Skip to content

The agent engineering platform.

fix(langchain): alias Bedrock providers in summarization token check (#37453)

ccurme36c381b14968
diff --git a/libs/langchain_v1/langchain/agents/middleware/summarization.py b/libs/langchain_v1/langchain/agents/middleware/summarization.py
index 6c1091152c..ac31da5b29 100644
--- a/libs/langchain_v1/langchain/agents/middleware/summarization.py
+++ b/libs/langchain_v1/langchain/agents/middleware/summarization.py
@@ -77,6 +77,23 @@ _DEFAULT_MESSAGES_TO_KEEP = 20
_DEFAULT_TRIM_TOKEN_LIMIT = 4000
_DEFAULT_FALLBACK_MESSAGE_COUNT = 15
+# Some providers tag emitted messages with a `model_provider` string that differs from
+# their LangSmith `ls_provider`. The reported-token check below compares the two, so we
+# accept known aliases per `ls_provider`.
+_LS_PROVIDER_ALIASES: dict[str, frozenset[str]] = {
+ "amazon_bedrock": frozenset({"bedrock", "bedrock_converse"}),
+}
+
+
+def _provider_matches(message_provider: str, model_ls_provider: str | None) -> bool:
+ if model_ls_provider is None:
+ return False
+ if message_provider == model_ls_provider:
+ return True
+ aliases = _LS_PROVIDER_ALIASES.get(model_ls_provider)
+ return aliases is not None and message_provider in aliases
+
+
ContextFraction = tuple[Literal["fraction"], float]
"""Fraction of model's maximum input tokens.
@@ -379,7 +396,10 @@ class SummarizationMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
and (reported_tokens := last_ai_message.usage_metadata.get("total_tokens", -1))
and reported_tokens >= threshold
and (message_provider := last_ai_message.response_metadata.get("model_provider"))
- and message_provider == self.model._get_ls_params().get("ls_provider") # noqa: SLF001
+ and _provider_matches(
+ message_provider,
+ self.model._get_ls_params().get("ls_provider"), # noqa: SLF001
+ )
):
return True
return False
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
index 1ae973c7ad..8dac1ff76c 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
@@ -6,6 +6,7 @@ import pytest
from langchain_core.callbacks import AsyncCallbackManagerForLLMRun, CallbackManagerForLLMRun
from langchain_core.language_models import ModelProfile
from langchain_core.language_models.base import (
+ LangSmithParams,
LanguageModelInput,
)
from langchain_core.language_models.chat_models import BaseChatModel
@@ -27,7 +28,10 @@ from pydantic import Field
from typing_extensions import override
from langchain.agents import AgentState
-from langchain.agents.middleware.summarization import SummarizationMiddleware
+from langchain.agents.middleware.summarization import (
+ SummarizationMiddleware,
+ _provider_matches,
+)
from langchain.chat_models import init_chat_model
from tests.unit_tests.agents.model import FakeToolCallingModel
@@ -1219,6 +1223,84 @@ def test_usage_metadata_trigger() -> None:
assert not middleware._should_summarize(messages, 0)
+def test_provider_matches() -> None:
+ """Direct equality matches, plus Bedrock aliases under amazon_bedrock."""
+ assert _provider_matches("anthropic", "anthropic")
+ assert _provider_matches("openai", "openai")
+ # Bedrock chat models tag messages with model_provider="bedrock" or
+ # "bedrock_converse" but trace under ls_provider="amazon_bedrock".
+ assert _provider_matches("bedrock", "amazon_bedrock")
+ assert _provider_matches("bedrock_converse", "amazon_bedrock")
+ # Non-matches
+ assert not _provider_matches("openai", "anthropic")
+ assert not _provider_matches("bedrock", "anthropic")
+ assert not _provider_matches("anthropic", None)
+
+
+class _MockBedrockChatModel(BaseChatModel):
+ """Mock model that mimics ChatBedrockConverse's ls_provider for tracing."""
+
+ @override
+ def _generate(
+ self,
+ messages: list[BaseMessage],
+ stop: list[str] | None = None,
+ run_manager: CallbackManagerForLLMRun | None = None,
+ **kwargs: Any,
+ ) -> ChatResult:
+ return ChatResult(generations=[ChatGeneration(message=AIMessage(content="Summary"))])
+
+ @property
+ def _llm_type(self) -> str:
+ return "amazon_bedrock_converse_chat"
+
+ @override
+ def _get_ls_params(self, stop: list[str] | None = None, **kwargs: Any) -> LangSmithParams:
+ return LangSmithParams(ls_provider="amazon_bedrock", ls_model_type="chat")
+
+
+def test_reported_tokens_trigger_for_bedrock_converse() -> None:
+ """Bedrock messages should satisfy the reported-token check.
+
+ Despite the model_provider/ls_provider mismatch (bedrock_converse vs.
+ amazon_bedrock), the reported-token check should still trigger summarization.
+ """
+ middleware = SummarizationMiddleware(
+ model=_MockBedrockChatModel(),
+ trigger=("tokens", 10_000),
+ keep=("messages", 4),
+ )
+ messages: list[AnyMessage] = [
+ HumanMessage(content="msg1"),
+ AIMessage(
+ content="msg2",
+ response_metadata={"model_provider": "bedrock_converse"},
+ usage_metadata={
+ "input_tokens": 7500,
+ "output_tokens": 2501,
+ "total_tokens": 10_001,
+ },
+ ),
+ ]
+ # reported token count (10_001) should override the supplied count of 0
+ assert middleware._should_summarize(messages, 0)
+
+ # mismatched provider should not engage
+ messages_other_provider: list[AnyMessage] = [
+ HumanMessage(content="msg1"),
+ AIMessage(
+ content="msg2",
+ response_metadata={"model_provider": "anthropic"},
+ usage_metadata={
+ "input_tokens": 7500,
+ "output_tokens": 2501,
+ "total_tokens": 10_001,
+ },
+ ),
+ ]
+ assert not middleware._should_summarize(messages_other_provider, 0)
+
+
class ConfigCapturingModel(BaseChatModel):
"""Mock model that captures the config passed to invoke/ainvoke."""