The durable fix is to make intent an explicit checkpoint instead of relying on Default’s narration: before opening a PR, require a short message with the objective, assumptions, and rejected alternatives. I work on Alook, where people and agents share the same room, so that checkpoint stays visible to the reviewer and the next agent: https://alook.ai. Concise can stay on; the review artifact becomes deliberate.