To support configuring how errors are handled in tools.

[ekzhu]

Should be part of the configuration parameter for FunctionTool.

[wistuba]

How do you imagine this to work? Like this?

caught_errors: Sequence[Exception] | Exception
stock_price_tool = FunctionTool(
    get_stock_price,
    description="Fetch the stock price for a given ticker.",
    caught_errors=caught_errors
)

where caught_errors defaults to Exception. Any error not in caught_errors is raised.

Better name for caught_errors would be good.

[ekzhu]

I think what you proposed is good. Perhaps name it "return_errors", as the error messages are being returned.

[wistuba]

Quick follow-up: this is not specific for the FunctionTool, right? This will change the Tool protocol and add a new property returned_errors, right? The logic dealing with exceptions, e.g., AssistantAgent._execute_tool_call assumes a BaseTool, not a FunctionTool.

[ekzhu]

Quick follow-up: this is not specific for the FunctionTool, right? This will change the Tool protocol and add a new property returned_errors, right? The logic dealing with exceptions, e.g., AssistantAgent._execute_tool_call assumes a BaseTool, not a FunctionTool.

Yes if it is a boolean field as I think the returned value itself is the error message.

Though I think the configuration should be on FunctionTool. it should be on the implementation of BaseTool to correctly specify the returned_errors.

[ekzhu]

It does seem like a breaking change... Is it possible to bypass BaseTool for now? Having the error raised directly from BaseTool, and use the is_error field in FunctionExecutionResult to inform the model of the error?

[ekzhu]

Ah.. Okay appologies for the confusion. I misunderstood what you meant by returned_errors -- I thought it meant the errors are returned as part of the ReturnT of BaseTool.

Please ignore my two comments above.

It makes sense. Let's move the conversation to the PR.