[Agent] Allow exposing some tool exceptions to the LLM

[valtzu]

Currently we always return the same static response in

ai/src/agent/src/Toolbox/FaultTolerantToolbox.php

Line 41 in 9f0eec0

return \sprintf('An error occurred while executing tool "%s".', $e->toolCall->name);

For example, if a tool execution is prevented due to AccessDeniedException, the LLM still sees this static error.

I think we could have some exception interface, like

interface ToolCallExceptionInterface 
{
    public function getToolCallResult(): ?string;
}

and then return that as tool call response instead of static message.

Use cases

1. The AI agent would be able to tell the user "Sorry mate, permission denied to perform the requested action".
2. The AI agent could auto-fix the issue of providing wrong parameters to the function call

[chr-hertel]

You mean as opt-in feature for a tool to throw?

[valtzu]

Yes, but not limited to tools, also related event listeners, like that IsGrantedToolListener and any userland listeners.

[chr-hertel]

true, understood, sounds good, yes 👍