Introduction

In the DeepSeek Harness (DSH) ecosystem, everything is pluggable. In scenarios requiring multimodal content processing, manually switching models in the model bar is both cumbersome and error-prone. dsh-vision-autoswitch addresses this pain point by providing an automated solution that detects the input content to decide whether to route requests to a vision model.

What It Is

This is a model inference plugin maintained by CultOfLuna. When a conversation input contains images, it temporarily routes the current turn to the vision model deepseek-v4-flash-vision-exp for processing. Once the turn ends, it automatically falls back to the model originally selected by the user. It does not alter the user’s model selection logic, so pure-text turns continue to be handled by the high-capability model.

Core Features

  • Automatic switch to a vision model for image inputs: When the input contains images, the current turn is automatically switched to deepseek-v4-flash-vision-exp for understanding.
  • Automatic fallback after the turn ends: Once image processing is complete, it automatically switches back to the user’s originally selected model without manual intervention.
  • Model bar reflects actual behavior: The model actually in use is synchronized in the model bar, so the UI matches the model that is actually served.
  • Preserves the original model selection: Pure-text turns are handled by the high-capability model, and the plugin does not write back to the original model selection.
  • Fallback target follows the user’s selection: The fallback target model is determined by the model currently manually selected by the user; it is not hard-coded.

Installation and Activation

Installation requires manual editing of the configuration file.

  1. Place the plugin files in the profile’s node_modules directory.
  2. Append the following content to the cordis.patch.yml file in that profile:
- insert:
    - id: vision-autoswitch
      name: 'dsh-vision-autoswitch'

After saving, DSH Desktop usually hot-reloads automatically; if plugin code changes are required, restart the application.

Typical Usage

In long conversations, the plugin continuously monitors each input turn. When a user uploads an image, the model bar displays deepseek-v4-flash-vision-exp, and the model begins processing the image content. When the turn ends, the model bar automatically falls back to show the user’s originally selected model (such as Pro or Flash). The entire process requires no manual clicking or switching by the user.

Use Cases and Notes

It is suitable for users who frequently handle image inputs. The plugin runs with the current DSH process permissions, so it is recommended to check the source code and license (MIT) before installation. Installation depends on environment components such as @deepseek-ai/cordis and @deepseek-ai/dsh-llm.

Summary

This plugin lowers the barrier to using multimodal conversations through automatic routing. For more details and updates, see the plugin directory or source repository.
* Plugin Directory
* GitHub Repository