DeepSeek Harness (DSH) follows the “everything is a plugin” philosophy in its ecosystem. In scenarios where fine-grained control of reasoning effort at the model level is needed (for example, reasoning_effort), manually editing configuration files is often not intuitive. The dsh-reasoning-effort plugin solves this problem by providing a visual configuration entry and protocol mapping capabilities.

Plugin Introduction

This plugin is maintained by Mu-scorpio and uses the MIT license. It primarily serves any model in DSH, allowing users to configure reasoning effort through a UI slider and providing a purple glow effect at the highest level.

Core Capabilities

  • Generality: Not limited to specific Providers. Any model connected to Harness can be configured.
  • Automatic discovery: Zero-configuration automatic scanning of all existing Providers in llm-pi-ai.
  • Level support: Supports seven levels: off, minimal, low, medium, high, xhigh, and max.
  • Configuration hierarchy: Supports Provider-level defaults, model-level overrides, and protocol (wire) value mapping.
  • UI interaction: The model selector is grouped and collapsible by Provider; the slider uses a thick track with discrete levels; the highest level includes purple glow, blinking, and trail effects.
  • State protection: Model selection is staged on the front end and committed only when the selector is closed, avoiding UI flicker.
  • Non-destructive behavior: Updates are appended to settings and do not overwrite existing reasoningEfforts.

Installation and Enablement

Run the following command in the terminal to install the prebuilt Bundle into the DSH web profile:

dsh plugin --profile web add -w dsh-reasoning-effort

After installation, restart Harness to make the configuration take effect.

Configuration

The plugin automatically discovers Providers in llm-pi-ai. If special mapping or exclusion is needed for a specific Provider, you can add override configuration in the Cordis patch:

- insert:
    - id: reasoning-effort
      name: dsh-reasoning-effort
      config:
        defaults:
          reasoning: medium
        providers:
          special-provider:
            api: openai-responses
            efforts:
              high: reasoning_high
            models:
              non-reasoning-model:
                disabled: true
          excluded-provider:
            disabled: true

Configuration Field Descriptions

Field Description
auto Whether to automatically discover all existing Providers. Defaults to true. When set to false, only explicitly listed items are processed.
defaults.api The protocol fallback value used when a Provider does not have its own api.
defaults.reasoning Supplies a default reasoning level for all Providers, but does not overwrite existing values.
defaults.efforts Overrides the generic wire value mapping for all Providers.
providers.<id>.api Selects a protocol default mapping for a specified Provider.
providers.<id>.reasoning Supplies a default reasoning level for a specified Provider.
providers.<id>.efforts Overrides the wire value mapping for a specified Provider.
providers.<id>.disabled Excludes a specified Provider.
providers.<id>.models.<model-id>.efforts Overrides the mapping for a single model.
providers.<id>.models.<model-id>.disabled Explicitly marks a single model as not using reasoning effort.

Development and Debugging

If you need to run from source or build it, execute the following steps:

pnpm install
npm run build
dsh web --patch ./cordis.yml

Notes

  • Security: The plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license before installation.
  • Compatibility: Existing reasoningEfforts are never overwritten.
  • Scope: It does not add Provider credentials, requests, tools, or telemetry data; it only enhances the settings contract.

This plugin focuses on managing the configuration of model reasoning effort. Actual model invocation is still handled by DSH itself.