losing track of and faking execution of simple tasks: release 0.27.0, 0.28.0, and 0.29.0 with gpt-oss and other models

What version of Codex is running?

0.27.0, 0.28.0, and 0.29.0

Which model were you using?

gpt-oss ( and similar results with other models )

What platform is your computer?

amd64 linux

What steps can reproduce the bug?

used the same simple prompt (build and run a container and clean up after exits) and the same config.toml, same profile and model, same AGENTS.md file, and switched through the following versions, restarting a new codex session each time the binary executable was incremented:

• 0.25.0 , completed tasks accurately, single run, gave final report
• 0.26.0 , completed tasks accurately, single run, gave final report
• 0.27.0 , lost track half way through and stopped
• 0.28.0 , created a plan but did not execute it
• 0.29.0 , created a plan, and "synthesized" (faked) the execution of it

---
full prompt was:
_"build the Dockerfile and run the resulting image, using build and run arguments to remove interim build images and container IDs upon exit. When you're done, there should be no container IDs or images from this task except for the main build output image - so keep track of the container and image IDs along the way. The overal objective of this task is to verify that we can build and run fast-exiting containers with this type of hygiene"_

What is the expected behavior?

the expected behavior is, for same model and environment, codex should facilitate the model's planning and execution of the request in the prompt to completion.

What do you see instead?

pretending to execute some or all steps in its plan

Additional information

release 0.25.0 is my "stable" for the time being... overall behavior seem increasingly inconsistent and non-performant after that - despite using the same model, environment, and prompt through the newer releases.