Responses duplicate when Stream assistant messages is on

[vertopolkaLF]

[jmbenedetto]

Confirmed from raw Codex session logs on macOS.

In the local session JSONL, each assistant message is recorded twice in sequence:

• event_msg with payload.type: "agent_message"
• response_item with payload.type: "message", role: "assistant"

I checked four sessions for the same repo on 2026-03-07 and all show a strict 1:1 duplicate-pair pattern:

• session A: 17 duplicate pairs
• session B: 21 duplicate pairs
• session C: 4 duplicate pairs
• session D: 6 duplicate pairs

So this does not look like the model independently generating the same text twice. It looks like the UI/log consumer is rendering both record types when stream assistant messages is enabled.

Environment details from the session metadata:

• Codex CLI 0.111.0
• observed with session sources exec and vscode

If useful, I can provide a small repro parser against the JSONL format, but the key signal is that each duplicated visible message corresponds to one agent_message record immediately followed by one assistant response_item with identical text.

[ishaanrajiv]

I can reproduce as well, on v0.0.4

From what I can see, here's the root cause:

• live assistant_text deltas are emitted and rendered incrementally during streaming
• later, item.completed may arrive with payload.detail containing the full assistant message text
• the completion path treats that detail as fallback content even when text was already emitted from streaming

As a result, the full assistant response gets appended again on completion

Have raised a PR for a fix: #381, but not sure if they are accepting external contributions.

[fabmnt]

I noticed the same bug on Windows 11. In the Codex app, the same thread looks fine, but in the t3code the response is duplicated (I have streaming responses enabled).