dsh-deepseek-vision

A DSH-native visual evidence workspace that gives text-only DeepSeek Harness models controlled access to images without changing the parent model.

Unlike provider-pool or CLI interception tools, this plugin keeps DeepSeek Harness in charge of models, attachments, sessions, and UI. It adds:

• see_image with latest, all, and explicit image selection
• one conversation-scoped vision analyst with follow-up memory
• structured summaries, question answers, exhaustive OCR, and uncertainties
• a read-only Evidence tab and per-call evidence cards
• a global Vision: … provider/model picker beside Choose Model
• live route changes; changing the route starts a new analyst

Requirements

• Node.js ^22.19.0 or >=24
• DeepSeek Harness 0.1.0-rc.6
• an image-capable model registered in the Harness catalog
• the DSH spawn subagent provider

The Harness must provide delegated-image prompt admission, model input modalities, the see-image-model settings namespace, and the Web conversation slots. This plugin cannot retrofit those contracts into an older release.

Do not mount this package while equivalent in-tree see-image-model, tool-subagent-image, or vision-picker rows are enabled. Duplicate services and tools will conflict.

Install from GitHub

This project is not published to npm. Build a checkout and add that local package to the Web profile:

git clone https://github.com/Argonaut790/dsh-deepseek-vision.git
cd dsh-deepseek-vision
corepack yarn install --frozen-lockfile
corepack yarn build
dsh plugin --profile web add .

The included cordis.patch.yml mounts the global route service and see_image; its package metadata exposes the Web picker and Evidence UI.

Configure

Open a conversation and select an image-capable route from the Vision: … chip. Models are listed only when the Harness catalog explicitly declares image input.