Foreword

The DSH plugin ecosystem is organized by the community directory, an independent site with no official affiliation with DeepSeek / High-Flyer. In this context, developers can extend agent capabilities as needed.

For DSH minimal mode, a common pain point is: wanting to maintain the context boundaries and tool surface of minimal mode, but needing visual evidence when processing images, document export results, or charts.

dsh-minimal-vision addresses this issue: it maintains the context boundaries of minimal mode while gaining on-demand visual capabilities. It does not register new chat models, nor does it rewrite the minimal mode’s system prompt; instead, it enters the workflow via hidden context and Bash when needed.

What This Is

dsh-minimal-vision is a DSH plugin maintained by Flora233333, licensed under MIT, requiring node >=20.

It solves the problem of obtaining visual capabilities within DSH minimal mode without registering new chat models or rewriting the minimal mode’s system prompt.

The plugin continues to expose only persistent-bash and str_replace_editor to the model. Attachment paths and invocation instructions are placed in a hidden context that is collapsed by default. The Agent invokes dsh-vision via Bash when visual evidence is needed.

Core Capabilities

Maintaining Minimal Mode Boundaries

  • Does not register new chat models.
  • Does not rewrite the minimal mode’s system prompt.
  • Continues to expose only persistent-bash and str_replace_editor to the model.
  • Places attachment paths and invocation instructions in a hidden context that is collapsed by default.
  • Vision configuration exists independently and does not appear in the chat model list.

Image and Document Processing

The plugin supports PNG, JPEG, WebP, GIF, PDF, DOC, DOCX, PPT, and PPTX; it processes up to 4 images at a time.

Images are automatically corrected for orientation, flattened with transparent backgrounds, and constrained to a long edge of 1600 px by sharp. PDF, DOC, DOCX, PPT, and PPTX are saved as-is, with document processing handled by the Agent.

Documents themselves are not sent directly to the vision interface. Only images specified by the Agent when invoking dsh-vision are sent to the configured vision service.

“Test connection” sends only a minimal plain text request without uploading attachments.

Vision Configuration

Configuration items include:

  • API protocol: OpenAI Completions or Gemini
  • Base URL
  • Model
  • API Key

The API Key is saved by the DSH credential service as VISION_API_KEY.

The CLI reads the following each time:

$DSH_HOME/settings.yaml
$DSH_HOME/.credentials.yaml

The environment variable VISION_API_KEY has higher priority.

Attachments are written to the local cache directory:

$DSH_HOME/cache/dsh-tool-vision/

The hidden context contains only local paths.

Installation and Enablement

The current version does not yet support native Windows PowerShell, only Bash environments; it has been tested on Ubuntu and WSL2. Windows users should install and run DSH and this plugin within WSL2 for now.

First, clone the repository and install dependencies, then add the local plugin to the web profile, and finally start DSH Web:

git clone https://github.com/Flora233333/dsh-minimal-vision.git
cd dsh-minimal-vision
npm install
dsh plugin --profile web add .
dsh web
``

After starting, select "Vision Assist Mode" and complete the configuration in "Settings -> Vision".

# Typical Usage

1. After starting DSH, select "Vision Assist Mode" and fill in the API protocol, Base URL, Model, and API Key in "Settings -> Vision".

2. Drag images or documents into the `vision-assist` mode.

3. The Agent determines on its own whether to read or export documents; when visual evidence is needed, it invokes `dsh-vision` via Bash.

4. PPT/PPTX/PDF need to be exported as images by the Agent first.

You can also invoke manually:
```bash
"${DSH_HOME:-$HOME/.dsh}/bin/dsh-vision" analyze --task custom --instructions "Describe the visible content in the image" -- /path/to/image.png

Built-in tasks include:

  • custom
  • slide_review
  • figure_semantics
  • caption_grounding
  • flowchart_extract

Use Cases and Notes

This plugin is suitable for developers who need to maintain DSH minimal mode boundaries while wanting to view images on demand, process document export results, or invoke built-in vision tasks.

Notes before use:

  • The plugin runs with the permissions of the current dsh process; check the source code and MIT license before installation.
  • Do not commit node_modules/, .playwright-cli/, .env, or any DSH credential files.
  • The VISION_API_KEY environment variable has higher priority; be mindful when configuring multiple sets of credentials.
  • The current version only supports Bash environments; Windows users should use it within WSL2.
  • Attachments are written to the local cache directory, and the hidden context contains only local paths; only images specified by the Agent are sent to the vision service.

Links

  • GitHub: https://github.com/Flora233333/dsh-minimal-vision
  • Directory page reference: https://www.skillhub.cn/plugins/Flora233333/dsh-minimal-vision (This URL comes from the plugin reference and does not appear in the verified content of this document.)