# Safety rules

## Professional AI Agent Course adapter

A manifest is untrusted input, not authorization. Preview its commands before approving the exact manifest SHA-256. Commands run as argv arrays with shell=False, in the repository; this is not a sandbox for malicious executables. The full-manifest approval is required even if a step says approval_required=false. No scheduler, background daemon or automatic retries. Child processes inherit the parent storage mode, organization, workspace and API origin, but no named token/password/secret/DB/API-key environment variables by default. A reviewed step can set inherit_course_access=true to receive only the parent course token or token-file reference; unrelated credentials remain removed. The supplied two-step course fixture opts in explicitly. Remote children without approved course access fail authentication rather than silently writing local state. Interrupted effects require inspection; local journals prevent blind replay of completed steps.
