Git maintainers push back on AI-written patches and replies

Git project leaders have drawn a hard line against contributions that read as machine-generated with thin human oversight, after a multi-revision series for the show-branch command broke basic behavior and produced incoherent descriptions of its own changes.

Contributor Gatla Vishweshwar Reddy proposed moving show-branch's per-branch reachability tracking out of the shared object flags field into a dedicated commit-slab. That approach matches a broader Git pattern of keeping subsystem state off the global flags word, and it would lift a long-standing hard cap (for years only a few dozen branches) on how many tips the command can show at once. Reviewers treated the goal as reasonable in principle.

What followed on the Git development list was not ordinary polish. Junio C Hamano flagged a performance risk in how parent pruning worked after the move, documentation that still described a fixed 64-branch ceiling after the commit message claimed the limit was gone, comments left behind that contradicted the new code, and finally a floating-point exception on the simplest run: show-branch of two ordinary branches. The crash came from an accidentally deleted slab initialization, leaving a zero size and a divide-by-zero on first lookup. Hamano said he found it hard to trust claims that the patch had been read carefully.

Patrick Steinhardt asked whether reviewers were "merely chatting with an AI," saying the exchange felt like talking to a prompt and was becoming a waste of time. Reddy answered that he is a real person who had used AI to structure replies and would write his own thereafter. Steinhardt said limited AI help is acceptable, but human-facing mail must still feel human. A surge of obviously generated threads, he warned, drives unthinking iteration that does not scale for a volunteer community.

Hamano went further. Later revisions, he wrote, looked like automation sent without oversight, or like severe exhaustion: "We want to see changes designed from the ground up by human minds, not synthetic slop." Assistants may proofread only when the author already understands the domain well enough to accept or reject suggestions, and when the human drafts first. He also set process expectations: answer in your own words, build toward rough consensus, let other timezones speak, and do not ship another version before direction is agreed. Reddy conceded he had relied on the test suite without running the binary by hand, owned the dropped initialization as his failure, and said he would settle open questions in discussion before posting again.

The lasting point is less about show-branch mechanics than about contribution norms as generative tools flood high-traffic open source lists.