Agent usage spend estimates no longer show "Unavailable" or a partial total for sessions that ran on the current Claude models (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5). Their rates were missing from the pricing table, so every token they produced was left unpriced.
Cache-read tokens are now priced at each model's published cache-read rate instead of a flat share of the input rate, which keeps long agent sessions from being overestimated.
The divider between a terminal and its code editor can now be dragged to resize them, in both the side-by-side and stacked layouts. The file tree's width can be dragged too, and both sizes are remembered per window.
The code editor now opens the moment you click it, instead of waiting for the terminal's current folder to be looked up first (which took seconds on remote devices).
When the file tree cannot read a folder, it now says why.
Disabling the device that ports are forwarded from now closes the forwards it already had open, instead of only stopping new ones. The device is not reconnected just to clean up.
Terminating a session on a device that cannot be reached no longer leaves its window stuck: a second prompt offers to close the window in the app anyway, while anything still running on the device keeps running.
Devices can be disabled in Settings → Devices. A disabled device stays configured but leaves every device picker, agent usage, sync, MCP checks, port forwarding, remote desktop and the canvas until it is enabled again. Quick Chat, forwarding and remote desktop choices that pointed at it fall back automatically.
Sync, remote desktop and the port manager's Forwarded tab only appear when an enabled remote device exists.
Agent usage no longer shows a red "fetch failed" card when a device cannot reach the usage service.
Cmd-clicking a link in any session (Ctrl-click off macOS) opens it in this machine's default browser, whichever device the session runs on. Plain URLs and agent-printed hyperlinks both work, and a plain click stays a normal terminal click.
