Introduction

In the DeepSeek Harness (DSH) development workflow, controlling model reasoning effort is a key step for tuning model output. Third-party controls often require manual state maintenance and can easily conflict with the host service. The dsh-effort-ultra plugin directly invokes DSH’s official session service to provide unified management of model selection and reasoning effort, and offers a Codex-like blue-purple gradient interface.

Plugin Role

This is a native reasoning effort control for the DeepSeek Harness composer. It is maintained by developer LoveIrishCoffee and aims to solve the lack of reasoning effort control in the native interface. The plugin does not rely on legacy plugin logic or modify the interface of other plugins; instead, it registers as an independent slot in the DSH composer system.

Core Features

The plugin provides a complete reasoning effort control solution with the following features:

  1. Six-Level Step Agent: supports Off, Low, Medium, High, Xhigh, and Ultra.
  2. Continuous Sliding and Explicit Menu: supports continuous dragging by holding the thumb, while a triangular button next to the level allows direct selection, making operation intuitive.
  3. Model Switching and State Synchronization: provides switching functionality in a multi-model catalog. All level changes are written to the current DSH session; after refreshing or switching sessions, the UI display is based on the actual host state.
  4. Follow Default and Off Logic: supports the “Follow Model Default” option; a reset arrow or menu can clear an explicit level; off truly submits the off level.
  5. Theme Adaptation: adapts to light and dark themes, and automatically disables star and shimmer effects when the system has “Reduce Motion” enabled.

Installation

Install it with the official plugin command:

dsh plugin --profile <你的 profile> add github:LoveIrishCoffee/dsh-effort-ultra

After installation, please restart DSH. If the plugin is later published to npm and listed, you can also install it by package name:

dsh plugin --profile <你的 profile> add dsh-effort-ultra

Usage

  1. Drag the Level: hold the thumb on the slider and drag it continuously; level changes are submitted in real time during dragging.
  2. Select a Level: click the triangular button next to the level and choose “Follow Model Default” or a specific level from the pop-up menu.
  3. Switch Models: when switching models inside the card, the plugin reads the model’s reasoning declaration from the official DSH catalog. If a model does not declare a default level, the UI automatically shows “Follow Model Default”.

Technical Implementation and Notes

  • Slot Occupation: The plugin occupies the conversation.input.model slot with priority -20. If an older third-party level control is installed at the same time, this plugin takes visual precedence.
  • Level Source: All available levels are taken from the reasoning declarations in the official DSH model catalog; the plugin does not guess or fabricate levels.
  • Dependencies and License: The plugin is released under the MIT license. Before installing, it is recommended to review the source code and license to ensure compliance with your project’s security policy.

Summary

dsh-effort-ultra fetches reasoning effort declarations from DSH’s official service and exposes them to users in a native way. For developers who need fine-grained control over model output, it is a stable and well-compatible tool.