Introduction¶

When troubleshooting model outputs in DeepSeek Harness (DSH), you often need to expand the Think row to see the complete Chain of Thought (CoT). Reading English reasoning content requires translation on the fly for Chinese teams, and repeatedly expanding and taking screenshots for translation during a single session is a costly process.

Below is an introduction to lzyxczw-cyber/dsh-thinking-cn, a configuration plugin. Instead of implementing a translation client itself, it mounts the experimental reasoning translation feature built into DSH and fixes the target language to Simplified Chinese. It is suitable for Chinese users who need to read the Chain of Thought content directly and for developers troubleshooting model outputs.

What is it¶

dsh-thinking-cn is a DSH bundle maintained by lzyxczw-cyber under the “Model Reasoning” category. It mounts two experimental packages built into DSH:

@deepseek-ai/dsh-experimental-translator
@deepseek-ai/dsh-experimental-client-ui-cot-translation

and fixes the configuration to provider: bing, targetLanguage: zh.

This package itself does not contain a translation client. Paragraph scheduling, switching between “View original” and “View translation”, and storing translation results per session are all built-in experimental features of DSH. After installing this bundle, you no longer need to manually find this feature in the Plugins page or manually switch the target language.

Core Features¶

1. After expanding the Think row, completed paragraphs are translated into Simplified Chinese in real-time as they arrive, while the unfinished tail remains in the original language.
2. The toolbar above the main content allows switching between “View original” and “View translation”; switching does not trigger additional model requests.
3. Translation results are stored in the session via plugin:translator/request and plugin:translator/result records. Re-expanding after collapsing, refreshing the page, or restarting the host will reuse existing results without calling the translation service again.
4. Translation goes through an anonymous Microsoft Edge (Bing) endpoint; no login, API key, or per-segment billing is required. Only the selected source text is sent, while the original reasoning content remains unchanged in the session and model input.

Installation and Usage¶

First, confirm that your DSH version includes the two experimental packages mentioned above (line 0.2.1, >=0.2.1-alpha.2 <0.3.0), then execute the installation:

dsh plugin add github:lzyxczw-cyber/dsh-thinking-cn

If the host prompts, restart the profile. In DSH Desktop, go to Settings → Plugins → Add and paste the repository address to install.

The declared DSH dependency version ranges in package.json are as follows:

"peerDependencies": {
  "@deepseek-ai/dsh-experimental-translator": ">=0.2.1-alpha.2 <0.3.0",
  "@deepseek-ai/dsh-experimental-client-ui-cot-translation": ">=0.2.1-alpha.2 <0.3.0"
}

Both peer dependencies are marked as optional, but the functionality itself depends on their existence.

Typical Usage¶

After installation, use it in the following four steps:

1. Send any prompt that generates reasoning, then expand the Think row.
2. Completed paragraphs in the expanded content will be translated into Simplified Chinese, while the unfinished tail remains in the original language.
3. Use the toolbar above the content to switch between “View original” and “View translation” without generating new requests.
4. Collapsing and re-expanding, refreshing the page, or restarting the host will all reuse stored results.

There are two ways to change the provider or target language. The default value is defined in cordis.patch.yml within the package; to adjust without modifying the package, override this line in your own profile patch:

- id: cot-translation
  name: '@deepseek-ai/dsh-experimental-client-ui-cot-translation'
  config:
    provider: bing      # bing | google | deepseek-account | deepseek-official
    targetLanguage: zh  # BCP 47 tag, or auto to follow the UI language

deepseek-account and deepseek-official initiate separate paid DeepSeek Flash requests and require account or API credentials to be configured first. The anonymous endpoint will not automatically switch to another provider on failure.

Use Cases and Notes¶

Suitable for users using the DSH Web or Desktop interface who need to read the Chain of Thought in Chinese directly, especially for developers troubleshooting model outputs. This feature is a client-side view; it only changes the display, not what the model sees.

Pay attention to a few points before use:

1. Only the expanded content body is translated; the single-line summary during folding remains in the original language.
2. Translation only starts when the content is expanded; folding does not trigger translation.
3. Anonymous Bing/Google endpoints may have rate limits or be unavailable; if they fail, the original text is retained and a Retry option is provided.
4. Translated Markdown or code formatting may differ from the original; the original is always accessible with one click.
5. This bundle mounts the same two row IDs as the official Reasoning translation experiment, so only one can be enabled. If you enable the same capability later, switch to the settings page of that one and set the target language to zh; do not rely on this patch.

The plugin runs with the permissions of the current DSH process; check the source code and license before installing. This plugin is MIT-licensed, and the bundled translation package is distributed by DeepSeek under their respective MIT licenses.

Conclusion¶

The value of this bundle lies in eliminating the step of manually finding and switching the target language, fixing the experimental reasoning translation in DSH to Simplified Chinese output, while retaining one-click access to the original.

Directory page: https://www.skillhub.cn/plugins/lzyxczw-cyber/dsh-thinking-cn
Repository: https://github.com/lzyxczw-cyber/dsh-thinking-cn