The first release in maintenance mode, and a security one: the apps take
rustix 1.1.6 (RUSTSEC-2026-0486, reachable through the updater's tar and
xattr code), the About page's data-only report links, and two weeks of
dependency updates; the repository gains Qodana with its first findings
fixed.
Version bumps: web/package.json and its lock, src-tauri/Cargo.toml and the
f2p-tracker entry in Cargo.lock, and tauri.conf.json (Android's versionCode
derives from it).
The first release under the changelog split: [Unreleased] becomes
changelog/v4/4.7.x/4.7.2.md (sub-headings up one level, an intro first,
and the "no release yet" phrases reworded), CHANGELOG.md keeps an empty
[Unreleased] and gains the release's row. The GitHub release's notes are
that file below its title and back link.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
changelog/v<major>/<major>.<minor>.x/<version>.md for each of the 42
releases from v1.0.0 to v4.7.1, and changelog/beta/ for pre-releases
(none has been published; its README says how one would be filed).
CHANGELOG.md becomes the index: what the layout is, [Unreleased] exactly
as it was, and a table per major version (version, date, edition) that
links each file.
Each file is its section, copied byte for byte except that its heading
is the file's H1 (so the sub-headings move up one level, outside code
fences), relative links gain ../../../ so they name the same repository
paths, and the --- separators between sections are gone. Checked by
reversing all three changes on every file and comparing with the
original section: 40 of 40 equal. Two releases had no section: v1.0.0
(its GitHub release notes, which the changelog pointed to as "Initial
stable") and v1.1.3 (its tag message).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README's link, the export-ignore beside CHANGELOG.md's, the watch-paths
note in DEPLOYMENT.md, and a CLAUDE.md rule for where release notes go
and how a release moves [Unreleased] into its own file.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rustix 1.0.0-1.1.5 marks as initialised the bytes an extended-attribute
call (getxattr, listxattr and their l/f forms) never wrote when given a
zero-length buffer - undefined behaviour from safe code. Fixed in 1.1.6
(2026-10-27). In the Tauri shell it comes in through async-io, polling,
zbus and tempfile, and the affected calls through tauri-plugin-updater ->
tar -> xattr. Dependabot raised no alert; an OSV querybatch over
Cargo.lock found it.
Cargo.lock: rustix's version and checksum only. `cargo update -p rustix
--precise 1.1.6` (cargo 1.99) also re-pointed thirteen unrelated
dependency edges to older windows-sys versions and one from syn 3 to
syn 2, so the entry was edited by hand instead; rustix 1.1.6's minimums
(bitflags 2.9.2, errno 0.3.10, libc 0.2.182, windows-sys >=0.52,<0.62)
are all met by what is locked. `cargo metadata --locked` (every platform)
and `cargo test --lib --locked` pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README: a Qodana badge and the workflow in the list of scanners.
DEPLOYMENT: what qodana.yml runs, the three secrets, the gate, and what
to remove if the licence ends. CLAUDE.md: the setup and the traps the
first runs found (no interpreter in native Python, the Rust image's
missing linker and libraries, plugins only in Docker, the lucide import
false positive). CHANGELOG [Unreleased]: the Qodana entry.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud's gate wants 80% coverage on new code, and the PR's changed
lines were at 71.8%: ten of them were never run by a test. Now they are:
- appdetails' failures told apart (no answer, 404, success: false, a
body that is not JSON) and the store page's two network_error paths;
- the leaderboards' stars, for a percentage and for "N/A", "No reviews"
and a missing field (both scripts);
- mark_dead_games.parse_players() on the stored formatted strings;
- backfill_discover.parse_released() on every date shape, the year-only
one included.
pytest: 501 passed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each qodana.yaml now fails its job on any problem (failureConditions,
any: 0), as in free-games-itchio-list, and names what it leaves out and
why:
- Python: DuplicatedCode in the two discovery and two leaderboard
scripts (sharing them is a refactor); broad excepts where a failed
request is "no answer" by design (steam_client, the Access diagnosis);
PyCharm's sorted() typing in edit_game.py; the escaped bracket pair in
normalize.py's `\[hover-add\]`.
- JS/TS: ES6UnusedImports (the Svelte plugin cannot resolve
`@lucide/svelte/icons/*` and calls all 122 icon imports unused; every
tsconfig has noUnusedLocals, enforced by svelte-check and tsc), the
admin's relative imports of the public app, `}` escapes kept for
unicode-mode parsing, fetcher.ts's throw-to-own-catch fallback, and
keyCode 229 for IME composition.
- Rust: CargoUnusedDependency, which cannot see the ::serde_json paths
generate_context!() expands to.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
No behaviour changes; each is what the inspection asked, or the type
information it lacked.
Python (scripts/):
- `return (a, b)` -> `return a, b` (12, steam_client, store_items,
backfill_discover).
- `log=print` parameters annotated `Callable[[str], None]`: PyCharm could
not type the overloaded builtin as a default and called every log(...)
"not callable" (10, budget, purge_unhealthy, refresh_anticheat,
refresh_peaks).
- fetch_full()'s docstring names all its parameters.
- Nested helpers no longer shadow the loop variables of the function
around them (purge_unhealthy judge(), update_reviews ask()).
- The leaderboards' bare `except:` catches what a reviews string can
raise (AttributeError, TypeError, ValueError), not SystemExit and
KeyboardInterrupt too.
- edit_game.py: unused `import json` dropped; mark_dead_games `[^\d]` ->
`\D`; the DLC section regex loses a group around the whole pattern
(only ever used with search()).
TypeScript / Svelte (web/):
- Two comma expressions written as statements (pipeline route's run
dedupe, CandidateDialog's Reopen button).
- shouldIgnore() returns its last condition instead of if/return.
- Resource's constructor-only private fields are readonly.
- The admin mock seed says HOUR, not 1 * HOUR.
Rust (web/src-tauri/): `serde` dropped - nothing in the crate or in what
generate_context!() expands uses it (only `::serde_json` paths are
generated). serde_json stays, with a comment: Qodana calls it unused
too, but the generated code needs it and the crate does not compile
without it.
Checked: pytest 467, typecheck, vitest 747, knip, cargo test --lib.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What the first runs showed, and the changes:
- Python: SDKs.json was empty - native mode finds no interpreter by
itself, and without one every `dict[int, str]` and `X | None` was an
"invalid type hint" (11 of the 50 problems). The job now builds .venv
from the hashed requirements and passes QODANA_PYTHON_PATH.
- Rust: "Analysis scope: 0 packages" and a Critical "Cargo project
loading" sanity problem: the image has no C linker and none of
Tauri's libraries, and runs as the runner's user, who can neither
apt-get install nor write the image's CARGO_HOME. Now `-u root`,
CARGO_HOME in the cached Qodana cache dir, and a bootstrap that
installs the build packages.
- JS/TS: installing the Svelte plugin fails in native mode with every
released CLI up to 2026.2.3 ("qodana.build.date system property must
be set"; fixed upstream as QD-16701 on 2026-11-05, unreleased), so
this project runs in the Qodana image, on the node_modules and
generated files the job prepared.
One workflow (qodana.yml) with a job per project, as in
free-games-itchio-list: the token check covers all three, triggers skip
data and Markdown-only commits, no PR comment (annotations, Qodana
Cloud and the report artifact carry the results), --repository-root .
for the subdirectory projects, and each job keeps its own cache keys.
The qodana.yaml files use `exclude` and name their Qodana Cloud project.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three workflows, one per Qodana Cloud project and token
(QODANA_TOKEN_PYTHON, QODANA_TOKEN_TS_JS, QODANA_TOKEN_RUST), each with
the qodana.yaml of its project directory:
- qodana-python.yml: Qodana for Python over scripts/, native mode on the
runner's Python 3.15 with the hashed requirements installed.
- qodana-js.yml: Qodana for JS over web/, native mode on Node 26 after
`npm ci --ignore-scripts` and `npm run build:admin`, with the Svelte
plugin so .svelte files are read.
- qodana-rust.yml: Qodana for Rust (EAP, Docker only) over web/src-tauri/,
with a stand-in web/dist so generate_context!() expands.
Each runs on pushes to main and pull requests that touch its files, and
by hand. The scan is skipped when its token is not available (forks,
Dependabot), as sonarcloud.yml does; the action is pinned by SHA, the
job's token can read contents and write only checks and PR comments, and
each workflow has its own cache keys so the three do not collide.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dependabot's npm-minor-patch group (#256): @lucide/svelte 1.72.0,
bits-ui 2.21.0, knip 6.45.1, rollup-plugin-visualizer 7.1.3, vite 8.3.8,
wrangler 4.162.0 - all released at least four days ago.
#256 failed Web CI on knip: 6.45 follows tsconfig.json's `extends` and
reports "$app" (from "$app/tsconfig", which `svelte-kit sync` generates
into node_modules/$app) as an unlisted dependency - an error. It cannot
be taken back with ignoreDependencies: knip matches those against a
package name, and getPackageNameFromModuleSpecifier() returns nothing for
a specifier starting with "$", so the item is never compared (and "$app"
would be read as a regex anyway). knip.json now ignores unlisted
dependencies in tsconfig.json only (ignoreIssues). The export backlog and
the three configuration hints are unchanged from main.
Local proof: typecheck, 747 tests, build + verify-dist --web (18,911
pages, precache 177 entries, as live), cf:check, knip, Tauri build +
verify-dist --tauri.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The project is stable: no new features, code changes only for security
advisories and runtime/dependency updates, and issues and pull requests
are for the data (add, correct, remove a game).
- README: status, build, quality, version, licence and runtime badges, a
"Project status" section, and a data-first Contribute section
- CONTRIBUTING, SECURITY, Contact, the PR template and CLAUDE.md state
the same policy
- Issue forms: add_game and edit_game are new; feature_request, feedback
and off_topic are removed; blank issues are off (config.yml)
- About page: "Report something" links the five remaining forms
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Dependabot #249 (coverage 7.17.2), #250 (action-gh-release 3.0.4,
sonarqube-scan-action 8.4.1) and #251 (twelve npm patch and minor
releases) were merged as they were. Nothing security-relevant, so the
entry waits in [Unreleased] for the next release.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The apps go from 2.5.0 to 2.5.1, which is the first app build with the
accessibility fixes from #244. The desktop app is also the first build that
can zoom.
Version bumps: web/package.json and its lock, src-tauri/Cargo.toml and the
f2p-tracker entry in Cargo.lock, and tauri.conf.json (Android's versionCode
derives from it).
The CHANGELOG's [Unreleased] section becomes v4.7.1, with an intro, and a
fresh [Unreleased] goes above it. ACCESSIBILITY.md now says what works for
the website and the apps 2.5.1, instead of "the apps get it with their next
release".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud's gate counted 6 new lines no test had loaded: the overlay state
and focusOnMount (lib/overlays.svelte.ts), and the palette shortcut's new
early return. overlays.test.ts drives them in Node: overlays.open follows
either overlay, focusOnMount focuses without scrolling, and Ctrl+K is left
alone under an overlay (palette shut, default not prevented) but toggles the
palette otherwise.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
src/lib/a11y.test.ts reads the source, like mobile-layout.test.ts. It
checks:
- the table's roles, counts and row indexes, and that no <a> is a row or
carries a fake link;
- every column has a name, and the cards are a list with positions;
- the panes have a minimum height;
- there is no colour-only kernel badge;
- the overlays are dialogs that focus themselves and make the page inert;
- the palette shortcut stands down, and back-to-top is inert while hidden;
- the zoom setting and its capability are present.
It also computes the theme's contrast with the WCAG formula, so a token
edit that drops below 4.5:1 for text or 3:1 for a control's edge fails.
ACCESSIBILITY.md moves the fixed items from "Known limitations" to "What
works today". The desktop and Android apps 2.5.0 predate them and get them
with their next release. CLAUDE.md lists the rules behind each fix, and
CHANGELOG [Unreleased] describes them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Computed with the WCAG formula, against the page, cards, and each colour's
own /12-/15 tint.
Light theme:
- Links, primary button labels, warnings and some badges measured
3.2-4.3:1. --primary goes 42% -> 34% (with --ring), --warning
38 -> 29, --destructive 47 -> 42, --success 32 -> 27 and --info 40 -> 35.
- --input goes from the --border value to 3.2:1, so text boxes and
drop-downs show an edge. --border stays light; card edges carry no
meaning.
Dark theme:
- --destructive 62 -> 67 brings red badges from 4.05 to 4.6.
- --input goes 19 -> 40% for the same 3:1 edge.
Desktop zoom: tauri.conf.json sets zoomHotkeysEnabled (Tauri's default is
false, which also turned off WebView2's zoom). capabilities/zoom.json, which
is desktop-only like updater.json, grants core:webview:allow-set-webview-zoom.
macOS and Linux zoom through a script Tauri injects that calls it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The terms screen was a plain div, rendered last in the page. Focus stayed
where it was, and Tab walked through the whole hidden page before reaching
its buttons. The human check had role="dialog", but neither moved focus
nor kept it.
Both now:
- report themselves to lib/overlays.svelte.ts;
- are role="dialog" with aria-modal, labelled by an h2 (each used to add a
second h1);
- move focus into themselves on mount (focusOnMount).
The layout makes everything behind them inert: the bare pages' wrapper,
the skip link and the shell. BackToTop moves inside the shell to go inert
with it.
Ctrl/Cmd+K stands down while they are open. It could open the palette
UNDER the gate (z-50 against z-[60]) and trap focus there, out of sight.
Also:
- BackToTop is inert while hidden: it only faded out, so it stayed a Tab
stop. Its JS smooth scroll follows prefers-reduced-motion.
- The command palette's search box gets a focus ring, drawn on its row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The games table:
- The rows were <a role="row"> with nothing around them, so a screen reader
read them as neither a link nor a table. The grid now has the complete
ARIA table: role="table" with aria-rowcount/aria-colcount, a rowgroup,
and aria-rowindex on every virtualised row. The icon-only columns have
names for screen readers (srLabel), and the non-sortable "Platforms"
header is plain text rather than a disabled button.
- A row is no longer a link wrapped around a fake one. The name is the
row's only Tab stop, and its link stretches over the row, so a click
anywhere still opens the game. The Steam icon is a real link, kept out
of the Tab order (a game's page has the same links). The tooltips sit
above the stretched link, and screen readers get their words.
- The row draws the focus ring, because the cells clip overflow.
The phone cards are a list with aria-setsize/aria-posinset.
Both panes keep a minimum height (min-h-64). Sized to the window alone,
they shrank to nothing at 300-400% zoom.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Kernel-level anti-cheat differed from the rest only by its badge colour,
red against amber, with the product name as the only text. That told a
colour-blind reader and a screen reader nothing.
games/AntiCheatBadge.svelte adds a shield icon and the words
"Kernel-level". They are visible on a game's page, and read only by screen
readers where a long product name already fills the badge (/random here;
the table and the cards in the next commit).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A read-through against the code found about forty claims that had drifted.
README:
- Dynamic game-count and version badges. The count is index.json's total.
The version is the newest v* release by semver, since an app release
published last would otherwise count as "latest".
- SvelteKit 3, every desktop package, and the shard layout without a last
shard number.
- The media host, which commits deploy, and the nine workflows the table
left out.
- The accessibility statement and form in the link lists.
SECURITY.md:
- Fix the broken bullet list.
- Describe the real stack, and pushes with the built-in GITHUB_TOKEN.
- Add to scope: the media host, /api/ingest/*, /api/updates/desktop and the
avatar origin.
AUTHORS.md and username.txt:
- Donations are in FUNDING.yml.
- The Add-games template is gone.
- Test devices come from pc_spec.
The issue forms lose the hash URLs, the GitHub Pages permalink, and the
editing and GPG options.
web/README.md, dev_env (EN + VI), the acknowledgements and the contact page
are updated to match. The last two are not consent documents, so nobody is
asked to accept the terms again.
CHANGELOG [Unreleased] lists all three commits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The report section gets a fourth form, Accessibility barrier, in a 2x2
grid. Below it are a short note and a link to ACCESSIBILITY.md on GitHub.
The statement is linked rather than rendered under /legal/: it is not a
legal document, and a file rendered there becomes a build input that the
deploy watch paths must list.
The stack list said SvelteKit 2 and a 3,600-row table. The footnote said
everything the page loads comes from this site, which is not true of the
beacon or the human check. The Vietnamese AI note also gets the last
sentence it was missing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ACCESSIBILITY.md sits at the root beside SECURITY.md. GitHub reads it from
.github/, the root or docs/ and shows it as the repository's Accessibility
tab. It covers:
- WCAG 2.2 AA as a target, not a claim.
- What works today and what does not, with a way around each limitation
where there is one.
- The browsers, apps and assistive technology actually used (keyboard,
Narrator, NVDA).
- How to report a barrier without saying anything about yourself.
- Replies on SECURITY.md's terms: first within a week, blockers first, no
fixed deadlines.
- The rules for contributors and coding agents.
.github/ISSUE_TEMPLATE/accessibility.yml files a report under the new
`accessibility` label, with no required final checkbox. The pull request
template is the short checklist, and completes the community profile.
CLAUDE.md says the statement is a claim about the code, and names the two
limitations that look solved but are not.
.gitattributes keeps the new file out of source archives like the other
community files. It also stops leaving out Contact, PRIVACY_POLICY and ToS:
the web build reads every legal document and stops when one is missing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Desktop and Android 2.4.1 -> 2.5.0 (package.json and its lock,
tauri.conf.json, Cargo.toml, the f2p-tracker entry of Cargo.lock); repo
public-facing version 4.6.1 -> 4.7.0. CHANGELOG: the random games page,
Leave site, Decline quitting the Android app, and the 2026-11-03
dependency pass that waited in [Unreleased]. No legal document changed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CLAUDE.md: why /random keeps its picks in memory and only the reader
rolls, the crypto sampling (not Math.random), why the route is top-level,
and the base key a plural needs; how Leave site finds the page before the
site, the watchdog and its three cancel events, the deliberate second click
to close, and why Decline quits through exit(). Both READMEs list the
random page and its modules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Web: a "Leave site" button on the terms card and on the declined notice.
No page may open the browser's own start page (chrome://newtab,
about:home), so lib/leave-site.ts goes back to the page before the site:
history.go(-(navigation.currentEntry.index + 1)) steps over every entry of
the site's contiguous same-origin run at once. history.go() is silent when
there is nowhere to go, so a 1.5 s watchdog (cancelled by beforeunload,
pagehide and popstate) shows "Nothing to go back to"; when every entry is
the site's that shows at once, and closing the tab is a second, explicit
click, because closing the last tab closes the browser window.
Apps: Decline called getCurrentWindow().close(), which on Android resolves
without closing anything (tao reports a destroyed window only from the
Activity's onDestroy), so Decline did nothing there. It now calls exit()
from @tauri-apps/api/app, which finishes the activity on Android and ends
the process on the desktop; the declined notice appears if the app is
still alive 2 s later. default.json trades core:window:allow-close, which
nothing else used, for core:app:allow-exit. The Leave button is web only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/random rolls one to five games from the catalogue (five to start with):
any type, online or offline, dead games only when asked, never one marked
unsafe. Roll again avoids the games on screen while enough others remain;
a smaller count keeps the first picks, a larger one adds to them.
- lib/random-pick.ts: crypto.getRandomValues with rejection sampling and a
sparse Fisher-Yates, never Math.random (SonarCloud S2245), plus the pool
rules, read off appidIndex so each game appears once.
- lib/random.svelte.ts: the picks as appids, module-level and memory only,
so Back from a game keeps them and no f2p:* key is needed (a new key
means a privacy-policy edit, which reopens the consent gate). Only the
reader rolls; a revalidation only drops a pick that left the catalogue
or stopped qualifying.
- The page, an entry in the sidebar, drawer and palette (nav.ts) and a
dashboard button. Top-level on purpose: the header search, isActive and
the PWA's globIgnores all treat /games/* specially. The empty state joins
verify-dist's BAKED_EMPTY_STATES; the sitemap ranks the page low.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Takes Dependabot's npm group (#236) and cargo group (#234) together: the
Tauri shell plugin moves to 2.6.0 on both sides, and the CLI refuses a
JavaScript/Rust pair on different minors. Adds the SvelteKit 3.1.1 and
adapter-static 4.0.1 patches (exact pins kept), @types/node 26.6.6, knip
6.44.1 and @lucide/svelte 1.66.0 - everything at least three days old;
newer releases wait for next week's Dependabot run.
wrangler 4.156.0's miniflare pins undici 7.30.1 itself, so the
"miniflare": {"undici": "^7.30.1"} override goes in the same commit, as
CLAUDE.md required. npm audit: only the known deepmerge low advisory.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- tauri-plugin-updater 2.14.0 -> 2.14.1 moves its Windows zip reader from
zip 4.6.1 to 9.2.0: closes Dependabot #82 (critical), #83, #84 and
RUSTSEC-2026-0367/0368/0369. Targeted cargo update; cargo test --lib ok.
- npm overrides workbox-build -> glob ^14.0.2 and filelist -> minimatch
^11.0.1 (GHSA-3g3m-2fjr-6pw4, GHSA-g3hv-75wg-m249, published 2026-10-30).
The precache lists the same 172 URLs as before.
- deepmerge (#81, low) has no fixed release: build-time only, dismissed.
- CLAUDE.md: when the new overrides and the undici one can go.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
appdetails writes "Free To Play"; SKIP_GENRE_TAGS said "Free to Play", and
the exact match let 1,357 of 12,854 games be stored with it as their genre
(the third most common), while the fallback to the first name could pick
it as well.
- fetcher.pick_genre() compares casefolded and never falls back to a price
model or a release state (NOT_GENRES); with nothing else it says
"Uncategorized". ingest uses it.
- Refresh Store Data replaces a stored one from appdetails: 150 games a
day, first after the page repairs, never under an override. It is the one
manual-field value the job writes; no person chose it. A dry run on the
first 150 (2026-10-30): Casual 91, Indie 24, Strategy 11, Simulation 11,
Uncategorized 10, three others; all answered, no override.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- docs/plan/19-all-languages.md: what kept them out, what every workflow
does with them, the re-sweep, and what is left.
- CLAUDE.md: discovery's language rules, ASCII appids, search folds but
storage never does, re-cut descriptions are not news.
- docs/ADMIN.md: the filters as they are now, re-sweeping a whole year,
the Languages line in the candidate dialog.
- README.md, docs/plan/11 (the decision changed), CHANGELOG [Unreleased].
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Without a byte order mark Excel reads a CSV in the system's legacy code
page, Japanese Windows included, and every Japanese, Chinese or Korean
name came out as mojibake. csvForExcel() prepends the mark; JSON is
unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The catalogue now holds games in every language.
- /games filters by language (?lang=, counted like the others); the names
are Steam's English ones as stored, data rather than i18n keys. About
says the catalogue covers every language.
- fold() (NFKC, then hiragana to katakana) on both sides of every search
box - the Fuse index (one getFn per key: the default getter hands back
objects for array fields), the palette, /developers, /publishers and
the anti-cheat list - so half-width, full-width, hiragana and katakana
match. Search only: stored text keeps its spelling.
- The palette takes one kanji, kana or Hangul syllable.
- The header search waits for an input method to finish composing before
it jumps to /games, and jumps with reset: false: the default reset sent
focus to <body>, so the next keystroke went nowhere.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Python's digit class also takes full-width and other scripts' digits.
A queued link in them gave an appid that missed every comparison and
made rotation.slice_of() raise, stopping every refresh. extract_appid
refuses a run of digits with a non-ASCII one in it; normalize_link()
reads NFKC first, so a Japanese keyboard's full-width appid is ASCII.
The blocklist uses ASCII digits and fullmatch, like removal.ts.
- truncate_description() backs off to just after a CJK sentence end as it
does to a space, and trims the CJK comma and full stop before the
ellipsis. Stored descriptions (all within the cap) never change, and
_fuller_cut() lets Refresh Store Data write a re-cut without calling it
news.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The candidate dialog lists the languages discovery recorded, with a
"No English" badge when English is not among them. A row queued before
discovery recorded languages shows "-", which is not the same as an
empty list (a game that declares none). Auto-approve is unchanged: such
a game is decided like any other.
The mock server's payloads carry languages too, and it queues one game
without English so the badge can be seen in dev.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every window swept before needs one more pass for the games without
English, and backfill windows are dispatched several at a time. The
discovery group kept ONE waiting run, so a third cancelled the one
already waiting - the lesson data-write learned in October.
- queue: max in both discovery workflows; pipeline.test.ts now holds
every shared concurrency group to one queue setting.
- Backfill Discovery takes max_pages (through env, default 40): a whole
recent year is about 30 pages of 100.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The store search pinned supportedlang=english, which left out every free
game that does not list English: 1,592 of 18,790 on 2026-10-30, most of
them Chinese, Japanese or Korean only (5118700 among them). Without the
parameter Steam filters on no language at all, with l=english or an
Accept-Language: ja header alike.
- fetch_search_page() sends no supportedlang, and sends l=english so the
row text ("Coming soon", month names) stays English by parameter.
- Candidates carry their languages (appdetails_languages()), a dry run
says which list no English, and every run counts them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on production after v4.5.1 deployed (2026-10-27, SIN): over one
keep-alive connection each budget held exactly - /api/no-such-route
121 x 404 then 429 with Retry-After, no-store and CORS; /api/human-check
19 then 429 - while requests on fresh connections, which Cloudflare
spreads over many servers, were never refused (175 human checks in 70 s
against a budget of 20). A probe Worker in a remote preview confirmed the
binding itself enforces (limit 5: 6 allowed, then refused).
So the layer stops browsers and connection-reusing scripts; a client that
reconnects for every request is the WAF rule's, which counted exactly
(250). The docs, CLAUDE.md, the module comment, wrangler.jsonc and the
unpublished v4.5.1 entry said "per location, approximate"; they now say
what was measured. The section 3b check used xargs (fresh connections) and
would have shown no 429 - it now sends everything over one connection.
No behaviour change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The section 3a check ran 300 separate curls through xargs -P 30. Each one
pays a TLS handshake, so from here it managed about 21 requests a second -
300 x 200 in 14 s - and never crossed 250 in 10 s: the rule looked broken
when it was not. One `curl --parallel` with its URL list on stdin reuses
its connections and does: 250 x 200 then 150 x 429, measured against the
live site-wide rule on 2026-10-27.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Worker-side rate limits behind the edge rule, and the recommended
site-wide WAF rule. Repo-only: nothing in the apps changed, so desktop and
Android stay at 2.3.0 with no new build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The zone's one Free-plan rate-limiting rule (IP, 10 s, block 10 s) stays
the first layer: it refuses a burst before the Worker is invoked and is the
only limit on the prerendered pages. The Worker now adds a second one with
the `ratelimits` bindings in wrangler.jsonc, 60-second windows and one
budget per surface, keyed by IPv4 address or IPv6 /64:
RL_DATA /api/data/* 240/min
RL_IMG /img/* 600/min
RL_API /api/activity, /api/updates/desktop, … 120/min
RL_HUMAN /api/human-check 20/min
RL_ADMIN /admin, /admin/*, /api/ingest/* 300/min
Checked after the country block and before every route and the Access
check; over a limit the answer is 429 + Retry-After: 60, no-store, with
CORS only where the route itself sends it (data, api). It fails open
(no binding, no CF-Connecting-IP, a throwing limit()) and logs no refusal.
Verified locally under wrangler dev: each budget refused exactly past its
limit, a second /64 address shared the first one's budget, pages were never
counted.
rate-limit.test.ts holds wrangler.jsonc to the code (binding names, period
60 = Retry-After, account-unique namespace ids) and checks that every
run_worker_first prefix draws on a budget. worker-configuration.d.ts gets
the five Env lines (and the IMG_MINT_PER_TICK value it had missed) without
regenerating the runtime types.
docs/SECURITY_SETUP.md section 3 describes both layers and recommends
widening the WAF rule to the whole zone except /_app/immutable/ at 250 per
10 s, counted from the build (a first visit sends 110-130 requests once
the service worker precaches); CLAUDE.md records the rules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SvelteKit 3 (PR #218) under the site and the apps, plus Tauri 2.12.5 from
Dependabot #214. Desktop and Android 2.2.2 -> 2.3.0 (package.json + lock,
Cargo.toml + the f2p-tracker entry in Cargo.lock, tauri.conf.json);
CHANGELOG section dated, fresh [Unreleased].
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The SvelteKit 3 fallback for a /#/ path that matches no route passed the
hash straight to location.replace(): CodeQL code-scanning alert #25,
client-side XSS, and an open redirect in practice - `#//other.example/x`
resolved to another site once goto rejected it.
The hash is now resolved against the page's origin and ignored when it
names another one, goto gets that URL, and an unmatched path goes to the
404 page at its fixed address (/error/404). Nothing from the hash reaches
location.* any more. Tests cover both off-site forms (`#//host`,
`#/\host`) and the 404 fallback; checked in a browser on `npm run dev`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- legacy-url: a /#/ path renders in place with { replace, reset: false },
and one that matches no route falls back to location.replace (kit 3's
goto rejects it).
- fallback-route: the re-render uses replace / reset: false / refreshAll
(noScroll and keepFocus throw in kit 3's dev build), and leaves routes
that hydrated or cannot fall back alone.
- params.legaldoc: every legal document matches; anything else does not.
- hooks.client: returns nothing for an ordinary error (kit 3 keeps its own
message), and says it is reloading only when the stale-chunk reload ran.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sveltejs/kit 2.70.3 -> 3.1.0, @sveltejs/adapter-static 3.0.10 -> 4.0.0.
Closes GHSA-27mw-82g8-phjj and GHSA-pfhg-q3qv-9q6r (dismissed in v4.4.1
as not applicable), and drops the cookie override kit 2 needed: npm audit
reports nothing.
- Config: svelte.config.js is refused by kit 3. Its options are now
sveltekitConfig(tauri) in build/sveltekit-config.ts, passed by
vite.config.ts as sveltekit(sveltekitConfig(IS_TAURI)), with tests for
both CSP flavours, paths.relative, the sitemap entry and the prerender
404 handler. The unused, now-deprecated "@" alias and typescript.config
hook are gone; tsconfig.json extends $app/tsconfig and owns `include`.
- $lib -> #lib (package.json imports), with .js on module imports.
tsconfig repeats the mapping in `paths`: svelte-check --tsgo only maps
.svelte imports through `paths`, and without it every #lib component
fell back to the *.svelte shim (props unchecked; proven with a planted
error).
- Param matchers: src/params/legaldoc.ts -> src/params.ts (defineParams).
vite build imports that file with Node's own import(), so it imports
#lib/legal.ts and legal.ts imports ./schema.ts by their real names.
- PWA: @vite-pwa/sveltekit never supported kit 3 and wrote the worker when
the SERVER build closed, which kit 3's environment build does first.
build/pwa.ts wires vite-plugin-pwa in a buildApp hook after kit's
prerender and before its adapter, with the same globs and URL mapping.
- goto options renamed (replace, reset: false, refreshAll); the /games
filter write is a shallow goto; a legacy /#/ path that matches no route
falls back to location.replace (kit 3's goto rejects it). Hook types
from @sveltejs/kit/hooks; the client handleError returns nothing to keep
kit's message.
- runed 0.35 (bits-ui) peers kit ^2: overridden to our kit, as for
echarts-wordcloud.
Verified against a kit 2 build of the same commit: same 11,037 files bar
hashed chunks, same non-immutable precache URLs, navigateFallback "/",
identical CSP per page (hashes masked), manifest and sitemap byte-equal,
same visible text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Desktop and Android 2.2.1 -> 2.2.2 (package.json + lock, Cargo.toml + the
f2p-tracker entry in Cargo.lock, tauri.conf.json), so the apps ship the
patched rustls/webpki stack their updater uses. CHANGELOG section dated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
npm (web/):
- undici 7.30.1 through a "miniflare" override (GHSA-xjj6-f33q-929c,
GHSA-4qv8-2p9j-2685, GHSA-f72v-6w45-q472). miniflare pins undici exactly
and wrangler 4.152.0, the newest, still ships 7.29.1. Local tooling only.
- zod 4.6.7 (GHSA-mrhx-c7xc-m3w5, via knip).
- Minor/patch: @lucide/svelte 1.61.0, @tanstack/svelte-virtual 3.13.48,
bits-ui 2.20.2, vite 8.3.5, wrangler 4.152.0, knip 6.43.0,
@tauri-apps/cli 2.12.4.
cargo (web/src-tauri/Cargo.lock, targeted updates only):
- rustls 0.23.46 (RUSTSEC-2026-0346..0349), rustls-webpki 0.103.16
(RUSTSEC-2026-0382), webpki-root-certs 1.0.11 (RUSTSEC-2026-0383),
tokio 1.54.1 (RUSTSEC-2026-0386), anyhow 1.0.104 (RUSTSEC-2026-0190),
event-listener 5.4.2 (RUSTSEC-2026-0221), plist 1.10.3 -> quick-xml
0.42.0 (RUSTSEC-2026-0194/0195). Dependabot had none of these.
- Left upstream: zip 4 (tauri-plugin-updater 2.14.0 asks for ^4), glib 0.18
(GTK 3), unmaintained proc-macro-error/zbus_names/zvariant.
pip: certifi 2026.10.23, coverage 7.17.0 (tests), with every file's hash.
CI: actions/checkout v4/v5 -> v6 and actions/setup-python v5 -> v6
everywhere (the Node 20 majors warned on every data job), rust-cache 2.9.3
by SHA. dependabot.yml says why those majors are taken by hand.
CLAUDE.md: the undici override and when to drop it; why SvelteKit stays on 2
(GHSA-27mw-82g8-phjj and GHSA-pfhg-q3qv-9q6r need setHeaders or use:enhance,
which the app does not have).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first run of the new Update Reviews on main asked appreviews about the
12 games GetItems counted at 0, and all 12 failed: since 2026-10-23
store.steampowered.com/appreviews/<id>?json=1 answers 302 to the store
front, from GitHub's runners as from anywhere else (it worked on 10-17).
The 38 games the ingest added that day were listed with "N/A".
- fetch_full() falls back to the GetItems summary when appreviews gives
none (reviews_from_store_api()), so an ingest gets the value the daily
refresh would write.
- update_reviews: a game never read ("N/A", nothing) with a count of 0
becomes "No reviews" without a question; the first count GetItems has
replaces it. A stored percentage is still never replaced by a 0 without a
second source.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/plan/18-pipeline-scale.md holds the measurements (9,847 games,
+~280 a day), what each job cost, what GetItems answers, the direction
and what is left. CLAUDE.md gains the rules for reviews, the purge,
budgets and the data-write group; README the new schedules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- queue: max on every data-write workflow (GitHub, 2026-05-07): up to
100 runs wait, first in, first out. The default kept one, and a run
queueing behind a running job cancelled it.
- ref: main on every checkout. A run's commit is fixed when it is
created, so a run that waited behind another data job wrote shards
from before that job's commit; Auto Update JSON failed four days in
ten on the rebase conflict in data/index.json.
- timeout-minutes everywhere, and JOB_BUDGET_MINUTES at least 20 below
it for the long loops (never for the ingest, which empties its queue).
- pipeline.test.ts fails on a data-write workflow missing any of these.
- The admin's Pipeline and Automation pages, docs/ADMIN.md and the
auto-approve comment no longer say a push evicts a waiting job; the
idle gate stays, so a batch's ingest starts at once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
It sent a HEAD request to every store page every five days, 5.5 hours at
9,847 games, and flagged 404/410. Steam sends a delisted app's page to
the store front, which answers 200: scripts/dead_links.log never existed
and no game ever became "delisted". The purge's GetItems screen sees
delisted games (unlisted) instead.
Removes the workflow, bash/dead_links.sh, scripts/check_dead_links.py
and DEAD_LINKS_LOG, and the workflow from web/shared/pipeline.ts and the
failure notifier's list. check_store_page() stays: it is the health
check's last resort.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Refresh Store Data: the page and appdetails cycles grow past 2,000
pages / 500 appdetails a day (7 -> 14, 30 -> 60) instead of the run,
the slice comes in a daily order, and the store step has a budget. A
game appdetails no longer calls free is listed in a ::warning::, the
case the purge's two-signal rule cannot see; nothing is written.
- Refresh All-Time Peaks: 30 -> 60 days past 400 pages a day, never
shorter, so SteamCharts is still asked at most once a month a game;
the budget stops it with a warning like a refusal does.
- Force Re-fetch: --part K/N, oldest last_updated first, each game
cleared on a copy that replaces it only once appdetails has answered;
no final loop stamping every game. Parts queue one after another.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The purge asked appdetails about every game, 5.3 hours at 9,847, and one
answer was enough to remove one - appdetails has answered success:false
for live games in bursts (2026-09-27).
GetItems now screens the whole catalogue in a few minutes (unknown,
hidden, unlisted, coming soon, not is_free; never type), and
check_game_health() must confirm each flag; "unavailable" is asked again
five minutes later. A failed batch or a network error never reads as
gone, blocklisted games are not asked about, and more than 500 flags or
100 removals stops the run before anything is written. Removals are
logged as before. A notice names earlier removals Steam has since undone.
The dry run on the real catalogue: 8 flagged, 3 removed (the three
unlisted games), 5 kept because appdetails calls them free. Daily now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Update Reviews asked appreviews about one game at a time, two seconds a
game. At 9,847 games it ran into GitHub's six-hour limit on 2026-10-23
and was stopped with nothing saved; five of its ten runs before that
were cancelled while waiting.
- SteamClient.fetch_items(): GetItems for a whole catalogue, a failed
batch costing only its own ids (the anti-cheat sweep now uses it).
- update_reviews.py reads review_summary() for every game, ~2 minutes,
daily. Steam floors its percentage; reviews_from_summary() floors
appreviews in integers too, so an ingest and the refresh agree. A
count of 0 is never written on GetItems' word (it said 0 for a game
with thousands): a stored percentage or "N/A" is asked on appreviews,
at most 300 a run. last_updated moves on a change, and on a 14-20 day
heartbeat that keeps the site's 30-day "stale" badge quiet without
rewriting every record every run.
- core/budget.py: JOB_BUDGET_MINUTES. Past it process_batch() starts no
further game, so a job saves what it did instead of being killed;
a ::warning:: when a loop used three quarters of its budget.
- rotation.cycle_for() / daily_order(): cycles that grow with the
catalogue as multiples of their base, and an order that changes daily.
- The player-count loops skip the batch pauses meant for store pages
(Top Online ~27 min instead of 48) and take a deadline, like
update_all_full().
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>