Introduction

DeepSeek Harness (DSH) adopts an “everything is a plugin” architecture. dsh-vision-plugin is a model inference plugin that adds image understanding capabilities to AI conversations. By configuring a vision model API, the plugin can transcribe user-uploaded image content into text for subsequent LLM processing.

Installation

Before installing, ensure that DSH is installed and meets the version requirement (recommended 0.1.7-rc.2 or a compatible version).

dsh plugin add @lp181818/dsh-vision-plugin

Configuration and Usage

After installation, you need to enable and configure the plugin in DSH settings.

Interface Configuration

  1. Open the DSH Web GUI.
  2. Go to Settings → Vision Model Configuration.
  3. Fill in the following information in the panel:
    - Enable Image Recognition: Main switch. Once enabled, AI can recognize images.
    - Base URL: Base URL of the vision model API (excluding /chat/completions).
    - Model ID: Model identifier used for vision tasks.
    - API Key: The actual key in use (from v1.2.0 onward, this is the sole key source).
    - Key Format: Select the authentication method matching the key provider (openai / anthropic / gemini / azure).
    - Fallback Model (optional): Automatically switches when the primary model encounters 429/5xx rate limiting after retries are exhausted.
    - Retry Count: Number of exponential backoff retries, range 0-10 (default 3).

Configuration File

You can also write the configuration directly in ~/.dsh/settings.yaml. Before saving, the plugin automatically tests connectivity and image input support.

vision-plugin:
  enabled: true
  baseUrl: https://openrouter.ai/api/v1
  modelId: qwen/qwen3.8-27b
  fallbackModelId: qwen/qwen3.8-flash   # Optional: fallback model
  maxRetries: 3                          # Optional: 0-10, default 3
  apiKey: sk-or-v1-...                   # From v1.2.0, key only comes from this field

Core Features

  • OpenAI-compatible requests: The plugin internally sends standard OpenAI-compatible POST {baseUrl}/chat/completions requests, providing broad compatibility.
  • Fault tolerance and retries: Supports exponential backoff retries (0-10 attempts) and fallback model switching, suitable for rate-limiting scenarios on aggregation platforms (such as OpenRouter and Volcano Ark ARK).
  • Multi-format key support: Supports four authentication formats: openai (default), anthropic, gemini, and azure.
  • Automatic testing: When saving the configuration, the plugin automatically tests connectivity and model support using an embedded image.

Compatibility and Notes

  • DSH version: Requires DSH 0.1.7-rc.2 or a compatible version.
  • Key changes: From v1.2.0 onward, the apiKeyEnv configuration item has been removed; the key source is unified to the API Key field on the settings page.
  • Supported services: Supports OpenRouter, OpenAI, Volcano Ark ARK, and any self-hosted gateway providing an OpenAI-compatible interface.
  • License: MIT License.

Summary

dsh-vision-plugin provides stable image understanding capabilities for DSH through a standard OpenAI-compatible interface. It features simple configuration, robust rate-limit retries, and key management, making it suitable for developers who need to process multimodal input.