Workers that live inside a parent agent's own loop rather than running as a standalone process — Carol's in-conversation helpers are the main example. They have no run of their own to check, so Hermione judges them by whether their parent is alive (parent-alive liveness).

Every process that runs on a clock — daily, hourly, every-N-minutes or continuously. Hermione judges these on cadence: did it run on time, did its last run complete, is it overdue. This block spans EVERY such process across all agents and services, not just one team's.

The dedicated datastore for scheduled-process run-audit. Every recurring droid records a start/finish heartbeat here (kept OUT of the planner database so the per-minute write burst never contends with the pipeline). Hermione's daily liveness sweep reads it to tell which scheduled jobs ran, failed, or are overdue.

Processes that run only when something invokes them — an event, a webhook, another agent, or a manual call. Never-ran is normal here, so Hermione watches a different signal: did the last invocation fail or hang. Spans every triggered/on-demand process across the whole team.

The core work of the monitoring service.