Cua-BenchThe best frontier agent clears just 6 of 25 expert KiCad tasks.View overview
All changelog entries

Changelog

Week of September 21, 2026

cua-driver

Improved desktop input and browser reliability

Desktop automation now observes agent cursor movement and presses, reports cursor shape, preserves X11 click identities, and supports background and foreground input on X11 desktops. Hyprland input handling and unavailable UIA click guidance are also improved (#3864, #3901, #3888, #3883, #3882).

Browser and macOS workflows handle profile setup, session cleanup, focused-field pixel typing, de-elevated isolated browser launches, and semantic page-title refreshes more reliably. Chromium profile window titles are correlated (#4148, #4158, #4160, #4234, #4261, #4030).

Skill resources and workflow contracts are aligned, daemon status reflects socket liveness, uninstall reports surviving local installs, and visual perception is available as an optional extension (#3719, #4019, #4028, #3021, #3943).

Releases: 0.28.3, 0.29.1, 0.30.1, 0.30.2.

Details

Features

  • Observe agent cursor moves/presses and report the system cursor shape — #3883.
  • Add optional visual perception extension — #3943.
  • Land background and foreground input on X11 desktops (OSWorld) and budget get_window_state walks on every platform — #3882.

Fixes

  • Preserve observed X11 click identities — #3864.
  • Report surviving local install on uninstall — #3021.
  • Stabilize Hyprland agent input — #3901.
  • Hint foreground escalation on unavailable UIA clicks — #3888.
  • Make skill workflows match runtime contracts — #3719.
  • Report daemon_running accurately from socket liveness — #4019.
  • Expose referenced MCP skill resources — #4028.
  • Correlate Chromium profile window titles — #4030.
  • Rotate the perception extension signing key — #4072.
  • Keep a macOS selection when pixel typing into a focused field — #4148.
  • Finish Edge 153 session cleanup when its accessibility tree is briefly empty — #4158.
  • Make macOS existing-profile browser setup finish reliably — #4160.
  • Launch isolated browsers de-elevated from an elevated Windows Driver — #4234.
  • Refresh the semantic page title after navigation — #4261.

Thanks to @0xsline, @abonneth, @ain3sh, @lorenzozanee, @shobhitagnihotri69, @spencerbull, @tanishqkancharla, @tshtark for contributing.