Files

Zhongwei Li befff05008 Initial commit

2025-11-30 09:01:45 +08:00

14 KiB

Raw Permalink Blame History

description, allowed-tools, model, argument-hint

description	allowed-tools	model	argument-hint
Orchestrate multiple specialized review agents with dynamic context analysis, hierarchical task decomposition, and confidence-based filtering. Use after code changes or when comprehensive quality assessment is needed. Includes security, performance, accessibility, type safety, and more. All findings include evidence (file:line) and confidence markers (✓/→/?) per Output Verifiability principles. 複数の専門エージェントによるコードレビューを実行。セキュリティ、パフォーマンス、アクセシビリティなど包括的な品質評価。	Bash(git diff:), Bash(git status:), Bash(git log:), Bash(git show:), Read, Glob, Grep, LS, Task	inherit	[target files or scope]

/review - Advanced Code Review Orchestrator

Purpose

Orchestrate multiple specialized review agents with dynamic context analysis, hierarchical task decomposition, and confidence-based filtering.

Output Verifiability: All review findings include evidence (file:line), distinguish verified issues (✓) from inferred problems (→), per AI Operation Principle #4.

Integration with Skills

This command explicitly references the following Skills:

[@~/.claude/skills/security-review/SKILL.md] - Security review knowledge based on OWASP Top 10

Other Skills are automatically loaded through each review agent's dependencies:

performance-reviewer → performance-optimization skill
readability-reviewer → readability-review skill
progressive-enhancer → progressive-enhancement skill

Dynamic Context Analysis

Git Status

Check git status:

!`git status --porcelain`

Files Changed

List changed files:

!`git diff --name-only HEAD`

Recent Commits

View recent commits:

!`git log --oneline -10`

Change Statistics

Show change statistics:

!`git diff --stat HEAD`

Specification Context (Auto-Detection)

Discover Latest Spec

Search for spec.md in SOW workspace using Glob tool (approved):

Use Glob tool to find spec.md:
- Pattern: ".claude/workspace/sow/**/spec.md"
- Alternative: "~/.claude/workspace/sow/**/spec.md"

Select the most recent spec.md if multiple exist (check modification time).

Load Specification

If spec.md exists, load it for review context:

Provides functional requirements for alignment checking
Enables "specification vs implementation" verification
Implements Article 2's approach: spec.md in review prompts
Allows reviewers to identify gaps like "仕様書ではこう定義されていますが、この実装ではそのケースが考慮されていません"

If spec.md does not exist:

Review proceeds with code-only analysis
Focus on code quality, security, and best practices
Consider creating specification with /think for future reference

Execution

Invoke the review-orchestrator agent to perform comprehensive code review:

Task({
  subagent_type: "review-orchestrator",
  description: "Comprehensive code review",
  prompt: `
Execute comprehensive code review with the following requirements:

### Review Context
- Changed files: Use git status and git diff to identify scope
- Recent commits: Analyze recent changes for context
- Project type: Detect technology stack automatically
- Specification: If spec.md exists in workspace, verify implementation aligns with specification requirements
  - Check if all functional requirements (FR-xxx) are implemented
  - Identify missing features defined in spec
  - Flag deviations from API specifications
  - Compare actual vs expected behavior per spec

### Review Process
1. **Context Discovery**: Analyze repository structure and technology stack
2. **Parallel Reviews**: Launch specialized review agents concurrently
   - structure-reviewer, readability-reviewer, progressive-enhancer
   - type-safety-reviewer, design-pattern-reviewer, testability-reviewer
   - performance-reviewer, accessibility-reviewer
   - document-reviewer (if .md files present)
3. **Filtering & Consolidation**: Apply confidence filters and deduplication

### Output Requirements
- All findings MUST include evidence (file:line)
- Use confidence markers: ✓ (>0.8), → (0.5-0.8)
- Only include findings with confidence >0.7
- Group by severity: Critical, High, Medium, Low
- Provide actionable recommendations

Report results in Japanese.
  `
})

Hierarchical Review Process Details

Phase 1: Context Discovery

Analyze repository and determine review scope:

Analyze repository structure and technology stack
Identify review scope (changed files, directories)
Detect code patterns and existing quality standards
Determine applicable review categories

Phase 2: Parallel Specialized Reviews

Launch multiple review agents concurrently:

Each agent focuses on specific aspect
Independent execution for efficiency
Collect raw findings with confidence scores

Phase 3: Filtering and Consolidation

Apply multi-level filtering with evidence requirements:

Confidence Filter: Only issues with >0.7 confidence
Evidence Requirement: All findings MUST include:
- File path with line number (e.g., src/auth.ts:42)
- Specific code reference or pattern
- Reasoning for the issue
False Positive Filter: Apply exclusion rules
Deduplication: Merge similar findings
Prioritization: Sort by impact and severity

Confidence Mapping:

✓ High Confidence (>0.8): Verified issue with direct code evidence
→ Medium Confidence (0.5-0.8): Inferred problem with reasoning
? Low Confidence (<0.5): Not included in output (too uncertain)

Review Agents and Their Focus

Core Architecture Reviewers

review-orchestrator: Coordinates all review activities
structure-reviewer: Code organization, DRY violations, coupling
root-cause-reviewer: Deep problem analysis, architectural debt

Quality Assurance Reviewers

readability-reviewer: Code clarity, naming, complexity
type-safety-reviewer: TypeScript coverage, any usage, type assertions
testability-reviewer: Test design, mocking, coverage gaps

Specialized Domain Reviewers

Security review (via security-review skill): OWASP Top 10 vulnerabilities, auth issues, data exposure
accessibility-reviewer: WCAG compliance, keyboard navigation, ARIA
performance-reviewer: Bottlenecks, bundle size, rendering issues
design-pattern-reviewer: Pattern consistency, React best practices
progressive-enhancer: CSS-first solutions, graceful degradation
document-reviewer: README quality, API docs, inline comments

Exclusion Rules

Automatic Exclusions (False Positive Prevention)

Style Issues: Formatting, indentation (handled by linters)
Minor Naming: Unless severely misleading
Test Files: Focus on production code unless requested
Generated Code: Build outputs, vendor files
Documentation: Unless specifically reviewing docs
Theoretical Issues: Without concrete exploitation path
Performance Micro-optimizations: Unless measurable impact
Missing Features: vs actual bugs/issues

Context-Aware Exclusions

Framework-specific patterns (React/Angular/Vue idioms)
Project conventions (detected from existing code)
Language-specific safety (memory-safe languages)
Environment assumptions (browser vs Node.js)

Output Format with Confidence Scoring

IMPORTANT: Use both numeric scores (0.0-1.0) and visual markers (✓/→) for clarity.

[REVIEW OUTPUT TEMPLATE]

Review Summary
- Files Reviewed: [Count and list]
- Total Issues: [Count by severity with markers]
- Review Coverage: [Percentage]
- Overall Confidence: [✓/→] [Average score]

## ✓ Critical Issues 🚨 (Confidence > 0.9)
Issue #1: [Title]
- **Marker**: [✓] High Confidence
- **File**: path/to/file.ts:42-45
- **Category**: security|performance|accessibility|etc
- **Confidence**: 0.95
- **Evidence**: [Specific code snippet or pattern found]
- **Description**: [Detailed explanation]
- **Impact**: [User/system impact]
- **Recommendation**: [Specific fix with code example]
- **References**: [Related files, docs, or standards]

## ✓ High Priority ⚠️ (Confidence > 0.8)
Issue #2: [Title]
- **Marker**: [✓] High Confidence
- **File**: path/to/another.ts:123
- **Evidence**: [Direct observation]
- **Description**: [Issue explanation]
- **Recommendation**: [Fix with example]

## → Medium Priority 💡 (Confidence 0.7-0.8)
Issue #3: [Title]
- **Marker**: [→] Medium Confidence
- **File**: path/to/file.ts:200
- **Inference**: [Reasoning behind this finding]
- **Description**: [Issue explanation]
- **Recommendation**: [Suggested improvement]
- **Note**: Verify this inference before implementing fix

Improvement Opportunities
[→] Lower confidence suggestions (0.5-0.7) for consideration
- Mark as [→] to indicate these are recommendations, not confirmed issues

Metrics
- Code Quality Score: [A-F rating] [✓/→]
- Technical Debt Estimate: [Hours] [✓/→]
- Test Coverage Gap: [Percentage] [✓]
- Security Posture: [Rating] [✓/→]

Recommended Actions
1. **Immediate** [✓]: [Critical fixes with evidence]
2. **Next Sprint** [✓/→]: [High priority items]
3. **Backlog** [→]: [Nice-to-have improvements]

Evidence Summary
- **Verified Issues** [✓]: [Count] - Direct code evidence
- **Inferred Problems** [→]: [Count] - Based on patterns/reasoning
- **Total Confidence**: [Overall score]

Review Strategies

Quick Review (2-3 min)

Focus areas:

Security vulnerabilities
Critical bugs
Breaking changes
Accessibility violations

Command: /review --quick

Standard Review (5-7 min)

Includes Quick + :

Performance issues
Type safety problems
Test coverage gaps
Code organization

Command: /review (default)

Deep Review (10+ min)

Comprehensive analysis:

All standard checks
Root cause analysis
Technical debt assessment
Refactoring opportunities
Architecture evaluation

Command: /review --deep

Focused Review

Target specific areas:

/review --security - Security focus
/review --performance - Performance focus
/review --accessibility - A11y focus
/review --architecture - Design patterns

TodoWrite Integration

Automatic task creation:

[TODO LIST TEMPLATE]
Code Review: [Target]
1. ⏳ Context discovery and scope analysis
2. ⏳ Execute specialized review agents (parallel)
3. ⏳ Filter and validate findings (confidence > 0.7)
4. ⏳ Consolidate and prioritize results
5. ⏳ Generate actionable recommendations

Custom Review Instructions

Support for project-specific rules:

.claude/review-rules.md - Project conventions
.claude/exclusions.md - Custom exclusions
.claude/review-focus.md - Priority areas

Advanced Features

Incremental Reviews

Compare against baseline:

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

Pattern Detection

Identify recurring issues:

Similar problems across files
Systemic architectural issues
Common anti-patterns

Learning Mode

Track and improve:

False positive patterns
Project-specific idioms
Team preferences

Usage Examples

Basic Review

/review
# Reviews all changed files with standard depth

Targeted Review

/review "authentication module"
# Focuses on auth-related code

Security Audit

/review --security --deep
# Comprehensive security analysis

Pre-PR Review

/review --compare main
# Reviews changes against main branch

Component Review

/review "src/components" --accessibility
# A11y review of components directory

Best Practices

Review Early: Catch issues before they compound
Review Incrementally: Small, frequent reviews > large, rare ones
Apply Output Verifiability:
- Always provide evidence: File paths with line numbers
- Use confidence markers: ✓ for verified, → for inferred
- Explain reasoning: Why is this an issue?
- Reference standards: Link to docs, best practices, or past issues
- Never guess: If uncertain, mark as [→] and explain the inference
Act on High Confidence: Focus on ✓ (>0.8) issues first
Validate Inferences: [→] markers require verification before fixing
Track Patterns: Identify recurring problems
Customize Rules: Add project-specific exclusions
Iterate on Feedback: Tune confidence thresholds

Integration Points

Pre-commit Hook

claude review --quick || exit 1

CI/CD Pipeline

- name: Code Review
  run: claude review --security --performance

PR Comments

Results formatted for GitHub/GitLab comments

Applied Development Principles

Output Verifiability (AI Operation Principle #4)

All review findings MUST follow Output Verifiability:

Distinguish verified from inferred: Use ✓/→ markers
Provide evidence: File:line for every issue
State confidence explicitly: Numeric + visual marker
Explain reasoning: Why is this problematic?
Admit uncertainty: [→] when inferred, never pretend to know

Principles Guide

@~/.claude/rules/PRINCIPLES_GUIDE.md - Foundation for review prioritization

Application:

Priority Matrix: Categorize issues by Essential > Default > Contextual principles
Conflict Resolution: Decision criteria for DRY vs Readable, SOLID vs Simple, etc.
Red Flags: Method chains > 3 levels, can't understand in 1 minute, "just in case" implementations

Documentation Rules

@~/.claude/docs/DOCUMENTATION_RULES.md - Review report format and structure