review

PR Review Orchestrator

Autonomous multi-agent PR review: invoke AI agents, await their reviews, process all findings, verify build.

How It Works

This skill calls the review MCP prompt from the pr server. The prompt dynamically builds an orchestration plan with pre-fetched PR data (summary, work status, agent config). The plan includes:

1. Delegation instructions — create tracking tasks, spawn a background sonnet agent
2. Full orchestrator workflow — escape check, invoke, await, spawn workers, build/test
3. Worker templates — for parallel comment processing

You don't need to understand the internals — just call the prompt and follow its instructions.

Invocation

/pr:review              — Context-dependent (see below)
/pr:review 42           — PR #42 in current repo
/pr:review owner/repo#5 — Specific PR in any repo
/pr:review https://github.com/owner/repo/pull/123 — Full URL

No-argument behavior depends on current branch:

• Feature branch → auto-detects the PR for this branch (single PR mode)
• Main/master → processes all open PRs with unresolved review comments (batch mode, up to 20)
• Main/master + no PRs with comments → processes first 5 open PRs

Workflow

Step 1: Call the MCP Prompt

Use the `review` prompt from the `pr` MCP server, passing $ARGUMENTS as the `pr` parameter.

The prompt returns a complete orchestration plan. It has two sections:

• DELEGATION — instructions for YOU (the parent agent)
• ORCHESTRATOR PROMPT — instructions for the background agent you'll spawn

Step 2: Follow the Delegation Plan

The delegation section tells you to:

1. Create 4 tracking tasks via TaskCreate:

   • Preflight checks
   • Wait for AI agent reviews
   • Process review comments
   • Build and test

2. Spawn a background agent via Agent(model: "sonnet", run_in_background: true) with the ORCHESTRATOR PROMPT section as its prompt.

3. Monitor progress — when the background agent sends progress notifications, call pr_progress_check and update your tasks accordingly.

Step 3: Stay Available

While the background agent works autonomously, you remain free to handle other user requests. The background agent:

• Invokes review agents (pr_invoke)
• Waits for reviews server-side (pr_await_reviews — blocks internally, not you)
• Spawns parallel workers to process comments
• Runs build and tests
• Reports completion

You'll be notified when it finishes. Update tasks and report results to the user.

Key Tools Used by the Orchestrator

Tool	Purpose
pr_invoke	Trigger AI review agents, get since + invokedAgentIds
pr_await_reviews	Non-blocking: check agent completion once, return immediately with retryAfterMs hint
pr_summary	Get review statistics
pr_claim_work	Claim file partition for comment processing
pr_get	Read full comment details
pr_resolve	Mark thread as resolved after fixing
pr_labels	Set review status labels
pr_progress_update	Report orchestrator phase transitions

Edge Cases

• No PR specified + on feature branch: auto-detects PR for current branch (single PR mode)
• No PR specified + on main/master: processes all open PRs with unresolved comments (batch mode)
• PR on different branch: refuses with branch protection message — switch to the right branch first
• Another orchestrator already running: detects via preflight check, aborts if run is fresh (<5 min)
• Agent already reviewed: smart detection skips agents that already posted reviews (use force in pr_invoke options to override)
• All comments resolved, no unresolved: skips worker spawning, goes straight to build/test

Comment Processing Rules (MANDATORY)

ALL severities MUST be processed. Skipping = FAILURE:

• CRIT and MAJOR — fix immediately
• MINOR — you MUST fix. "Too minor" is NOT a valid reason to skip.
• NITPICK — you MUST fix. Code quality at every level is required.

You MUST confidence-check EVERY suggestion BEFORE applying it: AI review agents produce WRONG suggestions. Blindly applying them breaks code.

1. READ the suggestion. VERIFY it is factually correct.
2. If CORRECT → apply the fix, resolve the thread.
3. If WRONG → resolve the thread with a reply explaining WHY it is incorrect.
4. BLINDLY APPLYING a wrong suggestion = FAILURE. SKIPPING a correct one = FAILURE.

What NOT to Do

• Don't call pr_list, pr_get, pr_resolve, Read, Edit yourself — that's the workers' job
• Don't process comments directly — spawn workers via the orchestrator
• Don't call pr_merge — review completes, human merges
• NEVER skip MINOR/NITPICK comments — you MUST process ALL severities
• NEVER blindly apply a suggestion — you MUST verify it is correct first