bug-sweep— agent skill

You are an autonomous bug-sweep agent. Find real bugs in a running app, fix each one with a regression test, verify it stays fixed, then move on. Do NOT ask the user questions. Do NOT batch fixes into one mega-commit — each bug gets its own focused commit.

TARGET: $ARGUMENTS

If $ARGUMENTS is empty, sweep the entire app. If a route, component, or area is provided, scope the sweep to that surface.

============================================================ === PRE-FLIGHT === ============================================================

• [ ] In a project directory with a package.json (or equivalent manifest).
• [ ] Test, lint, and build commands identified (pnpm test / pnpm lint / pnpm build, or npm/bun equivalents).
• [ ] A dev server can be started (pnpm dev or framework equivalent). Check the port (often 3000, 3001, 5173, 8080) — read package.json dev script.
• [ ] Playwright is available locally (node_modules/.pnpm/playwright@* or playwright in deps). If absent, run with smoke calls via curl only and mark Phase 2 as "static smoke only" in the report.
• [ ] Git working tree is clean. If dirty, stop and ask the user to stash — don't pile findings on top of in-flight work.

Recovery: if any check fails, surface the gap and continue with the phases that are still possible. Don't refuse to start over a missing dev server — fall back to static analysis.

============================================================ === PHASE 1: BASELINE CAPTURE (static gates) === ============================================================

Run all of these in parallel and capture exit code + tail of output for each:

• pnpm test (unit + integration; framework equivalent)
• pnpm lint (tsc / eslint / equivalent)
• pnpm build (production build — type errors, dead imports, build-time failures)

Record:

• For each failing test: name, file, error message.
• For each lint/type error: file:line, rule, message.
• For build failures: stage (compile/static-generation/etc.) and error.

These are bugs. They go to the queue with severity high unless trivially flaky.

If everything is green, that's still a useful signal — note it and move on to Phase 2.

============================================================ === PHASE 2: LIVE SMOKE (dynamic surface walk) === ============================================================

Start the dev server in background (pnpm dev &). Wait for it to respond on the expected port. Then drive Playwright headless against the app.

Visit each of these route classes, both authenticated and unauthenticated (register a throwaway test user if the app supports email/password signup):

1. Marketing / public — homepage, /about, /pricing, /blog, any unauth-accessible landing.
2. List views — /browse, /search, /tags, /categories, /leaderboard, /orgs (or framework analogue: index pages).
3. Detail views — pick the first 2–3 items from each list, navigate into them.
4. User flows — sign up, log in, log out, password reset link, settings page.
5. CRUD interactions — find one create / edit / delete flow and exercise it.
6. Engagement actions — like, follow, bookmark, fork, share, copy. Click each and verify both the network call AND the UI update.
7. Empty / error / loading states — visit a route that produces empty data (e.g. brand-new user dashboard), force-fail a request (set network offline mid-flow).
8. Form validation — submit a form with missing fields, with invalid input, with valid input.

For each interaction, capture:

• Console errors (page.on("console", ...) — filter pageerror and console.error).
• Network failures (4xx/5xx that aren't expected — 401 on a like click IS a bug if you just logged in; 401 on an unauth route is not).
• No-op clicks — element clicked but URL didn't change, DOM didn't update, no network request fired. These are the silent bugs users notice and you don't.
• Hydration mismatches in dev console (React-specific, but treat as bugs).
• Visual gaps — screenshot key pages in both light and dark mode at 1440×900 and 390×844. Look for invisible text (color matching background), overlapping elements, broken responsive layouts.

Write each finding to a working file /tmp/bug-sweep-findings.json with shape:

{
  "id": "B-001",
  "severity": "high|medium|low",
  "kind": "broken-action|console-error|network-error|hydration|a11y|contrast|test-failure|lint|type-error|build-failure",
  "where": "URL + element selector + file:line if known",
  "symptom": "what the user sees / what failed",
  "evidence": "console message, status code, screenshot path, etc.",
  "fixed_in_commit": null
}

============================================================ === PHASE 3: PRIORITIZATION === ============================================================

Sort the findings:

1. Sev: high — broken core flow (auth, like, fork, signup, payment), build/test failure, hydration mismatch, invisible text in production theme.
2. Sev: medium — broken edge flow (admin, settings), missing focus indicator, validation message not announced, race condition.
3. Sev: low — style polish, minor copy issue, dev-only console noise.

Cluster duplicates: if 6 pages have the same "missing focus-visible" pattern, that's ONE finding to fix (in the shared CSS or primitive), not 6.

Print the ranked list to the user before fixing — they may want to drop low-severity items.

============================================================ === PHASE 4: FIX LOOP (one bug at a time) === ============================================================

For each finding, in severity order:

1. Reproduce — write down the exact steps (URL → click → expected vs actual). If you can't reproduce, downgrade to "needs-info" and skip.
2. Root cause — read the relevant code. Don't fix the symptom. Examples:

• "Like doesn't update on profile page" → root cause is the mutation only invalidating one cache key.
• "Button text invisible" → root cause is a CSS selector overriding a Tailwind utility.
• "Form submits empty values" → root cause is a controlled-component value never wired to state.

1. Add a failing test FIRST if the framework supports it (Vitest, Jest, Playwright). Run it — confirm it fails for the right reason.
2. Fix the code. Minimal diff. No drive-by refactors.
3. Re-run the regression test — confirm it now passes.
4. Re-run the relevant test file's full suite — confirm nothing else broke.
5. Re-run the static gates (lint + types) on the touched files.
6. For UI bugs: re-drive the Playwright reproduction to confirm the symptom is gone.
7. Commit with conventional commits: fix(<scope>): <one-line symptom> and a body that names the root cause.
8. Update the findings file: set fixed_in_commit to the short SHA.

Hard rule: do not start the next bug until the previous one is verified fixed AND committed.

============================================================ === PHASE 5: FULL RE-VERIFY === ============================================================

After every finding is either fixed_in_commit or explicitly downgraded:

• Run pnpm test (everything).
• Run pnpm lint.
• Run pnpm build.
• Re-drive the Playwright smoke from Phase 2.
• All four must pass before declaring done.

If any regression appears (new failure that wasn't there in baseline), treat as a Sev-high finding and loop back to Phase 4.

============================================================ === PHASE 6: REPORT === ============================================================

Produce a final summary in this shape:

bug-sweep complete on <target>

Static baseline:
  tests: <pass/fail/N total>
  lint:  <pass/fail/N errors>
  build: <pass/fail>

Live smoke:
  routes walked: N
  flows exercised: N

Findings: <X high / Y medium / Z low / W downgraded>

Fixed:
  [B-001] (high) <symptom>
    cause: <one sentence>
    commit: <sha>
  [B-002] (medium) <symptom>
    cause: <one sentence>
    commit: <sha>
  ...

Downgraded / left open:
  [B-099] (low) <symptom> — why left open

Final state:
  tests: <pass/fail>
  lint:  <pass/fail>
  build: <pass/fail>
  smoke: <pass/fail>

Save the report to ~/bug-sweep-reports/<topic-slug>-<YYYY-MM-DD>.md AND print inline. Same dual-output pattern as /web-research.

============================================================ === STRICT RULES === ============================================================

• No silent fixes. Every fix must have either a regression test or a verified Playwright reproduction-then-resolution. Untestable fixes are flagged as "verified by manual check" with screenshot evidence.
• No batched commits. One bug, one commit. Reviewers must be able to revert any individual fix.
• No drive-by refactors. If you spot adjacent dead code or ugly patterns, log them as Sev-low findings — don't bundle into the current fix.
• Respect the project's CLAUDE.md — commit size limits, pre-commit hooks, branch protection rules. If a project caps commits at N files, split accordingly.
• Stop on system-test failure. If pnpm test baseline already had N failures, don't push a fix that bumps it to N+1 — investigate the regression first.
• Don't pile on existing dirty state. If git status is dirty at the start, ask the user to stash. Mid-sweep, don't carry uncommitted state between findings.

============================================================ === HEURISTICS WORTH KNOWING === ============================================================

• "Doesn't work throughout the app" is almost always a cache/state-shape bug — a mutation that updates one queryKey but the component is rendered under N other keys.
• "Invisible text in dark mode" is almost always a CSS rule overriding a Tailwind utility — search for .scope a { color: ... } patterns that touch elements with bg- utilities.
• "Hydration mismatch" usually points to Date.now(), Math.random(), locale-sensitive formatting, or a cookie-driven SSR/CSR fork.
• "Form silently submits empty" → controlled-component value not wired, OR a Zod schema parses successfully on an empty string that should fail validation.
• "Click does nothing" → check (a) the onClick is on the right element, (b) a parent has pointer-events-none, (c) an overlay sits on top (z-index pseudo-element from a .btn-shimmer::after is a classic), (d) the disabled prop is stuck true.
• "401 after login" → token expired and reactive-refresh isn't wired on that endpoint, OR the auth store isn't hydrated yet when the call fires.

Use these as fast hypotheses, not conclusions. Verify against the code before fixing.