AutoRAG-Research: Skill for Claude Code

/refactor -- 3-Agent PR Code Review Debate

You orchestrate a PR code review debate with 3 reviewer agents (Red/White/Green). Follow these 10 steps exactly.

Step 1: Verify PR Exists

Run:

gh pr view --json number,url,title,body,headRefName,baseRefName

If the command fails, exit immediately:

No PR found on the current branch. Create a PR first.

Extract and remember these values:

• PR_NUMBER -- the PR number
• PR_URL -- the PR URL
• PR_TITLE -- the PR title
• HEAD_REF -- head branch name
• BASE_REF -- base branch name

Step 2: Gather PR Diff and Context

Run these commands (in parallel where possible):

gh pr diff
gh pr diff --stat
git rev-parse HEAD
git remote get-url origin

Store results as:

• FULL_DIFF -- the full diff output
• DIFF_STAT -- the diff stat summary
• HEAD_SHA -- current HEAD commit SHA
• OWNER_REPO -- parse the remote URL to extract OWNER/REPO (handle both HTTPS and SSH formats)

If FULL_DIFF is empty, exit immediately:

Nothing to review. The PR diff is empty.

Step 3: Create Team and Spawn 3 Reviewers (Parallel)

3a. Create the team

TeamCreate(team_name="pr-review-{PR_NUMBER}")

3b. Create 3 tracking tasks

Create one task per reviewer:

• "Red Reviewer: Devil's Advocate analysis"
• "White Reviewer: Neutral Judge analysis"
• "Green Reviewer: Approval Advocate analysis"

3c. Spawn 3 reviewer agents in a SINGLE message (parallel)

Use the Task tool three times in the same message. Each agent:

• subagent_type: "general-purpose"
• team_name: "pr-review-{PR_NUMBER}"
• mode: "bypassPermissions"

Build each agent's prompt by filling in the appropriate role template below with FULL_DIFF, DIFF_STAT, PR_TITLE, and file list from diff stat.

Agent names and roles:

Name	Role
red-reviewer	Devil's Advocate -- always opposes, attacks security/race conditions/hidden assumptions
white-reviewer	Neutral Judge -- risk = probability x impact, pragmatic tradeoffs
green-reviewer	Approval Advocate -- champions strengths, only flags genuinely critical issues