Jev Windows Agent

It is a Windows UI Automation backend for LLM computer-use agents: click, type and navigate real Windows apps like Notepad, Settings, Spotify or File Explorer from a structured, JEV-driven decision loop instead of screenshot-and-guess. It extends arc-cua, keeping the original runtime, validation and JEV decision policy.

A planner or CUA hands off a bounded subtask; the package observes the desktop with UI Automation plus local OCR, builds the legal action space, makes the JEV decision, runs a freshness guard, executes the UI action and waits for the interface to settle. The target is fewer expensive reasoning calls per completed task, not fewer UI actions.

Import execute_payload and hand it an executor plus a payload like {goal, inputs, verification, constraints, max_actions}; the README's Spotify example returns {"status": "SUBTASK_COMPLETE", "actions_taken": 4}. Any GPT, Claude, Gemini, local or deterministic planner can generate that payload; the planner deliberately lives outside the package.

On Windows 11, `jev plan --yes --request "Open Apple Music and play my liked songs"` had the planner propose a one-step plan, then JEV live-rendered each decision (action, target, confidence, top candidates) while navigating the sidebar and playing Favourite Songs: SUBTASK_COMPLETE after 4 actions, with no frontier-model call in the loop.

arc-cua and its macOS backend are by shhivv, built by Isle (managed desktop environments for computer-use agents). This repository extends arc-cua with a WindowsUIABackend perception and execution backend, OpenRouter transport support and optional confidence gating, while keeping the original runtime and JEV policy untouched.

The README says any GPT, Claude, Gemini, DeepSeek or local planner, or any planner that can produce the subtask payload, via OpenRouter transport. Because the planner is outside the package, the payload format is the integration point.

The README shows a single demo run and no accuracy or reliability benchmarks; the Windows port's implementation notes are in docs/windows-backend-handoff.md. The extensions badge lists Python and both Windows and macOS support, and this repository's contribution is specifically the Windows backend.