openclaw - 💡(How to fix) Fix Cron startup catch-up fires duplicate when previous run was itself a late catch-up [1 participants]

Bug Description

The cron scheduler's startup catch-up logic (planStartupCatchup → isRunnableJob) has a structural flaw where a job that ran as a late catch-up can be incorrectly classified as "missed" on the next gateway restart, causing a duplicate execution.

Steps to Reproduce

1. A cron job is scheduled at 0 20 * * * in Europe/Berlin (20:00 CET = ~18:00 UTC)
2. Gateway is down at scheduled time → job doesn't fire
3. Gateway restarts at e.g. 01:56 UTC (next day) → catch-up fires correctly (lastRunAtMs = 01:56 UTC, nextRunAtMs set to next day's slot)
4. Gateway restarts again at e.g. 04:21 UTC the same day

Expected Behavior

Job should NOT fire — it already ran 2.5 hours ago and nextRunAtMs points to the next day.

Actual Behavior

Job fires again as a catch-up duplicate.

Root Cause

isRunnableJob() has two detection paths. Path 1 (nextRunAtMs >= nowMs) correctly says "not due." But Path 2 (startup-only, gated by allowCronMissedRunByLastRun) recomputes previousRunAtMs from the cron expression:

const previousRunAtMs = computeJobPreviousRunAtMs(job, nowMs);
const lastRunAtMs = job.state.lastRunAtMs;
return previousRunAtMs > lastRunAtMs;

At 04:21 UTC:

• previousRunAtMs = most recent 0 20 * * * in Europe/Berlin before now = ~18:00 UTC (same day)
• lastRunAtMs = 01:56 UTC (when catch-up actually ran)
• 18:00 > 01:56 → TRUE → classified as missed → duplicate fires

The comparison is purely timestamp-based and cannot distinguish "this run was fulfilling yesterday's slot via late catch-up" from "this run happened before today's slot."

Suggested Fix

When applyJobResult records a successful run, it could also store the cron slot timestamp that the run was fulfilling (e.g. lastFulfilledSlotMs). Path 2 would then compare previousRunAtMs > lastFulfilledSlotMs instead of previousRunAtMs > lastRunAtMs, preventing false-positive missed-run detection for late catch-ups.

Alternatively, Path 2 could be skipped entirely when nextRunAtMs is set and in the future (Path 1 already covers this case correctly).

Environment

• OpenClaw version: latest (pnpm global install)
• Cron expression: 0 20 * * * with tz: Europe/Berlin
• wakeMode: now (but irrelevant — Path 2 ignores wakeMode)