Testing the Agent

# Complete testing strategy for a production agent

# Phase 1: Unit tests — individual components in isolation
def test_check_eligibility_unit():
    """Test eligibility logic without calling the LLM or external APIs."""
    from datetime import date, timedelta

    def check_eligibility(order: dict) -> str:
        order_date = date.fromisoformat(order["date"])
        days_since  = (date.today() - order_date).days
        if days_since > 30:
            return "ineligible"
        if order["total"] >= 100:
            return "eligible_manual"
        return "eligible_auto"

    # Test all branches
    recent_small  = {"date": str(date.today() - timedelta(5)), "total": 49.99}
    recent_large  = {"date": str(date.today() - timedelta(5)), "total": 199.99}
    old_order     = {"date": str(date.today() - timedelta(45)), "total": 49.99}

    assert check_eligibility(recent_small)  == "eligible_auto"
    assert check_eligibility(recent_large)  == "eligible_manual"
    assert check_eligibility(old_order)     == "ineligible"
    print("✓ Phase 1: Unit tests passed")

# Phase 2: Integration test — full workflow with mock tools
def test_full_workflow_integration():
    """Run the complete workflow with mocked external calls."""
    mock_tools = {
        "get_order":    lambda order_id: {"status":"success","order":{"total":49.99,"date":"2024-01-10"}},
        "issue_refund": lambda order_id, amount: {"status":"success","refund_id":"REF-001"},
        "send_email":   lambda to, subject, body: {"status":"success"}
    }
    # Simulate a successful run through the full workflow
    order = mock_tools["get_order"]("ORD-123")["order"]
    refund = mock_tools["issue_refund"]("ORD-123", order["total"])
    email  = mock_tools["send_email"]("user@ex.com", "Refund confirmed", "...")
    assert refund["status"] == "success"
    assert email["status"]  == "success"
    print("✓ Phase 2: Integration test passed")

# Phase 3: Evaluation — score output quality
def test_output_quality():
    """Score agent output against expected criteria."""
    agent_output = {
        "outcome": "success", "refund_id": "REF-001",
        "message": "Your refund of $49.99 for order ORD-123 has been processed."
    }
    score = 0
    if agent_output.get("outcome") == "success": score += 1
    if agent_output.get("refund_id"):             score += 1
    if agent_output.get("message") and len(agent_output["message"]) > 20: score += 1
    if "49.99" in agent_output.get("message", ""): score += 1
    assert score == 4, f"Quality score {score}/4"
    print("✓ Phase 3: Output quality test passed (4/4)")

test_check_eligibility_unit()
test_full_workflow_integration()
test_output_quality()
print("\nAll phases passed!")