1.3.23
Decide what the agent may do before it does it
Every session used to run everything without asking, and that was the only way it ran. Now you pick — plan first, ask before each edit, or keep going as before — and when the agent needs you, you answer it on the card that asked.
New
- DesktopChoose how a session runs, from the pill under the composer. Plan explores and proposes before writing anything. Manual asks before every edit and command. Edit automatically works inside the worktree and asks before anything wider. Auto approves what passes a safety check and pauses on anything risky. Bypass permissions is still there — it is just no longer the only option. ⇧Tab cycles them without leaving the composer, and the mode you pick sticks to that conversation.
- DesktopWhen the agent needs permission, you answer on the tool card that asked for it, rather than hunting for a separate prompt. It can also stop and ask you a real question mid-run, and wait for your pick before it carries on.
- MobilePermissions and questions reach the phone too. Answer them from a sheet, and change the mode of a running session without going back to the desktop.
- DesktopA run in a workspace you made tells you when it finishes or fails, with nothing to switch on first. It stays quiet only while you are actually looking at that workspace — an app left open on a long run was the one case guaranteed to tell you nothing. Switch it off per device in Settings under Account.
Improved
- DesktopThe model menu says when it is offering built-in defaults instead of the models your account can actually use, and it goes and fetches the real list when you open a workspace. It used to stop trying shortly after the app started, and never say which list you were looking at.
Fixed
- EngineSigning a workspace in takes effect straight away. It could carry on using the previous credential indefinitely, with nothing to tell you it had.
- EngineOne workspace failing to start a session no longer stops every other workspace starting one.