[Bug]: Codex subagent completion output can interleave with main-agent output

Summary

Codex subagent completion output can interleave mid-token with main-agent assistant output in the visible chat UI.

The persisted messages/events appear to remain separate, but the rendered chat text can splice the two outputs together.

Example

Expected separate text fragments:

Include constraints: read-only, use a specific tool, don’t modify files.

I’m seeing a clear split: short diagnostic nudges during debugging...

Observed rendered text:

Include constraints: read-only, use a specific tool, don’tI modify’m seeing files a. clear split...

Reproduction Shape

1. Use T3 Code with Codex.
2. Start a main-agent turn that spawns a background/subagent.
3. Have the main agent emit commentary while the subagent is completing.
4. Observe the visible chat output when both messages arrive close together.

Expected Behavior

Main-agent messages and subagent completion messages should render as separate chat items, or otherwise remain textually isolated.

Actual Behavior

The UI can render a single garbled assistant output where text from the subagent completion and main-agent commentary are interleaved mid-token.

Evidence

In one observed case, local provider logs showed two separate item/completed events about 72 ms apart:

• one for main-agent commentary
• one for subagent final output

The local SQLite projection also contained separate assistant message rows. This suggests the issue is likely in stream/render/event assembly rather than model output.

Impact

This makes Codex subagent workflows hard to trust because final answers can appear corrupted even when the underlying stored events are distinct.

Related Issues

Possibly related to prior stream/rendering issues:

• #1147
• #349
• #2380
• #538

Reproduce the T3 Code scenario with a Codex main agent and background subagent, then trace the stream, render, and event-assembly path where the two item/completed events arrive close together. Compare the visible chat with provider logs and the local SQLite projection; done means main-agent commentary and subagent completion output remain separate chat items or textually isolated.