Battle-test the current change

Decide what can be run, run it, score how well the change is actually covered, and hand back the checks a human still has to do. $ARGUMENTS narrows the side (frontend/backend) and the git scope; blank means both sides, all pending changes.

Phase 1 - Triage (before running anything)

1. Resolve the scope: staged -> git diff --cached, unstaged -> git diff, last-commit -> git diff HEAD~1 (empty tree if there is no parent), blank -> both diffs. Add untracked files (git ls-files --others --exclude-standard): a brand-new module is invisible to git diff and is exactly the code least likely to be covered. Not a git repo (git rev-parse --git-dir fails) -> fall back to the files the user names, and say the scope came from them, not from git.
2. Empty-diff guard: no changes in scope -> say so and stop. Do not run a suite for nothing.
3. Detect what the repo actually has (package.json scripts, composer.json, Makefile, pytest/cargo/go, a dev script, a compose file, an emulator), then sort every check that matters for this diff into three buckets and print them before executing anything:
   • AUTO - runnable now, non-destructive, no credentials: test/typecheck/lint/build scripts, and single-file language checks (php -l through the Docker wrapper, tsc --noEmit, node --check).
   • BOOTABLE - runnable after starting something the repo owns: dev server, Docker stack, emulator. Start it, use it, stop only what this run started.
   • MANUAL - impossible here: prod or client credentials, a real payment, a real device, a third-party dashboard, an email or message actually arriving, a visual judgment call, or a production DB change.

Bucket by what the diff touches, not by what the repo owns. A backend-only diff earns no browser pass.

4. Blast-radius preflight, before the first command runs. A check that can reach real data or a real person is NOT auto, whatever the script is named. The default is refuse, not try:
   • Read the env the command will actually use (.env*, firebase.json / .firebaserc, the compose file, the test config) before running it. Pointed at a production project, a live API key, a shared staging DB, a real SMTP host or a third-party webhook -> MANUAL. Prefer the repo's emulator or test env; if the only env available is the real one, do not run it.
   • Anything that reaches a human (email, SMS, WhatsApp, push, a call to a partner API that bills or writes) is MANUAL even in "test mode", unless the repo pins a sandbox you verified.
   • Anything that mutates a store: seed, reset, migrate, truncate, --force, fixtures that drop collections. MANUAL unless it provably targets a local throwaway you started.
   • Rate-limited or paid third parties (Apollo, LinkedIn, HubSpot, OpenAI, a payment provider): MANUAL. Burning a client's quota to raise a score is not a trade worth making. When you cannot tell which env a command uses, it is MANUAL. State the reason in one line and move on; an unrun check costs points, and that is the correct outcome.

Phase 2 - Backend

Run AUTO first, then BOOTABLE. Keep the tail of a passing command, the full output of a failing one. PHP goes through the Docker wrapper; if Docker is down, that check moves to MANUAL instead of being skipped silently.

If the diff adds non-trivial logic that no suite covers, write ONE minimal runnable assertion for it (the smallest thing that fails if the logic breaks), run it, and say you added it.

Phase 3 - Frontend (only when the diff touches UI)

Delegate to /frontend-verify; do not reimplement browser driving. Cover its two gaps first:

• Preflight: playwright-cli --help. Missing -> the browser pass goes to MANUAL with the install command, the run continues.
• Which backend the page will talk to, checked before booting anything. A dev server is a browser pointed at whatever its env says. If that env is the live project (a real Firebase project id with no emulator flag, a prod API base URL, real keys), the browser pass is MANUAL: loading a route reads, writes and authenticates against real user data. Boot it only against the repo's emulator or sandbox env, and only by setting the flag the repo already provides (for example VITE_USE_EMULATORS=true with firebase emulators:start), never by editing .env. Cannot tell which backend it hits -> MANUAL.
• Dev server: /frontend-verify never starts one. Probe the base URL; if nothing answers, start the repo's dev script in the background and poll, never sleep-and-hope: timeout 30 bash -c 'until curl -sf http://localhost:3000 >/dev/null; do sleep 1; done'. Stop it by killing the port listener only: lsof -ti:3000 -sTCP:LISTEN | xargs kill. Never pkill -f node|vite|next, it can kill this session. A server that was already up stays up.

Phase 4 - Score out of 100

Start at 0, add only what was observed:

Green-but-shallow scores low: a passing typecheck with the new function never called is <= 30. Never round up to be encouraging.

Phase 5 - Report

Ran: 4 checks (3 green, 1 red)
  OK  pnpm typecheck, pnpm lint, vitest (28 passed)
  RED pnpm test - cart.spec.ts:41 expects the old status value
Battle-tested: 45/100
  The new webhook branch never ran; only the happy path is asserted.
Manual (2):
  1. Pay with a real card in Stripe test mode, confirm the order flips to paid.
  2. Apply the migration by hand before this ships.

Red is stated as red with its output, never softened. No MANUAL items -> Manual: none. A check that was skipped goes to MANUAL: silence reads as "checked and fine", which is a lie. Mention the disposable .frontend-verify/ directory if a browser pass wrote one.

Do NOT

The hard rule: this skill must never change real data or reach a real user. A check that can only run against the real thing does not run. It goes to MANUAL, it costs points, and that is the right answer. A lower score is always cheaper than a client's data.

• Never run anything that writes to a production database or a live client environment. No SQL against a production DB over SSH, ever; DB changes go to MANUAL to be applied by hand. A *deploy*.sh wrapper goes to MANUAL too, whatever it claims to test.
• Never point a check at prod to make it runnable. Do not edit .env, swap a config, or export a credential to get a suite to pass. Run it in the repo's sandbox or not at all.
• Never commit, push, or deploy. This skill reports.
• Never leave a process running that this run started.
• Never claim a check passed that was not executed.
• No new test framework, config file, CI wiring, or fixtures. Use what the repo already has.