About 4% of October's commits in 30 busy repos carry an AI stamp
Out of roughly 7,000 commits pushed to 30 busy open-source repos in October, about 4% carry a stamp saying an AI coding tool helped write them, and almost every one of those stamps comes from Claude Code. I picked the r
Out of roughly 7,000 commits pushed to 30 busy open-source repos in October, about 4% carry a stamp saying an AI coding tool helped write them, and almost every one of those stamps comes from Claude Code.
I picked the repos by hand from well-known, very active projects, and a third of them are AI or agent tools. On each default branch I went through every October commit apart from merges and counted the ones whose message has a trailer like Co-Authored-By: Claude or Assisted-by: .... The counting was done by ai-code-sbom (https://github.com/TrueHorison/ai-code-sbom), which I built after reading AT&T's supplier terms, which ask for an SBOM naming the AI tool behind each release. Per repo it was roughly:
git clone --no-checkout --shallow-since=2026-09-25 <repo>
python ai_code_sbom.py --repo <repo> --since <last September commit> --until <last October commit> --format md
My first list had Textualize/rich and elysia on it, and neither had a single commit in October, so they got swapped out, along with kamal, which didn't have enough history to cover the month.
Here's where the stamps turned up, plus a few of the zeros:
| Repo | Commits | AI-stamped | Share |
|---|---|---|---|
| sqlc-dev/sqlc | 12 | 6 | 50.0% |
| crewAIInc/crewAI | 193 | 76 | 39.4% |
| compiler-explorer/compiler-explorer | 174 | 61 | 35.1% |
| modelcontextprotocol/typescript-sdk | 38 | 6 | 15.8% |
| biomejs/biome | 407 | 48 | 11.8% |
| conventional-changelog/commitlint | 54 | 4 | 7.4% |
| pydantic/pydantic-ai | 1357 | 48 | 3.5% |
| bytecodealliance/wasmtime | 375 | 13 | 3.5% |
| openai/codex | 1282 | 0 | 0.0% |
| anthropics/claude-code | 56 | 0 | 0.0% |
| sveltejs/svelte | 40 | 0 | 0.0% |
| django/django | 117 | 0 | 0.0% |
| tokio-rs/tokio | 129 | 0 | 0.0% |
| All 30 | 7049 | 275 | 3.9% |
crewAI was close to 40%, and sqlc only beats it because it had 12 commits all month. The top three repos between them hold about two-thirds of every stamp I found, and if you take them out, the other 27 repos come to well under 2%.
Half the repos had no stamps at all. Svelte had none in 40 commits and django none in over a hundred, tokio was the same, and openai/codex pushed close to 1,300 commits in October without one. anthropics/claude-code itself had 56 commits and not one Claude Code stamp, which I'd guess (without checking) is squash merges.
Nearly all of it is Claude Code, 264 of the 275 stamped commits, and the rest is a handful of Cursor, GitHub's Copilot coding agent, and Assisted-by lines naming GPT or Gemini. A lot of that is defaults: Claude Code adds its co-author line unless you turn it off, so it was always going to win a count of stamps.
So 4% is a floor. A stamp only exists if the tool writes one and the person leaves it in, and plenty of tools and people don't. Squash merges and rebases that rewrite the message drop the trailers as well, so a repo that squashes pull requests into one clean message can show zero even when the branch was full of AI commits. The sample isn't random either, and it probably leans towards projects that are relaxed about AI tools.
Projects should keep the stamps
It's one line per commit, and it's the cheapest record of AI use there is that doesn't rely on someone squinting at the code and guessing. So I think projects should keep these lines, and the ones that squash-merge should carry the trailers into the squash message instead of letting them vanish.
Written with AI help.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.