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

Changelog

Week of June 15, 2026

cua-driver

  • Driver health reports — MCP exposes health_report for end-to-end diagnostics across driver startup, permissions, and backend state.
  • Native Wayland preview — Linux can opt into the native Wayland backend, with GTK4 coordinates, focused-widget typing, modal-dialog fixes, and set_config {key,value} persistence.
  • Linux ARM64 buildscua-driver-rs ships prebuilt aarch64 binaries for Linux.
  • macOS TCC and cursor fixes — The driver checks TCC as its own responsible process, ships screen-capture entitlements, stops idle overlay ticks, and shows the agent cursor by default.
  • Windows display fixes — DPI-aware captures scale zoom regions correctly, and idle overlay work stops when there is nothing to draw.
  • Hermes and terminal support — The MCP install surface detects ~/.hermes/skills, separates Hermes setup from the core driver, and falls back to key events when the type target is a terminal.

cua-sandbox

  • LiteLLM security updatecua-agent uses litellm 1.86.2 with fixes for CVE-2026-35030 and related issues.
  • Patched agent dependency — Sandbox agent images pick up the updated LiteLLM build used for model proxy calls.

cua-bench

  • KiCad benchmark datasetcua-bench adds cua-bench-kicad for KiCad and CAD tasks.
  • Registry dump path — The dataset registry export points at the monorepo location used by the benchmark tooling.
  • LiteLLM security updatecua-agent uses litellm 1.86.2 with fixes for CVE-2026-35030 and related issues.