Agent-Field/pr-af

What it solves

PR-AF is designed for deep, architectural code reviews rather than simple diff summaries. It aims to reduce false positives and identify complex, systemic vulnerabilities that linear analysis tools often miss, providing a high-fidelity gate for CI/CD pipelines.

How it works

Instead of a static script, PR-AF uses a dynamic pipeline that morphs its execution graph based on the PR's topology. It evaluates the diff through semantic, mechanical, and systemic lenses to spawn specialized, ephemeral reviewer agents. The process includes:

• Evidence Grounding: It verifies findings by pulling exact caller snippets and import context from the repository to ensure claims are grounded in actual code.
• Compound Vulnerability Synthesis: It clusters related risks across multiple files to identify larger systemic issues.
• Falsifiability Gates: The system attempts to invalidate findings (checking for existing mitigations or intended behavior) before they are posted as GitHub comments.

Who it’s for

It is built for engineering teams who need comprehensive architectural audits and high-accuracy CI/CD gating where depth and precision are more important than execution speed.

Highlights

• Benchmark Leader: Ranked #1 open-source reviewer on the Martian Code-Review-Bench for golden recall.
• Model Flexible: Supports various models, from cheaper models for routine PRs to frontier models like GLM-5.2 or Opus-class models for high-stakes reviews.
• Evidence-Based: Uses programmatic AST extraction to prune unsubstantiated claims.
• Cost-Effective: Approximately 10x cheaper per review than closed-source commercial tools.