| 1234567891011 |
- from app.commands.evaluate_agents import load_cases
- def test_smoke_and_full_agent_evaluation_datasets_cover_both_personas() -> None:
- smoke = load_cases("smoke")
- full = load_cases("full")
- assert {case["persona"] for case in smoke} == {"customer", "operation"}
- assert len(full) > len(smoke)
- assert all(case["expected_tools"] for case in full)
- assert all(case["expected_terms"] for case in full)
|