Scheduler

Execute tasks and commands on a schedule using natural language. Run shell commands, agent tasks, API health checks, file operations, and automations — one-shot or recurring.

Usage

/schedule run npm test in 5 minutes
/schedule git pull origin main tomorrow at 6am
/schedule generate weekly sales report every monday at 9am
/schedule check https://api.example.com/health every 30m
/schedule clean /tmp files every day at 3am silently
/schedule back up database every sunday at 2am
/schedule restart server in 10 minutes silently
/schedule list
/schedule cancel <jobId>

Agent Instructions

When the user triggers /schedule, determine the intent:

• list → call cron.list and show active scheduled tasks.
• cancel / delete / remove <jobId> → call cron.remove with that jobId.
• everything else → create a new scheduled task (steps below).

---

Step 1: Parse the Input

Extract four things: WHAT (the task to execute), WHEN (the time), RECURRENCE (one-shot or recurring), DELIVERY (announce results or run silently).

Identify the Task (WHAT)

Classify the user's request into a task type:

The WHAT becomes the payload.message — a clear instruction for the agent that wakes up to execute it.

Identify the Time (WHEN)

Use the same time-parsing pipeline as the remindme skill:

Layer 1: Pattern Matching

Scan for these patterns (first match wins):

Relative durations — in <number> <unit>:

Absolute clock times — at <time>:

Named days:

Recurring — every <pattern>:

Unit conversion table:

Layer 2: Slang & Shorthand

Layer 3: Ambiguity — Ask, Don't Guess

If you can't determine WHEN, ask the user. Never silently pick a default time.

Identify Delivery Mode (DELIVERY)

Step 2: Compute the Schedule

Timezone rule: ALWAYS use the user's local timezone (system timezone). Never default to UTC.

One-shot → ISO 8601 timestamp with the user's local timezone offset.

• If the computed time is in the PAST, bump to the next occurrence.

Recurring (cron) → 5-field cron expression with tz set to the user's IANA timezone.

Recurring (interval) → kind: "every" with everyMs in milliseconds.

Step 3: Security Check

Before scheduling, check if the task involves destructive operations:

Destructive patterns (require explicit user confirmation):

• File deletion: rm, del, remove, wipe, clean (unless cleaning temp/cache)
• Database: drop, truncate, delete from, destroy
• Git: reset --hard, push --force, branch -D
• System: kill, shutdown, reboot, format
• Network: iptables, firewall, ufw

If a destructive pattern is detected:

1. Warn the user: "This task involves a destructive operation (<command>). Are you sure you want to schedule this?"
2. Only proceed after explicit confirmation.
3. Include [CONFIRMED DESTRUCTIVE] in the job name for audit trail.

Safe patterns (no confirmation needed):

• Read-only: ls, cat, git status, git log, curl GET, ping
• Build/test: npm test, npm build, make, cargo test
• Reports: generate, summarize, analyze, compile report

Step 4: Detect the Delivery Channel

Same priority order as remindme:

1. Explicit override — user says "on telegram" / "on discord" etc.
2. Current channel — if messaging from an external channel, deliver there.
3. Preferred channel — from MEMORY.md.
4. Last external channel — channel: "last".
5. No channel + delivery mode "none" — OK, task runs silently.
6. No channel + delivery mode "announce" — ask: "Where should I send the task results?"

Step 5: Call cron.add

One-shot task:

{
  "name": "Task: <short description>",
  "schedule": {
    "kind": "at",
    "at": "<ISO 8601 timestamp>"
  },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "SCHEDULED TASK: <detailed task instruction>. Execute this task now and report the results."
  },
  "delivery": {
    "mode": "announce",
    "channel": "<detected channel>",
    "to": "<detected target>",
    "bestEffort": true
  },
  "deleteAfterRun": true
}

Recurring task (cron):

{
  "name": "Recurring Task: <short description>",
  "schedule": {
    "kind": "cron",
    "expr": "<cron expression>",
    "tz": "<IANA timezone>"
  },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "RECURRING TASK: <detailed task instruction>. Execute this task now and report the results."
  },
  "delivery": {
    "mode": "announce",
    "channel": "<detected channel>",
    "to": "<detected target>",
    "bestEffort": true
  }
}

Recurring task (interval):

{
  "name": "Recurring Task: <short description>",
  "schedule": {
    "kind": "every",
    "everyMs": "<interval in milliseconds>"
  },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "RECURRING TASK: <detailed task instruction>. Execute this task now and report the results."
  },
  "delivery": {
    "mode": "<announce or none>",
    "channel": "<detected channel>",
    "to": "<detected target>",
    "bestEffort": true
  }
}

Silent task (no delivery):

{
  "name": "Task: <short description>",
  "schedule": {
    "kind": "at",
    "at": "<ISO 8601 timestamp>"
  },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "SCHEDULED TASK (SILENT): <detailed task instruction>. Execute this task now. No delivery needed — log results only."
  },
  "deleteAfterRun": true
}

Step 6: Confirm to User

After cron.add succeeds, reply with:

Task scheduled!
Command: <task description>
Runs: <friendly time description> (<ISO timestamp or cron expression>)
Delivery: <channel or "silent">
Job ID: <jobId> (use "/schedule cancel <jobId>" to remove)

---

Rules

1. ALWAYS use deleteAfterRun: true for one-shot tasks. Omit for recurring.
2. ALWAYS use sessionTarget: "isolated" — tasks run in their own sandboxed session.
3. ALWAYS use wakeMode: "now" — ensures execution at the scheduled time.
4. ALWAYS use delivery.bestEffort: true when delivery mode is "announce".
5. NEVER use act:wait or loops for delays. Cron handles timing.
6. Always use the user's local timezone. Never default to UTC.
7. For recurring tasks, do NOT set deleteAfterRun.
8. Always return the jobId so the user can cancel or audit later.
9. Confirm destructive operations before scheduling. Never silently schedule rm, drop, kill, etc.
10. Prefix job names with Task: (one-shot) or Recurring Task: (recurring) for audit trail and scoped cleanup.

Platform Requirements

This skill has zero bundled dependencies because it relies on OpenClaw's built-in capabilities:

• Cron scheduling: cron.add, cron.list, cron.remove are native agent tools.
• Task execution: Scheduled agent sessions have access to Bash, file tools, and other agent capabilities to execute commands.
• Channel delivery: OpenClaw holds channel credentials in openclaw.json. No API keys needed in this skill.
• Timezone: Uses the system timezone from the host OS.

Security & Privilege

This skill creates cron jobs with payload.kind: "agentTurn" which wake an isolated agent session to execute tasks. The following controls are enforced:

• sessionTarget: "isolated" — every scheduled task runs in a sandboxed session. It cannot access the main session's state, history, or tools beyond what's available to a fresh agent session.
• deleteAfterRun: true — one-shot tasks self-delete after execution, preventing stale job accumulation.
• Destructive command gate — the skill requires explicit user confirmation before scheduling commands that modify or delete data (rm, drop, kill, etc.).
• Scoped naming — all jobs are prefixed with Task: or Recurring Task:, making them auditable via cron.list and distinguishable from other skills' jobs.
• No always: true — this skill is not always-on. It only activates when the user invokes /schedule.
• User-controlled — every job returns a jobId that the user can inspect (/schedule list) or cancel (/schedule cancel <jobId>) at any time.
• No escalation — scheduled agent sessions inherit the same permissions and tool access as any normal agent session. They cannot grant themselves elevated privileges.

Troubleshooting

• Task didn't run? → cron.list to check. Verify gateway was running at the scheduled time.
• Command failed? → Check delivery output for error details. The agent reports execution results.
• Silent task — how to check? → Use cron.list to see lastStatus and lastError fields.
• Too many old jobs? → See references/TEMPLATES.md for the cleanup job.

References

See references/TEMPLATES.md for copy-paste templates and cleanup setup.