This package runs hooks on every tool call, on your machine. That deserves a way to tell us when something is wrong with it.
Report privately through GitHub's security advisories on this repository: Security → Report a vulnerability. We acknowledge within 3 working days and aim to have a fix or a stated position within 14. Please do not open a public issue first.
| Version | Supported |
|---|---|
| 1.0.x | yes |
| 0.13.x | frozen at its tag; security fixes only until 2027-02-01 |
| < 0.13 | no |
- Releases are published from a tagged workflow using PyPI trusted publishing. No long-lived token on anybody's laptop.
- Build attestations are produced with the wheel, so
ai-eng doctorcan verify that the running wheel is the one that tag produced. - Auto-update is off by design. A version change is a committed diff in
.ai/config.toml, reviewed in a pull request, which is the real control against a compromised release.
The framework is a policy layer. It cannot protect you from credentials that already have every permission on every system reachable from the session. Per-repository tokens and a database user without superuser rights beat every line of code in this repository. That is stated in the design, not hidden in it.