The audit and its limits

What the three audit tiers check, how severities and the publish gate work, what waivers do, and why a clean report, including one with no findings, does not prove a skill is safe.

By The SKaaS team. Last updated . Markdown version

Every release gets an audit report. The report comes from up to three tiers. The first tier is the same rule engine that runs on your machine as skaas audit, so authors see its findings before they publish.

What are the three tiers?

What are the three tiers?
Tier What runs When Edition and plan
Static rules 59 rules in ten families, written for SKaaS Inline on every publish, in ms Every edition and plan
Deep scan Open-source scanners in a separate container After the commit, in seconds or minutes Cloud Team and above. Community edition through the deepscan Compose profile
Model-assisted review A model that reads the guide and the scripts, with your own model key On request, after the commit Commercial edition. Plan with the llm_review feature

The static tier is the only one that can stop a publish before the release exists. The other two run after the commit and add findings to the same report. See Deep scan and Commercial features.

What do the static rules check?

What do the static rules check?
Family What it covers
FMT Format: a valid name that matches its directory, a description, and well-formed optional fields.
RTE Routing: a description that says when to use the skill, is not too short or long, and defers to skills that exist.
SEC Secrets and hostile text: private keys, cloud and service tokens, hidden or bidirectional Unicode, instruction overrides, download-and-run, links by raw IP address.
PRV Privacy: a home-directory path, a private address, an email address, a host from the organization's denylist.
PRT Portability: an absolute path, a link to a file that is not in the release, a reference that leaves the package, a symbolic link.
SIZ Size: a body of more than 500 lines or about 5,000 tokens, and reference files nested more than one level deep.
DUP Duplicates: the same guide, or a near copy, elsewhere in the organization.
COL Collisions: another skill that competes for the same task.
LIC License: a public skill with no license.
TGT Destination: a key, a description length or a name that a target such as the Claude apps upload does not accept.

Offline, skaas audit runs every rule that needs only the files. It skips the rules that need the service, such as duplicates, collisions, the license check and the host denylist, and lists what it skipped in its JSON output. Add --target claude-apps-upload or --target claude-skills-api to check a destination too.

What are severities and the publish gate?

Findings are Critical, High, Medium, Low or Info. Each repository has a policy for what blocks a publish: Critical only, High and above (the default), or never. An organization sets a default and a repository can override it. A report is blocked when an open finding is at or above the policy, and a blocked publish tells you which finding stopped it. skaas audit --block-on and skaas publish --block-on choose the policy for one run on your machine.

A finding from the model tier never blocks a release on its own. It raises a review for a person to decide.

What does the deep scan do to a release?

It runs after the commit, so a release is never held back while it runs. If a deep finding is at or above the repository policy, the repository's deep scan action decides what happens: notify (the default) tells the administrators, and yank also withdraws the release from new installs. A repository can also treat an incomplete scan as blocking. An incomplete scan is stored as incomplete and is never shown as clean.

Can I waive a finding?

Yes, in the commercial edition. A waiver accepts one risk for a limited time. It needs a reason of at least 12 visible characters and an expiry.

Two checks come before the severity rule. The person needs admin access to the repository that holds the finding, and a key needs the repos:admin scope. An organization owner or administrator has admin access to every repository. A member has it only on a repository where an administrator gave the member the admin grant. A person with write access alone is refused for every severity, and the refusal says that the action needs admin access.

The table below is the default severity rule, which applies after those checks. An organization can set its own. A role in the table cannot waive a finding of that severity on a repository where it lacks admin access.

Can I waive a finding?
Severity Who may waive it (at least) Longest waiver
Critical Owner 7 days
High Admin 30 days
Medium Admin 90 days
Low Member 180 days
Info Member 365 days

When a waiver ends, the finding is open again. See Commercial features.

Does a clean report mean a skill is safe?

No. The auditor finds known patterns. It does not prove that a skill is safe, and no findings does not mean no risk. This holds for every tier: the scanners report what they know, the model can miss what it was not shown, and a skill can ask an assistant to do something harmful in plain words that no rule matches. Read what you install, pin releases, and treat skill content as data until a person has reviewed it. Review remains a human decision.

A report that says a tier did not finish tells you what is missing. Read it as "not examined", never as "no findings".

What does a finding tell me?

Its severity, the rule, the file and line, the text that matched, the release it was found in and a hint on how to fix it. A finding is open, waived or resolved, and it becomes resolved when a later release no longer has it.