Introduction

DeepSeek Harness (DSH) allows text-only models to write code, but in the product design loop of “design draft -> code -> verification,” the model cannot complete the “does it look right?” judgment step because it cannot see images. The dsh-design-qa plugin lends multimodal capabilities to a text-only model through a tool, giving it the “eye” needed to perform product design.

Core Features

  • Any text-only model can read images: Not limited to DeepSeek, any OpenAI-compatible endpoint that supports tool calling can use it.
  • On-demand invocation, no context impact: Image recognition is an independent tool; images do not enter the main model context, and no cost is incurred unless invoked.
  • Includes evaluation benchmarks: Contains 4 sets of fixtures and an evaluation dataset with 23 injected defects, used to verify the effectiveness of model judgments.
  • Automatic yield: When DeepSeek’s official vision capability goes live, the plugin automatically yields, and the model can use the native path.

Installation and Activation

Environment Requirements

  • DeepSeek Harness (DSH) must be installed and configured.
  • The Node.js version must be ^22.19 || >=24.
  • An Alibaba Cloud Bailian (BAILIAN) API Key is required.

Steps

  1. Obtain an API Key
    Visit the Alibaba Cloud Bailian Console to activate and create an API Key.

  2. Install the plugin
    Run the following command in the DSH terminal:

   dsh plugin --profile web add dsh-design-qa
  1. Apply the patch and restart
    Enter the plugin directory, set the environment variable, and execute the patch and restart:
   cd "${DSH_HOME:-$HOME/.dsh}/profiles/web/node_modules/dsh-design-qa"
   export BAILIAN_API_KEY=<你的 key>
   node install.mjs --route-only
   node install.mjs --restart

Note: Refreshing the browser is not considered a restart; the --restart command must be used.

Typical Usage

Basic Conversation

Paste an image directly into the DSH chat box and ask:

“What is this?”

Or send the absolute path or http(s) address of the image directly; the model will automatically call deepseek_vision.

Model Configuration

By default, the Alibaba Cloud Bailian qwen3.8-max model is used. To change it, configure the provider in settings.yaml and override the plugin’s provider and model settings in the profile’s cordis.patch.yml.

Evaluation and Benchmarks

The plugin includes built-in evaluation data with 4 sets of fixtures and 23 injected defects. Benchmark tests show that, in directed probe tests, the qwen3.8-max model performed best (24/24 passed). The tests found that being able to see ≠ being able to judge; if the model does not actively inspect, fabricates, or has unstable conclusions, it cannot be used in a judgment loop.

Notes

  • Requires Tool Calling: The target model must support tool calling.
  • Visibility ≠ judgment capability: The model may be able to see but cannot make effective judgments; it must be used in combination with specific scenarios.
  • HMR is not supported: After installing the plugin modifies the DSH core configuration, a cold start using --restart is required; refreshing the browser has no effect.
  • Cost control: A single image recognition call consumes approximately 2000 input + 400 output tokens. The cost is low, but pay attention to the API Key’s validity period.

Ecosystem Background

DeepSeek Harness (DSH) follows the philosophy of “everything is a plugin.” This plugin is maintained by community developer sunxin-ai and aims to address the visual task shortcomings of text-only models. The related source code and benchmark data can be viewed on GitHub.