dsh-desktop-operator

DeepSeek Harness agents excel at reasoning and multi-step planning, yet they lack the ability to directly drive a Windows desktop. dsh-desktop-operator bundles an adapted Open Computer Use native runtime, the DSH bridge layer, Windows x64/arm64 binaries, and all licensing into a single plugin package. Once installed, a full set of desktop-control MCP tools becomes available inside any Agent Preset with no adjacent source tree or second project to maintain.

Rather than simple mouse-and-keyboard simulation, the plugin replicates the complete Computer Use engineering pipeline. It captures window screenshots via Windows Graphics Capture even when the target window is fully occluded. It queries the UI Automation tree for accessibility-based element discovery and semantic search by name, AutomationId, control type, or class. It executes clicks, drags, text input, key presses, and UIA set_value, then strictly validates outcomes through post-conditions such as focus state, value equality, screenshot change, and window closure. Only verified results are reported as applied; anything uncertain returns unknown. High-risk actions like send, delete, purchase, upload, and install trigger DSH-native confirmation by default. Every state-changing call carries an idempotency key and a redacted audit-log entry that survives process restarts, preventing duplicate dispatch. Session leasing guarantees that only one agent holds the foreground input channel at a time, so concurrent requests fail explicitly instead of silently misfiring. A non-focusing status bar and an orange ring that tracks the real cursor keep the user fully aware at all times that the desktop is being controlled.

This plugin is ideal for developers and power users who need DSH agents to safely read, manipulate, and verify Windows desktop applications in end-to-end workflows such as office automation, cross-application data transfer, and desktop regression testing. Human-in-the-loop confirmation gates and auditable operation logs prevent irreversible actions from running unsupervised. macOS and Linux runtimes are not yet available, and long-term regression on Windows 11 and arm64 real devices is still in progress.