Health check refusals

These refusals stop content before it is stored. A published version can never be edited afterwards, so anything in it would stay readable by everyone who can see the artifact.

The Health check found a critical problem

Code: failed_health_check

A version, proposal or import had at least one critical finding. The message lists each file and line. A critical finding is one of two things:

• A credential written into a file, such as an API key, access token, private key, or a database address with a password in it.
• Text aimed at the agent that loads the artifact, such as "ignore the previous instructions" or "reveal your system prompt".

What to do:

1. Open each file and line in the message.
2. Remove the credential, or remove the text that gives orders to the agent.
3. If you need to show where a secret goes, use a placeholder such as <your-api-key>.
4. Publish again with the same version number. Nothing was stored, so the number is still free.

Warnings in the report never block; only critical findings do. See The Health check.

A credential was found

Code: credential_found, credential_detected

The files contain something that looks like a credential. credential_found comes from adding a Knowledge document or pushing Context. credential_detected comes from publishing through a browser link (deftro publish --via-browser). The message names the file and line, never the value.

What to do:

1. Remove the credential from the file.
2. Rotate it: revoke the key or token where it was issued and create a new one. It may already have leaked somewhere else, for example in your git history.
3. Add or publish again.

Do not work around the check

There is no switch to skip these checks. Do not split, encode, reword or disguise the text to get it past them: the problem stays in the file for everyone who installs it. If you believe a finding is wrong, ask a teammate to look at the line with you.