dsh-pc-pilot

When an AI model needs to actually move the mouse and type on a Windows desktop, the usual pain points are focus theft, inability to screenshot an occluded window, and re-traversing the entire UIA tree for every single action. dsh-pc-pilot gives DeepSeek Harness one unified desktop-and-browser tool surface so the model can observe and operate in the background while the user keeps working: background-safe input via UIA patterns, target-window WM messages, and Windows Graphics Capture; occlusion-immune screenshots that never copy the covering window; and a persistent CDP WebSocket for browser sessions. Every action is bound to a snapshot_id so the model always hits the right window and element.

Observability is first-class. Each observation assigns a stable element_id plus a monotonically increasing accessibility_revision, letting the model reason about UI changes across turns without re-guessing indices. The batched actions endpoint accepts up to 20 ordered steps, each with an optional when precondition and expect postcondition; execution halts at the first unmet gate with no hidden replay. A lightweight on-screen cursor arrow and a frosted status pill at the top of the screen show where the AI is acting — click-through, non-activating, and auto-hiding after three seconds.

This plugin suits developers who have already wired DSH into multi-window desktop workflows (form filling, data harvesting, cross-app orchestration) or who need rock-solid browser automation. It deliberately ships atomic operations and verification primitives rather than a workflow DSL; application-specific integrations and domain APIs belong in separate plugins that layer on top of the core.