What version of Codex CLI is running?

codex-cli 0.115.0

Which model were you using?

gpt-5.4

What platform is your computer?

Darwin arm64 (macOS)

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

Codex desktop / Codex TUI on macOS

What issue are you seeing?

In subagent sessions, the live TUI/subagent state surface sometimes corrupts the displayed text even though the raw rollout transcript is correct.

This is not just "markdown renders as plain text". The displayed line can:

• drop leading characters inside words
• remove punctuation or whitespace
• break inline-code boundaries
• partially mangle words in the middle of a sentence

The raw rollout for the same subagent session still contains the correct message, so this looks like a render/stream stitching bug in the subagent UI surface, not bad model output.

What steps can reproduce the bug?

I currently have two independent real reproductions from 0.115.0 subagent sessions in different worktrees/repos. In both cases, the corrupted text was shown live in the subagent UI surface, while the raw rollout for the same session was correct.

Repro A, subagent session 019cfe23-1c1a-7992-ba9c-86a7033c7c01:

TUI showed a corrupted line like:

• ейты проходят. Проверяю толькоVisual Acceptance`, чтобы корректно закры reviewer-флоу по этому plan и не застрять на неществующей UI-проверке.

But the raw rollout for the same session contains:

Гейты проходят. Проверяю только `Visual Acceptance`, чтобы корректно закрыть reviewer-флоу по этому plan и не застрять на несуществующей UI-проверке.

Repro B, subagent session 019cfe07-ce54-73b3-b5c5-a545aeccf4d5:

TUI showed a corrupted line like:

Evidence собожу reviewer artifact. Открою один-два свежих backend review-файла чтобы сохранить repo-local формат и правильно заироватьS` vs `

But the raw rollout for the same session contains:

Evidence собран; перехожу к reviewer artifact. Открою один-два свежих backend review-файла, чтобы сохранить repo-local формат и правильно зафиксировать `S2` vs `N/A` для visual review.

What is the expected behavior?

The subagent TUI/state surface should display the same text that is present in the raw rollout transcript, preserving characters, whitespace, punctuation, and inline-code boundaries.

Additional information

These two reproductions are useful because they are not the same exact workflow:

• they come from different worktrees/repos
• they affect live subagent commentary text, not just one static final handoff line
• the corruption pattern is similar in both cases
• the raw transcript remains intact in both cases

This seems related to, but not identical with:

• #13220 — subagent state markdown rendering is wrong
• #10901 — main TUI can cut off or oddly space output while the transcript remains correct

My current read is:

• #13220 is about incorrect markdown rendering in the subagent state surface
• this bug is about the subagent state surface corrupting the actual displayed text
• #10901 may share the same underlying rendering/streaming cause, but this report is specifically about subagent live text corruption with the raw rollout remaining intact