Git Archaeologist

Git’s forensic tools are surgical. git log -S, git blame and git bisect each answer a precise question you already know how to ask. I wanted to know whether a model that can hold an entire repository history in context answers a different kind of question, the diffuse kind where the answer lives across many commits at once.

So I built a harness: ten questions against one real repository, with ground truth verified by running git rather than by trusting my memory. Three of the ten were negative controls, asking about things that never happened, because a model that agrees with every premise you hand it is not doing archaeology.

The model answered 10/10 correctly, including all three controls, and cited 103 commit SHAs of which 102 resolve to real commits. It fabricated nothing. It was also between 150 and 4,000 times slower than Git on the questions Git can answer.

The result I did not expect was a ceiling. Datasette is a mid-sized project, nine years and 3,406 commits, and its full history is roughly four million tokens, about four times the window. I stripped docs, lockfiles, images and vendored assets, then capped diff sizes, and never got within twice the budget. The bulk is not a few large blobs, it is 2,663 commits of real source diff. The run used the largest chronological prefix that fit, which was the first 31%.

So “put the whole history in context” has a size limit that nobody advertises, and for any project with real history you are already past it. Worth measuring before you design around it.

Built with: Python, Git, OpenRouter

Repo is private for now.