In the Web canvas of DeepSeek Harness (DSH), model selection is typically handled by a single input box. When managing multiple models or configuring reasoning modes, this single view can feel cramped. The dsh-model-accordion plugin introduces a collapsible Provider structure, expanding the model selector from a single “seat” into a hierarchical list, keeping the interface clean while providing richer configuration options.

Plugin Positioning

dsh-model-accordion is a Web UI plugin developed by the maintainer lakabuji-arch. It sits within DSH’s plugin ecosystem, does not involve constructing backend services or model catalogs, and is responsible only for enhancing the interactive experience of the Web UI.

Installation and Activation

Install the plugin in a Web Profile environment:

dsh plugin --profile web add dsh-model-accordion

After installation, refresh the Web UI to load the new component.

Features

The plugin primarily implements the following interaction logic:

  1. Provider Accordion Structure: By default, Provider groups are collapsed and only expand when clicked to view specific models.
  2. Reasoning Mode Reading: The plugin reads each model’s reasoning mode (reasoning-effort) options from the resolved catalog metadata. A built-in official reasoning-effort table serves as a fallback, with the following priority: model’s own configuration > Provider-level default > built-in table.
  3. Data Flow: The selected model and reasoning mode are sent to the system through DSH’s shared model-directory service.
  4. UI Limitations: The popup height is constrained by the internal scrolling area, preventing the interface from stretching indefinitely.
  5. Theme Adaptation: Colors are defined using DSH’s semantic theme tokens, and switching between light and dark modes is supported.

Configuration and Usage

The plugin reads DSH’s configuration file to obtain model metadata. You can declare reasoning modes for specific models in settings.yaml.

For example, add custom reasoning-effort levels for a model not defined in the official table:

providers:
  my-relay:
    api: openai-completions
    baseURL: https://…
    models:
      - id: my-unknown-model
        api: openai-completions
        contextWindow: 200000
        maxTokens: 32000
        input: [text]
        reasoningEfforts:
          off: null      # 允许用户关闭思考
          low: low
          high: high
          max: max

The plugin prioritizes the model’s own reasoningEfforts configuration. If the model does not support the reasoning mode selected by the user, the option is silently discarded, and the system uses the model’s default value without throwing an error.

Notes

  1. Plugin Nature: This is a Web UI plugin. It does not add Host services, model providers, vision backends, or model catalogs. Provider availability and image session admission rules are still controlled by DSH and the installed Provider/Vision plugins.
  2. Compatibility: The plugin depends on the official model selection package and model-directory service provided by the DSH Web environment.
  3. Language Limitation: In the current version, status prompts and search box text are all in English. Full localization integration is planned for future releases.
  4. Permissions: The plugin runs with the permissions of the current DSH process. It is recommended to inspect the source code and license (MIT) before installing.