Claude Code's AI-to-AI checks went unread, so replies became 未読

The setup described here is a familiar pattern: rather than trusting a single assistant, the writer had Claude Code hand design questions and finished-work checks to Codex and Gemini. The reasoning is that one AI thinking alone misses things, so a second pair of eyes should catch them. The problem was not the answers themselves but the handling of them. Replies that came back fine and replies that came back empty were treated the same way — left sitting — and a failed request, the writer notes, left no record anywhere that it had been sent.

The writer's point is that an unread reply is equivalent to no check at all, and the more uncomfortable part is that the workflow felt like checking was happening. That gap between the appearance and the reality is what the redesign targets: replies stay flagged as unread, and the conversation cannot close until they have been read. The writer also flags the opposite risk — tightening too hard means a broken mechanism could halt all work, so a line has to be drawn between where to stop and where to let things through.

The specifics — where the three holes sat and what was done to prevent going through the motions of reading — are left to a follow-up piece.