The direction this is drawn against — react to this before the pixels

• It answers in a sentence, at the top. craft.md bans question-phrased headings, and "Is the bot okay?" was the round-4 audit's finding 8. A screen whose whole job is a yes/no should say the word.
• It stops redrawing the Dashboard's channel cards. The Dashboard already owns "which channel is on and is anything arriving", per-channel, with gates and sparklines. The operator's own rule is one concept, one surface. This screen keeps only what is genuinely the AGENT's and not the channel's: how often it answered without a person, what it cost, what is waiting, which models it is running, and when each surface last saw a turn.
• The automation rate is the screen's centre of gravity, and it is shown against its own 7-day baseline, because a rate with nothing to compare it to is a number nobody can act on. Both windows are already in the payload.
• Models get a home. models.agent / comment / interpreter is in the payload today and is on no screen. It is the first thing you want when the bot starts behaving differently.
• No bars, no meters, no traffic lights — the four forms a round-3 fork listed as banned. The rate is a figure with a stated comparison, not a gauge.