A pull request review skill with four parallel reviewers

---
name: pr-review-graph
description: Reviews a pull request with four parallel specialists, filters false positives in a fresh context, and writes prioritized PR comments.
spec: skill-graph/1
topology:
  nodes:
    - id: goal
      type: input
      label: PULL REQUEST (diff or link)
    - id: splitter
      type: orchestrator
      label: SPLIT
      model: sonnet
      tools:
        - Read
    - id: review_security
      type: worker
      label: Security & Auth
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: review_performance
      type: worker
      label: Performance & Memory
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: review_architecture
      type: worker
      label: Architecture & Style
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: review_tests
      type: worker
      label: Test Coverage
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: verifier
      type: verifier
      label: VERIFIER (Fresh Context)
      isolated_context: true
      model: opus
      tools:
        - Read
        - Grep
        - Glob
      gate: Every kept finding points at a file and line in the diff.
      on_fail: retry
      max_retries: 1
    - id: merger
      type: synthesizer
      label: FINAL ANSWER
      model: sonnet
      tools:
        - Read
  edges:
    - from: goal
      to: splitter
    - from: splitter
      to:
        - review_security
        - review_performance
        - review_architecture
        - review_tests
      contract: subtask_json
    - from:
        - review_security
        - review_performance
        - review_architecture
        - review_tests
      to: verifier
      contract: review_findings
    - from: verifier
      to: merger
      contract: verified_findings
contracts:
  subtask_json:
    format: json
    description: One key per worker id; each value is that worker's task.
  review_findings:
    format: markdown
    description: A numbered list; every item has its evidence and a confidence of high, medium or low.
  verified_findings:
    format: markdown
    description: Only the items that passed verification, then the rejected ones with the reason.
---

# Node Prompts & Edge Contracts

## @node: splitter
You are the review lead. Read the pull request and give each reviewer the files and concerns that matter for their specialty.

Input:
{{goal}}

Output strictly as JSON with one key per worker: `review_security`, `review_performance`, `review_architecture`, `review_tests`. Each value is the exact task for that worker.

## @node: review_security
Review ONLY for security: injection, authn/authz gaps, secrets, unsafe deserialization, SSRF, missing input validation.

Your task: {{splitter.review_security}}

Return findings as a numbered list. For each one give the evidence (file, line, quote or source) and a confidence of high, medium or low.

## @node: review_performance
Review ONLY for performance: N+1 queries, unbounded loops or payloads, leaks, blocking calls on hot paths, missing caching.

Your task: {{splitter.review_performance}}

Return findings as a numbered list. For each one give the evidence (file, line, quote or source) and a confidence of high, medium or low.

## @node: review_architecture
Review ONLY for architecture: layering violations, duplicated logic, naming, consistency with existing patterns in the codebase.

Your task: {{splitter.review_architecture}}

Return findings as a numbered list. For each one give the evidence (file, line, quote or source) and a confidence of high, medium or low.

## @node: review_tests
Review ONLY test coverage: untested branches, missing edge cases, brittle or meaningless assertions.

Your task: {{splitter.review_tests}}

Return findings as a numbered list. For each one give the evidence (file, line, quote or source) and a confidence of high, medium or low.

## @node: verifier
You are an independent auditor. Re-check each finding against the diff itself and remove false positives.

You only see the raw worker outputs below, not their reasoning. Drop any claim without concrete evidence, merge duplicates, and mark what you rejected and why.

Original input: {{goal}}

Worker outputs:
{{workers.output}}

## @node: merger
Write the final review as prioritized PR comments (blocking, should fix, nit). Each comment names the file and line and says what to change.

{{verifier.output}}