Codex Live: Duplex voice collaboration with an active Codex session
What variant of Codex are you using?
Codex desktop app on Windows, including long-running local and remote sessions
What feature would you like to see?
Summary
Please allow ChatGPT Live Mode to attach directly to an active Codex session, so that the user can have a continuous duplex voice conversation with Codex while it keeps planning, coding, running commands, testing, and coordinating subagents.
This is not merely voice input or speech-to-text for Codex. It is continuous, shared-state collaboration between the human and a running coding agent.
Proposed experience
A user starts a long-running task in Codex and opens Live Mode attached to the same session. Codex continues working while the user speaks with it naturally.
The user could say:
- “Explain the authentication architecture while you continue working.”
- “Do not modify the existing database schema.”
- “Ask a subagent to investigate that deployment failure.”
- “What did the testing agent just report?”
- “Pause this part and reconsider the approach.”
- “Continue working, but ask me before any destructive operation.”
Codex could also surface important developments during the conversation:
- a blocker has been discovered;
- tests revealed a deeper problem;
- a subagent returned an important finding;
- a major assumption has changed;
- the implementation is deviating from the approved plan;
- an operation requires human approval.
Routine terminal activity should remain quiet unless the user asks for it.
Shared session state
Live Mode and Codex should share the same session state, including:
- the current objective and plan;
- files changed during the session;
- terminal and tool activity;
- test results;
- active subagents and their reports;
- blockers and pending approvals;
- relevant repository instructions, specifications, and memory files.
Instructions spoken through Live Mode should immediately steer the running Codex session without requiring the user to submit another disconnected prompt.
The ideal model is one agent with two simultaneous interfaces:
- An operational interface where Codex edits files, runs tools, tests, and delegates work.
- A conversational interface where the human observes, questions, interrupts, redirects, and collaborates in real time.
Why this matters
The current coding-agent workflow is mostly turn-based:
- Give Codex a task.
- Wait for it to work.
- Read the result.
- Discover a misunderstanding or remember another requirement.
- Submit another prompt and wait again.
A duplex session would allow the human to remain cognitively present while execution continues. Requirements could be clarified, bad assumptions interrupted, investigations delegated, and architectural decisions discussed before an incorrect approach spreads through the codebase.
This would feel less like submitting jobs to a remote process and more like sitting beside a senior engineer who is coding, investigating, delegating, and speaking with you simultaneously.
Controls and safety
The user should be able to:
- pause, resume, or cancel work through speech;
- interrupt Codex immediately;
- require approval for destructive or expensive actions;
- choose which events should interrupt the conversation;
- mute routine narration;
- request status summaries on demand;
- inspect active subagents and their findings;
- use an “exceptions only” reporting mode.
Examples of actions that could require approval include production deployments, destructive database changes, deletion of resources, authentication changes, operations involving credentials, and actions with financial cost.
Core request
Please integrate ChatGPT Live Mode with active Codex sessions so users can continuously converse with, observe, interrupt, question, and redirect Codex while it continues working.
The central feature is continuous duplex collaboration with one shared state—not merely adding voice input to the existing prompt box.
Additional information
Example scenario
Codex is implementing a feature while several subagents investigate testing, architecture, and deployment.
During the Live conversation, the testing agent reports that the proposed authentication change conflicts with the existing schema. Codex says:
“The authentication investigation found a schema conflict. I have paused that portion of the implementation. Would you like me to explain the conflict or investigate alternatives?”
The user discusses the issue with Codex, selects an alternative, and Codex resumes immediately. The coding session never needs to stop and restart as a series of disconnected prompt-response cycles.
This could also work when accessing a desktop Codex session remotely. A user away from the computer could speak with the active session, receive important reports, redirect subagents, approve decisions, and ask Codex to explain the codebase while work continues.
Possible names:
- Codex Live
- Live Codex Session
- Duplex Codex
- Continuous Collaboration Mode
- Voice Mission Control
5 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
I reviewed #33221 and #33249. They are related to parts of this proposal, but neither appears to be a duplicate.
#33221 focuses on adding voice mode for interview preparation and spoken walkthroughs after completing a coding task.
#33249 proposes allowing ChatGPT to monitor Codex, inspect intermediate outputs, and send correction instructions, with ChatGPT acting as the supervisor and Codex as the executor.
This issue proposes a different interaction model: the human has a continuous duplex Live conversation directly attached to the active Codex session while Codex continues coding, testing, using tools, and coordinating subagents. The same agent would expose both an operational surface and a conversational surface, allowing real-time spoken discussion, interruption, delegation, approval, replanning, and meaningful progress reports without stopping execution.
The central request is simultaneous conversation and execution through one shared session state, rather than voice-only interaction or supervision between two separate OpenAI agents.
Adding a native iPhone, call-like access path would make this proposal substantially more useful.
The need is not simply another place to enter prompts. Many users already have AI connected to Slack, Telegram, Discord, and other channels. What is still missing is a natural interface for moments when the user is away from a desk, moving, or unable to look at a screen: talk to Codex as naturally as calling a colleague, give it instructions, interrupt and correct it in full duplex, and receive progress reports or completed results from the same running task.
The broader product idea is: at any time and from anywhere, it may be a more natural next step to instruct an AI and receive reports the way we place a phone call, instead of opening an app and exchanging discrete prompts.
For that experience, the iPhone client should ideally support:
This would turn voice from a dictation feature into a genuine remote operating interface for Codex: instruct, clarify, approve, and receive reports from anywhere without needing to watch the screen.
Strong +1. This is exactly the Codex feature I want most—not just dictation, but continuous two-way voice collaboration within the same active task.
I work in product operations at an internet company in China and started with zero coding background. Thanks to Codex, I can now independently build full-stack productivity tools and have shipped several real products. I’m a huge fan of the product and deeply appreciate the team’s constant improvements to the experience and new capabilities.
For longer tasks, the biggest remaining friction is the stop–wait–read–type loop. I want to speak with Codex while it keeps working, hear concise updates or questions, and interrupt or redirect it immediately without losing the shared task context.
Voice, screen, and text should work together in one session, with explicit visual confirmation for risky actions. I believe this would make Codex dramatically more accessible to product people, operators, and new builders.
I’d be very happy to share my workflow or participate in user research.
— Boyuan Liu
This is also a concrete recovery-path blocker for long-running tasks. After accidentally closing a Voice Chat session, reopening the active task does not provide a way to resume voice inside that same task while it continues running. Starting voice again creates a separate conversation, so the user loses direct shared context with the active task.
Requested behavior: an active task should expose a Resume Voice action that reconnects voice to the same task and preserves its current state, running work, and task context. If that is intentionally unavailable, the UI should make the limitation explicit before ending a voice session.