v0.1.0 — First public release

dsh-self-improvement v0.1.0

First public release. A DeepSeek Harness plugin for cross-session self-improvement: it captures what each session's workspace learns and feeds it back into later sessions, instead of letting that knowledge vanish at session end.

Highlights

• Persistent memory per workspace — survives restarts and session switches. Five kinds (fact / preference / lesson / resource / method), one file per kind.
• Importance-weighted injection — entries are ranked by importance + recency + hit count; when the budget runs out they are dropped by score, line-wise, with a trace — never head-truncated and never silently discarded.
• Retrospectives — immediate on agent/error (debounced and rate-limited) and again at session end; the output becomes memory, SOPs and a handoff brief for the next session.
• Sleep-time consolidation — while idle, induces higher-level principles, drafts hypotheses to verify and prepares prefetch points.
• SOP library as a real skill — reusable methods land in playbooks/ and are registered as a skill (progressive disclosure) with usage statistics.
• Configurable reuse scope — workspace (default, fully isolated), global (one shared library for all workspaces) or both (local accumulation plus a global supplement, promoting general entries at session end). Switching scope reports exactly which layer becomes invisible.
• Trust model — memory is graded by who wrote it, not what it claims: imperative statements are quarantined on write, unverified entries are capped when scored, and there is a final check at injection time.
• Improvement proposals — the plugin's suggestions about itself go through a state machine and require human approval; rejection reasons are written back to memory so the same advice is not repeated.
• Web panel — memory, queue, cost and proposals above the composer, with scope switching.
• Crash-safe — incremental snapshots while running, automatic recovery of unfinished sessions on startup, at-least-once retrospective queue with backoff and a dead-letter queue.
• Zero context cost — memory rides the system prompt; it consumes no conversation turns.

Install

dsh plugin add dxxCaO/dsh-self-improvement

Or place the directory under ~/.dsh/plugins-src/, where cordis.patch.yml registers it into the profile layer stack, then restart dsh web.

Requires Node ^22.19.0 || >=24.0.0, with @deepseek-ai/cordis ^4.0.1 and @deepseek-ai/dsh-tools >=0.0.1-rc.1 <0.2.0 as peers.

Quick start

The plugin starts working as soon as it loads. Useful entry points:

Command	Purpose
/selfip	Show the current reuse scope and config file locations
/selfip scope both	Switch the reuse scope
/memory <keyword>	Search memory
/forget <keyword>	Invalidate matching entries
/retro	Run a retrospective on the current session now
/promote	Promote general entries to the global library
/proposals	List and resolve pending improvement proposals

Data layout

Everything lives under the workspace itself, so it travels with the project and is easy to inspect:

<workspace>/self-improvement/
  config.json          workspace config
  memory/              memory (+ quarantine/)
  logs/                session logs, snapshots, handoff, pending queue, diagnostics
  playbooks/           SOPs
  proposals/           improvement proposals

---

dsh-self-improvement v0.1.0（中文）

首个公开版本。一个给 DeepSeek Harness 的跨会话自我改进插件：把每个会话工作区学到的东西沉淀下来并回灌到后续会话，而不是让这些经验在会话结束时消失。

主要能力

• 按工作区的持久记忆——重启、换会话都不丢。五类记忆（fact / preference / lesson / resource / method）分文件存放。
• 按重要性加权注入——按 重要性 + 新鲜度 + 命中次数 排序；预算耗尽时按分数逐行淘汰并留痕，绝不裁头、绝不静默丢弃。
• 复盘——出错时即时复盘（防抖 + 频率上限），会话结束时再复盘一次；产物变成记忆、SOP 与给下次会话的交接简报。
• 睡眠期巩固——空闲时归纳更高层原则、生成待验证假设、准备预取要点。
• SOP 作为真正的 skill——可复用方法落进 playbooks/，注册为 skill（渐进披露）并带使用统计。
• 可配置的经验复用范围——workspace（默认，完全隔离）、global（所有工作区共用一份库）或 both（本工作区积累 + 全局库补充，会话结束时提升通用条目）。切换范围时会明确告知哪一层变得不可见。
• 信任模型——按谁写的而非谁声称的分级：指令性语句写入即隔离，未验证条目评分封顶，注入期再做一次检测。
• 改进提案闭环——插件对自己的建议走状态机、需人工确认；否决理由回写记忆，避免重复提同类建议。
• Web 面板——输入区上方展示记忆、队列、成本与提案，可切换范围。
• 防丢失——运行中增量快照、启动时自动恢复未完成会话、至少一次语义的复盘队列（含退避与死信队列）。
• 零上下文成本——记忆走系统提示，不占用会话内对话轮次。

安装

dsh plugin add dxxCaO/dsh-self-improvement

或将本目录放入 ~/.dsh/plugins-src/（由 cordis.patch.yml 注册进 profile 层栈），然后重启 dsh web。

要求 Node ^22.19.0 || >=24.0.0，peer 依赖 @deepseek-ai/cordis ^4.0.1 与 @deepseek-ai/dsh-tools >=0.0.1-rc.1 <0.2.0。

快速上手

插件加载后自动开始工作。常用入口：

命令	用途
/selfip	查看当前经验范围与配置文件位置
/selfip scope both	切换经验复用范围
/memory <关键词>	检索记忆
/forget <关键词>	作废匹配的条目
/retro	立即对本会话做一次复盘
/promote	把通用条目提升到全局库
/proposals	查看/处理待批的改进提案

数据位置

全部落在工作区内部，随项目走、便于直接查看：

<工作区>/self-improvement/
  config.json          本工作区配置
  memory/              记忆（含 quarantine/ 隔离区）
  logs/                会话日志、快照、交接简报、待复盘队列、自诊断
  playbooks/           SOP
  proposals/           改进提案

---

Full documentation: README.md · README.en.md · MIT