Bug: occasional raw JSON subagent notifications render in Codex chat on GPT-5.4 high

What version of Codex CLI are you running?

Unknown from UI session; occurred in local Codex chat workflow.

Which model were you using?

GPT-5.4 with reasoning effort set to high.

What happened?

Occasionally, Codex chat renders what looks like an internal raw JSON payload instead of a normal assistant message. In this case it surfaced a subagent notification payload directly in the chat transcript.

Example symptom:

• raw JSON-like content starting with fields such as author, recipient, other_recipients, and content
• embedded <subagent_notification> XML-like wrapper
• user-visible output instead of a normal formatted assistant response

Steps to reproduce

I do not have a minimal deterministic repro yet. The issue was observed during a longer coding session that involved subagents and review/merge work.

Expected behavior

Internal subagent/event payloads should never be shown to the user. They should remain internal and be rendered as normal assistant text if surfaced at all.

Actual behavior

A raw JSON payload was shown directly in the chat UI.

Additional context

• Uploaded thread id: 019d298d-23ad-7730-89eb-081a702ce5d2
• Context: Codex chat session using GPT-5.4 high
• The visible payload appeared to be a subagent notification/result object

If helpful, I can provide more context from the thread above.