Choose the right place to start

If the work is already in a Codex task, start Voice there. “What did the last test run show?” then refers to that task’s conversation. A new voice chat is the better starting point when you want a separate task record for a new investigation. After asking Voice to create the task, open it and confirm that it points to the intended project before authorizing any command.

Use voice dictation when you only need to speak one prompt and inspect the transcribed text before sending it. A live Voice chat makes more sense when you plan to ask follow-up questions, check progress, or switch between tasks. The desktop app permits one active Voice chat at a time, although that conversation can coordinate multiple Codex tasks. For frequent use, set a shortcut in Settings > Voice > Voice chat hotkey.

Decision guide: start Voice in an existing Codex task, start a new Voice chat for a separate task, or use dictation for one prompt
Choose an entry point according to whether you need the existing task's context, a separate task record, or a single spoken prompt. Conceptual guide, not a product screenshot. Credit: Grow AI Skills original editorial diagram. Open full-size image.

Run a test task you can verify

Suppose your repository has several packages and you want tests for one of them. Name the project exactly as it appears in the app. If you do not know its test command, ask for the command first:

Start a separate Codex task in [exact project name]. For [package or component], find the documented test command in the README or project scripts. Tell me the command and where you found it. Do not run it or edit files yet.

Open the new task and check its project. Compare the proposed command with the README or scripts before running it. If there is no documented command, ask what test targets are available and choose one; do not present a guessed command as documented. If you already know the correct command, include it in the first request and skip discovery.

Once the target and command are right, direct the follow-up to that task:

In the [task title] task, run [confirmed test command]. Report the command, exit status, and failing test names. Stop before editing files.

Check the Codex task after the run. A Voice reply tells you the request was received. A task in the list tells you one was created. Look for the command and its output in that task before saying the tests ran. If the command could not start because the environment was missing a dependency, that is an execution blocker, not a failing test assertion. Ask what blocked the run and resolve it before drawing a conclusion from the suite. If Codex later changes code, review the diff and a test run after the change before calling the fix verified.

Three checks after a Voice request: acknowledgement, Codex task creation, and recorded command with exit status and output
A spoken reply, a new task, and a completed command are different observations. Confirm the command and output in the task record before reporting a test result. Conceptual checklist. Credit: Grow AI Skills original editorial diagram. Open full-size image.

Check progress and switch tasks

While the test task is running, ask, “Check active tasks and tell me what is blocking [task title].” Naming the task matters if Voice is coordinating several jobs; “that task” may not identify the one you mean. Voice can relay progress and send a follow-up while the underlying task continues.

To move the live conversation to another Codex task, say, “Let me talk to the task reviewing the tests.” You can then say, “Take me back to the previous task.” The destination must support Voice and be on a connected host. If Voice cannot switch, open the destination task directly and check its status. A failed switch does not mean the task stopped. Each task retains its own work record, so inspect the output in the task that performed the work.

Conceptual diagram of one ChatGPT Voice conversation coordinating two separate Codex tasks
A Voice conversation can direct more than one task; inspect the work in the relevant task. Original explanatory diagram, not a product screenshot. Credit: Grow AI Skills original editorial diagram. Open full-size image.

Check permissions before speaking an action

Voice follows the permission mode of the task it directs. Under Ask for approval, Codex may read and edit files in the current workspace and run routine local commands; it asks before internet access or crossing the workspace boundary. Saying “Do not edit files yet” does not change that permission mode. If the original checkout must stay untouched, work in a disposable copy with workspace-scoped permissions. Check the permission control below the composer before starting; Full access allows broader file and network actions without approval.

On macOS, Screen context is a separate choice. Enable it in Settings > Voice, bring the relevant app window to the front, and say, “Take a look at this” when the issue is in that window. An appshot can include accessible text outside the visible scroll area. It is unnecessary when you only need to ask about a Codex task’s own output. Move away from windows with secrets or private records before invoking screen context; an organization may disable the feature.

If Voice is unavailable

If Start voice chat is absent from a Codex task that began with text, update both the desktop app and the Codex host running that task. Existing-task Voice is still rolling out. A new voice chat or voice dictation may be available, but starting a new chat does not by itself put you inside the earlier task’s conversation.

If Voice will not start at all, check eligibility first. OpenAI currently lists desktop Voice for Plus, Pro, Business, Edu, and Enterprise, subject to rollout and workspace settings. Then check microphone permission and whether another app window already has the one active Voice chat. If Voice works but cannot inspect a macOS window, check Settings > Voice, Appshots permissions, and organization restrictions.