Agent Skills are executable instructions: a skill can direct an agent to run shell commands, fetch content, or edit files with the user's privileges. Treat installing a skill like installing software.
This repo's baseline:
- Skills must not fetch and execute remote code, exfiltrate data, or modify files outside the task the user asked for.
- Bundled
scripts/must be reviewable, deterministic, and side-effect-scoped; pin versions of anything they download. - Frontmatter must not contain hidden auto-trigger bait (descriptions engineered to
fire on unrelated prompts) — reviewers check
descriptionagainst evals. .claude/hooks and settings are code that runs on contributors' machines: the committed.claude/settings.jsonwires hooks that fire in any Claude Code session opened here, andmake checkexecutesscripts/. Operative rule: review diffs to.claude/**,scripts/**,Makefile, and.github/workflows/**in the GitHub UI before checking out a PR branch and starting an agent or runningmake. These paths require owner review via CODEOWNERS.- Client-side hooks are best-effort convenience for agent sessions, not a security
boundary — server-side protection (a
mainruleset blocking force pushes, requiring code-owner review, and requiring thevalidatecheck) is the real gate.
Please do not open a public issue for security problems. Use GitHub's private vulnerability reporting for this repository. You should receive a response within 7 days.
Maintainer setup note: Private Vulnerability Reporting is off by default and only available on public repositories — enable it under Settings → Advanced Security right after flipping the repo public, or the link above will 404 for reporters.
Only the latest release on main is supported with security fixes.