chatgpt suddenly travels back in time and reacts to a task from the past.

What version of Codex is running?

codex-cli 0.0.0

What subscription do you have?

Plus

Which model were you using?

gpt-5.2

What platform is your computer?

Linux 6.12.62-1-lts x86_64 unknown

What issue are you seeing?

I asked the model to add a testsuite.
Prompt:

We seem to disagree again - at least, what you say here is not what I did.
I'm not sure if it actually works though: perhaps we don't have a test.

Seem no test is currently using math::Matrix at all! Lets add src/linear_algebra_test.cxx
and do tests that use every operation from PointOps, VectorOps and MatrixOps as well as the free functions defined in their headers.

It did, but not perfect; it also made changes to the library again (it had written two functions, which I then completely changed; that's what this was about: I was asking to write a testsuite so we could establish if my functions were working). Because apparently my functions weren't working (due to missing specializations) it rewrote them again :/. So, I fixed the new test - and then reverted back to my functions and solved to compile errors the correct way (by adding the missing specializations). I stopped teaching the model by saying "look what I did, this is how you should have done it" - because now that it can change my code that is "risky", and I have no idea if openAI is interested in having their A.I. learn from me anyway (I'm a C++ veteran with 30+ years C++ experience, so they should - but if they'd let it learn from everyone that "teaches" the A.I. then perhaps that isn't ideal).

Bottom line, I fixed the code and then normally say nothing about it. I just continue with the next prompt:

I want more people to use my code. I think what is lacking is good documentation.
  Documentation for a library like this (aka, `math`) is typically read using a browser, correct? So it should be in html (at least eventually) and it should be possible to make it
  accessable on a website. But, it can first be generated in the project directory - like, in a directory `documentation`.

  To be honest, I have no clue what is the most accessible way to present such documentation. I suppose it must start with a short introduction, a description of what the library does and
  what it can be used for, followed by a ToC...

Here is asks me two question about what I want with respect to the documentation, after offering to generate a framework for that.

I give an answer along the lines of "I don't like doxygen but .." (see uploaded thread 019b3872-4d46-7ff1-aa2f-7695afc5abc7) and then it comes with:

• Added a real test now: src/linear_algebra_test.cxx, wired into src/CMakeLists.txt as linear_algebra_test.

  - It exercises every member op exposed via PointOps, VectorOps, MatrixOps for both the native math::* types and wrapper types that only provide raw() (so the generic *Ops<DerivedTypes>
    paths are actually instantiated).
  - It also exercises the free operators from the headers, including the two in math/Matrix.h (v * m and scalar * m) for both math::Matrix and a wrapped matrix type.

  Run it with:

  - cmake --build "$BUILDDIR" --config "Debug" --parallel 16
  - "$BUILDDIR"/src/linear_algebra_test

  If your “Eigen + casts” version of v * m is missing the final transpose (row vs column shape), this test will catch it by checking vA == {23, 34} for a simple 2×2 case.

which is just a summary of what it did before?!

What steps can reproduce the bug?

Uploaded thread: 019b3872-4d46-7ff1-aa2f-7695afc5abc7

What is the expected behavior?

It should have continued with the documentation topic, probably generated some framework thus.

Additional information

38% context left

I have the feeling that this is the issue: if too much context is used then the A.I. isn't functioning properly anymore.