DSH Compaction Weave v0.3.2-rc.1

简体中文

DSH Compaction Weave 为 DeepSeek Harness 提供上下文压缩，让长对话保留有效要求、当前进度和下一步，并通过增量摘要尽量复用已有缓存前缀。

本版内容

• 增量压缩与检查点合并，结合近期原话预算和低信息内容裁剪。
• 要求修订、取消与范围检查；保存工作报告及来源，支持历史和附件回查。
• 分页摘要、持久化接续、失败重试，以及对话中的错误原因和诊断位置。
• 摘要模型与重试参数设置页，完整中英文 README 和使用文档。
• 修复进度报告、否定要求、未来动作及跨批次未知状态引起的摘要误判，加入 7 份真实失败摘要重放。

测试与压缩效果

• 发布前本地验证：50 套 / 1359 项全部通过，含 5 套 / 93 项 DSH 宿主集成测试。
• GitHub Actions：Windows 与 Linux 的离线回归、项目检查和打包验证通过。
• 最近一轮官方接口实验使用 0.3.1-review.35 与 DeepSeek V4.1 Flash：4 次成功压缩的主对话输入平均减少 68.50%，压缩后平均保留 31.50%，取 3 次增量和 1 次合并的等权平均。
• 三次增量压缩后第 6 答缓存命中率分别为 98.31% / 99.11% / 99.20%；合并后观察到第 24 答为 96.08%。

逐次 token 数据、统计口径、失败记录与后续修复见 测试记录。

安装

适配 DSH 0.2.2-rc.1 / Node.js 24.9+。下载下面的 .tgz 文件，在 DSH 的“插件 → 添加插件”中输入安装包绝对路径，安装启用后重启；命令行也可安装：

dsh plugin --profile web add https://github.com/NightFrost42/dsh-compaction-weave/releases/download/v0.3.2-rc.1/dsh-compaction-weave-0.3.2-rc.1.tgz

将 web 替换为正在使用的 profile。源码包、SHA-256 校验清单和逐文件清单一并附在下方。

遇到报错、回答质量下降（“降智”），或压缩后忘记要求、进度和任务信息，欢迎 提交 Issue。请附版本、模型、复现步骤和相关对话片段；报错附错误码与诊断文件，提交前移除密钥和私人信息。

---

English

DSH Compaction Weave compacts context for DeepSeek Harness, retaining active requirements, current progress, and next steps while reusing stable prefixes through incremental summaries.

What's included

• Incremental compaction and checkpoint consolidation, recent-quote budgets, and low-information pruning.
• Requirement updates, cancellations, and scope checks; progress-source tracking and history/attachment lookup.
• Paged summaries, persistent continuation, configurable retries, and actionable error diagnostics.
• Summary-model and retry settings, with complete English and Chinese documentation.
• Fixes for summary misclassification around progress reports, negated requirements, future actions, and cross-batch uncertainty, including 7 complete real-summary regression replays.

Validation and measurements

• Local publication preflight: 50 suites / 1359 checks passed, including 5 DSH integration suites / 93 checks.
• GitHub Actions: offline regression, source checks, and release packaging passed on Windows and Linux.
• The latest official-API experiment used 0.3.1-review.35 with DeepSeek V4.1 Flash: across 4 successful compactions, main-conversation input fell by an average of 68.50%, with 31.50% retained, using the arithmetic mean of 3 incremental compactions and 1 consolidation.
• Cache hit rates on reply 6 after the three incremental compactions were 98.31% / 99.11% / 99.20%; observation after consolidation reached 96.08% on reply 24.

See Testing for per-compaction counts, methodology, failure records, and subsequent fixes.

Install

Use DSH 0.2.2-rc.1 and Node.js 24.9+. Download the .tgz asset, enter its absolute path under Plugins → Add plugin, enable it, and restart DSH. The command above also installs into a CLI profile; replace web with your active profile name.

Assets include the plugin, source archive, SHA-256 checksums, and a per-file manifest.

For errors, degraded answer quality, or requirements, progress, and task details forgotten after compaction, please open an Issue. Include versions, models, reproduction steps, relevant excerpts, and any error code or diagnostic file. Remove credentials and private information before posting.