Preface¶

DeepSeek Harness (dsh) treats models, tools, sessions and interfaces as detachable plugins. The official developer preview page sums this up in one sentence: Everything is a plugin. The model selector in the web input box is no exception—it corresponds to the slot conversation.input.model, and the built-in control can be overridden by other frontend plugins.

When switching models daily, the common reasoning strength (Effort) levels for DeepSeek series models are three settings: off, high, and max, usually presented as a dropdown menu in the UI. Someone in the community has already turned this set of options into a slider; dsh-huadongbianzuqi goes even further: it replaces these three levels with controls featuring knobs with avatars, mechanical rails, and dial scales, which the project itself calls the “Sliding Variable Resistor”. The name is a pun on “sliding rheostat”, and the repository README clearly states that this is an unserious but fully functional frontend plugin.

This article is compiled after cross-checking with the community directory page, GitHub repository README, package.json, license and source code: which UI block it replaces, how the levels are mapped, how to install it into the Web profile, and the permissions and asset declarations that must be reviewed before installation.

What is this¶

dsh-huadongbianzuqi is a frontend plugin for the DeepSeek Harness Web client. The GitHub repository is maintained by zjl88858, and the collaborator in the README is credited as “Shen Gui”. The community directory categorizes it under “Tools and Capabilities”, and the code license is MIT-0 (MIT No Attribution). The package name in package.json is dsh-huadong-bianzuqi, current version 0.1.0, and the primary language is JavaScript.

Its targeted problem is narrow: it takes over the model selection control in the input box, replaces the Effort options of DeepSeek models with a three-level slider, while still using Harness’s built-in modelDirectories service to select models and submit effort levels. The plugin does not implement model calls on its own.

You need to distinguish between two sites first. The source code of DeepSeek Harness is at deepseek-ai/deepseek-harness. The plugin discovery portal deepseek-harness-plugin.com is an independent community-run directory, and its about page clearly states that it has no affiliation, endorsement or sponsorship relationship with DeepSeek or High-Flyer, and does not host plugin code. The installation commands in the directory only point users to the maintainer’s repository.

When checking the star count, the directory page shows 6, while the GitHub API query on 2026-08-18 returned 7. The following content will use the GitHub first-hand data as the standard.

Core Features¶

The repository README and src/client.jsx clearly define the capability boundaries, which can be divided into the following points.

Replace Effort Dropdown with Slider¶

The plugin overrides the built-in model selection control via the conversation.input.model slot with a priority of -10. The model list is still retained in the popup; when reasoning levels are available, the Effort area displays the “Sliding Variable Resistor / REASONING DRIVE” control with a knob, rail, and scale. After uninstalling the plugin, the slot injection will be revoked, and the built-in control will automatically recover.

package.json declares that it injects into the web-side @deepseek-ai/dsh-client-ui-conversation and @deepseek-ai/dsh-client-ui-model-selection, with platform restrictions to web.

Frontend Display Names Only Exist in the Browser¶

The Effort IDs in the DeepSeek model directory will not be rewritten. The frontend display names are mapped as follows:

src/effort.js writes the original effort.id into reasoningEffort when submitting, then passes it to session.selectModel. Unit tests also cover this point: the slider submits the original ID such as max, and will not send “Lao Liang”, “Liang Zi” or “Liang Sheng” into the model request.

Unrecognized Effort IDs can still be used, and the UI will display the original name provided by the adapter. If the model declares reasoning but has no available levels, the popup will prompt “No optional reasoning levels for the current model”.

Interaction and Status¶

There are three operating methods listed in the README: mouse drag, rail click, and keyboard arrow keys. In the source code, the slider is a range input with role="slider", and the model list is a native <select>. The plugin retains loading, selecting and error prompts: when the directory loading fails or the switch is rejected, the UI will display the error returned by the directory, or fall back to “Model switch failed” or “Effort level switch failed”.

The avatars are bundled into inline data URLs during the build phase, and no additional static resource service is required at runtime. The host entry apply() in src/index.js is an empty function—all logic is in the browser-side client package.

Installation and Activation¶

The installation command given on the community directory page is as follows, run it in the DeepSeek Harness terminal:

dsh plugin add github:zjl88858/dsh-huadongbianzuqi

This is a web frontend plugin. The repository README requires installing it to the Web profile, then launching with that profile and refreshing the browser page:

dsh plugin --profile web add github:zjl88858/dsh-huadongbianzuqi
dsh --profile web

For reproducible installations, the directory page recommends pinning the commit hash:

dsh plugin add github:zjl88858/dsh-huadongbianzuqi#commit

As of 2026-08-18, the repository has the Git tag v0.1.0 (pointing to commit 19a359e8280fb63dd5c4695cd5d539de05abaf78). The README recommends installing via the tag:

dsh plugin --profile web add github:zjl88858/dsh-huadongbianzuqi#v0.1.0

Installing from Git will execute the project’s prepare script to generate the lib/ directory. pnpm 10 and above may reject the first build. Follow the terminal prompt to approve the build in that profile’s pnpm-workspace.yaml:

allowBuilds:
  dsh-huadong-bianzuqi: true

Then re-run the installation command. Approving the build means allowing this project’s Bun scripts to run on your local machine. The README explicitly requires pinning trusted tags or commits, and do not install unvetted feature branches. The build depends on Bun 1.2 or higher.

The README also mentions a method of “downloading pre-built tarballs from GitHub Releases”, with package names in the format dsh-huadong-bianzuqi-0.1.0.tgz, which does not require Bun during installation. When checking the repository’s Releases page, there are currently no published release assets. Until the maintainer officially releases the tarballs, do not treat that download link as an executable step; use the tag method above for reproducible installations.

To uninstall, use the package name from package.json, not the repository name:

dsh plugin --profile web remove dsh-huadong-bianzuqi

When developing from source, the repository provides:

bun run check

This command runs unit tests and generates lib/index.js and lib/client.js. Use bun pm pack to publish the tarball.

Typical Usage¶

After installation and running dsh --profile web, refresh your browser. The model control in the input box will be taken over by the plugin.

1. Click the model selection entry. When there are levels available, the button text will be Model Name + Display Name, for example, a label with “Liang Zi” or “Liang Sheng”; when no levels are available, only the model name is displayed.
2. Switch models in the “Model” dropdown at the top of the popup. The option format is Provider Name · Model Name.
3. When the current model provides an Effort list, a slider will appear below. Drag the knob, click the rail, or place focus on the slider and use the arrow keys to switch between the publicly available levels.
4. During the switching process, if the directory status is selecting, the control will enter a busy state; on failure, the error text will appear at the bottom of the popup.
5. It recognizes both off used by DeepSeek Harness and none that may be provided by other model directories, both displayed as “Lao Liang”.

The plugin only modifies the selection control in the browser. What is actually sent to the model is still the Effort ID from the directory, and the word “Liang Sheng” in the UI will not be included in the request.

Applicable Scenarios and Notes¶

This is suitable for users who are already using the DeepSeek Harness Web profile and whose current models publicly expose Effort levels via modelDirectories. It targets Web clients that provide conversation.input.model and modelDirectories; CLI or other profiles are not the target environment for this plugin. Sub-agent sessions (judged by sessions.subagentAddress in the source code) will not enable this control.

It is not a universal multi-level reasoning slider. There are other community plugins that adapt to the publicly available levels of the model or make more scale-based Effort sliders; the visual design and naming of dsh-huadongbianzuqi are a playful skin tailored for DeepSeek’s common three-level setup. The number of levels, their names, and whether they work still depend on the model directory, and the plugin will not bypass the limitations of the model or its deployment.

There are several things you need to review before installation:

• The plugin runs with the permissions of the current dsh process, and may execute code during installation. Both the directory page and the repository remind you: first check the source code and license; pin commits or tags for reproducible installations.
• The README defines this project as a joke project. The name, level names and visual design are for humorous expression, and do not represent the endorsement, participation or approval of this project by the characters in the images, DeepSeek or any other individuals or organizations.
• The maintainer states that they do not hold the portrait rights of the people in the avatar images provided with the repository. MIT-0 only covers the code and materials that the maintainer has the right to license, and does not cover the rights that third parties hold over the images and the depicted individuals. Before publicly disseminating, commercially using, redistributing or creating derivative works, you need to confirm the laws and authorizations in your jurisdiction on your own. If the rights holder wishes to remove the assets, the repository directs to contact via GitHub Issues.
• DeepSeek Harness is still in developer preview, and core plugins and APIs will change. This frontend slot will work only if the current Web client still provides conversation.input.model and modelDirectories.

Summary¶

dsh-huadongbianzuqi does one focused thing: in the DeepSeek Harness Web input box, replace the Effort dropdown with a three-level slider, with playful display names, while the original Effort ID is still submitted to Harness. It reuses the official model directory service, and the built-in control will return after uninstallation.

Directory page: https://deepseek-harness-plugin.com/zh-CN/plugins/dsh-huadongbianzuqi/

GitHub: https://github.com/zjl88858/dsh-huadongbianzuqi