qalisted

# QA Drive the real surface the way a user would and report what actually breaks — prove-on-artifact for behavior. Default is find-and-report; fix only when asked (`--fix`). ## Steps (paste as todos) 1. **Pick the surface and driver.** Web UI → a browser CLI available in the environment (`agent-browser`, Playwright — use what's installed; don't assume a bespoke daemon). CLI/TUI → run the binary. API → curl/HTTP. Name the driver you're using. 2. **List the critical paths** the change touches — the few user journeys that must work, plus the obvious edge cases (empty, error, boundary, unauthorized). 3. **Exercise each path on the real surface.** Perform the actions; capture evidence — screenshot, console/network log, exit code, response body. Observation is the test. 4. **Record defects with evidence.** For each: the steps, expected vs actual, and the captured artifact. No repro/artifact → not a confirmed defect. 5. **Report, ranked by severity.** Lead with blockers. If run with `--fix`, fix the root cause (see `investigate`), then re-run the failing path to confirm. ## Guardrails - A defect claim needs an artifact (screenshot/log/output). "Looks broken" without evidence isn't a finding. - Test behavior, not implementation. Drive the surface; don't grade the code. - Don't trigger destructive or irreversible actions (deletes, real payments, emails) while exploring — note them as untested rather than firing them.