Introduction

In the plugin ecosystem of DSH (DeepSeek Harness), a common practice is to decompose extension capabilities into plugins. For the Web chat interface, a specific issue is: users can already paste or upload images to the session, but the interface does not necessarily display the images directly, and text models do not necessarily understand the image content.

visual-review is a plugin designed for this scenario: it enables the DSH Web chat interface to directly display images and allows the model to invoke a visual tool to obtain a Chinese text description of the image.

What is This

visual-review is maintained by wang-bool and is licensed under MIT. It targets the DSH Web scenario, consolidating “image display” and “visual interpretation” into a single plugin.

Core Features

  • Directly render pasted or uploaded images in the DSH Web chat interface, supporting PNG, JPEG, WebP, and GIF.
  • Provides the visual_review tool, which invokes a vision multimodal model to return a Chinese text description of the image, covering text, objects, people, scenes, charts, and other content.
  • Dual engine: prioritizes the OpenAI-compatible multimodal chat/completions API; falls back to local Qwen3-VL-8B when the cloud is not configured.
  • Converts image blocks into text annotations with attachment IDs via request interception and agent/pre-step fallback, enabling text models to work alongside the visual tool.
  • Supports analyzing session images via attachment_id, or local files via image_path.
  • Provides the visual_review_config configuration tool to view or update cloud url, api_key, and model, and to switch engines with one click.

Installation and Environment

Using this plugin requires satisfying the DSH Web profile and Node.js version requirements first:

DSH Web profile
Node.js ≥ 18

If using the local engine, the following dependencies and local models are also required:

Python 3.10+
torch
transformers ≥ 4.57（支持 Qwen3-VL）
Pillow
本地 Qwen3-VL-8B-Instruct 模型

Execute the installation command under the DSH Web profile:

dsh plugin --profile web add visual-review

Configuration

Cloud Engine

Cloud configuration is stored at:

$DSH_HOME/visual_review.config.json

Cloud configuration items include:

url
api_key
model

You can also call visual_review_config in the session to view or update these fields. Leaving url, api_key, and model all empty will revert to the local engine.

This file contains plaintext API keys and should be added to .gitignore, with attention paid to file permission controls.

Local Engine

When the cloud is not configured, the plugin falls back to local Qwen3-VL-8B. In local mode, images are processed only locally and do not generate any external network requests.

Typical Usage

  1. Paste or upload an image in the chat box. The image will render in the bubble, and the model can automatically call visual_review to provide an image description.

  2. Ask the model to analyze an image on disk, for example:

请用 visual_review 分析 /path/to/photo.png
  1. Give the model more specific requests, for example:
读出这张图片里的所有文字
描述这个图表的趋势
  1. When you need to switch engines, call visual_review_config to modify the cloud configuration; leaving all three fields empty reverts to the local engine.

Use Cases and Notes

This plugin is suitable for scenarios using DSH Web where the text model needs to understand images with the help of a visual tool. If you already have an OpenAI-compatible multimodal interface, you can configure the cloud; if you want images not to be sent to external services, you can configure local mode.

Please note:

  • The /vr-image and /api/session.prompt routes only accept requests from the loopback address 127.0.0.1 / ::1; other sources return 403.
  • Cloud API keys are stored in plaintext in $DSH_HOME/visual_review.config.json. Do not commit them to the repository, and pay attention to permission control.
  • Local mode requires local models and dependencies. Before use, ensure Qwen3-VL-8B-Instruct is in place.
  • The plugin runs with the current DSH process and may access plugin configuration, local files, and cloud interfaces. Before installation, check the source code, license, and dependencies.

Related Links

The plugin directory here is an independent site with no official affiliation to DeepSeek / Huanfang.

  • GitHub：https://github.com/wang-bool/visual-review
  • Plugin Directory Page (Resource Clue)：https://www.skillhub.cn/plugins/wang-bool/visual-review