Introduction

The deepseek-v4-pro and deepseek-v4-flash models in DeepSeek Harness (DSH) currently support text input only. In the DSH environment, pasting a screenshot directly into a Pro chat will trigger an error saying that “image input is not supported.” The dsh-vision-pro-bridge plugin adds visual capabilities to the text-only V4-Pro model through a routing proxy.

Plugin Overview

The plugin is maintained by ShaineDemo and is licensed under the MIT License. It uses DeepSeek-exclusive routing and leverages the existing DEEPSEEK_API_KEY to first transcribe images to text with deepseek-v4-flash-vision-exp, then pass the transcribed text and local path to deepseek-v4-pro for answering.

Core Features

  • Zero third-party dependencies: It relies only on DSH’s core library without introducing external libraries such as OpenAI or Sharp.
  • DeepSeek-exclusive routing: Image processing is fully handled through DeepSeek’s official API, with a single provider.
  • Official image pipeline: It reuses the official adapter logic to ensure consistent behavior.
  • Content-addressed persistence: Images are cached in a specific local directory and passed to the model via stable paths.
  • Focused on the Pro model: It preserves the coding capabilities of V4-Pro while adding visual understanding.

Installation and Activation

Install the plugin using the DSH command line:

dsh plugin --profile web add dsh-vision-pro-bridge

After installation, restart the DSH Web service to load the plugin.

Typical Usage

  1. Select DeepSeek-V4-Pro (Vision Bridge) in the model selector.
  2. Paste or drag and drop a screenshot, then ask your question as usual.
  3. Set it as the default model by editing ~/.dsh/settings.yaml:
agent-default-model:
  provider: deepseek-vision-pro
  model: deepseek-v4-pro-vision

How It Works

The plugin registers a routing named deepseek-vision-pro (with model name deepseek-v4-pro-vision), which declares support for image input. When an image is submitted:
1. The plugin captures the image bytes and saves them to ~/.dsh/vision-pro-bridge/images/.
2. It calls the deepseek-v4-flash-vision-exp model to transcribe the image and generate a text description.
3. It concatenates the transcribed text with the local file path and processes it through the deepseek-v4-pro model.
4. DeepSeek’s raw network endpoint always only receives text and paths, never image data directly.

Use Cases and Notes

  • Use Cases: Suitable for scenarios such as layout analysis, OCR recognition, and UI review.
  • Notes:
    • Lossy transcription: Not suitable for scenarios requiring exact pixel coordinates.
    • Fixed model: The transcription model is fixed to deepseek-v4-flash-vision-exp.
    • No fallback chain: If the transcription call fails, the image becomes a placeholder; the conversation continues but without visual information.
    • Logs and cache: Activity logs are written to ~/.dsh/vision-pro-bridge-activity.l, and image cache is stored under ~/.dsh/vision-pro-bridge/images/.

Summary

This plugin provides developers who value lightweight setups and the DeepSeek ecosystem with a minimal visual solution. It does not rely on complex third-party services and fully reuses DeepSeek’s official pipeline, making it suitable for scenarios that require V4-Pro coding capabilities but occasionally need to inspect interfaces.

GitHub repository | Community directory