Adversarial PR Review

Adversarial PR Review

A multi-agent, adversarial pull request reviewer. An orchestrator pulls the real change from GitHub, delegates to a primaryReviewer that proposes findings, then to a challenger on a different model vendor that disputes those findings and hunts for missed bugs. The orchestrator arbitrates both sides into one calibrated decision — approve, comment, or request changes — instead of dumping every issue a single model noticed.

Built from the pipeline in our guide: AI Code Review: A Multi-Agent, Adversarial Guide.

What this workflow does

1. prInput takes a pull request URL (or POST {"text":"<pr-url>"} to the workflow)
2. orchestrator uses the GitHub MCP to read the PR diff, commits, comments, and CI status, plus a Sequential Thinking MCP to reason step by step
3. primaryReviewer (Claude Opus) returns structured findings, each with a severity from S0–S4, a confidence score, a file and line, and a suggested fix
4. challenger (GPT, a different vendor) critiques every finding — accept, reject, or needs-more-evidence — flags wrong severities, and adds high-impact issues the reviewer missed
5. The orchestrator merges both sides, drops rejected findings, applies agreed severity changes, and writes a final decision plus a suggested PR comment to finalReview

Why adversarial

A lone reviewer is confident, agreeable, and biased toward quick wins, so it both invents problems and misses real ones. Giving the challenger the explicit job of disputing the primary reviewer — on a different model so they do not share blind spots — filters weak findings and surfaces what a single pass would miss. The structure, not any single model, is what improves the output.

Use cases

• Self-hosted PR review that keeps proprietary code on your own infrastructure
• High-stakes merges where a missed bug costs more than the extra tokens
• Cutting false-positive noise from a single-agent reviewer

Setup

1. Add an LLM credential to orchestrator, primaryReviewer, and challenger — mix vendors so the challenger differs from the reviewer
2. Provide a GitHub token: the GitHub MCP reads it from the global variable githubPat
3. Replace your-github-username in the orchestrator prompt with your handle
4. (Optional) Add web-search and website-loader sub-workflows to the orchestrator for external context (CVEs, library status, linked RFCs)
5. Run on a real PR, then open the Traces tab to see every tool call, each sub-agent, tokens and latency, and where the two agents disagreed

Notes

The reviewer and challenger return structured JSON; the orchestrator's merge policy decides the verdict (any surviving S0/S1 finding means request changes). Swap models per role to trade cost against depth.