A cloud run can take the discussion's room with it, the way the room works on your own machine. The whole room is on by default in the start form on every device; in a command, end /goal cloud start with room.
Who does what
- The lead does the work, with the cloud's tools: reading and writing files and running commands in the sandbox. Its model is the run's model.
- The reviewers are the room's other seats. When the lead says the work is done, each one reads the objective, the lead's report and the files it wrote, and either agrees or lists what is wrong. Their points go back to the lead, and the work goes on until they agree or the room's review rounds are spent (at most three). The run's result says which seats agreed, or who still had concerns.
- The warden, when your Warden switch or your team asks for it, reads each action that changes something before it happens and can stop it, as it does on your Mac. A stopped action never runs, and the lead is told why.
The run's card shows who leads and reviews, how the latest review went and what the warden checked and stopped.
Which seats go
Only seats on a hosted API model can run in our cloud, through the run's own key and your credit. A seat on a CLI subscription such as Claude Code or Codex never runs there; nor does an agent on your computer, an outside agent, a custom endpoint or a Free Key seat. Those sit the run out, and the start says so by name before anything is sent.
The approval names every model the run may call: the lead's, the reviewers' and the warden's. The run's key can call those and no others, and all of them count against the same token budget and dollar ceiling. With The whole room off, the run uses the one model you name.