Codex stuck in Loops
Resolved 💬 5 comments Opened Feb 16, 2026 by thom-heinrich Closed Feb 16, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
I am an ultra heavy codex user, keeping track on basically every metric you can ablate from the tool.
Currently gpt 5.2 xhigh tends to a) forget where it was b) treat it own responses as user input and c) "solves" the same problems over and over again - this happens also with plan mode.
This leads at the moment to a drastically worse performance and the resolution rate is from ~40% in my repo down to around ~10% or lower. It struggles with basic tasks like browser automation, prompt compiling etc...
I am using Codex CLI as Pro User /w GPT 5.2 xhigh and the latest release.
5 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
GPT-5.3 Codex is better at this. Generally, use /new after each task (especially very different tasks) to ensure clean context management. GPT-5.3 is coming soon, and it will certainly include improvements to avoid unnecessary repetition.
It makes no sense to use xhigh in the standard. Especially not in long chat sequences. The OpenAI models have been very good at understanding context since 5.2, but they are not perfect yet.
Thanks for the bug report. We love to hear from codex power users.
What you're describing here is model behavior, not something that's under the control of the harness. As you noted, gpt-5.3-codex is better at this than gpt-5.2, and we continue to make improvements to newer-generation models.
If you see behaviors like this, it's helpful if you use the
/feedbackcommand to upload the session details. This helps the team responsible for training our models understand where improvements are needed. You can skip opening github bug reports for those issues.Came here to see if this was being spoken about.
I send my agent on some long-running task and tell it to do a sequence of steps (let's call them steps 1, 2 and 3). It completes most of those steps (1 and 2) and begins the longer-step (3).
It gets half way through 3 and compacts.
The agent then re-does step 1 after compacting when step 1 was already completed before, and it doesn't continue where it left-off with the half finished step 3.
At any size, reasoning constantly goes into an endless loop of analysis instead of writing code according to plan... model 5.3, sent feedback with ID 019caa2c-4ae3-7090-97e5-2d3fe438d8b6.