real fix for LangSmith RuntimeError (#314)

revert override() reversion

Bypassing review in order to expedite fix to LangSmith Studio

Author: yangm2

backend/pyproject.toml

@@ -66,11 +66,6 @@ markers = [
 filterwarnings = [
     # Third-party library using deprecated Pydantic v1 API internally.
     "ignore::pydantic.warnings.PydanticDeprecatedSince20:langsmith",
-    # langchain deprecated direct attribute assignment on ModelRequest in favour of
-    # request.override(), but override() triggers execution_info patching that crashes
-    # in LangSmith Studio before a run context exists. Direct mutation is intentional
-    # here; suppress until the library resolves that bug.
-    "ignore:Direct attribute assignment to ModelRequest:DeprecationWarning",
 ]
 
 [tool.coverage.run]

backend/tenantfirstaid/graph.py

@@ -84,11 +84,6 @@ class _SystemPromptFromContext(AgentMiddleware[Any, TFAContext]):
     location context (city/state) from the agent state, mirroring what
     the web app does via prepare_system_prompt().
 
-    Note: direct attribute assignment to ModelRequest.system_message is
-    deprecated by langchain in favour of request.override(), but override()
-    triggers execution_info patching that crashes in LangSmith Studio before
-    a run context exists. Direct mutation is the correct approach here until
-    the library resolves that bug.
     """
 
     def _build(self, request: ModelRequest[TFAContext]) -> SystemMessage:
@@ -108,16 +103,14 @@ def wrap_model_call(
         request: ModelRequest[TFAContext],
         handler: Callable[[ModelRequest[TFAContext]], ModelResponse],
     ) -> ModelResponse:
-        request.system_message = self._build(request)  # type: ignore[misc]
-        return handler(request)
+        return handler(request.override(system_message=self._build(request)))
 
     async def awrap_model_call(
         self,
         request: ModelRequest[TFAContext],
         handler: Callable[[ModelRequest[TFAContext]], Awaitable[ModelResponse]],
     ) -> ModelResponse:
-        request.system_message = self._build(request)  # type: ignore[misc]
-        return await handler(request)
+        return await handler(request.override(system_message=self._build(request)))
 
 
 def _build_system_message(
@@ -163,12 +156,15 @@ def create_graph(
         )
 
     # LangGraph deployment path: middleware injects the prompt from
-    # Studio's editable configuration panel.
+    # Studio's editable configuration panel. context_schema is NOT set here
+    # because this graph runs as a subgraph inside graph() — the outer graph
+    # owns the context and propagates it. Declaring it on both levels causes
+    # LangSmith to patch execution_info during __start__ before a run context
+    # exists.
     return create_agent(
         model,
         tools,
         middleware=[_SystemPromptFromContext()],
-        context_schema=TFAContext,
         state_schema=TFAAgentStateSchema,
         checkpointer=checkpointer,
     )

backend/tests/test_graph.py

@@ -95,6 +95,18 @@ def _make_middleware_request(
     request.runtime.context = TFAContext(system_prompt=prompt)
     request.state = {"city": city, "state": state}
     request.system_message = None
+
+    # override() must return a new mock whose system_message reflects the kwarg,
+    # mirroring the real ModelRequest.override() behaviour.
+    def _override(**kwargs):
+        child = MagicMock()
+        child.runtime = request.runtime
+        child.state = request.state
+        for k, v in kwargs.items():
+            setattr(child, k, v)
+        return child
+
+    request.override = _override
     return request