Skip to content

Security: roryeckel/hooknostic

SECURITY.md

Security policy

Reporting a vulnerability

Please do not open a public issue. Report privately through GitHub's security advisory form, which is visible only to the maintainer. If you would rather not use GitHub, or do not have an account, email hooknostic@2labz.com instead.

Include the Hooknostic version, the harness and its exact version (the output of hooknostic doctor covers both; hooknostic --version prints Hooknostic's version alone), and enough detail to reproduce. If you have a generated artifact or a hook payload that demonstrates the problem, attach it.

This is a single-maintainer project, so please allow a reasonable window for a reply before disclosing. You will get an acknowledgement of what was received and whether it is being treated as a vulnerability.

Supported versions

Before 1.0, only the latest published version is supported. Fixes ship in a new release rather than as patches to older ones.

What Hooknostic is responsible for

Hooknostic generates code that runs inside a coding agent's trust boundary, on the developer's machine, with that developer's permissions. It is a compiler for that code — not a sandbox and not a security boundary. That line determines what counts as a vulnerability here.

In scope — report these:

  • A generated artifact that writes outside the target's declared output directory, or that a build stages or commits outside it.
  • A shim that turns a hook's deny into an allow, drops a block, or otherwise reports an effect the dispatcher did not produce. A guard that fails open when it should fail closed is the highest-severity class this project has.
  • Tool input rewriting that produces a command other than the one the hook asked for — a codec that mis-encodes, drops a sibling field, or lands a rewrite under the wrong key.
  • A capability reported as supported when the generated artifact cannot actually deliver it, where a hook author would reasonably rely on it for a security decision.
  • Prototype pollution, injection, or unsafe deserialisation reachable from a harness payload, a tool name, or a config value.
  • A dependency vulnerability that is genuinely reachable through shipped code (the SDK, the CLI, or a generated artifact).

Out of scope — these are working as designed:

  • A hook you wrote doing something dangerous. Hooknostic runs your code; it does not judge it.
  • Anything about a harness's own trust, approval, or sandbox model. Hooknostic never modifies trust state on your behalf — Codex's per-hook trust, for example, still requires your interactive approval, and that is deliberate.
  • A harness bug that Hooknostic accurately reports. If hooknostic inspect states the limitation, the report belongs with that harness's maintainers, and we will help you word it.
  • Installing a generated artifact from a source you do not trust. Installation is a deliberate manual step precisely so that it is your decision.
  • Reports from automated scanners with no demonstrated reachable path.

Hardening you can rely on

  • Builds are reproducible from the version ranges in your config, never from whatever happens to be installed locally.
  • Builds are atomic: nothing is written until every selected target passes.
  • Artifact paths are validated before staging — relative, POSIX, no .., no absolute paths.
  • Routine npm publication uses GitHub Actions Trusted Publishing, without an npm token secret. The first-package bootstrap is performed by the owner with interactive 2FA from verified CI-built tarballs; it is intentionally unattested. See the release runbook.

There aren't any published security advisories