream::drainAck(). Replaces "let RAM grow
unbounded while GUI catches up" with kernel-natural
back-pressure.Pty::setReadEnabled(bool) — hook to pause/resume the read
notifier without tearing it down. Used by the back-pressure path.
Pty::write and Pty::resize moved into public slots: so they
can be invoked cross-thread via QMetaObject::invokeMethod.tests/features/:threaded_parse_equivalence — 11 input fixtures (plain ASCII,
mixed ANSI, DCS Sixel, APC Kitty, OSC 52 long payload, UTF-8
multibyte across chunk boundaries, DEC 2026 sync, OSC 133
markers, CSI with 32 params, UTF-8 bulk, 50 KB Print run) × 6
chunking strategies (whole, byte-by-byte, 16 KB worker flush,
7-byte, 3-byte, two pseudo-random seeds). Asserts action-stream
identity across strategies.threaded_response_ordering — DA1 / CPR / DA2 / mode-996 /
DSR-OK sequence asserted to fire in parse order across three
chunking strategies. End-to-end write order across the worker
boundary follows from this plus Qt's per-receiver FIFO.threaded_resize_synchronous — source-level contract that
every invokeMethod(m_vtStream, "resize", ...) uses
Qt::BlockingQueuedConnection, and that VtStream::resize /
::write are declared in a public slots: block.TerminalWidget PTY-write path unified through ptyWrite().
~30 call sites (keystrokes, bracketed paste, Kitty kbd protocol,
response callback, snippets, scratchpad) now route through a single
helper that branches on m_vtStream ? invokeMethod(QueuedConnection) : m_pty->write(). One-line diff per call site; correctness is
centralised.Qt::BlockingQueuedConnection.
TerminalWidget::recalcGridSize waits for the worker to finish
Pty::resize before returning, so the next paint always sees
consistent ws_row/ws_col vs. m_grid->rows()/m_grid->cols().
No 1-frame flicker at the new size with old PTY dims.elapsed field). Log write granularity is now per-batch
(coarser than per-PTY-read); users grepping the file minutes later
see no difference.Kill switch: set ANTS_SINGLE_THREADED=1 in the environment to force
the legacy single-threaded path while the new code proves itself. It
will be removed in a follow-up release once no regressions surface.
Theme: Nine long-standing user-reported bugs around the status bar,
scrollback integrity, session restore, and the Review Changes dialog —
fixed in one pass. The quiet centerpiece is the root-cause diagnosis of
the long tail of "Review Changes button does nothing" reports: a
0.6.26-era focus-redirect lambda was firing a singleShot(0) refocus
between a button's mousePress and mouseRelease, ripping focus off the
button mid-click and silently cancelling the clicked() signal that
QAbstractButton only emits when focus is continuous through release.
Also retires the ElidedLabel + stylesheet chip padding class of
status-bar bug (branch chip truncating to "…") by converting every
fixed-width status-bar chip to plain QLabel / QPushButton with
QSizePolicy::Fixed.
~/.config/ants-terminal/hooks/claude-forward.sh and merges
PreToolUse / PostToolUse / Stop / PreCompact / SessionStart entries
into ~/.claude/settings.json. Idempotent, preserves the user's
existing hooks, safe to re-run.cellAtGlobal / combiningAt now read from a snapshot of the screen
taken at scroll-up time rather than the live grid. The viewport is
completely immune to live screen writes, not just to deep-scroll
rows. Locked by new tests/features/scrollback_frozen_view/./audit are detected in the
transcript parser; new ClaudeIntegration::planModeChanged and
auditingChanged signals.refreshStatusBarForActiveTab() — single per-tab refresh entry
point covering branch, notification, process, Claude status, Review
Changes, and Add-to-allowlist. Replaces the scatter of direct calls
in onTabChanged where one miss meant a stale chip for 2 s.notification_timeout_ms config key (default 5000) with a
seconds spinner in Settings > General.tab_color_sequence config key — ordered colour fallback
independent of session persistence. UUID-keyed tab_groups is
retained for within-session drag-reorder stability.QSizePolicy::Fixed with
plain QLabel / QPushButton. Only the transient notification
slot is elastic and elides with "…". The
ElidedLabel + stylesheet chip padding miscalculation class of bug
is retired.git status --porcelain=v1 -b so ahead-of-upstream counts. The global
QPushButton:hover rule is now gated with :enabled so disabled
buttons no longer light up on hover. The dialog renders Status +
Unpushed + Diff sections; spec and feature test updated to match.WindowModal was tried and reverted — non-modal with button-disable
is the pattern that actually works.toolFinished / sessionStopped retraction path
across scroll-scan and hook detections, and duplicate-rule feedback
on prefill.focusChanged lambda queued a singleShot(0) to refocus
the terminal the moment focus touched status-bar chrome. That
zero-delay timer could fire between mousePress and
mouseRelease on a QPushButton, ripping focus away mid-click;
QAbstractButton requires continuous focus through release to emit
clicked(), so the click was silently dropped. Fix: the lambda
now exempts QAbstractButton focus targets and defers while any
mouse button is held down.ElidedLabel's
sizeHint fudge could not account for the chip stylesheet's
padding + margin. Converted to a plain QLabel with Fixed
sizePolicy.refreshStatusBarForActiveTab() wiring ensures state never bleeds
from a previous tab on onTabChanged.recalcGridSize ran on
inactive tabs before their geometry was laid out, shrinking grids
to 3×10 — fixed with a pre-layout guard; (b) TerminalGrid::resize
pushed every reflowed row into scrollback indiscriminately — now
skips entirely-blank rows.tab_groups only worked when session persistence was enabled,
because only the session-persist path re-used saved UUIDs. Added
an ordered tab_color_sequence fallback applied at startup once
tabs are constructed.Theme: Audit self-learning layer + contributor-facing docs + a
one-line CI hotfix for 0.6.31's OSC 133 verifier. The headline is a
per-rule fire/suppression tracker with an inline LCS-based
tightening suggester — the lightweight always-on counterpart to the
heavier weekly cloud routine documented in RECOMMENDED_ROUTINES.md.
Without instrumentation the 2026-04-16 one-shot FP-rate triage would
silently re-accrue as the codebase evolves; now each suppression is
recorded and the noisiest rules surface on demand in the audit dialog.
RuleQualityTracker (src/auditrulequality.{h,cpp}) records every
per-rule fire and user-suppression event from the project Audit
dialog into <project>/audit_rule_quality.json. The JSON is bounded
to a 90-day rolling window, capped at 50 000 records per category.
Tracker writes back on dialog destruction (RAII finalisation), and
force-flushes immediately on each suppression so user actions are
durable even on a hard exit.dropIfContains tightening when the heuristic suggester
has enough samples. Sorted by FP rate desc so the noisiest rules
surface first; rows with FP ≥ 50 % are highlighted.💡 <substring> looks like a common FP shape — consider adding it to <rule>'s dropIfContains in audit_rules.json). Pure-identifier substrings
are rejected so the suggester proposes rule-shape filters, not
project-noun filters.tests/features/audit_rule_quality/ covers fire +
suppression aggregation, the LCS suggester's positive and negative
paths (too-few-samples, pure-identifier rejection), and JSON
persistence round-trip. Headless model-only test, no Qt GUI.docs/RECOMMENDED_ROUTINES.md captures the four Claude Code
routines that earn a slot on the shared account quota for this repo:
nightly audit triage (diff-against-baseline, PR only NEW findings),
PR security review on main, weekly audit-rule self-triage
(cross-rule LLM analysis of audit_rule_quality.json, opens
tightening PRs — the deeper counterpart to the in-process tracker
above), and a monthly ROADMAP-item scoper. Each entry includes the
trigger config, environment setup, and the exact prompt text. Total
budget ~3-4 runs/day, leaving room for ad-hoc work and the Vestige
repo's own routines.QMessageAuthenticationCode::addData
in older Qt 6.x does not accept QByteArrayView; the 0.6.31 OSC 133
verifier used the view-taking overload and broke the Ubuntu job.
Switched to the (const char*, qsizetype) overload present in every
Qt 6.x version. Behaviour unchanged; the HMAC feature test still
passes. Purely a build-compat fix — users on 0.6.31 are unaffected
if they're already building locally.Theme: Security + audit signal-to-noise + two regressions. The headline is OSC 133 HMAC-signed shell integration (a 0.7.0 ROADMAP item shipped early): protects the command-block UI from forged shell- integration markers emitted by untrusted in-terminal processes. Also cuts the audit-tool false-positive rate from ~97 % to under 10 % via a two-pass triage of the Ants and Vestige audit runs, fixes the user- reported Review Changes click-does-nothing bug, and locks the 0.6.26 Shift+Enter "tab freezes" regression with a feature test.
OSC 133 HMAC-signed shell integration. When $ANTS_OSC133_KEY
is set in the terminal's environment, every OSC 133 marker
(A/B/C/D) MUST carry a matching
ahmac=<hex-sha256-hmac> parameter computed as
HMAC-SHA256(key, "<marker>|<promptId>[|<exitCode>]"). Markers
without a valid HMAC are dropped (no UI side-effects) and a forgery
counter increments, surfaced in the status bar via a throttled
warning ("⚠ OSC 133 forgery detected") with a 5-second cooldown.
The verifier closes the spoofing surface where any process running
inside the terminal — a malicious TUI, cat malicious.txt, anything
that writes to stdout — could otherwise mint OSC 133 markers and
pollute prompt regions, exit codes, and the command_finished
plugin event.
When $ANTS_OSC133_KEY is not set, the verifier is silent and OSC
133 behaves as in 0.6.30 (legacy permissive). No upgrade penalty
for users who don't opt in.
Shell hooks ship under
packaging/shell-integration/ants-osc133.{bash,zsh} and install to
${datarootdir}/ants-terminal/shell-integration/. The README in
the same directory walks through key generation
(openssl rand -hex 32), ~/.bashrc/~/.zshrc setup, threat
model, and verification steps.
Headless feature test
(tests/features/osc133_hmac_verification/) covers verifier OFF
back-compatibility, verifier ON accepting valid HMACs (including
uppercase-hex), and rejection of missing/wrong/promptId-mismatched/
exit-code-mismatched HMACs. Implements ROADMAP §0.7.0 → 🔒 Security
→ "Shell-side HMAC verification for OSC 133 markers."
Review Changes button silently swallowed clicks on a clean repo.
refreshReviewButton's clean-repo branch left the button
setEnabled(false); show() "as a hint that Claude edited
something." QAbstractButton::mousePressEvent drops clicks on
disabled buttons, so clicked() was never emitted and the 0.6.29
silent-return-with-flash guards inside showDiffViewer never
fired. The global QPushButton:hover rule isn't :enabled-gated
either, so the disabled button still highlighted on hover —
advertising itself as actionable while doing nothing. Net symptom
(user report 2026-04-17): "the Review Changes button is showing
and is active as it has an onmouseover event that highlights the
button. When I click the button though, nothing happens." Fix:
hide() on clean repo. The button reappears via the next
2-second refresh / fileChanged tick when a real diff appears.
New tests/features/review_changes_clickable/ locks the regression
with a source-grep guard against re-introducing the
setEnabled(false); ... show(); shape and a tripwire on the
global hover stylesheet remaining un-:enabled-gated.
Shift+Enter wedged the terminal tab in bracketed-paste mode
(0.6.26 regression). QByteArray("\x1B[200~\n\x1B[201~", 8)
truncated the 13-byte literal to 8 bytes, dropping the closing
[201~ end-paste marker and leaving an orphan ESC that kept
the shell in bracketed-paste mode and ate the next keystroke.
Switched to QByteArrayLiteral(...) so the size is derived from
the literal at compile time — closes the entire class of
"length argument drifts away from literal" bugs. New
tests/features/shift_enter_bracketed_paste/ locks the byte
sequence AND source-grep-guards against the
QByteArray(literal, <num>) shape recurring.
Status-bar allowlist button dedup missed the hook-path
container. The scroll-scan-path cleanup used
findChildren<QPushButton*>("claudeAllowBtn"), but the hook-path
permission handler creates the button as a QWidget container
(with child controls) that has the same objectName. A
QPushButton-typed findChildren skipped the container,
letting both buttons stack when a scroll-scan detection fired
while a hook-path container was already visible. Switched to
QWidget* for parity with the onTabChanged and hook-path
dedup lookups.
Two-pass triage of the Ants audit (docs/AUDIT_TRIAGE_2026-04-16.md)
followed by the Vestige 3D engine triage. 137 of 141 findings in the
Ants run were FPs; 100 % of four addFindCheck rules in the Vestige
run were FPs. Targeted rule tightenings:
memory_patterns regex inverted — only flags new X() /
new X(nullptr) / new X(NULL) (i.e. NOT parented). Any
identifier inside the parens is treated as a Qt parent and
suppressed. Kills 30 FPs from new QWidget(parent) /
new QAction(this) etc. that the previous substring blacklist
could only suppress when the parent expression matched a hardcoded
name.clazy --checks= drops qt-keywords. The project documents
bare signals: / slots: / emit as house style; flagging
them was ~48 FPs per run.clang-tidy auto-disables when no compile_commands.json
exists; collapses 'QString' file not found storms into a single
banner line. Kills 34 near-identical driver-error FPs.cppcheck excludes every build-* variant by name
(build-asan, build-debug, etc.). cppcheck was parsing
moc_*.cpp files in build-asan/ and tripping on their
#error "This file was generated using moc from 6.11.0" banners
— likely the source of the 30-second timeout on Compiler Warnings.qt_openurl_unchecked (±5 lines for startsWith("http" /
QUrl("https" / // ants-audit: scheme-validated),
insecure_http (±5 lines for startsWith scheme gates),
unbounded_callback_payloads (±10 lines for .truncate( /
.left( / .size() <= / constexpr int kMax),
cmd_injection (suppress on , nullptr) / , NULL)
exec-family terminators), bash_c_non_literal (±5 lines for
m_config. / Config::instance Config-trust-boundary references),
debug_leftovers (±8 lines for if (m_debug / if (on) /
#ifdef DEBUG debug-flag conditionals).secrets_scan RHS literal constraint — requires the right-
hand side to be a "…" string of ≥16 chars (or '…', or a
≥16-char unquoted token). Rejects variable-name LHSes with
constructor / variable RHSes (m_aiApiKey = new QLineEdit(tab),
m_apiKey = apiKey).long_files → Info; missing_ compiler_flags → Info. The 2026-04-16 triage flagged these as
"severity leak" — advisories that aren't bugs sharing a tier
with real bugs.__pycache__/, .claude/,
target/, .venv/, venv/, .tox/, .pytest_cache/,
.mypy_cache/. The previous static lists missed common
ecosystem scratch dirs.// ants-audit: scheme-validated inline marker recognized
in the context-window for qt_openurl_unchecked — closes the
last MAJOR-tier FP from the Ants triage where the OSC 8
openHyperlink openUrl was 35 lines below the OSC 8 ingestion
scheme allowlist (different file; ±5 line context can't span).Four small real findings from the prior Ants triage:
claudeallowlist.cpp:409 — std::as_const(segments) on the
value-captured QStringList (clazy range-loop-detach).auditdialog.cpp:3115 — hoisted QStringList parts out of the
finding-render loop (clazy container-inside-loop).elidedlabel.h:33, sshdialog.h:32 — getters return
const & not by value (cppcheck performance:returnByReference).Net measured effect on a clean re-scan: the same audit drops from ~141 to ≲10 findings, with no loss of real coverage.
Four addFindCheck rules produced 100/100 false positives when
ants-audit scanned the Vestige engine on 2026-04-16. Each is a
rule-shape bug rather than a find-site bug; all four are tightened
so the next scan of the same tree drops straight to zero noise for
these categories.
file_perms — World-Writable Files. Dropped the -perm -020
(group-writable) branch. Mode 664 is the default umask result on
most Linux distros, so the check was flagging every normal file
as a security finding. The rule now matches only true
world-writable (-perm -002, CWE-732).
src/auditdialog.cpp:708-720.header_guards — Missing Header Guards. Scan window
widened from head -5 to head -30 (copyright blocks at the
top of a header commonly pushed #pragma once past line 5, so
valid guards were read as missing). The _H-suffix requirement
on traditional #ifndef guards was also dropped — naming
conventions vary (FOO_HPP, FOO_GUARD, __FOO__) and the
suffix was over-fitted to a single house style.
src/auditdialog.cpp:1006-1026.binary_in_repo — Binary Files in Source. Prefer
git ls-files over find . when the project is a git checkout.
find was matching gitignored paths (__pycache__/*.pyc,
.claude/worktrees/…, target/…) because kFindExcl is a
static allowlist and can't keep up with every project's
.gitignore. The scan now falls through to the legacy find
command only when git rev-parse fails.
src/auditdialog.cpp:582-597.dup_files — Duplicate File Detection. Same fix as
binary_in_repo — prefer git ls-files, fall back to find.
md5sum runs only on git-tracked files ≥100 bytes, sidestepping
the Claude-managed-worktree and build-artefact duplication that
made the report useless.
src/auditdialog.cpp:561-579.Measured against the same Vestige tree:
| Rule | Before | After |
|---|---|---|
file_perms |
30 FPs | 0 |
header_guards |
20 FPs | 0 |
binary_in_repo |
30 FPs | 0 |
dup_files |
30 FPs | 0 |
The timeout-as-tool-health and security-rule self-exclusion
fixes cited in the same Vestige triage report landed earlier —
see the 0.6.30 Fixed section (auditdialog.cpp:162-165 for
timeouts, kGrepFileExclSec for self-exclusion).
Theme: Contributor-facing workflow infrastructure — project-level
Claude Code hooks + recipe-driven version bumping. No runtime or
behavior changes for end users. The ants-terminal binary is bit-
identical to 0.6.29 modulo the ANTS_VERSION macro string. The new
files are entirely tooling for contributors who clone the repo and
work on it through Claude Code.
.claude/settings.json wires a PostToolUse hook on
Edit|Write that delegates to
packaging/hook-on-cmakelists-edit.sh. The script short-circuits
unless the touched path ends in CMakeLists.txt, then runs
packaging/check-version-drift.sh and surfaces drift via a
systemMessage. Contributor benefit: the version-drift gap that
let 0.6.23 ship with stale packaging files can no longer survive
even an in-session edit — the agent learns about drift the moment
it edits CMakeLists, not at commit time..claude/bump.json — per-project recipe consumed by a
user-level /bump skill (lives in ~/.claude/skills/bump/, not
in this repo). Lists the 6 version-bearing files with templated
{OLD} / {NEW} / {TODAY} find/replace patterns, plus 4 todos
for content the recipe can't synthesise (CHANGELOG body, AppStream
<release> HTML, debian changelog block, build/test verification).
/bump 0.6.30 walks the recipe and runs check-version-drift.sh
as its post-check.packaging/hook-on-cmakelists-edit.sh — small bash dispatcher
the project hook calls. Stdin is the PostToolUse JSON payload;
stdout is empty on a clean state or {systemMessage:"…"} on
drift. Always exits 0 — never blocks an edit, only informs..gitignore narrowed .claude/ → .claude/* with explicit
!.claude/settings.json / !.claude/bump.json exceptions so the
team-shared workflow files commit while developer-local
.claude/settings.local.json and .claude/worktrees/ stay
ignored. Same pattern adopted in the Vestige 3D Engine repo.Theme: Status-bar reliability pass — nail down every user-reported status-bar symptom in one go: git branch chip elided to "…" when the statusbar had plenty of space, Review Changes button not appearing until a tab-switch, Add-to-Allowlist button silently no-op on save failure, Claude status label 2 s stale after tab-switch, Review Changes click "does nothing" with no feedback.
status_bar_elision
(tests/features/status_bar_elision/) — spec + test covering the
ElidedLabel elision policy. Three assertions: short text never
elides, over-cap text elides with a non-empty tooltip, and the
statusbar layout's QBoxLayout squeeze never drops short chips to
"…" (the user-reported regression vector). Fails against pre-fix
src/elidedlabel.h, passes on fix.ElidedLabel::minimumSizeHint + sizeHint now compute against
m_fullText, not QLabel::sizeHint() which reports the
displayed (possibly already elided) text. This is the root-cause
fix for the user-reported "git branch shows … instead of main"
regression: the previous 3-char minimum let QStatusBar squeeze the
chip below what "main" required, fontMetrics().elidedText()
returned "…" for the squeezed width, the shorter displayed text
fed back into sizeHint, and the widget was stuck on a fixed
point it couldn't escape. minimumSizeHint now returns
full-text-width + padding capped at maximumWidth, so short
text is guaranteed to fit and only genuinely over-cap text elides.refreshReviewButton() now fires on every 2 s status-bar tick
via the shared m_statusTimer. Previously only onTabChanged and
the Claude Code fileChanged hook triggered it, so a dirty repo
at startup (no tab-switch, no hook fired) left the Review Changes
button hidden. Also called once via QTimer::singleShot(0) at end
of the MainWindow constructor so the button appears immediately
on boot rather than 2 s later.ClaudeIntegration::setShellPid calls pollClaudeProcess()
immediately after arming the timer. Previously the Claude status
label was NotRunning for up to 2 s after every tab-switch to a
tab where Claude was actually running, until the next poll tick
caught up — user-visible as "status doesn't update right away."objectName "claudeAllowBtn" matching the scroll-scan
path's bare QPushButton; onTabChanged now searches for
QWidget (not just QPushButton) with that objectName so both
paths are cleaned on tab-switch. Hook-server path also listens on
every terminal for claudePermissionCleared instead of only
currentTerminal(), so a brief visit to another tab doesn't
orphan the button.m_claudePromptActive
so the Claude status label consistently reads "Claude: prompting"
while a prompt is live, and onTabChanged resets the flag so the
next tab doesn't inherit the previous tab's prompt state.showDiffViewer no longer has silent return branches. The
user-reported "Review Changes shows but click does nothing"
symptom was a combination of missing feedback when
focusedTerminal returned null, empty shellCwd, or the
underlying git diff returned non-zero. Each case now emits an
explicit showStatusMessage. Also combined worktree + index diff
into a single git diff HEAD (matches what refreshReviewButton
uses for its enablement probe) instead of two sequential calls
where the second was only reached on empty-worktree cases.ClaudeAllowlistDialog::saveSettings now returns bool and
all three callers (prompt-prefill auto-save,
QDialogButtonBox::accepted, Apply-button onApply) surface
failure to the user: the prefill path shows a status message with
the target path; OK/Apply surface an error in the dialog's
validation label and refuse to accept() so the user can retry.
Previously every QSaveFile open / write / commit failure was
swallowed — the user saw the "Rule added to allowlist" toast even
when the write had failed, then Claude Code re-prompted and the
user reported "Add to allowlist does nothing."Theme: End-to-end feature review — same-class overflow sweep across chrome widgets, quake-mode toggle fix, combining-character preservation on resize, disk-write durability hardening, and auto-profile regex caching.
ElidedLabel widget (src/elidedlabel.h) — QLabel that
truncates its full text with "…" to fit the current widget width,
re-eliding on every resize and exposing the un-elided string via
tooltip. Header-only; drop-in replacement wherever dynamic strings
could outgrow a bounded slot.combining_on_resize
(tests/features/combining_on_resize/) — exercises the three
invariants covering the new resize fix (I1 main-screen preservation,
I2 alt-screen preservation, I3 shrink-range eviction). Confirmed to
fail against pre-fix src/terminalgrid.cpp and pass against the fix.m_statusGitBranch, cap 220 px, elide right),
transient message (m_statusMessage, stretch, elide middle), and
foreground process (m_statusProcess, cap 160 px, elide right) —
now cap out and show "…" instead of growing unbounded. User report:
when the transient-message slot showed "Claude permission:
Bash(git log …)" it pushed the git-branch chip and process indicator
off the right edge. Middle-elide keeps both the leading label and
the trailing detail visible in the message slot.TitleBar::m_titleLabel — window title from OSC 0/2 is
user-controlled and can push the minimize/maximize/close buttons
off a frameless window. Middle-elide keeps "cmd" prefix and "cwd"
suffix both visible.MainWindow::m_claudeStatusLabel — "Claude: <tool-name>" in the
ToolUse state accepts an arbitrary tool name from the transcript
(MCP tools, custom tools); right-elide caps at 220 px.AuditDialog::m_statusLabel — "Running: <check-name>…",
"SARIF saved: <path>", and friends were unbounded; right-elide
keeps the dialog footer stable.checkAutoProfileRules previously compiled a fresh
QRegularExpression per rule on every 2 s tick — 10 rules × 30
ticks/min = 300 wasteful JIT compiles/min. Cached keyed on pattern
string in a function-local static QHash. Invalid patterns (from
mistyped user config) are now detected via QRegularExpression::isValid()
and surfaced via a one-shot status message instead of silently
never matching.m_quakeAnim QPropertyAnimation is reused across
hide/show toggles. The hide branch connected finished→hide() with
Qt::UniqueConnection, but Qt can't dedupe lambda-wrapped slots, and
the stale connection from the previous hide fired at the end of the
show slide-down — the window flashed in then vanished. Fix:
QObject::disconnect(m_quakeAnim, &finished, this, nullptr) before
every start(), and re-attach the hide() slot only on the hide
branch. Show branch has no finished listener (the animation simply
lands at its on-screen end value).TerminalGrid::resize()
(taken when cols is unchanged or when the alt-screen buffer is
being resized alongside main) walked TermLine::cells only and
default-constructed the combining side table — stripping accents,
ZWJ sequences, and variation selectors off every on-screen cell
whenever the user dragged the window edge. Symptom: filenames like
résumé.pdf or naïve.cpp in ls output mutated to
re?sume?.pdf / nai?ve.cpp on every resize. Fix: copy
TermLine::combining and softWrapped alongside cells; filter out
combining entries whose column exceeds the new width (shrink case).
Alt-screen TUIs (vim, less, htop) that render accented filenames
see the same bug and the same fix.fsync() before atomic rename.
Config::save, SessionManager::saveSession, and
SessionManager::saveTabOrder each follow the write-to-.tmpQFile::close flushes only
userspace buffers. On ext4 data=ordered (the common default), a
kernel crash or power loss between close() and rename() can
leave a zero-sized file replacing the previous good one —
silently losing config or a whole session's scrollback. Matches
the write-rename-fsync pattern used by SQLite/Git..spec, PKGBUILD, debian/changelog, man page,
metainfo.xml) being bumped in lockstep —
packaging/check-version-drift.sh correctly flagged the drift and
blocked the merge. All six files now agree at 0.6.28 and CI passes.Theme: clear-command scrollback fix, scroll-offset clamp,
tab-close visibility, Claude "prompting" status, and allowlist-button
lifecycle tightening.
clear now properly wipes scrollback. On TERM=xterm-256color
(set by PtyHandler), ncurses clear emits \E[H\E[2J\E[3J
(verified via clear | od -c). The 0.6.26 Ink-overflow-repaint
heuristic — "mode 3 within 50 ms of mode 2 = Ink frame-reset, preserve
scrollback" — false-matched this byte-burst because the 2J+3J pair
is identical to Ink's clearTerminal. User-reported symptom: after
clear, the scrollbar remained scrollable showing nothing, because
m_scrollback still held the pre-clear lines.
Disambiguator: cursor position at the moment 3J arrives. clear
emits H FIRST so cursor sits at (0,0); Ink emits H LAST so cursor
is still wherever the overflowing output left it (bottom of the
screen in practice). Treating cursor-at-origin as "user-initiated
clear" wipes scrollback on clear while still preserving it for Ink
— because Ink's overflow only triggers when output has filled past
the viewport, by construction the cursor is never at (0,0) when Ink's
3J fires. New user-clear scenario in
tests/features/scrollback_redraw/test_redraw.cpp pins the
behaviour; the existing ink-overflow-repaint scenario continues to
pass.
Scroll-to-bottom button no longer lingers after scrollback
shrinks. TerminalWidget::m_scrollOffset is a widget-side value
that wasn't re-clamped when TerminalGrid::m_scrollback.clear() ran
(via \E[3J). The scrollbar's setValue clamp hid the symptom for
the scrollbar itself, but m_scrollOffset > 0 was still the
show-button condition and the viewport-rendering base (viewStart = scrollbackSize - m_scrollOffset went negative, painting phantom
blank rows). Clamping m_scrollOffset to [0, scrollbackSize] in
updateScrollBar() — the single choke-point called after every
output batch — keeps the widget state consistent with the grid.
Tab close (×) button is now always visible. The theme
stylesheet had QTabBar::close-button { image: none; }, which hid
the × glyph entirely and left only an invisible hover hit-target. New
users didn't discover the close affordance unless they happened to
mouse over it. Removing the image rule lets Qt fall back to the
platform close icon (QStyle::SP_TitleBarCloseButton), which adapts
to the active palette. Hover still applies an ansi-red background
(%7) for "will-click" feedback.
"Add to allowlist" button now also clears the accompanying status
message on dismiss. Clicking the × button cleared the message, but
the claudePermissionCleared path (prompt disappears on
approve/decline) just deleted the button and left the
"Claude Code permission: …" text stuck in the status bar. Both paths
now call clearStatusMessage() symmetrically.
ClaudeIntegration's transcript-driven state machine is
unaware of the on-screen permission UI. For a user scrolled up in
history — who can't see the prompt directly — "idle" was
misleading. The prompt state is now tracked per-window as a boolean
overlay on top of the ClaudeState: any detected prompt flips the
label to yellow "Claude: prompting" regardless of the underlying
state, and the prompt-cleared signal reverts it to the base state.
Implementation consolidates the label-rendering switch into a single
MainWindow::applyClaudeStatusLabel() method so prompt and state
changes share one code path.Theme: UX polish — status bar consistency, focus handling, Ink overflow-repaint fix, tab-colour gradient.
CSI 2J + CSI 3J + CSI H + fullStaticOutput + output whenever its
rendered frame overflows the viewport (see ink/build/ink.js:705,
ansi-escapes/base.js:124 clearTerminal). Previously the CSI 3J
part of that burst wiped the user's conversation history and the
subsequent replay duplicated into scrollback. Now CSI 3J arriving
within 50 ms of CSI 2J on the main screen is classified as Ink's
frame-reset marker: scrollback is preserved, and the 0.6.22
suppression window is armed for the replay so the scroll-off lines
don't duplicate. Standalone CSI 3J (user-initiated clear -x or
similar) still wipes scrollback as the user requested. Two new
conformance scenarios — ink-overflow-repaint and standalone-3J
— in tests/features/scrollback_redraw/test_redraw.cpp pin both
behaviours.ClaudeIntegration::setShellPid emits
stateChanged(NotRunning) + contextUpdated(0) on every tab
switch. The stateChanged slot hid the bar; the contextUpdated(0)
slot unconditionally re-showed it. Fresh tabs, or tabs where Claude
was never started, now stay hidden until a real non-zero percent
is emitted.QPushButton
stylesheet so it matches the sibling "Add to allowlist" button.
Disabled state gets its own distinctive visual (dashed border +
italic + muted fg) so a clean-repo state reads clearly as
non-actionable.QStatusBar's size hint
tracks its tallest child, so the bar jumped when the transient
"Add to allowlist" button appeared and shrank when it went away.
A 32 px setMinimumHeight floor covering the default
QPushButton size keeps the bar stable regardless of transient
widget presence.QFrame::VLine painted in textSecondary (not border,
because border is nearly invisible against bgPrimary on
low-contrast themes like Gruvbox-dark and Nord). Gives a
deterministic visual boundary every theme renders correctly.QApplication::focusChanged. When focus lands on chrome widgets
(QStatusBar, QTabBar, bare QMainWindow) with no active modal,
keyboard focus is deferred-set back to the terminal one tick
later. Dialogs, menus, command palette, QLineEdit/QTextEdit
input widgets, and TerminalWidget descendants all retain focus
legitimately — the whitelist is specifically tuned to avoid
hijacking user-meant focus.QTabBar::tab:selected { border-bottom }) is preserved by
excluding the bottom 2 px from the gradient area.Theme: a user-reported scrollback regression from 0.6.24's
CSI-clear window extension, plus the two packaging-audit follow-ups
from the ninth periodic audit (audit_2026_04_15).
TerminalGrid::scrollbackPushed() frozen, which kept
TerminalWidget::onOutputReceived's scroll anchor from advancing
m_scrollOffset. The screen still scrolled on every LF, so the
viewport rows overlapping the screen got overwritten in place,
and — worse — the anchor never re-pinned to the user's reading
position in scrollback, so every subsequent push (once the
suppression let up) moved the viewport further from where the
user wanted to be. The 0.6.24 broadening to CSI 0J / CSI 1J
made the window arm during the far more common "home + erase to
end" repaint idiom, so the regression surfaced as near-continuous
viewport drift during Claude Code activity. Fix: when the user is
scrolled up (m_scrollbackInsertPaused == true), always push to
scrollback — even inside the CSI-clear window. The anchor then
advances m_scrollOffset by the push count, keeping viewStart
pinned to the user's reading position. The doubling-protection
window keeps its teeth for the at-bottom case (the only case where
the doubling symptom is observable). New feature test
tests/features/scrollback_redraw/test_viewport_stable.cpp
asserts content at viewport rows is invariant across streaming
for both the pause-only path and the CSI-clear path, across scroll
offsets from barely-scrolled-up to deep-in-history. spec.md
gains a §Viewport-stable companion contract documenting the
invariant and its scope (must-hold when viewport is entirely in
scrollback; partial guarantee when viewport overlaps the screen,
since screen writes are inherently visible there).packaging-version-drift rule now scans AppStream metainfo. The
ninth audit noted that
packaging/linux/org.ants.Terminal.metainfo.xml couples to every
release — its <release version="X.Y.Z"> list has to advance on
each bump — but the drift rule was not scanning it. Added to the
rule's coverage list; the grep extracts the newest declared release
(metainfo lists releases newest-first, so head -1 captures the
latest) and diffs it against CMakeLists.txt's project(VERSION).
Verified end-to-end: a contrived CMake bump to 0.9.99 now emits
six findings (five packaging recipes + the metainfo entry) instead
of five.packaging/check-version-drift.sh. The check used to live as a
~20-line bash string embedded in src/auditdialog.cpp. That meant
CI couldn't reuse it without duplicating the logic and risking a
new axis of drift. The script is now the single source of truth;
auditdialog.cpp invokes it with a minimal [ -x … ] && bash … || exit 0 wrapper (graceful no-op when the tree is missing the
script), and CI invokes it directly so a non-zero exit fails the
job. Audit dialog's parser continues to ignore exit codes — same
as before — so the CI-friendly non-zero-on-drift exit semantics
don't break the in-app view.Check packaging version drift step in
.github/workflows/ci.yml. Runs the script after every push and
pull request. A PR that bumps CMakeLists.txt without updating
every coupled file — spec, PKGBUILD, debian/changelog, man page,
AppStream metainfo, README — now fails CI before merge. Closes
the honor-system gap that let 0.6.23 ship with four stale
packaging files (and would have let this very release ship with
drifted metainfo if the new metainfo coverage had not been added
in the same commit).TerminalWidget::outputReceived signal after a 1 s grace
— but Claude Code v2.1+ repaints its TUI continuously (cursor blink,
spinner animation) even while the prompt is still visible, so the
button vanished within seconds of appearing. Introduced a stricter
claudePermissionCleared signal that fires only when the grid scan
detects the prompt footer is no longer on screen, and rewired both
the grid-scan and hook-stream button paths to retract against that
signal instead. The button now stays visible for as long as the
prompt is on screen.textSecondary. The 5e2ac58 fix gave the
disabled state an explicit textSecondary color — which works for
themes where textSecondary is ~70% luminance, but still fails WCAG
AA on themes like One Dark (#5C6370 on bgSecondary #21252B ≈
3:1 contrast). Switched to textPrimary + italic + dissolved
border: guarantees legibility across all 11 themes while the italic
still communicates "nothing to review right now."CSI 2J. The 0.6.22 fix armed the 250 ms
suppression window only on eraseInDisplay mode 2; a ninth-audit
probe found that CSI H; CSI 0J and CSI N;M H; CSI 1J
(with (N,M) at the bottom-right corner) produce the same
post-state — every visible cell cleared — and therefore exhibit
the same doubling bug on subsequent redraw. The window is now
armed for all three equivalent-effect shapes. Added two regression
scenarios to tests/features/scrollback_redraw/ (identical-repaint-0J
and identical-repaint-1J) that reproduce the bug pre-fix and pass
post-fix. Spec §Contract updated to document the broader invariant.CMakeLists.txt at 0.6.23 but the openSUSE spec,
Arch PKGBUILD, Debian changelog, and man page still reading 0.6.22
— the exact drift the packaging-version-drift audit rule (0.6.22)
was written to catch. Rule wasn't run on the 0.6.23 release commit.
Also added org.ants.Terminal.metainfo.xml release entries for
0.6.23 and 0.6.24 (AppStream metadata is not currently checked by
the audit rule — follow-up item).Theme: status-bar event-driven contract + three user-reported UX fixes (allowlist subsumption false-positive, tab-colour picker had no visible effect, Shift+Enter paste-image regression) and the first 0.7.0 performance item — SIMD VT-parser scan.
Bash(make * | tail * && ctest * | tail *)
opened the dialog with the right compound rule, but a pink warning
said *"Already covered by existing rule: Bash(make )" and refused
to add it — so the next compound invocation re-prompted. Root
cause: ClaudeAllowlistDialog::ruleSubsumes did a flat
narrowInner.startsWith(broadPrefix + " ") check, which was true
for any compound command whose first segment matched. Fix: split
the narrow rule on shell splitters (&&, ||, ;, |) and
require every segment to start with the broad prefix before
declaring subsumption. Matches the per-segment semantics
generalizeRule has used since 0.6.22. A compound broad rule
deliberately returns false rather than risk a false positive —
a duplicate rule is benign, a false-positive warning is not.
ruleSubsumes promoted from private to public so the feature
test can exercise the contract directly. Feature test in
tests/features/allowlist_add/ adds 16 cases including the exact
reproduction + simple-prefix regression guards + all four
splitters. See tests/features/allowlist_add/spec.md §C.m_tabColors[idx] and called setTabTextColor, but the
window-level stylesheet rule QTabBar::tab { color: %6; } beat
setTabTextColor on QSS specificity, so the colour was stored
and never rendered. Fix: new ColoredTabBar subclass (storage
via QTabBar::tabData — survives drag-reorder and auto-drops on
close, unlike an index-keyed map) + a trivial ColoredTabWidget
wrapper that installs it at construction (QTabWidget::setTabBar
is protected). paintEvent overlays a 3-px bottom strip in the
chosen colour after the base class paints, outside stylesheet
influence. MainWindow's context-menu handler now calls
m_coloredTabBar->setTabColor(idx, color) and the stale
m_tabColors map is removed. New feature test
tests/features/tab_color/ asserts round-trip through
setTabColor/tabColor, colour-follows-tab across moveTab
(drag reorder), and cleanup on removeTab with no index leak
into a newly-inserted tab. See tests/features/tab_color/spec.md.MainWindow::onTabChanged now calls
clearStatusMessage(), hides m_claudeErrorLabel, and
deleteLater's every QPushButton on the status bar with
objectName == "claudeAllowBtn" at the top of the handler —
before state-bar refresh (updateStatusBar, refreshReviewButton,
setShellPid) paints the new tab's state. Without the cleanup,
switching from "tab A asked permission" to tab B left the Allow
button visible inviting approval of the wrong prompt, and
theme-change notifications from tab A lingered for five seconds
on tab B. Contract codified in
tests/features/claude_status_bar/spec.md §D (new section).kittyKeyFlags() > 0, which emits CSI 13;2u for
Shift+Enter; some TUIs (Claude Code v2.1+ among them) treat that
sequence as "paste clipboard" when an image is present. Fix: move
the Shift+Enter literal-newline block to fire before the Kitty
encoder so the sequence Ants puts on the wire is always
\x16\n (readline quoted-insert + LF) regardless of the
negotiated Kitty flags. Ctrl+Shift+Enter's scratchpad continues
to take precedence; the new guard
!(mods & Qt::ControlModifier) makes that explicit.
src/terminalwidget.cpp keyPressEvent reorder.0x20 ..0x7E) via a 16-byte-wide SIMD lane, avoiding per-byte
state-machine work. Implementation in src/vtparser.cpp uses
SSE2 on x86_64 and NEON on ARM64 with a scalar tail / fallback;
a signed-compare trick (XOR 0x80 → two cmpgt_epi8s against
pre-computed bounds) flags any byte outside the safe range in
one movemask + ctz. TUI repaints (the workload that drove
the 0.6.22 scrollback-doubling fix) are the biggest winner —
thousands of printable bytes at a time now go through a tight
SIMD loop instead of four function calls per byte. New feature
test tests/features/vtparser_simd_scan/ asserts the emitted
action stream is byte-identical across whole-buffer, byte-by-byte,
and pseudo-random-chunk feed strategies over a 38-case corpus
including every offset-mod-16 alignment for an interesting byte
(regression guard against scan-boundary off-by-ones). See
tests/features/vtparser_simd_scan/spec.md.Theme: root-cause fix for the main-screen TUI "scrollback doubling"
regression, plus packaging catch-up (four recipes bumped to 0.6.22, three
missing AppStream <release> entries, openSUSE %post/%postun
scriptlets, Debian postinst/postrm). The scrollback fix is the
headline — Claude Code v2.1+ repaints its entire TUI on every state
update (not just /compact), and each repaint was pushing a full-screen
worth of duplicate content into scrollback. 0.6.21's scrollback-insert
pause only triggered while the user was scrolled up; this release adds
a sliding time window anchored to main-screen CSI 2J, so the
suppression works at the bottom too.
fileChanged signal without checking whether a diff
actually existed by the time the user clicked. Three-state design:
git diff --quiet HEAD (async, off the UI thread) determines
(a) hidden — no active terminal, no cwd, or cwd is not a git
repo; (b) visible + disabled — clean git repo (Claude edited
something but diff is gone); (c) visible + enabled — real diff
present. Re-checked on fileChanged, on tab switch, and after the
diff viewer finds empty output.generalizeRule() previously only handled cd X && cmd
(and even that dropped the cd prefix, producing rules that Claude
Code's allowlist matcher rejected on the next compound invocation).
User feedback: the "Add to allowlist" button would add a rule that
re-prompted on the very next turn. Rewritten to split on &&, ||,
;, | while preserving compound structure: Bash(cd /x && make y)
now generalises to Bash(cd * && make *) (matching the convention
already in the user's settings.local.json), Bash(cat f | grep x)
to Bash(cat * | grep *), etc. Safety denylist retained and
extended to ALL segments (any rm/bare-sudo/SUDO_* anywhere in
the chain blocks generalisation). Feature test in
tests/features/allowlist_add/ exercises 23 cases covering every
splitter variant and the denylist.ClaudeIntegration::setShellPid() re-pointed the watcher
to the new tab's shell PID but left the cached m_state /
m_currentTool / m_contextPercent from the previous tab intact
until the next poll tick (~1 s later). Tab A's "Claude:
thinking..." bled into tab B's status until polling caught up. Fix:
when setShellPid() receives a PID different from the current one,
immediately clear the cached state, drop the transcript watch, and
emit stateChanged(NotRunning, "") + contextUpdated(0) so the UI
reflects the switch within the current event-loop iteration. Paired
with a new feature test
(tests/features/claude_status_bar/spec.md) that verified the
fix both ways — fails against the un-fixed setShellPid, passes
with the one-line state-clear.tests/features/sgr_attribute_reset/spec.md for the invariant
asserted by CTest.CSI 2J,
eraseInDisplay(2)), open a 250 ms sliding window during which
scrollUp() drops the push-to-scrollback step. Each scrollUp during
the window extends it, so the window survives the entire repaint
burst regardless of length, but closes promptly when the app goes
quiet. Alt-screen bypasses scrollback already and is unaffected;
mode-3 erase (which clears scrollback by request) is also unaffected.
Defensively cleared on alt-screen entry. The 0.6.21 `m_scrollOffset0
pause (for users actively reading history) is retained — the two triggers compose (either one suppresses). File:terminalgrid.cpperaseInDisplay()+scrollUp(); header additions interminalgrid.h(m_csiClearRedrawActive,m_csiClearRedrawTimer,kCsiClearRedrawWindowMs`).
trivy fs with the vuln, secret,
and misconfig scanners in one invocation, severity-floor
MEDIUM. Output is piped through jq into the standard
path:line: severity: scanner/rule-id: title shape so
parseFindings() consumes it directly. Self-disables when
either trivy or jq is missing (the description text
explains how to install). Auto-selected when both are present.cmake_no_version_floor (find_package
REQUIRED with no version constraint), bash_c_non_literal
(bash -c <ident> injection-sink heuristic, restricted to lines
containing bash / sh so it doesn't flag grep -c), and
packaging_version_drift (cross-file consistency: extracts
CMakeLists.txt project(VERSION X.Y.Z) and diffs it against
every packaging recipe). Both regex-based rules ship with
bad/good fixture pairs under tests/audit_fixtures/; the
drift check is exercised by the audit dialog itself.tests/features/). New test
lane alongside the existing audit_rule_fixtures. Each feature
subdirectory ships a spec.md (human contract — reviewed like
code) and a C++ test executable that exercises the feature through
its public API and asserts the contract holds. GUI-free, fast,
labelled features so CI and local ctest -L features pick them
up. First test landed: scrollback_redraw — the regression guard
for this release's headline fix. Verified that it FAILS against
0.6.21 code (growth=77 vs. threshold=34, reproducing the doubling)
and PASSES against 0.6.22 — i.e. it would have caught the
incomplete 0.6.21 fix at commit time. See tests/features/README.md
for the authoring guide and rationale. CLAUDE.md updated.<release> was 0.6.17
(so 0.6.18/.19/.20/.21 were invisible to GNOME Software / KDE
Discover users reading the metainfo catalogue).%post/%postun scriptlets (update-desktop- database, gtk-update-icon-cache). Without these, minimal
Tumbleweed images won't see the launcher in the application menu
until next session restart. Matches Fedora/openSUSE packaging
guidelines for any package shipping a .desktop entry + hicolor
icons.Theme: audit-tool noise reduction, dialog theming, and a long-
standing main-screen-TUI scrollback corruption fix. Self-audit triage
turned up a GNU grep 3.12 argument-ordering quirk that silently
disabled the --include file-type filter, producing the bulk of the
CRITICAL/MAJOR false positives. Same release extends the QSS cascade
so every popup dialog inherits the terminal's active theme, and adds
a scrollback-insert pause that stops TUI redraws from interleaving
intermediate frames into the history the user is trying to read.
scrollbackPushed), but the history below their
read point filled with duplicated content and the "scrolled" view
drifted. Fix: TerminalGrid::scrollUp() now honours a new
scrollbackInsertPaused flag, set by TerminalWidget::onPtyData()
while m_scrollOffset > 0. Lines that would have been pushed into
scrollback during a paused window are simply dropped (they were
about to be overwritten in-place anyway). Flag auto-clears on the
next PTY packet once the user returns to the bottom.MainWindow::applyTheme() now ships
a comprehensive QSS covering QDialog, QLabel, QPushButton,
QLineEdit, QTextEdit / QPlainTextEdit / QTextBrowser,
QCheckBox, QRadioButton, QComboBox, QSpinBox /
QDoubleSpinBox, QGroupBox, QListWidget / QTreeWidget /
QTableWidget, QHeaderView, QScrollBar (v+h), QToolTip,
QProgressBar, and QDialogButtonBox. Every pop-up created with
the main window as parent (Settings, Audit, AI, SSH, Claude
Projects/Transcript/Allowlist, homograph-link warning, permission
prompts, QMessageBox, QInputDialog, QFileDialog, QColorDialog,
…) now paints with the terminal's active theme colors — no more
light-gray system defaults leaking through. Cached dialogs are
re-polished on theme switch so live widgets pick up new colors
without re-instantiation.--include=<glob> filter whenever a file-level
--exclude=<name> appears earlier on the command line — every file
under the search root is scanned regardless of extension. Our
addGrepCheck() builder put kGrepExclSec (which carried
--exclude=auditdialog.cpp --exclude=auditdialog.h) before
kGrepIncludeSource, so security scans that should have been scoped
to source files were also matching .md, .xml, CMakeLists.txt,
and the audit tool's own pattern-definition strings. Split the
exclude-dir and file-exclude portions, append file-excludes AFTER all
--include flags (including caller-supplied extras). Cuts the
CRITICAL cmd_injection false-positive count from 8 → 4 (remaining
hits are either real — execlp in the forkpty child, by design — or
intentional test fixtures).lineIsCode() over-counted string-only lines as code.
The comment/string-aware filter flagged continuation lines like
"system(), popen()…", "Security", as real code because the
string-delimiter character itself was treated as a code token.
Tightened to require an identifier- or operator-class character
outside strings/comments before a line counts as code. Drops the
remaining self-referential matches in auditdialog.cpp's own
pattern-definition arguments without affecting legitimate mixed
code+string lines like int n = f("foo");.invalidSuppression noise. cppcheck's own
--inline-suppr parser misreads documentation comments that mention
the literal cppcheck-suppress token (e.g. the passthrough-marker
docs in auditdialog.cpp) as actual suppression directives and
emits invalidSuppression errors. Added --suppress=invalidSuppression
to both cppcheck invocations — the category catches only meta-parsing
failures, never real code bugs, so losing it has no downside. Also
restructured the offending docs (auditdialog.{h,cpp}) so the token
no longer appears as the first word of a comment body.Theme: ROADMAP §H5 (Distribution-readiness bundle 5) — ready-to-submit
packaging recipes for the three mainstream non-Flatpak distros. Each
recipe drives the existing CMake install rules (no build-system forks,
no per-distro patches) and runs the audit-rule regression suite in its
%check / check() / dh_auto_test step. Fully additive, no runtime
code changes.
packaging/opensuse/ants-terminal.spec — openSUSE RPM spec
targeting Tumbleweed. Uses the core openSUSE CMake macros
(%cmake, %cmake_build, %cmake_install, %ctest, %autosetup)
so the file is close to portable to Fedora. BuildRequires declared
via cmake(Qt6*) pkgconfig-style entries so OBS resolves them
against whichever Qt6 stack the target project carries. %files
enumerates all fifteen install paths explicitly so a missing or
relocated artefact fails the OBS build instead of producing a
silently-incomplete package.packaging/archlinux/PKGBUILD — AUR recipe on the release
track (package name ants-terminal). check() runs ctest. sha256sums
is SKIP in the in-tree recipe with a comment pointing packagers at
updpkgsums since the upstream tarball doesn't exist until the tag
is pushed. A rolling -git variant is documented in
packaging/README.md (three-line diff: pkgname, source,
pkgver()).packaging/debian/ — Debian source tree: control, rules,
changelog, copyright, source/format. debhelper-compat 13
drives dh $@ --buildsystem=cmake with Ninja as the backend; DEP-5
copyright carries the full MIT license text.
DEB_BUILD_MAINT_OPTIONS = hardening=+all stacks dpkg-buildflags'
hardening wrappers on top of our CMake hardening flags. Suitable
for debuild -uc -us, a Launchpad PPA, or an eventual Debian ITP.packaging/README.md — one-page build / submission guide for
all three recipes: local-build recipe, OBS / AUR / PPA submission
flow, dch / osc vc / updpkgsums version-bump recipes, and
the -git variant diff for Arch.CMakeLists.txt project(... VERSION 0.6.19) → 0.6.20. The
single-source-of-truth macro ANTS_VERSION propagates to
setApplicationVersion, the SARIF driver, the dialog title badge,
and the HTML-export metadata on rebuild.packaging/linux/ants-terminal.1 .TH line now reads
ants-terminal 0.6.20 (was stuck at 0.6.18 — missed the 0.6.19
bump). groff ships the version string in its header/footer, so
man ants-terminal now matches ants-terminal --version again.Theme: ROADMAP §H4 (Distribution-readiness bundle 4 of 4) — shell
completions for the current QCommandLineParser surface, installed to
the conventional vendor locations for bash, zsh, and fish so distro
packages don't have to relocate them. Closes the H1–H4 distribution
slice; remaining packaging work (H5 spec/PKGBUILD/debian, H6 Flatpak,
H7 docs site) lives in 0.8.0+. Fully additive, no runtime code changes.
packaging/completions/ants-terminal.bash — bash-completion
function _ants_terminal registered via complete -F. Suggests the
full long/short option set; suppresses suggestions after --new-plugin
(the next token is a freeform plugin name with no useful enumeration).packaging/completions/_ants-terminal — zsh #compdef
function using _arguments, with mutually-exclusive groups so --quake
and --dropdown only offer the unspecified alias, and -h / -v
short-circuit further completion ((- *) exclusion).packaging/completions/ants-terminal.fish — fish complete
declarations, one per flag, with descriptions that mirror the manpage
one-liners. --new-plugin declared -r -x so fish knows it consumes
the next token but doesn't try to file-complete it.${datarootdir}/bash-completion/completions/ants-terminal,
${datarootdir}/zsh/site-functions/_ants-terminal, and
${datarootdir}/fish/vendor_completions.d/ants-terminal.fish. All
three are auto-discovered without the user sourcing anything by hand..github/workflows/ci.yml runs bash -n on the
bash file, zsh -n on the zsh file, and fish --no-execute on the
fish file so syntax regressions fail the build the same way the H2
AppStream / desktop-entry validators and H3 groff lint do.Theme: ROADMAP §H3 (Distribution-readiness bundle 3 of 4) — a
spec-compliant man 1 ants-terminal page so distro packagers and
users alike get the standard Unix discovery experience (man ants-terminal, apropos terminal, whatis ants-terminal). Fully
additive, no runtime code changes.
packaging/linux/ants-terminal.1 — groff -man
source covering NAME / SYNOPSIS / DESCRIPTION / OPTIONS
(-h/--help, -v/--version, --quake/--dropdown, and
--new-plugin <name> with its full validation contract) /
ENVIRONMENT (SHELL, HOME, XDG_CONFIG_HOME, XDG_DATA_HOME,
ANTS_PLUGIN_DEV, QT_QPA_PLATFORM) / FILES (config /
themes / plugins / sessions / recordings / logs / audit_rules.json
/ .audit_suppress) / EXIT STATUS (the four --new-plugin codes)
/ BUGS / AUTHORS / SEE ALSO (cross-refs to xterm, konsole,
gnome-terminal, tmux, ssh, forkpty(3), appstreamcli,
desktop-file-validate). Renders cleanly under both groff -man and
man -l.install(FILES … DESTINATION ${CMAKE_INSTALL_MANDIR}/man1). DESTDIR-staged install verified to
drop the page at …/share/man/man1/ants-terminal.1, which man-db
picks up automatically once the prefix is on MANPATH (true for
/usr and /usr/local out of the box)..github/workflows/ci.yml installs groff and
runs `groff -man -Tutf8 -wall packaging/linux/ants-terminal.1/dev/null
so syntax regressions in the man source fail the build the same wayappstreamcli/desktop-file-validate` do for the desktop entry and metainfo XML.
Theme: ROADMAP §H2 (Distribution-readiness bundle 2 of 4) — AppStream metainfo and a spec-compliant desktop entry so Ants appears in GNOME Software / KDE Discover catalogues once a distro packages it. Fully additive, no runtime code changes.
packaging/linux/org.ants.Terminal.metainfo.xml — AppStream 1.0
descriptor: id / name / summary / multi-paragraph
description / categories / keywords / url (homepage,
bugtracker, vcs, help, contribute) / provides/binary /
supports/control / content_rating (OARS 1.1, all-none) /
launchable tying to the .desktop id / releases with 0.6.14–
0.6.17 changelog summaries. metadata_license CC0-1.0,
project_license MIT. Validated by appstreamcli validate
(pedantic clean aside from the reverse-DNS uppercase hint, which
matches the convention used by org.gnome.Terminal and
org.kde.Konsole).packaging/linux/org.ants.Terminal.desktop — reverse-DNS app
id so it round-trips with the metainfo launchable. Fields:
Type=Application, Terminal=false, Categories=Qt;System;Terminal- Emulator; (exactly one main category — avoids the multi-listing
menu hint), Keywords= tightened, StartupWMClass=ants-terminal,
TryExec=ants-terminal. Two Desktop Action entries — NewWindow
and QuakeMode (wires --quake) — for right-click launcher
integration in most DEs. Validated by desktop-file-validate clean.include(GNUInstallDirs)) for the desktop
file (share/applications/), metainfo XML (share/metainfo/), and
the six hicolor icons renamed to ants-terminal.png under
share/icons/hicolor/<size>x<size>/apps/. DESTDIR=… staging works
out of the box for distro build roots..github/workflows/ci.yml installs
appstream + desktop-file-utils and runs appstreamcli validate --explain + desktop-file-validate on every push so schema drift
in the packaging files fails the build instead of landing in a
release tarball.make install with cmake --install build, adds a table of every path
the install rule lays down (bin/ desktop/ metainfo/ hicolor icons),
and notes DESTDIR=… support for packagers.ants-terminal.desktop.in + launch.sh for running
uninstalled from the build tree) and the distro workflow (the new
packaging/linux/ files installed via CMake). Dev path unchanged;
packaged path is new.packaging/linux/.Theme: ROADMAP §H1 (Distribution-readiness bundle 1 of 4) — community and security policy docs that distro packaging teams look for before accepting a package. Fully additive, docs-only.
CODE_OF_CONDUCT.md — Contributor Covenant 2.1 verbatim
with two reporting channels: a dedicated maintainer email and the
private GitHub Security Advisory (tagged [conduct] for triage) for
reporters who prefer GitHub's private-report flow. Email is listed
first per CoC convention — no GitHub account required. Ships
ROADMAP §H1.SECURITY.md — coordinated-disclosure policy: supported-versions
table, reporting channels (GitHub Security Advisory preferred;
signed encrypted email for reporters who need it), 48h / 7d / 30d /
90d disclosure timeline, severity rubric (Critical / High / Medium /
Low), in-scope / out-of-scope lists, and an acknowledgement of the
hardening already in the tree (image-bomb budgets, URI scheme
allowlist, plugin sandbox, OSC 52 quotas, multi-line paste
confirmation, compile-time hardening flags, ASan + UBSan CI).
Completes ROADMAP §H1. (The file itself landed in commit
4599813; this release changelogs it alongside the CoC.)ROADMAP.md — new 📦 theme plus a
rollup section covering bundles H1–H16 (security/CoC, AppStream +
desktop, man page, shell completions, distro packaging for
openSUSE / Arch / Debian, Flatpak, docs site, macOS, accessibility,
i18n, reproducible builds, SBOM, distro submissions, deb/rpm
signing). Nothing ships from this plan yet except H1; it exists so
later releases have a single source of truth for packaging work.CONTRIBUTING.md,
CODE_OF_CONDUCT.md, and SECURITY.md before the step-by-step
fork/PR instructions, so first-time contributors see the policy
docs before they see the mechanics.CONTRIBUTING.md intro references the CoC (participation
agreement) and SECURITY.md (private channel for sensitive
reports) so the rules aren't buried.Theme: the incremental reflow optimization from ROADMAP §0.7.0 Performance. Resizing a terminal with a long scrollback used to re-join every line into a logical-line buffer and re-wrap every logical line back into TermLines, allocating scratch vectors and copying cells cell-by-cell — O(scrollback) allocation churn on every width change. This release adds a fast path for the common case: standalone lines (not part of a soft-wrap sequence) whose content fits the new width get resized in place, skipping join/rewrap entirely.
TerminalGrid::resize(). The
new algorithm walks scrollback once:softWrapped == false, not mid-
sequence, content width ≤ new cols): tl.cells.resize(cols)
with default-attr padding when growing, or trim trailing blanks +
drop combining-char entries at now-removed columns when shrinking.
No allocation of scratch TermLines. No cell-by-cell copy.joinLogical + rewrap round-trip, unchanged. This
preserves correctness for trailing-space trim, combining-char
position merging, and wide-character boundary handling.
Typical scrollback is dominated by standalone lines ≤ the new width,
so the fast path handles most rows. Multi-line soft-wrap sequences
(long commands that wrapped at the prompt's edge) still round-trip
through the full logic so nothing regresses there.m_screenLines as before.simple copy today; scrollback hyperlink
spans are not re-wrapped at all (pre-existing behavior) — both
outside the scope of this optimization.PLUGINS.md does not need a bump.Theme: small security follow-up to 0.6.13 — catch a URI-scheme
injection vector in the new make_hyperlink trigger action that
self-review after shipping turned up. No ROADMAP transition; this is
a hardening patch on an already-shipped feature.
TerminalGrid::addRowHyperlink(). The
make_hyperlink trigger action expands $0..$9 backrefs from the
matched PTY output into the URL template — and PTY output is
attacker-controllable (SSH sessions, remote shells, any inner-tty
process). Without a scheme check, an innocent-looking template
like https://example.com/$1 could be weaponized when the
pattern matches an attacker-supplied string that expands to
javascript:alert(1) or data:text/html,.... The new check
mirrors the existing OSC 8 allowlist (http, https, ftp,
file, mailto); hyperlinks with any other scheme are silently
dropped, matching OSC 8's drop-on-bad-scheme semantics — text
still prints, it just isn't clickable.handleOsc(); kept in lockstep so both sources share the same
threat model.QSet added to terminalgrid.cpp includes for the allowlist
storage.Theme: close out the last deferred bullet from the 0.6.9 trigger-system
bundle — the iTerm2 HighlightLine, HighlightText, and
MakeHyperlink actions, which needed grid-cell mutation surgery that
chunk-level matching (checkTriggers) couldn't provide. This release adds
a new line-completion callback on TerminalGrid that fires from
newLine() with the row the cursor is leaving, and TerminalWidget
runs the grid-mutation subset of trigger rules against the finalized
line text there — so matches map to exact col ranges on a real row, and
mutations land before the row scrolls into scrollback.
highlight_line trigger action — when the pattern matches in
finalized output, recolors the entire line the match landed on.
Action Value format: "#fg" (fg only), "#fg/#bg" (both), or
"/#bg" (bg only). Empty-string slots mean "leave unchanged" —
so partial overrides compose cleanly with the line's existing SGR
attributes.highlight_text trigger action — same color-string format as
highlight_line, but recolors only the matched substring using
the regex's capture-0 span as the col range.make_hyperlink trigger action — turns the matched substring
into a clickable OSC 8-equivalent hyperlink. Action Value is a URL
template with $0..$9 backrefs expanded against regex capture
groups ($0 = whole match, $1..$9 = groups). Example: pattern
#(\d+) + template https://github.com/milnet01/ants-terminal/issues/$1
→ a clickable issue link on every #123 reference in output. Matches
iTerm2's MakeHyperlink contract.TerminalGrid::setLineCompletionCallback(…) — new public API.
Fired from newLine() before any cursor/scroll motion with the
screen row the cursor was leaving. Used by TerminalWidget to run
grid-mutation trigger rules; any other consumer can hook in for
per-line post-processing (e.g., future trigger-system actions that
need line-level context).TerminalGrid::applyRowAttrs(row, startCol, endCol, fg, bg) —
overrides per-cell fg/bg on an inclusive col range of a screen row.
Invalid QColor() leaves that channel untouched so callers can write
fg-only, bg-only, or both. Marks the row dirty for re-paint.TerminalGrid::addRowHyperlink(row, startCol, endCol, uri) —
pushes a hyperlink span onto the screen row's per-line
m_screenHyperlinks vector with the same shape the OSC 8 parser
uses. Auto-grows the vector if the row index is past the current
end. Rows scrolling into scrollback preserve the span via the
existing scrollUp move-semantics path.TerminalWidget::onGridLineCompleted(int screenRow) — dispatch
handler installed as the grid's line-completion callback. Iterates
rules, filters to the three new grid-mutation types, runs
globalMatch() on the finalized line text, and invokes the
appropriate applyRowAttrs / addRowHyperlink for each match (so
multiple hits on one line all apply).{"notify", "sound", "command"} list was hiding the
bell / inject / run_script types shipped in 0.6.9 from the
UI; 0.6.13 adds those plus the three new grid-mutation types so
every trigger rule supported by the backend is reachable from
Settings. The trigger-tab description was also expanded to document
the Action Value format for each type.TerminalWidget::checkTriggers() now skips the three new
grid-mutation types — those dispatch through onGridLineCompleted
instead. Routing them through both paths would double-fire and
match on raw PTY byte chunks (which don't map to grid cells)
instead of finalized line text.newLine() is called after a \n. The instant flag is not
honored for these three types (it wouldn't make sense — the mid-
line text isn't final, and half-matched mutations would flash
then get overwritten). Dispatch types (notify/sound/etc.) still
honor instant via the existing chunk-level matcher.{pattern, action_type, action_value, instant, enabled} rule shape from 0.6.9; the new
types just teach both checkTriggers and onGridLineCompleted
how to interpret action_type and action_value.highlight_text and make_hyperlink
both use globalMatch() so every occurrence on a line is
decorated (not just the first).PLUGINS.md does not need a
bump; trigger rules live in config, not in the ants.* API.Theme: close out the semantic-history ROADMAP item (which was already implemented but never promoted on the roadmap) and round out its editor coverage so the line/column jump works for editors beyond just VS Code and Kate.
TerminalWidget::openFileAtPath(). Ctrl-click
on a path:line:col capture in scrollback (compiler / linter /
stack-trace output) now opens the file at the cited location for:code, code-insiders, codium, vscodium
via --goto <path>:<line>:<col>-l <line> + -c <col> (col was previously ignored)<path>:<line>:<col> argv shapenvim, vim, vi, ex via +<line> <path>+<line>,<col> <path>idea, pycharm, clion, goland,
webstorm, rider, rubymine, phpstorm, datagrip via
--line <line> + --column <col>QFileInfo(editor).fileName() so absolute
paths (/usr/local/bin/code) and bare names (code) both work./proc/<pid>/cwd to resolve relative paths); this release just
documents the existing behavior and broadens the editor list. See
CHANGELOG.md §0.6.12.(?:^|[\s("'])((\.{0,2}/)?[a-zA-Z0-9_\-./]+\.[a-zA-Z0-9_]+(?::(\d+)(?::(\d+))?)?)
pattern in detectUrls() already captures path:line:col and was
the basis for the original semantic-history routing.+0 / --line 0 arg.Theme: ship the 0.7.0 security pair — a UI for auditing and revoking plugin capabilities, plus image-bomb defenses for the inline graphics decoders. Both items were the smallest, lowest-risk slice of the remaining 0.7.0 work; bundling them keeps the security theme closing out before the riskier threading/platform work begins.
manifest.json. Each
permission renders as a checkbox whose initial state reflects the
current grant in config.plugin_grants[<name>]. Unchecking + Apply
persists the revocation; takes effect at the next plugin reload
(matches the existing first-load permission prompt's grant
semantics). When no plugins are loaded — either Lua is not
compiled in or the plugin directory is empty — the tab shows a
guidance message pointing at PLUGINS.md. Permission descriptions
match the first-load dialog wording so users see the same
language across both surfaces. Closes
ROADMAP.md §0.7.0 → "Plugin capability audit UI".ImageBudget class (private to
TerminalGrid) tracks total decoded image bytes across every
inline-display entry (Sixel + Kitty T/p + iTerm2 OSC 1337) and
the Kitty graphics cache. Cap = 256 MB per terminal — when a
decode would push the total past the cap, the decoder rejects the
payload up front (Sixel) or post-decode (Kitty PNG / iTerm2
OSC 1337 — those don't carry pre-decode dimensions in the params
block) and surfaces an inline error in the warning color
([ants: <kind> WxH (NN MB) exceeds 256 MB image budget]). Every
eviction site (inline-image FIFO drain, Kitty cache eviction, Kitty
delete-by-id, Kitty delete-all) releases bytes back to the budget.
Alt-screen swap calls recomputeImageBudget() so the counter stays
accurate when the active and saved containers cross paths. Existing
per-image dimension cap (MAX_IMAGE_DIM = 4096) remains in force —
ROADMAP specified 16384 as the upper bound but our existing 4096 is
stricter. Closes
ROADMAP.md §0.7.0 → "Image-bomb defenses".SettingsDialog::PluginDisplay — lightweight POD struct
decoupling the Settings dialog from pluginmanager.h (and from the
Lua-gated build chain). MainWindow translates each PluginInfo into
this form before handing the list to the dialog via setPlugins(),
so the Plugins tab compiles and runs even without Lua support.TerminalGrid::writeInlineError(QString) — emits a short red
ASCII diagnostic line at the cursor (CR+LF after) for surfacing
image-bomb rejections without the disruption of a desktop
notification. Color matches the command-block status stripe red.TerminalGrid::recomputeImageBudget() — bulk-recompute helper
for the alt-screen swap path where individual add/release tracking
would be brittle. O(N) over the inline + cache vectors.SettingsDialog constructor now wires a "Plugins" tab between
"Profiles" and the OK/Cancel button row. applySettings() collects
checked permissions per plugin and calls
Config::setPluginGrants(name, granted) for every plugin in the
audit list — including the "all unchecked" case so revocation
persists as an empty grant array.MainWindow::settingsAction lambda now gathers the current plugin
snapshot from m_pluginManager->plugins() (or an empty list when
ANTS_LUA_PLUGINS is undefined) and calls setPlugins() on the
dialog every time it opens — so hot-reloads / new installs are
reflected without recreating the dialog.TerminalGrid Sixel / Kitty / iTerm2 decode paths and every
image-eviction site now consult / update the per-terminal
ImageBudget. RIS (ESC c) replaces the entire TerminalGrid
via *this = TerminalGrid(...), which naturally resets the budget
alongside the containers; no special-case code needed.config.plugin_grants[<name>] = [...] shape from manifest v2
(0.6.0). The new UI is additive: it reads what the first-load
permission prompt already wrote and exposes it for editing.QImage stored in both
m_inlineImages and m_kittyImages is counted twice even though
Qt's COW means the underlying bitmap is allocated once. We may
reject earlier than strictly necessary; we never reject too late.PLUGINS.md does not need a
bump. The audit UI sits entirely outside the ants.* API.clipboard.write, settings, and net mirror the first-load
dialog text in mainwindow.cpp.Theme: ship the command-block UI bundle from 0.7.0 §Features (shell integration). OSC 133 prompt regions already carried the data; this release wires the UI that makes blocks first-class citizens — per-block context menu, pass/fail status stripe, asciicast export of a single block. Also formalizes asciinema recording (the recorder itself has shipped — it's now marked ✅ on the roadmap).
\r to the PTY —
bypasses paste confirmation since it's your own history, not
external input), Fold / Unfold Output (toggles the existing
fold triangle for this specific region rather than the cursor
region), and Share Block as .cast… (exports one block to a
standalone asciicast v2 file). The menu header shows
Command Block (exit N) when the block has finished so users can
see pass/fail before acting. Actions are gated on block state —
Re-run requires commandEndMs > 0, Copy Output requires
hasOutput, and so on.exit_code == 0, red otherwise. Unlike the fold triangle it shows
on the most recent finished block too (no successor region
required). Gives at-a-glance scannability of long scrollback for
pass/fail.exportBlockAsCast(int, QString) — writes a standalone
asciicast v2 file with a single command block's output. Header =
current grid dimensions + commandStartMs timestamp. Event stream
= two records: the command echo at t=0.0, and the captured output
at t=duration (synthesized; we don't have per-byte timing for
historical blocks). Players replay as a prompt followed by the
output dump — honest about what we recorded. Shared via the
Share Block as .cast… context menu action; file extension
.cast per the asciicast v2 spec.TerminalWidget —
promptRegionIndexAtLine(int), commandTextAt(int),
outputTextAt(int), rerunCommandAt(int), toggleFoldAt(int).
Public so plugin code and future UI (marketplace, block sharing
service) can consume the same semantics.PromptRegion now carries three new fields — exitCode
(stored in the OSC 133 D handler alongside m_lastExitCode; lets
block UI paint per-region pass/fail instead of relying on the
grid-global most-recent value), commandStartCol (cursor column
at OSC 133 B fire — lets commandTextAt strip PS1 from the
extracted command), and outputStartLine (global line where OSC
133 C fired, so output extraction doesn't guess based on
endLine + 1). All additive; existing consumers unaffected..cast v2)" move from 📋 (planned) to
✅ (shipped in 0.6.10). Four bullets consolidated under the
command-block item — navigation (Ctrl+Shift+Up/Down) was
already wired in an earlier release, metadata display (durationPLUGINS.md not bumped.rerunCommandAt uses the same m_pty->write path the user's own
keystrokes use.commandTextAt
follows the region from endLine through outputStartLine - 1
(or the next region's startLine - 1 when no output was captured),
joining lines with \n. Trailing-space padding is stripped per
line.Theme: ship the trigger system bundle from 0.7.0 §Plugins — four items that share LuaEngine / PluginManager / OSC-dispatch plumbing and are cleaner to land together than to drip-feed across four releases. Sets up shell-aware plugin workflows for 0.7.
TriggerRule JSON schema gains an
instant boolean (iTerm2 convention — true fires on every PTY
chunk for in-progress matches like password prompts; false
defaults to firing only on newline-terminated chunks so settled
output doesn't double-fire mid-line) and three new action_type
values: bell (alias for sound — explicit per iTerm2's vocabulary),
inject (writes action_value directly to the focused PTY — useful
for "yes\n" auto-confirm rules), and run_script (broadcasts to
plugins as a palette_action event with payload
"<action_id>\t<matched_substring>"). Existing notify / sound /
command actions unchanged.ESC ] 1337 ; SetUserVar=NAME=<base64-value> ST (iTerm2 / WezTerm
convention) and fires a user_var_changed plugin event with payload
"NAME=value". Disambiguated from inline images by the byte after
1337; (S for SetUserVar, F for File). NAME capped at 128 chars,
decoded value capped at 4 KiB — this is a status channel, not a
transport. Lets a shell hook pipe state (git branch, k8s context,
virtualenv) into plugins without prompt-text scraping.ants.palette.register({title, action, hotkey}). Lua API for
plugins to inject Ctrl+Shift+P palette entries. title and action
required; hotkey optional (when set, registers a global QShortcut
that fires the same action). Entries appear as "<plugin>: <title>"
to keep the palette scannable across plugins. Triggering an entry
fires a palette_action event with action as payload, scoped to
the registering plugin only — broadcast events use a different
payload format ("<id>\t<match>" from run_script triggers).command_finished — fires on OSC 133 D; payload
"exit_code=N&duration_ms=N" (URL-form so plugins can split on
& without escaping).pane_focused — fires on tab switch; payload = tab title. Today
fires on tab changes; will cover within-tab pane focus when split-
pane focus tracking lands.theme_changed — fires after applyTheme() completes; payload =
theme name. Lets plugins swap palette/icon assets to match.window_config_reloaded — fires after Settings → Apply; payload
empty (plugins re-read settings via ants.settings.get on demand).user_var_changed — fires on OSC 1337 SetUserVar (see above).palette_action — fires when a registered palette entry is
triggered, or when a run_script trigger matches.TerminalGrid callbacks. New setCommandFinishedCallback,
setUserVarCallback. Wired by TerminalWidget to corresponding
Qt signals (commandFinished, userVarChanged, triggerRunScript)
which MainWindow forwards into PluginManager::fireEvent.MainWindow::rebuildCommandPalette()
collects menu actions and plugin-registered entries on every change
(plugin load/reload/unload, individual register calls). On
pluginsReloaded signal, all plugin entries are torn down so a
removed plugin's entries don't survive across reloads — init.lua
re-runs on reload and re-registers anything that should still appear.palette API,
the instant flag, the new trigger action types. The 0.7 roadmap
section in PLUGINS.md is updated to reflect what's now shipped vs
what remains.HighlightLine / HighlightText /
MakeHyperlink actions from the original ROADMAP item are deferred
— they require grid-cell mutation (per-cell color overrides + OSC 8
hyperlink injection from outside the parser) which is a bigger
surgery than the dispatch-level actions shipped here. The actions
shipped (notify, sound/bell, command, inject, run_script,
PostNotification via notify) cover the iTerm2 trigger doc's
most-used cases. Grid-mutation actions are tracked as a follow-up.pane_focused fires on tab switches today; once split-pane focus
tracking lands, the same event covers within-tab pane changes —
plugin handlers don't need to change.Theme: close the two remaining Qt-specific audit rules from 0.7.0 §Dev experience. With these in place, the entire 0.7.0 "Project Audit tool" lane is shipped — every motivating case surfaced by the 0.6.5 audit pass now has automated coverage.
unbounded_callback_payloads. Same-line regex flags
PTY / OSC / IPC byte buffers forwarded straight into a user-supplied
*Callback(…) without a length cap (.left() / .truncate() /
.mid() / .chopped() / .chop()). Motivating case: pre-0.6.5
OSC 9 / OSC 777 notification body shovelled the entire escape payload
(potentially MB) into the desktop notifier, which then crashed the
notification daemon and/or amplified a malformed-OSC DoS. Fix shipped
in 0.6.5; the rule prevents the regression. Severity Major.
terminalgrid.cpp is the canonical safe call site (filtered at
runtime by .left(kMaxNotifyBody)); rule produces zero findings on
our tree.qnetworkreply_no_abort. Detects the dangerous shape
from the pre-0.6.5 AiDialog incident: a 3-arg connect() to a
QNetworkReply signal whose third argument is a bare lambda. With no
context object, Qt cannot auto-disconnect when the lambda's captured
this is destroyed — owner closed mid-flight → reply fires later →
use-after-free. Pattern enforces the 4-arg form (sender, signal,
context, slot), which Qt protects via auto-disconnect on context
destruction. Severity Major. Covers readyRead, finished,
errorOccurred, sslErrors. Single-line only — multi-line connect
formatting is a known false-negative (rare in practice). Ants's own
AiDialog and AuditDialog use the safe 4-arg shape and are not flagged.bad.cpp with three @expect
markers each, good.cpp with the corresponding safe shapes that must
not match the regex. Wired through audit_self_test.sh so CI catches
regressions on every push (now 14 rules, 46 total checks pass).unbounded_callback_payloads,
qnetworkreply_no_abort) from 📋 to ✅. The Project Audit tool
lane in 0.7.0 is now fully shipped; remaining 0.7.0 work is
features (command blocks, asciinema), the trigger system, and the
performance / platform / security items.Theme: close the remaining grep-shaped items from 0.7.0 §Dev experience
in one sweep, plus fix the CI workflow file which — it turns out — had
never parsed (unquoted colon in a step name rejected the entire YAML).
silent_catch. Flags empty same-line catch (...) {}
handlers that swallow exceptions without logging or rethrow. Conservative
first cut (same-line empty body only); extending to multi-line / single-
statement trivial bodies needs grep -Pzo plumbing and is deferred.missing_build_flags. Nudges projects toward better
compile-time coverage by flagging absence of -Wformat=2, -Wshadow,
-Wnull-dereference, and -Wconversion in the top-level CMakeLists.txt.
Strips comment-only lines before matching so a CMakeLists that discusses
a flag in a comment (e.g. why it's disabled) doesn't cause a false negative.
Severity Minor — this is a nudge, not a bug. Correctly reports one finding
on our own tree: -Wconversion, which is intentionally off (noisy in
combination with Qt), documented in CMakeLists.txt.no_ci. Detects projects missing CI configuration
(.github/workflows/, .gitlab-ci.yml, .circleci/, .travis.yml,
Jenkinsfile). Severity Major — regressions ship silently without a CI
safety net.build-asan). New parallel GitHub Actions job
builds with -DANTS_SANITIZERS=ON (ASan + UBSan), runs ctest under
sanitizers, and smoke-tests the binary (--version, --help) with
QT_QPA_PLATFORM=offscreen so initialization-path bugs surface on every
push. detect_leaks=0 for now — Qt global singletons on fast-exit paths
look like leaks to LSan; re-enable once we have a real unit-test suite
exercising full app lifecycle.CONTRIBUTING.md. Short actionable guide derived from STANDARDS.md:
build modes, where tests live, step-by-step for adding an audit rule,
version-bump checklist, commit conventions, what-not-to-send.silent_catch — bad.cpp with three @expect markers,
good.cpp with logging/rethrow/return-with-value catch bodies that must
not match.name: Run cppcheck (audit rule: Qt-aware static analysis) had an unquoted : inside the scalar, which
GitHub Actions' YAML parser rejected as "mapping values are not allowed
here" — the workflow file was dropped entirely and no jobs ran for 0.6.5
or 0.6.6. Fixed by double-quoting the step name. The parse error was
silent on GitHub's side (workflow-file runs showed 0 jobs with a
one-line "workflow file issue" notice); caught here by running PyYAML's
strict parser over the file locally.Theme: close the first Dev-experience item queued in 0.7 after the 0.6.5 audit. One new audit rule + matching CI enforcement, backfill the two fixtures it immediately surfaced.
addGrepCheck. New audit_fixture_coverage
rule in the Project Audit dialog enumerates every addGrepCheck("id", …)
call in src/auditdialog.cpp, dedups by id, and reports any id without a
matching tests/audit_fixtures/<id>/ directory. Catches the exact gap we
hit in 0.6.5 — four rules shipped a cycle without regression fixtures.
Silent no-op on projects that don't follow this convention (grep yields no
ids). Output uses the parseFindings-friendly file:line: message shape so
findings link directly to the registration site.tests/audit_self_test.sh.
Belt-and-suspenders: the runtime check surfaces findings to the dev on next
audit run; the test-harness assertion blocks merges in CI when a new rule
id lacks either a fixture dir OR a run_rule line in the script. Mirrors
the runtime check exactly. Added to ctest output as
PASS/FAIL: fixture-coverage <id>.memory_patterns and qt_openurl_unchecked —
the two real gaps the new rule just surfaced. Each gets bad.cpp with
@expect markers and a good.cpp that avoids matching tokens (including
in comments — the pattern is case-sensitive and line-based, so comment
wording had to use UPPER-CASE or non-literal phrasing).addGrepCheck" item
from 0.7.0 §Dev experience to 0.6.6 shipped.Theme: second audit of the day. Ran the 5-phase prompt
(/mnt/Storage/Scripts/audit_prompt.md) against the post-0.6.4 tree. Five
subagent reports converged on three real High findings, three Medium, and a
batch of noise. Five subagent claims were rejected on Phase 4 verification
(documented in the commit body) — keeping the signal honest matters more than
the count.
handleOsc() was forwarding that whole
payload to the notification callback as a QString. A program inside the
terminal could spam the freedesktop notification daemon with multi-MB
titles. src/terminalgrid.cpp:761-790 now .left(N)-clamps before the
callback, matching the OSC 52 clipboard-cap pattern.m_sseLineBuffer already had this cap on the pipe side (src/aidialog.cpp:179);
m_streamBuffer — which holds parsed content — did not. A misbehaving
or hostile endpoint could stream past max_tokens indefinitely. Once
capped, we append a single [response truncated] marker and drop
subsequent chunks.readyRead / finished on a partially-destructed AiDialog. Explicit
abort() + deleteLater() closes the window.-Wformat=2, -Wshadow=local, and
-Wnull-dereference to the default warning set. -Wshadow=local was
picked over the full -Wshadow deliberately: the project uses
Qt-idiomatic void setData(QByteArray data) patterns where parameter
names legitimately match member accessors; flagging those would drown
the real int x = …; if (cond) { int x = …; } shadowing that the flag
exists to catch. Surfaced two legitimate local-vs-local shadows
(mainwindow.cpp:1170 and auditdialog.cpp:3062), both renamed.catch (…) blocks in OSC handlers now route to the existing
DBGLOG channel. Three sites in src/terminalgrid.cpp (OSC 133 D,
OSC 9;4, OSC 1337 param parse) write a diagnostic line when their
std::stoi throws, guarded by m_debugLog so there's zero cost when
debug isn't enabled. The two Kitty-graphics safeStoi/safeStoul
helpers stay intentionally silent — documented with a comment — because
they fire per-chunk in the APC parameter loop and logging would flood..github/workflows/ci.yml. Runs ctest plus the cppcheck Qt-aware
pass on every push to main and every PR. actions/checkout pinned by
40-char SHA (not tag) per STANDARDS.md §Supply chain. Workflow token
scoped to contents: read.tests/audit_fixtures/: todo_scan, format_string, hardcoded_ips,
weak_crypto. These four rules ship in auditdialog.cpp but had no
bad/good fixture pair, so regex drift would have shipped silently.
tests/audit_self_test.sh now covers 9 of the 9 unconditional
addGrepCheck() rules; ctest exercises 27 assertions (up from 23).launch.sh now starts with set -eu. Three-line wrapper, but free
hardening against silent failures if a typo lands in a future edit.-Wl cert-err33-c fix. src/ptyhandler.cpp:112's argv0
::snprintf return value was being ignored; now (void)-cast with a
comment explaining the truncation is a known-acceptable loss (the
written buffer is a login-marker, not a lookup key).Theme: Project Audit signal-to-noise. Addresses a reviewer-provided brief after a 2026-04-14 audit run surfaced 26 BLOCKER findings that were all false positives from ASCII section underlines in a vendored single-header library, and 30 MAJOR world-writable-file findings that were a FUSE-mount artefact.
Selection auto-scroll direction was inverted. Dragging a selection
past the bottom edge of the viewport scrolled up into older scrollback
instead of revealing newer lines below (and symmetrically at the top
edge). The auto-scroll timer in TerminalWidget applied
m_scrollOffset - delta with a signed direction, which inverted the
intent; flipped to + delta so the direction convention now matches
wheelEvent / smoothScrollStep — +offset reveals older content,
-offset reveals newer.
Conflict-marker regex no longer matches ASCII underlines. The
conflict_markers rule required exactly 7 sigil chars at start-of-line
but had no trailing anchor, so =========== (and longer banners in
vendored headers) still matched. Tightened to
^(<{7}|\|{7}|={7}|>{7})(\s|$). Also added |{7} for the diff3-style
merge-base marker that the old pattern missed entirely. A
conflict_markers self-test fixture now carries eight ASCII-underline
canaries in good.cpp so regressions are caught by the existing CTest
driver.
World-writable check on non-POSIX mounts. On filesystems that don't
enforce POSIX permission bits (FUSE, NTFS, vfat/exfat, CIFS, 9p) every
file reports as world-writable because the mount maps all Unix perms
to 0777. AuditDialog now probes stat -f -c %T <project> at
construction; when the result is in a known non-POSIX list the
file_perms check is replaced with an INFO note explaining the skip,
instead of producing ~every file in the tree as a finding.
Default exclude list broadened. kFindExcl and kGrepExcl now
exclude external/ and third_party/ in addition to the existing
vendor/, build/, node_modules/, .cache/, dist/, .git/,
.audit_cache/. Vendored code is not project-maintained, so findings
there aren't actionable and used to drown out real signal.
Category headers surface multi-tool corroboration. The HTML and
plain-text renders now append (N corroborated) to each category
header when one or more of its findings are flagged by ≥2 distinct
tools on the same file:line — the same signal that drives the ★ badge
and boosts the confidence score. Lets the downstream consumer
prioritise triage at a glance instead of having to scan every finding's
inline tags.
Confidence legend in the plain-text header. The conf N and ★
tags were opaque without documentation. A four-line legend now sits
in the handoff header explaining what the score means and how
corroboration feeds into it.
/tmp/ants-audit-*.txt export used by the "Review with Claude"
button now carries a BASELINE → VERIFY → CITATIONS → APPROVAL →
IMPLEMENT+TEST prompt between the project-docs block and the
findings list. Stops the downstream session from plunging straight
into fixes without verifying findings are real, cross-checking CVE
version ranges, or committing a regression test per fix. Aligns with
the project's existing no-workarounds rule (documented in
CLAUDE.md).Theme: Claude Code status responsiveness. Replaces the 2-second transcript poll with an inotify-driven event pipeline; state transitions now surface in ~50ms instead of up to 2s, with lower steady-state CPU.
Transcript state updates are event-driven. ClaudeIntegration already
had a QFileSystemWatcher on the active transcript but bypassed it with
an unconditional per-poll parse ("QFileSystemWatcher can miss rapid
changes on Linux"). That was over-defensive — QFSW's only real miss
modes are atomic rename-over (already handled by the re-add in
parseTranscriptForState) and full file replacement (now handled by a
slow backstop). The watcher's fileChanged signal now drives a 50ms
single-shot debounce timer that coalesces streaming-output bursts
(during an assistant turn Claude writes dozens of JSONL lines per
second) into at most ~20 parses/sec. Net effect: "Claude:
thinking…/compacting…/idle" transitions flip in ~50ms instead of up to
2000ms, and the CPU no longer re-parses 32KB of JSON every 2 seconds
when nothing has changed.
Process-presence poll kept at 2s, with a 20s transcript backstop.
Polling /proc for claude-code starting/stopping under our shell has
no clean event equivalent (we're not the direct parent, and netlink
proc-connector needs elevated caps), but it's cheap — ~40 bytes read
per cycle. The poll now also re-parses the transcript once every 10
cycles (~20s) as a defensive net for the rare file-replaced case where
QFSW loses its watch; the re-parse also re-arms the watch via the
existing addPath-if-missing check.
Theme: Claude Code status readability. Adds a dedicated "compacting" state
so the status bar doesn't flatten a multi-second /compact into generic
"thinking…".
ClaudeState::Compacting surfaces
when a /compact is in flight, rendered in magenta so it's visually
distinct from idle (green), thinking (blue), and tool use (yellow).
Detection is transcript-driven: the parser walks the existing 32KB tail
window looking for a user event whose string content contains
<command-name>/compact</command-name>, and only fires while no
subsequent isCompactSummary:true user event (the condensed-history
marker Claude Code writes when compaction finishes) has appeared. The
PreCompact hook path now routes through the same state, so terminals
that have the hook server wired up get the same indicator without the
2-second poll latency.Theme: scroll-anchor correctness. One-liner fix for a long-standing drift bug that only surfaced with a full scrollback buffer.
onPtyData diffed
scrollbackSize() before/after parsing, but once the buffer hits
scrollback_lines (default 50k) each push also pops a stale line from
the front, so the size delta is always zero and the anchor thinks
nothing happened. TerminalGrid now exposes a monotonic
scrollbackPushed() counter that keeps incrementing through the cap;
TerminalWidget diffs that instead. Below-cap behavior is unchanged.Theme: make the two features that already make Ants distinctive (plugins + audit) production-grade. Ships the full context polish + plugin v2 arc from the 0.6 roadmap — scrollback regex search, OSC 9;4 progress, multi-line paste confirmation, OSC 8 homograph warning, LRU glyph-atlas eviction, cell-level dirty tracking, per-plugin Lua VMs, manifest v2 with declarative permissions + capability-gated APIs, plugin keybindings, and hot reload.
.* (regex mode via QRegularExpression), Aa (match case),
and Ab (whole word, wraps the pattern in \b…\b). Alt+R / Alt+C /
Alt+W flip toggles without leaving the input. Invalid regex patterns
render the input with a red border and show the error in the match-
counter tooltip (!/!). Zero-width matches (\b, ^) are detected
and skipped to avoid infinite loops. Matches Kitty / iTerm2 / WezTerm
behavior; closes the Ghostty 1.3 gap.ESC]9;4;state;percent ST sequences. State 0 clears, 1 shows a normal
blue progress bar, 2 shows a red error bar, 3 shows an indeterminate
(full-width) lavender bar, 4 shows a yellow warning bar. Renders as a
3-pixel strip along the bottom edge of the terminal (above the
scrollbar) and as a small colored dot next to the tab title. Coexists
with the existing OSC 9 desktop-notification handler via payload
disambiguation (9;4;… vs. 9;<body>). Spec:
MS Terminal progress sequences.ESC [ C / ESC [ D arrow sequences to reposition the cursor. Capped
at 200 arrows to guard against runaway sends.glrenderer.cpp:compactAtlas.TermLine gains a dirty flag set by
every grid mutation (print, erase, scroll, alt-screen swap, resize).
TerminalWidget::invalidateSpanCaches() now does targeted eviction —
only URL / highlight caches for dirty lines get dropped, not the whole
map. Big win on high-throughput output that only touches a few lines.
(Full cell-level render-path partial-update is deferred; the dirty
primitive is in place for that work.)sudo , a curl/wget/fetch
piped into sh/bash/python/etc., or any non-TAB/LF/CR control
character. Policy is independent of bracketed paste — iTerm2 /
Microsoft Terminal got bitten by conflating the two
(MS Terminal #13014).
Config key: confirm_multiline_paste (default true). UI:
Settings → Behavior → "Confirm multi-line / dangerous pastes".github.com) that doesn't match
the actual URL host (github.evil.example.com), a confirm dialog
shows both and requires explicit opt-in. Strips leading www. for
fair comparison.lua_State, heap
budget (10 MB), and instruction hook (10 M ops). One VM stalling or
leaking globals can no longer affect others. PluginManager owns
a map of (name → LuaEngine *) and fans out events to each in turn.manifest.json. Plugins declare a
"permissions": ["clipboard.write", "settings"] array. On first
load, a permission dialog lists each requested capability and lets
the user accept / un-check individual grants. Grants persist in
config.json under plugin_grants; subsequent loads don't re-prompt
unless the manifest requests a new permission. Un-granted permissions
surface as missing API functions (not nil stubs) so plugins can
feature-detect with if ants.clipboard then … end.ants.clipboard.write(text) — gated by "clipboard.write"
permission. Writes to the system clipboard via QApplication.ants.settings.get(key) / ants.settings.set(key, value) — gated
by "settings" permission. Per-plugin key/value store, persisted
under plugin_settings.<plugin_name>.<key> in config.json. Manifest
can declare a "settings_schema" JSON-Schema subset (plumbing in
place; schema-driven Settings UI is a follow-up item)."keybindings" block:
{"my-action": "Ctrl+Shift+K"}. Firing the shortcut dispatches a
keybinding event to the owning plugin with the action id as the
payload. ants.on("keybinding", function(id) ... end) listens. The
shortcut lives on the MainWindow; invalid sequences log a warning.QFileSystemWatcher. Watches init.lua,
manifest.json, and the plugin directory itself; on change, debounced
150 ms, the plugin VM is torn down (fires unload event) and
re-initialized (fires load event). Enabled only when
ANTS_PLUGIN_DEV=1 — idle cost is zero otherwise.ants._version exposes ANTS_VERSION string so plugins can
feature-detect without hardcoding.ants._plugin_name exposes the plugin's declared manifest name.load, unload (lifecycle), keybinding (manifest
shortcuts). Added to eventFromString() in luaengine.cpp.ants-terminal --new-plugin <name> CLI. Creates a plugin skeleton
at ~/.config/ants-terminal/plugins/<name>/ with init.lua,
manifest.json (empty permissions array), and README.md templates.
Validates the name against [A-Za-z][A-Za-z0-9_-]{0,63}.ANTS_PLUGIN_DEV=1 env var enables verbose plugin logging on
scan/load, plus hot reload. Cached per-process via a static.--help / --version flags wired through QCommandLineParser
(previously --version set via setApplicationVersion but no CLI
handler existed).README.md — navbar now links to PLUGINS.md, ROADMAP.md,
and CHANGELOG.md. Plugins section points to PLUGINS.md as the
source of truth; keeps a quick-start summary table inline.CLAUDE.md, STANDARDS.md, RULES.md — cross-linked to the new
docs; plugin-system sections defer to PLUGINS.md for the author
contract and document internal invariants separately.PluginManager no longer exposes a single engine() accessor;
callers use engineFor(name) to get the per-plugin VM. Event
delivery fans out over all loaded engines.TerminalWidget::confirmDangerousPaste
before the bracketed-paste wrap — confirmation policy is
independent of PTY-side behavior.Project Audit — context-aware upgrade (this release's headline feature). The audit pipeline now collects and applies context at every stage, closing the gap with commercial tools (CodeQL, Semgrep, SonarQube, DeepSource).
// ants-audit: disable[=rule-id] (same line),
disable-next-line, and disable-file (first 20 lines). Foreign markers
are also honored for cross-tool compatibility: clang-tidy NOLINT /
NOLINTNEXTLINE, cppcheck cppcheck-suppress, flake8 noqa, bandit
nosec, semgrep nosemgrep, gitleaks #gitleaks:allow, eslint
eslint-disable-*, pylint pylint: disable. Rule-list parsing and
glob matching (google-*) are supported.moc_*, ui_*, qrc_*,
*.pb.cc/*.pb.h, *.grpc.pb.*, flex/bison outputs, /generated/
paths, and *_generated.* files are dropped before rendering.audit_rules.json. A new path_rules[] block
takes entries of {glob, skip, severity_shift, skip_rules[]} to
configure per-directory severity shifts and skips (e.g. tests/**
shifts everything down one severity band; third_party/** skips
entirely). Glob syntax supports **, *, ?.file:line finding,
cached per-file, rendered with the finding line highlighted. Exposed
via SARIF physicalLocation.contextRegion, embedded in the HTML
report's expandable [details] panel, and included in the Claude
review handoff.(file, line) with a 2 s timeout;
auto-disabled for non-git projects. Exported via SARIF
properties.blame bag (sarif-tools de-facto convention).highConfidence flag
with a weighted sum: severity × 15 + 20 cross-tool agreement + 10
external AST tool − 20 test-path, clamped, adjusted by explicit AI
verdicts. Displayed as a coloured pip in the UI; exported in SARIF
(properties.confidence) and HTML.[details] per finding. Click to reveal snippet
context and an inline "🧠 Triage with AI" action. Expanded state
persists across re-renders.semgrep is available, a
Security-category check runs p/security-audit plus language packs
(p/c, p/cpp, p/python, p/javascript, p/typescript). Picks
up project-local .semgrep.yml automatically. Complements clazy's
Qt-aware AST lane.{verdict, confidence, reasoning}; verdict badge rendered
inline, confidence score adjusted accordingly. Opt-in per click,
respects Config::aiEnabled.driver.version, HTML report meta line, and plain-text
Claude handoff now all carry ants-audit v<PROJECT_VERSION> pulled
from a single CMake add_compile_definitions(ANTS_VERSION=…) source
of truth.CMakeLists.txt project(... VERSION 0.5.0) propagates via
add_compile_definitions(ANTS_VERSION=…) to every caller. Four
previously-hardcoded "0.4.0" literals retired: main.cpp
(setApplicationVersion), auditdialog.cpp (SARIF driver),
ptyhandler.cpp (TERM_PROGRAM_VERSION env), and
claudeintegration.cpp (MCP serverInfo.version).audit_rules.json schema extended with path_rules[]. Existing
rules[] entries load unchanged; loader documents both blocks.physicalLocation.contextRegion with the
full ±3 snippet for every finding that has a file:line location,
plus properties.blame and properties.confidence. properties
also carries aiTriage when the user triaged that finding..conf-pip, .blame, .verdict, .snippet, .hit).
DATA.version now propagated through the inline payload.[conf N · blame · AI]
tags and a 3-line indented snippet per finding. Header gains a
Generator: line.dropFindingsInCommentsOrStrings still opt-in per check via the
existing kSourceScannedChecks set, but is now augmented by the
inline-suppression scan (applied to every finding with a
file:line). Pipeline order documented in STANDARDS.md.ai_enabled config and
refuses non-http/https endpoints.tests/audit_self_test.sh extended with 18 suppression-token
assertions covering every honored marker (ants-native,
NOLINT/NOLINTNEXTLINE, cppcheck-suppress, noqa, nosec, nosemgrep,
gitleaks-allow, eslint-disable-*, pylint: disable) plus negative
cases. Total: 23 rule tests passing.ANTS_VERSION compile definition wired via CMake
add_compile_definitions. main.cpp and auditdialog.cpp no
longer carry hardcoded version strings.README.md — expanded "Project Audit Dialog" section with examples
of inline suppression directives and audit_rules.json schema
including path_rules; semgrep added to optional dependencies.CLAUDE.md — pipeline order, confidence formula, generated-file
patterns, new design decisions around context-awareness.STANDARDS.md — context-awareness pipeline order invariants, blame
cache rules, AI triage config sourcing.RULES.md — three new audit rules covering inline-suppression
preference, path_rules vs. hardcoded paths, and "confidence is
display-only, not a gate".Project Audit dialog — full feature arc. Over the course of eight incremental commits, the audit panel evolved from a regex-heavy scanner into a structured, AST-aware, cross-validated tool.
file:line:checkId:title)..audit_cache/baseline.json; later runs highlight only new findings..audit_cache/trend.json (last 50 runs) and rendered as a delta
banner against the previous run.git diff --unified=0 HEAD~N.file:line by ≥2 distinct tools.connect-3arg-lambda,
container-inside-loop, old-style-connect, qt-keywords, etc.)
via clazy-standalone, reading the project's
compile_commands.json. Retires three FP-prone regex checks.//, /* */, or "strings".partialFingerprints, and per-finding
properties bag..audit_suppress (JSONL v2: {key, rule, reason, timestamp} per line).<project>/audit_rules.json. Flat
schema mirrors the internal AuditCheck struct. User rules run
through the full filter / parse / dedup / suppress pipeline.tests/audit_self_test.sh with per-rule
bad.*/good.* fixtures using // @expect <rule-id> markers;
count-based assertion.CLAUDE.md, STANDARDS.md, RULES.md, and CONTRIBUTING.md
prepended so Claude can weigh findings against documented rules..audit_suppress upgraded from v1 plain-key-per-line to v2 JSONL;
v1 still loads, first write converts in place with a migration
marker.audit_rules.json), not YAML —
QJsonDocument is built-in; flat schema's readability gap with
YAML is small enough that a parser dependency isn't worth it.audit_rules.json is a trust boundary — its command field runs
through /bin/bash unconditionally. Documented analogous to
.git/hooks: your repo, your commands, no sandbox.QDataStream +
qCompress binary serialization.ssh connection.ants API, instruction-count
timeout, event-driven handlers.TerminalGrid with scrollback + alt screen, PTY via
forkpty, QPainter rendering with QTextLayout ligature shaping.QSplitters, 11 built-in themes.