ONCUE / GUIDE
Permissions & review
Delegation works best when authority is explicit. Know what the agent can do and inspect what it did.
Give the task an appropriate scope
The current source includes permission modes and structured agent questions. Exact prompts and available choices depend on the deployed revision and selected harness.
Agent permission prompts are a cooperative control, not a separate security boundary. Sandbox isolation and operating-system permissions provide that boundary; the agent can read what its sandbox user can read.
Read requests in the context of your task. A request for more access is a decision point, not proof that the access is necessary.
Treat secret access as a separate decision
Model credentials, compute credentials, and repository development secrets have different roles. Only make the required secrets available through the intended configuration flow.
Do not put secret values into prompts, source files, screenshots, or demonstration recordings.
Inspect changes before publishing
Review the actual diff. Compare the changed behavior with the task, examine the recorded checks, and look for files or operations outside the intended scope.
The workflow supports preparing a draft pull request. Publishing must follow the user’s authorization and the repository’s policy. OnCue does not promise automatic merging.
Read verification precisely
Check results are evidence tied to a revision and a set of commands. Passing tests can coexist with incomplete coverage, missing checks, or behavior that needs manual review.
Use your repository’s usual review process before merging or deploying a change.