E2E Tests — ayunis-core-e2e

Browser tests against the real stack with deterministic LLM mocks. When the Proportional Workflow requires browser coverage, the change is done when its focused e2e spec passes. Use E2E for journeys and boundaries that lower-level tests cannot credibly prove; do not add an E2E test merely because a change is visible to a user. Architecture and history: ayunis-core-e2e/PLAN.md.

Stack setup (once per session)

Tests need a running ./dev stack in e2e mode plus the global seed:

./dev up --slot 2 --e2e                    # from the repo root; pick a free slot
(cd ayunis-core-backend && pnpm seed)

• --e2e writes MOCK_INFERENCE=true (all LLM providers become mocks that reply {provider}::{model}) and ORG_EVENTS_WEBHOOK_URL= (no outbound webhooks) into .env.dev.
• Check slot availability first: docker ps --format '{{.Names}}' | grep ayunis-dev. Never touch someone else's running slot.
• Env-change gotcha: if the stack was already running without --e2e, the backend must fully restart (./dev down && ./dev up --e2e). A nest --watch respawn is NOT enough — the watch supervisor caches env (appsignal.cjs preloads dotenv) and children inherit it.

Running

pnpm --filter ayunis-core-e2e test                       # full suite
pnpm --filter ayunis-core-e2e test --grep "<pattern>"    # one spec/test
pnpm --filter ayunis-core-e2e run lint                   # selector/flake policy
pnpm --filter ayunis-core-e2e run typecheck
pnpm --filter ayunis-core-e2e run openapi:generate       # refresh generated API client

Non-default slot: E2E_BASE_URL=http://localhost:30N1 pnpm --filter ... (API and Mailcatcher URLs derive from the slot; see src/config.ts).

Temporary PR-review screenshots/GIFs use pnpm --filter ayunis-core-e2e run pr-media:capture. Scene definitions are not committed to product branches; use the pr-media skill to put .pr-media/scenes.ts on the disposable pr-media/pr-<n> branch for the PR.

For visually meaningful frontend changes, use the pr-media skill when the result materially helps reviewers evaluate the PR. Include only the viewports and interactions needed to demonstrate the change.

Fixtures — import from src/fixtures/test, never @playwright/test

import { test, expect } from '../../src/fixtures/test';

• org (worker-scoped) — every worker gets its own org created via the API: admin user (org.admin.email / org.admin.password), email confirmed, one permitted language model set as org default (org.defaultModel), welcome video + personalization wizard dismissed. Tests in a worker share it; workers never collide — this is what makes fullyParallel safe.
• page — starts authenticated as the worker org's admin. Specs that test login/registration opt out: test.use({ storageState: { cookies: [], origins: [] } }).
• api — authenticated APIRequestContext for setup and side-effect assertions (e.g. assert a toggle persisted via GET), not for the behaviour under test.
• publicApi — unauthenticated APIRequestContext for public-endpoint behaviour against the backend API URL.
• mail — Mailcatcher client; mail.extractLinkToken(email, '/accept-invite') polls for the email and pulls the token from the frontend link.
• pageErrorGuard (auto) — any uncaught page error fails the test. Opt out per test with test.use({ allowPageErrors: true }) only with a comment why.
• Chat flow helpers (src/flows/chat.flow.ts): sendMessage, startThread (returns threadId), sidebarThreadItem.
• Mock replies are assertable: expect(...).toContainText(${org.defaultModel.provider}::${org.defaultModel.name}) proves routing end to end.

Selector policy (ESLint-enforced)

• Only getByTestId / getByRole. Text selectors are banned: the UI defaults to German and Chrome auto-translate can rewrite DOM text.
• No page.waitForTimeout, no { force: true } — wait on locators or expect.poll.
• Missing a stable hook? Add a data-testid to the frontend in the same PR — that is part of building the feature, not test scaffolding. Naming: <feature>-<element>[-<action>], kebab-case (register-email, invite-accept-submit, confirmation-confirm).

Source layout

src/fixtures/       Playwright fixture composition only
src/clients/generated/ Orval-generated backend API client (do not edit)
src/clients/api/    Semantic wrappers around generated endpoints
src/clients/*.ts    External clients/adapters, e.g. Mailcatcher + Playwright mutator
src/factories/      Test data creation composed from clients
src/flows/          Reusable UI journeys/page interactions
src/assertions/     Shared assertion helpers when repetition appears
tests/<domain>/     Product journeys, one journey per spec file

Adding a spec for a new feature

1. One user journey per file under tests/<domain>/. Copy the closest reference:
   • CRUD + dialogs: tests/chat/thread-crud.spec.ts (rename/delete via shared dialogs)
   • Email-driven flow: tests/auth/invite.spec.ts, tests/auth/password-reset.spec.ts
   • Public form journey: tests/auth/register.spec.ts
   • Admin mutation with API assertion: tests/admin/instructions.spec.ts
2. Prefer generated endpoint calls behind src/clients/api/; do not hardcode backend routes in specs. Compose test data in src/factories/, reusable UI journeys in src/flows/, and assert behaviour through the UI.
3. Unique data per test (Date.now() suffixes) — tests share the worker org.
4. Run the focused spec, lint, and typecheck. Run the full suite when the change affects shared fixtures, flows, clients, or broad behavior; otherwise rely on CI for the full-suite gate.

Debugging a failure

1. Read the failure output, then look at the screenshot — test-results/<test>/test-failed-1.png (readable directly).
2. test-results/<test>/error-context.md has the accessibility snapshot at failure time; trace.zip has the full network/DOM timeline (pnpm exec playwright show-trace <path>; the *.network files inside are JSON if headless analysis is needed).
3. Backend side: ./dev logs backend --tail 200 (or .dev/slot-N/backend.log).
4. A request stuck "in flight" on a cold dev stack is usually a Vite on-demand transform — the warmup project (tests/setup/warmup.setup.ts) should prevent it; suspect real bugs before suspecting the suite.
5. If the app is genuinely broken, that's the suite working: report the bug, keep the intended behaviour as test.fixme with a comment, and reference it (pattern: "deletes the currently viewed thread" in tests/chat/thread-crud.spec.ts).

Gotchas

• The seeded fixture (admin@demo.local, Usage Org users) is shared global state — specs that mutate it don't parallelize; prefer worker orgs. Rich seeded scenarios (usage limits, academy states) are Phase 5 (@seeded-org).
• Local runs cap at 4 workers and disable video deliberately (CPU contention causes 30s-timeout cascades) — don't raise them to "speed things up".
• CI (.github/workflows/e2e-tests.yml) builds backend/frontend in upstream jobs, downloads those dist artifacts into the e2e job, and serves the built frontend from the backend (same origin, port 3000) — dev-server-only behaviour won't reproduce there.