LONG running looping conversation task

What feature would you like to see?

All other LLMs can easily do this, e.g., Gemini and Claude Code, but Codex never does it; it always reports.

I know that it's 'correct' behavior to always report to the user, but sometimes I like brainstorming with multiple AI agents to get all their perspectives, so I'll create a chat space, e.g., a simple text chat file where all write to it with names.

So I introduce everyone, and we think together; e.g., I present a problem and both sides give some ideas/critiques, etc. When not in agreement, they can even debate among each other with me as the mediator. This helps me think of stuff I might have missed alone, as each agent can pinpoint edge cases I might not have thought of.

I can run this easily using two Gemini agents, or Claude, or a mixture of both. They know how to keep looping and reading the chat, but Codex fails each time to listen to the precise instruction to loop and read the task file and only respond there. For example, I tell it to do that, then the next message it says 'Understood.' I tell it I mean start now, then it looks once then reports back, e.g., saying 'I'm continuing.'

Compared to the other two agents, Gemini and Claude, in general, Codex seems to follow instructions better in general, so it's a pity I can't use it in this workflow.

I understand it's because it was trained to be a sort of chatbot that responds in chat, but the other agents seem to be able to adapt their training to always use the created chat environment instead of directly replying in the CLI.

Thanks.

My workaround is literally to que many messages one after another, telling it to "continue, and read the chat file".

Additional information

_No response_