GPT 5.5 under codex decides to make unilateral decisions that can cause breaking changes to existing code bases when given instructions after many compacted contexts

What version of Codex CLI is running?

codex-cli 0.130.0

What subscription do you have?

Plus

Which model were you using?

gpt-5.5 (reasoning high, summaries auto)

What platform is your computer?

Linux 6.17.0-23-generic x86_64 unknown

What terminal emulator and version are you using (if applicable)?

Konsole running zsh

What issue are you seeing?

GPT 5.5 under codex decides to make unilateral decisions that can cause breaking changes to existing code bases when given instructions after many compacted contexts. A number of times I have given codex instructions like "comment out this particular code so we can uncomment it later" and its response is always "delete the code and replace it with a placeholder" instead of following instructions..

This behavior happens in many different situations, mostly invovled with editing existing code, it makes a lot of unilateral decisions which were not really requested , even when a specific request was made which has nothing to do with its chosen action.

I have seen gpt 5.5 xhigh do things like modify files that did not need to be modified based on partial directions in AGENTS.md which referenced specific scenarios like

``When writing code in QHTML, if you run into a situation where the framework does not support the operation you are trying to perform in a declarative manner, rather than trying to do a javascript workaround, consider adding new features to the QHTML framework itself``

Which i interpret as create a new abstracted feature that will facilitate the simple creation whatever the problem is

Example: qhtml does not support the ability to loop over an array - codex was doing something like this
q-component somecomp {
q-property myprop
onReady {
for (var i=0; i<myprop.length; i++) { var elem = document.createElement("some-elem"); document.querySelector("somecomp").appendChild(some-elem) }
}
}

Which was causing the newly created elements to not be registered in the qhtml engine so then it started implementing all the lookups and everything right tthere in the onReady { } block like getContextNamedReferences(element) blah blah and it like re-built half the runtime inside of a simple component.

So I added that line about adding new features because I wanted codex to do something like

add a new language feature ``for (item in myprop) { some-elem { otherprop: item } }``

then use that syntax after creating the language feature..

and it does just that in GPT 5.3-codex and other previous versions.. but GPT 5.5 does not.. instead it resorted back to the creating massive javascript workarounds, only this time its deleting existing code that serves some other purpose because its name is similar to the thing its creating or changing and it removes it at the framework level then adds in the workaround specific to the q-component, whereas q-components are part of an entirely differente layer in the model that qhtml has

Just thought you might want to know about this strange behavior...

What steps can reproduce the bug?

Uploaded thread: 019dd681-c1c2-7de0-a7e6-0d16cd9f0ce1

What is the expected behavior?

in the case of the session from the transcript log, there are numerous situations where the behavior should have been to more closely follow the AGENTS.md and the prompt instructions.

Additional information

It is possible that this is just too complex for ChatGPT to be able to understand.. which it may very well be..