There should be no distinction between llm-judge and agent-judge. If there is such an agent-judge then please remove it. The llm-judge can be agentic if the target is an agent harness such as Claude or Codex or built in agentv harness.
For the built in agent harness in agentv which uses ai SDK, use the name agentv.
We can replace the judge target by adding command option --judge-target to override all judges to use the agent providers instead of LLM providers.