Permission-scoped Codex–ChatGPT messaging bridge with live transaction view

Summary

Please add an opt-in, permission-scoped backend messaging bridge between Codex and user-selected ChatGPT conversations or Projects.

This would provide a stable, auditable channel for routine messages and artifacts between two OpenAI products without relying on browser automation or merging their distinct permission models.

Product area: Codex App, ChatGPT conversations, and ChatGPT Projects
Scope: Version-independent product request

Problem

Today, communication between Codex and ChatGPT commonly requires manual copy/paste or browser-mediated automation. That introduces avoidable failure modes:

• expired or disconnected browser sessions;
• UI and rendering changes;
• incomplete or uncertain delivery;
• accidental duplicate submissions;
• loss of message identity and provenance;
• limited visibility into what was sent, where it went, and which permission allowed it.

Browser automation is useful for general web interaction, but it should not be the primary transport between two OpenAI products when a narrower backend channel could provide stronger authorization, auditability, reliability, and revocation.

Proposed design

1. Exact allowlists and separate permissions

Users should explicitly authorize exact ChatGPT conversation IDs or Projects. Permissions should be granted separately for:

• reading messages;
• sending messages;
• reading attachments;
• attaching Codex-generated artifacts.

Authorization for one conversation or Project must not imply access to any other ChatGPT history.

2. Live, user-visible transaction view

Users should be able to watch each exchange between Codex and ChatGPT in real time. The view should show:

• the actual message or artifact being exchanged;
• source and destination;
• direction of travel;
• permission used;
• timestamp;
• originating Codex task;
• message ID;
• queued, sent, delivered, failed, blocked, or cancelled status.

The user should be able to pause, cancel, or revoke the bridge while an exchange is underway.

3. Reliability and auditability

The transport should provide:

• stable message IDs;
• explicit delivery receipts;
• idempotency keys and duplicate protection;
• retry behavior that cannot silently duplicate a message;
• durable audit logs;
• clear failure reasons and recovery state.

4. Sensitive-action confirmation and revocation

The product should require confirmation before:

• transmitting sensitive files;
• expanding access to another conversation or Project;
• using a permission not already granted;
• performing an action outside the original authorization.

Users should have per-conversation revocation, per-Project revocation, and a global kill switch.

5. Strict authority separation

Messaging authorization must grant zero implied authority over:

• local files or folders;
• repositories or branches;
• terminal commands;
• computer-control permissions;
• network access;
• secrets or credentials;
• deployment or external-service actions.

Likewise, Codex filesystem, repository, terminal, computer, or network permissions must not imply permission to read or write ChatGPT conversations.

6. Privacy boundary

The live view should expose the communication transaction and its status. It should not expose private model reasoning, hidden chain-of-thought, credentials, authentication material, or security internals.

Example flow

1. The user authorizes one Codex task to exchange messages with one selected ChatGPT conversation.
2. The user grants send-message and read-response permissions, but no attachment access.
3. Codex prepares a message; the live transaction view shows its content, destination, permission, and queued status.
4. The message receives a stable ID and changes to delivered only after an explicit receipt.
5. A response from ChatGPT appears in the same transaction view with its own ID and provenance.
6. The user can pause the channel, revoke access, or use the global kill switch at any time.

Suggested acceptance criteria

• Users can allowlist exact ChatGPT conversations or Projects.
• Read, write, and attachment capabilities are independently controllable.
• Every exchange has a stable ID, status, provenance, and receipt.
• Duplicate submission is prevented across retries and reconnects.
• Users can see exchanges live and pause or cancel them.
• Sensitive or scope-expanding actions require confirmation.
• Revocation takes effect immediately.
• ChatGPT messaging authority and Codex local authority remain strictly independent.

Related requests

• #40142 proposes a one-click ChatGPT-to-Codex handoff with bidirectional task/results flow. This request is complementary: it focuses on the reusable permission-scoped transport, live transaction visibility, delivery semantics, revocation, and authority isolation.
• #30636 proposes an official ChatGPT transcript/Project import connector. This request differs by covering ongoing, bidirectional messaging rather than one-time import.