ruflo 3.63.0 adds controls for Claude Managed Agents that remain disabled until explicitly enabled. Dynamic workflows require a spending ceiling, while stopping sessions requires separate host permission and an allowed session. Provider completion stays labelled unverified, and unsigned learning outcomes remain quarantined. The maintainer acknowledges that live operation against Anthropic’s beta service has not been exercised.
ruflo 3.63.0 adds four MCP tools covering discovery, session control, learning and interruption for Claude Managed Agents. The functions require explicit opt-in, and dynamic workflows cannot be enabled without a budgetCents spending ceiling. The console shows budgets, blockers and worker counts without those panels themselves sending requests to the provider. The project released these controls on 25 October.[1]
Interruption or termination needs a host-side control switch as well as an explicitly allowed session. The provider’s current state is read again before action. A provider’s completion response stays labelled unverified. Learning accepts outcomes signed against the host’s configured public key, quarantining unsigned outcomes; routing remains unchanged even when the function proposes strategies or compares paired holdouts.[1]
The maintainer acknowledges that live execution against Anthropic’s beta service has not been exercised. The signing key shares the user account that runs verification, so the maintainer restricts checks to host-authored ones. Evidence binding is partial, crashed processes may leave locks needing manual removal, and sessions already in an error state cannot be controlled. The account of this release comes from its maintainer alone.[1]
