In the TUI, loss of the text bridge (truncated/no final output) and duplicate tool calls - vs one-off "codex exec" runs consistently printing complete json and a clean final response

What version of Codex is running?

0.20.0

Which model were you using?

gpt-oss

What platform is your computer?

linux kernel 6.8 , x86_64

What steps can reproduce the bug?

In the TUI, longer tool chains and processing seem lose the text bridge (truncated/no final output) and often show duplicate tool calls with one of them failing with the same syntax

...whereas "codex exec" ( one-off-run-and-return to shell with simple stdio in and out) consistently prints complete json and text of the interim tool chain and processing; and then a clean final response before exiting to the calling process/shell.

What is the expected behavior?

tool calls and interim processing output in the TUI is expected to be as stable and complete as what we see in a sigle codex exec run

What do you see instead?

with tui, if processing a prompt for over 600 to 1000 seconds or so before giving up back to the tui prompt, then continuing with the session is usually non-responsive until /new or exit-then-rerun the TUI

Additional information

_No response_