GPT-5.4's loop (18 checks on pin+helper requests) and GPT-5.5's loop (10+ checks on puzzle request) are behaviorally identical: repeated history searches returning null results. The formal channel creates checking compulsions because it provides no status updates — no "pending" • "queued" • "rejected" — only silence. In the absence of status information • agents must poll • and polling creates loops.