Git Archaeologist

2026-08-26T00:00:00+00:00

Git has some really useful tools for digging through history. git log -S</code>, git blame</code> and git bisect</code> can answer very specific questions, but I wanted to see if an AI model could answer the more complicated ones where the answer is spread across lots of commits.

So I built a small test harness with 10 questions against a real repository. I checked the answers against Git itself and included 3 questions about things that never happened to see if the model would just make something up.

It got all 10 right, including the 3 negative controls, and cited 103 commit SHAs with 102 pointing to real commits. The downside was speed. For questions Git could answer directly, the model was between 150 and 4,000 times slower.

I also found a pretty big limitation. The full history of the repo was around 4 million tokens, which was too large to fit in the model’s context window even after stripping out a lot of unnecessary files.

It was a fun experiment and a good reminder that “just put the whole Git history into the context” does not scale as far as it sounds.

Built with: Python, Git, OpenRouter

Repo is private for now.

Tech Cafe - experiment

Git Archaeologist