gouthamuks/ai-pr-reviewer ? reverse-engineered prompt

Build me a Python tool that reviews a single unified diff and tells me what looks risky or broken before I merge it.

I want a command line app that can take a diff file or staged changes, analyze it, and return a short review in text, markdown, or json. It should work offline with a simple rule based review when there’s no API key, and also support an AI powered review when a key is available. Please make it safe against weird prompt injection inside the diff, and keep the diff size capped so it does not try to chew through huge changes.

I also want a deterministic evaluation mode with a small labelled test corpus, plus a way to compare the baseline against recorded AI output so I can see precision and recall. If it helps, add a GitHub Action that runs the review on pull requests and writes the findings to the job summary. Use current docs online if you need to.