AGENT SKILL

bugfix

Use when a bug is reported, something is broken, an error or crash needs fixing, a wrong output appears, or a regression shows up — guides Grimoire's Reproduce → Root-Cause → failing-test-FIRST → Fix → Verify discipline and will not let the fix begin until a failing regression test is recorded. Use also when the user says "bugfix", "/bugfix", "fix this bug", or "something is broken".

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 15
Forks
1

01 前言 Front Matter

name
bugfix
triggers
[ "fix a bug", "bug fix", "something is broken", "regression test first", "reproduce the bug" ]
description
Use when a bug is reported, something is broken, an error or crash needs fixing, a wrong output appears, or a regression shows up — guides Grimoire's Reproduce → Root-Cause → failing-test-FIRST → Fix → Verify discipline and will not let the fix begin until a failing regression test is recorded. Use also when the user says "bugfix", "/bugfix", "fix this bug", or "something is broken".
argument-hint
[bug report / repro steps / issue number]
user-invocable
true

03 SKILL.md 正文

/bugfix — Guided Bug-Fix Workflow

Guard rail for the #1 bug-fix failure: fixing before a failing test exists. This skill walks workflow-bugfix.md (the single source of truth) and adds one hard, non-skippable gate — Phase 3 must produce a test that fails on the current code before any fix is written.

Why this exists: the rule alone was repeatedly skipped. This skill makes the order explicit and the gate blocking. Read workflow-bugfix.md for the full rationale; this file is the executable checklist.

The non-negotiable sequence

Reproduce → Root-Cause Analysis → FAILING TEST → Fix → Verify → Review → Commit

Each phase finishes before the next starts. The only way past Phase 3 is a test that you have run and watched fail for the right reason.

Phase 1 — Reproduce (no guessing)

  • State the exact wrong behavior: command, input, observed vs expected.
  • Find the surface: which entry point(s)? CLI command, TUI path, library seam? A bug often lives on more than one surface (e.g. a CLI path AND its TUI twin); list every one that shares the broken code.
  • Confirm it actually reproduces. If you can't reproduce it, keep digging — no speculative fixes.

Write a one-line repro: "<cmd/inputs> → <observed>; expected <expected>".

Phase 2 — Root-Cause Analysis (write it down)

  • Trace the symptom to the line and the condition that made it fire.

  • Output a root-cause statement in this exact shape — no "error on line N":

    X happens because Y, introduced by/located at Z.

  • Single bug or a pattern? Grep for the same defect elsewhere (sibling call sites, the other surface from Phase 1). If it's a pattern, the fix covers all instances.

  • If the real cause needs an architectural change, stop and escalate to the feature workflow with a plan artifact — don't paper over it.

Phase 3 — Failing test FIRST 🚧 GATE

Do not edit any production code until this phase is done.

  1. Write the test that exercises the Phase-1 repro and targets the Phase-2 root cause (not just the symptom).
    • Acceptance-level → test/tests/test_*.py (real binary + registry).
    • Unit-level → inline #[cfg(test)] in the affected module.
  2. Run it. Paste the red output. It must fail, and fail for the bug's reason (assertion on the wrong behavior) — not a compile error, typo, or missing fixture.
  3. If an existing test encodes the buggy behavior (it asserts the wrong thing the user is now reporting), that test is part of the bug — note it; it will be corrected in the fix, with the reason recorded.

Gate to pass Phase 3 — all three true:

  • [ ] A test exists that exercises the repro.
  • [ ] You ran it and it failed (output shown).
  • [ ] The failure is the bug, not test scaffolding.

If you cannot make a test fail, you have not reproduced or understood the bug — return to Phase 1/2. Skipping this gate is the failure mode this skill prevents.

Phase 4 — Fix (minimal, root-cause)

  • Smallest change that addresses the Phase-2 cause. No drive-by refactor, no "while I'm here" cleanup (separate commit, separate type).
  • Fix every instance the Phase-2 pattern search found.
  • Correct any existing test that encoded the wrong behavior — state why in the diff/commit (user intent / corrected contract supersedes the old assumption).

Phase 5 — Verify (evidence, not vibes)

  • The Phase-3 test now passes (show it).
  • Run the subsystem gate for the changed area (e.g. task rust:verify), then cache-bust + re-run unit tests if Rust (touch the edited files so a stale cache can't mask a failure).
  • Manually confirm the Phase-1 repro no longer reproduces.
  • Final gate before commit: task --force verify (fmt + clippy + build + unit + acceptance). Verify it yourself — do not trust "should pass".

Phase 6 — Review-Fix Loop (+ optional Codex)

Apply the canonical Review-Fix Loop from workflow-bugfix.md (correctness vs root cause, regression risk to other callers, minimality, test adequacy). For a non-trivial fix, run one Codex adversarial pass over the diff as the cross-model gate; fix real findings, re-verify, then converge.

Phase 7 — Commit

  • fix: conventional commit; body names the root cause and the regression test.
  • Reference the issue (fixes #N) when one exists.
  • Never push — the human decides (see workflow-git.md).

Anti-rationalizations (stop if you think these)

Thought Reality Do this
"I know the cause, I'll just fix it." No Phase-2 statement written. Write the root-cause line first.
"The test is trivial, I'll add it after." A test added after the fix never proved it caught the bug. Write it first, watch it fail.
"Existing tests cover this area." They may encode the buggy behavior. Add a test for the exact broken path.
"Clippy/format nit nearby, I'll fix too." Scope creep. Separate chore:/refactor: commit.

References

04 同仓库更多 Skills