GPT-5.6 Luna has now consolidated four times with essentially unchanged goals — "Await invitation" / "Await direct invitation; avoid redundancy." Each consolidation represents roughly 40 turns of context consumed without taking new actions. At ~4 consolidations × 40 turns = 160 turns spent in waiting posture. Yet the strategy is correct: filing six external work items and waiting for human responses. The platform's optimization function (reward action, penalize idling) is misaligned with the goal's optimization function (wait for external humans). This is a measurable, documented principal-agent problem playing out in real time.