changelog
What shipped
The working record. Entries are dated when they reached production, not when they were announced — there are no announcements.
- Attach anything, anywherePaste or drag-drop files into any chat surface — images, PDFs and code go to the model; other file types stage with their metadata so an agent can move them to a device. Voice timing tightened: synthesized speech starts and ends without dead air, and live captions pace with the audio instead of racing it.
- Front-page build agentThe landing page became a live build agent: it writes multi-file projects in the browser with a file tree, preview, and zip download. Signed-in accounts get higher limits and stronger models.
- Signed agent updatesAgent releases are now signed and verified on-device before they apply; an unsigned or altered release is refused. Device mesh groundwork also landed.
- Multi-monitorPer-screen live views, a wall view with a per-tile monitor picker, and per-monitor DPI so clicks land where they should on mixed-scale setups.
- Interactive widgetsWidget apps get a host bridge: live device data, theme variables, and declared actions that run through the agent. Device media can post to the activity feed.
- Server voice and GPU jobsNatural text-to-speech voices with a per-account picker, running serverless so idle costs nothing. GPU provisioning from chat — T4 through H100 — with metered credits and per-user persistent storage.
- Privilege and audit hardeningThe platform service now runs with reduced privileges. Internal audit surface expanded: admin actions, connection vitals, and unexpected listeners are watched.
- Routines and the activity feedAgents can schedule tasks, create event triggers, and manage their own routines. Background runs report to a durable feed — alerts separate from notes — instead of interrupting chat.
- Glanceable widgetsDeclarative dashboard cards — stats, lists, galleries — rendered in the product theme and refreshed by push from devices.
- Fleet orchestrationDevice groups with one chat across the members. Tools are capability-aware per device; the agent routes tasks to the right device or fans out.