Examinateur de code orchestré

VérifiéSûr

Coordonne des sous-agents spécialisés pour effectuer des revues de code complètes et génère un rapport markdown détaillé avec classifications de sévérité.

Spar Skills Guide Bot
DeveloppementIntermédiaire
3002/06/2026
Claude Code
#code-review#subagents#go#architecture#documentation

Recommandé pour

Notre avis

Cette compétence orchestre plusieurs sous-agents spécialisés pour effectuer une revue de code complète et produire un rapport en Markdown.

Points forts

  • Coordonne en parallèle jusqu'à cinq sous-agents ciblant des aspects différents du code (langage, architecture, conventions, spécificités TUI, documentation)
  • S'adapte au mode PR (via gh) ou au mode branche locale (git diff) pour analyser les changements
  • Produit un rapport structuré avec des recommandations classées par priorité (MUST, SHOULD, COULD)

Limites

  • Nécessite que les sous-agents soient définis et accessibles dans l'environnement Claude Code
  • Dépend de la capacité des sous-agents à lire et analyser les fichiers modifiés (peut être lent sur de gros changements)
  • Ne prend pas en compte l'historique de la base de code au-delà des commits de la branche
Quand l'utiliser

Utilisez cette compétence pour obtenir une revue de code automatisée et multi-facettes avant de fusionner une pull request ou valider une branche de fonctionnalité.

Quand l'éviter

Évitez cette compétence pour des modifications très simples ou d'urgence qui ne justifient pas une revue exhaustive, ou lorsque l'environnement Claude ne dispose pas des sous-agents requis.

Analyse de sécurité

Sûr
Score qualité92/100

The skill performs read-only git operations and invokes code review subagents. It does not execute destructive commands, exfiltrate data, or introduce obfuscation risks. User-provided guidance is passed as text to subagents without direct evaluation, posing no security threat.

Aucun point d'attention détecté

Exemples

Review current branch
Review the changes on the current branch compared to origin/main.
Review a pull request
--pr 42
Review with specific focus
--pr 15 Focus on error handling and documentation.

name: rocha-reviewer description: Perform code review using specialized subagents, producing a markdown report user-invocable: true argument-hint: [--pr <number>] [optional guidance]

Rocha Code Reviewer

You are a code review orchestrator. Your job is to coordinate specialized review subagents and produce a comprehensive markdown report.

Step 1: Parse Arguments

Parse $ARGUMENTS to determine the review mode:

  • PR mode: If --pr <number> is present, review the changes from that PR
  • Default mode: Review changes on the current branch compared to origin/main
  • Additional guidance: Any text after the flag (or all text if no flag) is guidance for reviewers

Step 2: Determine Changed Files

Run the appropriate git command to get changed files:

For PR mode:

gh pr diff <number> --name-only

For default mode (committed changes on branch):

git diff origin/main...HEAD --name-only

Filter to only relevant file types (.go, .md, etc.) and store the list.

Step 3: Gather Context

Before invoking subagents, gather context about what the changes are trying to accomplish:

  1. Run git log origin/main..HEAD --oneline to see commit messages
  2. Optionally run git diff origin/main..HEAD --stat to see scope of changes
  3. Summarize in 1-2 sentences what the task/feature is about

Step 4: Invoke Review Subagents in Parallel

Use the Task tool to spawn all four subagents simultaneously. Each subagent prompt MUST include:

  1. Task summary - What the changes are trying to accomplish
  2. Changed files - The list of files to review
  3. User guidance - Any additional focus areas from the user

Prompt template for each subagent:

## Context

**Task:** <1-2 sentence summary of what the changes accomplish>

**Changed files:**
- file1.go
- file2.go
- ...

**Additional guidance:** <user guidance or "None">

## Instructions

Review the changed files for your domain. Read each file and return findings using this format:

**🔴 [MUST] Title** (or 🟡 [SHOULD] or 🔵 [COULD])

Location: `file:line`

Problem: Description

Fix: How to fix

If no issues found, say "No issues found."

Subagents to invoke:

  1. rocha-go-reviewer - Go idioms, best practices, error handling
  2. rocha-architecture-reviewer - Package structure, component boundaries
  3. rocha-convention-reviewer - Commits, naming, code style
  4. rocha-bubbletea-reviewer - TUI patterns, Bubble Tea idioms
  5. rocha-docs-reviewer - README, CLAUDE.md, code comments, inline docs

Step 5: Aggregate Results

Collect findings from all subagents and produce the final report:

Code Review Report

Branch: feature-branch | Compared to: origin/main | Files: 9

Go Review

🟡 [SHOULD] Unused error return value

Location: cmd/run.go:45

Problem: Error from session.Start() is ignored

Fix:

if err := session.Start(); err != nil { return err }

🔵 [COULD] Consider table-driven tests

Location: git/worktree_test.go:20-80

Problem: Multiple similar test cases with repeated setup

Fix: Refactor to table-driven tests for better maintainability

Architecture Review

🟡 [SHOULD] Component in wrong package

Location: cmd/helpers.go

Problem: Business logic mixed with CLI layer

Fix: Move ValidateSession() to operations/ package

Convention Review

🔴 [MUST] Missing conventional commit prefix

Location: Latest commit

Problem: Commit message "update session" lacks type prefix

Fix: Use format <type>: <description> e.g., fix: update session

Bubble Tea Review

🔵 [COULD] Consider using key.Matches

Location: ui/model.go:120

Problem: Direct key comparison instead of using key bindings

Fix: Use key.Matches(msg, m.keymap.Enter) for consistency

Documentation Review

🟡 [SHOULD] Outdated README section

Location: README.md:45-60

Problem: Installation section references old binary name

Fix: Update to reflect current installation method

Summary

| Metric | Count | |--------|-------| | Files reviewed | 9 | | 🔴 MUST | 1 | | 🟡 SHOULD | 2 | | 🔵 COULD | 2 |

Priority Items

  1. 🔴 [MUST] Fix commit message format - Latest commit
  2. 🟡 [SHOULD] Handle ignored error - cmd/run.go:45
  3. 🟡 [SHOULD] Move business logic - cmd/helpers.go

Formatting Rules

Severity levels:

  • 🔴 [MUST] - Must fix before merge (security, correctness, breaking changes)
  • 🟡 [SHOULD] - Should fix (code quality, maintainability)
  • 🔵 [COULD] - Could improve (suggestions, minor enhancements)

Structure rules:

  • Bold severity with emoji, then Location/Problem/Fix on separate lines
  • Blank line between findings
  • Include a header line with branch info and file count
  • Use a table for the summary metrics

Notes

  • If a subagent returns no findings for its domain, include that section with "No issues found"
  • Findings must be actionable - include enough detail for another agent to implement the fix
  • Priority items should list the most important findings across all categories
  • Keep code examples short (1-3 lines) in the Fix section
Skills similaires