The problem: I kept making "mega-commits" with unrelated changes because I get in a flow state and then end up with a huge number of changes I forgot to commit.

The solution: I built a tool that analyzes your staged git changes, clusters them by semantic similarity using embeddings + hierarchical clustering, and creates separate commits for each cluster.

How it works:

- Parses diffs into semantic chunks using tree-sitter

- Generates embeddings (OpenAI text-embedding-3-small)

- Clusters with single-linkage hierarchical clustering

- Interactive dendrogram UI to adjust cluster threshold

- Creates one commit per cluster

Currently Mac-only (uses self-written https client with kqueue), but happy to add cross-platform support if there's interest. Also planning local LLM/embedding support.

Looking for feedback!