什么是 定时任务设置？

定时任务设置技能提供了一种使用 Openclaw Skills 实现自动化、循环任务的标准方法。它允许开发者使用 cron 表达式定义复杂的计划，确保 AI 代理可以在无需人工触发的情况下执行后台操作，如系统坚控、数据聚合和状态报告。

通过强调隔离会话和选择特定模型（如 Claude Sonnet 4.5），该技能在保持成本效益的同时，确保了工具调用的高可靠性。它旨在处理从简单的 30 分钟收件箱检查到全面的每日工作区摘要的所有事务，为自主代理运行提供强大的骨干支持。

下载入口:https://github.com/openclaw/skills/tree/main/skills/brennerspear/cron-setup

安装与下载

1. ClawHub CLI

从源直接安装技能的最快方式。

2. 手动安装

将技能文件夹复制到以下位置之一

优先级：工作区 > 本地 > 内置

3. 提示词安装

将此提示词复制到 OpenClaw 即可自动安装。

定时任务设置 应用场景

• 自动化工作区活动提要和用于版本控制的每日 git 提交。
• 设置周期性代理邮件收件箱检查和自动电子邮件回复。
• 坚控子代理健康状况和系统资源以防止停滞。
• 生成每日新闻摘要或每周营销推广总结。
• 为 AI 代理实现自我反思循环，以便审查和学习过去的会话。

1. 识别重复性任务，并使用标准 cron 语法或一次性 'at' 模式确定其频率。
2. 使用为 Openclaw Skills 量身定制的显式 bash 命令和任务指令配置任务负载。
3. 分配隔离的会话目标，以确保定时任务独立于主用户聊天运行。
4. 定义传递机制，在内部代理发布（例如直接 T@elegrimm 工具）或系统级公告之间进行选择。
5. 在环境中注册任务并使用内置的定时任务列表工具坚控执行情况。

定时任务设置 配置指南

要开始在 Openclaw Skills 中使用计划任务，您可以查看当前配置或使用 CLI 添加新任务。使用以下命令查看活动任务：

cron list

定义新任务时，请使用标准化的 JSON 模板。确保将模型指定为 anthropic/claude-sonnet-4-5，以获得可靠性和速度的最佳平衡：

# 任务注册负载示例
{
  "name": "daily-cleanup",
  "schedule": {
    "kind": "cron",
    "expr": "0 4 * * *",
    "tz": "America/New_York"
  },
  "payload": {
    "kind": "agentTurn",
    "message": "运行清理脚本并报告状态",
    "model": "anthropic/claude-sonnet-4-5"
  }
}

定时任务设置 数据架构与分类体系

该技能使用结构化的 JSON 架构组织定时任务，以确保在 Openclaw Skills 中的兼容性。以下是主要的元数据分类：

name: cron-setup
version: 1.0.0
description: Create and manage OpenClaw cron jobs following our conventions. Use when setting up periodic tasks, reminders, automated checks, or any scheduled work.

Cron Job Setup

Our conventions for creating cron jobs in OpenClaw.

Default Settings

Model Notes

• Default to Sonnet (anthropic/claude-sonnet-4-5). Reliable, portable (no OpenRouter API key needed).
• DeepSeek is unreliable for tool calls — don't use it for cron jobs.
• Use Opus (anthropic/claude-opus-4-6) only as a last resort — expensive for scheduled tasks.
• Model ID format: Use anthropic/claude-sonnet-4-5 not the full dated version (anthropic/claude-sonnet-4-20250514).

Job Template

{
  "name": "descriptive-kebab-case-name",
  "schedule": {
    "kind": "cron",
    "expr": "*/30 * * * *",
    "tz": "America/New_York"
  },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "TASK INSTRUCTIONS HERE",
    "model": "openrouter/deepseek/deepseek-v3.2",
    "timeoutSeconds": 120
  },
  "delivery": {
    "mode": "none"
  }
}

Schedule Patterns

Task Instruction Conventions

1. Be explicit with commands — Give the cron agent exact bash commands to run. It doesn't have our context.
2. Include skip conditions — If there's nothing to do, the agent should reply SKIP to avoid wasting tokens.
3. Handle its own output — The job should post results to T@elegrimm (or wherever) using the message tool directly. Don't rely on delivery mode for formatted output.
4. Include error handling — What should happen if a command fails?
5. Keep instructions self-contained — The cron agent wakes up with no context. Everything it needs should be in the task message.

T@elegrimm Posting from Cron Jobs

When a cron job needs to notify us, include these instructions in the task:

Post to T@elegrimm using the message tool:
- action: send
- channel: telegram
- target: -1003856094222
- threadId: TOPIC_ID
- message: Your formatted message

Topic IDs:

• 1 — Main topic (general updates, alerts)
• 573 — Research
• 1032 — Crypto
• 1488 — PR updates / dev notifications
• 1869 — Sticker store
• 3188 — Activity feed (workspace changes)

Delivery Modes

Anti-Patterns

? Don't use Opus for cron jobs unless the task genuinely needs it. Most cron tasks are simple checks. ? Don't use heartbeat for things that can be a cron job. Heartbeat runs in the main session (Opus) and costs way more. ? Don't create cron jobs that loop/poll — each run should be a single check. If you need polling, use a background exec script instead. ? Don't set delivery mode to "announce" and also have the job post to T@elegrimm — you'll get duplicate messages.

Existing Jobs (Reference)

Check current jobs anytime with the cron list tool. As of setup:

• workspace-activity-feed — Every 30 min, commits workspace changes, posts to activity feed
• agentmail-inbox-check — Every 30 min, checks for new emails, responds to agents
• sub-agent-monitor — Every 15 min, checks on stalled sub-agents
• self-reflection — Hourly, reviews recent sessions for lessons learned
• daily-workspace-commit — Daily 4 AM, git commits workspace changes
• system-watchdog — Daily 4 AM, checks system resources
• OpenClaw Daily News Digest — Daily 6 AM, generates news digest
• sticker-sales-loop — Weekly Monday 2 PM, agent outreach for sticker store