Rule #1: Heartbeats drift. Cron is precise.
This skill provides the definitive guide for managing time in OpenClaw 2026.2.15+. It solves the "I missed my reminder" problem by enforcing a strict separation between casual checks (heartbeat) and hard schedules (cron).
SystemBehaviorBest ForRiskHeartbeat"I'll check in when I can" (e.g., every 30-60m)Email checks, casual news summaries, low-priority polling.Drift: A "remind me in 10m" task will fail if the heartbeat is 30m.Cron"I will run at exactly X time"Reminders ("in 5 mins"), daily reports, system maintenance.Clutter: Creates one-off jobs that need cleanup.
Rule: Never use act:wait or internal loops for long delays (>1 min). Use cron:add with a one-shot at schedule.
While Cron is precise, execution depends on the Gateway Heartbeat (typically every 10-60s). A job set for :00 seconds will fire on the first "tick" after that time. Expect up to ~30s of variance depending on your gateway config.
Use this payload structure for "remind me in X minutes" tasks.
Key Features (v2026.2.15+):
nextRunAtMs corruption and "Add-then-Update" deadlocks are resolved.deleteAfterRun: true).CRITICAL: Push Notifications vs. Silent Logs
For push-notification reminders (Reliable):
json
{
"name": "Remind: Water",
"schedule": { "kind": "at", "at": "2026-02-06T01:30:00Z" },
"payload": {
"kind": "agentTurn",
"message": "DELIVER THIS EXACT MESSAGE TO THE USER WITHOUT MODIFICATION OR COMMENTARY:\n\n💧 Drink water, Momo!"
},
"sessionTarget": "isolated",
"delivery": { "mode": "announce", "channel": "telegram", "to": "1027899060" }
}
For background logs (Silent):
json
{
"name": "Log: System Pulse",
"schedule": { "kind": "every", "everyMs": 3600000 },
"payload": {
"kind": "systemEvent",
"text": "[PULSE] System healthy."
},
"sessionTarget": "main"
}
Pre-2026.2.15, the "Add-then-Update" pattern caused deadlocks. While this is now stabilized, it is still best practice to pass all parameters (including wakeMode: "now") directly in the initial cron.add call for maximum efficiency.
Note: As of v2026.2.14, OpenClaw includes maintenance recompute semantics. The gateway now automatically cleans up stuck jobs and repairs corrupted schedules.
Manual cleanup is only needed for:
deleteAfterRun: false.sessionTarget: "main"? (CRITICAL)#Sub-agents ( isolated) often have restricted tool policies and cannot call gateway or delete other cron jobs. For system maintenance like the Janitor, always target the main session via systemEvent so the primary agent (with full tool access) performs the cleanup.
For cron to work, the agent must know its time.
MEMORY.md.Timezone: Cairo (GMT+2)Problem: If you say "I'll wait 30 seconds" and end your turn, you go to sleep. You cannot wake up without an event. Solution: If you need to "wait" across turns, you MUST schedule a Cron job.
act:wait).wakeMode: "now".If you have old cron jobs using these patterns, update them:
Legacy (Pre-2026.2.3)Modern (2026.2.15+)"schedule": {"kind": "at", "atMs": 1234567890}``"schedule": {"kind": "at", "at": "2026-02-06T01:30:00Z"}``"deliver": true in payloadNot needed - announce mode handles delivery"sessionTarget": "main"``"sessionTarget": "isolated" (default behavior)Manual ghost cleanup requiredOne-shots auto-delete ( deleteAfterRun: true)cron.update after cron.addSingle-step cron.add with all properties
"My reminder didn't fire": Check cron:list. Verify the at timestamp is in the future (ISO 8601 format). Ensure wakeMode: "now" is set.
"Gateway Timeout (10000ms)": This happens if the cron tool takes too long (huge job list or file lock).
~/.openclaw/state/cron/jobs.json and restart the gateway if it's corrupted."Job ran but I didn't get the message": Ensure you are using the Strict Instruction Pattern with agentTurn + announce mode for proactive pings.
"The reminder message has extra commentary": The subagent is being conversational. Use the strict prompt pattern: "DELIVER THIS EXACT MESSAGE TO THE USER WITHOUT MODIFICATION OR COMMENTARY:\n\n💧 Your message here"