Preface

The context compaction mechanism in DeepSeek Harness (DSH) is used to handle long conversations by compressing earlier dialogue into checkpoints to save tokens. The official default backend is dsh-compaction-basic, whose summarization prompt is a generic English template and has two main problems: first, it can easily lose exact values (such as version numbers, file paths, command strings, and error strings), which is fatal for quantitative or programming agents; second, it forces English output, causing a fragmented reading experience in Chinese conversations.

dsh-compaction-pro is a high-fidelity, bilingual, recursive context compaction backend plugin. It only overrides the core summarize() hook, while all other pressure/retention/token metering strategies fully inherit the official implementation, serving as an upgraded replacement for dsh-compaction-basic.

Installation and Enablement

Install it using the official command-line tool:

dsh plugin --profile web add dsh-compaction-pro

After installation, you must register it in the configuration file and disable the built-in backend; otherwise, a service conflict may occur.

  1. Open the configuration file ~/.dsh/profiles/web/cordis.patch.yml.
  2. Add the compaction-pro plugin definition.
  3. Explicitly disable compaction-basic.

The example configuration is as follows:

# 1) 禁用内置后端，防止冲突
- id: compaction-basic
  disabled: true

# 2) 注册 dsh-compaction-pro
- id: compaction-pro
  name: 'C:/Users/<you>/dsh-plugins/dsh-compaction-pro/src/index.ts'
  config:
    faithful: true
    recursive: true
    summaryLanguage: auto
    thresholdRatio: 0.8
    retainRatio: 0.16
    maxTokens: 8192

Core Features

The core capabilities of dsh-compaction-pro focus on improving “summarization quality”:

  • High fidelity: Enforces retention of exact values, paths, commands, and identifiers, and never rewrites them.
  • Bilingual support: The summary language follows the conversation language (auto mode) and supports Chinese output.
  • Recursive chunking: For large content segments, it first compresses chunks and then merges them, avoiding information truncation caused by the maxTokens limit.
  • Decision restoration: It not only preserves key decisions but also includes the reasoning behind them and the rejected alternatives.
  • Tool result retention: A dedicated “Key tool outputs” section preserves returned structures, important return values, or file snippets.
  • Custom instructions: You can replace the built-in summarization prompt entirely through the customInstruction parameter.

Configuration Parameters

The plugin behavior is controlled via the YAML configuration file. The main parameters are:

Configuration Type Default Description
faithful boolean true Enables high-fidelity mode and preserves exact values.
summaryLanguage string 'auto' Summary language. Supports 'en', 'zh', or 'auto' (follows the conversation).
recursive boolean true Enables recursive chunking for processing large content segments.
customInstruction string — Custom summarization prompt. Completely overrides the default logic.
thresholdRatio number 0.8 Context window ratio that triggers compaction.
retainRatio number 0.16 Ratio of the recent tail retained in context.
maxTokens number 8192 Maximum token limit for generated summaries.
summarizationProvider / summarizationModel string — Specifies the model used for summarization. If left empty, the conversation routing model is reused.

Notes

  1. Service conflict: This plugin and the built-in dsh-compaction-basic share the same ctx.compaction service interface and cannot be enabled simultaneously. You must explicitly disable compaction-basic in the configuration file.
  2. Early development stage: The plugin is in the dsh v0.1.0-rc.6 development stage. Upstream APIs may contain breaking changes, so pay attention to version updates.
  3. Type definitions: Because some internal types are not officially exported, the plugin’s type definitions mirror the official internal types. If upstream contracts change, the plugin may become invalid or crash.

Summary

dsh-compaction-pro improves the summarization prompt to solve the problems of losing exact values and language fragmentation caused by DSH’s default compaction. By inheriting the official strategies, it achieves effective compression without distortion.

  • Plugin homepage: https://github.com/helibeiqi/dsh-compaction-pro
  • Directory page: https://www.skillhub.cn/plugins/helibeiqi/dsh-compaction-pro