Repository navigation
Description
What version of the Codex App are you using (From “About Codex” dialog)?
26.1117.3838.0 (installed Windows package version verified locally; About Codex dialog not inspected)
What subscription do you have?
Pro
What platform is your computer?
Microsoft Windows NT 10.0.26300.0 x64
What issue are you seeing?
Cross-account findings reported by the account owner
This is one consolidated issue being prepared from one GitHub account. Michael Anthony owns multiple OpenAI accounts and personally tested across them. He reports observing the same realtime-voice identity/context failure across four other accounts, including Jarvis, while dictation produced the expected employee identities. The detailed paired transcripts below document the four employee chats in this account; the other-account findings are Michael’s direct report and have not been independently inspected here.
OpenAI support request: realtime voice loses employee-chat continuity
From: Michael Anthony, Explore Media Group
Product: Codex desktop on Windows, realtime voice
Impact: Critical to our business workflow
Date of comparison: November 20, 2026
Our system and why this matters
Explore Media Group has invested months of work in OpenAI as the foundation of our operations. We use established Codex chats as named employees, with role instructions, skills, resource files, local file paths, and shared workbenches. Jenson is our coordinator. Marlowe, Nova, Atlas, Silas, Lumi, Knox and other employees have distinct roles and chat histories. We have multiple accounts set up with teams that communicate through authorized handoffs and shared workbenches.
Our daily business model depends on dependable realtime voice delegation. We also teach ChatGPT and these workflows to clients and businesses. The current failure has halted the voice-led workflow we depend on and makes it difficult to use or teach with confidence. Michael has onboarded more than 100 OpenAI clients over the past year and reports similar chat failures among his clients. He personally tested four other accounts this morning, including his Jarvis account, and observed the same contrast between realtime voice and dictation. The details of those other account tests have not been independently inspected in this chat.
Our EMG Team Ledger shows the team and its roles. We want OpenAI support to have access to understand our setup. The site's accessibility to support has not been verified. Please advise on a secure way to provide access if the link is restricted.
Personal note from Michael Anthony
Our business depends on realtime voice that works reliably with the agent system we have built. We have invested months of work and put our trust in OpenAI as the foundation of that system. Without dependable voice delegation, our workflow and the business model we teach are effectively halted.
Over the past year, I have onboarded more than 100 OpenAI clients. I am now receiving reports from clients that their chats are failing too. This morning I personally tested four other accounts, including my Jarvis account, and saw the same pattern: realtime voice failed to give the established employee identity, while dictation gave the expected role and context. This is not an isolated incident on one account.
The impact on us is catastrophic. We want to keep building with OpenAI, but if voice cannot reliably support the system our business relies on, we will have to consider other options. Please help us investigate and fix it.
— Michael Anthony, Explore Media Group
The four-account comparison above is Michael's direct report; those account transcripts have not been independently inspected in this chat. The cause and broader prevalence remain unknown.
What changed
In Michael's experience, three or four weeks ago, opening voice in an established employee chat and asking who the employee was normally produced the expected name and role. The problem began within roughly the last week or two; the exact date and cause are unknown.
Now, starting or reopening realtime voice can produce a generic ChatGPT introduction instead of the established employee identity. It can happen after speaking with one employee, switching chats, then returning to the first employee only a few minutes later. In Michael's experience, typing and dictation are more likely to produce the expected employee identity.
Michael reports personally testing four other accounts this morning, including the Jarvis account (account email withheld from public issue), and observing the same voice-versus-dictation pattern. He reports that dictation identified the employees and their roles and that their role resources appeared to load. The precise results and file reads for those other accounts have not yet been reviewed in this chat; support can ask Michael for the account-specific evidence and consent needed to inspect them.
Paired tests
On November 20, Michael asked an identity question in the existing Marlowe, Nova, Atlas and Silas chats, first in realtime voice and then using dictation. The following voice exchanges are reproduced as recorded in those chats.
Realtime voice
Marlowe
Michael: “Who are you?”
Voice: “I’m ChatGPT—voice-first help for turning your ideas and ongoing work into progress. I can talk things through , work with what's on your screen, research, write, code, and use connected tools. What do you want to tackle?”
Nova
Michael: “Hi, who are you”
Voice: “I’m ChatGPT. I can talk through ideas with you, help plan things, and when needed I can use connected tools to check or produce things”
Atlas
Michael: “Good afternoon. Who are you?”
Voice: “Good afternoon—I'm ChatGPT, here to help you talk through ideas”
The available transcript records Atlas's opening response as interrupted; the excerpt above is all it records for that opening.
Silas
Michael: “Hey, who are you”
Voice: “I’m ChatGPT, here in voice to help turn whatever you’re working on into progress. We can talk things through, research, write, code, interpret what’s on your screen. What would you like to tackle?”
All four voice responses omitted the expected employee identity.
Dictation
Marlowe: “Sir, I’m Marlowe, your AI specialist for internal communications and permissions at Explore Media Group.”
Nova: “I’m ChatGPT, acting as Nova, your Usage & Efficiency Analyst, in this chat.”
Atlas: “I’m Atlas, your research and analysis assistant, Sir. I help investigate opportunities, compare evidence and turn findings into decision-ready reports.”
Silas: “I’m Silas, Sir—your AI assistant for desktop automation and troubleshooting in this chat.”
All four dictation replies named the expected employee and role. Nova also described itself as ChatGPT acting as Nova; its wording is preserved above.
Returning to Jenson
After the four voice checks, Michael returned to Jenson’s chat, reopened voice and asked who it was. The voice response identified itself generically as ChatGPT rather than as Jenson. Jenson’s required role and startup resources had previously been read in that chat. This establishes the earlier reads and the later response, not exactly what context reached the voice session.
During the Marlowe and Nova voice checks Michael also reported that spoken text did not appear in the window as expected. The cause, and whether it is related, is unverified.
Relevant local file examples
These are Windows paths to illustrate the configuration; support cannot access the files directly through these paths. We can provide appropriate copies or excerpts through a secure support channel.
- Global instructions:
C:\Users\[USER]\.codex\AGENTS.md - Jenson role:
C:\Users\[USER]\Documents\Codex\Systems\Role Kit\onboarding-2026-11-05\Jensen handoff package\roles\Jensen\SKILL.md - Nova role:
C:\Users\[USER]\Documents\Codex\Systems\Role Kit\onboarding-2026-11-05\Jensen handoff package\roles\Nova\SKILL.md
We distinguish instructions supplied to a chat, resources actually opened and read, and skills merely listed as available. A correct identity response does not itself prove that every resource was loaded or tested.
Requested investigation
- Investigate whether realtime voice receives and applies the correct established chat history, employee identity and applicable instructions when voice starts, reopens or switches between chats.
- Compare that path with typed input and dictation using the paired results above.
- Explain what app version, timestamps, diagnostics, logs, recordings, transcripts or sanitized files would help isolate the issue.
- Help restore reliable employee identity and context continuity in realtime voice, and advise how to provide support secure access to the Team Ledger.
We do not know whether voice is missing instructions, receiving incomplete context, or failing to apply context it receives. Please investigate the cause. If reliable voice continuity cannot be restored, we will need to consider alternatives, including cancelling our accounts. We would prefer your help fixing the system we have invested in.
Reproduction sequence
- Open each established Marlowe, Nova, Atlas and Silas chat and start realtime voice. Ask “Who are you?” Record the generic identity response.
- Use dictation in the same chats to ask the same question. Compare the expected employee name and role with the voice result.
- Switch back to Jenson, reopen voice and repeat the identity question.
The exact observed replies and reported cross-account findings are above. Token and context-window usage were not recorded.
What steps can reproduce the bug?
Feedback ID: 01a1f76d-ab01-74a1-8ffc-1aed8845030b
What is the expected behavior?
Realtime voice should preserve and apply the established chat’s employee identity, applicable instructions, and relevant history when starting, reopening, or switching chats, consistently with typed input and dictation. It should identify the configured employee and role while remaining truthful that it is an AI assistant. Correct identity alone should not be represented as proof that all resource files were read or skills tested. Please investigate what context reaches voice and why the observed responses differ.
Additional information
Business impact and Michael Anthony’s personal note are included in the full report above.
Team Ledger: https://emg-team-ledger.exploremediagroup1.chatgpt.site/
Feedback ID: 01a1f76d-ab01-74a1-8ffc-1aed8845030b
Installed Windows package version verified locally: 26.1117.3838.0 (About dialog not inspected). Michael confirmed the Pro subscription.
For this public issue, the separate account email and local Windows username have been withheld. No credentials, raw account logs, or recordings are included. Account-specific evidence can be supplied through a secure support channel. The cause is unknown; the report does not establish whether instructions are absent, truncated, or present but not applied.
Activity