Commands in your workspace finish during server updates
When the AI runs a command in your workspace while we ship an update, the command now finishes and its output reaches the chat.
User impact: Targeted change
If the AI runs commands in your workspace, a server update no longer cuts a running command off or loses its output. No action needed.
Until now, a server update cut the connection to a running command after twenty seconds. The model got a note that the command might or might not have run, and the output was gone. Usually it ran the command again just to find out.
Now the command keeps running inside your workspace on its own. When the update hands your reply to the new server, that server collects the output, exit code and run time from the workspace and carries on with the reply. Each command runs once.
Two smaller fixes came with it. A command that starts something in the background, such as a dev server, no longer holds up the reply until the time limit. And stopping a reply in the middle of a command now cleans up the processes it left behind. New workspaces also clean up processes that end later on.
Short commands did not get slower. In our measurements they finish about 20 milliseconds sooner than before.