AI agent: every page, round 3
From his critique of round 2 (777f9cf3). Every screen now answers two questions below its frames: what each field drives in today's code, and how many clicks a real task takes. The frames work: switches toggle, rows open, tabs switch, the moderation slider drags. Run a task and the frame marks the next click and counts them. Nothing is built until the next critique.
The usability metric
Clicks from the AI agent area to a published change, tried once in the Playground where the change alters a reply. Typing is not counted; a drag counts as one click. "Now" is read from today's screens' code (design/agent-model.md section 4, not timed in a browser) or says why the task cannot be done. "Proposed" is measured: the task runner below counts the clicks, and a browser script (core3-tasks.mjs) replays every task and checks the count.