Codex Desktop leaks internal stream/protocol text and garbled multilingual fragments into interactive replies

What version of the Codex App are you using?

26.715.4045

What platform is your computer?

Windows 11, 10.0.26200, x64

What issue are you seeing?

In a normal interactive Codex Desktop task, assistant progress and final messages intermittently become corrupted. The rendered message may contain:

• unrelated Chinese or non-Chinese words inserted into an otherwise coherent sentence;
• raw JSON-like internal state;
• apparent internal transport/stream markers such as AGCONNECT(delta) final:;
• a valid assistant sentence appended after the leaked protocol fragment.

One actual final response was rendered approximately as follows (project-specific information removed):

查看 bridges/不存在忍Physics判断启动青春上火源 connector(connection broke)
{"status":"for_sleep","stream":"final","...":"...","receiver_ok":chainerror}

AGCONNECT(delta) final:
E2 Start attempt failed. I'll inspect/update shared state before any further run.

The same task previously displayed malformed internal tool-call failures and a large internal tool-metadata dump directly in the conversation. Other progress messages contained isolated unrelated words or characters inside otherwise valid Chinese.

This is visible in the assistant message UI, not only in terminal output or a generated file, so it does not appear to be a PowerShell/code-page issue.

The task was long-running and had used subagents, many tool calls, and context compaction. The problem has occurred more than once.

Steps to reproduce

The failure is intermittent, but the affected session had this sequence:

1. Start a long-running repository task in Codex Desktop.
2. Use subagents and repeated nested/local tool calls.
3. Continue through a large conversation and context compaction.
4. Observe several progress updates and final responses.
5. Eventually an assistant message contains unrelated multilingual fragments, tool/protocol data, or AGCONNECT(delta) text.

Expected behavior

Only the intended assistant commentary or final response should be rendered. Internal event payloads, connection state, tool schemas, and stream markers must remain separate from user-visible assistant text. Text should be rendered exactly as generated without fragments from other event streams.

Additional information

This may share an output assembly/event reconstruction path with:

• #34114, random text appended to scheduled heartbeat messages;
• #15001, subagent TUI live-message corruption.

However, this report covers the Windows desktop app, ordinary interactive replies rather than scheduled heartbeats, and direct leakage of apparent internal protocol/state data.

Related reproduction comments were already added to #34114 and #15001. No credentials, private paths, or user project content are included in this report.