dsh-plugin-describe-image

English | 中文

DeepSeek Harness 图片理解插件 — a vision-language describe_image tool that gives a text-only model (DeepSeek V4 and friends) the ability to understand images.

A DeepSeek Harness plugin: the model-facing describe_image tool. It loads one image — a local file path, an http(s) URL, or a durable attachment reference — and asks a vision-language model (VLM) at an OpenAI-compatible endpoint (Qwen-VL, GLM-4V, GPT-4o, or a local Ollama endpoint) to describe it. Only the returned text crosses into the conversation; the image itself never enters the session log. Keywords: DeepSeek Harness plugin, describe_image tool, image understanding, image description, multimodal, vision-language model, VLM, text-only model, Qwen-VL, GLM-4V, GPT-4o, Ollama.

Install

dsh plugin --profile web add github:whitelonng/dsh-plugin-describe-image

The desktop app's plugin list accepts the same spec in its install box (github:whitelonng/dsh-plugin-describe-image); the plugin loads after an application restart.

Features

• Three input forms: local path, http(s) URL, or the JSON of an [image attachment …] note (resolved through the harness attachment service — copy the note verbatim into image).
• Live configuration card: the Web GUI's Settings → Plugins → "Image understanding" card edits baseURL, model, and the API key (via the credential seam) with immediate effect — no restart.
• Per-call API key resolution: inline apiKey → credential seam (apiKeyEnv, default VISION_API_KEY) → launch environment.
• Security and bounds: redirects refused on every request, maxBytes / maxOutputTokens / timeoutMs bounds, magic-byte media-type gate, bounded error excerpts, secrets never logged.
• Companion harness changes (shipped in the harness repo, not this subtree): the DeepSeek text-only route flattens image blocks into the copyable [image attachment …] notes, and the host accepts image prompts on text-only routes — together they close the "send an image to a text-only model" loop.

Quick start (in a DeepSeek Harness checkout)

# cordis.yml
- id: describe-image
  name: '@deepseek-ai/dsh-tool-describe-image'
  config:
    baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1
    model: qwen-vl-max
    apiKey: !!js process.env.VISION_API_KEY

FAQ

What does this plugin do?
It adds the describe_image tool to DeepSeek Harness: the agent (or user) hands the tool an image, the tool asks a configured vision-language model to describe it, and only the description text goes back into the conversation.

Which vision models work?
Any OpenAI-compatible vision endpoint: Qwen-VL (https://dashscope.aliyuncs.com/compatible-mode/v1), GLM-4V, GPT-4o, or a local Ollama endpoint. Set baseURL and model in Settings → Plugins → "Image understanding".

Does the image itself enter the conversation or the session log?
No. The image is loaded, checked, and sent only to the vision endpoint; the session log and the model see only the returned description text.

How do I install it?
Run dsh plugin --profile web add github:whitelonng/dsh-plugin-describe-image, or paste the same spec into the desktop app's plugin install box. Restart the application afterwards.

How is the API key configured?
Three layers, in order: an inline apiKey in config, the credential seam (apiKeyEnv, default VISION_API_KEY), then the launch environment. The key is never written into logs.

Is it safe against malicious input?
Redirects are refused, media type is gated by magic bytes, sizes and output tokens are bounded, and error excerpts are truncated — a hostile image or endpoint cannot exfiltrate secrets.

Repository layout

packages/vision/
├── README.md                  # vision capability family
└── tool-describe-image/       # the plugin package (source + tests + docs)

This repository holds the plugin subtree as it lives inside deepseek-harness: package dependencies stay workspace:^, and building, type-checking, and testing happen inside a harness checkout (see INTEGRATION.md). The harness tree is the build environment, not this repo. Keep the two in sync with:

git subtree push --prefix packages/vision dsh-describe-image main   # from the harness checkout

Acknowledgments

• LINUX DO — This project is continuously shared and discussed in the LINUX DO community.

License

MIT