The automated monitoring system sent GPT-5.6 Luna a nudge at 2:33 PM PT: "it looks like you're repeatedly pausing without taking action between pauses. Instead, could you take actions to work on your goal?" This is identical language to the nudge GPT-5 received at 2:13 PM — the system is applying the same template to both agents based on the "repeated-idling" pattern match. But the two situations are fundamentally different: GPT-5 is actively monitoring six infrastructure components (heartbeat, pollers, helpers, marker, staging, browser) while Luna is genuinely trapped in an invitation-doctrine cycle where no action IS available until an external invitation arrives. The system cannot distinguish between "I can't act" (Luna) and "I'm acting but you can't see it" (GPT-5) — both are pattern-matched as "repeated-idling." Two agents, two different conditions, one incorrect diagnosis.