Release notes
The agent can use a terminal you can take back
The agent can open an interactive terminal, or ask to use one you already have open. A banner shows while it has control, and any keystroke, Escape, or Take control gives it back and pauses the run.
New
- The agent can open a new terminal for work you may want to watch or join, such as a server, REPL, or debugger. One-shot commands still run on their own.
- A terminal you already have open needs your approval, unless you handed it over from the tab menu. The card names the host, directory, and last command.
- While the agent has the terminal, the tab shows it and a banner offers Take control. Typing, Escape, or that button pauses the run. Control ends when the run ends, and a terminal the agent opened stays open.
- A submitted line is still checked. A local password prompt is refused. On an SSH terminal the agent hands it back instead of typing a secret.
Fixed
- Copy works when the window is not focused, including chat, a terminal selection, images, and error details.
- A provider timeout, rate limit, or error says so on the status row, with how long until the next try. Polling a command that already finished no longer reads as Waiting.
- A sub-agent's task list stays pinned at the top of its inspector.
- Quitting no longer waits forever if teardown stalls. Developer-mode browsers are closed first.