Introduction

The default DeepSeek adapter in DeepSeek Harness (DSH) declares the model as text-only, and image uploads are blocked by MODEL_DOES_NOT_SUPPORT_IMAGES. The sjakdhasdh/dsh-vision plugin aims to address this pain point by adding visual recognition capabilities to DSH. It supports local image paths and web image URLs, calls vision large models through the OpenAI-compatible format, and defaults to Alibaba Cloud Bailian’s qwen3.7-flash model.

Core Features

  • Input support: Supports local absolute image paths (such as C:\a.png) or web image URLs.
  • Vendor decoupling: Uses the OpenAI-compatible format and is not bound to a specific vendor; defaults to Alibaba Cloud Bailian.
  • Configuration priority: Plugin config > environment variables > default values.
  • Zero dependencies: No additional runtime dependencies; uses only the Node built-in fetch.

Installation and Activation

Run the following command in the plugin directory to complete the installation:

dsh plugin --profile web add ./dsh-vision

After running the command, restart DSH. The vision tool can then be used in a new session.

Configuration Methods

The plugin supports configuration via environment variables or the profile patch layer.

Method 1: Environment Variables

Set the following variables in the runtime environment:

export DASHSCOPE_API_KEY=sk-xxx
export VISION_MODEL=qwen3.7-flash-2026-07-15
export DASHSCOPE_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1

Method 2: Profile Patch Layer

Add the configuration in ~/.dsh/profiles/<name>/cordis.patch.yml:

- id: dsh-vision
  config:
    apiKey: sk-xxx
    model: qwen3.7-flash-2026-07-15
    baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1

Usage Instructions

The model automatically calls the vision tool. When calling it, provide the following parameters:

Parameter Required Description
image ✅ Local absolute image path or http(s) URL
prompt ❌ Recognition requirement, defaults to “Please describe the content of this image in detail”

Notes

  • Model limitation: The default DeepSeek adapter in DeepSeek Harness declares the model as text-only, and image uploads are blocked.
  • Direct image pasting: To allow users to paste images directly (image blocks are rendered as [image attachment: sha256:...] markers), a small patch is required for dsh-llm-deepseek (see official documentation or community discussions).
  • License: This plugin is open-sourced under the MIT License.

Ecosystem and Resources

DSH’s philosophy is “everything is a plugin.” This plugin is an independent project provided by community developer sjakdhasdh, with no official affiliation with DeepSeek / High-Flyer.