prc — PR critic

A small, model-agnostic code review tool. Reviews a git ref and emits a structured verdict: approve, request changes, or discuss — with per-issue severity, category, location, and suggested fix.

prc is opinionated about three things most LLM review tools get wrong:

1. Explicit completion verdict. Every review ends with a single top-level decision. No "LGTM with these 14 suggestions."
2. Author response is part of the protocol. Every blocker and major issue expects an accept, refute, or clarify. Single-word refusals are rejected by the protocol, not by social pressure.
3. Model-agnostic. Any OpenAI-compatible endpoint works — hosted (OpenAI, OpenRouter, Together, Groq) or local (vLLM, llama.cpp, LM Studio, ollama).

30-second quickstart

uv sync
export PRC_MODEL=gpt-4o-mini
export PRC_API_KEY=sk-...
uv run prc HEAD

That reviews the most recent commit in the current repo and writes a markdown verdict to stdout.

Pointing at a local model:

export PRC_MODEL=qwen2.5-coder:7b
export PRC_API_BASE=http://localhost:11434/v1
export PRC_API_KEY=ollama
uv run prc HEAD~3..HEAD --format json

Try it without an API key using the built-in mock LLM:

PRC_MOCK=1 uv run prc HEAD

What ships in v0.1.0

• prc <git-ref> [--mode quick] [--format md|json]
• Sequential understand → correctness → verdict pipeline.
• Structured verdict with severity / category / location / suggested fix.
• Author response data types ready for the v0.2 interactive loop.
• Bench-A scaffolding with 10 hand-curated injected-bug fixtures.

See CHANGELOG.md for the not-yet-shipped list.

Honest positioning

prc is not a replacement for CodeRabbit, Cursor Review, or Greptile if what you want is broad, polished, auto-posted PR commentary on a busy shared repo. Those tools have richer integrations, web UIs, and inline-comment ergonomics that this project does not try to match.

prc is a small, local-first tool for one specific case: you want a single structured verdict on a change, and you want the protocol itself to push back on hand-wave answers from either side. It runs against any OpenAI-compatible model, has no hosted dependency, costs nothing beyond the model call, and the entire pipeline fits in a few hundred lines you can read in one sitting.

Configuration

Variable	Purpose	Default
PRC_MODEL	Model identifier passed through to the provider.	(required, unless PRC_MOCK)
PRC_API_KEY	Bearer token. Optional for local servers.	(none)
PRC_API_BASE	OpenAI-compatible base URL.	https://api.openai.com/v1
PRC_MOCK	If 1, use a built-in fixed mock LLM.	(unset)

Development

uv sync
uv run pytest
uv run ruff check .
uv run mypy src/

CI runs the same checks on every push.

License

MIT. See LICENSE.