Testing

Three levels, each with a different job.

bun run test:unit    # vitest: unit + runes projects, node env
bun run test:e2e     # playwright, chromium

Unit tests

Vitest, node environment, *.test.ts sitting next to what it tests.

This level covers the framework-agnostic logic — the parts where a subtle bug is invisible in a screenshot:

• sync/reconcile.ts — merge order and conflict resolution
• watch/stream-format.ts — parsing quality, codec, size out of addon labels
• player/codec-support.ts, player/external-player.ts — probes and deep links
• addons/catalog-queries.ts — orchestration, against a fake client
• services/container.ts — scope rules, including the ones that must throw
• server/safe-fetch.ts — the SSRF address blocks

Keep these pure. A class earns a test by having state; a pure function earns one by having edge cases, and most of these are the second kind.

Construct services with fakes, or provide() them into a throwaway container, rather than vi.mock-ing a module path.

Rune modules

A second Vitest project, runes, runs *.svelte.test.ts files with the Svelte plugin, so .svelte.ts modules compile. It exists for rune state that is logic rather than UI : the sync store's queue, flush debounce, grace period, cursors, owner scoping, cross-tab broadcast and profile-switch races.

It stays in the node environment. The store needs IndexedDB (fake-indexeddb), BroadcastChannel (native in Node) and a document that fires visibilitychange (a twelve-line stub in src/lib/sync/test-setup.ts); a full DOM shim would be a much larger dependency for that.

Two things that bite:

• The store is a module-level singleton. Each test gets a fresh module (vi.resetModules()) and a fresh IDBFactory, and every instance a test creates is detached afterwards. One left attached keeps its listener on the shared document and answers for the next test.
• Under fake timers, vi.waitFor advances the clock between checks, which fires the store's own timers. Use vi.advanceTimersByTimeAsync(0) to flush, or move only Date with vi.setSystemTime when the behaviour is time-based.

End-to-end

Playwright, in e2e/. After any UI or route change, run it — bun run check and bun run lint do not catch a bad reactive access, a hydration mismatch, or a broken remote call.

It runs against a production build on :3000 (bun run build && bun run start), not vite dev: a cold dev-server compile made the run flaky. It reuses an existing server on :3000 and starts one otherwise, so a bun run dev on :5173 is untouched either way.

The reuse is blind: a leftover bun run start from an older build answers the run, and the suite passes against code you no longer have. Before a verification run, stop anything on :3000 or rebuild it (lsof -ti :3000 | xargs kill).

It needs NUVIO_TEST_EMAIL and NUVIO_TEST_PASSWORD in .env — see .env.example.

The whole suite shares one auth token (e2e/auth.ts memoises the password grant), because the real api.nuvio.tv rate-limits. Do not re-run the full suite gratuitously; run the spec you touched.

Add a spec when you add a screen or a flow.

Zero console errors

Every spec that loads a page uses collectRuntimeErrors (e2e/errors.ts) and asserts the list is empty. An uncaught exception or a genuine console.error fails the test.

Do not widen the IGNORE list to make a test pass — fix the error. The only pre-approved ignores are third-party asset 404s: posters, favicons, net::ERR_.

Two timing notes:

• A page with a playing <video> never reaches networkidle, so bound that wait: waitForLoadState("networkidle", { timeout: 8000 }).catch(() => {}).
• Give async errors a waitForTimeout beat before asserting, or you assert before the error has been raised.

Accessibility

e2e/a11y.spec.ts runs axe over every main route and a couple of open-overlay states, asserting zero WCAG 2 A/AA violations, plus skip-link and focus-on-navigation checks.

New screens go in its pages list. Fix what it flags rather than filtering the rule — a filtered rule is a permanent exemption nobody revisits.

Coverage

bun run test:unit:coverage

Istanbul. Coverage is a diagnostic, not a target: a number going up because a test exercised a line without asserting anything is worse than the number staying put.

CI

• Code Quality runs the prek hook set on every PR.
• Test runs the unit suite.
• Nuvio API contract runs daily, checking the committed spec snapshot against the live one. See The Nuvio API spec.
• Docker builds linux/amd64 and linux/arm64 per PR and on main.