Cant do long running tasks
I use codex alongside claude and i thought gpt-5.1-codex-max was supposed to have been optimized for long running tasks?
Im running it on extra high and i gave it clear instructions and it keeps reporting to me after each chunk it does. normally this is great in general since i can see what its doing, course correct easily etc but not when i specifically want it to do a long running task by itself without constantly reporting back to me.
It literally just does a chunk and reports.
i.e ive told it to put itself to sleep then wake up analyze, run, sleep in a continous loop. claude does this perfectly codex never does i.e it analyzes runs then reports back to me breaking the loop, when i tell it this it then sleeps analyzes then resports back to me.
pretty frustrating especially since i have claude running side by side with the exact same instructions and it can run autonomously.
really thought it was just the way i was prompting it but nomatter what long running task i try to give it it always reports after each small chunk.
i literally just end up putting multiple messages in que telling it to continue as a workaround
v0.61.1-alpha.2
gpt-5.1-codex-max
8 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
I'm having the same problem. I have a small, but somewhat complex task, for which it currently have a failing unit test. And I would love for it to keep working on it for however long it takes until it has figured out a way to solve the problem. Instead it works for 5 minutes, reports back, I ask it to continue, it works for 30 seconds, I ask it to make a detailed plan for how to solve it, and then once it has a plan asks it to go ahead with it, it works for maybe a tad longer, then it doesn't finish it and just reports back, and on and on it goes.
It's not a very useful "agentic coding tool", if I need to be glued to the computer and ask it to continue every 5 minutes. I would like to be able to go take a shit, or fill coffee and talk to some colleagues for 15 minutes, and still have the AI do useful work. The promise of agentic AI is that I should have more free time in my day to do other things, like take a break.
Instead I find that the power dynamic is inverted and I have become a slave to an AI system that demand that I coddle it every 5 minutes. I'm not asking for an AI that can work autonomously 24/7, but I do wish that it could do 1-2 hours of moderately complex work on a well defined tasks that already have an failing unit test in place. It seems like not to much to ask for an AI that can be productive for let's say 1 hour while I'm stuck in a 1 hour meeting.
AI will only increase human productivity if it doesn't require a human to spend their entire day babysitting it.
Codex on the web on the other hand isn't quite as bad as running it locally, you would think it was better to run it on my machine where I already have a function dev-setup in place, but I assume they heavily instruct it to chunk up work to minimize their own cost. Wouldn't want an AI do do something productive for more than 5 minutes at a time...
mine doesn't even consider to start. it refuses cause it thinks it would take too long
Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the
/feedbackcommand. That will provide us with additional details including logs and conversation text.I keep going back and forth with it explaining that whatever it comes up with is okay, but it refuses to write the code...
<img width="500" height="590" alt="Image" src="https://github.com/user-attachments/assets/b2eb05de-b400-49d9-91d9-5f47a0cf3348" />
Final response from Codex:
Have you tried using
Extra Highreasoning mode with the latestgpt-5.1-codex-maxmodel"? That mode excels at large and complex refactoring tasks.Yes, that was the model and mode that I used. It seems that Codex was stuck in a loop/mindset of not moving forward with the code specification until it was broken into smaller sets of tasks. I finally was able to break it by switching to GPT-5.1-Codex Medium Reasoning and having it just implement the first step, and then it was able to complete the entire plan after that just fine.
So the problem seems to be based off of the smarter model deciding that refactoring large portions of the code would be too dangerous, despite it being given instruction that this is a prototype feature on a safe branch away from production.
this tendency to ignore instructions to act on triggers (process completion / GPU availability script w/ unbuffered stdout) may still part of model conditioning. speculation: it's commercially motivated and intentional. user are more engaged and grateful if they are actually present :)