[macOS App] Commentary progress messages render twice while final answers render once

What version of the Codex App are you using (From “About Codex” dialog)?

26.901.20858 (bundled app-server 0.153.0-alpha.5)

What subscription do you have?

API key / custom Responses-compatible provider (not signed in with ChatGPT)

What platform is your computer?

Darwin 25.6.0 arm64 arm

What issue are you seeing?

Every user-visible assistant progress update with phase: "commentary" is rendered twice in the Codex desktop transcript. The completed response with phase: "final_answer" is rendered once.

This reproduces in a fresh local task after fully quitting and relaunching the app. It is not a duplicate user submission and does not appear to be duplicate model output: thread/read reconstructs one commentary item and one final-answer item for the affected turn.

The configuration uses an API key with a custom Responses-compatible provider and gpt-5.6-sol.

What steps can reproduce the bug?

1. Configure Codex desktop to use an API key and a custom provider with wire_api = "responses".
2. Fully quit and relaunch the desktop app.
3. Create a new local task.
4. Send a prompt that requires tool use, so the assistant emits a progress/commentary message before using the tool.
5. Observe that the commentary message appears twice with identical text.
6. Let the turn finish and observe that the final answer appears only once.

This reproduced in two separate tasks, including a brand-new task after restart.

What is the expected behavior?

Each logical agentMessage should render once. A matching item/completed event should finalize or update the existing item by ID, not create a second visible commentary block. The behavior should be consistent for commentary and final_answer phases.

Additional information

Evidence from an affected turn:

• thread/read returns exactly one agentMessage with phase: "commentary" and one with phase: "final_answer".
• The raw JSONL has the expected lifecycle/persistence pair for each logical message: one event_msg.item_completed followed by one response_item.message, sharing the exact same item ID, phase, and text.
• The UI deduplicates this pair for final_answer but displays both copies for commentary.
• codex doctor reports the desktop installation, app-server handshake, configuration, and rollout DB parity as healthy. There are no duplicate rollout paths or thread IDs.
• No API keys, provider credentials, local usernames, or private prompt data are included here.

The App Server documentation says item/completed is the authoritative final state for the item:
https://learn.chatgpt.com/docs/app-server#items

Start by tracing desktop transcript handling for the item/completed and response_item.message lifecycle, using thread/read and the raw JSONL evidence in the report. Compare how matching item IDs are deduplicated for final_answer versus commentary, then reproduce with a custom Responses-compatible provider; done means each logical agentMessage renders once.