Finn763/pit-stop

pit-stop 是做什麼的？

F1 pit-stop for codebases: drive in, get fixed, come out faster. One instruction runs read → find → fix → verify → report with no mid-run questions. The user only says go; you return once with a report.

The run (in order, no skipping)

1. Load: read README/AGENTS.md (or equivalent) + git status + recent-commit hot spots. Write one line first: MODE = what counts as a finding here, what is rejected outright. Missing a key fact? Ask ONCE, then proceed under strictest defaults.
2. Find: scope before you scan (named target > git hot spots > full tree). Every finding needs path:line evidence. Skip recorded decisions unless friction justifies reopening. Tag each: delete/stdlib/native/yagni/shrink/security/obs, strength Strong/Worth/Speculative. Speculative = report only, never fix. Nothing found: say Lean already. Ship. — never pad.
3. Propose: Strong items only. Each = symptom + evidence + impact + minimal fix + cost. Plus a NOT-doing list with its trigger conditions.
4. Fix: review→fix loop with an independent reviewer (different angle beats same eyes); stop at clean or --max-rounds (default 3). Cross-round ledger, no repeat fixes. Two stagnant rounds → escalate to "needs human", stop. See references/guardrails.md.
5. Report: header Generated fully automatically by pit-stop. Four blocks — changed / verified (tool output) / unverified / remaining. No claim without fresh tool output. Secret/token/PII redaction in everything pasted. See templates/report.md.

Iron laws

• Evidence before claims, always. No verification run in this turn = no success claim.
• A subagent's "done" is not done. Read its diff before reporting.
• Fuzzy spots downgrade, never guess. Write "needs human", keep moving.
• Push/commit never automatic. Changes wait in the workdir for one explicit word.
• Single spend fuse: one phase burning past $20 → stop phase, write report.

Details: references/phases.md, references/guardrails.md, references/verification.md.