[Linux desktop] Chat mode Personalization shows Codex global instructions instead of ChatGPT custom instructions

What variant of Codex are you using?

ChatGPT desktop app for Linux (preview), using Chat mode rather than Codex mode.

What issue are you seeing?

In the new ChatGPT desktop app for Linux, Settings → Personalization → Custom instructions shows my Codex-oriented global agent instructions (# GLOBAL AGENT — LEAN RULES) even while I am using ChatGPT / Chat mode.

However, the ChatGPT web app shows a different set of ChatGPT Custom Instructions under Settings → Personalization → Custom instructions. Those web instructions are intentionally general-purpose (accuracy, associations/examples, comparative tables), while my Codex instructions are intentionally engineering-specific (repository discovery, AGENTS.md, minimal diffs, validation, risk classification, etc.).

I use ChatGPT Chat and Codex as distinct environments and do not want the Codex global instruction set applied to ordinary ChatGPT conversations.

The Linux desktop UI currently makes it unclear whether:

1. the Codex instructions are actually being injected into Chat-mode conversations;
2. the field is only displaying the wrong instruction source;
3. desktop instructions are host-local and layered on top of account-level ChatGPT Custom Instructions; or
4. Chat and Codex currently share this Personalization field by design.

The wording under the field says:

Give ChatGPT extra instructions and context for all chats on this host.

That makes the scope especially ambiguous, because it appears to describe a host-level ChatGPT setting while displaying content that belongs to my Codex configuration.

Steps to reproduce

1. Sign in to ChatGPT on the web.
2. Open Settings → Personalization → Custom instructions and configure ChatGPT-specific instructions.
3. Have a separate Codex global instruction set configured for coding/agent work.
4. Open the new ChatGPT desktop app on Linux using the same account.
5. Select ChatGPT, then Chat mode (not Codex).
6. Open Settings → Personalization.
7. Observe that Custom instructions contains the Codex global agent instructions rather than the ChatGPT Custom Instructions shown on the web.

Expected behavior

ChatGPT Chat and Codex should have independently understandable instruction scopes.

Ideally:

• Chat mode should use/display the user's ChatGPT Custom Instructions (or clearly expose a separate host-local Chat instruction layer).
• Codex mode should retain its Codex/global-agent/project instruction hierarchy.
• The UI should make precedence and scope explicit if multiple instruction layers exist (for example: account-level ChatGPT instructions, host-level instructions, Codex global instructions, project AGENTS.md, and conversation context).
• Editing Chat personalization should not accidentally overwrite or alter Codex engineering instructions, and vice versa.

Why this matters

The two products/modes can legitimately require opposite steering. For example, general ChatGPT sessions may intentionally encourage creativity, interdisciplinary associations, analogies, experimentation, and speculative exploration, whereas Codex sessions may intentionally prioritize constrained scope, minimal diffs, repository-specific discovery, and targeted validation.

Sharing or ambiguously exposing the same instruction field makes it risky to edit either configuration and makes it difficult to know which instructions are actually governing a Chat session.

Additional information

The Linux desktop app is very new, so this may simply be an early preview UI/configuration issue.

I searched the current openai/codex issues for related Chat/Personalization/Custom Instructions reports and did not find an obvious exact duplicate.

Screenshots are available showing:

• ChatGPT web: general-purpose Custom Instructions.
• Linux desktop app while in ChatGPT/Chat mode: Codex GLOBAL AGENT — LEAN RULES in the Personalization field.

No source file, test, or code entry point is identified. Start by reproducing the behavior through Settings → Personalization in ChatGPT desktop Linux preview, comparing Chat and Codex modes with the web app; done means the instruction source and scope are clearly separated or explicitly explained, without unintended changes between modes.