[app][macOS] New Realtime Voice conversations do not receive text context sent after startup

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

26.814.41407 (build 6720)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.5.0 arm64 arm (macOS, Apple Silicon)

What issue are you seeing?

New Realtime Voice conversations start without the text context sent in the same task. The user can create a new Voice conversation and send a text instruction, but the initial voice interaction behaves as though that instruction is not present. Later, the text agent processes the instruction independently.

This leaves the text and voice parts of one task with different effective context. The normal UI does not allow an existing text task to switch into live Voice, so creating a new Voice task and then sending its instructions is the required workflow; that workflow currently cannot reliably establish context for Voice.

Local diagnostics indicate that Realtime starts before the provisional task is fully materialized. No task identifiers, transcripts, or private logs are included here.

What steps can reproduce the bug?

1. Open Codex Desktop on macOS.
2. Start a new Realtime Voice conversation.
3. Send a short text instruction that defines the first voice response or the current task context.
4. Begin the voice interaction.
5. Compare the first voice response with the text instruction and with the later text-agent response.

The issue has reproduced across multiple newly created Voice conversations.

What is the expected behavior?

Voice and text in the same task should share one conversation context. A text instruction sent in a newly created Voice task should be available to the first relevant voice response, and the text agent and Voice should follow the same instruction set.

If the product requires Voice to start before a task can receive text, the client should materialize the task and synchronize the text context into the Realtime session before generating a response that depends on it.

Additional information

This is related to, but distinct from, #35500, which covers the inability to switch an existing text task into live Voice. This report covers the remaining required flow: a new Voice task is created successfully, but text context sent after startup is not reliably applied to the voice interaction.