Introduction

The extension method of DSH is plugin-based: breaking down additional capabilities into installable, enableable plugins. If a DSH session needs supplementary visual capabilities, developers can integrate a vision plugin into DSH instead of writing this logic separately into the application.

Below is an introduction to dsh-vision, maintained by xiaoshihou514.

What is this

dsh-vision is a native, minimal DSH plugin designed to add visual capabilities to DeepSeek Harness.

  • Plugin name: xiaoshihou514/dsh-vision
  • Maintainer: xiaoshihou514
  • License: MIT
  • Goal: To supplement DSH with vision capability

Core Features

According to the README, dsh-vision currently provides the following capabilities:

  1. Adds vision capability.
  2. Supports calling the free Zhipu API.
  3. Supports using the local Qwen-VL 2B visual small model.

This only lists the capabilities explicitly written in the README and does not elaborate on unverified features such as image editing, generative vision, or private model integration.

Installation and Activation

First, install the DSH core:

npm i -g @deepseek-ai/dsh

The README also provides a plugin installation command:

dsh plugin --profile web add "github:xiaoshihou514/dsh-weixin#main"

Please note: This command references dsh-weixin#main, whereas the subject of this article is dsh-vision. Until the maintainer confirms the plugin address or the README is updated, do not execute this command directly as the final installation command for dsh-vision.

Typical Usage

The usage of dsh-vision can be selected based on the visual backend:

  1. Online free API: Using the Zhipu API.
  2. Local model: Using the local Qwen-VL 2B.

The README does not provide specific configuration fields, API parameters, or model paths, so configuration details are not elaborated here.

Suitable Scenarios and Notes

Suitable for these scenarios:

  • You are already running sessions in DSH and wish to supplement visual capabilities.
  • You wish to prioritize using the free Zhipu API or deploy the visual model locally on Qwen-VL 2B.

Confirm the environment before installation:

  • Node engine requirement: ^22.19.0 || >=24
  • The package declares @deepseek-ai/cordis and @deepseek-ai/dsh-* series peer dependencies
  • The plugin runs with the permissions of the current dsh process; before installation, check the source code, dependencies, and license

The license is MIT. If you plan to integrate it into a production environment, it is recommended to verify the visual results in a limited number of sessions first before deciding whether to enable it long-term.

Links

  • Community Directory: https://www.skillhub.cn/plugins/xiaoshihou514/dsh-vision
  • GitHub: https://github.com/xiaoshihou514/dsh-vision

Note: The community directory is an independent site with no official affiliation to DeepSeek / Fenxiang, nor is it an official app store.