Run project QA
Read .agents/skills/qa/SKILL.md and .agents/qa/config.json, then run project-qa run --root <repo>. If absent, use setup. --filter <id-substring> limits a follow-up run and repeats, so run a fixed CI case list with one --filter per case in one invocation. --base-url <url> selects a temporary website target without editing project settings. Check report.json for the effective baseUrl, tested revision and selected cases, then confirm the browser destination; API and auth URLs remain separately configured.
Use deterministic scripts. Interactive investigation can repair a failed journey, then put repeatable steps in its tracked suite. Never replace failed live behavior with mocked success.
Reports distinguish pass, fail and blocked, record Git revision, and keep screenshots private. Exit 0 means selected cases passed, 1 means failure, 2 means blocked. Start a fix from the failure details the runner prints and records, listed in the Runner API. When several journeys fail together, fix the first one, since later ones often share its cause. Report omitted cases and configured gaps separately. A passing journey proves its assertions, not uninspected visual quality or runtime health. Mouse or tap-only journeys do not prove touch holds, drags or scrolling. For a UI audit, use the installed ui-regression skill when available and report functional, rendered-quality and runtime verdicts separately.
Read sessions before recovering login, telemetry before production traffic, voice before microphone or speech journeys, and surface requirements for relevant components. Runner API defines executable cases and optional browser-network-cli diagnostics, including installation when needed.
Verify cleanup of created data and tabs. Keep saved sessions for the next run. Do not close personal browser tabs, log out providers after each test, or run suites sharing a stateGroup concurrently. Parallel runs each need their own server port and QA_STATE_DIR.
For native macOS permission onboarding, read the native workflow and complete its capture-readiness steps before GUI work. The agent must inspect real journey screenshots before claiming UI acceptance. Ask the user only for an OS action they must perform. The shared driver supplies exact AX window selection, native mouse input and scoped captures. Projects own selectors and app outcomes. Run nonmutating protocol and signed-fixture checks in CI, with installed grants only in a separately authorized manual-live run.