[app][macOS] Existing text tasks cannot switch to live Voice without losing task context

No files or tests are named. Start by tracing the existing text-task flow and the entry point for starting live Voice from a new task, then compare how each session is attached to task history. Done means an existing text task can start and end live Voice without losing its thread history, context, objective, or state.

Summary

A Codex task that was started in text mode cannot switch into live Voice. The user must create a new empty task and choose Start new voice chat before sending any text. In an existing text task, only dictation is available.

For long-running architecture and requirements discussions, this effectively forces the user to abandon or manually transfer substantial task context just to continue the same conversation by voice.

Environment

• Platform: macOS
• Codex desktop app: 26.721.41059 (build 5848)
• Observed: 2026-07-27

Steps to reproduce

1. Open Codex desktop on macOS.
2. Start a normal text task.
3. Exchange enough messages for the task to accumulate meaningful context.
4. Try to start a live Voice conversation inside that same task.

Actual behavior

There is no way to start live Voice in the existing task. The available microphone interaction is dictation only. Starting live Voice requires a new empty task, so the active task's conversational context is not preserved in the Voice session.

Expected behavior

An existing text task should expose Start voice chat and transition into live Voice while preserving the same task/thread history, context, objective, and state. Ending Voice should return to the same text task without creating a disconnected conversation.

Why this matters

The need for Voice often becomes clear only after a discussion grows long or complex. Architecture interviews and requirement discovery may run for many turns before speaking becomes more efficient than typing. Requiring users to predict this before the first message makes Voice unavailable exactly when it becomes most useful.

This also creates an unnecessary distinction between:

• a task that started in Voice and can continue conversationally; and
• an otherwise identical task that started with one text message and can never enter live Voice.

Suggested direction

Allow an existing Codex task to attach a live Voice session to its current thread, preserving all existing context and task state.

Related broader proposal: openai/codex#33966 (duplex voice collaboration with an active Codex session). This report focuses on the narrower, currently reproducible UX gap of not being able to switch an already-started text task into Voice.