Responses dropped in TUI: "OutputTextDelta without active item"

What version of Codex CLI is running?

codex-cli 0.116.0

What subscription do you have?

Enterprise

Which model were you using?

gpt5.4

What platform is your computer?

Darwin 25.3.0 arm64 arm

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

iTerm2

What issue are you seeing?

Having issues seeing replies from the model when using gpt5.4. I can see output if I switch to a different model (gpt-codex5.3 in this case). I asked it to debug the issue and generate a report of its findings. Here is the output of that investigation

### Version / Environment

• Codex CLI: codex-cli 0.116.0
• OS: macOS (zsh)
• First observed: March 24, 2026 (America/Los_Angeles)
• Reproduced repeatedly on March 24–25, 2026

### Summary
When using gpt-5.4, turns complete but no assistant message is rendered in the shell UI.
The logs consistently show parser errors:
codex_core::util: OutputTextDelta without active item.

### Repro

1. Start Codex with gpt-5.4 (or switch to it via /models).
2. Send a simple prompt (example: test?).
3. Observe no visible assistant response.
4. Switch to gpt-5.3-codex.
5. Send a simple prompt again.
6. Assistant response appears normally.

### Expected
Assistant response is shown for gpt-5.4 like other models.

### Actual
For gpt-5.4, task completes but no assistant message is shown.

### Evidence

• Repeated log error (example):
• 2026-03-25T04:21:51.767031Z ... model=gpt-5.4 ... OutputTextDelta without active item
• Same session behavior:
• gpt-5.4 turns end with last_agent_message: null
• gpt-5.3-codex turns in same environment produce normal output_text
• Session-level counts from rollout archives:
• gpt-5.4: 14/14 turns had no assistant message
• gpt-5.3-codex: 7/7 turns had assistant message

### Workaround
Pin model to gpt-5.3-codex (works reliably in this setup).

Related issues you can link in the report for context:

• https://github.com/openai/codex/issues/7377
• https://github.com/openai/codex/issues/6279
• https://github.com/openai/codex/issues/5081

What steps can reproduce the bug?

1. Start Codex with gpt-5.4 (or switch to it via /models).
2. Send a simple prompt (example: test?).
3. Observe no visible assistant response.
4. Switch to gpt-5.3-codex.
5. Send a simple prompt again.
6. Assistant response appears normally.

What is the expected behavior?

Seeing the assistant's response

Additional information

The workaround of pinning to a different model works. Additionally if I pin a different model and tell the assistant to run the query with gpt5.4, it tells me that a particular response is coming from 5.4 and I can see that response (though unsure if the assistant is lying to me)