# CodeHawk

> AI-powered pull request review pipeline

CodeHawk is an AI-powered pull request review pipeline that runs in CI. A two-phase architecture separates agent analysis from deterministic posting: Phase 1 reads the diff and writes findings.json; Phase 2 validates, scores with a penalty-based star rating, posts inline comments, and gates the build. It supports Azure DevOps and GitHub, fix verification on re-push, and developer controls for intentional code.

- **Page:** https://siddharthdeshpande.com/opensource/codehawk
- **Repository:** https://github.com/dsiddharth2/codehawk
- **Language:** Python
- **Tech stack:** Python, OpenAI, Docker, Azure DevOps

## Stats

- **2-phase** — review architecture
- **ADO + GH** — VCS supported
- **0–5★** — penalty scoring
- **Docker** — CI-ready image

## What's inside

1. Phase 1: an LLM agent reads the PR diff and writes structured findings
2. Phase 2: validates, scores, posts inline comments, and gates CI
3. Fix verification on re-push — fixed, dismissed, or still-present
4. Developer controls: # cr: intentional, .codereview.md, gate thresholds
5. Model-agnostic: o3, GPT-4, Claude, Gemini, and more
6. Per-run token cost tracking and a dry-run mode that never posts