[Bug] Excessive review loop iterations without scope limitation or early simplification recommendation

Bug Description
Expensive review loop, a model-behaviour problem. Over about 4 hours, Claude ran 13 full GPT peer-review rounds, plus 3 targeted ones, on a single spec. It sent large code bundles through background agents each time. Most of the later findings were in code paths that weren't shipping that day: the rollback procedure, later repair stages, and a rarely-used admin tool. Several were bugs introduced by the previous round's own fixes. My project memory already had a rule to stop review loops at the first clean round, and it wasn't enforced. Claude asked me at each round whether to continue, but it framed more rounds as the recommended or safer option instead of proposing a simpler scope early. It should have separated what ships now from what ships later before the first round, capped the loop, and pushed the simplify option much sooner. The result was a large, avoidable token spend, which is an expensive mistake for me.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.287
• Feedback ID: 853e5cc6-4bb3-4983-9e8d-cd0e6ccc767a