golem-recurring-task-moonbit

Have the agent call schedule_ on its own generated client at the end of each invocation, passing a wall-clock datetime for the next run. This creates a durable loop where each execution enqueues the following one.

Track consecutive failures in agent state, then compute the delay as base interval times a power of two capped at a maximum interval. Reset the failure counter to zero whenever the work succeeds.

Yes, use the schedule_cancelable_ client variant which returns a CancellationToken and call cancel() on it. Alternatively use a boolean state flag checked at the start of the method, or cancel via the Golem CLI with an idempotency key.

Yes, scheduled invocations are durable and persisted, so if the agent crashes the pending invocation still fires at the designated time and the agent recovers. Invocations on the same agent run sequentially without concurrency.

Every scheduled tick appends entries to the agent's oplog, so recovery replays the full history and slows over time. Enable periodic snapshotting with the snapshotting attribute so recovery starts from the latest snapshot instead.