llm-as-judge 1

dsh host plugin: every agent turn's final assistant output is reviewed by a separate LLM graded 1-100; below-threshold scores steer the agent with concrete feedback. The score is shown in the finalized answer's action row (conversation.chat.assistant-acti