add execution status to FunctionCall and example for history compaction with openai realtime#5024
add execution status to FunctionCall and example for history compaction with openai realtime#5024
Conversation
…on with openai realtime
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
| tail_items: list[llm.ChatItem] = [] | ||
| for item in self.realtime_llm_session.chat_ctx.items: | ||
| if item.created_at > last_updated_at or ( | ||
| item.type == "function_call" and item.status != "completed" |
There was a problem hiding this comment.
it's not a good ux to ask user to care about this, but I am not sure if it's a good idea to add the guard of inflight tool call into the update_chat_ctx. maybe create a new api for truncate operation for realtime session?
There was a problem hiding this comment.
To me, the biggest concern is that if you are unlucky with timing and another chat context update happens at the same time as a new function call arrives, you can see get the problem.
If we had a different method for sending precise context updates using the _update_chat_ctx_lock, that could be a solution to the problem as well. So long as the example also provides a way to update local contest without update_chat_ctx.
| extra={"item_id": item.id}, | ||
| ) | ||
| fnc_call = llm.FunctionCall( | ||
| id=item.id, |
There was a problem hiding this comment.
nit: use item_id for consistency
| compacted_history.items.extend(tail_items) | ||
| logger.info(f"keeping {len(tail_items)} items from realtime session") | ||
|
|
||
| await self.update_chat_ctx(compacted_history) |
There was a problem hiding this comment.
It seems that the original system message is missing. I tried this example, and the agent started speaking Spanish after compaction.
3 participants
fix #5021