Pre-release
See where commits declare AI assistance
Review repository setup and commit declarations with confidence levels and the evidence behind each signal.
Review the evidence
Section titled “Review the evidence”- Repository setup. See which coding tools have configuration or instruction files in the tree.
- Commit history. The CLI reads local Git metadata. In Keryn Web, request a history read for GitHub repositories to see declared involvement per repository and across an organization.
- Confidence. Review which trailer, actor, or other evidence produced a signal. A commit without a declaration stays unclassified. The declared share is a lower bound and does not measure how much code a tool wrote.
Keryn Web shows organization and repository totals. The CLI leaves developer rows out by default. The guide creates a small example repository and shows the recorded result, using synthetic commit metadata.
Measured accuracy
Section titled “Measured accuracy”We measured commit declarations against 6,917 commits in Keryn’s own repositories in November 2026. The expected tool for each commit came from the session log of the coding agent that pushed it, not from the commit text. The 252 negatives were commits written by deploy and metrics scripts. The labels were fixed before the detector ran.
| Tool | Commits | Precision | Recall |
|---|---|---|---|
| Claude Code | 95 | 96.6% (95% CI 90.5–98.8%) | 89.5% (81.7–94.2%) |
| Codex CLI | 6,570 | no declarations | 0.0% |
| Script-generated commits | 252 | 3 false declarations (1.2%) | n/a |
- Codex CLI left no trace. In this setup it added no trailer, author identity or note, so there was nothing to declare. Tools that leave no marker are not counted, which is why every share is a lower bound.
- Claude Code misses were commits written in Claude Code sessions that carried no Claude co-author trailer.
- False declarations came from an older deploy script that added a Claude co-author trailer to generated commits.
- Limits. The sample covers one organization, two agents and one month. It includes no human-only commits and no other tools, and it does not measure pull requests or repository setup files. We will add tools and organizations as labelled data becomes available.