dao-awa/dsh-computer-use-native ↗★ 0
dsh-computer-use-native
Windows computer use for DeepSeek Harness built for a vision model: screenshot the real desktop, read the interface from pixels, and click what you see. Drives Win32 through koffi FFI with no native addon and no accessibility tree, and posts input to a window's own queue so the desktop focus is left alone. 适合需要通过截图和像素识别直接控制 Windows 桌面的自动化任务。
설치
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:dao-awa/dsh-computer-use-nativeConfiguration
Set these under the plugin's config in cordis.yml:
| Field | Default | Meaning |
|---|---|---|
maxEdge | 1568 | Longest edge of a returned screenshot, in pixels. Lower spends fewer tokens per capture; higher makes small UI text legible. |
compressionLevel | 6 | PNG effort, 0–9. Measured difference between 6 and 9 is about 1% of file size, so raising it rarely pays. |
typeDelayMs | 0 | Pause between typed characters. Raise it for applications that drop fast input. |
providerName | native-win32 | Name recorded in the computer-use registration. |
Measured on a 2582×1550 browser window, PNG bytes by maxEdge:
maxEdge | Image | Bytes |
|---|---|---|
| 1024 | 1024×615 | 533 KB |
| 1280 | 1280×768 | 762 KB |
| 1568 | 1568×941 | 1.08 MB |
| 1920 | 1920×1153 | 1.49 MB |
A route that normalizes images for its own token budget may reduce these before they reach the model.