यत् पिण्डे, तत् ब्रह्माण्डे (yat piṇḍe tat brahmāṇḍe)

People ask me how I choose what model to route queries to. It's simple. Claude gets knowledge work anything below that would be degrading Claude equivalent to having John von Neumann clean my toilet

yeah codex is realy bad at this. i've resorted to using a hermes worker with a custom supervisor profile (also using gpt-5.4) with a cron job that scans all of my current active codex sessions (in zellij) every 60m and nudges them forward by sending a message to initiate new turn when a nudge make sense. although a bit hacky in terms of tooling, this is working surprisingly well so far for overnight runs and 24*7 execution. will write a post on this soon. over the course of implementation, i've come to realize that my setup resembled openai's symphony but with more flexibility reagrding state transitions and agent choices etc. before coming up with this, i tried to give codex explicit instructions to implement a loop where it periodically monitors the output of a long running job and address the outcomes accordingly. but it consistently failed; it tried to use custom bash scripts first and couldn't get the loop working even when i tried to make it use subagents explicitly. i am often tempted to use claude code's `/loop` which seems like the perfect UX for this usecase but i have come to appreciate codex's reasoning and problem solving more for complex tasks like you mentioned in the other reply. rooting for @thsottiaux @OpenAIDevs @ChatGPTapp to improve codex on these product details to give the harness/UX the models deserve :)