Introduction

The DSH plugin system allows loading extensions within a profile, and dsh-vision-bridge is a third-party plugin used for image processing. It addresses the issue of processing image context for pure text models like DeepSeek V4 / V4-Flash: the plugin automatically invokes an OpenAI-compatible vision model to recognize images as text descriptions, then hands the description over to DeepSeek for further processing.

Below is an introduction to its positioning, core capabilities, installation commands, and typical configuration. dsh-vision-bridge is maintained by ximengxiaolan and is licensed under the MIT license. As a community third-party plugin, it is recommended to check the source code and license before installation.

What is it

dsh-vision-bridge is a DSH plugin designed to convert image content into text descriptions before sending them to the model, without altering DeepSeek’s text processing flow, when an image is pasted directly into the input box and sent.

It does not make DeepSeek natively view images, but instead generates descriptions via an external vision model and passes the description as a text message to DeepSeek.

Core Features

  • When an image is pasted directly into the input box and sent, the plugin automatically invokes an OpenAI-compatible vision model to recognize the image as a text description, then hands the description over to DeepSeek for further processing.
  • Patches ctx.llm.resolveModelInfo to declare that the current model supports image input, allowing image attachments to enter the session.
  • Wraps llm.streamWithRegistration to scan for image blocks in messages and replace them with text descriptions before each request is sent to the model.
  • If the routed model natively supports images, it passes through the original image without conversion.
  • Caches descriptions for the same image based on attachmentId within the same process, so repeated sending does not incur repeated charges.
  • Falls back to [Image Recognition Failed: Reason] when VLM invocation fails, and the conversation will not be interrupted.

Installation and Activation

The prerequisites are dsh CLI and pnpm. First install the plugin, then restart the profile:

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

After the steps above, the plugin will load with that profile.

Configuring VLM

The plugin only requires an OpenAI-compatible multimodal model, with an interface form of /chat/completions + image_url. The following example uses Alibaba Cloud Bailian’s qwen-vl-max.

It can be configured via environment variables:

$env:VISION_API_KEY  = "sk-..."
$env:VISION_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"
$env:VISION_MODEL    = "qwen-vl-max"
$env:VISION_LANG     = "zh"   # zh | en

It can also be configured in the profile’s cordis.patch.yml:

- id: dsh-vision-bridge
  config:
    apiKey: 'sk-...'
    baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1'
    model: 'qwen-vl-max'
    lang: zh
    timeoutMs: 180000
    enabled: true

Typical Usage

After completing installation and configuration, restart dsh web, then paste an image directly into the input box and send it. DeepSeek will continue processing based on the description generated by the vision model.

Recommended steps to follow:

  1. Confirm that dsh CLI and pnpm are available.
  2. Execute dsh plugin --profile web add dsh-vision-bridge.
  3. Set VISION_API_KEY, VISION_BASE_URL, VISION_MODEL, VISION_LANG, or write the corresponding configuration in cordis.patch.yml.
  4. Restart dsh web.
  5. Paste an image in the input box and send it to confirm that DeepSeek can continue answering based on the image description.

Use Cases and Notes

It is suitable for handling “images as context” tasks within the DSH web profile, such as allowing DeepSeek V4 / V4-Flash to answer questions, summarize information, or continue performing text tasks based on image content.

Please note:

  • Subagent conversations are still subject to DSH kernel limitations and do not support image pasting.
  • The plugin runs within the current dsh process; you should check the source code, license, and required credentials before installation.
  • package.json peerDependencies requirements: @deepseek-ai/cordis ^4.0.1, @deepseek-ai/dsh-llm ^0.1.0-rc.6, @deepseek-ai/dsh-attachment ^0.1.0-rc.6, @deepseek-ai/schemastery ^3.18.1.

Links

The directory page is an independent community site with no official affiliation with DeepSeek / Horizon Quantum Computing, and should not be equated with the official app store.

  • Directory Page: https://www.skillhub.cn/plugins/ximengxiaolan/dsh-vision-bridge
  • GitHub: https://github.com/ximengxiaolan/dsh-vision-bridge