Shutdown and errors

The bridge observes every failure through onError without interrupting delivery, and service.stop() owns the full shutdown order. See Channel agent for service setup and the full option reference.

Errors

onError(error, context) observes every failure with a context.stage of filter, prepare, acknowledge, interaction, agent, or delivery. Observed errors never interrupt delivery: the user receives errorMessage (default Sorry, I couldn't process that message., disable with errorMessage: false). An outcome with no text and no attachments is answered with emptyResponseMessage (default I couldn't produce a response.).

No failure reply is sent when the run's signal was already aborted, or when even the placeholder message could not be delivered — those failures are only observed through onError.

Graceful shutdown

service.stop() signals in-flight runs to abort and waits for (drains) queued conversation work before stopping the adapter — queued work is not discarded, it runs to its abort check. Stop the service before closing the databases it may still use:

try {
  await service.stop()
} finally {
  interactionStore.close()
  // await memoryClient.close() // when the agent uses a closable memory store
}

Make shutdown idempotent when registering both SIGINT and SIGTERM handlers. On shutdown, an already-sent streaming placeholder is deleted or replaced with a short (interrupted) note instead of being left dangling.

Continue with

• Approvals and interactions — durable stores worth closing on shutdown.
• End-to-end guide — the full shutdown checklist in a complete program.