Watch a cloud run

A cloud run streams into a cloud pane, the same kind of pane a local agent runs in. You watch the output, send a follow-up, stop the run, and open the pull request from there. The pane's output is durable, so you can close it and open it again later and the run is still there.

While it prepares

Before any output appears, the pane tells you what it is doing: checking that the run can start, preparing the machine, then starting the agent. If a step takes a while, the pane says so rather than going quiet, so you can tell waiting from stalled.

Once the agent starts, the pane streams the run live.

Send a follow-up

While a run is live, you can message the agent from the pane. If it cannot take input at that moment, the pane tells you and holds your message until it can be delivered.

Stop a run

Use Stop to end a live run. A stopped run keeps whatever output it captured up to that point.

When the run ends

A finished run opens a draft pull request. Open it to review the change. See review agent work for how to read one.

To keep going, use Continue. Coldtea starts a linked run that carries this run's transcript and result, so the next one builds on it. Continuing with the same agent keeps the model and reasoning effort you used; switching agents falls back to that CLI's default. You can also start a fresh run from the task.

If a run does not start, the pane explains why: setup that is still missing, a failure message, or that it was stopped before any output arrived. Fix the setup and start again. See start a cloud run for the prerequisites.