@vcxmug/dsh-vision

Native vision (image recognition) for DeepSeek Harness agents.

DeepSeek's own models have no image input yet. This plugin gives any agent eyes on demand: it registers two tools that talk directly to any Responses API-compatible multimodal model endpoint (no MCP bridge, one composition row):

Tool	Purpose
vision_describe	analyze a local image path or http(s) URL, with an optional question and detail level (auto/low/high, default high)
vision_list_models	list the models served by the configured endpoint

Requirements

• Node 18+ (global fetch).
• A multimodal model endpoint that speaks the OpenAI Responses API, and an API key — either pasted into the Web settings form, or a JSON file under $HOME (a file like {"API_KEY": "sk-..."} configured via apiKeyFile/apiKeyField).

Install (local, no npm publishing)

npm pack ./packages/dsh-vision
npm install -g ./vcxmug-dsh-vision-0.1.0.tgz

Add a pure mount-point row to your agent preset ($DSH_HOME/.agent-presets//agent.cordis.yml, or edit the preset in the Web UI: Settings → Agent presets):

- id: vision
  name: '@vcxmug/dsh-vision'

Configure in the Web UI: Settings → Plugins → dsh-vision (baseUrl, model, apiKey, apiKeyFile, detail, timeouts). Changes apply on the next tool call — no session restart.

Start a NEW session with that preset. The model sees vision_describe and vision_list_models and uses them whenever an image must be understood.

How it works

• Host-side plugin, one fetch call per invocation — no shell, no MCP.
• Local images are base64-encoded inline (data: URL) with a byte-exact encoder; URLs are passed through (the endpoint fetches them server-side).
• Per-call timeout and caller cancellation are honored ().