Convert_to_openai_messages Drops Invalid Tool Calls From Exported Conversations
The convert_to_openai_messages utility in langchain-core omits AIMessage.invalid_tool_calls from the converted OpenAI message, while the corresponding ToolMessage replies are retained, leading to orphaned tool_call_id references in exported conversations.
The conversion logic for AIMessage only reads message.tool_calls (valid tool calls) and never checks message.invalid_tool_calls, so any failed tool calls are silently dropped from the assistant message. This creates an inconsistency because the subsequent ToolMessage still references the dropped call ID.
1. Create an AIMessage with empty content and invalid_tool_calls containing a single entry with name='weather', args='{"city":', id='failed_call', error='Invalid JSON'. 2. Create a ToolMessage with content='Invalid arguments' and tool_call_id='failed_call'. 3. Call convert_to_openai_messages([message, reply]). 4. Observe that the assistant message dict is {'role': 'assistant', 'content': ''} with no tool_calls key, while the tool message retains the reference to 'failed_call'.
Fixing Code Block
from typing import Sequence, List, Dict, Any, Optional, Union
from langchain_core.messages import (
BaseMessage, HumanMessage, AIMessage, SystemMessage, ToolMessage,
FunctionMessage, ChatMessage
)
def _format_invalid_tool_call(call: Any) -> dict:
"""Format an InvalidToolCall object or dict for OpenAI messages."""
if isinstance(call, dict):
call_id = call.get("id")
name = call.get("name")
args = call.get("args", "")
else:
call_id = getattr(call, "id", None)
name = getattr(call, "name", "")
args = getattr(call, "args", "")
return {
"id": call_id,
"type": "function",
"function": {
"name": name,
"arguments": args,
},
}
def convert_to_openai_message(message: BaseMessage) -> dict:
"""Convert a LangChain message to OpenAI format."""
if isinstance(message, HumanMessage):
return {"role": "user", "content": message.content}
elif isinstance(message, AIMessage):
ai_dict: dict = {"role": "assistant", "content": message.content}
tool_calls: List[dict] = []
if message.tool_calls:
tool_calls.extend([
{
"id": tool_call["id"],
"type": "function",
"function": {
"name": tool_call["name"],
"arguments": tool_call["args"],
},
}
for tool_call in message.tool_calls
])
if message.invalid_tool_calls:
tool_calls.extend(_format_invalid_tool_call(call) for call in message.invalid_tool_calls)
if tool_calls:
ai_dict["tool_calls"] = tool_calls
# Preserve other standard AIMessage fields if present
if getattr(message, "audio", None):
ai_dict["audio"] = message.audio
if getattr(message, "function_call", None):
ai_dict["function_call"] = message.function_call
return ai_dict
elif isinstance(message, SystemMessage):
return {"role": "system", "content": message.content}
elif isinstance(message, ToolMessage):
return {
"role": "tool",
"content": message.content,
"tool_call_id": message.tool_call_id,
}
elif isinstance(message, FunctionMessage):
return {"role": "function", "content": message.content}
elif isinstance(message, ChatMessage):
return {"role": message.role, "content": message.content}
else:
raise ValueError(f"Unknown message type: {type(message)}")
def convert_to_openai_messages(messages: Sequence[BaseMessage]) -> List[dict]:
return [convert_to_openai_message(message) for message in messages]
The fix adds handling for AIMessage.invalid_tool_calls in the convert_to_openai_message function. It formats each invalid call into the OpenAI tool_call structure using the call's id, name, and original arg string, and merges them with any valid tool_calls. This ensures that the assistant message carries all tool call metadata, preventing orphan references in subsequent ToolMessages.
Edge Case Audit
Merging invalid_tool_calls into the tool_calls list means that if the converted messages are sent directly to the OpenAI API, the API may reject the request because the arguments field contains invalid JSON. This could break existing pipelines that relied on the old behavior of stripping invalid calls before sending. Additionally, if an AIMessage has both valid and invalid tool calls, they are combined into a single list, which may confuse downstream consumers expecting only valid calls. Before deploying, test with real API calls. Rollback: revert the code change or pin langchain-core to the previous version if unexpected API errors occur.
