Introduction¶

It is common to ask questions in Chinese and receive answers in Chinese within DeepSeek Harness (DSH). However, if the model displays the entire intermediate reasoning process in English, you are left staring at the screen and mentally translating, making it difficult to determine whether it is checking files or planning in the right direction. Waiting minutes for the result only to discover a wrong first step wastes time and tokens. The official experimental package allows for post-hoc machine translation of the Chain of Thought into Chinese, but this is a display-layer solution. Below, we introduce the community plugin zhong-wen-thinks: it injects native language instructions at the beginning of the system prompt to encourage the model to think in Simplified Chinese from the reasoning phase onward, facilitating real-time reading and allowing for interruptions and corrections during the thought process.

What is this¶

zhong-wen-thinks is a Cordis plugin for DSH, maintained by yhy854, categorized as Model Reasoning in the SkillHub directory. The current npm package version is 0.1.0, licensed under MIT. The plugin does not introduce third-party npm dependencies; instead, it injects a fixed Chinese thinking instruction via ctx.systemPrompt without modifying the DSH kernel.

Core Features¶

Simplified Chinese Reasoning Display¶

Once installed and enabled, processes such as the Chain of Thought (CoT), task planning, analysis before and after tool calls, and self-correction should ideally be output in Simplified Chinese; code, commands, paths, variable names, and original error messages should remain in English to prevent command failures due to mistranslation.

System Prompt -900 Injection Slot¶

DSH concatenates system prompts in order of numbers. This plugin inserts at -900, following HARNESS_IDENTITY (-1000) and preceding other rules. The README indicates this is an earlier slot accessible to external plugins, used to increase the priority of the Chinese thinking instruction.

Preventing English Drift¶

The instructions are written in the TEXT constant in index.js, making the logic visible externally: it requires that Chinese thinking takes priority over other language habits and reminds the model to switch back to Chinese thinking after reading English logs or error messages.

Integration with Official Plugin System¶

The plugin follows the official DSH plugin management system: it can be enabled or disabled via the panel without needing to modify configuration files separately. When disabled, the aforementioned instructions are no longer injected.

Installation and Usage¶

Install via the plugin panel in the DSH desktop app; do not rely on CLI concatenation commands not mentioned in the documentation.

1. Open the DeepSeek Harness desktop app and click “Plugins” on the left.
2. Click “Add Plugin” in the top right corner.
3. Enter any one of the lines below in the input box to confirm installation.

NPM Package Name:

zhong-wen-thinks

GitHub URL:

https://github.com/yhy854/zhong-wen-thinks

After restarting DSH as prompted by the panel, the plugin takes effect in newly opened sessions. It can be disabled in the plugin panel when not needed daily.

Typical Usage and Effects¶

The following comparison is taken from the project README to illustrate the difference in Chain of Thought display before and after enabling the plugin, not to guarantee consistency in every single inference.

Disabled (Example):

Thinking: The user wants to fix the build error. Let me inspect package.json first and check the tsconfig moduleResolution...

Enabled (Example):

用户：帮我看看这个项目为什么构建失败
思考中：
  先看 package.json，确认 Node 版本要求。
  报错说是 ESM 导入问题，多半是 tsconfig 的 moduleResolution 没配对。
  我去读一下 tsconfig.json 验证这个猜测。

If a wrong direction is discovered during the thinking process, you can pause in the DSH session and add clarifications, rather than waiting until the entire round ends to change your approach.

Suitable Scenarios and Notes¶

Who is it for: Chinese developers who primarily use DSH for coding and engineering tasks and wish to understand the model’s reasoning process in real-time.

Compatibility: The README states compatibility with DSH kernel 0.2.x; the peerDependencies in package.json is @deepseek-ai/dsh-system-prompt >=0.2.0-0 (including rc pre-releases).

Known Limitations (excerpted from README):

1. This is a prompt constraint, not an API or kernel-level “hard lock” on reasoning language; after continuously ingesting long pure English logs, English reasoning may still briefly appear.
2. If the first round of the Preset trims the prompt down to only Persona, the plugin section might be trimmed out in the first round, but it can be restored in subsequent stages.
3. If combined with other settings complete: true or plugins that overwrite the entire system prompt, the instructions from this plugin may be overwritten.
4. Forcing non-English reasoning may have a minor impact on theoretical accuracy for a few extreme science problems; in daily engineering tasks, the benefits of readability and timely correction usually outweigh this.

If you need native English reasoning with only Chinese displayed on the interface, consider the official experimental package @deepseek-ai/dsh-experimental-client-ui-cot-translation (post-hoc machine translation display); the README states these can be used together.

Security & Permissions: DSH plugins run under the current DSH process permissions. Please review the GitHub source code and the MIT license yourself before installing to ensure they fit your environment and compliance requirements. The community directory SkillHub is an independent site with no official affiliation to DeepSeek / 幻方.

Conclusion¶

zhong-wen-thinks uses the -900 system prompt slot to prioritize “thinking and displaying the Chain of Thought in Simplified Chinese,” solving the common “English CoT blind box” issue for DSH users. It can be installed using the package name or GitHub URL and turned on or off at will. Directory page: https://www.skillhub.cn/plugins/yhy854/zhong-wen-thinks; Source code and documentation: https://github.com/yhy854/zhong-wen-thinks.