Git rejects non-trivial AI patch, restates LLM contribution ban

Git maintainers have rejected a performance patch that used OpenAI Codex for a substantial implementation change, reaffirming project policy against non-trivial LLM-written contributions.

Arijit Banerjee proposed speeding up promisor link recording in index-pack, the path that builds pack indexes during partial clones. Instead of parsing every reconstructed non-blob object under a shared lock just to enumerate outgoing links, the change would read those links straight from object buffers on worker-local maps and merge them afterward. Banerjee reported a median 15% drop in end-to-end wall-clock time for a blobless clone of linux.git (about 19% in the promisor index-pack phase), with no material change on full clones.

brian m. carlson refused the patch on sight. Project rules, he wrote, allow only trivial machine-generated edits such as spelling fixes or missing include guards. Anything larger must be discarded and rewritten by hand without recourse to the generated code. Analysis with an LLM is fine; submitting substantial LLM-written code is not.

Git maintainer Junio C Hamano backed that reading. The Developer's Certificate of Origin still requires humans who know the origin of their work and can take responsibility for it, and the project will reject contributions that look AI-generated, bloated, or poorly understood. He noted the community wants to tread carefully on licensing and review load, and pointed to practices elsewhere, including LLVM's rejection of extractive contributions whether or not AI is involved.

Banerjee said he was not trying to bypass policy and asked whether the submission could stand as an RFC to preserve the technical result. Carlson’s answer was no: keep the investigation if desired, but ship only a human-authored rewrite.