Introduction¶

The programming interface of DSH usually contains only one workspace. If you want to use the Claude Science Chinese interface, independent runtime environment, model service settings, and artifact downloads right next to the assistant without opening a separate CSSwitch, the plugin introduced in this article, ysx12138/dsh-science-panel, does exactly that.

What is this¶

ysx12138/dsh-science-panel is a community plugin maintained by Ysx12138, categorized as a client, with a description of “Claude Science Chinese Panel for DeepSeek Harness”. It embeds a locally running Claude Science instance into the DeepSeek Harness main area, complete with a Chinese interface, model service settings, and document downloads, without the need to install a separate CSSwitch.

It is necessary to clarify the boundaries: this is a community-developed draft and has no official affiliation with Anthropic, DeepSeek, OpenCode, or CSSwitch. The project includes a localization dictionary, backend source code, and macOS Apple Silicon backend and Windows x64 gateway builds. The Science core, model service, and API Key are three separate things: the plugin is responsible for the working interface and bridging, while model capabilities come from the selected service.

Scope of Support¶

• Verified on Apple Silicon Mac: isolation of Science startup, stop, real-time conversation, old conversation preview, Chinese display, and document download.
• Windows 11 x64 only provides preview adaptation and has not yet completed acceptance testing within the actual DSH graphical window.
• Intel Mac and other systems are not supported by the current build.
• See package.json for version dependencies: engines.dsh is >=0.2.1-rc.5 <0.3.0, and Node requires >=20. Before installation, ensure your local DSH version is within this range.

Installation and Activation¶

First, determine your local DSH version. The installation source in the command is the repository address given in the README; do not concatenate it yourself based on the plugin name.

1. Install and open DeepSeek Harness. See engines.dsh in package.json for the version range.

2. In DSH, go to “Plugins” -> “Add Plugin” -> “Install Third-Party Plugin”, enter the GitHub repository address and install:

https://github.com/Ysx12138/dsh-science-panel

After the npm package is published, you can also enter dsh-science-panel.

3. Open “Claude Science” on the left side of DSH. On first use, a model settings dialog will pop up: by default, select OpenCode / Qwen3.8 Flash, fill in the API Key for that service and save. Users who already have an Anthropic official API Key can also select the corresponding preset.

4. The plugin will automatically look for existing Science or CSSwitch cache. If none exists, it will download and install the official Science via your computer’s network or the configured system proxy, and then start it automatically. Please keep DSH open during the download process.

The order of these steps is: Install plugin -> Open panel -> Configure model service -> Wait for first download and startup.

Model Service and Keys¶

Presets will automatically fill in the service address and model name. To use other services that support the Anthropic Messages interface, select “Custom”, then enter the service address, model name, and corresponding Key:

Service Address: <Custom Service Address>
Model Name:   <Custom Model Name>
Key:      <API Key for the service>

You need to re-enter the Key when switching to different services to avoid sending the original service’s key to the new service. The Key is saved in a local private file with permissions restricted to the current user, but the file content is plaintext; requests will be sent to the selected model service and may incur costs.

Currently, only the Anthropic-compatible interface is supported. OpenCode’s DeepSeek V4.1 Flash uses the OpenAI-compatible interface in its official model list and cannot be used directly as an Anthropic preset for this project; therefore, Qwen3.8 Flash, which supports the Anthropic interface in the list, is selected by default. Availability of models, service permissions, and pricing are subject to the service provider.

Mac users can run the check script in the cloned directory to confirm the local architecture and whether Science is in place. This command only checks and does not install or modify the application:

./scripts/check-prerequisites.sh

Download, Signing, and Runtime Environment¶

The downloaded program checks Anthropic’s developer signature, and existing files will not be overwritten. Installing on Mac to the “Applications” folder requires write permission for the current user.

Windows downloads the official standalone EXE to %USERPROFILE%\.dsh-science-backend\runtime-windows and starts it using a standalone data directory, without the need for CSSwitch, Bash, or WSL. Using the official software and account must comply with their terms of service and availability requirements. CSSwitch’s external skill transaction installation bridge has not yet been ported to Windows and will be explicitly rejected when called.

Windows downloads and runs depend on the built-in SQLite in Node.js 22.13 or higher, which runs in the DSH background. Windows uses system user file permissions to protect the configuration and uses system mutexes to protect startup and exception recovery, preserving the Windows system user directory while only isolating the Science data directory.

Localization Capabilities and Boundaries¶

The Chinese dictionary covers common navigation, buttons, tooltips, and dynamic states; content not covered by the dictionary will still be displayed in English. Code and input areas will not be rewritten by the localization script, and original conversation and document data will not change due to interface translation.

The backend appends Chinese output requirements based on the original system instructions, covering task progress, questions, and multiple-choice options. Whether the model follows them still depends on the model’s capability; if the user explicitly requests another language, the user’s request is followed.

The plugin runs with the permissions of the current DSH process. Science runs in a separate directory under DSH. The project only provides the interface and runtime bridging, and does not include personal history sessions, private skills, MCP configurations, or academic service keys.

Verified and Pending Verification¶

Verified on existing Apple Silicon Mac: isolation of Science startup, stop, real-time conversation, old conversation preview, Chinese display, and document download; verified in an independent DSH test configuration that local directories can be used to install and register plugins. Automatic download and installation have been verified through system proxy, real official download, signature check, and temporary directory installation; copying existing CSSwitch cache and starting it independently have also been verified. Automated tests use mock services and test keys.

End-to-end testing from plugin installation to real model response for brand-new macOS users has not yet been completed. When testing presets with existing OpenCode keys, the service returns a Cloudflare HTTP 403 (1010), no model response is received, and it is impossible to determine whether the key is valid based on this; Anthropic official presets have not been verified with real keys. Model services may change addresses, protocols, or permissions.

Windows automated testing uses GitHub’s Windows Server 2025 environment, covering model gateway compilation, Windows file locks, Chinese command forwarding for mock model services, and the independent startup, login link, stop, and restart of the official Science core. The DSH embedded interface, real model conversation, and article download on Windows 11 physical machines have not yet been verified. Tests use temporary directories and fake keys and do not contain personal data.

Applicable Scenarios and Notes¶

Suitable for developers who want to use the Claude Science Chinese interface and configure model services directly within DSH; for those who do not intend to maintain CSSwitch or switch back and forth between editors and standalone applications.

Note three points. First, the plugin runs with the permissions of the current DSH process; before installing, check the source code and license. Second, the backend is a temporary signature build without an Apple Developer identity signature and may be subject to macOS installation restrictions. Third, do not commit user keys, historical data, or the application core itself to the repository.

Development and License¶

npm test runs the plugin tests. backend-source/ contains the corresponding CSSwitch modified source code; third-party/ contains licenses, binary sources, and collected dependency declarations. CSSwitch uses the MIT license. Dependency declarations and license lists have been collected, but binary dependencies, frontend resources, and distribution licenses after packaging still require a full review.

Conclusion¶

The value of this plugin lies in converging the Claude Science Chinese panel and model service settings into the DSH main area, saving the overhead of maintaining a separate CSSwitch. Currently, it is still a community preview draft; macOS functionality is relatively complete, while Windows side only reaches the verification level of the gateway and independent startup of the core. Before connecting to real models, you need to verify the service address, interface, and permissions yourself.

• Directory Page: https://www.skillhub.cn/plugins/Ysx12138/dsh-science-panel
• GitHub: https://github.com/Ysx12138/dsh-science-panel