ui-hunt

Load the qa skill. Target: $ARGUMENTS (else the current repo).

1. python3 ~/.claude/skills/qa/history.py <repo>: the top surface × mechanism rows are the hunt list.
2. For each top row, enumerate every instance in code (shared component → all its consumers; status mapping → every branch; list → empty/error/partial states). Count first.
3. React/TanStack repos: hunt.py with the repo's .qa/invariants.json. Other stacks: parallel agents, one per top surface, each verifying candidates at file:line.
4. Prove in the browser with qa verify <repo> --base URL / playwright-cli, driving the ui-stress states (empty, error, slow, long text, 320px, denied).
5. Group findings by shared cause: one fix at the shared component, then re-measure every consumer. Report: verified defects (P1–P3, file:line, what the user sees), the shared causes, and what stayed unmeasured.

Close with (these were the follow-ups every studied run of this command got): acted X of N over the leads, each one test written / fixed (commit) / issue (URL) / dismissed with a reason; the live state of any background agent or workflow, read now rather than from memory ("are the workflows running though or not???"); and, when work continues in another session, /wrap step 8's continuation prompt. Nothing is "on main" until ~/.claude/commands/bin/repo-hygiene.sh --landed says so.