In the development workflow of DeepSeek Harness (DSH), model outputs often mix multiple languages, making them inconvenient to read and process further. The dsh-reply-language plugin forces the model to maintain a uniform reply language across all outputs (including explanations, summaries, main content, and error messages) through system prompt injection, while preserving the original form of code and technical terms.

What Is This

This is a DSH bundle in the model inference category. It is maintained by CAI-MH and released under the MIT license. Its core function is to inject language constraints during the inference process to ensure consistent model output.

Core Features

  • Enforced unified language: Injects a system prompt that requires the model to use the specified language in explanations, summaries, main content, and even error messages.
  • Preserved technical terms: Code snippets, command-line instructions, file paths, API names, and proper nouns are kept as-is and not translated.
  • Runtime switching: Provides two model tools that allow you to switch or query the currently required reply language at any time during a conversation.

Installation and Enablement

Use the official command to install it:

dsh plugin --profile web add dsh-reply-language

Alternatively, add it to the dependencies in the profile’s package.json and list it under dsh.profile.bundles, then restart dsh web.

Typical Usage

  • Switch language: In a conversation, call the reply_language_set tool with the target language name (for example, 中文, English, 日本語, Français). The setting takes effect immediately.
  • Query current language: Call the reply_language_get tool to view the reply language currently enforced for the model.

Applicable Scenarios and Notes

  • Zero configuration: The plugin requires no configuration file. The default language is Simplified Chinese; it can be changed at runtime through the tools.
  • Host-only: This is a Host-only bundle. It does not provide a client UI; all operations are completed through tool calls.
  • Priority logic: The system prompt registered by the plugin has an order of 900. It overrides language hints in user messages until explicitly changed through the tools.

Summary

By intervening in language control at the inference stage, this bundle can effectively improve output stability in multilingual scenarios. For more details, see the DSH Plugin Directory or the GitHub repository.