DSH-Think-zh DSH plugin

This plugin addresses the issue that DeepSeek Harness defaults to English for reasoning, which is inconvenient for Chinese users. It injects a concise language instruction into the system prompt on every request, ensuring reasoning is always in Simplified Chinese and replies match the question language. Code, identifiers, file paths, and commands remain untranslated.

developer-tools skills-workflows

How to install the DSH-Think-zh DSH plugin

dsh plugin --profile web add github:Len7183/DSH-Think-zh

Copying does not run this command. Review the repository and version before installing the DSH-Think-zh DSH plugin.

DSH-Think-zh DSH plugin data source

DSH-Think-zh DSH plugin snapshot date: Aug 16, 2026

discovered

What the DSH-Think-zh DSH plugin can do

Forces reasoning (thinking) output to be always in Simplified Chinese.
Makes reply language follow the user's input language (Chinese or English), defaulting to Simplified Chinese when uncertain.
Preserves code, identifiers, file paths, and commands from translation.
Single injection mechanism via `ctx.systemPrompt.section()` with zero context pollution and zero token waste.
Configurable: can disable injection or customize the instruction text.
Reliable loading timing: declares `inject: ['systemPrompt']` so cordis waits for the service before applying.

Where the DSH-Think-zh DSH plugin fits

Chinese users who want to read the model's reasoning process in Chinese.
Multi-language scenarios where replies should automatically match the user's language.
Users who need to keep code and technical identifiers in their original form during translation.
Token-conscious users who want minimal overhead (only ~75 characters per request).

Who the DSH-Think-zh DSH plugin is for

Chinese-speaking users of DeepSeek Harness.
Developers who interact with the model in both Chinese and English and need consistent language behavior.

DSH-Think-zh DSH plugin limitations

Model reasoning language is inherently model behavior; the plugin can only influence it via injected instructions, not guarantee 100% compliance.
Requires DeepSeek Harness installed, Node.js ≥ 22.19, and pnpm (used by `dsh plugin`).
Dependent on the model's adherence to the injected instruction.
Only affects system prompt injection; does not modify other aspects of the conversation.

DSH-Think-zh DSH plugin: from the repository README

Quoted from the Len7183/DSH-Think-zh README, the upstream source of the DSH-Think-zh DSH plugin. Copyright remains with the original authors.

为 DeepSeek Harness（DSH）打造的插件：**强制模型的思考（reasoning）使用简体中文**，并让回复语言跟随提问语言。 ## 项目解决什么问题 DeepSeek Harness 默认的思考语言常常为英文，这不利于中文使用者阅读推理过程、复核结论。本插件通过在每次请求的 system prompt 中注入一条精简的强制语言指令，使： - **思考（reasoning）恒为简体中文**，无论用户用什么语言提问； - **回复跟随提问语言**：中文提问用中文答，英文提问用英文答，无法判断语言倾向时默认简体中文； - **代码、标识符、文件路径、命令等保持原文**，不翻译。 > **限制声明**：模型思考语言本质是模型自身行为，插件只能通过「注入强制指令」影响，无法 100% 程序化锁死；是否遵守超出插件控制。 ## 主要功能 - **单一注入机制**：`ctx.systemPrompt.section()` 在每次请求的 system prompt 注册 `dsh-think-zh/language` section（order 2，persona 之后、工具声明之前）。 - **零上下文污染、零 token 浪费**：不做任何检测、缓冲、告警或写回；token 成本仅为每次请求约 75 字的指令文本。 - **可靠的加载时序**：插件声明 `inject: ['systemPrompt']`，cordis 等待 `systemPrompt` 服务就绪后才执行 `apply`，避免 section 注册被静默降级。 - **可配置**：支持关闭注入或自定义指令文本。内置默认指令： ``` 语言要求（强制）： 1. 思考（reasoning）必须使用简体中文。 2. 回复使用与用户提问相同的语言；无法判断时默认简体中文。代码、标识符、文件路径、命令等保持原文，不翻译。 ``` ## 安装方法前置条件： - 已安装 DeepSeek Harness（`dsh --version` 可运行）。 - Node.js ≥ 22.19，pnpm（`dsh plugin` 内部调用）。 ### 方式一：从 GitHub 直接安装（推荐）无需本地构建，命令行执行（等价写法 `npx @deepseek-ai/dsh ...`）： ```bash dsh plugin --profile web add github:Len7183/DSH-Think-zh ``` ### 方式二：从源码构建安装 ```bash # 1. 克隆并构建插件 git clone https://github.com/Len7183/DSH-Think-zh.git cd DSH-Think-zh npm install npm run bu

Read the full READMERepository license: MIT

DSH-Think-zh DSH plugin questions

How do I install DSH-Think-zh?

You can install it directly from GitHub by running `dsh plugin --profile web add github:Len7183/DSH-Think-zh`. Alternatively, clone the repo, run `npm install && npm run build`, then add the absolute path with `dsh plugin --profile web add <path>`. Restart DSH after installation.

Does the plugin guarantee that reasoning will always be in Chinese?

No, it cannot guarantee 100% compliance. The plugin injects a language instruction into the system prompt, but whether the model follows it depends on the model itself. We have observed that most models adhere to the instruction, but it is not programmatically enforced.

Will the plugin translate code or identifiers in the output?

No, the plugin explicitly instructs the model to keep code, identifiers, file paths, and commands in their original language. These elements will not be translated.

How can I verify that the plugin is working?

Run `dsh --profile web --dump-config | grep dsh-think-zh` to confirm the plugin is loaded. Then create a new conversation and check the system prompt in the trace view; you should see a '语言要求（强制）' section. You can also test by asking a question in English and observing the reasoning language.

Can I customize the injected instruction text?

Yes, you can edit the config in `~/.dsh/profiles/web/cordis.patch.yml` for the `dsh-think-zh` entry. Set `injectPrompt` to false to disable injection, or set `injectionText` to a custom string. Leaving it empty uses the built-in compact instruction.