Preface

DeepSeek Harness (DSH) is a local-first web UI for interacting with models such as DeepSeek. During inference, models often output chain-of-thought, task cards, and answers in Chinese or other languages. For developers who need to inspect a model’s reasoning process, reading the original text directly can present a language barrier. dsh-think-translate aims to solve this problem. It translates the “Think” lines, task cards, and answers in the DSH web UI into a user-selected language among eight supported languages, and it operates entirely at the display layer.

Core Features

This plugin focuses on displaying translated content in the user interface without interfering with the model’s execution.

  • Real-time translation: Supports real-time, streaming translation of chain-of-thought, task cards, and answers.
  • Multilingual support: Supports eight languages: Chinese, English, Japanese, Korean, Spanish, French, German, and Russian.
  • Local-first: Defaults to local Ollama as the translation provider, supporting qwen2.5:7b / 14b or a custom model.
  • Privacy preservation: Offline-first; translated text is processed locally only and does not enter the model context.
  • Zero context cost: As a pure display-layer plugin, the model still sees the original text, and translated content does not consume the context window.
  • Code safety: Automatically identifies and skips translation of code-related text such as file paths, commands, URLs, and regular expressions.
  • Intelligent segmentation: Chunks text based on paragraphs and sentences to preserve translation quality.
  • Failover mechanism: Includes a 3x backoff retry mechanism and can automatically switch to Google or Bing when the local model is unavailable.

Installation and Activation

Run the following command in a terminal to install it:

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

After installation is complete, restart the DSH web interface to activate the plugin.

Usage Example

  1. Open Settings and find the Think Translation option.
  2. Select the desired target language (for example, “Japanese”).
  3. Select the translation provider:
    • Local model: On first selection, the plugin automatically downloads the qwen2.5:7b / 14b model.
    • Google / Bing: No extra configuration is required; use it directly.
  4. Send the model a prompt that can trigger thinking, then expand the Think line in the interface to view the translated content.

Notes

  • Display-layer only: Translation happens only in the browser. The original text is retained in the logs and model context and does not affect the model’s input or output.
  • Client and host configuration: Changes to the plugin’s client-side settings require a page refresh to take effect. If you modify host-side configuration, you need to restart the web interface.
  • Model download: On first use of local Ollama, the plugin automatically downloads the model. Ensure your network connection is stable.
  • Proxy support: When using Google or Bing translation, the plugin automatically uses the system proxy or VPN and uses a Node CONNECT tunnel to bypass anti-crawler mechanisms.

Conclusion

dsh-think-translate solves the language barrier during a model’s reasoning process through a pure display-layer approach, while preserving privacy and incurring zero cost. For more information and source code, visit the GitHub repository.