reader/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What this is

Self-hosted book library + media server (books, movies/TV, music, audiobooks). Three independent parts, no root package.json:

• server/ — Node ≥22 (CI uses 24), TypeScript run directly via tsx, Fastify, built-in node:sqlite.
• web/ — React 19 + Vite client. The same build is served by the server and bundled into the Android app.
• android/ — Kotlin WebView shell (+ native pager, TTS, Media3 playback). Contract with the web bundle is in android/contract.md.

Docs (mostly Chinese) live in docs/; docs/architecture.md explains why things are designed the way they are and docs/ui.md is the UI spec — read the relevant section before changing shelf/library/reader behaviour.

Commands

Run from repo root (or cd into the package). CI runs install → typecheck → test → build per package: GitHub Actions .github/workflows/ci.yml (PRs and every branch push; also UI review, Android shell check + debug APK) and CNB .cnb.yml. main is the long-lived development branch and never releases on push: .github/workflows/release.yml runs only for v* tags or a manual run, runs CI first, then publishes the multi-arch (amd64 + arm64, native runners) image to Docker Hub hectorqin/reader-pro (x.y.z/x.y/latest + sha-* for tags; branch name + sha-* for manual runs); it needs the DOCKER_USERNAME/DOCKER_PASSWORD secrets. android/scripts/check-shell.sh checks the JS bridge contract: only @JavascriptInterface methods count, and each must be declared in web/src/android-bridge.d.ts (method syntax, name?(): T for optional).

npm ci --prefix server && npm ci --prefix web

# Dev: two terminals. DATA_DIR must NOT be inside BOOKS_DIR (server refuses to start).
BOOKS_DIR=/tmp/reader-dev-books DATA_DIR=/tmp/reader-dev-data npm run dev --prefix server   # :5888
npm run dev --prefix web    # http://localhost:5174, proxies /api to READER_SERVER (default 127.0.0.1:5888)

npm run typecheck --prefix server && npm test --prefix server && npm run build --prefix server
npm run typecheck --prefix web    && npm test --prefix web    && npm run build --prefix web

Single tests:

• Server (node:test): cd server && npx tsx --test test/media-api.test.ts (add --test-name-pattern="..." to filter).
• Web has two runners. Most files are Vitest: cd web && npx vitest run test/foo.test.ts. A few files written for node:test (api, asset-url, paginator, window, windowed-toc) are excluded from Vitest in vite.config.ts and run with npx tsx --test test/<file>. When adding a node:test file, add it to both the test script and the Vitest exclude list.

UI review (real Chromium via CDP; needs npx playwright install chromium or CHROME_PATH): npm run ui:review --prefix web (builds, then screenshots + measured assertions into git-ignored artifacts/ui-review/react-app/, result in verification.json). Also ui:review:reader and ui:sources. Never commit screenshots to docs/. Run this and look at the screenshots after layout changes.

Docker image (context must be repo root): docker build -f server/Dockerfile -t reader:local . Android: sh android/scripts/build-web-assets.sh && cd android && ./gradlew :app:assembleDebug. Icons: add path to web/tools/icons/paths.mjs (use this project's own English names — test/icons.test.ts rejects names borrowed from other icon sets), then npm run icons.

Server architecture

• main.ts wires an AppContext (config, db, Scanner, services) and passes it to http/app.ts buildApp; routes are in http/routes/*. Tests build the same app against a temp BOOKS_DIR/DATA_DIR (see test/api.test.ts).
• Paths: every filesystem access goes through lib/paths.ts resolveInside() (throws PATH_TRAVERSAL); the scanner does not follow symlinks.
• Book identity: id = sha256(lower(dc:identifier) + '\0' + sha256(fileBytes)), not the path — renames/moves keep progress. book_files (many) → books (one); missing files are flagged missing=1, not deleted.
• Scanning (indexer/scanner.ts): mtime+size first, content hash only if that changed; concurrent scan() calls share one Promise.
• Metadata precedence: manual overrides (metadata_overrides, applied at read time in ShelfService.toDto) > embedded > provider (fills gaps) > filename parsing. The scanner may freely rewrite books because manual edits live elsewhere. Filename parsing is deliberately conservative: without evidence, the whole string stays the title and author stays empty.
• Book DTOs are assembled only in ShelfService (toDto/getMany); other services (sync, browse) call it instead of building their own rows. "On shelf" means user_books.hidden = 0 AND a non-missing file exists — keep that definition consistent everywhere.
• Sync (services/sync.ts): last-writer-wins, note deletes are tombstones (deleted=1), each push batch is one transaction, and the POST response is the merged pull (the client must not GET afterwards).
• Schema (db/schema.ts): plain SQL; changes must be additive (MIGRATIONS_SQL, columns with defaults). Self-hosted users upgrade late and older builds may run against newer DBs.
• Sources/plugins (sources/): local library, OPDS, and process-based extension plugins (plugin-manager.ts, process-plugin.ts; example in examples/plugins/demo-chapters).
• Media (media/) is isolated: separate media.db, started asynchronously in a worker after reading is already serving. Routes return 503 MEDIA_STARTING / MEDIA_UNAVAILABLE while the reading side keeps working. Scans don't hold reader.db locks; publishing atomically swaps in a full snapshot in its own thread (failed/partial scans never replace the catalog); read-only catalog queries run on a separate query thread. Permissions are re-checked on every access/stream. reader.db and media.db (plus signing key) must be backed up/restored as a pair. See docs/media-implementation.md, docs/media-write-isolation.md.

Web architecture

• main.tsx creates AppRuntime (app/runtime.ts: platform choice Web/Android, ReaderApi, MediaApi, player, SyncEngine, OfflineStore, settings, QueryClient), then renders providers + RouterProvider.
• Hash routing only (createHashRouter, app/router/routes.tsx): neither the server nor the file:// WebView can serve client-side fallbacks. New links use #/....
• Single bundle, no code splitting, base: './' (see vite.config.ts) because Android loads it from file://. Don't introduce dynamic chunks.
• Server data via React Query; shared client state via Zustand; long-lived resources (player, theme, sync) owned by Providers/services and cleaned up there. New pages go in features/<domain>/pages; cross-domain code goes through shared/ or app/. Don't add to the old Screen dispatch (app.ts, app/legacy-compat.ts exist only for Android diagnostics/old tests).
• Reader is imperative: features/reader/pages/ReaderPage.tsx mounts ReaderScreen (ui/reader-screen.tsx) in an effect. Book content lives in a Shadow DOM with pagination measured via the DOM (ui/reader-view.ts, render/). Don't rewrite the stage as a React subtree; only the chrome is React.
• Theming: ThemeController (shared/theme/theme.ts) writes data-theme/data-mode on <html>; components use --color-* tokens from styles/theme-tokens.css (--reader-* and media aliases map onto them). Read/save theme via readThemePreference/saveThemePreference/resolveAppTheme (and useThemeState from shared/theme/use-theme.ts in React), never storage directly. New colors/spacing/durations become tokens, not literals.
• UI is built from the project's own components (ui/toolkit.tsx, ui/sheet.tsx as the only modal type, ui/action-menu.tsx for every ⋯ menu, ui/dialog.tsx, ui/pager.tsx) with Lucide/own icon font. Mantine was removed; some docs still mention it.
• Lists are server-paginated with the page in the URL (docs/ui.md §3.5). After a write, re-fetch and only then report. The client must not reconstruct facts the server already knows (e.g. book↔file joins); write by book id.
• Offline mirror (store/offline.ts) is additive-only; removals must call offline.removeBooks explicitly.