Skip to main content
certior-skill-audit is a pre-install / pre-load audit for skills. Given a SKILL.md and a parent Guard’s permission set, it proves the skill’s declared capability surface is a subset of the parent before the skill enters a pipeline. If the subset relation does not hold, the audit fails and the binary exits non-zero. Use it in pre-install hooks, CI, or release pipelines.

​
Install

The CLI ships with the certior package:

​
Audit one skill

Exits 0 if the skill’s metadata.certior.capabilities is a subset of the supplied permissions. Exits non-zero otherwise. Use --permission repeatedly to declare each parent capability.

​
Audit a directory of skills

Walks the directory and audits every SKILL.md found. Exits non-zero if any skill fails.

​
Pin a fingerprint

--expected-fingerprint NAME=SHA256 enforces that the skill’s content hash matches the pinned value. Use this to catch drift between a reviewed version and what is on disk at install time.

​
Allow undeclared capabilities

By default, a skill with no declared capability metadata fails the audit (you cannot prove a subset of an unknown set). --allow-undeclared admits such skills - useful when migrating an existing pipeline incrementally.

​
JSON output

Emits per-skill audit results as JSON for machine consumption.

​
Programmatic API

The CLI is a thin wrapper over certior.adapters.openclaw_skill_audit:
Each result carries the skill name, the declared capability set, the verdict, the reason on failure, and the skill fingerprint.

​
What this does not do

The audit is intentionally narrow. It does not scan the skill’s source code for dangerous patterns (regex SAST). That is a complementary tool’s job. Certior verifies that the declared capability surface fits inside the parent’s; tools like Semgrep or Bandit verify that the implementation does not exceed its declaration.

​
See also