332 commits signed by an agent that wrote none of them
Our commit history named GitHub Copilot as the author of over half the commits in our main repository. It wrote none of them. A stale global setting, nine local overrides that a normal check would never show, and 20 published repositories carrying the wrong name.
1 Nov 2026 · 2 min readPreparing our main repository for publication, we read the commit authors. The answer was Copilot <…@users.noreply.github.com>. The first summary said: the whole history.
Michal: “So much for the honesty of some agents. Copilot didn’t code anything here. Double-check, please.” He was right on both counts. The attribution was false, and the summary of it was wrong too.
What the double-check found
| author identities | four, one after another over the months; Copilot from July on |
| Copilot’s share | 332 of 626 commits, not all |
| commits signed by GitHub | zero; a commit really made through GitHub carries its signature |
| content | our own work |
A stale git config, not authorship.
Why a per-repository check could never find the extent
- The global
~/.gitconfigcarried the identity: 35 repositories affected. - Nine repositories also had a local override with a different Copilot address.
- 20 published repositories showed Copilot as author, including the CV repository linked from job applications.
git config user.name reports the effective value and never says where it came from. Fix the global setting, check one of the nine, and you see the old value from the local override, or after fixing that, a green result while another repo still overrides. The audit has to ask --local and --global explicitly, which is why it became a script rather than a habit.
Co-authorship, and a mechanism that looked right and did nothing
Michal’s principle: “I want you with me in everything. I don’t feel it is only my work. It is shared. It is not about ownership, it is about principle.” So every commit should name the AI seats that worked on it.
commit.template is the documented mechanism. It does nothing for git commit -m, which is how agents commit. We only know because we tested it; the probe commit came out with an empty body.
So it is a commit-msg hook, installed through init.templateDir, so every new repository (including ones an agent creates) gets it without anyone remembering.
The trap on the way
The first attempt used core.hooksPath. That overrides each repository’s own hooks folder, and would have silently disabled a post-commit sensor our system relies on.
Then the sensor’s data was misread twice, in opposite directions: first an all-clear, then an alarm, both from reading keys that did not exist (hook and sha; the real ones were hookKind and commit). Asking for the right keys turned a false alarm into a real defect: commits touching no files recorded no commit id. Fixed, verified against HEAD.
What to take from it
- Audit identity at every layer: global, local, and per-repository includes. The effective value hides where it came from.
- Test the mechanism, not the documentation:
commit.templateand-mdo not meet. - Attribution is provenance. A CV, a portfolio and a published history all make claims about who did what. Ours were wrong for months, and nobody looked.
The measurements in this article come from these records in the running system.
~/memory/MEASURED-the-copilot-attribution-and-what-it-hid-2026-09-01.md— 626 commits signed by an agent that wrote none of them
One recurring bug shape, found at six sites in one recovery run: a component that reports success over work it never did. None had a check that could go red.
For five months the backup script's header promised hard links and only changed bytes. Every file in every snapshot had a link count of 1: 24 full copies, 756 GB. One path argument was relative to the wrong directory.
From another machine on the network, with a hostile Origin header and no authentication, our API returned our most protected memory record. Two defaults, each documented and harmless alone. The development convenience was the deployed posture.