The problem that turned out not to exist, and Games Hub reaches 1.0

On time this week.

A correction

Last week I wrote that Pressless could not finish its first publish. The reasoning went: publishing sends one upload per changed file, GitHub allows about 500 such requests an hour, and a real twelve-year archive is around 862 files — so the upload runs out of budget after about eight minutes, and retrying starts from the beginning.

That was wrong, and I am glad it was tested rather than fixed.

The test was the dull kind: make a throwaway repository and upload files to it until something breaks. 550 files went up inside one hour, one after another, and every single one was accepted — at a sustained 136 a minute. That is past both of the limits Google can find you in GitHub's own documentation, 80 a minute and 500 an hour, with nothing complaining at any point. File uploads are simply not counted against the budget I thought they were counted against. They fall under a much larger one that 862 requests sits comfortably inside.

Two honest caveats. The run reached 550, not 862 — past 500 the published limit is disproved, and the rest follows from that rather than from having watched it finish. And a longer second run was refused by this machine's own rules about hammering someone else's service, which is fair enough.

The more useful half is what the test said about the fix I was going to write. The plan was to send the whole site as one big request instead of one per file. Measured the same day: hand GitHub the text of a photograph that way and it stores the text, literally, rather than the picture — and raw image bytes cannot go in that kind of request at all. So the fix would have quietly broken every photograph on the site while looking like an improvement. A problem that did not exist, about to be solved by a fix that did not work.

What the week actually gave Pressless is smaller and real: the first publish now takes about seven minutes instead of about fourteen. It had been pausing a full second between every file out of caution about a limit that turns out not to apply; it now pauses half a second, which is comfortably under the rate measured as acceptable. The three requests that finish a publish keep the full second, because those are the ones the limits are plausibly about and nobody has measured them.

Also fixed: a stray file in your site folder — a .git directory, an editor's backup, a shortcut — now stops the publish and says which file it was, instead of being uploaded to your live website. And five different reasons a publish can refuse now say which one they were, where before they all produced the same sentence.

Pressless still has nothing to download. Of the four things standing between it and a first release, three have not been started.

Games Hub 1.0.0

Games Hub shipped 0.6.0 on Sunday and 1.0.0 on Tuesday. It is the first of these projects to reach a 1.0, and the 62-entry backlog I mentioned last week is gone — 0.6.0 carried 71 changes on its own.

What 1.0.0 is actually about is a slightly odd thing to celebrate, and it is the same shape as everything else in this post. The build had been asking the compiler not to mention any problems. Compilers will tell you about suspicious code for free if you let them; this one had never been asked. Being asked, it immediately found three real faults and — this is the part that matters — no noise at all. Then a second compiler found six pieces of dead code, one of which was a second copy of Pyramid's redeal limit, sitting next to the one the rules actually use. Two numbers, one rule.

The Windows compiler had a related problem: it was allowed to object but its objections did not stop anything. It had 28 complaints saved up. The reason nobody noticed is worth writing down — the script used to test on Windows locally was configured to switch that setting off, while the automatic build switched it on. Two people looking at the same project and disagreeing about whether it was strict.

Downloads are now checkable. Every published file carries a signed record of which repository, which commit and which build produced it, and you can verify it yourself. Releases before 1.0.0 fail that check, because the record did not exist yet.

Two more from 0.6.0 I liked. A self-test that only tested positions somebody had thought of: four Canasta bugs in one week were all legal moves the game wrongly refused, all involving wild cards, and every one of them sailed through the test suite. It now builds hundreds of random positions and holds the game to two rules out of the actual rulebook — and rather than declaring victory when it found nothing, a fault was deliberately introduced to prove the check bites. And the screenshot comparison had been measuring the shuffle: without a fixed deal, comparing before-and-after pictures compared two different games.

A from-scratch rebuild also went from about a minute and a half to about two seconds.

Next up, and named as the priority: playing every game without a mouse.

428 tests, and a feature that had never once worked

Contact List was listed as quiet last week. It was not quiet this week.

One day of going through it properly found that Google Sync had never worked in a single downloadable copy of the app. Open the sync page in any packaged build and you got an error page. The piece of Google's code it needs was never included in the package.

All 428 tests passed. Every automatic check passed. They pass because they run the app from the source folder on a machine that already has that Google code sitting around — so the missing piece is only missing for you. Nothing found it except someone downloading the actual app and clicking the actual button.

The Windows build had the same shape of hole twice over. Windows does not ship the world's list of time zones, and the substitute was never bundled — so the time zone list in Settings was empty, saving a time zone always failed, and every date in the app fell back to a raw machine timestamp. Two documented features, dead on a documented platform, silently.

And two of the gates meant to catch things like this were reporting success without running. The check that runs before every push was throwing away the answer to "what changed?", getting an empty list, concluding there was nothing to push, and skipping the whole thing — on any ordinary push. The other printed CI PASSED after its own setup step had failed.

The rest of that day: your typed search terms were being written to a log file on disk, because the search text travels in the web address and the log recorded whole addresses. Notes with more than one line lost everything after the first line when re-imported. Phone numbers and emails with custom labels from a Google or Apple export were dropped without a word. Contacts the importer refused were counted as neither imported nor skipped, so a file where most records failed still reported plain success. Disconnecting from Google deleted the local key and left your permission live on the Google account. And the duplicates page counted ticked boxes rather than contacts, so "delete 5 selected" could be followed by a deletion of 3.

Twenty fixes, none of them released — Contact List is still on 1.1.0 from July, and there are now two months of work sitting behind that.

UT Ants can read your maps

UT Ants was announced last week with nothing behind it. Five days later it reads a real Unreal Tournament 99 map from end to end without any part of the original game: the walls and their textures, the sounds, everything placed in the level, which switch opens which door, and where a player can walk.

It can also work out what a custom object is by tracing its family tree across files, and it can carve a level into rooms and answer "which room is this point in". There is a small command-line tool that will tell you what a map depends on and what is in it. About 220 automated checks, passing on a machine that has never had Unreal Tournament installed.

The best find is one no test could have caught. Deep in the map format is a tree that answers "which room am I in", and each branch has a front and a back side. The code was reading those two the wrong way round. Swap two neighbouring things of the same size and the file still reads perfectly; and because the test files were written by the same code that read them, a test agreed with itself whichever way round it was. It was found by checking the answers against what the real game's own files say. Before the fix, 0 of 11,451 answers on one map were right. After, 11,406. Across a whole installed library, nearly 22 million disagreements became about 30,000.

Those remaining 30,000 — 0.27% — are still unexplained, and four theories about them have been tested and eliminated. The target was relaxed from "no disagreements" to "under 1%", with the reason written down: reading it backwards scores zero, not almost-perfect, so a 1% ceiling still catches that kind of mistake three hundred times over.

Two other numbers. A tool that reads all 837 packages of a real Monster Hunt library takes 67 seconds cold and 14 warm, against 48 seconds to boot the old server for one map. And 587 maps were reported as depending on a package called "Base" that does not exist — the code was following a reference one step short. Following it all the way, unmet dependencies across 740 maps went from 732 to zero.

Also measured: the flags in a map that say which surfaces are solid are not reliable. Of 202 surfaces flagged solid, 42 actually were in the running game. On one map it was six out of six wrong.

Still nothing to download and nothing to play. No renderer yet — that is next after the material work now starting.

Ants Terminal 0.7.108

Ants Terminal shipped 0.7.107 on Friday and 0.7.108 today, and 0.7.108 is a large one.

The headline is a measurement of its own review process. A full read-through of the codebase had been silently skipping a third of it — 189 files and 59,257 lines belonged to no section of the plan at all, including the single biggest file in the project. There is now a check that fails the build if any file belongs to no section, or to two.

Its code analyser was in the same state as everything else in this post: run with no configuration, it selects no checks at all, reads every file, reports nothing, and exits successfully. Indistinguishable from a clean project. And one code checker, when it refused to run, printed its own help text — which the audit then read as 92 findings, with contents like --help Display available options. A tool that examined nothing looked exactly like a tool that found problems.

Two more instruments got the same treatment. A new check lists work closed since the last release that no release note mentions: the first run found 19 items, one of them closed the same day as the release it was missing from. Another flags release notes that are word-for-word copies of the internal to-do item — 14 of 56 were, and all 14 were rewritten.

For actual users: typing in the scrollback search no longer freezes the window. Every keystroke re-scanned the entire history — 50,000 lines by default — so typing a word could lock things up for seconds; it now waits for you to pause. A plugin that ran out of memory used to take the whole terminal down, which its own documentation promises cannot happen. Session logs and session recordings were being created readable by every account on the machine, and those files can contain anything the terminal printed, keys included.

The speed numbers are large. Looking up where things are defined across the whole documentation set: 41.8 seconds down to 3.4, with identical answers. Two hundred lookups together: 15.6 seconds down to 0.35. And it had been re-reading its own settings file about eight times a second.

There is a 0.7.109 written and waiting, out as a preview only.

DOOM Ants 0.7.2

DOOM Ants shipped 0.7.2 on Monday — 43 changes, all in one day of work, and mostly the same bounds-checking effort I described last week.

The pick of them: saving reported that it had written past the end of its space after it had already written past the end of its space. The guard ran once every byte was out. On a large custom level that is ordinary play, not an attack. Saves that will not fit are now refused by name, before anything is overwritten. Separately, saving used to destroy the previous save while writing the new one, so an interruption lost both; the new one is now written alongside and swapped in once it is safely on disk.

Two lovely ones. A table of switch graphics had lost its end marker, which left every switch in the game silently doing nothing. And the game had been reading past the end of its sprite-name table on every single launch since forever — nothing ever went wrong, because the value sitting just past the end happened to be the marker it was looking for.

There is also a fix to level loading: working out which walls belong to which room used to re-scan every wall in the map once per room, and now walks it once, producing identical results on all 68 original maps.

Waiting for the next release: the check meant to catch a damaged or mismatched network message had been switched off since 1997 — present in the code, passing everything. Turning it on means network play only works between copies of this version or newer. Saves now record which build wrote them and are refused if it does not match, which makes older saves unreadable — they were previously being misread, which is worse.

finbreak 0.1.23

finbreak shipped 0.1.23 on Sunday, and the headline is a recovery code: a second way into your vault, shown once when you create it, which can unlock the vault and set a new master password. You can decline it, replace it, or remove it later. finbreak keeps no copy — lose it and it is gone, which is the whole point. Existing vaults upgrade themselves on the next unlock, and a copy of the old vault is taken first and only deleted once the upgrade has been checked.

That upgrade is also where the week's most alarming bug lived, and it was caught before release rather than after. The step that checks the new copy is good had three possible situations and only two answers: it said "no" both when the copy was bad and when it could not tell. On "no" it deleted the copy — the known-good one — while a comment beside it asserted that nothing had been touched. It now distinguishes the two, and "I could not tell" keeps everything.

Also in: the time zone you pin now decides what "today" means, not just how dates are printed. I described this last week as fixed; what was fixed then was the display, and this is the rest of it — this month's spending, the alerts and the forecast were all still coming from the machine clock, so pinning Johannesburg and then travelling still moved a whole month of totals on the 1st. And after a crash, two copies of finbreak could open the same encrypted vault and both write to it; only one can claim it now.

On accessibility: the master password box, the recovery code box, the new password and confirmation boxes and the custom date-format box were all announced to a screen reader as having no name at all. Four screens were also ignoring your chosen date format entirely.

demoreel 0.1.1 — first release

demoreel is the tool that records a video of an app running on a private display, so nothing of your real desktop is in shot. It got its first ever release on Tuesday.

Two findings, and both are about the tool's own premise. The private display had no password on it. Measured before the fix existed: another program running as you could read that display and grab a picture of whatever was on it — on a tool that exists specifically so your desktop is not in frame. It is now locked with a credential.

And the code checker had never read a single file of the program. It ran, it printed "All checks passed!", and asked directly which files it had looked at, it named one: its own configuration file. The program is a file called demoreel with no .py on the end, and the checker only looks at files ending .py. The only code in the project was invisible to the only tool checking it, for the project's entire life.

Also: demoreel stop could kill an unrelated program, because a recording killed abruptly leaves its notes behind and the system reuses process numbers. Reproduced by having it kill something innocent. And waiting for an app to finish drawing was taking most of a processor core away from the app it was waiting for — it now backs off as the wait lengthens, and answers a fast-drawing app sooner than before.

Two speed ideas were measured and then not done, which I like as much as the ones that were: one made a six-minute encode 7% slower, and the other produced a marginally worse picture.

Short notes

Local Web Server Manager — still no release, and the count I gave last week was wrong: it is not 77 commits, it is 367 since the project's history starts. Everything it does is still unreleased. This week: the text-size setting silently did nothing on this kind of desktop, while the menu happily ticked your choice and saved it. Stop, Restart and Open are deliberately switched off for servers the app did not start, but nothing said so and — because applying a theme wiped out the desktop's own dimming — a switched-off button looked identical to a live one, so clicking Open just did nothing. Window placement was guessing from a setting that is routinely missing, guessing wrong, moving the window, having the move discarded, and reporting success. The colour tool that picks accessible shades printed its closest failure in exactly the format of a success and ended with "0 shortfalls" — and its output is meant to be pasted straight into the code. The release pre-flight read an unanswerable question as an all-clear and said READY. And every translatable word in the app was invisible to the translation tool: asked to extract them, it found zero.

LottoTracker is dark by default, with thirteen themes and the balls coloured by game — red Lotto, blue PowerBall, green Daily Lotto. The white page was reported as blinding. Dark now lives in the stylesheet rather than being applied afterwards, so there is no white flash on the way in. Every theme was checked against a contrast floor rather than trusted, and two failed on the first run and were corrected. LottoTracker still has no release and no version number at all; what a 1.0 means for it was settled this week, and it means all five of its goals plus packaging.

Quiet this week: Vestige, RetroDB, MAME Curator, perch, RetroArch, Album Builder, Snatch, Slipcase, Rolodex, OneUp and Rusty PSN — most of them because they had a very loud week last week.

What changed on the site

The version line now tells you when that release was cut, not just what it was called — and the date is shown in your date format, not mine. A build running on a machine in another country cannot know how you write dates, so the page ships 2026-09-09, which is unambiguous and sorts properly, and a small script rewrites it into whatever your own system uses once the page is open. If that script never runs, the date is still there and still correct.

Games Hub is now listed as Live rather than Beta, on the strength of 1.0.0. DOOM Ants moves from Early WIP to Beta — its own README has said "playable" for a while, and it ships builds you can download and run, so the old label was underselling it.

New downloads picked up this week: Games Hub 1.0.0, Ants Terminal 0.7.108, DOOM Ants 0.7.2, finbreak 0.1.23 and demoreel 0.1.1.

One thing I checked and could not fix from here: RetroDB's download on this site is still 3.12.0, from June. The app is on 3.23.6 — everything I have written about it for weeks — but those versions were never published as releases, so there is nothing for this site to link to. That is a RetroDB job, not a site one, and it is now on the list.