MeowLynxSea/dsh-cot-summerization
30Last commit Aug 16, 2026
dsh-cot-summerization DSH plugin
This DSH plugin replaces the raw chain-of-thought output from reasoning models with a concise, polished summary. It intercepts the reasoning stream, sends it to a summarizer model, and displays only the summary in the UI. The original reasoning is preserved in the model's history for multi-turn inference, but never exposed to the user.
How to install the dsh-cot-summerization DSH plugin
dsh plugin add github:MeowLynxSea/dsh-cot-summerizationCopying does not run this command. Review the repository and version before installing the dsh-cot-summerization DSH plugin.
dsh-cot-summerization DSH plugin data source
dsh-cot-summerization DSH plugin snapshot date: Aug 16, 2026
discovered
What the dsh-cot-summerization DSH plugin can do
- Streaming chunked summarization with configurable chunk size and interval
- Near-duplicate detection using bigram Dice coefficient (threshold 0.65) and longest common substring
- Support for custom summarization styles (concise, descriptive, wenyan, custom) and language
- Adaptive chunking that adjusts to real-time stream rate and summarizer RTT
- Optional typewriter mode for character-by-character display of summaries
Where the dsh-cot-summerization DSH plugin fits
- Hide lengthy reasoning chains from end users while maintaining model performance
- Create a cleaner, more professional UI for AI chat applications
- Protect the model's internal reasoning process from being directly visible to users
- Enable a 'mystery' experience similar to closed-source models
- Reduce visual clutter in debugging or logging by replacing CoT with compact summaries
Who the dsh-cot-summerization DSH plugin is for
- Developers building open-source AI chat interfaces with DSH
- Users who prefer a cleaner output without seeing raw reasoning steps
- Teams wanting to offer a premium 'hidden reasoning' experience while staying fully open-source
dsh-cot-summerization DSH plugin limitations
- Requires an additional API call to a summarizer model, increasing cost and latency
- Short reasoning chains (under 32 characters) are passed through untouched
- Non-reasoning models (without reasoning streams) are not affected
- Depends on DSH's LLM channel configuration; summarizer model must be accessible
- Typewriter mode serializes the entire stream, causing output to be queued
dsh-cot-summerization DSH plugin: from the repository README
Quoted from the MeowLynxSea/dsh-cot-summerization README, the upstream source of the dsh-cot-summerization DSH plugin. Copyright remains with the original authors.
**DeepSeek Harness 插件:把模型的原始思维链拦截在半空,改写成一份体面的摘要再端给你。** 闭源模型花钱隐藏思维链,开源模型免费全裸。 我们识别出了开源生态最后一块体验短板,并亲手为它补上—— 方向可能和大家期待的不太一样,但确实是补上了。 </div> <img src="assets/divider-wave.svg" width="880" alt=""> ## 🎯 项目使命 <img src="assets/logo.svg" width="200" align="right" hspace="16" alt="logo:被锁住的眼睛"> 长期以来,「隐藏思维链」是闭源大厂的专属奢华配置:不可关闭、不可审计、不可验证,但很贵。开源模型则把每一步推理都摊开在你面前——坦诚,透明,像一份没整理过的草稿纸。 **我们相信这不应该是一道单选题。** 本插件以工业级严谨,在完全开源、完全可审计的 harness 里,复刻了闭源生态最具标志性的体验:让你看不见思维链。不同的是,我们的隐藏是开源的、可配置的、可关闭的——业界首次把「不可见」做得如此可见。 三大支柱: | 支柱 | 立场 | | :--- | :--- | | 🛡️ **模型隐私权** | 思维链是模型的内心世界。你不会翻别人的日记,除非付费。 | | 🧹 **视觉卫生** | 原始 CoT 又长又乱,充满「等等,我再想想」。得体的人不看这些。 | | 💎 **神秘感平权** | 闭源模型靠神秘感溢价。现在开源用户也能免费获得——本地酿造,纯手工。 | <br clear="right"> ## ✨ 核心特性:宣传名与现实对照 以下每一项,源码均可查。这很讽刺,我们接受。 | 宣传名 | 现实 | | :--- | :--- | | **Information Never Existed™** | 原始 reasoning 增量在 `llm/stream` 拦截层被直接吞掉——UI、流式 chunk、落地 transcript 三处均无原文。模型可见历史除外:那里有一份 model-only 的 surface 替换事件,悄悄把原文还给了模型(推理性能要紧)。 | | **夏日渐进蒸馏** | 流式分段摘要:每凑满 `chunkChars`(500 字)或 `chunkIntervalMs`(8 秒)触发一次,切分点优先落在句边界;开启 `adaptiveChunk` 后,分块大小会随流速率与总结器 RTT 动态缩放。 | | **BioGram™ 重叠度量** | bigram Dice 系数,阈值 0.65。该阈值经历了 0.8 → 0.7 → 0.65 的科学调参过程。 | | **连续核心追踪引擎** | 最长公共子串,滚动数组实现,O(n
Read the full READMERepository license: MIT
dsh-cot-summerization DSH plugin questions
Can I still see the original reasoning chain if I need to?
Yes, but only through specific configurations. If you set `onError` to `pass-through`, the original chain will be shown when summarization fails. Additionally, the raw reasoning is stored in the model history via a model-only surface replacement event (unless you disable `preserveRawForModel`). However, the default behavior is to hide it from the UI and transcript.
What happens if the summarizer model fails?
By default, the plugin shows a placeholder (hide) when the summarizer fails. You can change this with the `onError` setting to `pass-through` to let the original reasoning chain through. The plugin also has a timeout setting (default 30s) to prevent hanging.
Does this plugin send my data to external servers?
The reasoning chain is sent to the LLM provider configured in DSH's own LLM channel. If you configure a local model (e.g., via Ollama), the data never leaves your machine. The plugin does not introduce any additional network calls beyond the DSH framework.
How does this differ from simply folding the reasoning chain in the UI?
Folding only hides the content visually; the raw data is still present in the HTML and transcript. This plugin intercepts the reasoning at the stream level, so the original text never reaches the UI or transcript. The model's history does contain the original chain for performance, but the user-facing output is completely replaced.
Can I use this plugin with non-reasoning models?
No, it only affects reasoning models that produce a `reasoning` stream. If the model does not output a reasoning chain, the plugin passes the stream through unchanged.