Please report security issues privately via GitHub → Security → Report a vulnerability on this repository rather than opening a public issue. We'll acknowledge reports within 72 hours.

Some skills bundle executable code in scripts/. Because users may run these on their own machines, every script goes through review before merge. Maintainers check that scripts:

1. Do exactly what the skill's documentation says, and nothing else.
2. Make no network calls that aren't explicitly disclosed in the SKILL.md.
3. Contain no obfuscated, encoded, or minified logic — code must be readable as-is.
4. Never read, collect, or transmit credentials, tokens, or personal data.
5. Don't write outside the working directory without clearly documenting it.

PRs whose scripts can't be fully understood on review will be rejected regardless of how useful the skill is. Reviewers may ask you to simplify.

Review is a strong filter, not a guarantee. Before installing any skill that contains scripts/, skim the code yourself — the whole library is plain text, so this takes a minute. The validator (python scripts/validate.py) also runs a basic secret/credential scan on every PR, but it is a heuristic, not a substitute for review.

Every generated catalog.json entry publishes:

• source_url: the canonical repository location for the skill.
• license: the license identifier applied to the package.
• security_profile: instructions-only or includes-scripts.
• permissions: declared network, filesystem, and command-execution expectations.
• content_digest: a SHA-256 digest over every relative path and file byte in the skill folder.

The digest changes when instructions, benchmarks, references, scripts, or assets change. Rebuild the catalog with python scripts/build_catalog.py and compare the digest before and after an update when auditing package contents. A matching digest proves byte-for-byte package identity under this repository's digest algorithm; it does not prove that the instructions are safe.

Permission declarations describe what a skill may need to complete its documented workflow. They are validated metadata, not a sandbox or authorization grant. The agent runtime and user remain responsible for approving each tool call.

This policy covers the contents of this repository only. Issues in Claude, Claude Code, or other agent runtimes should be reported to their respective vendors.