review（审查）

Code Review Workflow

Name

han-core:review - Multi-agent code review with confidence-based filtering

Synopsis

/review [arguments]

Description

Multi-agent code review with confidence-based filtering

Implementation

Automated multi-agent code review with confidence-based filtering. Runs parallel specialized review agents to identify high-confidence issues across quality, security, and discipline-specific concerns.

Overview

This command orchestrates multiple review agents in parallel to provide comprehensive code review:

• General Quality: Code correctness, maintainability, testing
• Security: Vulnerabilities, auth/authz, input validation
• Discipline-Specific: Frontend, backend, or specialized concerns

Key Features:

• ✅ Parallel agent execution for speed
• ✅ Confidence scoring (0-100) with ≥80% threshold
• ✅ False positive filtering
• ✅ Automatic de-duplication
• ✅ Consolidated findings report

How It Works

1. Preparation

Gathers context about the changes:

# Review scope of changes
git diff --stat main...HEAD

# Get full diff
git diff main...HEAD

# Check recent commits
git log main...HEAD --oneline

2. Parallel Review Agents

Launches multiple independent agents in parallel (single message, multiple Task calls):

Core Reviewer (Always Runs)

• Agent: han-core:code-reviewer skill
• Focus: General quality, correctness, maintainability, testing
• Filters: Confidence ≥80%, false positive filtering
• Output: Categorized issues (Critical ≥90%, Important ≥80%)

Security Reviewer (Always Runs)

• Agent: security:security-engineer
• Focus: Security vulnerabilities, auth patterns, input validation
• Checks: SQL injection, XSS, CSRF, auth bypass, secrets exposure
• Output: Security issues with severity and confidence scores

Discipline-Specific Reviewer (Auto-Selected)

Auto-detection based on changed files:

Manual override: Specify agent explicitly if auto-detection is incorrect.

3. Consolidation

Merges findings from all agents:

1. Collect all issues from parallel agents
2. De-duplicate identical findings
3. Filter for confidence ≥80%
4. Categorize by severity:

- 🔴 Critical (confidence ≥90%): Must fix before merge - 🟡 Important (confidence ≥80%): Should fix before merge

1. Format with file:line references

4. Report

Presents consolidated findings:

## Review Summary

Total files changed: X
Lines added: +X, removed: -X
Review agents: 3 (Core, Security, Frontend)

---

## Findings

### 🔴 Critical Issues (Must Fix)

**[Issue]** - `file.ts:42` - **Confidence: 95%**
- Problem: ...
- Impact: ...
- Fix: ...

### 🟡 Important Issues (Should Fix)

**[Issue]** - `file.ts:89` - **Confidence: 85%**
- Problem: ...
- Impact: ...
- Suggestion: ...

---

## Verification Status

- [ ] All automated checks passed
- [ ] Security review: 1 critical issue
- [ ] Quality review: 0 issues
- [ ] Frontend review: 2 important issues

---

## Decision: REQUEST CHANGES

Critical issues must be resolved before approval.

---

## Next Actions

1. Fix SQL injection vulnerability at `services/user.ts:42`
2. Add error handling to `components/UserForm.tsx:89`
3. Re-run /review after fixes

Usage

Review current branch

/review

Reviews all changes from main branch to HEAD.

Review specific PR

/review pr 123

Uses gh CLI to fetch PR #123 and review changes.

Review with specific agents

/review --agents security,performance

Override auto-detection and run only specified agents.

Review specific files

/review src/services/payment.ts

Review only specified files instead of entire diff.

Confidence Scoring

All findings include confidence scores to reduce noise:

Filtering rules:

• ❌ Pre-existing issues (not in current diff)
• ❌ Linter-catchable issues (automated tools handle these)
• ❌ Code with lint-ignore comments
• ❌ Style preferences without documented standards
• ❌ Theoretical concerns without evidence

Agent Details

Core Reviewer (han-core:code-reviewer)

Dimensions:

1. Correctness - Does it solve the problem?
2. Safety - Security and data integrity
3. Maintainability - Readable, documented, follows patterns
4. Testability - Tests exist and cover edge cases
5. Performance - No obvious performance issues
6. Standards - Follows coding standards

Red flags (never approve):

• Commented-out code
• Secrets/credentials in code
• Breaking changes without coordination
• Tests commented out or skipped
• No tests for new functionality

Security Reviewer (security:security-engineer)

Focus areas:

• Input validation and sanitization
• SQL injection prevention
• XSS/CSRF protection
• Authentication and authorization
• Secrets management
• API security (rate limiting, CORS)
• Dependency vulnerabilities

Severity levels:

• Critical: Exploitable vulnerabilities
• High: Security pattern violations
• Medium: Potential security concerns

Discipline-Specific Reviewers

Each specialized agent brings domain expertise:

Frontend (presentation-engineer):

• Accessibility (WCAG compliance)
• Responsive design
• Performance (bundle size, lazy loading)
• User experience
• Component patterns

Backend (backend-architect):

• API design (RESTful, GraphQL)
• Error handling and validation
• Database transactions
• Caching strategies
• Scalability concerns

Database (database-designer):

• Query optimization
• Index usage
• Migration safety
• Data integrity
• Schema design

Integration with Workflows

Part of /feature-dev workflow

Phase 6: Review (from /feature-dev)
  ↓
  Calls /review command
  ↓
  Reports findings
  ↓
  User fixes issues
  ↓
  Re-run /review until clean

Standalone usage

# Make changes
git add .

# Review before commit
/review

# Fix issues

# Review again
/review

# If clean, commit
/commit

PR review automation

# Fetch PR
gh pr checkout 123

# Review changes
/review

# Comment on PR
gh pr comment 123 --body "$(cat review-findings.md)"

Advanced Features

Redundant Review for Critical Code

For high-risk changes (auth, payments, security), run redundant reviewers:

/review --redundant

This runs:

• 2x Core reviewers (independent evaluations)
• 2x Security reviewers (double-check vulnerabilities)
• 1x Discipline-specific reviewer

Consensus logic: Report issue only if ≥2 reviewers agree (reduces false positives).

Historical Context Review

Include git history analysis:

/review --with-history

Adds:

• Blame analysis: Who wrote original code?
• Change patterns: Frequently modified files (potential hot spots)
• Regression risk: Areas with past bugs
• Commit context: Related commits and their impact

Custom Agent Teams

Define custom agent combinations:

/review --team security-critical

Uses pre-defined team from .claude/review-teams.json:

{
  "security-critical": [
    "han-core:code-reviewer",
    "security:security-engineer",
    "security:security-engineer", // redundant
    "infrastructure:devops-engineer"
  ]
}

Configuration

.claude/settings.json

{
  "review": {
    "confidenceThreshold": 80,
    "autoSelectAgents": true,
    "enableRedundancy": false,
    "includeHistory": false,
    "maxIssuesPerCategory": 10
  }
}

Project-specific standards

Review agents check these files for project standards:

• CLAUDE.md - Project-specific guidelines
• CONTRIBUTING.md - Contribution standards
• .github/PULL_REQUEST_TEMPLATE.md - PR requirements

Best Practices

DO

• ✅ Run /review before creating PR
• ✅ Fix critical issues (≥90%) before requesting human review
• ✅ Re-run /review after fixing issues
• ✅ Trust agent consolidation (de-duplication)
• ✅ Let agents run in parallel for speed

DON'T

• ❌ Ignore critical findings
• ❌ Report issues with <80% confidence
• ❌ Run agents sequentially (use parallel)
• ❌ Second-guess agent findings without evidence
• ❌ Skip re-review after fixes

Troubleshooting

"No issues found" but code has problems

Likely causes:

• Issues have <80% confidence (adjust threshold?)
• Pre-existing issues (not in current diff)
• Automated tools already catch them

Solutions:

• Check linter/type checker output
• Run with --confidence-threshold 70 to see filtered issues
• Manually review automated tool results

Too many low-value findings

Likely causes:

• Agents reporting medium-confidence issues
• No project-specific standards documented

Solutions:

• Verify confidence threshold is ≥80%
• Document standards in CLAUDE.md
• Use false positive filters

Agents disagree on same issue

Normal: Different perspectives are valuable

Resolution:

• Higher confidence score wins
• Security concerns override others
• Consolidation chooses most specific finding

See Also

• /feature-dev - Full feature development workflow (includes review)
• /commit - Smart commit after review passes
• han-core:code-reviewer - Core review skill documentation
• security - Security agent details