review-pull-request
À propos
Cette Compétence Claude effectue une revue automatisée et complète d'une pull request GitHub en utilisant la CLI GH. Elle analyse les différences et l'historique des commits, vérifie les contrôles CI/CD, et soumet des retours structurés avec des niveaux de sévérité tels que 'bloquant' ou 'suggestion'. Utilisez-la lorsqu'une PR vous est assignée pour garantir un examen approfondi avant de solliciter des relecteurs humains ou d'auditer le code fusionné.
Installation rapide
Claude Code
Recommandénpx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/review-pull-requestCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Review Pull Request
Review GH PR end-to-end — understand change → submit structured feedback. Uses gh CLI for all GH interactions + produces severity-leveled review comments.
Use When
- PR ready for review + assigned to you
- Second review after author addresses feedback
- Self-review before req others
- Audit merged PR for post-merge quality
- Want structured review process not ad-hoc scanning
In
- Required: PR id (number, URL,
owner/repo#number) - Optional: Review focus (security, perf, correctness, style)
- Optional: Codebase familiarity (familiar, somewhat, unfamiliar)
- Optional: Time budget (quick scan, std, thorough)
Do
Step 1: Understand Ctx
Read PR description + understand what change accomplishes.
- Fetch PR metadata:
gh pr view <number> --json title,body,author,baseRefName,headRefName,labels,additions,deletions,changedFiles,reviewDecision - Read title + description:
- What problem does PR solve?
- What approach did author take?
- Specific areas author wants reviewed?
- Check PR size + assess time req:
PR Size Guide:
+--------+-----------+---------+-------------------------------------+
| Size | Files | Lines | Review Approach |
+--------+-----------+---------+-------------------------------------+
| Small | 1-5 | <100 | Read every line, quick review |
| Medium | 5-15 | 100-500 | Focus on logic changes, skim config |
| Large | 15-30 | 500- | Review by commit, focus on critical |
| | | 1000 | files, flag if should be split |
| XL | 30+ | 1000+ | Flag for splitting. Review only the |
| | | | most critical files. |
+--------+-----------+---------+-------------------------------------+
- Review commit history:
gh pr view <number> --json commits --jq '.commits[].messageHeadline'- Commits logical + well-structured?
- History tells story (each commit coherent step)?
- Check CI/CD status:
gh pr checks <number>- All checks passing?
- If failing, note which → affects review
→ Clear understanding of what PR does, why exists, how big, CI green. Ctx shapes review approach.
If err: PR description empty/unclear → note as first feedback. PR w/o ctx = review antipattern. gh cmds fail → verify auth (gh auth status) + repo access.
Step 2: Analyze Diff
Read actual code changes systematically.
- Fetch full diff:
gh pr diff <number> - Small/medium PRs: read entire diff sequential
- Large PRs: review by commit:
gh pr diff <number> --patch # full patch format - Each changed file eval:
- Correctness: Code does what PR says?
- Edge cases: Boundary conditions handled?
- Error handling: Caught + handled appropriately?
- Security: Injection, auth, data exposure risks?
- Perf: Obvious O(n^2), missing indexes, mem issues?
- Naming: New vars/fns/classes named clearly?
- Tests: New behaviors covered by tests?
- Take notes as read, classifying each by severity
→ Set of obs covering correctness, security, perf, quality for every meaningful change. Each obs has severity.
If err: diff too large to review effectively → flag: "This PR changes {N} files and {M} lines. I recommend splitting it into smaller PRs for more effective review." Still review highest-risk files.
Step 3: Classify Feedback
Organize obs into severity levels.
- Classify each obs:
Feedback Severity Levels:
+-----------+------+----------------------------------------------------+
| Level | Icon | Description |
+-----------+------+----------------------------------------------------+
| Blocking | [B] | Must fix before merge. Bugs, security issues, |
| | | data loss risks, broken functionality. |
| Suggest | [S] | Should fix, but won't block merge. Better |
| | | approaches, missing edge cases, style issues that |
| | | affect maintainability. |
| Nit | [N] | Optional improvement. Style preferences, minor |
| | | naming suggestions, formatting. |
| Praise | [P] | Good work worth calling out. Clever solutions, |
| | | thorough testing, clean abstractions. |
+-----------+------+----------------------------------------------------+
- Each Blocking explain:
- What's wrong (specific issue)
- Why matters (impact)
- How to fix (concrete suggestion)
- Each Suggest explain alternative + why better
- Keep Nits brief — one sentence enough
- Include ≥1 Praise if anything positive stands out
→ Sorted feedback list w/ clear severity. Blocking has fix suggestions. Ratio: few Blocking, some Suggest, minimal Nit, ≥1 Praise.
If err: everything seems blocking → PR may need rework not patch. Consider req changes at PR level vs line-by-line. Nothing wrong → say so — "LGTM" valid when code good.
Step 4: Write Comments
Compose review w/ structured actionable feedback.
- Write review summary (top-level):
- One sentence: what PR does (confirm understanding)
- Overall: approve, req changes, comment
- Key items: list Blocking (if any) + top Suggest
- Praise: call out good work
- Write inline comments for specific code locations:
# Post inline comments via gh API gh api repos/{owner}/{repo}/pulls/{number}/comments \ -f body="[B] This SQL query is vulnerable to injection. Use parameterized queries instead.\n\n\`\`\`suggestion\ndb.query('SELECT * FROM users WHERE id = $1', [userId])\n\`\`\`" \ -f commit_id="<sha>" \ -f path="src/users.js" \ -F line=42 \ -f side="RIGHT" - Format feedback consistent:
- Start each comment w/ severity tag:
[B],[S],[N],[P] - Use GH suggestion blocks for concrete fixes
- Link to docs for style/pattern suggestions
- Start each comment w/ severity tag:
- Submit review:
# Approve gh pr review <number> --approve --body "Review summary here" # Request changes (when blocking issues exist) gh pr review <number> --request-changes --body "Review summary here" # Comment only (when unsure or providing FYI feedback) gh pr review <number> --comment --body "Review summary here"
→ Submitted review w/ clear actionable feedback. Author knows exactly what to fix (Blocking), consider (Suggest), what went well (Praise).
If err: gh pr review fails → check perms. Need write access or be requested reviewer. Inline comments fail → fall back to all feedback in review body w/ file:line refs.
Step 5: Follow Up
Track resolution.
- After author responds or pushes updates:
gh pr view <number> --json reviewDecision,reviews - Re-review only changes addressing feedback:
gh pr diff <number> # check new commits - Verify Blocking resolved before approving
- Resolve comment threads as issues addressed
- Approve when all Blocking fixed:
gh pr review <number> --approve --body "All blocking issues resolved. LGTM."
→ Blocking verified fixed. Conversation resolved. PR approved or further changes req'd w/ specific remaining items.
If err: author disagrees → discuss in PR thread. Focus on impact (why matters) not authority. Disagreement persists on non-blocking → yield gracefully. Author owns code.
Check
- PR ctx understood (purpose, size, CI status)
- All changed files reviewed (or highest-risk for XL PRs)
- Feedback classified by severity (Blocking/Suggest/Nit/Praise)
- Blocking has specific fix suggestions
- ≥1 Praise for positive aspects
- Review decision matches feedback (approve only if no Blocking)
- Inline comments ref specific lines w/ severity tags
- CI/CD checks verified (green before approval)
- Follow-up done after author revisions
Traps
- Rubber-stamping: Approving w/o reading diff. Every approval = assertion of quality.
- Nit avalanche: Drowning author in style prefs. Save nits for mentoring; skip in time-sensitive reviews.
- Miss forest: Reviewing line-by-line w/o understanding overall design. Read description + commit history first.
- Block on style: Formatting + naming almost never blocking. Reserve Blocking for bugs, security, data integrity.
- No praise: Only pointing problems = demoralizing. Good code deserves recognition.
- Scope creep: Commenting on code not changed in PR. Pre-existing issues → file separate issue.
→
review-software-architecture— system-level architecture review (complementary)security-audit-codebase— deep security analysis for security-sensitive PRscreate-pull-request— other side: creating PRs easy to reviewcommit-changes— clean commit history makes PR review easier
Dépôt GitHub
Compétences associées
llamaguard
AutreLlamaGuard est le modèle de Meta, doté de 7 à 8 milliards de paramètres, conçu pour modérer les entrées et sorties des LLM selon six catégories de sécurité comme la violence et les discours haineux. Il offre une précision de 94 à 95 % et peut être déployé avec vLLM, Hugging Face ou Amazon SageMaker. Utilisez cette compétence pour intégrer facilement le filtrage de contenu et des garde-fous de sécurité dans vos applications d'IA.
cost-optimization
AutreCette compétence de Claude aide les développeurs à optimiser les coûts du cloud grâce au redimensionnement des ressources, aux stratégies d'étiquetage et à l'analyse des dépenses. Elle fournit un cadre pour réduire les dépenses cloud et mettre en œuvre une gouvernance des coûts sur AWS, Azure et GCP. Utilisez-la lorsque vous devez analyser les coûts d'infrastructure, redimensionner les ressources ou respecter des contraintes budgétaires.
quantizing-models-bitsandbytes
AutreCette compétence quantifie les LLMs en précision 8 bits ou 4 bits à l'aide de bitsandbytes, permettant une réduction de 50 à 75 % de la mémoire utilisée avec une perte de précision minime. Elle est idéale pour exécuter des modèles plus volumineux sur une mémoire GPU limitée ou pour accélérer l'inférence, prenant en charge des formats comme INT8, NF4 et FP4. La compétence s'intègre à HuggingFace Transformers et permet l'entraînement QLoRA ainsi que l'utilisation d'optimiseurs en 8 bits.
dispatching-parallel-agents
AutreCette compétence Claude déploie plusieurs agents pour enquêter et résoudre simultanément 3 problèmes indépendants ou plus. Elle est conçue pour des scénarios impliquant des défaillances non liées qui peuvent être résolues sans état partagé ni dépendances. La capacité fondamentale est la résolution de problèmes en parallèle, en assignant un agent par domaine problématique indépendant afin de maximiser l'efficacité.
