fix: deprecate setattr on ModelCallRequest (#34022)

* one alternative considered was setting `frozen=True` on the dataclass,
but this is breaking, so a deprecation is a nicer approach

.gitignore

@@ -163,3 +163,6 @@ node_modules
 
 prof
 virtualenv/
+scratch/
+
+.langgraph_api/

libs/langchain_v1/langchain/agents/middleware/context_editing.py

@@ -10,6 +10,7 @@ chat model.
 from __future__ import annotations
 
 from collections.abc import Awaitable, Callable, Iterable, Sequence
+from copy import deepcopy
 from dataclasses import dataclass
 from typing import Literal
 
@@ -238,10 +239,11 @@ class ContextEditingMiddleware(AgentMiddleware):
                     system_msg + list(messages), request.tools
                 )
 
+        edited_messages = deepcopy(list(request.messages))
         for edit in self.edits:
-            edit.apply(request.messages, count_tokens=count_tokens)
+            edit.apply(edited_messages, count_tokens=count_tokens)
 
-        return handler(request)
+        return handler(request.override(messages=edited_messages))
 
     async def awrap_model_call(
         self,
@@ -266,10 +268,11 @@ class ContextEditingMiddleware(AgentMiddleware):
                     system_msg + list(messages), request.tools
                 )
 
+        edited_messages = deepcopy(list(request.messages))
         for edit in self.edits:
-            edit.apply(request.messages, count_tokens=count_tokens)
+            edit.apply(edited_messages, count_tokens=count_tokens)
 
-        return await handler(request)
+        return await handler(request.override(messages=edited_messages))
 
 
 __all__ = [

libs/langchain_v1/langchain/agents/middleware/model_fallback.py

@@ -92,9 +92,8 @@ class ModelFallbackMiddleware(AgentMiddleware):
 
         # Try fallback models
         for fallback_model in self.models:
-            request.model = fallback_model
             try:
-                return handler(request)
+                return handler(request.override(model=fallback_model))
             except Exception as e:  # noqa: BLE001
                 last_exception = e
                 continue
@@ -127,9 +126,8 @@ class ModelFallbackMiddleware(AgentMiddleware):
 
         # Try fallback models
         for fallback_model in self.models:
-            request.model = fallback_model
             try:
-                return await handler(request)
+                return await handler(request.override(model=fallback_model))
             except Exception as e:  # noqa: BLE001
                 last_exception = e
                 continue

libs/langchain_v1/langchain/agents/middleware/todo.py

@@ -194,12 +194,12 @@ class TodoListMiddleware(AgentMiddleware):
         handler: Callable[[ModelRequest], ModelResponse],
     ) -> ModelCallResult:
         """Update the system prompt to include the todo system prompt."""
-        request.system_prompt = (
+        new_system_prompt = (
             request.system_prompt + "\n\n" + self.system_prompt
             if request.system_prompt
             else self.system_prompt
         )
-        return handler(request)
+        return handler(request.override(system_prompt=new_system_prompt))
 
     async def awrap_model_call(
         self,
@@ -207,9 +207,9 @@ class TodoListMiddleware(AgentMiddleware):
         handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
     ) -> ModelCallResult:
         """Update the system prompt to include the todo system prompt (async version)."""
-        request.system_prompt = (
+        new_system_prompt = (
             request.system_prompt + "\n\n" + self.system_prompt
             if request.system_prompt
             else self.system_prompt
         )
-        return await handler(request)
+        return await handler(request.override(system_prompt=new_system_prompt))

libs/langchain_v1/langchain/agents/middleware/tool_selection.py

@@ -255,8 +255,7 @@ class LLMToolSelectorMiddleware(AgentMiddleware):
         # Also preserve any provider-specific tool dicts from the original request
         provider_tools = [tool for tool in request.tools if isinstance(tool, dict)]
 
-        request.tools = [*selected_tools, *provider_tools]
-        return request
+        return request.override(tools=[*selected_tools, *provider_tools])
 
     def wrap_model_call(
         self,

libs/langchain_v1/langchain/agents/middleware/types.py

@@ -94,6 +94,31 @@ class ModelRequest:
     runtime: Runtime[ContextT]  # type: ignore[valid-type]
     model_settings: dict[str, Any] = field(default_factory=dict)
 
+    def __setattr__(self, name: str, value: Any) -> None:
+        """Set an attribute with a deprecation warning.
+
+        Direct attribute assignment on `ModelRequest` is deprecated. Use the
+        `override()` method instead to create a new request with modified attributes.
+
+        Args:
+            name: Attribute name.
+            value: Attribute value.
+        """
+        import warnings
+
+        # Allow setting attributes during __init__ (when object is being constructed)
+        if not hasattr(self, "__dataclass_fields__") or not hasattr(self, name):
+            object.__setattr__(self, name, value)
+        else:
+            warnings.warn(
+                f"Direct attribute assignment to ModelRequest.{name} is deprecated. "
+                f"Use request.override({name}=...) instead to create a new request "
+                f"with the modified attribute.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+            object.__setattr__(self, name, value)
+
     def override(self, **overrides: Unpack[_ModelRequestOverrides]) -> ModelRequest:
         """Replace the request with a new request with the given overrides.
 
@@ -446,7 +471,14 @@ class AgentMiddleware(Generic[StateT, ContextT]):
 
                 ```python
                 def wrap_tool_call(self, request, handler):
-                    request.tool_call["args"]["value"] *= 2
+                    modified_call = {
+                        **request.tool_call,
+                        "args": {
+                            **request.tool_call["args"],
+                            "value": request.tool_call["args"]["value"] * 2,
+                        },
+                    }
+                    request = request.override(tool_call=modified_call)
                     return handler(request)
                 ```
 
@@ -1337,7 +1369,7 @@ def dynamic_prompt(
                 handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
             ) -> ModelCallResult:
                 prompt = await func(request)  # type: ignore[misc]
-                request.system_prompt = prompt
+                request = request.override(system_prompt=prompt)
                 return await handler(request)
 
             middleware_name = cast("str", getattr(func, "__name__", "DynamicPromptMiddleware"))
@@ -1358,7 +1390,7 @@ def dynamic_prompt(
             handler: Callable[[ModelRequest], ModelResponse],
         ) -> ModelCallResult:
             prompt = cast("str", func(request))
-            request.system_prompt = prompt
+            request = request.override(system_prompt=prompt)
             return handler(request)
 
         async def async_wrapped_from_sync(
@@ -1368,7 +1400,7 @@ def dynamic_prompt(
         ) -> ModelCallResult:
             # Delegate to sync function
             prompt = cast("str", func(request))
-            request.system_prompt = prompt
+            request = request.override(system_prompt=prompt)
             return await handler(request)
 
         middleware_name = cast("str", getattr(func, "__name__", "DynamicPromptMiddleware"))
@@ -1469,7 +1501,7 @@ def wrap_model_call(
                     pass
 
                 # Try fallback model
-                request.model = fallback_model_instance
+                request = request.override(model=fallback_model_instance)
                 return handler(request)
             ```
 
@@ -1632,7 +1664,14 @@ def wrap_tool_call(
             ```python
             @wrap_tool_call
             def modify_args(request, handler):
-                request.tool_call["args"]["value"] *= 2
+                modified_call = {
+                    **request.tool_call,
+                    "args": {
+                        **request.tool_call["args"],
+                        "value": request.tool_call["args"]["value"] * 2,
+                    },
+                }
+                request = request.override(tool_call=modified_call)
                 return handler(request)
             ```
 

libs/langchain_v1/tests/unit_tests/agents/middleware/core/test_composition.py

@@ -230,9 +230,7 @@ class TestChainModelCallHandlers:
         test_runtime = {"test": "runtime"}
 
         # Create request with state and runtime
-        test_request = create_test_request()
-        test_request.state = test_state
-        test_request.runtime = test_runtime
+        test_request = create_test_request(state=test_state, runtime=test_runtime)
         result = composed(test_request, create_mock_base_handler())
 
         # Both handlers should see same state and runtime

libs/langchain_v1/tests/unit_tests/agents/middleware/core/test_decorators.py

@@ -90,8 +90,7 @@ def test_on_model_call_decorator() -> None:
 
     @wrap_model_call(state_schema=CustomState, tools=[test_tool], name="CustomOnModelCall")
     def custom_on_model_call(request, handler):
-        request.system_prompt = "Modified"
-        return handler(request)
+        return handler(request.override(system_prompt="Modified"))
 
     # Verify all options were applied
     assert isinstance(custom_on_model_call, AgentMiddleware)
@@ -277,8 +276,7 @@ def test_async_on_model_call_decorator() -> None:
 
     @wrap_model_call(state_schema=CustomState, tools=[test_tool], name="AsyncOnModelCall")
     async def async_on_model_call(request, handler):
-        request.system_prompt = "Modified async"
-        return await handler(request)
+        return await handler(request.override(system_prompt="Modified async"))
 
     assert isinstance(async_on_model_call, AgentMiddleware)
     assert async_on_model_call.state_schema == CustomState

libs/langchain_v1/tests/unit_tests/agents/middleware/core/test_tools.py

@@ -79,8 +79,8 @@ def test_middleware_can_modify_tools() -> None:
             handler: Callable[[ModelRequest], AIMessage],
         ) -> AIMessage:
             # Only allow tool_a and tool_b
-            request.tools = [t for t in request.tools if t.name in ["tool_a", "tool_b"]]
-            return handler(request)
+            filtered_tools = [t for t in request.tools if t.name in ["tool_a", "tool_b"]]
+            return handler(request.override(tools=filtered_tools))
 
     # Model will try to call tool_a
     model = FakeToolCallingModel(
@@ -123,8 +123,7 @@ def test_unknown_tool_raises_error() -> None:
             handler: Callable[[ModelRequest], AIMessage],
         ) -> AIMessage:
             # Add an unknown tool
-            request.tools = request.tools + [unknown_tool]
-            return handler(request)
+            return handler(request.override(tools=request.tools + [unknown_tool]))
 
     agent = create_agent(
         model=FakeToolCallingModel(),
@@ -163,7 +162,8 @@ def test_middleware_can_add_and_remove_tools() -> None:
         ) -> AIMessage:
             # Remove admin_tool if not admin
             if not request.state.get("is_admin", False):
-                request.tools = [t for t in request.tools if t.name != "admin_tool"]
+                filtered_tools = [t for t in request.tools if t.name != "admin_tool"]
+                request = request.override(tools=filtered_tools)
             return handler(request)
 
     model = FakeToolCallingModel()
@@ -200,7 +200,7 @@ def test_empty_tools_list_is_valid() -> None:
             handler: Callable[[ModelRequest], AIMessage],
         ) -> AIMessage:
             # Remove all tools
-            request.tools = []
+            request = request.override(tools=[])
             return handler(request)
 
     model = FakeToolCallingModel()
@@ -244,7 +244,8 @@ def test_tools_preserved_across_multiple_middleware() -> None:
         ) -> AIMessage:
             modification_order.append([t.name for t in request.tools])
             # Remove tool_c
-            request.tools = [t for t in request.tools if t.name != "tool_c"]
+            filtered_tools = [t for t in request.tools if t.name != "tool_c"]
+            request = request.override(tools=filtered_tools)
             return handler(request)
 
     class SecondMiddleware(AgentMiddleware):
@@ -257,7 +258,8 @@ def test_tools_preserved_across_multiple_middleware() -> None:
             # Should not see tool_c here
             assert all(t.name != "tool_c" for t in request.tools)
             # Remove tool_b
-            request.tools = [t for t in request.tools if t.name != "tool_b"]
+            filtered_tools = [t for t in request.tools if t.name != "tool_b"]
+            request = request.override(tools=filtered_tools)
             return handler(request)
 
     agent = create_agent(

libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_context_editing.py

@@ -82,16 +82,23 @@ def test_no_edit_when_below_trigger() -> None:
         edits=[ClearToolUsesEdit(trigger=50)],
     )
 
+    modified_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call wrap_model_call which modifies the request
+    # Call wrap_model_call which creates a new request
     middleware.wrap_model_call(request, mock_handler)
 
-    # The request should have been modified in place
+    # The modified request passed to handler should be the same since no edits applied
+    assert modified_request is not None
+    assert modified_request.messages[0].content == ""
+    assert modified_request.messages[1].content == "12345"
+    # Original request should be unchanged
     assert request.messages[0].content == ""
     assert request.messages[1].content == "12345"
-    assert state["messages"] == request.messages
 
 
 def test_clear_tool_outputs_and_inputs() -> None:
@@ -115,14 +122,19 @@ def test_clear_tool_outputs_and_inputs() -> None:
     )
     middleware = ContextEditingMiddleware(edits=[edit])
 
+    modified_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call wrap_model_call which modifies the request
+    # Call wrap_model_call which creates a new request with edits
     middleware.wrap_model_call(request, mock_handler)
 
-    cleared_ai = request.messages[0]
-    cleared_tool = request.messages[1]
+    assert modified_request is not None
+    cleared_ai = modified_request.messages[0]
+    cleared_tool = modified_request.messages[1]
 
     assert isinstance(cleared_tool, ToolMessage)
     assert cleared_tool.content == "[cleared output]"
@@ -134,7 +146,9 @@ def test_clear_tool_outputs_and_inputs() -> None:
     assert context_meta is not None
     assert context_meta["cleared_tool_inputs"] == [tool_call_id]
 
-    assert state["messages"] == request.messages
+    # Original request should be unchanged
+    assert request.messages[0].tool_calls[0]["args"] == {"query": "foo"}
+    assert request.messages[1].content == "x" * 200
 
 
 def test_respects_keep_last_tool_results() -> None:
@@ -167,21 +181,26 @@ def test_respects_keep_last_tool_results() -> None:
         token_count_method="model",
     )
 
+    modified_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call wrap_model_call which modifies the request
+    # Call wrap_model_call which creates a new request with edits
     middleware.wrap_model_call(request, mock_handler)
 
+    assert modified_request is not None
     cleared_messages = [
         msg
-        for msg in request.messages
+        for msg in modified_request.messages
         if isinstance(msg, ToolMessage) and msg.content == "[cleared]"
     ]
 
     assert len(cleared_messages) == 2
-    assert isinstance(request.messages[-1], ToolMessage)
-    assert request.messages[-1].content != "[cleared]"
+    assert isinstance(modified_request.messages[-1], ToolMessage)
+    assert modified_request.messages[-1].content != "[cleared]"
 
 
 def test_exclude_tools_prevents_clearing() -> None:
@@ -215,14 +234,19 @@ def test_exclude_tools_prevents_clearing() -> None:
         ],
     )
 
+    modified_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call wrap_model_call which modifies the request
+    # Call wrap_model_call which creates a new request with edits
     middleware.wrap_model_call(request, mock_handler)
 
-    search_tool = request.messages[1]
-    calc_tool = request.messages[3]
+    assert modified_request is not None
+    search_tool = modified_request.messages[1]
+    calc_tool = modified_request.messages[3]
 
     assert isinstance(search_tool, ToolMessage)
     assert search_tool.content == "search-results" * 20
@@ -249,16 +273,23 @@ async def test_no_edit_when_below_trigger_async() -> None:
         edits=[ClearToolUsesEdit(trigger=50)],
     )
 
+    modified_request = None
+
     async def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call awrap_model_call which modifies the request
+    # Call awrap_model_call which creates a new request
     await middleware.awrap_model_call(request, mock_handler)
 
-    # The request should have been modified in place
+    # The modified request passed to handler should be the same since no edits applied
+    assert modified_request is not None
+    assert modified_request.messages[0].content == ""
+    assert modified_request.messages[1].content == "12345"
+    # Original request should be unchanged
     assert request.messages[0].content == ""
     assert request.messages[1].content == "12345"
-    assert state["messages"] == request.messages
 
 
 async def test_clear_tool_outputs_and_inputs_async() -> None:
@@ -283,14 +314,19 @@ async def test_clear_tool_outputs_and_inputs_async() -> None:
     )
     middleware = ContextEditingMiddleware(edits=[edit])
 
+    modified_request = None
+
     async def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call awrap_model_call which modifies the request
+    # Call awrap_model_call which creates a new request with edits
     await middleware.awrap_model_call(request, mock_handler)
 
-    cleared_ai = request.messages[0]
-    cleared_tool = request.messages[1]
+    assert modified_request is not None
+    cleared_ai = modified_request.messages[0]
+    cleared_tool = modified_request.messages[1]
 
     assert isinstance(cleared_tool, ToolMessage)
     assert cleared_tool.content == "[cleared output]"
@@ -302,7 +338,9 @@ async def test_clear_tool_outputs_and_inputs_async() -> None:
     assert context_meta is not None
     assert context_meta["cleared_tool_inputs"] == [tool_call_id]
 
-    assert state["messages"] == request.messages
+    # Original request should be unchanged
+    assert request.messages[0].tool_calls[0]["args"] == {"query": "foo"}
+    assert request.messages[1].content == "x" * 200
 
 
 async def test_respects_keep_last_tool_results_async() -> None:
@@ -336,21 +374,26 @@ async def test_respects_keep_last_tool_results_async() -> None:
         token_count_method="model",
     )
 
+    modified_request = None
+
     async def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call awrap_model_call which modifies the request
+    # Call awrap_model_call which creates a new request with edits
     await middleware.awrap_model_call(request, mock_handler)
 
+    assert modified_request is not None
     cleared_messages = [
         msg
-        for msg in request.messages
+        for msg in modified_request.messages
         if isinstance(msg, ToolMessage) and msg.content == "[cleared]"
     ]
 
     assert len(cleared_messages) == 2
-    assert isinstance(request.messages[-1], ToolMessage)
-    assert request.messages[-1].content != "[cleared]"
+    assert isinstance(modified_request.messages[-1], ToolMessage)
+    assert modified_request.messages[-1].content != "[cleared]"
 
 
 async def test_exclude_tools_prevents_clearing_async() -> None:
@@ -385,14 +428,19 @@ async def test_exclude_tools_prevents_clearing_async() -> None:
         ],
     )
 
+    modified_request = None
+
     async def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal modified_request
+        modified_request = req
         return AIMessage(content="mock response")
 
-    # Call awrap_model_call which modifies the request
+    # Call awrap_model_call which creates a new request with edits
     await middleware.awrap_model_call(request, mock_handler)
 
-    search_tool = request.messages[1]
-    calc_tool = request.messages[3]
+    assert modified_request is not None
+    search_tool = modified_request.messages[1]
+    calc_tool = modified_request.messages[3]
 
     assert isinstance(search_tool, ToolMessage)
     assert search_tool.content == "search-results" * 20

libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_fallback.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import warnings
 from typing import cast
 
 import pytest
@@ -45,7 +46,7 @@ def test_primary_model_succeeds() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     def mock_handler(req: ModelRequest) -> ModelResponse:
         # Simulate successful model call
@@ -70,7 +71,7 @@ def test_fallback_on_primary_failure() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     def mock_handler(req: ModelRequest) -> ModelResponse:
         result = req.model.invoke([])
@@ -95,7 +96,7 @@ def test_multiple_fallbacks() -> None:
 
     middleware = ModelFallbackMiddleware(fallback1, fallback2)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     def mock_handler(req: ModelRequest) -> ModelResponse:
         result = req.model.invoke([])
@@ -119,7 +120,7 @@ def test_all_models_fail() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     def mock_handler(req: ModelRequest) -> ModelResponse:
         result = req.model.invoke([])
@@ -136,7 +137,7 @@ async def test_primary_model_succeeds_async() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     async def mock_handler(req: ModelRequest) -> ModelResponse:
         # Simulate successful async model call
@@ -161,7 +162,7 @@ async def test_fallback_on_primary_failure_async() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     async def mock_handler(req: ModelRequest) -> ModelResponse:
         result = await req.model.ainvoke([])
@@ -186,7 +187,7 @@ async def test_multiple_fallbacks_async() -> None:
 
     middleware = ModelFallbackMiddleware(fallback1, fallback2)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     async def mock_handler(req: ModelRequest) -> ModelResponse:
         result = await req.model.ainvoke([])
@@ -210,7 +211,7 @@ async def test_all_models_fail_async() -> None:
 
     middleware = ModelFallbackMiddleware(fallback_model)
     request = _make_request()
-    request.model = primary_model
+    request = request.override(model=primary_model)
 
     async def mock_handler(req: ModelRequest) -> ModelResponse:
         result = await req.model.ainvoke([])
@@ -305,3 +306,46 @@ def test_model_fallback_middleware_initialization() -> None:
     # Test with multiple fallback models
     middleware = ModelFallbackMiddleware(FakeToolCallingModel(), FakeToolCallingModel())
     assert len(middleware.models) == 2
+
+
+def test_model_request_is_frozen() -> None:
+    """Test that ModelRequest raises deprecation warning on direct attribute assignment."""
+    request = _make_request()
+    new_model = GenericFakeChatModel(messages=iter([AIMessage(content="new model")]))
+
+    # Direct attribute assignment should raise DeprecationWarning but still work
+    with pytest.warns(
+        DeprecationWarning, match="Direct attribute assignment to ModelRequest.model is deprecated"
+    ):
+        request.model = new_model  # type: ignore[misc]
+
+    # Verify the assignment actually worked
+    assert request.model == new_model
+
+    with pytest.warns(
+        DeprecationWarning,
+        match="Direct attribute assignment to ModelRequest.system_prompt is deprecated",
+    ):
+        request.system_prompt = "new prompt"  # type: ignore[misc]
+
+    assert request.system_prompt == "new prompt"
+
+    with pytest.warns(
+        DeprecationWarning,
+        match="Direct attribute assignment to ModelRequest.messages is deprecated",
+    ):
+        request.messages = []  # type: ignore[misc]
+
+    assert request.messages == []
+
+    # Using override method should work without warnings
+    request2 = _make_request()
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")  # Turn warnings into errors
+        new_request = request2.override(model=new_model, system_prompt="override prompt")
+
+    assert new_request.model == new_model
+    assert new_request.system_prompt == "override prompt"
+    # Original request should be unchanged
+    assert request2.model != new_model
+    assert request2.system_prompt != "override prompt"

libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_todo.py

@@ -83,14 +83,21 @@ def test_adds_system_prompt_when_none_exists() -> None:
     middleware = TodoListMiddleware()
     request = _make_request(system_prompt=None)
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     middleware.wrap_model_call(request, mock_handler)
 
-    # System prompt should be set
-    assert request.system_prompt is not None
-    assert "write_todos" in request.system_prompt
+    # System prompt should be set in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt is not None
+    assert "write_todos" in captured_request.system_prompt
+    # Original request should be unchanged
+    assert request.system_prompt is None
 
 
 def test_appends_to_existing_system_prompt() -> None:
@@ -99,16 +106,23 @@ def test_appends_to_existing_system_prompt() -> None:
     middleware = TodoListMiddleware()
     request = _make_request(system_prompt=existing_prompt)
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     middleware.wrap_model_call(request, mock_handler)
 
-    # System prompt should contain both
-    assert request.system_prompt is not None
-    assert existing_prompt in request.system_prompt
-    assert "write_todos" in request.system_prompt
-    assert request.system_prompt.startswith(existing_prompt)
+    # System prompt should contain both in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt is not None
+    assert existing_prompt in captured_request.system_prompt
+    assert "write_todos" in captured_request.system_prompt
+    assert captured_request.system_prompt.startswith(existing_prompt)
+    # Original request should be unchanged
+    assert request.system_prompt == existing_prompt
 
 
 @pytest.mark.parametrize(
@@ -137,13 +151,20 @@ def test_todo_middleware_on_model_call(original_prompt, expected_prompt_prefix)
         model_settings={},
     )
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal captured_request
+        captured_request = req
         return AIMessage(content="mock response")
 
     # Call wrap_model_call to trigger the middleware logic
     middleware.wrap_model_call(request, mock_handler)
-    # Check that the request was modified in place
-    assert request.system_prompt.startswith(expected_prompt_prefix)
+    # Check that the modified request passed to handler has the expected prompt
+    assert captured_request is not None
+    assert captured_request.system_prompt.startswith(expected_prompt_prefix)
+    # Original request should be unchanged
+    assert request.system_prompt == original_prompt
 
 
 def test_custom_system_prompt() -> None:
@@ -152,13 +173,20 @@ def test_custom_system_prompt() -> None:
     middleware = TodoListMiddleware(system_prompt=custom_prompt)
     request = _make_request(system_prompt=None)
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     middleware.wrap_model_call(request, mock_handler)
 
-    # Should use custom prompt
-    assert request.system_prompt == custom_prompt
+    # Should use custom prompt in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt == custom_prompt
+    # Original request should be unchanged
+    assert request.system_prompt is None
 
 
 def test_todo_middleware_custom_system_prompt() -> None:
@@ -181,13 +209,20 @@ def test_todo_middleware_custom_system_prompt() -> None:
         runtime=cast(Runtime, object()),
     )
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal captured_request
+        captured_request = req
         return AIMessage(content="mock response")
 
     # Call wrap_model_call to trigger the middleware logic
     middleware.wrap_model_call(request, mock_handler)
-    # Check that the request was modified in place
-    assert request.system_prompt == f"Original prompt\n\n{custom_system_prompt}"
+    # Check that the modified request passed to handler has the expected prompt
+    assert captured_request is not None
+    assert captured_request.system_prompt == f"Original prompt\n\n{custom_system_prompt}"
+    # Original request should be unchanged
+    assert request.system_prompt == "Original prompt"
 
 
 def test_custom_tool_description() -> None:
@@ -235,13 +270,20 @@ def test_todo_middleware_custom_system_prompt_and_tool_description() -> None:
         model_settings={},
     )
 
+    captured_request = None
+
     def mock_handler(req: ModelRequest) -> AIMessage:
+        nonlocal captured_request
+        captured_request = req
         return AIMessage(content="mock response")
 
     # Call wrap_model_call to trigger the middleware logic
     middleware.wrap_model_call(request, mock_handler)
-    # Check that the request was modified in place
-    assert request.system_prompt == custom_system_prompt
+    # Check that the modified request passed to handler has the expected prompt
+    assert captured_request is not None
+    assert captured_request.system_prompt == custom_system_prompt
+    # Original request should be unchanged
+    assert request.system_prompt is None
 
     # Verify tool description
     assert len(middleware.tools) == 1
@@ -390,14 +432,21 @@ async def test_adds_system_prompt_when_none_exists_async() -> None:
     middleware = TodoListMiddleware()
     request = _make_request(system_prompt=None)
 
+    captured_request = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     await middleware.awrap_model_call(request, mock_handler)
 
-    # System prompt should be set
-    assert request.system_prompt is not None
-    assert "write_todos" in request.system_prompt
+    # System prompt should be set in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt is not None
+    assert "write_todos" in captured_request.system_prompt
+    # Original request should be unchanged
+    assert request.system_prompt is None
 
 
 async def test_appends_to_existing_system_prompt_async() -> None:
@@ -406,16 +455,23 @@ async def test_appends_to_existing_system_prompt_async() -> None:
     middleware = TodoListMiddleware()
     request = _make_request(system_prompt=existing_prompt)
 
+    captured_request = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     await middleware.awrap_model_call(request, mock_handler)
 
-    # System prompt should contain both
-    assert request.system_prompt is not None
-    assert existing_prompt in request.system_prompt
-    assert "write_todos" in request.system_prompt
-    assert request.system_prompt.startswith(existing_prompt)
+    # System prompt should contain both in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt is not None
+    assert existing_prompt in captured_request.system_prompt
+    assert "write_todos" in captured_request.system_prompt
+    assert captured_request.system_prompt.startswith(existing_prompt)
+    # Original request should be unchanged
+    assert request.system_prompt == existing_prompt
 
 
 async def test_custom_system_prompt_async() -> None:
@@ -424,13 +480,20 @@ async def test_custom_system_prompt_async() -> None:
     middleware = TodoListMiddleware(system_prompt=custom_prompt)
     request = _make_request(system_prompt=None)
 
+    captured_request = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal captured_request
+        captured_request = req
         return ModelResponse(result=[AIMessage(content="response")])
 
     await middleware.awrap_model_call(request, mock_handler)
 
-    # Should use custom prompt
-    assert request.system_prompt == custom_prompt
+    # Should use custom prompt in the modified request passed to handler
+    assert captured_request is not None
+    assert captured_request.system_prompt == custom_prompt
+    # Original request should be unchanged
+    assert request.system_prompt is None
 
 
 async def test_handler_called_with_modified_request_async() -> None:

libs/langchain_v1/tests/unit_tests/agents/test_response_format.py

@@ -784,8 +784,7 @@ class TestDynamicModelWithResponseFormat:
                 handler: Callable[[ModelRequest], CoreAIMessage],
             ) -> CoreAIMessage:
                 # Replace the model with our custom test model
-                request.model = model
-                return handler(request)
+                return handler(request.override(model=model))
 
         # Track which model is checked for provider strategy support
         calls = []

libs/partners/anthropic/langchain_anthropic/middleware/prompt_caching.py

@@ -94,14 +94,6 @@ class AnthropicPromptCachingMiddleware(AgentMiddleware):
         )
         return messages_count >= self.min_messages_to_cache
 
-    def _apply_cache_control(self, request: ModelRequest) -> None:
-        """Apply cache control settings to the request.
-
-        Args:
-            request: The model request to modify.
-        """
-        request.model_settings["cache_control"] = {"type": self.type, "ttl": self.ttl}
-
     def wrap_model_call(
         self,
         request: ModelRequest,
@@ -119,8 +111,12 @@ class AnthropicPromptCachingMiddleware(AgentMiddleware):
         if not self._should_apply_caching(request):
             return handler(request)
 
-        self._apply_cache_control(request)
-        return handler(request)
+        model_settings = request.model_settings
+        new_model_settings = {
+            **model_settings,
+            "cache_control": {"type": self.type, "ttl": self.ttl},
+        }
+        return handler(request.override(model_settings=new_model_settings))
 
     async def awrap_model_call(
         self,
@@ -139,5 +135,9 @@ class AnthropicPromptCachingMiddleware(AgentMiddleware):
         if not self._should_apply_caching(request):
             return await handler(request)
 
-        self._apply_cache_control(request)
-        return await handler(request)
+        model_settings = request.model_settings
+        new_model_settings = {
+            **model_settings,
+            "cache_control": {"type": self.type, "ttl": self.ttl},
+        }
+        return await handler(request.override(model_settings=new_model_settings))

libs/partners/anthropic/tests/unit_tests/middleware/test_prompt_caching.py

@@ -82,12 +82,17 @@ def test_anthropic_prompt_caching_middleware_initialization() -> None:
         model_settings={},
     )
 
+    modified_request: ModelRequest | None = None
+
     def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal modified_request
+        modified_request = req
         return ModelResponse(result=[AIMessage(content="mock response")])
 
     middleware.wrap_model_call(fake_request, mock_handler)
     # Check that model_settings were passed through via the request
-    assert fake_request.model_settings == {
+    assert modified_request is not None
+    assert modified_request.model_settings == {
         "cache_control": {"type": "ephemeral", "ttl": "5m"}
     }
 
@@ -162,13 +167,18 @@ async def test_anthropic_prompt_caching_middleware_async() -> None:
         model_settings={},
     )
 
+    modified_request: ModelRequest | None = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal modified_request
+        modified_request = req
         return ModelResponse(result=[AIMessage(content="mock response")])
 
     result = await middleware.awrap_model_call(fake_request, mock_handler)
     assert isinstance(result, ModelResponse)
     # Check that model_settings were passed through via the request
-    assert fake_request.model_settings == {
+    assert modified_request is not None
+    assert modified_request.model_settings == {
         "cache_control": {"type": "ephemeral", "ttl": "1h"}
     }
 
@@ -237,13 +247,18 @@ async def test_anthropic_prompt_caching_middleware_async_min_messages() -> None:
         model_settings={},
     )
 
+    modified_request: ModelRequest | None = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal modified_request
+        modified_request = req
         return ModelResponse(result=[AIMessage(content="mock response")])
 
     result = await middleware.awrap_model_call(fake_request, mock_handler)
     assert isinstance(result, ModelResponse)
     # Cache control should NOT be added when message count is below minimum
-    assert fake_request.model_settings == {}
+    assert modified_request is not None
+    assert modified_request.model_settings == {}
 
 
 async def test_anthropic_prompt_caching_middleware_async_with_system_prompt() -> None:
@@ -268,13 +283,18 @@ async def test_anthropic_prompt_caching_middleware_async_with_system_prompt() ->
         model_settings={},
     )
 
+    modified_request: ModelRequest | None = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal modified_request
+        modified_request = req
         return ModelResponse(result=[AIMessage(content="mock response")])
 
     result = await middleware.awrap_model_call(fake_request, mock_handler)
     assert isinstance(result, ModelResponse)
     # Cache control should be added when system prompt pushes count to minimum
-    assert fake_request.model_settings == {
+    assert modified_request is not None
+    assert modified_request.model_settings == {
         "cache_control": {"type": "ephemeral", "ttl": "1h"}
     }
 
@@ -300,12 +320,17 @@ async def test_anthropic_prompt_caching_middleware_async_default_values() -> Non
         model_settings={},
     )
 
+    modified_request: ModelRequest | None = None
+
     async def mock_handler(req: ModelRequest) -> ModelResponse:
+        nonlocal modified_request
+        modified_request = req
         return ModelResponse(result=[AIMessage(content="mock response")])
 
     result = await middleware.awrap_model_call(fake_request, mock_handler)
     assert isinstance(result, ModelResponse)
     # Check that model_settings were added with default values
-    assert fake_request.model_settings == {
+    assert modified_request is not None
+    assert modified_request.model_settings == {
         "cache_control": {"type": "ephemeral", "ttl": "5m"}
     }