Plans, approvals and audit
The human approves a semantic plan, not the model’s prose and not every internal command vops needs to execute.
Immutable plans
A plan records:
- objective and target;
- ordered semantic steps;
- normalized non-secret inputs;
- expected effects;
- risk and reversibility information;
- success criteria;
- rollback limits;
- a content hash.
Approval binds to that content. A changed manifest, image, target or material input creates a different plan and requires a new decision.
Secret values are supplied at execution through the credential boundary and do not become reviewable plan text.
Review pending requests
vops agent approvals list --status pendingBefore deciding, check:
- the objective still matches the user’s request;
- repository, target and environment are correct;
- every expected effect is acceptable;
- the plan says what verification will prove;
- rollback is actually supported for the affected capability;
- no secret value appears in the plan.
Then decide as the local human:
vops agent approvals approve <approval-id> \ --reason "Staging target and effects reviewed"
vops agent approvals deny <approval-id> \ --reason "The plan targets the wrong host"The agent-facing MCP surface does not expose approval.
Follow operations
vops agent operations listvops agent operations list --session <session-id>vops agent operations show <operation-id>vops agent operations cancel <operation-id>An operation moves through explicit states such as queued, running and verifying before reaching a terminal result.
Cancellation is a request. A provider call or host action may have crossed a non-interruptible point. Read the operation result to learn which effects completed.
Verification is part of success
Execution success means the requested provider or host action returned. Application success additionally needs evidence such as service state, health, DNS and an external HTTPS request where relevant.
An agent should not report “deployed” from an apply result alone when a verify capability exists.
Audit
The local audit records session, policy, plan, approval, operation and revocation events in sequence. Stored and returned data is redacted before it reaches the event stream.
The audit answers:
- who requested the operation;
- under which objective and session;
- which policy decision occurred;
- which plan was approved;
- what execution and verification reported;
- when authority was paused or revoked.
It is operational evidence, not a backup of the application or provider.
Rollback language
Rollback is capability-specific. Application rollback is unavailable until vops has a durable previous-revision restore primitive. Do not promise rollback because a plan marks an operation “reversible: conditional.”