Across multiple chats and multiple repos, the new codex model is spending a lot of time running python commands to get information but those commands often fail:
$ python - ‘<<PY’ import torch, sys path=saves/trl_qwen3_vl_4b_naive/training_args.bin ‘args=torch.load(path)’ ‘sys.stdout.write(ftype={type(args)}\n)’ PY
It then wonders why there is no output from the code and goes in circles trying to produce output.
eventually running stuff like:
$ python - ‘<<PY’ ‘print(hello)’ PY
and
$ python -c ‘print(’“'”‘test’“'”‘)'