dsh-reasoning-support: DSV4.1 Reasoning Support and Full Agent Tool Retention

Introduction

The DSV4.1 model in DeepSeek Harness (DSH) can sometimes solve problems in minimal mode, but it may omit conditions in full agentic tool and project contexts. The dsh-reasoning-support plugin aims to close this gap by introducing a streamlined analysis process, multimodal context support, and a bounded acceptance/repair loop, bringing the reasoning performance as close as possible to minimal mode while retaining the full agentic toolchain.

Core Features

The plugin provides the following core capabilities:

  • Streamlined Analysis Pipeline: Adds pre-analysis and review stages before the final response, using additional model calls to validate the result.
  • Multimodal Context Support: Supports images and attachments in auxiliary pipelines for screenshot-driven creation and engineering tasks.
  • Adaptive Tool Context: Limits the tool list by default, retaining only native tools and tools discovered on demand to prevent excessive tools from interfering with reasoning.
  • Bounded Acceptance and Repair: Sets limits on the number of repair and evidence-collection rounds to prevent infinite loops.

Installation and Activation

Before installation, ensure the environment meets the following requirements:
* Node.js >= 24
* DSH version 0.1.7-rc.2 through 0.2.0 (0.3.1 uses the new declarative preset API)

Run the following command to install the plugin:

dsh plugin --profile web add dsh-reasoning-support

After installation, restart DSH to activate the new plugin. When creating a new session, select the Reasoning Support preset.

To make it the default preset for new sessions, add the following to the agent-preset-registry configuration entry:

selectedDefault: reasoning-support

Configuration and Typical Usage

Before using this plugin, declare image input support in the DSV4.1 model configuration. Using llm-pi-ai as an example, add the input field to the existing entry:

input: [text, image]

This plugin is suitable for everyday Q&A, screenshot-driven creation tasks, and engineering tasks that require processing documents or source files.

Limitations and Notes

The plugin has the following limitations during operation:

  1. Repair Rounds: At most 2 repair/evidence-collection rounds are allowed, after which the review stage must begin.
  2. Auxiliary Call Limitations: A single auxiliary call has a timeout of 150 seconds. Once the reported auxiliary usage reaches 2, no new auxiliary calls will be initiated.
  3. Image Count: Each auxiliary request supports up to 16 images.
  4. File Processing: Scanned PDFs require page images or OCR processing to be validated. Text files encoded in UTF-8 are supported by default, with a size limit of 2 MiB.
  5. Tool Discovery: When discovering tools on demand, each tool_search returns up to 5 results; the most recent 12 additional tools are retained (configuration range 5–100).

Ecosystem Context

Everything in DeepSeek Harness is a plugin. dsh-reasoning-support is a DSH profile bundle; after installation, there is no need to manually place configuration files, and the plugin will be loaded automatically. For more details, see the GitHub repository or the community directory.

羽毛球分组比赛记分
小程序二维码

欢迎使用《羽毛球分组比赛记分》微信小程序

Xiaoye