Introduction

In the native model-switching logic of DeepSeek Harness, when reasoningEffort is not explicitly specified in the routing configuration, switching models falls back to undefined. In the context of qwen, this is equivalent to enable_thinking: false, causing thinking capability to be silently disabled. dsh-reasoning-effort-slider aims to solve this issue through a UI component and a memory mechanism.

Plugin Scope

dsh-reasoning-effort-slider is a DeepSeek Harness web client plugin maintained by CroissanTTs. It injects a reasoning-effort slider into the chat input bar, supports per-session adjustment, and remembers each model’s selection, automatically inheriting and adapting to it when switching models.

Core Features

  1. UI Injection: Injects the reasoning-effort slider into the chat input bar, and displays it only when the selected model declares support for reasoning effort.
  2. Per-Session Independent Adjustment: Dragging the slider changes only the reasoning effort of the current session and does not affect other sessions.
  3. Multi-Model Support: Supports models with declared reasoning-effort support, including Bailian, DeepSeek, and OpenAI o-series.
  4. Automatic Memory and Inheritance: Remembers the last selected level for each provider/model combination. When switching models, the plugin automatically writes the remembered effort to the new model and constrains it within the maximum level supported by the target model.

Installation and Enablement

To install this plugin, run the following command:

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

Usage Mechanics

  1. Slider Logic: The slider follows DSH’s seven-level logic (off/minimal/low/medium/high/xhigh/max), but displays only the specific levels supported by the current model.
  2. Smart Hiding: If a model supports only one level (such as kimi-k3) or only off, the slider is hidden automatically.
  3. Switching Adaptation: When no level has ever been selected, it displays the gray “Default” state. When switching models, the plugin first tries to retain the original level; if the target model does not support that level, it is automatically downgraded to the closest available level supported by the target model (and no lower than the target model’s minimum level).
  4. Preventing Silent Disabling: Because the slider never displays off, this mechanism cannot silently disable thinking.

Notes

  1. Configuration Semantics: In the context of qwen, configuring it as undefined is equivalent to disabling thinking.
  2. Permissions and Security: The plugin runs with the permissions of the current DSH process; it is recommended to review the source code and license before installation.
  3. Relationship Description: This is a generic plugin and can also be used as an optional dependency of dsh-bailian-models.

Summary

This plugin fills the gap in DSH’s reasoning-effort management during model switching through a simple slider and memory mechanism. For more information, see the plugin directory https://www.skillhub.cn/plugins/CroissanTTs/dsh-effort-slider or source repository https://github.com/CroissanTTs/dsh-effort-slider.