I would like to report an issue I experienced with Codex yesterday and overnight.
Yesterday, I asked Codex to make a relatively small change to an existing project. Instead of completing the change, Codex apparently entered a continuously repeating loop.
For hours, the pattern looked roughly like this:
Codex stated that it was going to make certain changes or fixes.
This was followed by context compaction.
After the compaction, Codex again explained what it was going to change.
Another context compaction occurred.
Then essentially the same process started again.
What made this particularly noticeable was that Codex sometimes brought previously completed issues and changes back into the workflow. It appeared to lose track of the actual project state and repeatedly reconsider or redo work that had already been completed.
When I checked this morning, Codex had apparently spent around 9 hours running in this loop.
The bigger problem is that during this time my available weekly usage went from 100% to 0%, while the original task was still not completed.
It looks as if the combination of repeated context compactions and a long-running agent task may have caused Codex to lose track of its current progress, resulting in the same steps being planned or performed repeatedly.
This is especially problematic because such a loop can apparently continue for many hours and consume the entire available usage allowance without actually completing the requested task.
I think it would be useful for Codex to have additional safeguards for situations like this, for example:
Detecting when the same tasks or steps are being repeated multiple times.
Stopping or asking the user for confirmation when no meaningful progress has been made for an extended period.
Better preservation of task and project state across context compactions.
A warning when an unusually large amount of usage is being consumed by a single task.
Automatic termination of obviously repetitive agent loops.
Has anyone else experienced similar behavior, especially during long-running Codex sessions involving multiple context compactions?
I’m actually still a complete beginner when it comes to programming and have very little coding knowledge myself. So far, I’ve mainly used Codex for a small private project: I built a simple website for managing work schedules and vacation planning, with Supabase as the backend.
I have read some of the Codex Tips and Tricks threads, at least partially. However, I have to admit that I don’t understand many of the more technical details yet, especially when the discussions go deeper into programming concepts.
My English skills are also quite limited, which sometimes makes it even harder for me to fully understand the documentation and technical discussions.
That’s actually one of the reasons why Codex is so useful for me: it allows me to build and learn things that I probably wouldn’t be able to create on my own with my current programming knowledge.
Usually, I send ChatGPT a screenshot of what I want to reply to, explain in my own words what I would like to say, and then ask it to formulate the response for me in both English and German. That makes communication in communities like this much easier for me, especially because my English is limited.
ChatGPT is also a huge help for me in everyday life for another reason. Due to my physical limitations, I can only move one index finger, for example, so typing longer texts or doing many things manually on a computer can be very time-consuming for me.
With ChatGPT, I can describe what I want to do in just a few words, provide screenshots for context, and get help turning that into a complete message, explanation, or even instructions for a project.
So for me, AI is not only a tool for programming. It is also a form of accessibility that makes using a computer, communicating with others, and working on my own projects much easier.
Yes, exactly — I did the same thing for this post as well. I hope that doesn’t come across as rude or impersonal.
As I mentioned, ChatGPT is an enormous help for me because of my physical limitations and also because of my limited English. It allows me to express what I actually want to say without having to type everything myself.
I still decide what I want to say and give ChatGPT the context and my own thoughts. If the generated response doesn’t quite fit, I change parts of it, shorten it, or ask for another version until it reflects what I actually mean.
So the ideas and opinions are still mine — ChatGPT mainly helps me put them into words and makes communication much easier for me.
Unfortunately, I couldn’t observe the full 9 hours because I went to sleep and expected Codex to eventually finish the task on its own.
After my usage limit reset, I tried to continue and finish the exact same task in the same chat. This time I watched it more closely and noticed very quickly that it wasn’t making real progress. It kept going through the same pattern of thinking, context compression, and then thinking again without actually completing the requested changes.
So I stopped the run and opened a completely new chat inside the same project. In the new chat, Codex worked perfectly again and completed the requested changes very quickly.
That makes me suspect that the problem was related to the context window of the old chat constantly being at or near its limit. What confused me was that I assumed Codex would mainly need the context relevant to my latest change request. I didn’t realize that such a large amount of the previous project/chat history could continue to influence the context and repeatedly trigger compression.
So in my case, simply starting a fresh chat in the same project immediately solved the problem.
