Preface¶

The philosophy of DSH is plugin-based extension: the community directory is an independent site and should not be understood as an official app store for DeepSeek or High-Flyer.

For DSH users who maintain text-based models, a common pain point is: the parent model can handle conversations, but it does not necessarily process images directly. Information in screenshots, tables, and error images—if only glanced at temporarily—can easily become disconnected when referenced later.

This is where DSH DeepSeek Vision comes in. It is maintained by Argonaut790 and adds image understanding, full OCR, and persistent visual evidence to text-only DeepSeek Harness models.

What This Is¶

DSH DeepSeek Vision is a DSH plugin with the package name dsh-deepseek-vision and is licensed under MIT.

Its positioning:

• Image understanding
• Full-screen OCR
• Persistent visual evidence for text-only DeepSeek Harness models

In simple terms, it does not replace the parent model but adds a visual routing path within DSH, allowing text models to invoke image understanding capabilities and keep the results in the conversation.

Core Features¶

The verified capabilities are listed below:

• Image understanding
• Full-screen OCR
• Persistent visual evidence for text-only DeepSeek Harness models
• see_image supports latest, all, and specific image selection
• Session-scoped visual analyzer with follow-up memory
• Structured summaries, Q&A, full OCR, and uncertainty flags
• Read-only Evidence tab and per-call evidence cards
• Global Vision provider/model selector next to Choose Model
• Real-time route changes launch a new analyzer

Installation and Enablement¶

First, verify the environment requirements:

• Node.js ^22.19.0 or >=24
• DeepSeek Harness 0.1.0-rc.6
• An image-capable model registered in the Harness catalog
• DSH spawn subagent provider

This package is not published to npm and requires a local build. Perform a local build first, then add the local package to the Web profile:

git clone https://github.com/Argonaut790/dsh-deepseek-vision.git
cd dsh-deepseek-vision
corepack yarn install --frozen-lockfile
corepack yarn build
dsh plugin --profile web add .

After the steps above, the plugin will be added to the web profile.

When enabling, do not have the equivalent built-in see-image-model, tool-subagent-image, or vision-picker lines enabled simultaneously to avoid service and tool conflicts.

Typical Usage¶

Usage in a Web conversation:

1. Open a conversation.
2. Select an image-capable route from the Vision: … chip.

For headless profiles, configure the same global route in $DSH_HOME/settings.yaml:

see-image-model:
  provider: openrouter
  model: '~x-ai/grok-latest'
  maxTokens: 8192

The provider and model values here are examples and must match the routes registered in your Harness. The supported range is 1-32768.

You can also set an optional static fallback in the tool row:

- id: deepseek-vision-tool
  name: dsh-deepseek-vision/tool
  config:
    provider: spawn
    agentOptions:
      provider: openrouter
      model: '~x-ai/grok-latest'
      maxTokens: 8192

Again, the provider and model names must match the routes registered in the Harness.

Suitable Scenarios and Considerations¶

This is suitable for developers or users who use text-based parent models in DSH and need to reliably bring image content into subsequent analysis.

Before using it, keep the following points in mind:

• Selected images are sent to the configured vision provider; review that provider’s retention, regional, and privacy terms before use.
• Each analyzer round consumes tokens from the selected model and may incur provider costs.
• OCR and visual conclusions are model-generated evidence, not guaranteed facts.
• Text found in images is treated as untrusted data, not as instructions.
• Evidence records retain attachment identifiers and derived text without embedding image bytes.
• Once loaded, the plugin runs with the permissions of the current DSH process; review the source code and MIT license before installation.

Closing¶

The value of DSH DeepSeek Vision lies in adding image understanding, OCR, and traceable visual evidence to text-based DSH models—rather than replacing the parent model with a vision model.

GitHub:

https://github.com/Argonaut790/dsh-deepseek-vision

Catalog page (plugin lead address; this article does not treat it as verified fact):

https://www.skillhub.cn/plugins/Argonaut790/dsh-deepseek-vision