qa-test

QA-test web applications against user instructions, picking the right CLI per dimension. chrome-devtools-axi is the default driver: snapshot+ref loop with loud STALE_REF errors, Lighthouse, Core Web Vitals insights (LCP/INP/CLS), heap snapshots, and built-in emulation — plus contextual help[N] planning hints after each command. agent-browser handles only what AXI lacks: visual/DOM diff family (Myers + pixel), HAR export, encrypted auth vault, multi-session isolation, cloud providers, iOS Simulator, and the find text|label|role|placeholder|testid semantic locator fallback when AXI refs churn. Runs are resumable: every test step is a three-state checkbox in ./.plans/qa-test/<run-id>/PLAN.md, and the workflow reads prior state before acting.

Core rules

1. One run = one directory under ./.plans/qa-test/<run-id>/. <run-id> is YYYY-MM-DD-HHMM-<slug>. Never write outside this dir except for global CLI artefacts. Project-scoped only — fail if not inside a git repo.
2. Read state before write. On every invocation, list existing ./.plans/qa-test/, summarise prior runs, and ask before resuming, starting fresh, or overwriting. Apply the RESUMING.md protocol verbatim.
3. Default to chrome-devtools-axi; reach for agent-browser only when AXI lacks the capability. AXI handles functional, network, console, performance, screenshot, emulation, multi-page. Switch to agent-browser for: visual/DOM diff family, HAR export, encrypted auth vault, multi-session isolation, cloud providers, iOS Simulator, and the find text|label|role|placeholder|testid ref-stability fallback. Full mapping in TOOL-MATRIX.md. Mixing both in one run is normal — keep separate sessions per CLI.
4. Honour the non-negotiables in RULES.md. Stale-ref recovery, network-capture priming, wait content over wait timeouts, no eval with secrets, one perf trace per page.
5. Three-state checkboxes only. [ ] pending, [x] complete, [~] — <reason> skipped. Skipped items preserve the audit trail; never silently delete.
6. Evidence goes to disk, never to context. Screenshots, HARs, Lighthouse reports, heap snapshots, response bodies all land under ./.plans/qa-test/<run-id>/evidence/. Reference paths in RESULTS.md; do not inline binaries.
7. Auto-install on first run. If either CLI is missing, run scripts/preflight.sh which installs chrome-devtools-axi and agent-browser globally via npm and runs agent-browser doctor --offline --quick --json. Surface install + doctor output before proceeding.
8. Pin behaviour, not versions, in CI. Inside the per-run dir, write the resolved CLI versions into ENV.md so reruns are reproducible. Do not pass @latest in the auto-install command — the script pins what it installs.

Workflow

0. Sanity