Overview
PR-AF is an open-source agentic code review system built on AgentField, ranked #1 open-source code reviewer on Martian Code-Review-Bench. Instead of shallow diff summaries, it builds a task-specific review plan per PR, spawns focused reviewer agents, grounds findings in code evidence, and challenges its own results. It is model-flexible, running cheaper models for routine PRs and frontier models for high-stakes reviews.
Key Features
- Task-specific review plans with spawned focused reviewer agents.
- #1 open-source reviewer on Code-Review-Bench golden recall (0.706 with GLM-5.2).
- Self-hosted API with Docker; trigger reviews via CLI, curl, CI, or other agents.
- Model-flexible routing from cheap models to Opus-class frontier models.
Use Cases
- Deep automated code review for pull requests in CI.
- Self-hosted review gates with open models for compliance.
- Agentic review intelligence grounded in code evidence.
Technical Details
- Plan-then-review pipeline: per-PR review plan → spawned focused reviewer agents → evidence-grounded findings → self-challenge stage that verifies results.
- Model routing by PR stakes: cheaper models for routine reviews, GLM-5.2 for open-model CI gates, Opus-class frontier models for major PRs.
- Self-hosted Docker API; reviews triggered via CLI, curl, CI, or other agents.
- Reproducible benchmarking: public per-PR judge verdicts and reproduction scripts.