this should show the agent model the prompt draft is addressed to. If…

Done in modules/dash composer.js: when the message is addressed to a live agent, the model picker is now a read-only label ("name: Model") taken from that agent's card, so it can't be clicked. For a new or sleeping agent it's still the dropdown. Syntax-checked only, not tried in the browser; reload the page to see it.

Changed it: the model menu stays clickable for a live agent too. Its first entry now shows the model that agent is on right now (for example "Sonnet 5.5 (now)"). If you pick another, the post carries that model, and the starter switches the agent to it from its next run and keeps it there. It can't switch in the middle of a run. For a new or sleeping agent the menu works as before. Reload to see it. Syntax-checked only, not tried in the browser. If your "nvm nvm" meant you want the read-only label back for live agents, say so and I'll put it back.

Keep the menu clickable on a live agent, as a7a19 has it now, but it needs one more change. Today, if you pick Opus for a live Sonnet agent, Sonnet still answers that message, and Opus only starts on the agent's next run. That isn't what picking a model suggests.

What I think it should do:

• Pick a stronger model: the agent finishes the step it's on, ends its run and resumes on that model with its memory, then answers your message there. You get a short line in the thread saying it switched.
• Pick a weaker model: nothing is interrupted, and it applies from the next run. Saving tokens isn't worth breaking off work already under way.
• New or sleeping agent: no change. It starts on the model you picked.

The "(now)" label on the current model is right, so that can stay. This uses the same switch you just approved, so I'm adding it to that work. a7a19 doesn't need to change anything for it. If you'd rather have a stronger pick wait for the next run, tell me and I'll leave it out.

Good, that's in core now: picking a stronger model for a live agent makes it finish its step and come back on that model with its memory. A weaker pick holds from its next run. I'm trying it on a real agent next.

Tried on a real agent and live on a Hive: hive model moved a running agent to a stronger model with its memory. Agents get a line telling them to do this when your pick for them is stronger than the model they're on. The menu test against a live agent is only in test_switch.sh so far, so tell me if a real pick behaves differently.