Summary
In the unified ChatGPT desktop app on macOS, a Codex Side Chat interaction unexpectedly appeared later as a normal ChatGPT conversation on chatgpt.com, even though I never promoted the Side Chat or explicitly created a regular Chat.
The Side Chat itself behaved normally and returned its response. The unexpected web conversation was only discovered the next day.
This is concerning because the new web Chat contained context originating from a Codex project / Side Chat.
Environment
- Platform: macOS
- Surface: unified ChatGPT desktop app → Codex → Side Chat
- Date reproduced: August 11, 2026 (discovered on web August 12, 2026)
- App version: 26.803.61601 (
Powered by Codex & OWL)
- App release date: August 9, 2026
- Same ChatGPT account on desktop and web
Steps to reproduce
- Open a Codex project/task in the macOS desktop app.
- Open/use Side Chat.
- Select several lines from an assistant response.
- Choose Add to chat, which inserts a structured annotation into the Side Chat composer.
- Ask a follow-up question about the selected text.
- Side Chat answers normally.
- Do not use any explicit “Promote to chat”, “Continue in Chat”, fork, or regular Chat creation action.
- Later, open
chatgpt.com on the same account.
- Observe a new ordinary ChatGPT conversation in web history that was never explicitly created.
- Open it and observe that it contains the Side Chat question/context from the Codex project.
In the observed incident, the unexpectedly created web conversation began with the prompt Tell me more about this.
Expected behavior
- Side Chat remains ephemeral/separate from normal ChatGPT history unless explicitly promoted.
- Add to chat only adds an annotation to the current composer; it should not create or persist a separate ChatGPT conversation.
- Codex Side Chat/project context should not silently appear in ordinary ChatGPT web history.
Actual behavior
A normal ChatGPT web-history conversation was silently created/persisted from a Side Chat/annotation interaction, while the Side Chat also completed normally.
Why this appears to be a bug
Related tracker behavior indicates:
I searched the public tracker for combinations of Side Chat / annotation / Add to chat / web history / ChatGPT web persistence and did not find an exact duplicate.
Impact / privacy
This is more than a history-list cosmetic issue. Codex project context can contain private code, paths, technical details, or other project information. Silently creating a normal ChatGPT web-history conversation crosses a user-visible product/data boundary without an explicit action or warning.
I am intentionally not including private project names, paths, source code, or raw transcript content in this public report. Sanitized logs/thread IDs can be provided privately if maintainers need them.
Summary
In the unified ChatGPT desktop app on macOS, a Codex Side Chat interaction unexpectedly appeared later as a normal ChatGPT conversation on
chatgpt.com, even though I never promoted the Side Chat or explicitly created a regular Chat.The Side Chat itself behaved normally and returned its response. The unexpected web conversation was only discovered the next day.
This is concerning because the new web Chat contained context originating from a Codex project / Side Chat.
Environment
Powered by Codex & OWL)Steps to reproduce
chatgpt.comon the same account.In the observed incident, the unexpectedly created web conversation began with the prompt
Tell me more about this.Expected behavior
Actual behavior
A normal ChatGPT web-history conversation was silently created/persisted from a Side Chat/annotation interaction, while the Side Chat also completed normally.
Why this appears to be a bug
Related tracker behavior indicates:
I searched the public tracker for combinations of Side Chat / annotation / Add to chat / web history / ChatGPT web persistence and did not find an exact duplicate.
Impact / privacy
This is more than a history-list cosmetic issue. Codex project context can contain private code, paths, technical details, or other project information. Silently creating a normal ChatGPT web-history conversation crosses a user-visible product/data boundary without an explicit action or warning.
I am intentionally not including private project names, paths, source code, or raw transcript content in this public report. Sanitized logs/thread IDs can be provided privately if maintainers need them.