test_agent_evaluations.py 577 B

123456789101112131415
  1. from zbt.commands.evaluate_agents import DEFAULT_EVALUATION_H5_MOBILE, load_cases
  2. def test_smoke_and_full_agent_evaluation_datasets_cover_both_personas() -> None:
  3. smoke = load_cases("smoke")
  4. full = load_cases("full")
  5. assert {case["persona"] for case in smoke} == {"customer", "operation"}
  6. assert len(full) > len(smoke)
  7. assert all(case["expected_tools"] for case in full)
  8. assert all(case["expected_terms"] for case in full)
  9. def test_customer_evaluation_uses_a_dedicated_h5_account() -> None:
  10. assert DEFAULT_EVALUATION_H5_MOBILE != "18800000001"