dsh-chinese-thinking

dsh-chinese-thinking

A DeepSeek Harness plugin that makes every agent — the main session and every subagent — reason in Simplified Chinese and reply in Chinese.

It registers a language policy in the host's own system prompt, so it covers every workspace, every new session, and every subagent without relying on a project AGENTS.md.

Honest limitation: a prompt is a soft constraint. It strongly steers the model, but it cannot prove that every token of reasoning is Chinese. For a display-level guarantee, enable the host's Reasoning translation plugin as a second layer.

What it injects

By default, at the end of the system prompt (after the deployment persona suffix):

## 语言要求（dsh-chinese-thinking）
- 内部推理（reasoning / thinking）一律使用简体中文书写，不要用英文思考。
- 面向用户的回复使用简体中文。
- 代码、命令、标识符、文件路径、报错原文保持原样，不要翻译。
- 引用英文资料时保留英文原文，并附中文说明。

Install

Client (Electron): Plugins → + Add plugin, paste dsh-chinese-thinking, install, then Enable now.

CLI profile:

dsh plugin --profile web add dsh-chinese-thinking

Configure

Set options in the profile's cordis.patch.yml:

- id: dsh-chinese-thinking
  name: dsh-chinese-thinking
  config:
    enabled: true
    reasoningLanguage: 简体中文
    replyLanguage: 简体中文
    keepVerbatim: true
    keepCitations: true
    extra: ''
    order: 10300

Option	Default	Meaning
enabled	true	false loads the plugin but injects nothing
reasoningLanguage	简体中文	Thinking language; any language name works
replyLanguage	简体中文	Reply language; empty string leaves it alone
keepVerbatim	true	Keep code, commands, paths, errors verbatim
keepCitations	true	Keep English citations and add Chinese notes
extra	''	One extra custom bullet
order	10300	Position in the system prompt (larger = later)

Why not just AGENTS.md

A project AGENTS.md only applies inside that project, and a subagent or a fresh unrelated session may never read it. This plugin registers a global system-prompt section, so it applies everywhere; the two can coexist.

Self-test

npm run selftest

Drives the plugin with a fake Cordis context: default config registers one section, enabled: false registers none, custom language and options render as expected.

Limitations

• Prompt-level soft constraint: the model may still slip into English, especially in long or English-heavy sessions.
• It does not translate reasoning that already exists; use the host's Reasoning translation plugin for that.
• Only runs while DeepSeek Harness is running.

License

MIT