code-review-pr

name	code-review-pr
description	Automated code review for pull requests using multi-perspective analysis and confidence-based scoring. Supports multiple LLM reviewers (Claude Code, GPT, Cursor) contributing findings to a shared format, plus a validator to aggregate and deduplicate. Use when reviewing PRs, code changes, contributing findings, or running /code-review, /code-review --validate.

Code Review (PR) - Multi-Agent

Automated code review with multi-perspective analysis and confidence scoring. Supports 3 reviewers + 1 validator workflow: Claude Code, GPT, and Cursor each contribute findings; the validator aggregates and deduplicates into a final report.

Modes

Mode	Command	What it does
Review	/code-review	Run review, write findings to .code-review/findings-{agent}.json
Validate	/code-review --validate	Read all findings, deduplicate, merge, output final report

Shared Findings Format

All agents write to .code-review/findings-{agent}.json. Use the agent slug: claude-code, gpt, or cursor.

File structure:

{
  "version": "1.0",
  "agent": "cursor",
  "timestamp": "2025-02-10T12:00:00Z",
  "pr": { "branch": "feature-x", "sha": "abc123..." },
  "findings": [
    {
      "title": "Missing error handling for OAuth callback",
      "reason": "CLAUDE.md compliance",
      "confidence": 85,
      "file": "src/auth.ts",
      "lineStart": 67,
      "lineEnd": 72,
      "sha": "abc123...",
      "url": "https://github.com/owner/repo/blob/abc123.../src/auth.ts#L67-L72",
      "detail": "CLAUDE.md says 'Always handle OAuth errors'"
    }
  ]
}

Review Mode (Contributing)

When the user says "code review" or /code-review:

1. Identify agent: Use cursor, claude-code, or gpt (ask if unclear)
2. Gather context: CLAUDE.md, .cursorrules, AGENTS.md
3. Get changes: git diff or gh pr diff
4. Audit from 4 perspectives:
   • Compliance (CLAUDE.md / guidelines)
   • Compliance (redundant pass)
   • Bugs (in changed code only)
   • History (git blame, context-based issues)
5. Score each issue 0-100 (see Confidence Scoring)
6. Filter: Keep only issues with confidence >= 80
7. Write to .code-review/findings-{agent}.json
8. Output a summary to chat (e.g., "Wrote 3 findings to .code-review/findings-cursor.json")

Validate Mode (Aggregating)

When the user says "validate review" or /code-review --validate:

1. Read all .code-review/findings-*.json files
2. Deduplicate:
   • Same file + overlapping line range + similar title/reason → treat as same issue
   • Merge: keep highest confidence, list which agents found it
3. Boost confidence when 2+ agents agree: add 5 per agreeing agent (cap at 100)
4. Filter: Keep only merged confidence >= 80
5. Output final report in review format (see below)
6. Optional: Post to PR with --comment if gh available

Confidence Scoring

Score	Meaning
0	Not confident, likely false positive
25	Somewhat confident, might be real
50	Moderately confident, real but minor
75	Highly confident, real and important
100	Absolutely certain, definitely real

For CLAUDE.md issues: only score 80+ if the guideline explicitly mentions the violation.

What to Filter Out (Do Not Report)

• Pre-existing issues not introduced in this PR
• Code that looks like a bug but isn't
• Pedantic nitpicks
• Issues linters will catch
• General quality issues (unless in CLAUDE.md)
• Issues where code has explicit lint-ignore comments

Final Report Format (Validator Output)

## Code review (validated)

Found N issues (from cursor, claude-code, gpt):

1. [Issue title] ([reason]) — found by: cursor, claude-code

[GitHub blob URL]

2. ...

Link Format

https://github.com/owner/repo/blob/[full-sha]/path/file.ext#L[start]-L[end]

When to Skip Review

• Closed or draft PRs
• Trivial / automated PRs
• PR already has a recent review
• No changes to review

Setup (Tell Everyone)

1. Create .code-review/ in repo root (add to .gitignore if findings should stay local)
2. Install this skill into .cursor/skills/code-review-pr/ (or personal ~/.cursor/skills/)
3. Agent identifier: When running review, specify agent: "I'm using Claude Code" or "run as gpt" so the correct findings-{agent}.json is written

Workflow

# Person 1 (Claude Code): contribute findings
/code-review
# -> writes .code-review/findings-claude-code.json

# Person 2 (GPT): contribute findings
/code-review
# -> writes .code-review/findings-gpt.json

# Person 3 (Cursor): contribute findings
/code-review
# -> writes .code-review/findings-cursor.json

# Anyone: validate and get final report
/code-review --validate
# -> reads all 3, merges, deduplicates, outputs final report

# Optional: post to PR
/code-review --validate --comment

Requirements

• Git repository (GitHub preferred for link format)
• gh CLI for PR details and posting (optional)
• .code-review/ directory (create on first run)

Additional Resources

• For troubleshooting and configuration, see reference.md
• For findings JSON schema, see schema.md