ashishkosana/review-lens

review-lens is a code review tool that uses a large language model to review a git diff, but is built to avoid the common problem of an AI reviewer confidently flagging things that are not actually wrong. It reviews changed code through four separate lenses at once, correctness, security, performance, and test coverage, then runs a second adversarial pass that tries to disprove each finding before showing it to you. Only findings that survive that self-checking step get reported, which trades catching a few less obvious issues for far fewer false alarms. You can try it instantly without an API key using a built in demo mode, which shows example findings like a SQL injection caused by building a query with a raw string, or a missing check for a database row that does not exist. For real use on your own code, you set an Anthropic API key and point the tool at a diff, either your current uncommitted changes, a specific commit range, or a saved diff file. Results can be printed to the terminal, formatted as markdown for a pull request, or posted directly as a GitHub pull request comment through an included GitHub Action, which can also fail a check based on how severe the findings are. Internally the project is built around a small interface for talking to the language model, so most of the code, including diff parsing, ranking findings, and rendering output, is tested without needing an API key or network access at all. Only one small part of the code actually calls the Anthropic API. The tool also includes an evaluation harness with a labeled set of example diffs, some containing known bugs and some clean, used to measure precision and recall and prove that the verification pass genuinely reduces false positives rather than just sounding like it should. The README is clear that the pass or fail check the tool can produce is a convenience, not a security guarantee, since a language model reading untrusted diff text can in theory be talked out of flagging a real problem, so a human should always stay in the loop rather than merging purely because the check passed. The project is licensed under MIT.