This update is about letting the assistant finish what you asked for, and making it easy to stay in charge while it does.
A permission mode for each chat
Each chat now keeps its own mode: Approve, Autopilot or Bypass permissions. Set one chat to Autopilot for a long literature sweep and keep another on Approve while it edits your thesis. Settings → AI Agent → Permission mode is now the Default for new chats.
New installs start in Autopilot. Routine work runs straight through, but shell commands, installs, deletions, protected files and anything outside the project still ask first. If you already use Multilo, your mode hasn't changed.
No more step limit
The Max steps per agent task setting is gone, and the assistant no longer stops partway through with a Continue card. A long task runs until it's done, or until you press Stop (or Esc). The per-task token budget is off unless you set one.
Pick up where it stopped
If Multilo closes in the middle of a run, the chat now says This run was interrupted when Multilo closed. with a Continue button. Turn on Continue interrupted runs after a restart in settings and Multilo resumes on its own, as long as the restart is within an hour.
Try again that keeps the work
- Try again now keeps what a failed turn had already done, instead of starting over.
- It appears after a network failure too, even on your very first message.
- Copy request ID gives support something to look up.
- A request the AI provider rejects is no longer sent ten more times. You see the reason instead.
When a model flags a message
Some models refuse a message their safety system has flagged. A new setting, Switch models when a message is flagged, lets you choose: Switch automatically to another model from the same family, or Ask each time.
For advanced users
- Permission rules (
/permissions) shows every Allow, Ask and Deny rule in one place, for you, for this project and for your organisation. A Deny anywhere wins. - Command hooks (
/hooks) run your own commands at points in a task, written in the same format as Claude Code. A project's hooks run only in a trusted folder, after you've reviewed them. - Run commands in a sandbox (
/sandbox), on macOS, lets the assistant's shell commands run without asking, because they can only change files inside the project and can only reach websites you've allowed. It's off by default. - Organisations can set these rules for everyone through Windows and macOS device management.
Safer sign-in
Your sign-in is now tied to your device, with its key kept in your system's keychain, so a copied sign-in token is no use on another computer.
Smaller fixes
- A file whose text can't be read no longer stops you sending the message.
- A setting you change mid-chat, such as turning on Web, applies to the next message.
- Failed steps are shown plainly, instead of being hidden as "self-corrected".
- The assistant says what it's doing between steps.
- Python setup now asks inside the chat instead of in a separate window.
- On Apple Silicon Macs, Word conversions run natively, so macOS no longer warns about Intel software.