from sage_reliability_lab.agent import CareNavigationAgent from sage_reliability_lab.clustering import cluster_failures from sage_reliability_lab.evaluation import evaluate_trace from sage_reliability_lab.synthetic_data import generate_synthetic_calls def test_batch_pipeline_runs(): agent = CareNavigationAgent() calls = generate_synthetic_calls(8) traces = [agent.run(call.transcript, call_id=call.call_id, expected_intent=call.expected_intent) for call in calls] evals = [evaluate_trace(trace, expected=call) for trace, call in zip(traces, calls)] clusters = cluster_failures(evals) assert len(evals) == 8 assert all(0.0 <= item.overall_score <= 1.0 for item in evals) assert isinstance(clusters, list)