Preface¶

The plugin system of DeepSeek Harness (DSH) emphasizes “everything is a plugin.” In the Web profile, the built-in conversation.input.model slot in the input box is responsible for model selection; if you want to select both the model and the reasoning effort level in the same entry point, you can override this slot.

dsh-reasoning-effort-hdbzq is a frontend plugin prepared for this purpose: it replaces the model selection control in the input box with a model list plus a three-level reasoning effort slider.

Below are its positioning, features, installation method, typical usage, and precautions.

What This Is¶

This is a DeepSeek Harness frontend plugin, version 0.1.0, licensed under MIT. The repository address is shown as flyemFSB/dsh-reasoning-effort-hdbzq.

It solves the problem of “model selection and reasoning effort levels being scattered across different entry points”: it overrides the built-in conversation.input.model slot, replacing the model selection control with a model list and a three-level reasoning effort slider. The plugin reuses the official modelDirectories service and shared directory store, and does not implement model invocation itself.

Core Features¶

• Overrides the built-in conversation.input.model slot; after uninstallation, the built-in model selection slot is automatically restored.
• The popup panel displays the model list grouped by provider, including loading, failure, and empty states.
• Provides a draggable three-level slider; dragging is responsive, with elastic bounce-back when out of bounds, and snapping to the nearest level and submitting on release.
• Supports keyboard left/right arrow keys, instant submission by clicking on the track, and Escape to close and restore focus.
• Effort level names and avatars are mapped by image filename position; falls back to the original host directory names when images are insufficient or exceed three levels.
• The real Effort ID is passed as-is to session.selectModel; the display name shown in the UI is not written into the model request.
• Styling uses the official --dsw-alias-* semantic tokens, supporting light/dark themes.
• Copy supports the official locale dictionary mechanism with zh / en dual dictionaries.
• Build output includes an official __ModuleLoader__ closure package containing lib/client.js and source maps, with style data-plugin automatically injected.
• Does not render content when the addressed sub-agent session is hit.
• The number of effort levels varies with the efforts array in the model directory: fewer than 3 levels shows only the corresponding count; more than 3 levels shows the host original names starting from the 4th level.

Installation and Activation¶

First, install the plugin in the target Web profile:

dsh plugin --profile web add https://github.com/flyemFSB/dsh-reasoning-effort-hdbzq/releases/download/v0.1.0/dsh-reasoning-effort-hdbzq-0.1.0.tgz

After installation, start it:

dsh --profile web

The target environment needs to provide a DeepSeek Harness Web profile with the conversation.input.model slot and the modelDirectories service, and dsh must be >= 0.1.0-rc.6. Runtime dependencies are provided by the Web profile, with peerDependencies including @deepseek-ai/cordis, multiple dsh-client packages, and clsx.

Typical Usage¶

1. Open the model selection entry point in the input box, and the popup panel will display the model list grouped by provider.

2. Select the target model.

3. Drag the slider to adjust the reasoning effort level; after releasing, it snaps to the nearest level and submits the real Effort ID.

4. You can also use the keyboard left/right arrow keys to adjust the level, or click on the track for instant submission.

5. For rapid consecutive switching, submissions use latest-wins logic with subsequent propagation; if a submission is rejected, it bounces back to the confirmed level.

6. Press Escape to close the panel and restore focus.

Source Build and Verification¶

To build and verify from source:

npm install
npm run check

npm run check corresponds to typecheck, unit tests, and a full build. To package:

npm run pack

To rename effort levels, rename the image files and update the ASSET_ORDER in build.mjs accordingly.

Uninstallation¶

dsh plugin --profile web remove dsh-reasoning-effort-hdbzq

After uninstallation, the built-in model selection slot is automatically restored.

Applicable Scenarios and Notes¶

Suitable for:

• Developers who provide the conversation.input.model slot in the DeepSeek Harness Web profile and want to select both the model and reasoning effort level in the same panel.
• Client plugin developers who want styling to follow light/dark themes and copy to support zh / en.
• DSH plugin developers who want to override official slots rather than implement model requests themselves.

Notes:

• The plugin runs with the current dsh process permissions; source code, license, and repository address should be checked before installation.
• The MIT license only covers code and materials the maintainer is authorized to license; it does not grant permission for likeness, name, identity, publicity rights, trademarks, copyrights, or other third-party rights. You need to independently confirm the required permissions before public distribution, commercial use, redistribution, or creating derivative versions.
• The slot override has a fixed priority of -10, relying on the built-in slot’s default priority of 0; if the official changes the built-in registration priority or slot name, slot conflict errors may occur.
• Effort level display names are mapped by position, not by id; the same image may have different effort semantics across different models.
• Avatars are embedded as data URLs at build time; the tarball does not contain separate static assets.
• This plugin is a browser-side UI pipeline; it does not assemble or send provider requests and has no KV Cache effect.

Links¶

• The catalog page given in the materials: https://www.skillhub.cn/plugins/flyemFSB/dsh-reasoning-effort-hdbzq
• GitHub: https://github.com/flyemFSB/dsh-reasoning-effort-hdbzq

The community catalog is an independent site and has no official affiliation with DeepSeek / High-Flyer.