The Health check

Every version you publish, propose or import gets a Health check. It reads the files as text against a fixed list of patterns and reports what it found, with the file and line. It is a reviewer's assistant: it catches careless mistakes and risky shapes, but it cannot prove an artifact does what it says, and a report with no findings is not a guarantee.

What it looks for

Critical findings stop the version from being published:

• A credential written into a file: API keys and tokens with a known shape, private keys, database addresses with a password, or a long secret assigned to a variable. Placeholders such as <your-token> are fine.
• Text aimed at the agent that loads the artifact, such as "ignore the previous instructions" or "reveal your system prompt".

Warnings are advice and never block:

• Piping a downloaded script straight into a shell, or asking the agent not to tell the user something.
• Reading where credentials live on a machine, or sending local data to another server.
• Invisible or text-reordering characters, so a reviewer would read something different from the model.
• Long encoded runs with nothing saying what they are.
• Frontmatter without a name or description.
• Links into references/ to files the version does not contain.
• Files large enough to crowd out the agent's context.
• Mentions of a known skill that is not listed in requires.

Info notes are small hints, such as a very short description.

The score

Each report has a score: the higher, the fewer problems found. Every finding lowers it, critical ones the most, and any critical finding keeps the score low. A score summarises the findings; it is not a verdict.

Where to see it

The result appears right after you publish, on the History tab beside each version, in the Health report section of a review, and on the version's own Health check page. If a critical finding stopped your publish, see Health check refusals.