UI End-to-End Tests (Playwright)

Status (2026-10-01): the bundled local web dashboard was removed from the CLI (#5530), together with the tests/ui/ Playwright suite, its fixtures and the pytest-playwright test dependency. The CLI has no web UI to test. A replacement UI built on the planned Mission Status Read API (#5528) will live in its own repository and carry its own browser-driven suite. Restore the old suite from git history (mission playwright-ui-e2e-bootstrap-01KWX72W, issue #1008) if it is useful there.

Why browser-driven tests

Backend and API tests cannot catch a bug where the server responds correctly but the browser renders it wrong. That is exactly what shipped in PR #970: 338 backend tests and every architectural test passed while the dashboard's WP-card click-through modal silently dropped the agent identity, because no test layer exercised the click-then-render flow. CLAUDE.md's "Never claim the frontend works without Playwright proof" rule comes from that incident.

Lessons to carry into any future UI suite

Boot the app hermetically. Run it in-thread on an ephemeral port against a synthetic project root; never reuse a singleton that detaches a process or kills siblings on a shared port range.
Seed fixture data in the exact shape the reader consumes, and keep that synthetic shape in one shared fixture so every test agrees with it.
Wait for real state, never a fixed sleep.
Assert the pre-interaction baseline first, so a later "populated" assertion proves a real transition.
Scope every assertion to the container under test, never page-global: a field that also renders elsewhere on the page makes a page-global check pass even when the feature under test is broken.
Prove non-vacuity before committing: break the render path (not the fixture data) and confirm the test fails, then revert.
Run headless (PWHEADLESS=1) and record the command and result in the PR.