Foreword

DSH’s plugin ecosystem allows processing logic to be hooked into the harness’s request chain. dsh-cot-summari**z**ation focuses on reasoning display: it does not directly show the original Chain of Thought (CoT), but instead swallows the raw reasoning increments at the llm/stream interception layer, then has a small model compatible with Chat Completions generate a summarized Chain of Thought, flowing into the UI’s Think row as a reasoning block.

Compared to directly displaying the original Chain of Thought, this plugin changes the interface display to a summary generated by a small model. The original Chain of Thought can still be recovered in the model’s visible history, avoiding the impact of the summary on multi-turn reasoning in the Agent Loop.

Plugin Positioning

  • Maintainer: MeowLynxSea
  • Category: extension
  • License: MIT
  • Positioning: A DSH plugin for hiding the interface display of the original Chain of Thought and displaying a summarized Chain of Thought generated by a small model.
  • Hiding Mechanism: Open-source, auditable, and toggleable.

Core Capabilities

  1. Swallows raw reasoning increments at the llm/stream interception layer, ensuring the UI, streaming chunks, and persisted transcript do not retain the original text.
  2. Rewrites a summarized Chain of Thought via a small model, flowing it into the UI’s Think row as a reasoning block.
  3. Supports streaming chunked summarization, triggerable by chunkChars and chunkIntervalMs, and supports adaptiveChunk for dynamic chunking.
  4. Supports near-duplicate elimination using a bigram Dice similarity threshold of 0.65, and detects core repetitions using the Longest Common Substring (LCS).
  5. Supports the Nekomimi boundary protocol, performing boundary recognition and normalized comparison for clauses ending in meow, ~, 〜, or ～.
  6. Configurable provider and model for the DSH’s own LLM channel; follows the current request when left empty.
  7. Supports hide, pass-through, and drop error handling strategies, and allows short original CoT text to pass through.
  8. Recovers the original Chain of Thought in the model’s visible history, avoiding the impact of the summary on multi-turn reasoning in the Agent Loop.

Cost and Data Path

  • The additional cost is +1 small model call.
  • The original Chain of Thought will be sent to the DSH configured provider/model.
  • If avoiding external network access is required, the DSH provider can be pointed to a local model.
  • When the main call is aborted, the summary request is aborted synchronously and is subject to timeoutMs.
  • When typewriter mode is enabled, the stream is serial, and the reply body and persisted data will queue after the typing finishes.

Installation and Enablement

Install the plugin:

dsh plugin add github:MeowLynxSea/dsh-cot-summerization

If installing from a local directory:

cd ~/.dsh/profiles/web
pnpm add file:/path/to/dsh-cot-summerization

To disable the plugin, simply patch out the cot-summarizer entry in the profile’s cordis.patch.yml.

Web Client Configuration

Adjust the following configuration items in Web Client → Settings → cot-summarizer:

enabled
style
language
minReasoningChars
maxSummaryChars
timeoutMs
incremental
chunkChars
streamReasoningBlock

Where:

  • enabled controls the master switch.
  • style and language affect the summary output.
  • minReasoningChars is the threshold for allowing short original CoT text to pass through, defaulting to 32 characters.
  • maxSummaryChars limits the summary length.
  • timeoutMs constrains the summary waiting time.
  • incremental, chunkChars, and streamReasoningBlock affect streaming chunked summarization and the Think row display.

Applicable Scenarios and Notes

Suitable for the following scenarios:

  • Wanting the original Chain of Thought not to be displayed directly in the UI.
  • Wanting a summary generated by a small model that is closer to “thinking in progress”.
  • Needing to keep the original Chain of Thought in the model’s visible history.

Notes:

  • This plugin runs with the current DSH process permissions; you should check the source code and license before installing.
  • The original Chain of Thought will pass through the DSH configured LLM channel; first confirm if the provider/model meets external network access requirements.
  • Verified sources do not provide the star count or the full directory page URL; this article does not cite them.

Links

https://github.com/MeowLynxSea/dsh-cot-summerization