leo-claude-mktplace/plugins/pr-review/skills/review-workflow.md

Multi-Agent Review Workflow

Overview

PR reviews use specialized agents that analyze different aspects of code changes. The coordinator dispatches to relevant agents and aggregates findings.

Agent Roles

Agent	Focus Area	File Patterns
Security Reviewer	Vulnerabilities, auth, injection	*.ts, *.js, *.sql, *.py
Performance Analyst	N+1, inefficient algorithms, memory	*.ts, *.js, *.py, *.go
Maintainability Auditor	Complexity, readability, DRY	All code files
Test Validator	Coverage, edge cases, assertions	*.test.*, *_test.*, *_spec.*

Dispatch Logic

1. Analyze changed files - Identify file types and patterns
2. Select agents - Match files to relevant agents
3. Parallel dispatch - Send review tasks to agents
4. Collect findings - Aggregate results from all agents
5. Filter by confidence - Apply threshold (default: 0.7)
6. Generate report - Structure findings by severity

Skip Patterns

Don't dispatch agents for:

• *.md, *.txt - Documentation only
• *.json (config) - Unless security-sensitive
• Generated files - *.min.js, *.d.ts, etc.

Finding Structure

Each finding includes:

{
  id: "SEC-001",
  category: "security" | "performance" | "maintainability" | "tests",
  severity: "critical" | "major" | "minor" | "suggestion",
  confidence: 0.0 - 1.0,
  file: "src/api/users.ts",
  line: 45,
  title: "SQL Injection Vulnerability",
  description: "Detailed explanation...",
  fix: "Suggested code fix..." (optional)
}

Verdict Logic

Condition	Verdict
Any critical finding	REQUEST_CHANGES
2+ major findings	REQUEST_CHANGES
Only minor/suggestions	COMMENT
No significant findings	APPROVE

Confidence Thresholds

Reference: skills/review-patterns/confidence-scoring.md

Range	Label	Action
0.9 - 1.0	HIGH	Must address
0.7 - 0.89	MEDIUM	Should address
0.5 - 0.69	LOW	Consider addressing
< threshold	Filtered	Not shown

Default threshold: 0.7 (MEDIUM and above)

Aggregation Rules

When multiple agents find the same issue:

1. Keep highest confidence score
2. Merge descriptions
3. Use most severe category
4. Deduplicate fixes