AI Killed Git Commits: So I Stopped Publishing Them

Today, the author shipped contenox 1.0.0 as a single commit into an empty repository, rather than pushing a tag on top of a thousand commits. The 957 commits that got them there are still publicly available in the old repository. The author explains that GitHub's workflow relies on four assumptions: a commit being a unit of human intent, a pull request being a unit of review, history being provenance, and timestamps being labor.

These assumptions no longer hold true for a tree written by agents. The author describes the structure of their repository, with 957 commits over a year, most named Checkpoint, Fix tests, Snapshot WiP. They also discuss how the median file size remained unchanged while the number of files and packages decreased. The author emphasizes that a commit stream is not history, but a log, with no information about human decisions.

They then explain how they shifted their publishing method to release-focused, private monorepos and public release mirrors. This new approach moves the unit of publication to releases, with every release signed and carrying complete trees, allowing for readable diffs. The author also outlines the process for publishing releases, including creating release notes, building binaries, and pushing tags.

They emphasize that this change is driven by the lack of valuable information in intermediate commits, rather than secrecy or technical necessity. The author acknowledges that this approach may not appeal to everyone, but it does make releases more reviewable and focuses on the differences between proprietary and open-source publishing methods.