Preface

The Web interface for DeepSeek Harness (DSH) allows configuring reasoning effort for models, but the native DeepSeek adapter has limited support for reasoning levels and usually supports only off / low / high / max. For scenarios using a custom provider with llm-pi-ai, more granular control is needed. The dsh-reasoning-effort plugin provides a configuration and selection solution for this: it defines available levels on the settings page and embeds a slider in the model selector to adjust them in real time.

Features

The plugin integrates reasoning effort configuration and selection into the DSH Web UI. It mainly includes the following:

  1. Settings page configuration
    The plugin adds a “Reasoning effort” option to the DSH Web UI settings. You can select the reasoning levels exposed by each custom provider model on this page. Configurable levels include off, minimal, low, medium, high, xhigh, and max.

  2. Model selector slider
    A reasoning effort slider is embedded in the composer’s model selector. After selecting a model, you can drag the slider to choose the reasoning level for the current session. The slider supports mouse dragging, keyboard arrow keys, and wheel operations.

In addition, the plugin only lists models explicitly configured under llm-pi-ai.providers.<provider>.models and does not display directory providers.

Installation and Enablement

The plugin is installed as a local package. Restart DSH Web after installation.

dsh plugin --profile web add ./dsh-reasoning-effort-0.2.0.tgz

Typical Usage

  1. Configure the model list: First add the model list to the custom provider.
  2. Open the settings page: Go to DSH Web settings, find the reasoning effort configuration, and confirm the target model and its supported levels.
  3. Use the slider: Select the model in the chat interface’s model selector, then drag the embedded slider to choose a specific reasoning level.

Use Cases and Notes

  • Adapter compatibility: The seven levels defined by this plugin (off / minimal / low / medium / high / xhigh / max) form the level set for llm-pi-ai custom providers. The native DeepSeek adapter supports only off / low / high / max; if other values are passed, the adapter returns UNSUPPORTED_REASONING_EFFORT. Therefore, use the levels of this plugin only for configuring custom providers.
  • Configuration persistence: All configuration is saved through the DSH settings service, so manual editing of configuration files is not required.