AI Agent Hub
Back to plugins
dsh-think-summary preview

dsh-think-summary

Model Inference Updated 2026.08.19

Run the following command in DeepSeek Harness:

dsh plugin install oakcakerolls/dsh-think-summary

Paste the following prompt into your AI chat to install this plugin:

Install this plugin in DeepSeek Harness by running dsh plugin install oakcakerolls/dsh-think-summary; the source repository is available at https://github.com/oakcakerolls/dsh-think-summary.

About this plugin

When a model unlocks extended thinking, thousands or even tens of thousands of tokens of raw chain-of-thought flood the conversation stream. The result is unreadable, unstructured, and impossible to revisit after the fact. dsh-think-summary tackles exactly that: it watches the reasoning stream in real time, segments it at natural semantic boundaries rather than arbitrary byte cuts, and produces per-segment summaries so the model's internal deliberation becomes a trackable, reviewable record you can reference in future turns.

The pipeline runs through four stages: detect, segment, summarize, and refine. Detection uses a CJK-aware lightweight token counter that kicks in once a configurable threshold is crossed, while silently filtering out auxiliary calls such as title generation or compression to avoid false triggers. Segmentation is Markdown-structure-aware: fenced code blocks stay atomic, tables are never split mid-row, lists break only at item boundaries, and headings or structural markers serve as strong split signals. Summarization first yields an instant zero-cost heuristic result, then optionally refines eligible segments using the session provider's smallest available model, with configurable input trimming, a bounded concurrency pool, and automatic fallback on timeouts. Code blocks and tables support a three-state policy of ignore, keep-and-skip, or keep-and-refine, so they never pollute the summary surface by default. Results appear in three layers: a live scrolling panel above the composer, a collapsible summary bar at the tail of each reply, and a dedicated session-level tab listing every recorded segment with dual token counts. Nothing is written back into the conversation context, and persistence uses atomic writes so your history survives restarts.

This plugin is built for developers and power users who regularly steer extended-thinking models and need to inspect, audit, or share the reasoning process. If you ask deep inference questions, want a post-hoc replay of what the model actually thought, or prefer a clean distilled summary over a wall of raw tokens, dsh-think-summary is made for you.

Screenshots

Use Cases

  • Watch per-segment summaries of a model's reasoning unfold live above the composer
  • Replay and audit a model's chain-of-thought after the fact to pinpoint key decision points
  • Share clean distilled summaries with your team instead of scrolling raw token walls

Best For

  • Developers who work daily with extended-thinking models like o1 or R1
  • Engineering teams that need to audit or explain model reasoning to stakeholders
  • AI application builders who want readable, shareable records of model thought processes