test_summarization_before_model_uses_unscaled_tokens_for_cutoff() — langchain Function Reference

Source Code

libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py lines 1031–1054

def test_summarization_before_model_uses_unscaled_tokens_for_cutoff() -> None:
    calls: list[dict[str, Any]] = []

    def fake_counter(_: Iterable[MessageLikeRepresentation], **kwargs: Any) -> int:
        calls.append(kwargs)
        return 100

    with patch(
        "langchain.agents.middleware.summarization.count_tokens_approximately",
        side_effect=fake_counter,
    ) as mock_counter:
        middleware = SummarizationMiddleware(
            model=MockChatModel(),
            trigger=("tokens", 1),
            keep=("tokens", 1),
            token_counter=mock_counter,
        )
        state = AgentState[Any](messages=[HumanMessage(content="one"), HumanMessage(content="two")])
        assert middleware.before_model(state, Runtime()) is not None

    # Test we support partial token counting (which for default token counter does not
    # use use_usage_metadata_scaling)
    assert any(call.get("use_usage_metadata_scaling") is False for call in calls)
    assert any(call.get("use_usage_metadata_scaling") is True for call in calls)