Make request_user_input tool available in code mode

This is currently by design. I've changed your bug report into a feature request.

+1. I need this feature, it will be useful

+1
We should be able to benefit from the request_user_input tool when using custom Planning/Brainstorming instructions.
For example I am heavily using Obra Superpowers and it would be great to include request_user_input into the brainstorming skill. Please re-think this design choice, thanks in advance.

My use case is that I often want to be doing workflows that require a large number of user questions (categorise 50 items, for example) and it's too much to do all at once. What would be possible is to do one by one - codex asks the question, user selects the answer, codex does the task, codex asks question 2, etc. Would be a very nice way of working with codex spark I think.

I have reflected on this and I think this feature wants to be it's own "Interactive" mode in addition to Plan and the default. And then once you're adding more modes, you may as well open up an interface to users to allow custom modes and user engagement methods.

please make this feature available in other modes aside from plan mode. its a non negotiable in my opinion. there's so much utility for it outside of planning. thank you so much!

there are literally dozens of tweets on X praising this mode in Claude, and they are not talking about using them inside of plan mode

https://x.com/dani_avila7/status/2022157433733341254?s=20
https://x.com/bytes032/status/2023005871157227888?s=20
https://x.com/SarahChieng/status/2024555081056620658?s=20
https://x.com/andrewmccalip/status/2024332410545225784?s=20
https://x.com/pdrmnvd/status/1993451295152484409?s=20
https://x.com/heykojo/status/2008284215452848445?s=20
https://x.com/d4m1n/status/2014379167492559190?s=20
https://x.com/justinmfarrugia/status/2012179986719867182?s=20
https://x.com/omarsar0/status/2007095834744144102?s=20
https://x.com/whereisxavier/status/2012472071591207122?s=20
https://x.com/peterpme/status/2015081381307498708?s=20
https://x.com/nurijanian/status/2011630310434488633?s=20
https://x.com/alxfazio/status/1987163032128999764?s=20
https://x.com/30Thrd/status/2017277769382609142?s=20
https://x.com/sairahulxGTM/status/2024015008930074890?s=20
https://x.com/rpenacastro/status/2019724509679607838?s=20
https://x.com/pHequals7/status/2018421004641632417?s=20
https://x.com/gjohnsx/status/2017292285214371985?s=20
https://x.com/waydegilliam/status/2016335814595379261?s=20

There are so many more of these.

I understand that some people might not want it to randomly pop up and interrupt their session.

They might want it to be more autonomous. In that case, maybe you can either:

• bring back pair programming mode
• or create a feature flag so it doesn't pop up in code mode unless its on

or something like that.

there has to be a way for you to add it back without upsetting people

for example, you could tell it not to use request user input tool unless: you're in plan mode, or the user explicitly asks you to.

there are many solutions, i'm sure you can find a great one!

This is very useful feature for some advanced skills.When I was using Claude code, I created some skills that require model to ask user complex and detailed questions during execution mode, but this skills cannot adapt to codex due to current issue.

I also think that it could be helpful outside of plan mode. For example to fill up forms, wizards and as @zsh2401 said, it could allow more complex skills interactions

This is a very good idea! 👌

Concur. I often wish Codex would stop coding to ask me a question rather than continuing without direction.

This would be a great addition to help me when I get brain fog

This thread is getting busy - if it's of interest, I have implemented myself a custom mode by forking the codex repo and it works quite well. With any solution, however, the more important thing is successfully prompting the agent to request input from the user, and this is a challenge as they are RL'd to keep going and going (for good reason). We currently have the ability to use the default mode to get the agent to ask you for input, it just comes in the form of a usual response, rather than the nice tool interface.

This would make working with codex so incredibly collaborative and feel so much more powerful. Please implement 🙏🏽

Please add this.

This is very useful tool to chat with ai and discuss this, not only in plan mode

We NEED this!

+1 should not need plan mode to use this

• 1 - the UX is nicer than Claude too.

+1. This will guide Codex toward the intended approach and improve code quality.

+1 - Will be a very nice addition, we can add a config constant so those who "may not want to interrupt 'the flow'" can disable it

Yes, add a pair programming mode flag please.

I'm constantly switching between Claude and Codex because of this one missing feature. Would be great if it got added to Codex so I didn't have to keep breaking my flow.

👍

This is something we also badly need. Sometimes the spec changes during the implementation, and you can create skills that can surface contradictions. Allowing agents to grill the user for confusing instructions mid-implementation is quite helpful. Of course this should be something that can be turned off.

Thank you so much to everyone who commented and reacted. You're all amazing.

The feature is live in 0.106.0: #12735

enable feature flag default_mode_request_user_input = true to enable it.

The models are explicitly guided to make assumptions and only stop if it is blocked so it should not be disruptive at all. You can invoke it by asking it to ask you questions or say ("use request user input tool to interview me about xyz").

This will enable to you to make some fun skills as well, like custom plan modes, interview modes, and more.

The Codex team is the best in the world.

Thank you @aibrahim-oai @charley-oai and @etraut-openai for getting this across the finish line so quickly.

We really appreciate it. I know a lot of people are happy to hear about this!

Many thanks, guys, for you work!

Awesome! Thank you for listening to our feedback.

Is it currently possible to do something like this in a non-plan-mode prompt or inside a skill?

At the start, temporarily prefer request_user_input over assumptions when any of these would materially change …

If so, could someone correct this example using canonical language to enable—and later disable—this proactive elicitation behavior?

This does not work properly for me at the moment
I added default_mode_request_user_input = true to config.toml
I do not see this option in https://developers.openai.com/codex/config-reference
For same prompt (it's complex agent skill) it works after switching to plan mode

@dydzio0614 it's most likely because you have to put it under the [features] section in config.toml. I had the same issue but https://github.com/openai/codex/issues/15293#issuecomment-4125030821 cleared it up