How to tell Codex NOT to ask for confirmation all the time?

[haimat]

I am sometimes a bit frustrated on how often Codex CLI gets back to me as a user with infos and updates, even though I just want it to work. A typical scenario in that context is that for example I tell the agent to work on this or that feature to implement in our code base and it starts working, then after a few minutes it gets back to me telling me about the first step that it has done now, waiting for my comment. I tell it fine, but you don't need to inform me about such intermediate updates, just work until you either are finished with the task or you have a critical question to ask. Then it tells me "OK, no more confirmation", only to get back to me after a few minutes with the next update 😞

I tried various prompts and variations, trying to tell Codex just to work, without stopping, only if it really needs something from the user. But no luck, i always gets back to me after a few minutes and waits for my confirmation or whatever. This is a pitty, because if I have to guide and confirm the agent every few minutes, then it is not really that much of a help to me. I want it to work on its own until a task is done. I know from Claude CLI using similar tasks and prompts that it can work for much longer without interruption, sometimes even for 2-3 hours, until the task is really done. This is much more of a help for us.

Do you guys have any experiences on this? Is this behaviour "hard-wired" into the Codex models, or is there better prompting for what I am trying to achieve?

[dick-longest]

The best thing to do is to insult him like crazy, I make an average git commit that takes about three rounds of conversation, first I will ask him to commit according to the specification, then he will tell me that he is in the sandbox and cannot execute the commit command without authorization, and then I start looking for permissions to open the sandbox, and after some time of searching and not being able to find it, I ask him and he will tell me that the sandbox is not controlled by me, it is triggered by him calling the command Then I had another conversation asking him to commit, and he would tell me that he was about to call the command to let me authorize and then do nothing, and then I told him again that I would already authorize him to execute the command, and then he would trigger the command pop-up, and I could choose to execute the command, which is behavior without any AGENTS.md, and the behavior will be worse with AGENTS.md However, with constant repetition and multiple rounds of abuse, it will become more normal, but other problems may arise, such as writing incorrect formatting such as \n in the commit body 🫤 When I complete a piece of code of about 50 lines, I usually copy and send the same content about 20 times in advance, and then promptly add it when he asks me to continue, even when it's just a meaningless request.

[Chapoly1305]

I genuinely don’t understand the purpose of the full auto mode if it keeps asking questions!