Some of you were struggling with slow and local models timing out compaction in Hermes Agent, so we enabled streaming on the backend for those tasks and timeout will not be dynamic. So long as tokens are coming in, it'll go until completed