Optional Action Execution
Nimi can place an optional fail-closed boundary around external or delegated actions. This boundary is not a prerequisite for Local AI, LocalAgent, Conversation, Memory, Knowledge, or ordinary Runtime readiness.
Safety Order
When an action surface enables the full execution boundary, it uses this order:
discover → dry-run → verify → commit → audit| Stage | Safety outcome |
|---|---|
discover | Return only actions visible to the current scoped principal |
dry-run | Produce a typed proposal without committing side effects |
verify | Check the proposal against current intent and policy |
commit | Apply the authorized action with idempotency protection |
audit | Record the outcome and lineage |
A failed verification never becomes a commit. A commit whose audit outcome is uncertain fails closed. High-risk actions cannot use a path that omits the safeguards required by their owner.
Owner Boundary
The action owner defines the allowed operation, input, output, risk class, and authorization result. Runtime may provide a delegated gateway, approval step, and output firewall without becoming the owner of Realm or App truth. Nimi Home or another host may present approval UI without becoming the action authority.
General Workflow, MCP, A2A, or a public Action Registry are not implied by this boundary. If a future adapter uses it, that adapter remains separately owned and non-blocking.
Reader Scenario: A Verified Write
An optional external-action surface proposes a write.
- Discovery returns the scoped operation.
- Dry-run produces a typed proposal with no side effects.
- Verification confirms current intent and policy.
- Commit applies the action once.
- Audit records the terminal result.
If verification fails at step three, the flow stops before commit and returns a typed refusal.