bugfix: Bug in which context_id wasn't stable across REPL messages#480
Conversation
| while True: | ||
| input_text = input(colored(text="Input: ", color="blue")) # noqa: F821 | ||
| output = responder(input_text) | ||
| output = responder(prompt=input_text, context_id=self.context_id) |
There was a problem hiding this comment.
@eob I like this PR. And this seems totally fine. Do we need to be concerned at all with people trying to use the REPL for functions that aren't prompt and tripping on the context_id arg, or should we enforce that you must take context_id as an arg, or some other option.
There was a problem hiding this comment.
Yeah that's a really good question. And I think the problem you're observing extends even to the prompt: str argument, right? There's an unstated requirement about the signature of responder.
Proposal: what if we allowed ToolRepl and AgentRepl to make these kinds of assumptions now, since that's probably the 90% case, but then we planted a stake in the ground (your enforcement suggestion above?) about the "chat I/O contract" that an AgentService must, at minimum, provide.
Here, I suppose that's implicitly:
prompt(prompt: str, context_id: str) => List[Block]
though I imagine if we were being proactive about defining it, we'd probably want to extend that a bit:
prompt(prompt: Optional[str, Block], context_id: str) => List[Block]
Or even something further..
There was a problem hiding this comment.
I think we should standardize. Maybe with some nasty combo of lists
2 participants
No description provided.