Preface

In DeepSeek Harness, the native “Think” line usually only shows a single-line stream or a large block of raw text. When reading long reasoning chains, this presentation makes it difficult to quickly grasp the key logic. The dsh-thinking-summary plugin captures the model’s reasoning stream and generates per-segment summaries displayed in the chat history, replacing the original “Think” line.

Plugin Introduction

This is lyuwen/dsh-thinking-summary, a readable thinking display plugin for DeepSeek Harness. It interacts with the Host through a remote service, breaking down and summarizing long reasoning streams to help developers examine the thinking process more clearly.

Installation and Enablement

Use the dsh plugin command to add the plugin to the web profile:

dsh plugin --profile web add github:lyuwen/dsh-thinking-summary

After installation, restart dsh web for the changes to take effect.

Core Features

The plugin divides the reasoning display into three stages, presented dynamically in the chat history:

  1. Collapsed summary: The default state. It shows a one-line summary for the most recently completed segment, and displays “Thinking…” while the first segment is in progress.
  2. Expanded summary chain: Click the title to expand it and display a vertical chain of summaries for the reasoning flow, including the “Show raw thinking” link.
  3. Fully expanded raw text: Click the link to view the complete raw thinking text, and click “Show summaries” to return to the summary view.

Paragraph Grouping Mechanism

Summaries are calculated based on paragraphs separated by blank lines, but fragments below the threshold are merged forward. Fragments that have not reached the character threshold accumulate until approximately 160 characters to form one summary unit. This prevents the model from frequently outputting very small paragraphs and generating an excessive number of summaries.

Rendering Support

Client-side rendering supports code, tables, lists, and links. However, it only supports a lightweight subset of Markdown and does not support mathematical formulas, footnotes, or file mention chips.

Configuration and Usage

  1. Select the summarization model: In Settings → General, select the model used to generate summaries. This model is stored in Host memory and will be reset to the default value (DeepSeek-V4-Flash) after restarting dsh web.
  2. View raw content: In the expanded summary chain, click the “Show raw thinking” link to view the raw text.
  3. Return to summaries: In the fully expanded state, use “Show summaries” to return to the summary view.

Notes

  • Historical reasoning is not re-summarized: The plugin only summarizes reasoning content observed during streaming. Reasoning that has already been settled in history is not re-summarized (unless the Host already has a cache).
  • Model selection limitation: The summarization model is stored in Host memory and reverts to the default value after the process is restarted.

Summary

This plugin is suitable for scenarios where you need to examine the model’s reasoning process in detail while keeping the interface clean. It balances information density and readability through segmented summaries and a three-level display.