Introduction

When using third-party models in DeepSeek Harness (DSH), it is often necessary to declare two things: what thinking levels the model supports, and whether the model accepts image input. dsh-better-reasoning-effort provides in-app editing interfaces for these two types of declarations, and offers a quick reasoning-effort slider in the official model menu.

What is this

HaoyueQin/dsh-better-reasoning-effort is a DSH plugin maintained by HaoyueQin, licensed under MIT. It places the editing for third-party model reasoning intensity (thinking levels) and image input modalities into the official Models page card, and provides automatic adaptation based on the model knowledge base and wire-protocol reasoning.

DSH’s philosophy is “everything is a plugin”; the community directory is an independent site with no official affiliation to DeepSeek / Hyperbolic.

Below are the core features, installation status, typical usage, and notes.

Core Features

Editing in Models Page Cards

The plugin injects two editing blocks into the expanded area of each model row on the official Models page:

  • Reasoning effort
  • Input modalities

These two blocks are saved by the same pair of Apply / Reset buttons, utilizing the official settings.mutate save workflow.

One-click Auto-adapt

Auto-adapt uses the built-in model knowledge base and wire-protocol reasoning to fill in recommended thinking levels and wire spellings with one click.

It can also probe the provider RAW /models list and combine signals; suggestions are labeled with high, medium, or low confidence.

Staging Unsaved Rows

For unsaved rows, Stage can be used to stage declarations. The documentation lists two scenarios:

  1. provider create card
  2. New model rows under a saved provider

When the row or provider is saved, the declaration is automatically committed.

Image Input Declaration

Through the Image input checkbox, you can declare that a manually created model supports image input. This declaration affects:

  • Composer attachments
  • read-image tool
  • proxy gating

Unchecking the box narrows the declaration to text-only.

Quick Slider in Official Model Menu

The plugin provides a quick reasoning-effort slider inside the official model menu. The documentation notes the source as HanaAyane's dsh-reasoning-effort. The official Composer’s bottom-right model · effort display remains unchanged.

Installation and Enablement

Currently, verified documentation does not provide the original text of the official installation command, so the installation command is not constructed based on the name. It is recommended to check the source code and license before enabling; the plugin executes with the current DSH process permissions.

Repository: https://github.com/HaoyueQin/dsh-better-reasoning-effort
License: MIT
engines: node ^20.0.0 || ^22.0.0 || >=24

Typical Usage

Below, following examples from verified documentation, perform in-page editing first, then auto-adapt or stage to commit.

  1. In the official Models page card, edit each model’s reasoning effort levels and image input support.
  2. Use Auto-adapt to one-click fill recommended levels and wire spellings from the model knowledge base and wire-protocol reasoning.
  3. For unsaved provider create card or new model rows, use Stage; the declaration is automatically committed after saving.
  4. Use the Image input checkbox to let custom models support image attachments, the read-image tool, and proxy gating.
  5. Manually copy the context window/max output hints displayed as read-only in Zoned suggestion to the official capacity input box.

Applicable Scenarios and Notes

Suitable for scenarios in DSH where you want to maintain third-party models and complete the configuration of thinking levels, wire spellings, and image input support within the official Models page.

Points to Note:

  • Staging cannot express a deliberate unset; to persist a declaration of “nothing”, you must first save the row, then clear all levels and Apply.
  • The context window/max output in Zoned suggestion are hints only and will not auto-fill.
  • Endpoint probe credentials are resolved server-side and never echoed.
  • Currently, verified documentation could not find star count, clear category tags, directory page URL, or the original text of the official installation command.
  • The README is truncated, and subsequent parts did not appear in the documentation; installation and complete usage may not be covered.

Conclusion

After the steps above, dsh-better-reasoning-effort places the third-party model’s reasoning intensity and image input declarations back into the official Models page, and covers the main editing paths with Auto-adapt, Stage, and the menu slider.

GitHub:

https://github.com/HaoyueQin/dsh-better-reasoning-effort

Currently, verified documentation does not provide the directory page URL, so the directory page link is not listed.