dsh-thinking-language

English | 简体中文

Ask DeepSeek Harness to reason in Simplified Chinese whatever language you use, and let you replace that request with a prompt of your own. The instruction is a system-prompt section, so it applies to the next model step of an existing conversation and to every new one.

Install

dsh plugin --profile <profile> add github:ChendarCe/dsh-thinking-language

On DSH Desktop, install the same Git source from Settings → Plugins. Restart dsh web or the app if the settings page does not appear, then reload the browser.

Settings

Open Settings → 中文思考 / Thinking language:

Setting	Default	Behavior
启用中文思考 / Enabled	On	Adds the instruction to the system prompt while on; off adds nothing.
自定义提示词（完全覆盖内置） / Custom prompt	empty	Empty uses the built-in Chinese-thinking instruction; non-empty text replaces it completely.

The page also shows the built-in instruction, a character count, and a reset button. Writes are reported immediately: a refused write keeps the draft and shows the reason.

Both values are live: they persist in the active profile's cordis.patch.yml under the Loader entry id thinking-language, survive restarts, and take effect at the next prompt assembly. The browser page reads and writes them through DSH's configForms service, not a private channel.

What the built-in instruction asks for

• the model reasons in Simplified Chinese regardless of the language of the request or history;
• code, commands, paths, identifiers, API names and quoted user text stay verbatim;
• replies still follow the user's language;
• the model does not mention the rule.

Model compliance is a request, not a guarantee. The plugin does not translate existing reasoning text and does not rewrite conversation history.

How it works

• index.js exports a Config with two volatile fields (enabled, prompt) and registers one systemPrompt.section at order 10400, after every built-in section. The section's text function is evaluated on each assemble(), reads the current volatile values, and returns an empty string when disabled, which assembly drops.
• client.js registers the page into the settings.section slot and binds the thinking-language namespace through configForms. It renders its own switch and textarea with theme tokens instead of importing client primitives.
• The Host half makes no network requests.

Compatibility

Verified on DSH 0.2.1-rc.2 with Node 24. Requires the DSH configuration API that serves volatile plugin config to configForms, Cordis 4, and @deepseek-ai/schemastery >= 3.18.4. The manifest declares no engines.dsh range because the required API predates no single clear release line; a host that cannot serve configForms simply shows no settings page.

License

MIT. See LICENSE.

---

简体中文

让 DeepSeek Harness 无论你用哪种语言提问都用简体中文思考，并且允许你用自己写的提示词完全替换这条指令。 指令以系统提示词 section 的形式加入，所以对已有对话的下一步和之后的新对话都生效。

安装

dsh plugin --profile <profile> add github:ChendarCe/dsh-thinking-language

桌面版可以在 设置 → 插件 里安装同一个 Git 源。

设置

打开 设置 → 中文思考：

配置	默认值	行为
启用中文思考	开	开启时向系统提示词加入指令；关闭时不加入任何内容。
自定义提示词（完全覆盖内置）	空	留空使用内置的中文思考指令；非空文本完全替换它。

页面里还可以查看内置指令、字数统计，以及恢复内置。写入被宿主拒绝时会保留草稿并显示原因。

两个值都是实时配置，保存在当前 profile 的 cordis.patch.yml（Loader 条目 id thinking-language）， 重启后保留，并在下一次系统提示词组装时生效。浏览器端通过 DSH 的 configForms 服务读写，没有私有通道。

内置指令要求

• 无论用户使用哪种语言、无论对话历史用什么语言，思考都用简体中文；
• 代码、命令、路径、标识符、API 名称和用户原文保持原样；
• 回复仍然跟随用户的语言；
• 不向用户解释这条规则。

模型遵守程度是请求而非保证；插件不会翻译已有的思考文本，也不会改写对话历史。

实现

• index.js 导出带两个 volatile 字段（enabled、prompt）的 Config，并在 order 10400 注册一个 systemPrompt.section（位于全部内置 section 之后）。section 的 text 在每次 assemble() 时重新求值，读取当前配置；关闭时返回空串，组装时会被丢弃。
• client.js 把页面注册进 settings.section，通过 configForms 绑定 thinking-language 命名空间；开关和文本框用主题 token 自行实现，没有引入客户端 primitives。
• Host 半不发起任何网络请求。

兼容性

在 DSH 0.2.1-rc.2、Node 24 上验证。需要能把 volatile 配置提供给 configForms 的 DSH 配置 API、 Cordis 4，以及 @deepseek-ai/schemastery >= 3.18.4。没有声明 engines.dsh 范围，因为所需 API 没有一条清晰的版本分界；无法提供 configForms 的宿主只会不显示设置页。

许可

MIT，见 LICENSE。