Feature request: Live Voice as a first-class Codex orchestration surface

The request

I want live voice in Codex to become a serious operating surface for real work—not a separate, lightweight way to talk to Codex.

Right now, voice is useful for conversation. The bigger opportunity is to let it run the same kind of focused work a good Codex task can run: keep context, follow a working style, use the skills I have approved, and coordinate work through my own setup.

This is not a request to expose every model knob to everyone. OpenAI can keep choosing the sensible default. What I am asking for is deliberate control when it matters.

Make orchestration the centre of live voice

The hero feature should be native orchestration.

When I start a live-voice session, I should be able to say: “Work as my operations lead for this task,” select a saved style, give it the right task context, and let it use the skills and workflows I have approved. It should be able to move between voice and text without becoming a different, forgetful assistant.

For me, that means voice can become the front door to Codex: I speak the goal, Codex keeps the thread, coordinates the work, asks only for meaningful decisions, and brings the result back. That would be a huge jump in usefulness.

What I would like to see

1. A real voice-session setup screen

   • Start fresh or deliberately resume a specific task.
   • Pick a saved working style / instruction profile for that session.
   • See what context, skills, tools, and connections the session will use.
   • Avoid accidentally reopening an old, oversized task just because it is the most recent one.

2. First-class orchestration

   • Voice sessions should use approved Codex skills, workflows, and task capabilities—not stop at chat.
   • The session should preserve task continuity across voice and text.
   • It should be able to coordinate bounded work and report back clearly, with the same permission model users expect from Codex.

3. Secure connection to the user's own harness

   • Let users connect a voice session to their personal context and local execution setup—for example, a knowledge system such as My Brain, or a local harness.
   • Make every connection explicit and permissioned. I want visible control, not hidden access.

4. Model and voice controls where supported

   • Keep “Automatic” as the default.
   • Where OpenAI allows it, offer an advanced model choice in the voice-session picker.
   • Let users select the voice and set a practical voice working style without having to repeat it every session.

Why this matters

Live voice is the most natural way to start work when I am away from the desk or on my phone. If it can launch a properly configured Codex task—with my context, my working style, and my approved capabilities—it stops being a side feature and becomes the fastest way into serious work.

I would be happy to test this and provide direct product feedback.

— Hammad

The issue does not name files, tests, or entry points. Start by locating any existing live-voice, session, and orchestration surfaces, then narrow the broad request to one explicitly scoped behavior with a clear permission and continuity path. Done would require an agreed implementation scope and tests for that selected behavior.