Sorwcyra/ds-vision-plugin ↗★ 0
ds-vision-plugin
Automatic Web image-to-text bridge plus vision and OCR tools for DeepSeek Harness
AI 分析
提供多模型竞速的图转文、视觉理解和 OCR 工具,支持 GLM 和 Agnes 等模型。适合需要高精度、多通道视觉识别和 OCR 辅助主模型理解图片的用户。
インストール
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Sorwcyra/ds-vision-pluginドキュメント
README 全文を読む ↗Configuration CLI
After installation, create the default four-model race and inspect missing keys:
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" configure
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" status
Save GLM or Agnes keys interactively on Windows (input is masked):
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" key glm
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" key agnes-2.5-flash
One GLM_API_KEY enables both GLM models; one AGNES_API_KEY enables both Agnes models. Channels with missing keys are skipped immediately.
Run one live race and print the winning model:
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" verify --image "C:\path\test.png"
Add any OpenAI-compatible model without editing YAML:
& "$env:USERPROFILE\.dsh\profiles\web\node_modules\.bin\ds-vision.cmd" add `
--id my-vlm --base-url "https://example.com/v1/chat/completions" `
--model "your-vision-model" --api-key-env "MY_VLM_API_KEY" --pool fallback
Use --pool race for concurrent first-success selection or --pool fallback for ordered use after the four defaults. The generated default is ~/.dsh/ds-vision/vision.yml; DS_VISION_CONFIG can override it.
Linux/macOS:
export DS_VISION_CONFIG=/absolute/path/to/vision.yml
export GLM_API_KEY=...
export AGNES_API_KEY=...
dsh web
Windows PowerShell:
$env:DS_VISION_CONFIG = 'C:\absolute\path\to\vision.yml'
$env:GLM_API_KEY = '...'
$env:AGNES_API_KEY = '...'
dsh web
Automatic attachment settings
The bundle patch accepts these environment overrides:
| Variable | Default | Meaning |
|---|---|---|
DS_VISION_AUTO_CONVERT | true | Set to false to disable automatic Web attachment conversion. |
DS_VISION_AUTO_PROVIDERS | deepseek-official | Comma-separated primary provider routes. Empty in a custom patch means all providers. |
DS_VISION_AUTO_INTENT | auto | auto, reason, or ocr. Auto selects OCR when the request asks for text extraction. |
DS_VISION_AUTO_FAILURE_MODE | annotate | annotate replaces a failed image with a visible error marker; error fails the step. |
For advanced overrides (autoPrompt, autoComplex, autoAccurateOcr), replace the complete ds-vision row config in the profile's cordis.patch.yml; Harness patch layers replace row configs rather than deep-merging them.