e2e

E2E testing (Playwright)

Context

The Playwright suite that ../testing.yml runs against every PR's Amplify preview. ~50 spec files cover the app end to end; this is a representative slice showing the infrastructure around the tests, not the full suite.

What's here

• playwright.config.ts — worker count derived from CPU cores, three Chromium projects (default, @serial-tagged tests forced to a single worker, everything else parallel), CI-only blob/junit reporters for sharding, and global setup/teardown for provisioning a test workspace.
• support/infra-classify.ts (+ infra-classify.test.ts) — a small, dependency-free module that decides whether a failing test was a real regression or transient infra (a 5xx on our own hosts, a dropped connection). It's deliberately free of any Playwright import so it can be unit-tested with Vitest, and its guarantees (a real failure always blocks; infra noise never does) are tested directly rather than only observed through flaky CI runs.
• support/fixtures-base.ts — a custom test/page fixture that blocks the app's live-updates (SSE) connection so waitForLoadState('networkidle') can resolve, reuses worker-scoped auth state, and wires the infra classifier in: it watches responses/failed requests during a test and, on an unexpected failure, attaches the observed 5xx/network errors so the release-gate reporter can tell the two failure modes apart.
• support/fixtures-parallel.ts — worker-scoped authentication plus per-worker unique-data generators (generateTestId, generateTestEmail, createEmployeeData), so parallel workers never collide on the same test data.
• support/auth.ts — login/logout helpers with resilient selectors (test ID → accessibility role/label → CSS fallback, in that order) and gateway-error detection, so a flaky login doesn't masquerade as a broken test.
• support/navigation.ts — retry logic for hash-based SPA routes: a cold page.goto() to a hash URL doesn't always trigger the client router on CI, so this detects a redirect-to-dashboard and recovers via a sidebar click or a client-side hash change before falling back to a full re-navigation.
• tests/team.spec.ts — one representative spec built on the fixtures and navigation helper above.

Omitted

• The other ~45 spec files — they follow the same pattern as team.spec.ts (navigate via a helper, assert on visible content) applied to different pages, so including more would repeat the same shape.
• global-setup.ts / global-teardown.ts, support/workspace.ts, support/parallel-auth.ts, support/env.ts, and setup/* — these provision a real test workspace and credentials against internal infrastructure and aren't generalizable outside that environment. Because they're omitted, the files here won't run standalone (consistent with the rest of this repo).
• reporters/verdict-reporter.ts — the CI-facing consumer of infra-classify.ts's output; omitted since the classifier itself is the interesting part.

Anonymization

Real hostnames, the live-updates subdomain, and localStorage/cookie key names have been replaced with generic placeholders. navigation.ts and team.spec.ts originally asserted on real (German) product copy for a specific feature; both are rewritten here around a generic "team roster" page with English content checks, since the literal UI text was more identifying than the anonymized component/prop names used elsewhere in this repo.

Design decision

infra-classify.ts is kept free of any @playwright/test import even though its only two callers are Playwright fixtures/reporters. Keeping the classification logic pure means the load-bearing guarantees (a real failure always blocks the release gate; infra noise never does) are covered by fast Vitest unit tests against literal Playwright error strings, rather than only being verifiable by reproducing flaky CI runs.