| name | scheduler-operations-playbook |
| description | Create, inspect, update, run, diagnose, and verify Prometheus scheduled jobs, including cron/interval timing, ownership, expected outputs, delivery, retries, and run history. Use for scheduler operations; do not use for ordinary foreground tasks or generic background-agent work. |
Scheduler operations
Scheduled work is durable automation. A job is not healthy merely because its scheduler entry exists or its last run says “success.” Verify execution, output, and delivery separately.
Inspect before mutating
- List or inspect the target schedule and confirm its identifier.
- Review timing, timezone, enabled state, owner, model/capabilities, prompt, expected outputs, and delivery configuration.
- Read recent run history and inspect the latest real artifact or delivery evidence.
- Distinguish scheduler state from worker execution and user-facing delivery.
Create or update safely
- Make the job prompt self-contained for a future cold start.
- Include the concrete objective, required inputs, output location or delivery target, verification criteria, and failure behavior.
- Use a schedule-owner agent only when the workflow truly needs durable ownership.
- Patch only intended fields on existing jobs; preserve IDs, ownership, delivery, and unrelated configuration.
- Confirm timezone and the next expected run in user-readable terms.
Verify
For a new or materially changed job, use a safe immediate run when possible. Then verify:
- the run started under the expected owner;
- required tools/model capabilities were available;
- the intended artifact changed or was created;
- delivery reached the configured destination;
- run history reflects the actual outcome.
Do not infer success from tool count, a generated final message, or a stale artifact timestamp.
Triage
- No run: inspect enablement, schedule shape, timezone, and scheduler health.
- Run failed: inspect the first actionable tool/model/auth error.
- False success: compare expected outputs with current files or external state.
- Missing delivery: separate completed work from notification/channel failure.
- Retry loop: stop repeated identical attempts and fix the underlying capability, prompt, or auth issue.
- Long-running work: inspect durable task/process state instead of assuming the foreground timeout cancelled it.
Read details only when needed
- Read detailed-guide.md for field-level operations, incident patterns, heartbeat diagnostics, and delivery rules.
- Read a matching example under
examples/ only when creating, updating, or recovering that kind of schedule.
- Read
references/workflows/scheduler-operations-playbook.md only for historical workflow details not covered here.
Report the schedule change, next run, immediate-test result, artifact evidence, and delivery status independently.