Tool policies
Select explicit decisions by Tool or capability and apply them in trusted runtime composition.
Declare what may run
import type { ToolPolicy } from "@smoothie-framework/runtime/governance";
const policy: ToolPolicy = {
capabilities: {
"documents.read": "auto_approve",
"billing.refund": "require_human",
"admin.delete": "deny",
},
};Apply the policy when composing the Agent runtime or execution context. The three decision values are auto_approve, require_human and deny. For a locked-down run, denyAll blocks Tool calls.
Policies name capabilities, not arbitrary argument-level access rules. The binding must still enforce ownership and resource authorization against trusted request context.
Keep the available Tool set explicit
An Agent run may narrow its declared Tool superset with activeTools and toolChoice; it cannot introduce an undeclared operation. Changing policy must not silently widen a persisted run's execution authority.
For application Tools, follow Define, bind and execute tools. For development-agent grants, use CLI Tools and permissions; those are a separate execution surface.