fix(sdk): patch invalid tool calls (#3386)

Co-authored-by: Aarushi Singh <aarushi07.singh@gmail.com>

libs/deepagents/deepagents/middleware/patch_tool_calls.py

@@ -3,7 +3,7 @@
 from typing import Any
 
 from langchain.agents.middleware import AgentMiddleware, AgentState
-from langchain_core.messages import AIMessage, ToolMessage
+from langchain_core.messages import AIMessage, AnyMessage, ToolMessage
 from langgraph.runtime import Runtime
 from langgraph.types import Overwrite
 
@@ -20,25 +20,27 @@ class PatchToolCallsMiddleware(AgentMiddleware):
         answered_ids = {msg.tool_call_id for msg in messages if msg.type == "tool"}  # ty: ignore[unresolved-attribute]
 
         if not any(
-            tool_call["id"] not in answered_ids for msg in messages if isinstance(msg, AIMessage) and msg.tool_calls for tool_call in msg.tool_calls
+            tool_call["id"] is not None and tool_call["id"] not in answered_ids
+            for msg in messages
+            if isinstance(msg, AIMessage)
+            for tool_call in (*msg.tool_calls, *msg.invalid_tool_calls)
         ):
             return None
 
-        patched_messages = []
+        patched_messages: list[AnyMessage] = []
         for msg in messages:
             patched_messages.append(msg)
-            if isinstance(msg, AIMessage) and msg.tool_calls:
-                patched_messages.extend(
-                    ToolMessage(
-                        content=(
-                            f"Tool call {tool_call['name']} with id {tool_call['id']} was "
-                            "cancelled - another message came in before it could be completed."
-                        ),
-                        name=tool_call["name"],
-                        tool_call_id=tool_call["id"],
-                    )
-                    for tool_call in msg.tool_calls
-                    if tool_call["id"] not in answered_ids
-                )
+            if not isinstance(msg, AIMessage):
+                continue
+            for tool_call in (*msg.tool_calls, *msg.invalid_tool_calls):
+                tool_call_id = tool_call["id"]
+                if tool_call_id is None or tool_call_id in answered_ids:
+                    continue
+                name = tool_call["name"] or "unknown"
+                if tool_call.get("type") == "invalid_tool_call":
+                    content = f"Tool call {name} with id {tool_call_id} could not be executed - arguments were malformed or truncated."
+                else:
+                    content = f"Tool call {name} with id {tool_call_id} was cancelled - another message came in before it could be completed."
+                patched_messages.append(ToolMessage(content=content, name=name, tool_call_id=tool_call_id))
 
         return {"messages": Overwrite(patched_messages)}

libs/deepagents/tests/unit_tests/test_end_to_end.py

@@ -14,7 +14,7 @@ from langchain.agents.middleware.types import ModelRequest, ModelResponse
 from langchain.tools import ToolRuntime
 from langchain_core.language_models import LanguageModelInput
 from langchain_core.language_models.fake_chat_models import GenericFakeChatModel
-from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 from langchain_core.runnables import Runnable
 from langchain_core.tools import BaseTool, tool
 from langgraph.channels.delta import DeltaChannel
@@ -3578,3 +3578,48 @@ class TestDeltaChannels:
         state = agent.get_state(config)
         files = state.values.get("files", {})
         assert any("hello.txt" in k for k in files)
+
+
+def test_invalid_tool_call_patched_on_next_turn() -> None:
+    # Turn 1: model truncates and emits an invalid tool call (no matching ToolMessage
+    # will be produced because agents only route on ``tool_calls``).
+    # Turn 2: the middleware must patch the dangling call before the model is re-invoked.
+    fake_model = FakeChatModelWithHistory(
+        messages=iter(
+            [
+                AIMessage(
+                    content="",
+                    invalid_tool_calls=[
+                        {
+                            "id": "call_truncated",
+                            "name": "search",
+                            "args": '{"query": "weath',
+                            "error": "Unterminated string at line 1 column 17",
+                            "type": "invalid_tool_call",
+                        }
+                    ],
+                ),
+                AIMessage(content="Recovered."),
+            ]
+        )
+    )
+    checkpointer = InMemorySaver()
+    agent = create_deep_agent(model=fake_model, checkpointer=checkpointer)
+    config: dict = {"configurable": {"thread_id": "patch-invalid-tool-calls"}}
+
+    agent.invoke({"messages": [HumanMessage(content="Run a tool")]}, config)
+    result = agent.invoke({"messages": [HumanMessage(content="Try again")]}, config)
+
+    # The second model call must see the dangling invalid_tool_call paired with a ToolMessage.
+    second_call_inputs = fake_model.call_history[1]["messages"]
+    synthetic = next(
+        (m for m in second_call_inputs if isinstance(m, ToolMessage) and m.tool_call_id == "call_truncated"),
+        None,
+    )
+    assert synthetic is not None, "PatchToolCallsMiddleware did not inject a ToolMessage for invalid_tool_calls"
+    assert "could not be executed" in synthetic.content
+    assert "malformed or truncated" in synthetic.content
+    assert synthetic.name == "search"
+
+    # Final state must also expose the patched ToolMessage.
+    assert any(isinstance(m, ToolMessage) and m.tool_call_id == "call_truncated" for m in result["messages"])