Cron Helper

OpenClaw cron 定时任务配置、诊断和修复。

When to Use

• User says "配置cron", "定时任务", "周期性任务", "每天X点"
• Cron jobs not firing or not sending messages
• Discussing heartbeat vs cron
• Debugging scheduled task failures

⚠️ CRITICAL: Cron Is Stateless — Design for Persistence

Every cron job fires in a brand-new isolated session with ZERO memory of previous runs.

The agent has no conversation history, no memory of what it did last time, and CANNOT ask the user questions. This means:

❌ Broken Cron Pattern (无状态 = 无用功)

Prompt: "检查邮件，看看有没有重要的"

→ Agent checks email, finds nothing, outputs "没有新邮件"
→ Nothing saved. No state recorded.
→ Next run: same check, same "没有新邮件", forever identical
→ If something WAS important last run, nobody knows now

✅ Correct Cron Pattern (落盘 = 有状态 = 跨次累积)

Prompt: "检查未读邮件。将新邮件摘要追加到 workspace/memory/email-log.md。
        如果发现标记为紧急的邮件，立即发送通知到飞书群。
        上次检查过的邮件ID记录在 workspace/memory/email-cursor.json。"

→ Agent reads cursor to know where it left off
→ Agent checks NEW emails since cursor
→ Agent writes updated cursor
→ Agent appends summary to log
→ Next run: picks up from cursor, never re-processes

Statefulness Design Rules

Every cron prompt MUST follow at least one of these patterns:

Mandatory State Checklist (写在 prompt 里)

Before writing ANY cron prompt, ensure it answers:

• Where does this run save its results? (file path)
• How does the next run know what the previous run did? (cursor/log/state file)
• What happens if the agent can't decide something? (must have a default action, CANNOT ask user)
• Is the output actionable even without human follow-up? (if not, the prompt is incomplete)

⚠️ Critical: Timeout Must Be ≥ 300s

CLI 默认 timeout=30s，这对任何 real task 都太短了。每次创建 cron job 时必须显式设置 timeout。

必须在 payload 和 CLI 中同时指定：

# CLI 用 --timeout-seconds 参数
openclaw cron add --timeout-seconds 600 ...

# payload 中也要写
"payload": { "kind": "agentTurn", "message": "...", "timeoutSeconds": 600 }

经验教训：ceo-agent 的每日随想分析任务默认 300s，前2天跑成功了，但数据积压后持续超时4天。改为 1200s 后解决。

Prompt Self-Sufficiency Rules

A cron prompt is a standalone instruction to an amnesiac agent. It must contain:

1. Complete context — Don't assume the agent knows anything from previous sessions
2. Explicit file paths — Where to read state, where to write state
3. Decision authority — The agent must be able to decide and act autonomously
4. Fallback behavior — What to do when things go wrong (not "ask the user")

Good prompt template:

执行 [任务描述]。
- 读取上次状态：[文件路径]
- 执行检查/操作：[具体步骤]
- 保存本次状态：[文件路径]
- 如果发现 [异常情况]，[自动处理方式]
- 将结果追加到 [日志路径]

Bad prompt:

检查一下系统状态  ← 太模糊，没有落盘，没有状态

Cron vs Heartbeat

Choose Cron when: precise timing, self-contained task, no conversation context needed Choose Heartbeat when: bulk checks across areas, needs recent conversation context

Cron Job Structure

Required Fields

{
  "id": "daily-report-001",
  "agentId": "report-agent",
  "name": "每日报告",
  "enabled": true,
  "schedule": { "kind": "cron", "expr": "0 10 * * *", "tz": "Asia/Shanghai" },
  "payload": { "kind": "agentTurn", "message": "..." },
  "delivery": { "mode": "announce", "channel": "feishu", "to": "chat:oc_xxx" }
}

Every job MUST have: id, agentId, name, enabled, schedule, payload.kind="agentTurn", payload.message, delivery.channel, delivery.to.

Cron Expression Format

┌────────── minute (0-59)
│ ┌────────── hour (0-23)
│ │ ┌────────── day of month (1-31)
│ │ │ ┌────────── month (1-12)
│ │ │ │ ┌────────── day of week (0-6, Sun=0)
│ │ │ │ │
* * * * *

Common patterns:

• 0 10 * * * — Daily at 10:00
• */5 * * * * — Every 5 minutes
• 0 */2 * * * — Every 2 hours
• */30 0-8 * * * — Every 30min during 0:00-8:30 (night hours)
• 0 9,17 * * * — Twice daily at 9AM and 5PM

Common Fatal Errors

Safety: NEVER Modify Jobs Without These Steps

1. Backup first: cp ~/.openclaw/cron/jobs.json ~/.openclaw/cron/jobs.json.backup-$(date +%Y%m%d-%H%M%S)
2. Precise matching: Only modify EXACTLY what user asked (not all jobs)
3. Test jq filter before applying: Run read-only first to verify scope
4. Write to temp file, validate, then atomic mv (never jq redirect to same file)
5. Validate after write: See scripts/validate-jobs-syntax-v2.sh

Precise Matching Rules

Prompt Design Examples

Example 1: Stateful Daily Report

{
  "payload": {
    "kind": "agentTurn",
    "message": "生成今日学习报告。1) 读取 memory/learning-queue.json 获取待研究主题队列。2) 研究队首主题，撰写笔记保存到 learning/YYYY-MM-DD.md。3) 从队列中移除已完成主题，追加新发现到队列。4) 将报告摘要发送到飞书群。如果队列为空，主动从 trending topics 中选择一个加入队列。"
  }
}

Example 2: Idempotent Health Check

{
  "payload": {
    "kind": "agentTurn",
    "message": "检查服务健康状态。读取 memory/last-health-status.json 作为基线。对每个服务执行健康检查。如果状态与基线不同：更新基线文件，发送变更通知到飞书群。如果状态相同且全部正常：更新 last-check 时间戳，不发消息（避免噪音）。如果连续3次失败：发送告警并记录到 memory/alert-log.md。"
  }
}

Example 3: Broken (Anti-pattern)

{
  "payload": {
    "kind": "agentTurn",
    "message": "检查系统状态"
  }
}

Why broken: No state file, no decision authority, no persistence. Every run produces identical output.

Diagnostic Workflow

# 1. Check gateway status
openclaw gateway status

# 2. Validate jobs.json syntax
cat ~/.openclaw/cron/jobs.json | jq .

# 3. Run validation script
~/.openclaw/skills/cron-helper/scripts/validate-jobs-syntax-v2.sh

# 4. Check gateway logs for cron errors
openclaw gateway logs | grep -i cron

For detailed troubleshooting, see references/troubleshooting.md. For full syntax specification, see references/syntax-guide.md. For validation scripts, see scripts/.

New Job Creation Template

{
  "id": "TODO-unique-id",
  "agentId": "TODO-agent-id",
  "name": "TODO-任务描述",
  "enabled": true,
  "schedule": { "kind": "cron", "expr": "TODO * * * *", "tz": "Asia/Shanghai" },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "TODO-完整自包含指令：读取[状态文件]→执行[操作]→保存[状态文件]→如果[异常]则[自动处理]",
    "timeoutSeconds": 300
  },
  "delivery": { "mode": "announce", "channel": "feishu", "to": "chat:oc_TODO" }
}

Before submitting: Run scripts/validate-jobs-syntax-v2.sh on the new config.