gacha-event-tracker/AGENTS.md

AGENTS.md

This file provides guidance to coding agents working in this repository. It is the working agreement: what this project is, the constraints it holds to, and the rules that are not visible from the code alone. Read it before changing anything.

CLAUDE.md points here, so Claude Code picks it up too — keep the guidance in this file and leave that one a pointer.

Read the docs before changing the thing they describe

This file is the working agreement, not the specification. docs/ holds the reasoning, and it is written for whoever touches that area next — reading the relevant one first is the difference between repairing a rule and rediscovering it the expensive way.

Doc	What it settles	Read it before
docs/PRD.md	What the product is for, feature by feature (F1–F14), and the quality bar for dates	Changing behaviour a reader can see, or arguing something is out of scope
docs/DATA-MODEL.md	GachaEvent, the SQLite tables, every localStorage key space, the export format	Touching src/shared/schema.ts, an ID scheme, a stored key, or the game/reset table
docs/INGESTION.md	The six pipeline stages, the parser/adapter/merge layering, date formats, the review gate	Adding a source, writing or repairing a parser, or changing the fetch runner
docs/ARCHITECTURE.md	Process shape, file layout, request paths, offline and update mechanics	Moving files, adding a route, or changing the service worker
docs/FEEDBACK.md	What readers actually said about the first release, and the work it argues for	Deciding what to build next
docs/SOURCES.md	Which sites publish a usable schedule for the games we still do not cover, and what is wrong with the ones that do not	Picking the next game to add, or assessing a source request

Two rules that follow from that:

• The docs are part of the change. A change that makes a sentence in docs/ false is not finished until that sentence is fixed. They are the only record of why, so drift costs the next agent the whole reasoning, not just a detail.
• When this file and a doc disagree, that is a bug — say so. Neither one silently wins. This file summarises; the doc holds the argument, so fix whichever is actually wrong rather than reconciling them in your head and moving on.

What this is

A web app that aggregates live and upcoming events across popular gacha games, plots them on a calendar, sorts them by end date or by what the reader is partway through, tracks day-by-day progress on events that repeat daily, and lets a user mark events completed.

Status: working app, refreshing itself on a schedule. Schema, eight parsers, nineteen sources across eighteen games, the full interface, offline support, a static server, a Docker image and CI all exist and are tested. The refresh runner (bun run refresh) fetches, caches raw snapshots and rebuilds the feed; .github/workflows/refresh.yml runs it twice a day and commits only when a page actually changed. The SQLite layer and the review queue are still specified in docs/ but not built, so the feed is a static JSON file built from snapshots, falling back to checked-in fixtures.

Three constraints that shape everything

1. No accounts, no logins, no user records. Completion state lives in the browser's localStorage, keyed by event ID. There is no user table and no session. Any request implying "sync across devices" is solved with export/import JSON, not a server-side user.
2. No LLM in the pipeline. Event data is extracted by deterministic code-based parsers only. There is no Anthropic dependency, no API key, and no per-run inference cost. A source that cannot be parsed deterministically does not get an adapter — see docs/INGESTION.md § No LLM.
3. A server is allowed (Bun) and owns fetching, parsing, and SQLite. The client only ever calls this app's own /api/*.

Stack

Layer	Choice
Runtime / server / bundler / test runner	Bun 1.3 (Bun.serve, bun:sqlite, bun test, bun build)
UI	React 19 + TypeScript (strict) + Tailwind
Storage	SQLite via bun:sqlite (gitignored — *.sqlite)
Validation	Zod — one schema module shared by server and client

The only runtime dependency is zod. Do not add a bundler, test runner, HTTP client, or HTML parsing library — Bun covers all four. tsconfig.json runs strict plus noUncheckedIndexedAccess, exactOptionalPropertyTypes, noUnusedLocals and noUnusedParameters.

The last two are there because this project has no linter and does not want one: a binding nothing reads is the only kind of dead code a reader cannot tell from a deliberate seam, and tsc already walks every file. Three had accumulated behind strict — an unread type View import, and a useGameMeta() resolver in two components that each draw their colours from somewhere else. A genuinely unused parameter is spelled with a leading underscore, which both flags already exempt.

Commands

bun install
bun test                      # full suite, offline, no network, no build needed
bun run typecheck             # tsc --noEmit
bun run dev                   # build then serve on :3000
bun run build                 # feed + css + js + static into public/

# Fetch sources and refresh the snapshots. Makes real requests — see § Scraping
# conduct before running it, and prefer --dry-run.
bun run refresh --dry-run
bun run refresh --only genshin-game8-events

# Run one source against its fixture (offline, free)
bun run parse genshin-game8-events fixtures/genshin/game8-events-2026-08-14.html
bun run parse endfield-wikigg-events fixtures/endfield/wikigg-events-2026-08-15.html --json

# Single test file / single test
bun test test/dates.test.ts
bun test --test-name-pattern "year-less"

# Hosting under a subpath (GitHub Pages)
BASE_PATH=/gacha-event-tracker/ bun run build

Tests must never need build output. They run before bun run build in CI; anything reading public/ must create its own fixture tree instead.

bun run parse ... --json is also how .expected.json fixtures are regenerated after an intentional parser change. Regenerating them makes the test self-consistent, not correct — always re-verify a sample against the live page afterward.

Current state of the code

src/shared/       schema.ts (the contract), time.ts, daily.ts, effort.ts, games.ts, feed.ts
                  custom.ts — reader-authored games and events, and their key spaces
src/ingest/       html.ts, dates.ts (sixteen formats), merge.ts, sanitize.ts, robots.ts, snapshots.ts
  parsers/        game8.ts, wikigg.ts, akwiki.ts, fandom.ts, bawiki.ts, holodori.ts, iopwiki.ts,
                  stellasora.ts — keyed by SITE, not game
  adapters/       index.ts — SOURCES registry binding url+game+parser, and the sanitize seam
src/client/       React app, service worker, manifest
  state/          progress, daily log, ignores, prefs, sort — all localStorage
                  gameOrder.ts — the reader's game order, and the A–Z fallback; pure
                  useCustom.ts — the reader's own games and events (PRD F13)
                  lens.ts — who sees which rows (focus, outstanding, next-to-expire); pure
                  zoom.ts — the timeline's scale ladder; pure
                  lanes.ts — how the timeline stacks: a lane per game, or one deadline queue; pure
                  theme.ts — dark or light, and what a game hue reads as on each
scripts/          build-feed.ts, build-static.ts, parse-fixture.ts (offline), refresh-sources.ts (fetches)
serve.ts          static server + /api/health
test/             831 tests
fixtures/<game>/  raw HTML + .expected.json per source — pinned, kept forever
snapshots/        current page per source, rewritten by refresh — see its README

Not yet built: the SQLite layer and the review UI. Everything upstream of them runs as files on disk.

Domain rules that are not obvious from the code

These come from how gacha games actually schedule things, and they cause most bugs here:

• Store every timestamp as UTC ISO 8601. Sources publish in a mix of UTC+8, server-local, and "after maintenance".
• Banner ends are usually global and simultaneous; event ends are usually per-region. Character banners end at one instant worldwide; story/login events end at each region's daily reset (Asia / America / Europe differ by hours). regionScoped and regionEnds exist for this — do not collapse them into one timestamp.
• endsAt: null is a correct, expected value. An event whose end is genuinely unannounced gets endsAt: null and endPrecision: "unknown". Never invent a plausible date to satisfy a non-null type. This is the worst failure mode this codebase has, because the user's entire reason for visiting is trusting the end date.
• A date with no time of day is stored as 00:00Z, and that is a placeholder, not an instant. Most sources print August 19, 2026 and nothing else, so dates.ts returns precision: "day" at UTC midnight because it has to return something. Counting down to it literally turns the placeholder into a claim the source never made — that the day opens in UTC — and retires an event up to nine hours before the game does, while the reader is standing in the game watching a longer timer. So clockFor (src/shared/time.ts) resolves a day-precision boundary to the reset that opens that game-day on the reader's server, via dayStartMs: the same clock daily.ts keys every tick by, and the only fact we hold about a game's day. Two boundaries are never re-anchored — a regionEnds value, which exists precisely because the source stated an instant per server, and an event the reader typed in, which readerInstant already resolved in their own timezone. This is a reading of the printed date, not an invented time, and it changes nothing stored: the feed, every event ID and the parsers are untouched, so it is one resolution at the point where the region is finally known. The same clock governs the other end of the pipeline. A parser whose page has no trustworthy "ongoing" heading decides currency against ctx.now itself, and comparing the 00:00Z placeholder to now retires a row hours before clockFor calls it over for anybody — the reader watches a deadline they were counting down to vanish on its last day. So latestBoundaryMs answers the same question for the last region, and bawiki.ts and two branches of fandom.ts ask it. Nothing stored changes: it is one comparison, not a resolved boundary written to the feed.
• Patch cycles are ~6 weeks. Any event over 180 days is a parse error, not a long event. The validator and the tests both reject it.

Working on parsers

• Parsers are pure. No network, no Date.now(), no randomness — time arrives as ctx.now. This is what makes fixture tests meaningful; a parser that reads the clock cannot be tested.
• Skip, never guess. Every function in dates.ts returns null rather than inferring a missing year, month, or end. readColumnTable drops a row it cannot date. An omitted event is a recoverable disappointment; a confidently wrong date is the failure this product exists to prevent.
• Parsers are keyed by site, not game. One game8 parser serves nine sources and fandom four; wikigg, akwiki, bawiki, holodoriwiki, iopwiki and stellasorawiki serve one each — the first two share a host family and have entirely different templates, and the last two are both Miraheze wikis whose page templates have nothing in common. Adding a source for a known site is one SOURCES entry; a new site is a parser module.
• A source may publish more than one region's schedule. Arknights' wiki lists CN and Global on every row, five months apart. Publish the one our readers are on and skip the row that lacks it — a CN date on a Global calendar is a confidently wrong date, not a near miss.
• Game8 has no single template. Eight shapes are known and a page may mix them: label/value detail tables, column tables, image-grid schedules (unsupportable), combined label+range+blurb cells, rowspan Start/End pairs, labelled Start: … End: … cells, <hr>-separated date pairs, and two schedules laid side by side in one <table> under a spanning label row. Full table in docs/INGESTION.md. Before assuming a new Game8 page will work, dump its structure and check every table — Endfield was written off as undatable on a pass that only inspected its Duration rows, and its real events were further down the page.
• The header row is the row that dates rows, not the first one. Game8's banner pages put the Standard and Paid schedules side by side inside one <table> and label the pair Standard Banners | Banner | Rating | Availability | Paid Banners | …. That row is not merely unhelpful, it is plausible — it contains both column words, so it resolves and puts the range at an index no data row has, and the whole table yields nothing with no error anywhere. So readColumnTable falls back to row 1 only when row 0 produced nothing, which is what keeps every page that parses today parsing identically. Verified rather than assumed: the change was diffed across all pinned fixtures and every live snapshot, and moved no existing event.
• Some Game8 wikis schedule banners, not events, and head their sections accordingly — List of All Banners, All Current Banners, and a Previous Banners back catalogue that previous events does not match. All three are in the section vocabulary now. The finished rows sit directly below the live ones and are dated identically, so that exclusion is the only thing between the calendar and a year of expired banners.
• Check what fences a section off. Inclusion is decided by headings, and the level varies: Persona 5 hides fifty finished events behind nothing but an <h4>Finished Events</h4> in a collapsed accordion, while Genshin uses h4 for sub-headings inside one event. So h4 gates sections but never names one — an unrecognised h4 must leave the current event title alone.
• Prefer a source that states machine-readable times. wiki.gg emits ISO timestamps with a timer per server region, which is the only reason regionEnds carries real data anywhere.
• Silent drops are the dangerous failure. A date format the parser does not recognise makes events vanish with no error. Abbreviated months (Apr. 29 - May 13, 2026) are supported for exactly this reason. When adding a source, compare the parser's event count against an independent count of the page.

Event IDs are localStorage keys

`${game}:${slugify(title)}:${startsAt.slice(0, 10)}`
→ "genshin:mutual-aid-in-bloom-into-the-frostlands:2026-08-12"

Changing slugify or eventId in src/shared/schema.ts — including seemingly cosmetic changes to the slug rules — silently orphans every completion mark every user has, with no server-side recovery, because the server never had the data. If it must change, ship a client-side migration that remaps old keys and keep it for at least a year. Use the schema-guardian agent on any such change.

Two more key spaces have the same property, for the same reason:

• dailies:<game> (dailiesId in src/shared/daily.ts) keys a game's standing daily chore. Two segments, so it cannot collide with an event ID.
• Game-day keys (dayKey) are YYYY-MM-DD in server-reset space, not UTC — the day rolls at 04:00 local server time. They are storage keys and they are compared with < and sorted, so the format is fixed. Changing the reset hour or the offsets moves every reader's streak by a day. The clock those keys are cut on — RESET_HOUR_LOCAL, serverOffsetUtc, resetHourFor, resetShiftMs — lives in time.ts, not daily.ts, because the countdown resolves day-precision boundaries on the same grid (§ Domain rules). Ticks are no longer its only caller, so a change there now moves a reader's streak and every undated end date at once. A game whose server map differs lists the affected regions in resetOffsets (games.ts) — Endfield serves Europe off the Americas machine, so europe is UTC-5 there and its reset is 09:00 UTC, not 03:00. Keep that override per region: a blanket per-game offset drags the regions that do have their own server onto someone else's clock. A game that rolls on a different hour says so in resetHourLocal instead — Reverse: 1999 resets at 05:00, not 04:00, so its day rolls at 10:00 UTC on its single UTC-5 server. Do not encode that as a bent resetOffsets value: shifting a game's stated server offset to land the right instant would misreport the server clock to everything else that asks for it. Both fields are absent for every game that takes the default, which is why adding the second one moved nobody's day keys. Neither field can express a server whose offset shifts: Fate/Grand Order's English server runs on US Pacific, which observes daylight saving, and one fixed number is wrong for half the year in either direction — so fgo takes the default and games.ts says why. Reaching for a value anyway would re-label day keys twice a year, which is the one thing this whole section exists to prevent. Every day-key function takes an optional game — anything reading or writing a tick must pass it, or it writes under one clock and reads under another. A day that drops out of dailyDays renders no pip, so a tick on it becomes unreachable; check real fixture windows before changing an offset.

The sanitizer at the ingest boundary recomputes an event ID only when a sanitized title actually changed and the ID was minted the standard way. If a change to it starts moving IDs on real fixtures, that is a data-loss bug, not a diff to regenerate.

Scraping conduct

Sources are community wikis. Treat them as a guest would:

• Honor robots.txt; set a descriptive User-Agent with a contact URL.
• One request per source per refresh cycle, minimum 6 hours apart. --force sets that floor aside for one run, and only a person may pass it — it is refused on an unattended run, because a schedule that forces every cycle is just a shorter interval with extra steps, and the interval is the obligation. What makes it defensible is what it does not change: conditional headers still go out, so a page that has not moved costs the host a 304 rather than a re-serve, and per-host spacing, robots, the one-request-per-source rule and the no-retry rule all still apply. Prefer it with --only: forcing nineteen sources to re-ask a question they answered an hour ago is the behaviour this bullet exists to prevent, whatever flag authorised it. The run names every source it asked early, and a run that was due anyway is never reported as forced.
• Space requests to one host, honouring its Crawl-delay and defaulting to 2s. Nine of the nineteen sources are game8.co pages, so the per-source floor alone still permits one cycle to arrive as nine back-to-back requests to a single site — which is the shape an edge network throttles, and what a burst looks like from the far end regardless of our intent.
• Send If-None-Match / If-Modified-Since; treat 304 as "skip, unchanged".
• Cache raw snapshots so re-parsing never re-fetches. Iterate against fixtures, not the network.
• Record sourceUrl on every event and surface attribution in the UI.

Note that game8.co disallows GPTBot and Google-Extended in robots.txt — it has opted out of AI-training crawlers. Our use is a low-rate personal aggregator with attribution and no model training, and no User-agent: * rule applies to our paths. Keep it that way: do not raise the fetch rate, and do not add an LLM that consumes page content.

game8.co does not answer a GitHub Actions runner (confirmed 2026-08-17). Its edge returns 202 Accepted with a bot-management body to every one of the nine game8 sources, from the first scheduled cycle onward — last confirmed: never — while the same URLs return 200 and parse cleanly from a normal address. So robots.txt permits us and the network does not, and those nine games have only ever been built from checked-in fixtures in CI.

The per-host spacing above does not fix this and was not meant to: a 202 on the very first request of a cycle is address reputation, not rate. Do not work around it. Browser-shaped headers, a proxy, or a residential egress would each be defeating a deliberate access control, which is the same reason uma.moe was declined below — and unlike uma.moe we would be doing it to a host whose robots.txt was welcoming, which makes it worse, not better. The legitimate options are to run the refresh from an address game8 will serve, or to find those games another source.

A source whose ToS forbids automated access does not get an adapter. Flag it and ask.

Sources assessed and declined (2026-08-17, extended 2026-08-19), so these are not re-litigated each pass:

Source	Verdict
azurlane.koumakan.jp	Declined. Content-Signal: ai-input=no — an explicit refusal of collecting content as model input, which is what capturing a fixture to read amounts to. Stronger than game8's or wiki.gg's signal. Find Azur Lane another source
uma.moe	Declined. Data comes from an API behind a Cloudflare Turnstile proof header; an adapter would mean defeating a deliberate access control. The robots.txt is permissive, but the gate is not in robots.txt
reverse1999.fandom.com	Built (2026-08-17), via api.php, not the wiki page — see § Fandom below
bluearchive.fandom.com	Declined. Fetches and parses fine; the page is the problem. Its Event/Event_List is a JP-server archive whose newest entry ended 2026-02-18, so all 88 rows are history and it yields zero live or upcoming events. An adapter would put an empty lane on the calendar and, because the runner rejects a body that parses to nothing, report a broken source forever. Same failure as the Infinity Nikki Game8 page, further along
bluearchive.wiki	Built (2026-08-17), from the rendered /wiki/Events page — see § Blue Archive below
fategrandorder.fandom.com	Built (2026-08-18), via api.php like Reverse: 1999 — but off Event_List_(US), not Event_List, which is the Japanese server. See § Fandom below
holodori.wiki	Built (2026-08-18), from the rendered /wiki/Events page. Miraheze again, so the same call as Blue Archive; CC BY-SA 4.0, no Content-Signal, no Crawl-delay for *
prydwen.gg, gametora.com	Cleared, unbuilt. User-agent: * allows the paths we would want. prydwen sets Crawl-delay: 10, far below our one-per-6h
iopwiki.com	Built (2026-08-19), Girls' Frontline 2 — see § IOP Wiki below. robots.txt is two lines, User-agent: * and Crawl-Delay: 20, no Disallow anywhere
stellasora.miraheze.org	Built (2026-08-19), from the front page's Current Banners module and not /wiki/Banner_List — see § Stella Sora below
game8.co/games/Chaos-Zero-Nightmare	Built (2026-08-19). Zero parser work — the existing game8 parser reads it. The ninth game8 source, so fixture-backed in CI from day one
game8.co/games/Umamusume-Pretty-Derby	Built (2026-08-19), off the stable List of All Banners page, not the monthly release-schedule pages whose URL changes every month. Cost a widening of game8.ts's section and column vocabulary — see § Working on parsers
nikke-…-international.fandom.com	Built (2026-08-19), via api.php like Reverse: 1999 and FGO. Its robots.txt was read in a browser and is the standard Fandom file — see § Fandom below. Richest schedule of anything added in this pass: story events and dated pickup banners, with the reset clock evidenced on the page
infinity-nikki.fandom.com	Built (2026-08-19), replacing the Game8 page for Infinity Nikki, which had been stale since August 2025. Same standard Fandom robots.txt. Published at day precision: the page states a wall clock and no zone for it — see § Fandom
infinitynikki.miraheze.org	Declined. Exists and serves robots.txt, but the wiki is abandoned — front page last edited 11 February 2025 and /wiki/Events returns a permission error. Checked as a replacement for the stale Infinity Nikki Game8 page
prydwen.gg/infinity-nikki	Declined. 404 — prydwen does not cover Infinity Nikki
grayravens.com (Punishing: Gray Raven)	Declined. Conduct is fine; the data is not. The whole 626 KB /wiki/Events page contains exactly one date range, written as prose, one event per six-week patch
guardian-tales.fandom.com	Declined. Parses fine and contains no 2026 date at all — newest dated entry is 2025. The bluearchive.fandom.com failure again: parses cleanly to nothing live
blhx.fandom.com, azurlane-archive.fandom.com	Declined. The two Fandom alternatives to the declined koumakan wiki are dead archives — Event_Calendar stops in 2021, and the archive wiki's headings have nothing under them. Azur Lane still has no source
Aether Gazer	Do not build. The developer confirmed no further content updates after 23 July 2026, with store listings removed 17 October 2026. The wiki dates nothing anyway — Event_Guide_List is an image gallery. A lane that will be empty by winter
marisaimpact.com (Honkai Impact 3rd)	Declined (2026-08-19). Conduct is clear — its robots.txt is comments only, with no directive and no Content-Signal, and the page answers our own User-Agent with a 200. The data is the problem: the schedule is a grid of week columns headed Estimated date for Regional Servers under a page that says Based on CN server, it states no year anywhere, and it lives at a per-version URL — /calendar89 is v8.9 and expires on 20 August, with no stable route to the current one. docs/SOURCES.md § 13
game8.co hubs for Black Beacon, Brawl Stars, Destiny: Rising, Diablo Immortal, Epic Seven, Fire Emblem Shadows, Gundam UC Engage, Mongil: Star Dive, Pokémon Champions, Pokémon UNITE, Tower of Fantasy	Declined (2026-08-19). All thirteen hubs in that sweep exist and answer 200; these eleven have no usable schedule. Six are abandoned wikis whose newest page is 2021–2025 — the Infinity Nikki failure mode, a source that parses perfectly and publishes history. Gundam's calendar prints ends with no starts, so no event ID; Pokémon UNITE is fresh but its template fails canParse, which is the check working. Per-game evidence in docs/SOURCES.md § 12
game8.co/games/MementoMori, game8.co/games/fire-emblem-heroes	Assessed, not yet built (2026-08-19) — the two live finds of that sweep, and proposals rather than decisions. MementoMori parses today with no parser change; FEH has the freshest page of any source here and needs a ninth Game8 column shape plus a ruling on the 180-day rule, which one real seven-month banner breaks. See docs/SOURCES.md §§ 12a–12b

The Infinity Nikki lane was rebuilt on a live source, and its Game8 source was retired. game8.co/games/Infinity-Nikki/archives/487445 stopped being updated on 31 August 2025 — it mentions the year 2026 zero times — and had been publishing five year-old events with endsAt: null, which the app renders as live-with-unknown-end indefinitely. That is worse than an empty lane and it is the failure this product exists to prevent, arriving through a source that looks perfectly healthy to the runner, because a stale page is not a broken one: no failure streak, no annotation, no broken tier. Nothing in the pipeline catches this. When adding a source, check when the page was last updated, not only whether it parses.

The lane now comes from infinity-nikki.fandom.com at day precision (§ Fandom above), and the Game8 entry is gone from SOURCES rather than kept as a second opinion — a source whose every row is wrong is not corroboration. Its fixture stays in fixtures/nikki/, because it is the only page here carrying Game8's labelled Start: … End: Permanent shape and test/adapters still drives it through the parser directly as a regression test.

.github/ISSUE_TEMPLATE/feature_request.yml points readers at that table by heading, so a source request can be checked against it before anyone writes it up — the loudest feedback on the first release was "not enough games" (docs/FEEDBACK.md), which makes this the request that arrives most. Keep the heading if the section moves.

wiki.gg hosts (arknights, endfield) carry Content-Signal: search=yes, ai-train=no, use=reference with Allow: /, and disallow ClaudeBot and other AI crawlers by name. Our fetcher is neither: it trains nothing, and no LLM reads the page content — constraint 2 is what keeps that true, so it is load-bearing here and not only a cost decision. Note also that Reverse: 1999, Blue Archive, Umamusume and Nikke have no wiki.gg wiki — those subdomains 401.

Fandom: read the API, never the page. reverse1999.fandom.com/wiki/Events answers some non-browser clients with a Cloudflare managed challenge — HTTP 403, Just a moment…, "Enable JavaScript" — and /robots.txt is on the same challenged route, though our own client is served both (see the correction below). Browser-shaped headers or a JS-executing client would get past a challenge and must not be used: that is defeating a deliberate access control, the same reason uma.moe was declined above.

What makes this source legitimate anyway is that the wiki publishes a second, sanctioned surface. Its robots.txt — read in a browser, where it serves fine — has no Disallow: / for * and explicitly allows /api.php?action=, and that endpoint answers our real User-Agent with a 200 and a JSON body. So the adapter fetches api.php?action=parse&page=Events, with no impersonation anywhere: our own headers, on a path the site put in writing. The only namespaces * is refused are Special:, User:, Template: and Help:, none of which we want; parsers/fandom.ts skips Special: links for that reason.

The Fandom 403 is per-address and per-client, and one of those was measured wrong — 2026-08-19. For one day this section recorded that 403 as Fandom's posture, holding across every wiki and every address, which made those four sources permanently unschedulable. Two separate variables were folded into that one claim, and they have to be held apart:

• The client. The original measurement was curl, and curl is not what fetches. On one address, in the same minute:

  Client	GET /robots.txt
  curl, no User-Agent override	403, 8 tries out of 8
  curl, a Chrome User-Agent	403, 8 tries out of 8
  curl, our real User-Agent	403
  Bun fetch, our real User-Agent	200, and the real file, on all five hosts

  There, RobotsCache returns allowed / robots.txt ok for all four sources with no override, and the file it reads is byte-for-byte the standard Fandom one recorded above.

• The address, which the client does not rescue. The scheduled run a60eb66 (17:46 UTC, on ubuntu-latest) reported skipped_robots for all four Fandom sources using that same Bun client, in the same cycle that nine game8 sources took the usual 202 CloudFront. So a served address is a precondition, not a detail: a claim that these four now refresh on a schedule is a claim about the address the runner has, and it must be re-measured there rather than inherited from here.

  [self-hosted, safe-ip] was then measured, and it fixed neither edge. The first cycle after that switch reported 403 (an interstitial challenge) for all four Fandom sources; the next one, once the 6h floor was up, reported 202 CloudFront for all nine game8 sources. So that address is challenged by Cloudflare and refused by CloudFront alike, and the runner move bought nothing on its own — worth recording plainly, because "use a different runner" is the first idea anybody has here and it has now been tried.

  What did fix the Fandom half was the cron passing --assume-robots-on-403 (below): the following cycle put all four through at 200, two of them with fresh bytes. game8 has no equivalent and cannot get one — its 202 is the page itself being withheld, not a permission we hold and cannot re-read, so there is nothing for a recorded decision to stand on. Those nine remain fixture-backed in CI until they are fetched from an address CloudFront serves or replaced with another source.

Three things worth keeping from the episode, because each one is a trap:

• curl is the wrong instrument for a question about this pipeline. The 403 is a Cloudflare managed challenge (cf-mitigated: challenge, cType: 'managed'), and a managed challenge is scored largely on the TLS and HTTP/2 fingerprint of the client. Bun's stack passes where curl's does not. Verify a fetch gate with bun -e and the real RobotsCache, never with a shell client, or you will record a conclusion about curl and file it under Fandom.
• The User-Agent is not the lever, in either direction. Adding a browser User-Agent to curl changes nothing — 403 either way — so a 200 that appears when you drop the header is the challenge's probabilistic half, not a header you tuned. fategrandorder answered 403 once and then 200 seven times running to an identical request. This matters because the obvious reading of such a result is "shape the headers until it works", which is precisely the impersonation the top of this section forbids. It also does not work.
• The challenge is route-scoped. api.php is exempt at Fandom's edge and answers everything; /robots.txt, /wiki/* and / are the challenged routes. So a challenge on the HTML page tells you nothing about the surface the adapter actually reads.

A new Fandom source still needs that wiki's own robots.txt read and recorded here before it is built — read it with the runner's client, which can. That is what Nikke's clearance on 2026-08-19 rests on, and the file is the standard Fandom one: no Disallow: / for *, /api.php?action= explicitly allowed, only Special:, User:, User_talk:, Template:, Template_talk:, Help: and UserProfile: refused, and the named AI crawlers it blocks (GPTBot, CCBot, OAI-SearchBot, ImagesiftBot) are not us.

Should a challenge ever return, it is still not a licence to work around it. An adapter can be written and fixture-backed from any address, because api.php?action=parse answers our own User-Agent with a 200; what a challenged address cannot do is pass the robots gate, which fails closed and skips. The answer then is the one below — a person refreshing from an address Fandom serves, on a permission recorded by hand — and never browser-shaped headers.

And no, a managed challenge cannot be waited out or solved on our side. Asked directly on 2026-08-19, so it is not re-tested each pass:

• Delays do nothing, because neither block is a rate limit. Three independent reasons, and the run of 17:46 UTC on 2026-08-19 supplies the first two:

  • Every one of these failures was a first request to that host in the cycle. robots.txt is fetched once per host per run, so each of the four Fandom 403s was first contact with a different host. genshin-game8-events is the first game8 source in the order and it took a 202 before any other game8 request existed. A mitigation on request one cannot have been caused by the pace of requests two through nine.
  • The spacing is already there. game8's robots.txt states no Crawl-delay, so the runner applies DEFAULT_HOST_GAP_MS (2s) between all nine game8 requests — and all nine still 202. Cycles are ≥6h apart and game8's lastConfirmedAt is still never, so months of 12-hour spacing has not moved it either.
  • Neither status is a rate response. Rate limiting is 429 or 503 with Retry-After. What arrives is 202 with a CloudFront bot-management body, and 403 with cf-mitigated: challenge and no Retry-After. Those are identity verdicts, not pace verdicts.

  Spacing requests is an obligation we owe the host regardless, and it stays. It is simply not a lever on this, and adding more of it would only slow the run down while changing nothing.

• Carrying cookies does nothing. curl with a cookie jar across three requests is 403 every time. The edge issues __cf_bm (bot management) and never cf_clearance, which is the only cookie that marks a challenge as passed.

• Earning cf_clearance is the access control. It is issued only after the /cdn-cgi/challenge-platform/ script runs and returns a valid proof from a browser environment. Executing that to obtain the cookie is defeating a deliberate access control — the exact ground uma.moe was declined on, where the gate was a Turnstile proof header. A gate not written in robots.txt is still a gate.

• It does not even work. A real headless browser pointed at these hosts gets a managed challenge that never resolves, so the option being refused on conduct is also the option that fails.

The legitimate lever is the address — but the one alternative tried here was not a better address. refresh.yml ran on [self-hosted, safe-ip] from 2026-08-19 and was switched back to ubuntu-latest on 2026-08-20, because that runner was challenged by Cloudflare and refused by CloudFront alike (§ the per-address measurement above). So this remains the route for game8 and it is still unwalked: an address CloudFront serves, or another source. Do not read the revert as the idea being wrong — read it as that particular address not being the one.

--assume-robots-on-403 is the one concession to that, and it is deliberately the narrowest thing that helps: bun run refresh --assume-robots-on-403 treats an interstitial challenge on /robots.txt itself as the permission recorded above rather than failing closed. It is not a workaround for a host that turned us away — it never overrides a robots.txt we could read, so a file that disallows us still says no, and it does nothing at all for game8.co, whose robots.txt reads fine and welcomes us while its edge refuses the pages. Every host it applied to is named in the run's warnings, so it stays a thing somebody decided this morning rather than a default. Nothing else relaxes: one request per source, six hours apart, spaced per host, no retries.

Who may pass it is a question about the person, not the machine. Both overrides are refused on an unattended run and available to a person, and runAttendance (scripts/refresh-sources.ts) draws that line: a local shell is a person, a workflow_dispatch is a person and GitHub records which one in GITHUB_ACTOR, and a schedule — or any other runner event — is not. This replaced a blanket isCi() check, which asked the coarser question and got the case that matters wrong: a dispatch sets CI=true, so somebody clicking "Run workflow" was refused exactly as the cron was, and the overrides were unreachable from the workflow at all. The run prints which override was used and who authorised it.

The two overrides are not symmetrical, and --force is the stricter one. A schedule may never force: asking every cycle before the interval is up is a shorter interval with extra steps, and the interval is the obligation. That one stays reachable only from a person or a dispatch.

The cron does pass --assume-robots-on-403, and that is a decision with a named cost (taken by the repository owner, 2026-08-20). Four Fandom sources skip on a challenged robots.txt every cycle, and the alternative was four calendars sitting as stale as the last manual run — which for a product whose whole promise is a trustworthy end date is the worse failure. So the schedule stands on the permission recorded in this section, and the owner re-reads those files by hand over time rather than the code pretending to.

Be exact about the risk that buys, because it is not "we might be crawling against robots.txt":

• A plain 403 still fails closed. The challenge-or-refusal split is what guarantees that much, so a host that actually turns us away still stops the run.
• What is invisible is a robots.txt edited to disallow us. From a challenged address the file never arrives, so a withdrawal would look exactly like the challenge we already expect. Nothing in the pipeline can catch that; only the manual re-read can.
• The per-cycle warning is therefore the compensating control, not a nicety. Every run names each host it stood on a hand-recorded permission for, says the file was not read, and tells the reader to re-read it — as a ::warning annotation on the run page, so it survives a completely green cycle where nothing else draws the eye. test/refresh.test.ts pins both the one-warning-per-host rule and that it reaches the run page. Do not demote it to a log line, and do not let a future change make it conditional: it is the only thing that ever prompts the re-read this arrangement depends on.

A 403 is two answers wearing one status code, and only one of them is covered. A managed challenge means "we cannot tell what you are" — the question a human answers by reading the file in a browser, which is the whole basis of the concession. A bare 403 means "you are forbidden", and that is a host declining us, which no permission recorded on our side may talk over. The flag claimed this distinction from the day it was written and could not actually draw it, so it excused both; isInterstitialChallenge (src/ingest/robots.ts) now decides, on Cloudflare's own cf-mitigated: challenge header with the challenge page's markers as a fallback. A 403 whose body cannot be read is unclassifiable, and unclassifiable is not challenged. This only ever narrows what the flag opens — nothing that passed the gate before stops passing it.

One consequence to keep in mind: because /robots.txt is unreadable from a challenged address, the robots gate fails closed there and the source is skipped. That is a warning line rather than a broken build — skipped_robots does not touch the failure streak, and the run only hard-fails if every source is blocked — so the scheduled refresh simply never updates this game, and the feed falls back to the checked-in fixture. Refreshing it means running bun run refresh from an address Fandom serves, which is how its first snapshot was taken.

Four Fandom templates now, and the third states its zone in a column header. The Nikke wiki's Event page is Event | Start(UTC+9) | End(UTC+9) | Archived(?) for story events and Nikke | Start(UTC+9) | End(UTC+9) for pickup banners. That header is the safety property, not a convenience: no date in any cell carries an offset, so a table whose Start/End columns stop naming a zone must be refused rather than read as UTC — the Blue Archive hazard, arriving one column to the left, and canParse asserts the lookup. Two more things about it:

• Every title is an image, and the newest row is the one without one. Names come from the wrapping <a title="Project Matis">, but an event whose logo has not been uploaded yet renders as a red link reading File:Persona on Frontline logo.png — so a reader that only understood <a title> would silently drop today's live event and publish a calendar missing what is on now. The file name is the fallback, and a test pins that exact row.
• A start with no clock keeps the day the page printed. Story events state a bare date on the start and a clock on the end; converting the bare one from UTC+9 would move it to the previous calendar day, and the start's day is half an event ID. That is the Fate/Grand Order rule below, applied to the opposite gap — there, a zone with no clock; here, a clock on only one side.

The fourth is Infinity Nikki, and it is published at day precision on purpose. infinity-nikki.fandom.com (the unhyphenated name 301s to it) heads its Current Events and Upcoming Events tables Event | Duration | Description | Type, and every duration reads July 20, 2026 04:00 – August 10, 2026 03:49 — a full date and a wall clock on both sides, and no zone anywhere on the page for that column. The only zone statements are prose elsewhere dating version launches (UTC-7) and a note that rewards reset at 04:00 (Server Time); the durations do run 04:00 → 03:59, which only lands on a reset boundary if the column is server-local. Strong, and circumstantial.

So parseZonelessClockRange reads the clock and throws it away, publishing the printed date at day precision. That invents nothing and treats these cells exactly as every Game8 date is already treated. Converting instead would mean picking an offset, and the offset moves the day: July 16, 2026 20:00 read as UTC-7 is 2026-07-17T03:00Z, and the start's day is half of every event ID this game will ever have. If the wiki's editors ever state the zone on that column, this source can carry exact instants and should.

Two shapes to know: Permanent Events and Past Events share the page and are fenced off by heading, and titles come from a link's title attribute — which means they must be entity-decoded by hand, because an attribute never passes through text() and Alison's Travel Shop would otherwise become a slug, and a slug is a localStorage key. The sanitiser catches exactly that, and a parser needing repair on its own fixture is a parser with a bug.

The second Fandom source's page is chosen, not obvious. fategrandorder.fandom.com publishes two schedules: Event_List opens "This page lists all Events in Fate/Grand Order Japan", and Event_List_(US) is the English server. They run months apart, each links the other, and reading the Japanese one on an English calendar is the akwiki CN column again — it was how this source first landed, and every date it published was a JP date. The adapter is pointed at page=Event_List_(US) and a test asserts it; parsers/fandom.ts carries the reasoning.

Three more things about that page, all of them ways to publish or lose a date:

• Its sections are fenced by pictures. ONGOING EVENTS, FUTURE EVENTS and PAST EVENTS are banner images with the label drawn in a positioned <div> over them — no heading, no id. Only the ongoing section is parsed, and canParse asserts both of the dividers that bound it, so a redesign fails the source rather than emptying the lane.
• The other two sections cannot be dated, and that is the whole reason they are skipped. FUTURE EVENTS gives an ETA of August 2026 — a month with no day, and a day is half an event ID. PAST EVENTS is 111 monthly tables that state no year anywhere; the Japanese page's equivalents carry it in a MMYYYY table id, which is a difference easily assumed away.
• Every duration names a zone and no clock — August 12, 2026 ~ August 26, 2026 PDT. So the boundaries stay on the day the page states rather than being shifted into UTC: there is no time of day to anchor a conversion to, and the start's day is part of the event ID. That PDT is also the evidence that the English server is one machine on US Pacific — see games.ts, where it does not become a resetOffsets entry, because Pacific observes daylight saving and that field holds one fixed number.

Blue Archive: the page, never the API — the opposite call to Fandom. bluearchive.wiki is a Miraheze wiki, and Miraheze's robots.txt disallows /w/ and /*?action=. So the route parsers/fandom.ts takes is the one that is closed here, and the rendered /wiki/Events page is the surface * is allowed — it answers our own User-Agent with a 200, no Content-Signal, and no Crawl-delay for us. Special: is disallowed too, which is why parsers/bawiki.ts skips those links exactly as the Fandom one does.

Three things about that page are worth knowing before touching it, all of them ways to publish a confidently wrong date:

• It states JP and Global in separate tabs, and the Japanese one runs four to nine months ahead. Same hazard as the CN column on akwiki, same answer: publish Global only. The tab's nav button carries the id tabber-Global_version-label and sits above both panels, so a reader that slices from the first id match reads the Japanese schedule while believing it read ours.

• There are three Global tabs, not one — the schedule, plus Mini-Event and Joint Firing Drill further down, whose ids are the same name with _2 and _3. The parser finds the schedule by its Name (EN) header rather than by position, and canParse asserts that lookup, so a renamed tab or column fails the run instead of quietly emptying the lane.

• The page states no time of day and no timezone anywhere. The schedule's dates are bare YYYY-MM-DD, which is honest day precision. Its five other tables (Mini-Event, Reward campaigns, Attendance bonuses, Guide missions, Joint Firing Drill) do carry a wall clock — 08/12/2026 11:00 — but name no zone for it, and three of the five do not say which server they describe. Those are deliberately unparsed, and that second clause is the whole reason: a clock whose server is unknown cannot even be labelled with a day, because you do not know whose day it is. Attendance bonuses would be a real dailies source if a zone is ever stated. For the same reason ba has no resetOffsets: Blue Archive Global does run one worldwide server, but nothing in this source says on what clock.

  This rule governs a clock with no known server, not a clock with no stated zone — a distinction worth drawing precisely, because it was drawn the wrong way once. As first written it said rounding such a cell to a day "does not save it", which would also have condemned every Game8 date in this repository: those state no zone either, and are published at day precision without anyone minding. The Infinity Nikki wiki (§ Fandom) is the case that forced the correction — one worldwide service, a clock, no zone — and it is read at day precision on the printed date, which invents nothing. What is still forbidden is converting an unzoned clock by picking an offset, because the offset moves the day and the start's day is half an event ID.

hololive Dreams: the same Miraheze call as Blue Archive, and the opposite data. holodori.wiki is Miraheze too, so /wiki/Events is the surface * is allowed and /w/ and ?action= are closed — parsers/holodori.ts takes the route bawiki.ts takes, for the reason it takes it. What differs is the quality of what is there, and three things are worth knowing:

• It states its timezone on every cell. Every boundary is 08/17/2026 8:00PM (JST), which makes this the only wiki source here publishing exact precision on both sides without a per-region timer. parseSlashClockZone requires the zone rather than defaulting to UTC, so a row that ever loses it drops out instead of landing nine hours off. That is also where holodori's resetOffsets of UTC+9 comes from — evidenced, not assumed; see docs/DATA-MODEL.md.
• Inclusion is fenced by an <h2>, and the two tables are identical. Current Events and Past Events have the same columns, so a reader that took every wikitable would put the back catalogue on the calendar with nothing to mark it. Rows are checked against ctx.now on top of the heading, because "Current" is maintained by hand and goes stale before anyone moves a row.
• Every event title is still a red link. The wiki has no article for any of them yet, so each links to ?action=edit&redlink=1 — a create-page form, and a ?action= URL this wiki's robots.txt disallows. The parser refuses a href with a query and falls back to the events page; when the articles exist, they get linked with no change.

Two rows on the page are not events and are meant to be missing. Beginner Mission runs Game Launch → Unknown: no start means no event ID, and a permanent tutorial chore is not what a calendar of deadlines is for. An Unknown end is kept, though — that is endsAt: null, and unlike bawiki.ts this parser does not drop a started-but-undated row, because the heading has already said the event is running.

IOP Wiki: the Server column is the whole safety story. iopwiki.com/wiki/GFL2_Events is the best date material here after wiki.gg — every row states an exact instant on both boundaries and names the zone (2026-08-06 13:00 - 2026-08-26 22:59 (UTC)), so parseIsoClockRangeUtc converts nothing and both sides are exact. Three things about it:

• CN, EN and JP rows share one table, and the Chinese schedule runs about a year ahead. This is the akwiki CN-column hazard verbatim and gets the same answer: publish EN, skip the rest. It would be wrong by months on a row that otherwise looks perfect.
• Betas is a section, not an event type. Closed beta rows are dated exactly like everything else and would parse cleanly onto a calendar of things nobody can play. Fenced on the <h2>.
• The page is an archive, 145 rows back to 2023, so inclusion is decided against ctx.now as in bawiki.ts. The lane is therefore thin by design — one live event on a quiet week is the truth, not a gap.

The zone requirement is deliberate: parseIsoClockRangeUtc refuses a row that loses its (UTC) rather than assuming it, exactly as parseSlashClockZone does. GFL2 takes no resetOffsets: its EN boundaries land on three different clocks (22:59, 08:59 and 02:59 UTC), which is a patch window rather than a reset hour — Arknights and Reverse: 1999 each earned an override from a single boundary their whole page agreed on.

Stella Sora: the front page, not the article — the opposite call to Blue Archive. Miraheze again, so /wiki/ is open and /w/ and ?action= are closed. But this wiki publishes its schedule twice, and the fuller surface is the worse one:

• /wiki/Banner_List has 55 clean rows with full wall clocks and states no timezone anywhere.
• The front page's Current Banners module emits the same instants as real <time datetime="2026-08-17T20:00-07:00"> elements.

The two agree exactly, which is strong evidence the table is UTC and is still only evidence — so we read the surface that says what it means and pay for it in coverage: four live banners instead of a full history. If an editor ever states the zone on Banner_List, that page becomes the better source immediately. Two traps in the markup: the template writes its BEM underscores as __, so a selector written against the name a browser shows finds nothing at all; and banner names are red links to ?action=edit&redlink=1, which robots.txt disallows, so a href with a query is refused and the page URL stands in — the holodori.ts rule.

Stella Sora takes no resetOffsets either, and for the opposite reason to most: it states an offset outright, and the offset is -07:00 — US Pacific, which shifts by an hour twice a year. That is the Fate/Grand Order problem arriving through a source that looks like it answered the question.

scripts/refresh-sources.ts enforces all of the above in code — the 6h floor (except under the opt-in --force above), one request, no retries, conditional headers, per-host spacing, robots (failing closed when robots.txt cannot be read, except under the opt-in --assume-robots-on-403 described in § Fandom, which covers a challenged 403 and never a plain refusal). The two overrides differ on who may pass them: --force is refused on an unattended run (runAttendance — a person or a workflow_dispatch, never a schedule), while --assume-robots-on-403 is passed by the cron by standing decision and is reported by name in the run's warnings — an override that reports nothing is one nobody withdraws. Anything that would make it fetch more often is a change to this section first.

A source down is a warning; a source down for days is a broken build. One wiki failing must never blank a calendar or stop the sources that did answer from being committed — so a failure is exit 0 and the previous snapshot stands. But a source that has failed BROKEN_AFTER_FAILURES (3) cycles running is not having a bad afternoon: that game's calendar has been quietly built from a checked-in fixture for a day and a half. The runner reports those as broken — a GitHub annotation, a row in the job summary with the status code, and a broken step output — and refresh.yml fails the run on it in a final step, after the commit and the CI dispatch. Exiting non-zero from the runner instead would skip the commit and throw away the pages that did arrive. This tier exists because six of seven sources failed every cycle for three days behind a green tick; a warning nobody opens the log to read is not a signal.

Untrusted input

Every string on an event came from a page we do not control. src/ingest/sanitize.ts is the trust boundary and it is wired into toAdapter() in src/ingest/adapters/index.ts, which is the single seam every source passes through — do not sanitize inside a parser, and do not add a code path that reaches parser.parse directly. Parsers stay pure readers of one site's markup.

The sanitizer never touches a date, cleans rather than drops (a title that sanitizes to nothing is the only drop), and logs every repair and drop by default. See docs/INGESTION.md § Stage 2.5.

Events that repeat daily

Some events are twenty small jobs on twenty deadlines, not one job with an end date, and a missed day is unrecoverable. src/shared/daily.ts decides dailiness from what the source published — type: "login", or "daily"/"check-in"/"7-day" wording — and never from a game's habits or an event's length. It adds no schema field, so the feed contract is untouched.

• The day rolls at 04:00 server time (RESET_HOUR_LOCAL), per region. Getting this wrong ticks the wrong box for four hours every night.
• An unannounced end yields no checklist, not a checklist of guessed length — the endsAt: null rule applies here exactly as it does to a countdown.
• A tick is never removed except by the reader, including ticks outside the window the feed now claims. A source quietly moving a date must not erase a fortnight's streak that exists nowhere else.
• A repeating event the reader marked done leaves the strip. They have said there is nothing left to do; keeping a tickable chip for it is the app arguing with them. Their logged days are untouched, so unmarking it brings the chip and the streak straight back.
• Detection is a guess, not a verdict — and it ships off. prefs.detectDaily defaults to false and the control is labelled experimental: wording is a weak signal and gets it wrong in both directions, so a new reader opts in rather than out. The default moves nothing for an existing reader, whose stored prefs wins. The reader can mark any event as repeating, or unmark one detection got wrong (progress.daily, resolved by resolveDaily), whether the guessing is on or off. Store an override only when it disagrees with detection — recording agreement would freeze today's guess and stop a better parser from ever reaching that event. Neither control ever deletes a mark or a logged day, so both are reversible.

Events the reader entered themselves

No adapter list covers a ten-game player, so a reader can define a game and type in events (PRD F13, src/shared/custom.ts, src/client/state/useCustom.ts). They join the same lists, timeline, sort, filters, progress, ignore and daily stores as scraped events. Four rules:

• Their ids live in their own spaces: mygame:<slug> and myevent:<random>. Never ${game}:${slug}:${date} — a reader can type a scraped event's exact title and date, and that collision would silently share one completion mark and one streak between two events. Random also means renaming their own event never moves its id. dailies, mygame and myevent are reserved first segments and none may ever become a GameId; a test pins this.
• Nothing they type enters the ingest pipeline. sanitize.ts and merge.ts are for pages we do not control. Their events are not fetched, parsed, merged, scored or quarantined.
• A hand-entered date is never attributed to a source. No sourceUrl, no source link, and the row and detail sheet both say it is theirs. "I don't know when it ends" is an offered answer, for the same reason the parsers are forbidden from guessing one.
• They are in the export. These exist in one browser and nowhere else, so an export without them is a lossy backup. Import merges by id and never removes.

A lane may now be a game the reader invented, so gameMeta is a context resolver (metaFor, pure and total) rather than a direct lookup — a lane can outlive its game when an import carries an event whose game did not come with it.

Retiring a game, a source or a page must never cost the reader a row they typed. We retire things routinely — a source moves, a page goes stale, a game shuts down — and their events are the only copy in existence. Nothing in the client deletes: no store prunes against the feed, knownGames only ever appends, and metaFor renders a lane whose game is gone rather than dropping it. The one place this could break is the load path. useCustom reads through validRecords, which drops a record that fails its schema, and the survivors are what the next write persists — so a record that stops parsing is not hidden pending a fix, it is deleted from the device by the act of opening the app. That is why CustomEvent.game is z.string() and not GameId: narrowing it to the enum reads like a tightening and would arm every future game removal to erase reader data on next launch. test/custom.test.ts § retiring a game, a source or a page pins it.

Shipping a new version

The shell is cached cache-first, so a reader with the tab open keeps the bundle they first loaded. An old app presented as current is the same failure as old events presented as current, so a waiting version is disclosed and reloaded on a tap (PRD F14, docs/ARCHITECTURE.md § Shipping a new version to an open page). Four things hold it up:

• sw.js must not skipWaiting() on install. It activates only on the skip-waiting message the reader's tap sends. Claiming an open page unasked runs the old bundle against the new cache and says nothing.
• __BUILD__ must stay in sw.js. scripts/build-static.ts substitutes a hash of the built shell for it, which is what makes a deploy's worker bytes differ and therefore detectable. It throws if the placeholder is gone — do not "fix" that by dropping the substitution. There is no CACHE_VERSION bump ritual any more; the cache name is a namespace, and per-build names would discard the stored feed an offline reader is reading.
• The feed is not part of the build id. It changes twice a day and needs no reload; announcing it as a new version teaches readers to dismiss the notice unread.
• The app never reloads itself. Someone may be mid-way through typing an event in.

Conventions

• Commit straight to main. This is a solo repo and its history is a single line; do not open a branch for a change unless asked for one. Committing still waits to be asked.

• Only ever commit your own work. Stage the files you changed, by path, and nothing else. The working tree may already hold edits, untracked fixtures or a half-finished experiment that someone else — the user, or another agent — put there and has not decided about yet; git add -A, git commit -a and git stash all sweep those into your change or out of sight. Authorship in the log then says you wrote something you never read, and the commit stops being the one coherent change the bullet below asks for. If unrelated changes are in the way, say what you see and leave them alone.

  Another agent may be working in this tree right now, not merely before you. The index is shared and it moves under you: a git status that was clean when you started can hold four staged files by the time you commit, and none of them yours. So read git status --porcelain immediately before every commit and treat anything you did not touch as a stop sign, and when the index already holds someone else's staged work, commit with an explicit pathspec — git commit -- <your paths>, which takes the working-tree content of exactly those paths and leaves the rest of the index where its owner left it. This is not hypothetical: on 2026-08-19 a session ran a sweeping commit that swallowed a 64-line source assessment another session had just finished, and published it under a one-line message about snapshot freshness. Both changes were fine; the log stopped being true.

  If you find your work inside someone else's commit, say so and ask before rewriting it. The fix is a reset --soft and two pathspec commits, and it is quick — but the commit you would be rewriting is theirs, the session that wrote it may still be running, and racing it for HEAD costs more than the mixed message does. Ask, then split.

• Commits are self-contained and succinct. One coherent change per commit, typechecking and passing tests on its own — a feature spanning layers splits as model → store → UI → docs, each step green by itself, even when that means widening a type in the model commit that only the UI commit uses. docs/FEEDBACK.md makes the same argument for adapters specifically: do not batch six games into one commit, because each one is a fixture and a test that has to prove itself. Succinct is about the message, not the change: a one-line subject in plain English, and a body that says why — the reasoning a diff cannot show — rather than listing the files it touched. Never reformat code the change did not touch; a formatter the project does not run buries a 100-line change in a 550-line diff.

• Zod schemas are the single source of truth for types. Derive with z.infer<>; never hand-write an interface that duplicates a schema.

• Every adapter ships a fixture in fixtures/<game>/ and a test asserting parsed output. This is how a source silently changing shape gets caught.

• Keep old fixtures when a source changes shape — the old one is the regression test proving the parser still handles the previous format. Fixtures are pinned and permanent; snapshots/ is the current page and gets overwritten. Do not conflate them.

• Colour is a token, and every token has two answers. The app ships dark and offers light (PRD F15), and the whole difference is a set of custom properties re-struck under :root[data-theme="light"] in styles.css. So a component names ink, hairline or soon and never a colour: a literal — bg-white, a hex, a hardcoded scrim — is a component that looks right in one theme and wrong in the other, and nothing will tell you which. The two things that genuinely cannot be tokens are the game hues, which are data (games.ts, and the reader's own), and the pre-paint script in index.html; both are handled in src/client/state/theme.ts and pinned by test/theme.test.ts. Two rules follow. The heat ramp must stay readable on whichever ground it is on — it carries meaning, and the dark theme's amber is 1.9:1 on paper, so each step is re-struck rather than reused. And the dark theme does not move: readableHue returns dark untouched by construction, because adding a theme is not a licence to redraw the one that shipped.

• The page is two columns past lg, and the split is the one below. What the page tells the reader to do — the next deadlines, tonight's dailies — pins to a rail on the left and stays put while the lists it shows them scroll beside it. Below that breakpoint it is one column in the same order. The focus bar goes at the top of that rail rather than full-width above both columns: it narrows what the rail holds, and a wide row of chips above everything pushes the headline deadline down the page. On a phone, and on the timeline, which has no rail, it is back at the top of the page — one render site per view, never both at once. The rail's rule belongs to the panel, not the column: the panel is short and the list is long, so a full-height divider would spend most of its length walling off a gap.

• Truncating a list is not re-sorting it. Each section shows LIST_CAP rows and offers "show all N". The rows below the cut keep their place in the order, stay counted in the header, and stay on the timeline — so the sorting groups, it never reorders rule below holds for what is hidden exactly as it does for what is shown. Expanding is per-visit state, not a stored preference: it is something a reader does while reading one list, not a statement about how the app should work.

• A game we add arrives switched off, and knownGames absent means unrecorded. Adding a source is our decision; a reader who plays two games did not ask for the other twelve. So a lane missing from prefs.knownGames is new to them and is hidden on sight (adoptNewLanes, PRD F8). The trap is the other reading: every install from before this existed has no knownGames at all, and treating that as "has been offered nothing" switches off every game they already read. Seeding records what is on their screen and changes nothing else. Lanes they invented (mygame:) are recorded but never hidden.

• The order games appear in is the reader's, and absent means unplaced. Nothing used to decide it: the focus bar, the settings list, the timeline's lanes and the dailies strip all rendered App's games, which is whichever game happened to hold the first event row. orderGames (src/client/state/gameOrder.ts) is now the one rule, and every surface goes through it — sorted on the name and never the LaneId, because the id is not what is printed, and through localeCompare, because < files hololive Dreams after every capitalised game. An absent prefs.gameOrder means the reader has never placed a game, not an empty order, so they get the alphabetical rule — the knownGames distinction again, and the same trap. Two properties are load-bearing rather than incidental. The result is always a permutation of the lanes it was given: a game dropped there is indistinguishable on screen from one the reader switched off, and switching it back on would not bring it back. And a lane it does not name trails the ones it does, which is what keeps a game we add later out of the middle of a hand-made order, and what keeps a retired source's slot warm rather than pruning it. games itself is left in feed order on purpose: adoptNewLanes diffs it and knownGames is seeded from it, so ordering it at source would let a display preference reach the code that hides a reader's games.

• Reordering is a settings-only affordance. The focus bar and the dailies strip are the fastest tap targets in the app, and a drag target on top of a tick target costs somebody a streak the first time it misfires — so the live surfaces are never draggable. Both a handle and ↑↓ buttons ship, because touch fires no drag events at all: the arrows are the mechanism and the drag is the pointer fast path, and being ordinary buttons is what makes the whole thing reachable by keyboard and screen reader without a second implementation of the same interaction.

• The settings panel is groups that state their own answer, and they ship closed. Five <details> groups — games, reading, what you see, your own, your progress — each with its current state on the summary line (Europe · Dark, 17 of 18 on · A–Z, plus finished, not started). It was one open block in two columns, which is readable at four games and not at eighteen, where the game list is eighteen rows of four controls sitting above the checkbox somebody came to tick. The state line is the half that makes collapsing honest: without it every question about how the app is set up costs a click, so the closed panel has to be a report and not a menu. Two rules follow. Do not default a group open to make something inside it findable — the empty states name the group as well as the switch, which is what makes that unnecessary. And the summaries are derived from prefs at render, never stored, so they cannot drift from the controls they describe. Native <details> for the reason the reorder arrows are ordinary buttons: keyboard and screen reader reach it without a second implementation. summary is not an a, button or [tabindex], so it needs its own :focus-visible rule in styles.css — the shared one does not reach it.

• A game's dailies stay together, and grouping never re-sorts inside a group. dailyGroups (Dailies.tsx) emits a game's standing chore followed by that game's repeating events, in the order those arrived; it replaced [...chores, ...repeating], which put a game's chore and its own login event at opposite ends of the strip. Collapsed, the grouping is adjacency only — no per-game headings, because that strip is the part of the page answerable in ten seconds and a heading per game makes it the tallest block on it, pushing "next to expire" down the page.

• A day already gone is still tickable, and the window bounds display and never data. People play at midnight and tick at breakfast, so catchUpDays (src/shared/daily.ts) offers the last CATCH_UP_DAYS game-days for a standing chore and for an event whose end was never announced — the two cases dailyDays cannot answer. It never returns a day past today: a tick claims you did it, and a future pip is a control for a claim that cannot be true, so it is absent rather than disabled. A tick older than the window stays logged and keeps counting toward the streak — nothing prunes the log against a window, here or anywhere else.

• Which view opens is the reader's answer. prefs.view is asked once on the first run (PRD F8) and written by the tabs from then on. It was component state, which meant a reader who preferred the timeline was put back on the list by every reload, with nothing to blame but the app forgetting. The stored answer is theirs; do not add a heuristic that overrides it.

• A list row is one target. The event row opens the event and does nothing else — status, effort, notes and the daily checklist all live in the detail sheet. A second control inside a full-bleed row target is a mis-tap waiting to happen, and a decorative chevron says "this opens" without adding a second stop for keyboard and screen-reader users.

• Sorting groups, it never reorders within a group. Every mode falls back to endingSoonestFirst, so choosing one can never cost the reader the deadline order the product exists for.

• Telling the reader to do something is not the same as showing it to them. showCompleted and showIgnored decide what they can look at; the "next to expire" headline and the dailies strip are instructions, so both drop anything done or ignored regardless (outstanding in src/client/state/lens.ts). Being pointed at a job you already finished is the bug either way. For the same reason "next to expire" reads the minimum end date rather than the head of the list, which under "doing first" is a different event entirely.

• The page states its own age unprompted, and reads it off the data. The footer says when event data last refreshed on every load, not only past the two-day threshold — a page silent about its age reads as current, and "how old is this?" has to be answerable before a countdown is worth trusting (PRD F7). freshness() in src/shared/feed.ts is the one definition: it takes the newest lastSuccessAt and never generatedAt, which is a build stamp that would call a fixture-backed calendar minutes old, and it treats a game as only as fresh as its oldest source, so one live wiki cannot vouch for a stalled sibling. Given that thirteen sources cannot be fetched from CI at all (§ Scraping conduct), this disclosure is the only thing standing between a reader and a confidently stale calendar — do not let a future change source it from the build clock.