It’s a mix of those “super prompts” ppl talked about a couple of months ago … and also my response to the LLM API closing the connection if I idle too long or saturate my tokens in a fast moving agent flow