ghbrain

I built ghbrain to extract activity from merged GitHub pull requests and turn it into data that people can use in different ways. It brings together the issue, commits, changes and review conversations behind a piece of work.

A merged pull request can become a training example, with the author as the agent, commits as implementation steps and reviewers as users giving feedback.

Review conversations could also help distill evals or create an agent skill for code reviews. The project includes an evaluation flow that compares a generated change with accepted work.

I separated collection from transformation so the same data can be reused for different purposes. Built-in extractors produce agent-style transcripts, and applications can provide their own formats. Both stages are available as a library and command-line tool, so developers can use the built-in workflow or integrate it into their own.