[Feature] Process new instructions while codex performs work to steer the model
Resolved 💬 5 comments Opened Dec 12, 2025 by guidedways Closed Dec 14, 2025
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What feature would you like to see?
This is something I've wanted for a long time but with the newer GPT 5.2 ExHigh model I feel this is needed even more:
Please add support for processing instructions _while_ codex performs, so that we can steer the model or help guide / clarify things while it's still investigating, planning an implementation
Additional information
Use case:
- Ask codex to investigate and plan on fixing a certain bug. Up front you're aware this isn't a one-shot; it'll take careful planning and you may have some idea on how this should be done
- You noticed codex is gathering _facts_ and describing what it's doing and how it plans to do a certain thing
- What codex is about to do is probably either a bad idea, or will result in duplicate code (in case you already have some code present in an existing custom framework etc) - you'd like to steer the model by saying
take a look at xyz class for existing methods
Right now, codex queues up messages and processes them once it's done, but by then it's too late as you may in fact want it to do something differently, requiring a major re-write. If you hit Cancel and tell it to do something differently, many times this seems to break the _flow_ and results in either a fresh new investigation or just a half baked solution.
5 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
You can interrupt codex at any time with ctrl-c and enter new guidance. It will then pick up where it left off. Does that meet your needs?
@etraut-openai not quite - you need to try this with claude code to see what I mean. Instead of stopping the model mid-way and risking confusion (at times it's running a bash command or doing something and cancelling ongoing work feels like the wrong thing), I'd like to _nudge_ the model or give it additional
hints(or guidance) as it works.For instance, Claude Code has supported this since forever - when you steer it as it works, many times the model is able to take in new guidance passively as a _by the way_ thing while it continues to do what it's doing and would use your guidance a few steps down the line. For instance, after I've given a prompt I may remember to say:
please run the XYZ unit test first to establish a baseline and then run it again after the fixThis isn't a new instruction, it's additional guidance that I'd like the model to take into consideration as it carries out work.
Here's another example:
Checking file inclusion into Xcode projectThis is where I wanted to just say
Don't edit the xcode project file - the new file will automatically get included as we're using Folders within xcodeI was too nervous to hit CTRL+C because the last time I did, it seemed to _restart_ its investigation even though I simply added
Only focus on changes in XYZ.swifteven though that's what I had originally implied but it began looking at the wrong changeset.Being able to steer the model live would help with all these situations and would avoid cancellation anxiety.
I think this feature request is a duplicate of #4312, which already has some upvotes. I'm going to close this one to consolidate. If you think that #4312 doesn't quite capture what you're looking for, feel free to add to it.