Introduction

The plugin ecosystem of DeepSeek Harness (DSH) emphasizes modularity and extensibility. When trying to use an automation plugin based on computer-user, developers often encounter a silent loading failure: when the DSH version is higher than 0.1.2, the plugin fails to load, causing all computer_* tools to become unavailable. In addition, the original plugin returns a screenshot path instead of the image itself, which limits the ability of vision models to read the screen directly. computer-user-vision is an unofficial fork designed to fix the issues above and provide DSH on Windows with more precise screen control and image interaction capabilities.

What Is It

This plugin is an unofficial fork of computer-user, designed specifically for DeepSeek Harness (DSH) running on Windows. It fixes the silent loading failure issue on DSH 0.1.2 and later versions, returns screenshots as real images (rather than file paths) to vision models, and provides precise image-to-screen pixel mapping.

Core Features

  • Loading and Compatibility Fixes: Resolves the plugin loading failure issue on DSH 0.1.2+ by changing named imports to namespace imports and adding feature detection, ensuring compatibility with both new and old API versions.
  • Enhanced Visual Interaction: computer_screenshot returns real image data instead of a path, and includes the screen_per_pixel attribute to ensure that vision models can identify screen positions accurately. Vision mode retakes screenshots according to the vision_max_pixels configuration to avoid blurring.
  • UI Automation Support: Supports clicking UI elements directly by ref or name, without needing to manually calculate pixel coordinates. The newly added computer_elements tool can enumerate controls in the current window.
  • Window and Input Control: Provides computer_list_windows and computer_activate_window for precise window management. Input tools (typing, key presses, scrolling, and dragging) support the expect_window parameter, which checks whether the focused window matches before sending input, preventing input chaos.
  • Safety and Awareness: Includes a built-in control indicator that displays a stop button and a global hotkey. Supports PerMonitorV2 DPI awareness, resolving coordinate offset issues on high-resolution displays.

Installation and Enablement

  • Environment Requirements:
    • Operating system: Windows (only win32 is supported).
    • Node.js: Version ^22.19 or >=24 is required.
  • Installation Method:
    • According to the README text, no specific installation command is provided. It is recommended to configure it through the plugin directory page or by directly accessing the GitHub repository. The plugin follows the MIT license.

Typical Usage

  • Take a Screenshot and Display a Grid:
    computer_screenshot({ grid: 100 })
  • Input with Window Check:
    // Only type the text when the current window title contains "Notepad"
    { "text": "hello", "expect_window": "记事本" }

Applicable Scenarios and Notes

This plugin is designed specifically for DeepSeek Harness running on the Windows platform. Because it involves screen capturing and input control, the plugin runs with the permissions of the current DSH process. Before installation, it is recommended to review the source code to confirm that its behavior meets your expectations.

Conclusion

By fixing loading compatibility issues and introducing precise control mechanisms based on UI Automation, computer-user-vision fills the gap in visual automation scenarios for DSH on Windows. For further information or contributions, you can visit the plugin directory or the GitHub repository.