メニュー
Fix a bug in this repository the safe way — reproduce first, failing regression test, root cause via the maps, surgical fix, review, ledger entry. Use whenever the user reports broken, wrong, or crashing behavior.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | fix-bug |
| description | Fix a bug in this repository the safe way — reproduce first, failing regression test, root cause via the maps, surgical fix, review, ledger entry. Use whenever the user reports broken, wrong, or crashing behavior. |
The contract: no fix before a reproduction, no reproduction without a failing test, no "fixed" without the regression test green, no closing without the ledger row and the gotcha recorded.
Draft ai/lab/specs/BUGFIX_<name>.md from BUGFIX_TEMPLATE.md: symptom
(observed vs expected), exact reproduction steps. Get the user's OK.
Turn the reproduction into a regression test and watch it fail. A bug you cannot make a test fail for is a bug you do not understand yet. The test is permanent — it outlives the fix.
ai/guide/MODULE_MAP.md → which modules; note Stability of every suspect.ai/guide/FEATURE_MAP.md gotchas + ai/lab/WORKLOG.md → was this area
recently touched, or is the "bug" deliberate behavior a past ADR explains?Root cause in a file with Stability frozen or ? ⇒ stop and ask the human.
Record their approval in the bugfix doc before proceeding.
Name the root cause in the bugfix doc — the defect, not the symptom. Then make the smallest diff that turns the regression test green. Match the license headers of neighboring files. No drive-by refactors, no layout changes, no dependency additions while "in there".
Run the regression test (now green), then the suites covering the touched area. Report command run · exit status · failures verbatim. Red or unrun ⇒ not fixed.
Request /review-change on the diff (fresh context — not this session).
Append a bugfix row to ai/lab/WORKLOG.md linking the bugfix doc and review;
add the gotcha line to FEATURE_MAP so the next agent knows the trap. All
[inferred] — tell the user which tags await their [verified] flip.
Bootstrap the ai/ knowledge layer for a repo that has none — draft the MODULE_MAP, diagrams, and guide docs as [inferred] for a human to audit, touching no source code. Use when the maps are still placeholders or the user asks to cold-start, bootstrap, or onboard a repository.
Bootstrap the ai/ knowledge layer for a repo that has none — draft the MODULE_MAP, diagrams, and guide docs as [inferred] for a human to audit, touching no source code. Use when the maps are still placeholders or the user asks to cold-start, bootstrap, or onboard a repository.
Deep, judgement-based adversarial code audit — hunts for stale cross-references, unescaped interpolation, platform gaps, generated-file ownership conflicts, and cross-module consistency rot that no mechanical check can catch. Read-only; writes a dated findings report.
Kit-maintainer diagnostic (LOCAL-ONLY, not distributed) that audits the README and docs/ for staleness against the repo's real command/skill/CLI roster, stamped tree, and cross-doc consistency. Read-only; produces a severity-ranked findings report.
Run verification and drift checks on the AI knowledge-base to identify missing documentation or stale references.
Deep-mine the source code to discover implemented features; writes ai/analysis/FEATURE_CATALOG.md.