Imported from TriliumNext/Trilium (.claude/skills/developing-standalone/SKILL.md). Install upstream with npx skills add TriliumNext/Trilium --skill developing-standalone. Copyright stays with the author.

Developing the standalone (in-browser) build

apps/standalone runs the entire Trilium stack in the browser with no server: the ordinary apps/client UI on the page, and @triliumnext/core — plain JS, the same code the server runs — inside a dedicated Web Worker on top of SQLite compiled to WASM (@sqlite.org/sqlite-wasm, not sql.js) persisted in the origin's OPFS through the SAHPool VFS. It ships as a PWA to app.triliumnotes.org (Cloudflare Pages, deploy-app.yml, PR previews included) and is the webDir of the Capacitor mobile app (developing-capacitor-mobile skill). There is no apps/standalone-desktop.

Layout

apps/standalone/
  src/index.html, main.ts         # page bootstrap: SW registration, leadership claim, worker start, standaloneApi
  src/desktop.ts                  # re-exports apps/client/src/desktop — the client entry
  src/sw.ts                       # service worker: caches assets, forwards local API requests to the leader tab
  src/leader_election.ts          # Web Lock decides which tab owns the database worker
  src/local-bridge.ts             # page ⇄ worker protocol (LOCAL_REQUEST/RESPONSE, WS_MESSAGE, HTTP_REQUEST, backup/restore)
  src/local-server-worker.ts      # the worker: opens SQLite, initializeCore(), routes requests
  src/ios-interceptors.ts         # iOS Capacitor only (see mobile skill)
  src/error-overlay.ts            # full-screen surface for fatal startup failures
  src/lightweight/                # browser implementations of every core provider + standalone-only routes
  src/services/                   # image provider, Capacitor HTTP handler
  src/stubs/                      # officeparser entry + Buffer polyfill, empty stubs for pdfjs/tesseract/puppeteer
  src/test_setup.ts               # boots core with the real lightweight providers under vitest/happy-dom
  public/_headers, manifest.webmanifest, favicon.ico
  e2e/multi_tab.spec.ts, extra_window.spec.ts
  vite.config.mts                 # root is src/; aliases, stubs, asset copies, worker + SW build, vitest config

Request flow

1. main.ts requires a service worker (secure context: HTTPS or localhost — otherwise it throws a message explaining that), registers ./sw.js with scope /, reloads once if the SW installed but doesn't control the page yet, and exposes window.standaloneApi (restore.importBackup, backup.downloadDatabase) — the client's escape hatch for things carrying a whole file, the twin of desktop's window.electronApi.
2. Leader election (leader_election.ts): the OPFS SAHPool VFS takes exclusive sync-access handles, and createSyncAccessHandle exists only in a dedicated worker, so exactly one tab per origin can own the database. A Web Lock (trilium-standalone-db, held for the tab's lifetime) picks it; only the winner calls startLocalServerWorker(); when it closes the browser releases the lock and a waiting tab is promoted. Where Web Locks are missing (some WebViews) the tab assumes it is alone.
3. Service worker (sw.ts): cache-first for hashed assets, network-first otherwise; requests under LOCAL_API_PREFIXES (/bootstrap, /api/, /sync/, /search/, from local-bridge.ts) are forwarded with LOCAL_FETCH to the leader client, found via a cached id, WHO_IS_LEADER / LEADER_REPLY probing, and one retry on NOT_LEADER (leadership can flip while a request is in flight). Backup downloads stream over a MessageChannel from the leader. It must let /_trilium_native_http/ fall through (Android native proxy) and has a defensive guard for capacitor://.
4. The worker (local-server-worker.ts): waits for the previous worker's OPFS access handles to be released (waitForSahPoolRelease — a reload races the browser's asynchronous release, and installOpfsSAHPoolVfs() reacts to an acquisition failure by deleting the pool directory, so the pool must be acquirable before it runs — a wait that reaches its deadline ends the boot instead of installing anyway), then installs the SAHPool VFS once — failure is a fatal startup error painted by error-overlay.ts, never an in-memory fallback (TRILIUM_INTEGRATION_TEST=memory builds seed a fixture into the pool, so writes survive reloads within a test). It then calls initializeCore({...}) with the lightweight/* providers, registers createConfiguredRouter() (BrowserRouter mimicking Express, path params + query strings), then answers LOCAL_REQUEST messages. DbLock (lightweight/db_lock.ts) serialises createAsyncRoute handlers (imports) that hold a transaction across awaits against the single connection — synchronous routes keep a zero-overhead path. WS_MESSAGE replaces the WebSocket in both directions. Anything that escapes after startup is reported as WORKER_ERROR for one background task; before startup it means the app never came up and error-overlay.ts paints it (the async worker init is invisible to main.ts's try/catch).

The provider set (src/lightweight/)

initializeCore() takes every environment-specific capability as a provider; the browser build supplies its own for each. A new provider or provider method in core needs an implementation here, or standalone will not boot — the server-side one in apps/server/src/*_provider.ts is the pair to mirror.

Standalone-only routes (integrity checks, the LLM channel, upload handling that arrives as a buffer instead of a temp file) are registered in lightweight/browser_routes.ts; everything else is core's shared route table (adding-internal-api-route skill).

Bundle discipline (vite.config.mts)

The worker's top-level imports are its startup cost — prefer dynamic imports for anything not needed to answer /bootstrap (the LLM skills and restore paths already do this).

• officeparser is aliased to src/stubs/officeparser_entry.ts, which loads the package's Node ESM entry plus the buffer polyfill (buffer_polyfill.ts, must be imported first) instead of the prebuilt browser monolith that inlines pdfjs (2.7 MB → 0.4 MB); pdfjs-dist, tesseract.js, puppeteer are aliased to stubs/empty.ts. office_preview.spec.ts guards the alias against upstream layout changes.
• @sqlite.org/sqlite-wasm and @triliumnext/core are excluded from optimizeDeps; the sqlite .wasm and OPFS async proxy are copied unhashed so the worker can resolve them by convention; apps/server/src/assets/** (minus the User Guide) is copied to server-assets/.
• The service worker keeps a stable sw.js URL (registering a hashed one would orphan the old worker); everything else under src/ is content-hashed. public/_headers matches: immutable for /src/*, must-revalidate for /sw.js, plus Cross-Origin-Opener-Policy: same-origin.
• Vite root is src/, base is "" — paths in config and coverage globs resolve relative to src/.

Where a dynamic import actually splits — and where it does nothing

Core is not code-split the way its module layout suggests. Rollup folds the packages/trilium-core modules into a single ~2 MB chunk (named after an arbitrary constituent, e.g. abstract_provider-<hash>.js) that ~33 core modules — becca_loader, browser_routes, crypto_provider, the migrations — import statically, and the worker pulls in at startup. Anything reachable from that graph lands in the eager chunk no matter how the leaf that uses it imports it.

So converting a static import to await import() inside a core module often defers nothing. Lazy-loading sax in enex.ts produced a byte-for-byte identical standalone bundle — same chunk content hash, same total — because sax was reachable statically elsewhere in the eager graph. The prediction that it would save ~12–15 KB came from reasoning by analogy to another module; it did not survive an A/B build, and was reverted.

It is not a universal no-op, though. await import("officeparser") inside services/office_preview.ts did split out (officeparser.browser-<hash>.js, ~2.6 MB, fetched lazily), because nothing else in the eager graph imported officeparser — that one dynamic import was the dep's only path in.

The rule: a dynamic import inside core splits a dep out only if the dep has no other static importer in the core graph. Deferring at the route or service boundary (the import.ts route that statically imports enex.js) is the reliable lever. Either way, confirm with an A/B build — compare chunk hashes and grep the built output for a real import( — and don't bother at all for a ~40 KB dep inside a chunk that always loads.

Running

pnpm standalone:start                                # vite dev
pnpm --filter standalone build                       # dist/ (needs NODE_OPTIONS=--max-old-space-size=4096, set by the script)
pnpm --filter standalone start-prod                  # build + vite preview on :8888
TRILIUM_INTEGRATION_TEST=memory pnpm --filter standalone build   # OPFS DB seeded from a fixture (what e2e uses)
pnpm --filter standalone e2e                         # Playwright: multi-tab leader handoff, extra windows

?safeMode and other TRILIUM_* toggles are passed as URL query parameters (platform_provider.ts).

Testing

pnpm --filter standalone test <pattern> — Vitest, happy-dom, real sqlite-wasm. test_setup.ts shims wasm loading (happy-dom's fetch refuses file:// and its Response breaks instantiateStreaming) and boots core with the real lightweight providers, so a spec exercises the same code path as the browser. local-server-worker.ts is excluded from coverage on purpose (it boots a second core). This suite is also the second runtime for every packages/trilium-core spec — green under pnpm --filter server test is not proof; the cross-runtime traps (capability gaps, provider differences, serialize(), backup) are in the writing-unit-tests skill.