How to handle follow-up actions when the agent is aborted during a tool invocation.

[phi-friday]

Question

First, thank you for providing the ag_ui integration feature—I’m using it very well.

However, while recently testing the abort function, I encountered the following error:

data: {"type":"RUN_STARTED","threadId":"553632f0-09a6-4053-8c95-61b85b03a1b9","runId":"8735d904-7b17-4adc-b7fe-3dca71bf0fb8"}

data: {"type":"RUN_ERROR","message":"status_code: 400, model_name: gpt-4, body: {'message': 'No tool output found for function call call_tYFQBQ4RTmvnbPkV40KWCKCh.', 'type': 'invalid_request_error', 'param': 'input', 'code': None}"}

Right after the error occurred, agent.messages is as follows.

[
  {
    "id": "019a9502-57fb-7209-a94f-c47d2b2f3647",
    "role": "user",
    "content": "test"
  },
  {
    "id": "d9d864d5-3adb-4c92-b42d-489b65e68639",
    "role": "assistant",
    "toolCalls": [
      {
        "id": "call_tYFQBQ4RTmvnbPkV40KWCKCh",
        "type": "function",
        "function": {
          "name": "retrieve_documents",
          "arguments": "{\n  \"search_query\": \"test\"\n}"
        }
      }
    ]
  },
  {
    "id": "019a9502-d49b-7159-bfc5-dc3142d2255b",
    "role": "user",
    "content": "test"
  }
]

I executed abort while a tool was running.
After it was aborted, when I entered a new message, there was still a message indicating a tool had been called, but there was no result from the tool in the message, so all subsequent workflows seemed to fail.

It seems to be related to this issue #2050.
Is there a good way to resolve this?

For a new message, is the best approach to check whether a tool call has results; and if not, to explicitly enter a failure result?
If this is an issue that should be handled on the frontend rather than by pydantic-ai, just let me know.

Additional Context

No response

[dsfaccini]

hey @phi-friday , thanks for raising the issue! indeed when you abort a tool call and run the agent again the assistant sees that it called a tool but doesn't see any output from that tool, so it fails with a 400.

For a new message, is the best approach to check whether a tool call has results; and if not, to explicitly enter a failure result?

to your question on how to best solve this, checking every tool call for a result is a bit defensive, you could instead catch the error on the server and check the message like except ModelHTTPError as e: if 'tool output' in e.message: message_history = clean_orphaned_tool_calls(previous_history)

def clean_orphaned_tool_calls(message_history: list[ModelMessage]) -> list[ModelMessage]:
    if not message_history:
        return message_history

    cleaned_messages = []
    i = 0

    while i < len(message_history):
        msg = message_history[i]

        if isinstance(msg, ModelResponse):
            tool_call_parts = [p for p in msg.parts if isinstance(p, ToolCallPart)]

            if tool_call_parts:
                has_returns = False

                if i + 1 < len(message_history):
                    next_msg = message_history[i + 1]
                    if isinstance(next_msg, ModelRequest):
                        tool_return_parts = [
                            p for p in next_msg.parts if isinstance(p, ToolReturnPart)
                        ]
                        call_ids = {tc.tool_call_id for tc in tool_call_parts}
                        return_ids = {tr.tool_call_id for tr in tool_return_parts}
                        has_returns = bool(call_ids & return_ids)

                if not has_returns:
                    i += 1
                    continue

        cleaned_messages.append(msg)
        i += 1

    return cleaned_messages

but you could also handle it on the client if you're more comfortable with that.

[DouweM]

@phi-friday Processing the history to remove "dangling" tool calls is indeed the way to go, either on the server side (like @dsfaccini suggested) or on the client side.

[phi-friday]

@dsfaccini @DouweM

Thank you very much for your kind response.
I believe that even if an error occurs during the chat, it should not affect the user's request, so I think I need to handle messages more strictly by considering abort situations.
Thank you for providing detailed explanations and especially for presenting sample code on how to handle this.
Have a great day, both of you!