26.7.0

Highlights

• Deep Research: a new task type that runs multiple searches in parallel, in bounded waves, and compiles the results into a structured markdown report.
• Tasks can now be cancelled mid-run by the assistant itself, not just by the user, making it easier to stop work that's no longer needed.

Fixes

• Fixed chat cards in scoped conversations incorrectly grabbing focus, and made the sidebar's busy indicator track task state more reliably.
• Fixed a race condition where chat cards could appear before the message they belonged to.
• Fixed turn limits so they apply correctly to Deep Research tasks without affecting other task types.
• Improved reliability of Deep Research activation, including refreshing its state and hiding verify/accept controls when they don't apply.
• Fixed the confirmation dialog for dispatching tasks with unsaved changes, corrected the effort picker's selected-state display, and cleaned up runtime log alignment.

Improvements

• Tightened markdown heading styles for a more compact, readable layout.
• Polished the top bar controls and chat transition animations.
• Refreshed several UI components and improved theme consistency throughout the app.