Skip to content

The agent engineering platform.

feat(langchain): support custom token_counter in ContextEditingMiddleware (#39754)

syyy4437f266278d78
diff --git a/libs/langchain_v1/langchain/agents/middleware/context_editing.py b/libs/langchain_v1/langchain/agents/middleware/context_editing.py
index b70c3287cd..19727c79c6 100644
--- a/libs/langchain_v1/langchain/agents/middleware/context_editing.py
+++ b/libs/langchain_v1/langchain/agents/middleware/context_editing.py
@@ -196,12 +196,14 @@ class ContextEditingMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT,
edits: list[ContextEdit]
token_count_method: Literal["approximate", "model"]
+ token_counter: TokenCounter | None
def __init__(
self,
*,
edits: Iterable[ContextEdit] | None = None,
token_count_method: Literal["approximate", "model"] = "approximate", # noqa: S107
+ token_counter: TokenCounter | None = None,
) -> None:
"""Initialize an instance of context editing middleware.
@@ -212,10 +214,38 @@ class ContextEditingMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT,
token_count_method: Whether to use approximate token counting
(faster, less accurate) or exact counting implemented by the
chat model (potentially slower, more accurate).
+
+ Ignored when `token_counter` is provided.
+ token_counter: Optional custom function counting tokens for a
+ sequence of messages. Takes precedence over
+ `token_count_method` when provided, mirroring
+ `SummarizationMiddleware`. Useful when the built-in
+ approximation is inaccurate for a workload (e.g. CJK-heavy
+ conversations) or when a provider-specific tokenizer should
+ be used without implementing
+ `get_num_tokens_from_messages`.
"""
super().__init__()
self.edits = list(edits or (ClearToolUsesEdit(),))
self.token_count_method = token_count_method
+ self.token_counter = token_counter
+
+ def _resolve_token_counter(self, request: ModelRequest[ContextT]) -> TokenCounter:
+ """Select the token counter used to evaluate edit triggers."""
+ if self.token_counter is not None:
+ return self.token_counter
+
+ if self.token_count_method == "approximate": # noqa: S105
+ return count_tokens_approximately
+
+ system_msg = [request.system_message] if request.system_message else []
+
+ def count_tokens(messages: Sequence[BaseMessage]) -> int:
+ return request.model.get_num_tokens_from_messages(
+ system_msg + list(messages), request.tools
+ )
+
+ return count_tokens
def wrap_model_call(
self,
@@ -235,18 +265,7 @@ class ContextEditingMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT,
if not request.messages:
return handler(request)
- if self.token_count_method == "approximate": # noqa: S105
-
- def count_tokens(messages: Sequence[BaseMessage]) -> int:
- return count_tokens_approximately(messages)
-
- else:
- system_msg = [request.system_message] if request.system_message else []
-
- def count_tokens(messages: Sequence[BaseMessage]) -> int:
- return request.model.get_num_tokens_from_messages(
- system_msg + list(messages), request.tools
- )
+ count_tokens = self._resolve_token_counter(request)
edited_messages = deepcopy(list(request.messages))
for edit in self.edits:
@@ -272,18 +291,7 @@ class ContextEditingMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT,
if not request.messages:
return await handler(request)
- if self.token_count_method == "approximate": # noqa: S105
-
- def count_tokens(messages: Sequence[BaseMessage]) -> int:
- return count_tokens_approximately(messages)
-
- else:
- system_msg = [request.system_message] if request.system_message else []
-
- def count_tokens(messages: Sequence[BaseMessage]) -> int:
- return request.model.get_num_tokens_from_messages(
- system_msg + list(messages), request.tools
- )
+ count_tokens = self._resolve_token_counter(request)
edited_messages = deepcopy(list(request.messages))
for edit in self.edits:
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_context_editing.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_context_editing.py
index 0220e078ce..12051cbdac 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_context_editing.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_context_editing.py
@@ -468,3 +468,113 @@ async def test_exclude_tools_prevents_clearing_async() -> None:
assert isinstance(calc_tool, ToolMessage)
assert calc_tool.content == "[cleared]"
+
+
+def test_custom_token_counter_forces_clearing() -> None:
+ """A custom token counter is used to evaluate edit triggers."""
+ tool_call_id = "call-custom"
+ ai_message = AIMessage(
+ content="",
+ tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
+ )
+ # Tiny output: approximate counting would stay far below the trigger.
+ tool_message = ToolMessage(content="12345", tool_call_id=tool_call_id)
+
+ _state, request = _make_state_and_request([ai_message, tool_message])
+
+ counted_batches: list[int] = []
+
+ def fake_counter(messages: Sequence[BaseMessage]) -> int:
+ counted_batches.append(len(messages))
+ return 1_000
+
+ middleware = ContextEditingMiddleware(
+ edits=[ClearToolUsesEdit(trigger=50, keep=0, placeholder="[cleared]")],
+ token_counter=fake_counter,
+ )
+
+ modified_request = None
+
+ def mock_handler(req: ModelRequest) -> ModelResponse:
+ nonlocal modified_request
+ modified_request = req
+ return ModelResponse(result=[AIMessage(content="mock response")])
+
+ middleware.wrap_model_call(request, mock_handler)
+
+ assert counted_batches, "custom token counter was not called"
+ assert modified_request is not None
+ cleared_tool = modified_request.messages[1]
+ assert isinstance(cleared_tool, ToolMessage)
+ assert cleared_tool.content == "[cleared]"
+
+
+def test_custom_token_counter_takes_precedence_over_method() -> None:
+ """When provided, the custom counter overrides `token_count_method`."""
+ tool_call_id = "call-custom-low"
+ ai_message = AIMessage(
+ content="",
+ tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
+ )
+ # Large output: approximate counting would exceed the trigger and clear.
+ tool_message = ToolMessage(content="x" * 200, tool_call_id=tool_call_id)
+
+ _state, request = _make_state_and_request([ai_message, tool_message])
+
+ middleware = ContextEditingMiddleware(
+ edits=[ClearToolUsesEdit(trigger=50, keep=0, placeholder="[cleared]")],
+ token_count_method="approximate", # noqa: S106
+ token_counter=lambda _messages: 10,
+ )
+
+ modified_request = None
+
+ def mock_handler(req: ModelRequest) -> ModelResponse:
+ nonlocal modified_request
+ modified_request = req
+ return ModelResponse(result=[AIMessage(content="mock response")])
+
+ middleware.wrap_model_call(request, mock_handler)
+
+ assert modified_request is not None
+ untouched_tool = modified_request.messages[1]
+ assert isinstance(untouched_tool, ToolMessage)
+ assert untouched_tool.content == "x" * 200
+
+
+async def test_custom_token_counter_forces_clearing_async() -> None:
+ """Async version: a custom token counter is used to evaluate edit triggers."""
+ tool_call_id = "call-custom"
+ ai_message = AIMessage(
+ content="",
+ tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
+ )
+ tool_message = ToolMessage(content="12345", tool_call_id=tool_call_id)
+
+ _state, request = _make_state_and_request([ai_message, tool_message])
+
+ counted_batches: list[int] = []
+
+ def fake_counter(messages: Sequence[BaseMessage]) -> int:
+ counted_batches.append(len(messages))
+ return 1_000
+
+ middleware = ContextEditingMiddleware(
+ edits=[ClearToolUsesEdit(trigger=50, keep=0, placeholder="[cleared]")],
+ token_counter=fake_counter,
+ )
+
+ modified_request = None
+
+ async def mock_handler(req: ModelRequest) -> ModelResponse:
+ nonlocal modified_request
+ modified_request = req
+ return ModelResponse(result=[AIMessage(content="mock response")])
+
+ await middleware.awrap_model_call(request, mock_handler)
+
+ assert counted_batches, "custom token counter was not called"
+ assert modified_request is not None
+ cleared_tool = modified_request.messages[1]
+ assert isinstance(cleared_tool, ToolMessage)
+ assert cleared_tool.content == "[cleared]"