DeepSeek Harness (DSH) adopts a plugin-based architecture. During inference, models typically output their chain of thought in their native language (such as Chinese). This forces users to frequently switch languages when reviewing the model’s thinking process. dsh-think-translate is a display-layer translation plugin designed specifically for the DSH Web interface. It translates the chain of thought, task cards, and answers into the target language in real time, without modifying the raw text received by the model.

Core Capabilities

The plugin provides the following core features:

  • Multilingual Support: Supports 8 target languages (Chinese, English, Japanese, Korean, Spanish, French, German, Russian).
  • Real-time Translation: Real-time streaming translation of the chain of thought, task cards, and answers.
  • Zero Context Cost: As a pure display-layer plugin, the model always receives the original text, and translated text does not occupy the context window.
  • Code Safety: Code snippets such as file paths, commands, URLs, and regular expressions are not translated.
  • Smart Chunking: Supports paragraph- and sentence-aware chunking to ensure high translation quality for long texts.
  • Localization Support: Includes a local Ollama provider and supports automatic model downloads.
  • Customization and Discovery: Supports OpenAI- and Anthropic-compatible endpoints and automatically discovers providers in DSH configuration.

Installation and Enablement

Install the plugin using DSH’s command-line tool:

dsh plugin --profile web add dsh-think-translate

After installation is complete, restart the Web process.

Typical Usage

  1. Open Settings, and go to the Think Translation tab.
  2. Select the language to translate into from the target language dropdown menu. The settings panel, thinking lines, and task cards switch synchronously.
  3. Use dragging and checkboxes to manage the translation provider chain. The order in the list is the execution order for translation.

Notes

  • No Configuration Required: The plugin itself requires no additional configuration.
  • Version Compatibility: In the DSH 0.2.0-rc.2 environment, ensure that version 1.2.5 or higher is installed.
    • Version 1.2.6 fixes the issue of duplicated chain-of-thought and answer text in DSH 0.2.
    • Version 1.2.7 optimizes local settings and sets Ollama as the highest-priority provider.
  • Local Model Issue: If the local model list is empty, check the ollama list command or the OLLAMA_MODELS environment variable to confirm that the service is running properly.