agent-governance

Configure an ActorResolver to establish identity, then rely on DefaultToolAuthorizer, which intersects the actor's roles with each tool's required roles. The check runs twice: once filtering tools offered to the model and again at invoke time.

The governance read-model routes are only mounted when governanceAuthorize is configured. Setting up the store alone is not enough; add a gate such as an ADMIN role check, or explicitly opt into open reads with a predicate returning true.

No. HeaderActorResolver trusts client-sent x-actor-id and x-actor-role headers, so any caller can impersonate any actor. Use it only behind a gateway that strips and re-sets those headers; otherwise use AuthActorResolver over ctx.auth.user.

Replace the authorizer binding with authzToolAuthorizer from @adonis-agora/agent/authz, which consults each tool's ability annotation via the authz Bouncer adapter. Both the offered-tools filter and invoke-time re-check follow the same seam, so they stay consistent.

The dashboard is a pure consumer of /agent/governance/* and refuses to mount without governanceAuthorize configured. Its dashboard.authorize option only gates the SPA shell; the data gate must be set separately for panels to work.

evaluateOwnership deliberately returns 404 for unknown ids so a response never confirms that a record the caller does not own exists. Known ids owned by someone else return 403, while owners and privileged actors get 200.