Jobs
A queued Run can arrive at its worker twice. The first challenge in adding durable work is to decide which delivery owns the right to execute it, and what remains recoverable if the worker dies. In Core, Ghouls shows how one Run moves from admission to claim, settlement, and restart.
Follow perform_run
and the worker tests.
Look for test_concurrent_redelivery_executes_graph_exactly_once and
test_stale_delivery_sequence_is_skipped_before_claim. The worker claims the admitted Run
and its exact delivery sequence before executing the graph. For broker connection, enqueue,
and abort behavior, follow the queue services
and their tests.
Define the payload, admission, failure, recovery, and terminal evidence with the owner of the new work. Integrate worker startup and shutdown with the host lifecycle. Check current authority again at protected effects, including any host operation. Workers owns those obligations. The current extension context offers no generic job store.