We also rejig error handling slightly so we only ask for status once,
even if the endpoint fails. This should help to avoid inconsistencies.