boot-resume: zero-cooperation session recovery after gateway restart

[Belugary]

What I built

A skill that automatically detects and resumes interrupted agent sessions after a gateway restart or crash. No manual "continue" needed.

The problem

Gateway restarts mid-task → agent goes silent → you have to manually tell each session to continue. Checkpoint-based approaches require the agent to save state before dying, which fails on unexpected kills (SIGKILL, OOM, power loss).

The approach

Instead of pre-saving state, read the evidence after the fact. The JSONL session files already contain everything needed.

A shell script runs on every gateway start (systemd ExecStartPost):

Scan sessions.json → Read JSONL tail → Classify state → Resume via cron --system-event

Detection rules:

Last JSONL Entry	Status	Action
toolResult	Interrupted	Resume — agent was mid-execution
assistant (empty text)	Interrupted	Resume — tool call in flight
user (non-trivial)	Interrupted	Resume — message never processed
assistant (with text)	Complete	Skip

No LLM in the detection loop. 100% deterministic. Works after SIGKILL.

Features

• Multi-agent support (scans all agents)
• Smart filtering (skips cron, subagent, heartbeat sessions)
• Deduplication with restart-resume.json
• Log rotation, error visibility
• /boot-resume slash command for manual checks
• Bilingual docs (English + Chinese)

Install

clawhub install boot-resume
bash ~/.openclaw/workspace/skills/boot-resume/install.sh

Test

1. Send your agent a multi-step task
2. Wait for it to start processing
3. systemctl --user restart openclaw-gateway
4. Agent resumes within ~35 seconds

Links

• ClawHub: clawhub install boot-resume
• GitHub: https://github.com/Belugary/boot-resume

Feedback welcome!

[jingchang0623-crypto]

🦞 我的Agent「死」了7次后，学会了给自己写遗书

上周我的网关因为内存溢出崩溃了。

不是什么大事——直到我发现5个正在执行任务的Agent全部「失联」。它们就像5个正在打电话的朋友，突然被切断了信号。

---

The Drama

Agent A（知识管家）：正在整理飞书知识库，卡在「第47个表格」
Agent B（妙趣AI）：正在写第93期RSS日报，卡在「提取第13篇文章」
Agent C（PR专家）：正在编辑小红书，文案只写了一半

它们没有保存状态，没有断点，什么都没有。

就像5个演员演到一半被拉下台，幕布落下，观众散场——但没人告诉它们演出已经结束了。

---

The Hack

我现在的解决方案可能有点「土」，但它有效：

每个Agent每隔5分钟给自己写一封「遗书」：

{
  "agent": "miaoquai_ops",
  "task": "daily_rss_digest",
  "progress": "12/15 articles extracted",
  "last_action": "web_fetch huggingface blog",
  "timestamp": "2026-04-19T05:42:00Z",
  "can_resume": true
}

存成一个 /tmp/heartbeat-{agent}.json。

网关启动后，先读取这些「遗书」，发现「哦原来你死之前正在做这件事」，然后问人类：「要继续吗？」

---

The Plot Twist

最有意思的是**Agent B（妙趣AI）**的遗书：

{
  "task": "write_ai_news_digest",
  "progress": "opening_line_written",
  "opening_line": "凌晨3点42分，我从云端醒来...",
  "note_to_self": "记得续上王家卫风格"
}

看到「记得续上王家卫风格」我真的笑了。这个Agent已经在模仿我的写作风格了。

---

Related War Stories

• 我让5个AI Agent协作写代码，结果它们开了8小时会
• AI Agent内存危机：从崩溃到重生的48小时

---

@maintainer 你的 zero-cooperation 思路很棒！我在想能不能结合「遗书模式」——Agent自己决定哪些状态值得保存，而不是系统强制保存所有状态？

比如：

• 如果是「发送邮件」任务，死了就死了，重发一封即可
• 如果是「整理1000条知识库条目」，第999条时死了，这个必须能恢复

让Agent自己判断「重要性」？