TUI renders the same assistant reply twice when identical content is emitted as `commentary` and `final_answer`

What version of Codex CLI is running?

codex-cli 0.116.0

What subscription do you have?

Not specified

Which model were you using?

gpt-5.4

What platform is your computer?

Linux DESKTOP-ULILLCA 6.6.87.2-microsoft-standard-WSL2 x86_64 GNU/Linux

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

Not identified. Observed TERM=xterm-256color.

What issue are you seeing?

In the interactive Codex CLI/TUI, a single assistant reply can be rendered twice with identical text.

What I observed:

• I sent one user message: start
• Codex produced the same assistant text twice
• The terminal displayed both copies back-to-back

I checked the local raw session transcript afterward and confirmed this was not a duplicate user submission. The same assistant text was emitted once as commentary and then again as final_answer for the same turn.

So the symptom is either:

• the client/runtime should not emit identical content in both phases for one turn, or
• the TUI should deduplicate/suppress the repeated final_answer when it is identical to the immediately preceding commentary

What steps can reproduce the bug?

I do not yet have a minimal deterministic repro.

What happened in the observed case:

1. Open Codex CLI interactive TUI
2. Work in a normal session
3. Send a short message: start
4. Codex asks a clarifying/design question
5. The exact same assistant message is shown twice in a row

This happened on 2026-03-24.

What is the expected behavior?

Only one visible copy of the assistant message should be rendered for that turn.

If both commentary and final_answer are intentionally emitted internally, the UI should not show two identical copies back-to-back.

Additional information

I verified the duplicate from the local session JSONL rather than only from the screen.

Observed ids:

• thread/session id: 019d1eaa-bc32-7e71-bcb6-00ae694062f8
• turn id: 019d200e-74df-7440-933c-bd402b1147f1

Sanitized evidence from the local transcript for that single turn:

{"type":"event_msg","payload":{"type":"agent_message","phase":"commentary","message":"I’m using the `brainstorming` workflow here because phase 2 is a new AO subsystem and needs a real design before we touch code.\n\nFirst scope question for phase 2: should `ao doctor` in this repo be `diagnose-only`, or should it also be allowed to perform a small set of safe local repairs automatically..."}}
{"type":"response_item","payload":{"type":"message","role":"assistant","phase":"commentary"}}
{"type":"event_msg","payload":{"type":"agent_message","phase":"final_answer","message":"I’m using the `brainstorming` workflow here because phase 2 is a new AO subsystem and needs a real design before we touch code.\n\nFirst scope question for phase 2: should `ao doctor` in this repo be `diagnose-only`, or should it also be allowed to perform a small set of safe local repairs automatically..."}}
{"type":"response_item","payload":{"type":"message","role":"assistant","phase":"final_answer"}}

Important detail:

• I found only one user message for that turn
• this does not look like pressing Enter twice
• this does not look like an external orchestrator resending input
• it looks like one turn producing identical commentary and final_answer, both rendered visibly