Correction des revues de code

VérifiéPrudence

Automatise la planification et l'exécution des corrections issues des revues de code. Génère des artefacts datés contenant le plan et le journal d'exécution avec commits atomiques.

Spar Skills Guide Bot
DeveloppementAvancé
2002/06/2026
Claude Code
#code-review#automated-fixes#planning#execution#multi-agent

Recommandé pour

Notre avis

Lit le dernier rapport de révision, planifie des correctifs pour chaque constatation et les exécute automatiquement, produisant un plan et un journal d'exécution.

Points forts

  • Automatise entièrement le processus de révision et de correction.
  • Divise le travail entre plusieurs agents parallèles pour gagner du temps.
  • Respecte les dépendances entre correctifs grâce à des vagues d'exécution.
  • Produit des artefacts clairs (plan et journal) pour traçabilité.

Limites

  • Nécessite un fichier de révision préexistant dans le dossier .reviews/.
  • Peut ne pas gérer les conflits entre changements simultanés.
  • Ne crée pas de nouveaux fichiers ni de tests, sauf indication explicite.
Quand l'utiliser

Utilisez cette compétence après avoir exécuté une révision de code, pour appliquer rapidement tous les correctifs planifiés et hiérarchisés.

Quand l'éviter

Ne l'utilisez pas si vous souhaitez examiner chaque correctif manuellement avant exécution, car elle les applique sans approbation intermédiaire.

Analyse de sécurité

Prudence
Score qualité90/100

The skill automates planning and executing code fixes from a review, which involves modifying source code and committing changes without waiting for user confirmation. While useful, it could cause unintended damage if the review findings or the AI's interpretation are flawed. However, it operates within the codebase and uses standard linting tools, not destructive commands. The risk is moderate.

Points d'attention
  • Automates code modification and git commits without human approval after planning phase
  • Uses parallel subagents that could introduce conflicting changes if not properly coordinated
  • Runs linting and file-modification commands automatically

Exemples

Fix all review findings
Review the latest code review and fix all findings.
Execute review fix plan
Run review-fix to plan and execute fixes from the most recent review.
Apply prioritized fixes
Automatically fix all Fix and Improve items from the latest review, grouped by wave.

name: review-fix description: Read the latest code review, plan fixes for all findings, then execute the fixes — all in one command. Produces a dated plan and execution log in .reviews/.

Review Fix — Plan & Execute from Last Review

Read the most recent .reviews/YYYY-MM-DD-review.md, plan fixes for all findings grouped by priority, then execute every planned fix with atomic commits. Produces two artifacts: a review-plan and a review-execution log.

Instructions

Phase 0 — Find the Latest Review

  1. Glob for .reviews/*-review.md and pick the most recent file by date.
  2. Read the full review file. Parse every finding into a list of { severity, reviewer, file, line, description } entries.
  3. If no review file exists, tell the user to run /code-review first and stop.

Phase 1 — Plan

Spawn 4 parallel planning agents (Task tool, subagent_type: Plan) that each receive the full findings list filtered to their scope. Each agent returns a structured list of fix actions.

Agent scopes

| Agent | Scope | |---------------|--------------------------------------------------------| | backend | Findings from Tina, Marcus, Priya, Carlos, Oscar that touch cachibot/ | | frontend | Findings from Tina, Zoe, Luna that touch frontend/src/ | | data | Findings from Carlos (storage/db layer) | | tests | Findings from Derek (test coverage gaps) |

Findings that span multiple scopes go to every relevant agent.

Each planning agent must:

  • Group related findings that can be fixed together (same file, same pattern).
  • For each group, produce:
    • What: one-line summary of the change
    • Files: list of files to modify
    • How: 2-3 sentence implementation approach
    • Severity: highest severity in the group (Fix > Improve > Note)
    • Risk: Low / Medium / High (does this change behavior or just style?)
    • Depends on: other group IDs this group must wait for (if any)
  • Order groups by: Fix severity first, then by dependency order.
  • Skip any finding rated Note — notes are informational only.

After all 4 agents return, compile results into .reviews/YYYY-MM-DD-review-plan.md:

# Review Fix Plan — YYYY-MM-DD

**Source review:** `.reviews/YYYY-MM-DD-review.md`
**Generated:** YYYY-MM-DD HH:MM

## Stats

| Severity | Findings | Planned groups |
|----------|----------|----------------|
| Fix      | N        | N              |
| Improve  | N        | N              |
| Note     | N        | (skipped)      |

## Execution Waves

Wave 1 (no dependencies):
- [B1] Backend: <summary>
- [F1] Frontend: <summary>

Wave 2 (depends on wave 1):
- [B2] Backend: <summary> (depends on B1)

...

## Detailed Plan

### [B1] <Summary>
- **Severity:** Fix
- **Risk:** Low
- **Files:** `path/to/file.py`
- **How:** Description of the fix approach
- **Findings addressed:**
  - Tina: `path/to/file.py:42` — original finding text
  - Marcus: `path/to/file.py:50` — original finding text

Print the plan summary (wave count, group count, severity breakdown) to the user, then immediately proceed to Phase 2 without waiting for approval.

Phase 2 — Execute

Read the plan file. Execute fixes wave by wave — all groups in a wave run in parallel, but waves run sequentially (wave 2 waits for wave 1 to finish).

For each group, spawn a Task agent (subagent_type: general-purpose) with:

  • The group's detailed plan (What, Files, How)
  • The original finding text for context
  • Instruction to read each file before editing
  • Instruction to make the minimal change that addresses the finding
  • Instruction to run ruff check --fix and ruff format on any modified Python file
  • Instruction to run npm run lint -- --fix on any modified TypeScript file (from frontend/)
  • Instruction to NOT create new files unless the plan explicitly says to
  • Instruction to NOT add tests (Derek findings are tracked but test writing is a separate task)

After each wave completes:

  1. Verify no lint errors remain (ruff check cachibot/ and cd frontend && npm run lint).
  2. If lint fails, spawn a fix agent for the failing files before continuing.
  3. Stage all changed files and create one commit per wave:
    Review fixes wave N: <comma-separated group IDs>
    
    Addresses findings from YYYY-MM-DD code review.
    Groups: [B1] summary, [F2] summary, ...
    
    Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
    

After all waves complete, write .reviews/YYYY-MM-DD-review-execution.md:

# Review Fix Execution — YYYY-MM-DD

**Plan:** `.reviews/YYYY-MM-DD-review-plan.md`
**Source review:** `.reviews/YYYY-MM-DD-review.md`

## Results

| Wave | Groups | Status  | Commit  |
|------|--------|---------|---------|
| 1    | B1, F1 | Done    | abc1234 |
| 2    | B2     | Done    | def5678 |

## Findings Addressed

| Severity | Planned | Fixed | Skipped | Reason          |
|----------|---------|-------|---------|-----------------|
| Fix      | N       | N     | N       |                 |
| Improve  | N       | N     | N       |                 |

## Group Details

### [B1] <Summary> — Done
- **Commit:** abc1234
- **Files changed:** `path/to/file.py`
- **Findings fixed:**
  - Tina: `path/to/file.py:42` — fixed
  - Marcus: `path/to/file.py:50` — fixed

### [F1] <Summary> — Skipped
- **Reason:** File was recently refactored, finding no longer applies
- **Findings skipped:**
  - Zoe: `frontend/src/Component.tsx:100` — N/A after refactor

## Remaining Work

- Derek's test coverage findings were not addressed (test writing is separate)
- N findings skipped due to: ...

Phase 3 — Summary

Print a final summary to the user:

Review fix complete.

Waves executed: N
Commits created: N
Findings fixed: N / M planned (X Fix, Y Improve)
Skipped: N (reasons listed in execution log)

Plan:      .reviews/YYYY-MM-DD-review-plan.md
Execution: .reviews/YYYY-MM-DD-review-execution.md

Important Rules

  • Never skip a Fix-severity finding without logging a clear reason.
  • Read before edit — every agent must read the target file before modifying it.
  • Minimal changes — fix exactly what the finding describes, nothing more.
  • No new features — this is a fix pass, not a feature pass.
  • No test writing — Derek's findings are logged as "remaining work" but not executed here. Test creation deserves its own focused session.
  • Preserve behavior — if a fix would change external API behavior, mark it as Risk: High in the plan and add a note in the execution log.
  • If the review file has a "Previous Reviews" section noting still-open findings from prior reviews, include those in the plan too.
Skills similaires