When is an AI coding CLI suitable for a CI job that creates a draft patch?

When is an AI coding CLI suitable for a CI job that creates a draft patch?

1. Its non-interactive mode has documented failure semantics, bounded runtime, narrow permissions, and machine-checkable output
2. Its interactive output looks correct when copied from a developer terminal
3. The job searches the generated prose for the word done
4. The process has access to production credentials in case they become useful