gacha-event-tracker/docs/ARCHITECTURE.md

Architecture

Shape

One Bun process serves the static React build, exposes a read-only JSON API, and runs the ingestion scheduler on a timer. SQLite is the only datastore. There is no auth layer because there are no users.

                    ┌──────────────────────────────────────────────┐
   game wikis  ───► │  Bun process                                 │
   news pages       │                                              │
                    │   scheduler (every 6h)                       │
                    │        │                                     │
                    │        ▼                                     │
                    │   ingest pipeline                            │
                    │   fetch → parse → merge → validate           │
                    │        → reconcile → gate and publish        │
                    │        │                    │                │
                    │        │                    └──► quarantine  │
                    │        ▼                          │          │
                    │   ┌─────────────┐                 │          │
                    │   │  SQLite     │◄────────────────┘          │
                    │   └─────────────┘      /review (127.0.0.1)   │
                    │        │                                     │
                    │        ▼                                     │
                    │   GET /api/events        static React build  │
                    └────────┬─────────────────────────┬───────────┘
                             │                         │
                             ▼                         ▼
                        browser fetch            browser render
                                  │
                                  ▼
                        localStorage: progress, daily, ignored,
                        prefs, the reader's own games and events
                                         ← never leaves the device

That is the design, and the scheduler half of it is not built yet. What runs today is scripts/refresh-sources.ts on a Gitea Actions cron, writing snapshots to disk, and a feed built from those files rather than from SQLite — see § Today. The stages, the layering and the review gate are unchanged by that; only what wakes them up and where they land is.

The pipeline makes no third-party API calls beyond fetching source pages. There is no inference anywhere, at ingest time or in a request path.

Layout

src/
  server/
    index.ts            Bun.serve entry: static + API + scheduler bootstrap
    routes/
      events.ts         GET /api/events, /api/events.json
      games.ts          GET /api/games
      review.ts         /review UI + approve/reject (localhost-bound)
      health.ts         GET /api/health
    db/
      client.ts         bun:sqlite handle, WAL, pragmas
      migrations/       NNN-name.sql, applied in order at boot
      queries.ts        all SQL lives here — no SQL in route handlers
  ingest/
    scheduler.ts        timer + jitter + per-source lock          [not built]
    pipeline.ts         the 6 stages, orchestration only          [not built]
    html.ts             flat-table HTML reader (no dependency)    ✓ built
    dates.ts            ten deterministic date formats            ✓ built
    merge.ts            cross-source dedupe, corroboration        ✓ built
    sanitize.ts         the trust boundary — stage 2.5            ✓ built
    robots.ts           robots.txt parsing, matching, host cache  ✓ built
    snapshots.ts        raw snapshot cache + conditional headers  ✓ built
    validate.ts         zod parse + calendar sanity rules         [not built]
    reconcile.ts        diff vs published, confidence, conflicts  [not built]
    parsers/
      types.ts          SourceParser interface                    ✓ built
      game8.ts          game8.co article calendars                ✓ built
      wikigg.ts         wiki.gg mp-event templates                ✓ built
      akwiki.ts         arknights.wiki.gg mrfz-wtable             ✓ built
      fandom.ts         Fandom via the action=parse API           ✓ built
      bawiki.ts         bluearchive.wiki schedule (#eventtable & tabber) ✓ built
      holodori.ts       holodori.wiki Current/Past Events tables   ✓ built
      index.ts          parser registry                           ✓ built
    adapters/
      types.ts          Adapter interface, ParseContext           ✓ built
      index.ts          SOURCES registry, parseGame(), sanitize   ✓ built
  shared/
    schema.ts           zod schemas — the contract, both sides    ✓ built
    time.ts             clocks, urgency, region resets, captions  ✓ built
    games.ts            per-game name, hue, and reset clock       ✓ built
    daily.ts            which events repeat, and game-day keys    ✓ built
    effort.ts           effort estimates and the runway heuristic ✓ built
    custom.ts           the reader's own games and events (F13)   ✓ built
    feed.ts             the /api/events.json wire contract        ✓ built
  client/                                                         ✓ all built
    main.tsx            render + service worker registration
    App.tsx             shell, views, filters, onboarding gate
    api.ts              typed feed fetch, schemaVersion refusal
    sw.js               offline: shell cache, feed fallback
    manifest.webmanifest, icon.svg
    components/
      NextUp.tsx        the next three deadlines    (PRD F2)
      EventRow.tsx      row + meter + caption       (F2, F3)
      Meter.tsx         the depletion meter
      Legend.tsx        what the bars and colours mean
      Timeline.tsx      the board: pinned axis, lanes, bars   (F1)
      EventDetail.tsx   detail sheet, ignore action (F9)
      ProgressControls  status, effort, note        (F12)
      Dailies.tsx       today's strip, per-game reset clocks
      DailyChecklist    one repeating event's whole run
      Fireworks.tsx     the burst when the last daily lands
      GameFocus.tsx     one game at a time          (F4a)
      Controls.tsx      games, region, export/import(F4, F5, F6)
      Welcome.tsx       first-run games and view    (F8)
      Toast.tsx         undo an ignore
      UpdateNotice      a newer app is installed and waiting  (F14)
      YourOwn.tsx       the reader's own games, in settings   (F13)
      CustomForms.tsx   the game and event forms behind it    (F13)
      Colophon.tsx      credit, disclaimer, repo link
    state/
      storage.ts        namespaced, versioned localStorage
      useMarkSet.ts     ignores (and the superseded completions shape)
      useProgress.ts    status, effort, note, daily override  (F12)
      useDailyLog.ts    which game-days are ticked off
      usePrefs.ts       region, filters, focus, view, theme, onboarding flags
      theme.ts          dark/light: resolving it, applying it, hues on paper
      useCustom.ts      the reader's own games and events     (F13)
      gameMeta.tsx      lane id → name, label, hue; resolves custom lanes too
      sort.ts           deadline order, or what you're partway through
      lens.ts           who sees which rows — focus, outstanding, next-to-expire
      zoom.ts           the timeline's scale ladder; pure
      lanes.ts          how the timeline stacks — lanes, or one deadline queue; pure
      useAppUpdate.ts   is a newer build waiting, and taking it   (F14)
serve.ts                static server + /api/health              ✓ built
scripts/
  build-feed.ts         fixtures → public/data/events.v1.json     ✓ built
  build-static.ts       shell + worker into public/, build-stamped ✓ built
  parse-fixture.ts      run one adapter offline                   ✓ built
  refresh-sources.ts    fetch, cache, rebuild — the only network   ✓ built
fixtures/<game>/        checked-in raw HTML + expected parse output
snapshots/              the current page per source, rewritten by refresh

Request paths

Route	Purpose	Notes
GET / + assets	React SPA	Served from the bun build output
GET /api/events?from&to&game	Filtered feed	ETag + Cache-Control: public, max-age=300
GET /api/events.json	Whole published feed	Cheap; the client mostly uses this and filters locally
GET /api/games	Game metadata: id, name, color, lastUpdatedAt	Drives the freshness badges (F7)
GET /api/health	Per-source last-success, quarantine depth	For an operator, not the UI
GET /review	Quarantine review UI	Bound to 127.0.0.1 only
POST /api/review/:id/approve | /reject	Promote or discard a quarantined event	Same binding

Why /review needs no auth

Bun.serve runs two listeners: the public one on 0.0.0.0:PORT with the SPA and /api/*, and a second on 127.0.0.1:ADMIN_PORT with /review and /api/review/*. The review routes are not registered on the public listener at all — they are unreachable from off-box, so there is nothing to authenticate. This is the mechanism that satisfies "no logins" without leaving an open admin endpoint on the internet.

This is load-bearing. If someone later puts a reverse proxy in front of the admin port, or merges the two listeners "to simplify", the review UI becomes a public write endpoint. Any change in that area needs an explicit auth story first.

Data flow, concretely

1. Scheduler wakes every 6h (± jitter). For each source not fetched within its minIntervalMs, it acquires a per-source lock row and enqueues a run.
2. Pipeline executes the six stages in docs/INGESTION.md. Every stage writes to ingest_runs so a failure is diagnosable after the fact without re-running.
3. Publish upserts into events by stable ID, bumping version and updatedAt when any field changed. Events that vanish from a source are not deleted — they are marked status = 'delisted' so a source outage cannot silently empty the calendar.
4. Client fetches the feed, joins it against localStorage by event ID — progress, ticked days, ignores, and the reader's own events — and renders. That join is client-side only; the server never learns what the reader completed, skipped, or typed in.

Concurrency and failure

• One in-flight run per source, enforced by a lock row with a stale-lock timeout of 15 minutes.
• A source that fails keeps its previously published events. A failed run never deletes or blanks data — worst case, the game's lane goes stale and gets a warning badge (F7).
• Three consecutive failures for one source raises its health to failing in /api/health, and is what the refresh runner reports as broken. It never stops the schedule or the commit — a wiki being down for an afternoon is normal — but it does fail the run afterwards, because at two cycles a day that game's calendar has been served from a checked-in fixture for a day and a half.
• Raw snapshots are cached by content hash, so a parser change is always evaluated offline against stored pages rather than by re-fetching.

Deployment

Single process, single SQLite file, no external services at all.

PORT=3000
ADMIN_PORT=3001            # bound to 127.0.0.1
DATABASE_PATH=./data/events.sqlite
INGEST_INTERVAL_MS=21600000
INGEST_ENABLED=true        # false for local UI work — never touches the network
CONFIDENCE_THRESHOLD=0.8
BASE_PATH=/               # trailing slash; set when hosting under a subpath

INGEST_ENABLED=false is the default for local development. Frontend work should run against a seeded SQLite file and cost nothing.

Today

serve.ts serves public/ plus /api/health, and the feed is generated offline by bun run build:feed from snapshots/, falling back to fixtures/ for a source with no snapshot. It emits exactly the shape /api/events.json will, so the real server slots in without the client changing. Reads are confined to public/ by resolving the path and checking it stays inside the root — string-matching .. is not enough, because encodings and URL normalisation both change what the string looks like.

scripts/refresh-sources.ts (bun run refresh) is what fills snapshots/, and it is the only code here that touches the network. .gitea/workflows/refresh.yml runs it at 05:27 and 17:27 UTC and commits only when a page actually changed. It stands in for the unbuilt scheduler and enforces the same conduct in code — the 6h floor, one request, no retries, conditional headers, per-host spacing, robots failing closed. A source that has failed BROKEN_AFTER_FAILURES (3) cycles running is reported as broken and fails the run after the commit; see AGENTS.md § Scraping conduct for why that ordering is load-bearing.

Dockerfile builds and serves this; the image runs typecheck and tests during build, ships no source or toolchain, and runs unprivileged. .gitea/workflows/ci.yml and .gitlab-ci.yml run the same gates and publish it, while .github/workflows/ci.yml builds and deploys to GitHub Pages. GitHub only needs to build and serve, not refresh data sources; the scheduled refresh and container image publishing run exclusively on Gitea Actions.

Gitea Actions is GitHub-Actions-compatible, not GitHub Actions. That compatibility makes GitHub-shaped answers look right and then fail in specific places, so ask what Gitea does before changing CI, the registry or secrets. The directory name .github/ is not evidence of which forge runs a workflow: Gitea reads .gitea/workflows/ natively, and test/refresh.test.ts asserts on the workflow files by path. The divergences found so far, each one paid for once:

• The registry does not take the Actions token. A Gitea package belongs to a user or org, not a repository, so secrets.GITHUB_TOKEN carries no authority over it and permissions: packages: write authorises nothing. The image push logs in with an access token with the package scope, stored as REGISTRY_TOKEN — the comments in .gitea/workflows/ci.yml carry the detail.
• uses: takes full URLs (https://github.com/actions/checkout@v4); a bare reference resolves against the instance's DEFAULT_ACTIONS_URL.
• No gh CLI on the runner. refresh.yml dispatches ci.yml through Gitea's REST API (${{ github.server_url }}/api/v1) with curl.
• A push made with the Actions token does trigger workflows, unlike on GitHub, which suppresses them. refresh.yml dispatches ci.yml anyway, and the per-ref concurrency group absorbs the duplicate.
• No Pages. Gitea's deploy artefact is the container image, built without BASE_PATH; Pages is the GitHub mirror's job alone.
• Gitea push-mirrors to GitHub — see AGENTS.md § Conventions on pushing.

Hosting under a subpath

Assets resolve against a <base href> substituted at build time, the feed URL resolves against document.baseURI so deep links work, and the service worker derives its paths from its own registration scope. BASE_PATH=/gacha-event-tracker/ bun run build for any host that serves the app from a subpath (such as GitHub Pages); without it a subpath deploy 404s on every asset. GitHub's .github/workflows/ci.yml sets it automatically for Pages deployments, while Gitea's container image serves from /.

Offline

The service worker caches the shell and webfonts (cache-first) and the feed (network-first, falling back to the last copy seen). Countdowns run off the device clock, so the app stays useful with no network. Offline state is surfaced in the header and above the footer — stale data must never be presented as current.

First load

The critical path on a cold visit is short and worth keeping that way: index.html → styles.css + main.js in parallel → React mounts → the feed renders. Two things about it are deliberate.

The bundle is built with --production (package.json → build:js). Without that flag Bun ships React's development build, which is 222 KB of extra JavaScript, validates every element creation at runtime, and leaves StrictMode double-invoking effects — so fetchFeed runs twice and every reader downloads the feed twice. It shipped that way until 2026-08-20. See AGENTS.md § Commands for the trap in the obvious-looking alternative, which builds cleanly and does not run.

The feed is preloaded from the shell. fetchFeed cannot start until main.js has downloaded, parsed and mounted, so the one request the page exists to make was starting last — measured on a 4G/4×-CPU profile, the bundle finished at ~400 ms and the feed then ran to ~570 ms. Nothing about that URL depends on the bundle, so <link rel="preload" as="fetch"> puts it on the wire in the first few milliseconds and it is already in cache when React asks: the events are on screen at ~470 ms instead, and the loading state is gone rather than merely brief.

The crossorigin attribute on that link is load-bearing. as="fetch" must match the credentials mode of the fetch() that follows it, and a mismatch is not a no-op — the browser discards the preload and fetches the feed a second time, which is worse than not preloading. If that link is ever edited, count the feed requests in a real browser; nothing in the test suite can see this.

Compression is the server's job, and serve.ts — which is what the Docker image runs, and so what serves the deployed site — does it by negotiating accept-encoding per request. It matters more than it sounds: the bundle is 344 KB raw and 100 KB gzipped. A static host put in front of the built public/ instead must do its own; GitHub Pages did this transparently while it was the deploy target, which is why nothing here ever had to.

The theme, before the bundle arrives

The page is drawn dark by default and light when the reader has asked for it (PRD F15). Which one is decided by one attribute on <html>: styles.css holds the dark tokens on :root and re-strikes them under :root[data-theme="light"], so nothing in React knows a theme exists and no component holds a colour of its own.

Setting that attribute is the one part React cannot do in time. It mounts after the bundle has downloaded and parsed, which on a cold cache is long enough to show a reader who chose light a dark page, on every single load. So a small inline script in index.html reads the same localStorage prefs key the app does, sets the same attribute, and updates <meta name="theme-color"> — before first paint, and with a try/catch so storage being unavailable costs the reader the default theme rather than the page.

That makes the theme's ground colour a fact written in three files that cannot import each other: the stylesheet, state/theme.ts (which needs it for the meta tag), and the shell. test/theme.test.ts pins the three together rather than trusting them to be edited at the same time.

Shipping a new version to an open page

A cache-first shell is what makes the offline story work and what makes a deploy invisible: the reader this app is built for leaves the tab open for days, so a new game, a repaired parser or a corrected date reaches their device and then sits there unused. Presenting an old app as current is the same failure as presenting old events as current, so it is disclosed the same way.

build:static ──► sw.js stamped with a hash of the built shell
                        │
browser byte-compares sw.js on navigation, hourly, and when the tab is
revealed (registration.update)
                        │
   bytes differ ──► new worker installs, precaches, and WAITS
                        │
   registration.waiting ≠ null and a controller exists
                        │
   UpdateNotice: "A new version of Event Clock is ready."  [Reload] [×]
                        │
   Reload ──► postMessage {type:"skip-waiting"} ──► worker activates
          ──► controllerchange ──► location.reload()

Four properties this depends on, each of them load-bearing:

• The worker never calls skipWaiting() on install. Claiming an open page unasked leaves the running bundle and the cached shell on two different builds, with nothing on screen saying so. It activates only on the message the reader's tap sends. A first install has no worker to wait for and activates immediately regardless — and is deliberately not announced, since nothing is being replaced.
• The build id is derived, not remembered. scripts/build-static.ts hashes the built shell (index.html, main.js, styles.css, sw.js source) and substitutes it for __BUILD__ in the worker, so any shell change alters the worker's bytes and is therefore offered. The predecessor was a hand-bumped CACHE_VERSION, which had already been forgotten once. The substitution throws if the placeholder is gone, because the failure mode is silent.
• The feed is not part of the id. It is rewritten twice a day and served network-first, so new events reach an open page without a reload. Calling that a new version would teach readers to dismiss the notice unread.
• The cache name does not move with the build. Everything in it is refetched (cache: "reload", since none of these URLs are fingerprinted) on install, so a per-build bucket would buy nothing and would discard the stored feed — the copy an offline reader is reading.

src/client/state/useAppUpdate.ts holds the client half; sw.js and the hook cannot import each other, so test/update.test.tsx pins both ends of the skip-waiting handshake and the placeholder.

Deliberate non-choices

• No ORM. bun:sqlite plus hand-written SQL in queries.ts. The schema is five tables.
• No Redis / job queue. The scheduler is a timer and a lock row. Restarting the process resumes cleanly because state is in SQLite.
• No server-side rendering. The feed is small and cacheable; a static SPA is enough.
• No websockets. Events change on a scale of hours; a 5-minute cache is more than adequate.

Reader application (2026-10-16 redesign)

App.tsx exports ReaderApp; only the v2 reader store writes production progress. Legacy hooks remain as compatibility helpers, validators and tests, not mounted writers. ReaderContent coordinates shared Checklist/Timeline/Routines/Archive/Settings, account-aware detail adapters and timed archive transitions. components/redesign/ contains the feature pages. redesign.css scopes Relaxed and Structured via data-interface, with one application tree and independent theme/density.

state/reader/ owns migration, schema, reducer, store, clocks, accounts, archive, backups, restore, imports, routines and event selectors. Operations re-read the persisted envelope under Web Locks; an IndexedDB readwrite transaction serializes browsers without Web Locks. Failed locking/storage never publishes an optimistic success. Storage events update other tabs. Field timestamps protect newer edits and tombstones; account writes rebase changed routine choices. Legacy deltas reconcile on startup into the original migrated accounts; raw keys remain preserved.

Clock revisions freeze automatic metadata and overrides. Imported historical clocks are excluded from active clock resolution; history resolves each explicit revision. Display timezone lives in DisplayTimeContext and never changes the feed.

Events and redeem codes use separate validated static feeds. build-codes.ts reads pinned snapshots or fixtures offline and preserves firstObservedAt in snapshots/codes-first-seen.json. The refresh runner combines event adapters and count-only code sources while sharing robots, floors, per-host spacing and snapshot gates. No fake GachaEvents represent codes. Both feeds are network-first in the worker; manual refresh bypasses the browser cache. Shell update waiting/activation is unchanged.

Backups are outside the reader envelope. Restore validates first, writes a recovery export, prepares a pending transaction, then replaces live state. Startup finishes an interrupted transaction before opening the reader. Browser UI restore uses the same cross-tab lock as ordinary edits.

Ordinary HTTP / WireGuard hosts

Reader identity minting uses state/reader/ids.ts: 128-bit crypto.getRandomValues tokens for accounts, operations, reset revisions and backups. The redesign's initial randomUUID calls were a startup bug on ordinary HTTP network hosts; unlike localhost, these are not browser secure contexts, even when reached through a WireGuard tunnel. Existing IDs are unchanged and valid. Initialization preserves migration recovery before minting, so the failed first visit did not discard legacy bytes.

Regression verification uses a Chrome test hostname mapped to loopback with isSecureContext false, randomUUID unavailable and actual getRandomValues/IndexedDB available. Existing-save migration, new setup, editing, additional servers, reset override/revert, backup and reload all pass there.

Presentation correction (2026-10-17)

ChecklistEvent/NextDeadline render the approved list hierarchy; TimelinePage owns the shared axis for event spans and routine periods. RoutinesPage groups by reset or account, with native RoutineEditor dialogs. ArchiveDetail reads saved archive snapshots and never routes historical progress through the live event editor. Settings keeps one shared tree for both interfaces, with separate default-view/custom-content pages and direct game profiles. Structured is the absent-value default at prepaint, preference normalization, setup and appearance application. No state keys, event identities or stored clock revisions changed for this presentation correction.

Classification at the reader boundary (2026-11-01)

src/shared/classification.ts preserves source identity while resolving logical content destinations, pass families and TCG categories. The wire feed retains its original games/IDs. state/reader/contentRouting.ts normalizes classified state under the existing lock/save path, with persisted origin/counterpart mapping and timestamp rebasing. Source inventory is retained in the reader so pass cycles work with ordinary Events disabled. Current inventory takes precedence over remembered snapshots for live cycles; Archive snapshots keep their historical dates and clock references. No new route, fetch, localStorage key or server-side reader record is introduced.