session-end

🌙 session-end

A structured closing ritual for AI coding sessions.
Stop losing context. Stop repeating yourself next conversation.

---

每次对话结束时，AI 会忘记一切。session-end 是一套 6 步收尾流程——进度、经验、待办全部写入记忆，下次对话开口就能接着来。

When the session ends, AI forgets everything. session-end is a 6-step closing ritual — progress, lessons, and todos are saved so the next conversation picks up right where you left off.

---

What it does / 做什么

/session-end

Step	EN	中文
1	✅ Task Audit	逐条核查本次任务完成情况
2	🔍 Retrospective	找出失误（不是总结成果）
3	🧠 Decide what to update	判断哪些记忆/规则需要更新
4	📝 Execute updates	执行记忆、CLAUDE.md、Skill 更新
5	✨ Memory compression	压缩整理记忆文件（Claude Code: dream skill）
6	📋 Next session plan	写下下次对话的 Top 5 待办

The retrospective step is intentional: find mistakes, not just celebrate wins. Known pitfalls walked into again, wrong tool choices, undefended scripts — all get surfaced and written to memory so they don't repeat.

复盘步骤刻意聚焦于错误：已知的坑又踩了、工具选择绕路了、脚本没有防御性——全部记录，下次不再重犯。

---

Install / 安装

Claude Code (plugin — recommended)

claude plugins install github:zeshuochen/session-end

Then invoke with /session-end.

To also install the automatic trigger hooks, merge hooks/hooks.json into your ~/.claude/settings.json under the "hooks" key.

Claude Code (manual)

mkdir -p ~/.claude/skills/session-end
curl -o ~/.claude/skills/session-end/SKILL.md \
  https://raw.githubusercontent.com/zeshuochen/session-end/main/skills/session-end/SKILL.md

Other agents / 其他 Agent

The skill is plain markdown — paste skills/session-end/SKILL.md into your agent's context, system prompt, or skill system as-is. Steps 1–4 and 6 are fully agent-agnostic. Step 5 references dream, a Claude Code–specific skill; skip or replace with your own memory tool.

---

Hooks / 自动触发（可选）

The hook watches every message for trigger phrases and injects structured instructions automatically.

Trigger phrase	Effect
end session / 结束对话	Injects instruction to invoke /session-end immediately

Install: merge the contents of hooks/hooks.json into your ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python \"${CLAUDE_PLUGIN_ROOT}/hooks/trigger.py\"",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

The hook requires Python 3 (stdlib only, no dependencies).

---

Memory system / 记忆系统

session-end is most powerful paired with a persistent memory structure:

~/.claude/
├── CLAUDE.md                    ← global rules and conventions
└── projects/<project>/
    └── memory/
        ├── MEMORY.md            ← index of all memory files
        ├── project_*.md         ← project state
        ├── feedback_*.md        ← lessons and corrections
        └── reference_*.md       ← external resource pointers

Each session-end run updates these files. The start session hook reads them back on next open. No context lost.

---

Agent compatibility / 兼容性

Agent	Skill	Hooks	Notes
Claude Code	✅ /session-end via plugin	✅ UserPromptSubmit	Full support incl. dream compression
Gemini CLI	✅ via activate_skill	⚠️ Adapt format	See gemini-extension.json
Copilot CLI	✅ Paste as context	❌ No hook system	Manual invoke only
Cursor	✅ Add to .cursorrules	⚠️ See .cursor-plugin/	Manual invoke
Codex / others	✅ Paste as context	❌	Manual invoke

---

License

MIT © Zeshuo Chen