Core developers
Approvals, control and audit
A client requests an operation. The Host authenticates the caller, checks policy and prepares the exact effect for review. Keeping these steps distinct lets a reader inspect a run without gaining permission to start or cancel one.
Where the backend participates
The account backend distributes scoped information such as published organization catalogs and registered Host advertisements. Its readers authenticate and check current membership or device grants. When a client selects a published role, the destination Core still resolves the live workspace, provider session and local authority. A catalog revision identifies the configuration a reviewer saw. The original operation receipt identifies the effect the Host recorded.
What a reviewer authorizes
A prepared operation identifies its registered operation kind and version, immutable input, authenticated principal, target Host instance and applicable resource generation. For a project or session action, the responsible owner checks that workspace or session against current state. The review covers that specific request. Core's shared decision vault issues a single-use reference and the owner checks it again at admission.
For a remote turn, approval also binds the source and destination identities and the selected conversation. For a local team, the review names members, roles and permitted recipients. A directory entry, account membership or policy setting alone cannot issue either approval. Revoking a grant closes future admissions under that grant. Controlling an existing resource requires an independently authorized operation on the owned live handle.
After a lost reply
Core records intent and consumed approval evidence before dispatching an external effect. If the acknowledgement is lost, keep the original operation ID and query its owner. A provider's original receipt can identify a turn that was accepted before the connection failed. A terminal result requires evidence that the turn finished. An unknown outcome stays unknown across restart until the responsible owner can establish more.
For a managed launch or control operation, read its durable audit record locally:
cleetus-host inspect operation "$OPERATION_ID"
The record includes the launch state, hold and revision, plus the latest control action and state where recorded. Compare instanceID with originalInstanceID when shown. A previous instance's receipt remains useful history. It cannot be used to adopt a PID. Likewise, a recorded leader exit alone gives no proof that all descendants settled.
Task completion and provider setup have their own original-receipt readers. Use the reader for the operation you submitted. Task completion example · OS enforcement evidence · Account security