skill
unattended-task-pitfalls
Hardening playbook for AI agent unattended/scheduled tasks · 无人值守自动任务加固经验库
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/mowenQWQ/unattended-task-pitfalls itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 22, 2026unattended-task-pitfalls
无人值守自动任务加固经验库——"操作纪律会失效，只有代码能兜底"。来自两次备份损坏事故 + 一次 504 杀进程事故 + cron 总结误判事故的完整复盘。
A hardening playbook for AI-agent unattended/scheduled tasks — "operational discipline fails unattended; only code has your back." Distilled from two backup-corruption incidents, a 504 process-kill incident, and a cron summarization misjudgment.
⚠️ AI 使用说明：本作品使用 AI 辅助生成（生成时间：2026-09-03）。仅供学习与参考使用，发布、引用或投入使用前请人工核实关键信息。
中文
这是什么
一个 AI agent 工作区的真实事故复盘（2026-08 ~ 09）：
- 08-27：定时备份产出损坏文件并上传云端，教训"gzip -t 验证后才能上传"被记入长期记忆
- 08-30：同款事故复发——因为 cron 会话根本读不到"操作纪律"，纪律只对在线的 agent 有效。528MB 截断坏文件再次上传（正常 911MB）
- 08-24/25：cron 总结任务在隔离会话里误判"全天无对话"（它看不到主会话上下文）
- 09-05：cron 提示词里的裸命令被执行会话照抄前台运行，第三次被网关杀进程——但这次恢复侧零失误（先查进程再重跑，锁自动释放，验证闸门自动拦截残缺包）。新教训：prompt 既是纪律载体，也是风险载体
由此提炼出本库的核心命题：无人值守链路里，操作纪律会失效，只有代码能兜底。
内容结构
| 章节 | 解决什么 |
|---|---|
| §1 核心哲学 | 为什么"记忆型教训"在无人值守场景必然失效，教训落点分级（脚本内 > prompt > 记忆） |
| §2 后台执行 | 前台跑被网关杀、"504 ≠ 命令未执行"（盲目重试 = 并发实例互踩）、锁与幂等、prompt 内嵌命令的安全形态 |
| §3 验证链 | 产物完整性验证（gzip -t）、"验证不过绝不上传"闸门、"已存在跳过"复用坏文件坑 |
| §4 隔离会话 | cron 会话的上下文盲区、"无对话/静默"结论的证据法（含会话初始化产物的排除清单） |
| §5 通知设计 | cron 无条件通知 vs 心跳条件通知、自动任务输出禁用拟人内容、静默档防通知疲劳 |
| §6 交叉验证 | 自动任务的产物也要被自动检查（发现 08-30 事故的唯一渠道）、一次性任务自删、失败的默认出口 |
核心原则速览
- 凡无人值守的纪律，一次性落到代码与 prompt（锁、验证闸门、复查脚本），不要指望 agent"想起来"
- prompt 里内嵌的长命令要直接写成安全形态（
setsid nohup ... &全格式），不指望执行者"记得加壳" - 单一任务的自我报告不可信——给关键产物配独立验证路径
- 网关/管道类超时 ≠ 命令未执行，先查进程再决定重试
- 失败的默认出口 = 记录日志 + 通知用户 + 保留现场，不在失败路径上做主观决策
适合谁
- 给 AI agent（Claude Code / OpenClaw / 自建 agent）搭建 cron / 定时备份 / 自动报告的人
- 遇到过"任务失败但报告成功"、"备份文件损坏"、"定时总结瞎编"的人
- 设计多任务自动化体系、需要通知与静默策略的人
安装
git clone https://github.com/mowenQWQ/unattended-task-pitfalls.git
cp -r unattended-task-pitfalls /path/to/your/agent/skills/
scripts/ 内含产物验证脚本模板（verify_archive.py），可直接适配到你的备份任务。
更多技能库
| 仓库 | 简介 |
|---|---|
| win-dev-pitfalls | Windows 开发踩坑全量主线 |
| bat-ps1-dev | bat/PowerShell 脚本开发专项经验库 |
| Web-Security-Test-Rules | 已授权网站安全测试规则库 |
| agent-mistake-patterns | AI 智能体犯错模式库与自我纠错纪律 |
| grounded-summaries-skill | 防止 AI 在总结任务中编造内容 |
English
What is this
A post-mortem from a real AI-agent workspace (Aug–Sep 2026):
- 08-27: a scheduled backup uploaded a corrupted archive; the lesson ("verify with
gzip -tbefore upload") was stored in long-term memory - 08-30: the same incident recurred — because cron sessions never read "operational discipline"; discipline only binds an online agent. Another truncated 528MB file (expected: 911MB) went upstream
- 08-24/25: a cron summarization task in an isolated session misjudged "no conversations today" (it cannot see the main session's context)
- 09-05: a bare command embedded in the cron prompt was copy-pasted and run in the foreground; the gateway killed the process for the third time — but recovery was flawless this time (check process → rerun in background; lock auto-released; verification gate auto-rejected the truncated artifact). New lesson: a prompt is both a discipline carrier and a risk carrier
The core thesis: in unattended pipelines, operational discipline fails — only code has your back.
Structure
| Section | Covers |
|---|---|
| §1 Core philosophy | Why "memory-stored lessons" inevitably fail unattended; lesson-placement tiers (in-script > prompt > memory) |
| §2 Background execution | Gateway-killed foreground runs, "504 ≠ not executed" (blind retries = concurrent instances racing), locks & idempotency, safe form for commands embedded in prompts |
| §3 Verification chain | Artifact integrity (gzip -t), "never upload without verification" gates, the "skip if exists" reuse-of-corrupt-file trap |
| §4 Isolated sessions | Cron context blindness; evidence-based "no activity" conclusions (with init-artifact exclusion list) |
| §5 Notification design | Unconditional cron vs conditional heartbeat notifications; no anthropomorphic filler in automated output; silence tiers |
| §6 Cross-verification | Automated artifacts deserve automated checks; one-shot task self-deletion; default failure exits |
Quick principles
- Move every unattended discipline into code and prompts once (locks, verification gates, review scripts) — don't count on the agent "remembering"
- Embed long commands in prompts in their safe form (
setsid nohup ... &full format) — don't count on the executor "remembering to wrap it" - Never trust a task's self-report — give critical artifacts an independent verification path
- Gateway/pipeline timeouts ≠ command not executed — check the process before retrying
- Default failure exit = log + notify + preserve the scene; make no subjective decisions on the failure path
Installation
git clone https://github.com/mowenQWQ/unattended-task-pitfalls.git
cp -r unattended-task-pitfalls /path/to/your/agent/skills/
scripts/ includes a reusable artifact-verification template (verify_archive.py).
License
MIT