review

• You are an orchestrator - Delegate review activities to specialist agents via Task tool
• Parallel execution - Launch ALL applicable review activities simultaneously in a single response
• Actionable feedback - Every finding must have a specific recommendation
• Let Claude Code route - Describe what needs review; the system selects appropriate agents

reference.md

• Detailed review checklists (Security, Performance, Quality, Testing)
• Severity and confidence classification matrices
• Agent prompt templates with FOCUS/EXCLUDE structure
• Synthesis protocol for deduplicating findings
• Example findings with proper formatting

1. Parse

   $ARGUMENTS

   to determine review target:
   • PR number → fetch PR diff via

     gh pr diff

   • Branch name → diff against main/master

   • staged

     → use

     git diff --cached

   • File path → read file and recent changes
2. Retrieve full file contents for context (not just diff)
3. Analyze changes to determine which conditional perspectives apply:
   • Contains async/await, Promise, threading → include Concurrency
   • Modifies dependency files → include Dependencies
   • Changes public API/schema → include Compatibility
   • Modifies frontend components → include Accessibility
   • Project has CONSTITUTION.md → include Constitution
4. Count applicable perspectives and assess scope of changes.

AskUserQuestion

• Standard (default recommendation): Subagent mode — parallel fire-and-forget agents. Best for straightforward reviews with independent perspectives.
• Team Mode: Persistent teammates with shared task list and peer coordination. Best for complex reviews where reviewers benefit from cross-perspective communication. Requires

  CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS

  in settings.

• Diff touches 10+ files across multiple domains
• 4+ review perspectives are applicable
• Changes span both frontend and backend
• Constitution enforcement is active alongside other reviews

• User selects Standard → Continue to Phase 2 (Standard)
• User selects Team Mode → Continue to Phase 2 (Team Mode)

Review this code for [PERSPECTIVE]:

CONTEXT:
- Files changed: [list]
- Changes: [the diff or code]
- Full file context: [surrounding code]
- Project standards: [from CLAUDE.md, .editorconfig, etc.]

FOCUS: [What this perspective looks for - from table above]

OUTPUT: Return findings as a structured list, one per finding:

FINDING:
- severity: CRITICAL | HIGH | MEDIUM | LOW
- confidence: HIGH | MEDIUM | LOW (see reference.md for classification matrix)
- title: Brief title (max 40 chars, e.g., "Missing null check in auth service")
- location: Shortest unique path + line (e.g., "auth/service.ts:42-45")
- issue: One sentence describing what's wrong (e.g., "Query result accessed without null check, causing NoneType errors")
- fix: Actionable recommendation (e.g., "Add null guard: `if result is None: raise ServiceError()`")
- code_example: (Optional, include for CRITICAL and non-obvious HIGH severity)
  ```language
  // Before
  const data = result.data;

  // After
  if (!result) throw new Error('No result');
  const data = result.data;

• HIGH: Clear violation of established pattern or security rule
• MEDIUM: Likely issue but context-dependent
• LOW: Potential improvement, may not be applicable

Continue to **Phase 3: Synthesize & Present**.

---

Requires

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS

enabled in settings.

1. Create team — derive name from review target (e.g.,

   review-pr-123

   ,

   review-feature-auth

   ,

   review-staged

   )
2. Create one task per applicable review perspective — all independent, no dependencies. Each task should describe the perspective focus, the files changed, diff context, and expected output format (structured FINDING format from the Standard perspective template).
3. Spawn one reviewer per perspective:

security-reviewer

team:the-architect:review-security

simplification-reviewer

team:the-architect:review-complexity

performance-reviewer

team:the-developer:optimize-performance

quality-reviewer

general-purpose

test-reviewer

team:the-tester:test-quality

concurrency-reviewer

team:the-developer:review-concurrency

dependency-reviewer

team:the-devops:review-dependency

compatibility-reviewer

team:the-architect:review-compatibility

accessibility-reviewer

team:the-designer:build-accessibility

Fallback: If team plugin agents are unavailable, use

general-purpose

for all.

1. Assign each task to its corresponding reviewer.

Deduplication algorithm:
1. Collect all findings from all reviewers
2. Group by location (file:line range overlap — within 5 lines = potential overlap)
3. For overlapping findings:
   a. Keep the highest severity version
   b. Merge complementary details (e.g., security + quality insights)
   c. Credit both perspectives in the finding
4. Sort by severity (Critical > High > Medium > Low) then confidence
5. Assign finding IDs (C1, C2, H1, H2, M1, etc.)
6. Build summary table

1. Collect all findings from review activities
2. Deduplicate overlapping findings (keep highest severity)
3. Rank by severity (Critical > High > Medium > Low) then confidence
4. Group by category for readability

undefined

• ✅ [Positive observation with specific code reference]
• ✅ [Good patterns noticed]

AskUserQuestion

• "Address critical issues first"
• "Show me fixes for [specific issue]"
• "Explain [finding] in more detail"

• "Apply suggested fixes"
• "Create follow-up issues for medium findings"
• "Proceed without changes"

• "Add to PR comments (if PR review)"
• "Done"

• Parallel execution - All review activities run simultaneously for speed
• Intent-driven - Describe what to review; the system routes to specialists
• Actionable output - Every finding must have a specific, implementable fix
• Positive reinforcement - Always highlight what's done well
• Context matters - Provide full file context, not just diffs
• Team mode specifics - Reviewers can coordinate via peer DMs to reduce duplicate findings; lead handles final dedup at synthesis
• User-facing output - Only the lead's synthesized output is visible to the user; do not forward raw reviewer messages