review

name	review
description	Reviews code for quality, security, correctness. Triggers: code review, quality review, security review, review PR, review branch.
user-invocable	true
effort	high
argument-hint	[target: branch, pr, file path, or staged changes]
agent	code-reviewer
context	fork
allowed-tools	Read, Grep, Glob, Bash

Code Review

$ARGUMENTS

Reviews code changes for quality and issues.

Changed files context

• Changes: !git diff --stat main...HEAD 2>/dev/null || git diff --cached --stat 2>/dev/null || echo "no changes detected"

Signal Collection (never stop at the first red)

Collect every failing signal up front, then review the diff in full anyway:

Signal	How to read it
Merge conflict with base	gh pr view --json mergeable,mergeStateStatus or git merge-tree
Failing CI checks	gh pr checks or the platform equivalent
Lint / typecheck failure	the project's own commands

Each failing signal becomes a blocker finding. None of them ends the run.

A review that aborts on the first red signal spends the whole cycle repeating what the tracker already displayed, while the finding that would have told the author something new never gets written. One invocation produces the most complete picture of the change that it can.

Automated Diff Analysis

Before starting manual review, run the diff analyzer script to get a structured risk assessment:

python3 ${CLAUDE_SKILL_DIR}/scripts/diff-analyzer.py [base_branch]
# Default base branch: main
# Example: python3 ${CLAUDE_SKILL_DIR}/scripts/diff-analyzer.py develop

The script outputs JSON with:

• files: each changed file with additions, deletions, category (security/test/config/migration/infra/docs/logic), and risk level
• risk_score: overall assessment (high/medium/low)
• hotspots: top 5 files by additions
• secrets_scan: potential secret leaks detected in added lines
• test_coverage_estimate: whether test files accompany logic changes (good/partial/none)
• parallel_review_recommended: boolean flag

If the script reports parallel_review_recommended: true, use the Parallel Review (Agent Teams) mode below.

---

Parallel Review (Agent Teams)

For significant PRs or large changesets, create a parallel review team:

Create an agent team to review [target]:
- Teammate 1 (security-auditor): "Review for security vulnerabilities, auth issues,
  injection risks, secret leaks. Report with severity ratings." Use Opus.
- Teammate 2 (performance-optimizer): "Check for N+1 queries, memory leaks,
  unnecessary allocations, caching opportunities. Report with impact ratings." Use Opus.
- Teammate 3 (test-engineer): "Validate test coverage, edge cases, mock quality,
  missing assertions. Report coverage gaps." Use Opus.
Each reviewer should report findings independently. Do NOT modify files.

After all reviewers complete:

1. Synthesize findings into unified Code Review Report
2. Prioritize by severity (blocker > major > minor > nit)
3. Issue verdict per the verdict rule below — not by impression

When to use: PRs with >5 files changed, cross-module changes, security-sensitive code. READ-ONLY: No teammate should modify files during review.

---

Sequential Review (Default)

1. Reads changed files
2. Analyzes for issues
3. Checks best practices
4. Reports findings

Review Scope

Target	What's Reviewed
(none)	Staged changes
branch	Branch vs main
pr	Pull request changes
file.ts	Specific file

Review Checklist

Code Quality

• Clear naming
• Proper error handling
• No code duplication
• Appropriate abstractions

Security (OWASP Top 10)

• A01: Proper auth/authorization on all endpoints
• A02: No weak crypto, HTTPS for external comms
• A03: Input validation, parameterized queries, output encoding (XSS)
• A04: Threat model assumptions documented for new features
• A05: No debug mode, default credentials, or verbose errors in prod config
• A06: Dependencies checked for known CVEs
• A07: No hardcoded secrets, session management correct
• A08: Integrity checks on deserialized data, CI/CD pipeline safety
• A09: Security-relevant events logged (without PII)
• A10: External URL handling validates scheme/host (SSRF prevention)
• Cross-scope replay: can an identifier from one tenant/user/org be replayed in another?
• Fails closed wherever the path affects security, money, or data retention

API / Contract Changes

• Backward compatibility preserved (no silent breaking changes)
• API versioning updated if contract changed
• Schema validation on request/response
• Error responses follow project convention
• Wire-level contracts checked, not just code signatures: HTTP routes, webhook payloads, event/queue schemas

Concurrency / Async

• Shared mutable state protected (locks, atomics, channels)
• No fire-and-forget promises without error handling
• Database transactions scoped correctly (no long-held locks)
• Race condition risk assessed for concurrent access paths

Migrations / Schema Changes

• Migration is reversible (has rollback path)
• No table locks on large tables during peak hours
• Data backfill handles NULL/missing values
• Indexes added for new query patterns

Performance

• No N+1 queries
• Appropriate caching
• No memory leaks
• Optimized loops

Testing

• Tests for new code
• Edge cases covered
• Mocks appropriate

Severity & Verdict

Tier	Meaning	Merge impact
blocker	Causes damage: data loss, security hole, money, corruption	Blocks merge, no exceptions
major	Real defect that will bite in production	Blocks merge unless waived in writing
minor	Should be fixed, not worth blocking on	Does not block
nit	Polish, taste, style	Does not block

Verdict rule — apply it mechanically, do not negotiate with yourself:

• any blocker → REQUEST_CHANGES
• any major without a documented waiver (who waived it, why, what the follow-up is) → REQUEST_CHANGES
• only minor / nit → APPROVE
• the change cannot be classified from the diff → NEEDS_DISCUSSION, and state what would resolve it

Severity describes impact, confidence describes certainty — they are independent axes. A finding with confidence < 6 is reported at the tier its evidence supports and is never promoted to blocker on suspicion alone.

Output Format

## Code Review Report

### Summary
- **Files Changed**: [count]
- **Lines Added**: [+count]
- **Lines Removed**: [-count]
- **Issues Found**: [count]
- **Overall Confidence**: [1-10] — how confident the reviewer is in the assessment

### Findings

#### Blocker
- **[file:line]**: [issue]
  - Severity: blocker | Confidence: [1-10]
  - Evidence: [specific code reference and reasoning]
  - Suggested fix: [code]

#### Major
- **[file:line]**: [issue]
  - Severity: major | Confidence: [1-10]
  - Evidence: [specific code reference and reasoning]
  - Suggested fix: [code]

#### Minor
- **[file:line]**: [issue]
  - Severity: minor | Confidence: [1-10]
  - Evidence: [line number + reasoning]

#### Nit
- **[file:line]**: [suggestion]
  - Severity: nit | Confidence: [1-10]

### Confidence Guide

| Score | Meaning |
|-------|---------|
| 9-10 | Certain — verified via code, tests, or documentation |
| 7-8 | High — strong evidence, minor assumptions |
| 5-6 | Medium — plausible issue, needs author confirmation |
| 3-4 | Low — speculative, based on patterns not proof |
| 1-2 | Guess — flag for discussion, don't block on this |

### Positive Notes
- [What's good about the code]

### Verdict
[APPROVE / REQUEST_CHANGES / NEEDS_DISCUSSION]

State which clause of the verdict rule produced it, e.g.
"REQUEST

Common Rationalizations

Excuse	Why It's Wrong
"Small change, quick scan is enough"	Small changes introduce subtle bugs — apply consistent review regardless of size
"Tests pass, so the code is correct"	Tests validate specific scenarios, not all behaviors — verify missing coverage
"It's just a refactor, no need for deep review"	Refactors change invariants — verify behavior preservation, not just compilation
"The author is senior, they know what they're doing"	Seniority doesn't prevent mistakes — review the code, not the person
"We're in a hurry, ship it"	Rushed reviews create tech debt that costs 10x more to fix later

Self-Evaluation (LLM-as-Judge)

After completing the review, perform a self-evaluation pass:

Check for Blind Spots

1. Did I verify, or assume? — For each finding, confirm you read the actual code (not inferred from context)
2. Did I miss the inverse? — If you flagged X as a problem, did you check if NOT doing X is also a problem elsewhere?
3. Did I anchor on the first issue? — Review whether early findings biased you toward similar patterns, missing different issue classes
4. Did I check the unhappy path? — Error handling, edge cases, failure modes — not just the golden path
5. Did I flag uncertainty? — Findings with confidence < 6 should be clearly marked as "needs author input"

Calibrate Confidence

• If all findings are confidence 7+, you may be overconfident — re-examine the weakest finding
• If any finding lacks a file:line reference, downgrade it or remove it
• If you found zero issues, state what you specifically checked (not "looks good")

READ-ONLY

This skill only analyzes. It does NOT modify any files.

Related Skills

• Issues found? → /debug to trace root causes
• Missing tests? → /tdd to add test-first coverage
• Security findings? → /cve-scan for dependency vulnerabilities
• Architecture concerns? → /analyze for deeper code quality metrics