# ChatGPT / Codex web and desktop comparison

Version 1.0.1 · October 3, 2026 · SOURCE_REVIEWED

This record applies the ecosystem analysis to a concrete comparison. It separates published contracts from the account-level trials required to establish a defect. It does not claim a matched-session experiment.

## Evidence classes

- **Documented:** official source retrieved during the October 3 investigation.
- **Observed:** controls visible in the supplied screenshot.
- **Reported:** the user's account of settings, style and chat-saving differences.
- **Inference:** a possible explanation requiring an additional trace.
- **Proposed:** an expected behavior or repair contract.
- Evidence classes describe provenance, not validation state. Canonical evidence states are **SOURCE_REVIEWED** for the documentation/screenshot review, **HYPOTHESIS** for causal explanations, **SPECIFIED** for repairs, and **AWAITING_LOCAL_RUN** for unexecuted runtime validation. **NOT_RUN** is an execution status only.

## Comparison record

| Question | ChatGPT Web / Work | Codex Web / Cloud | Codex Desktop developer experience | Evidence and limit |
|---|---|---|---|---|
| What is the visible product? | Chat and Work are user-facing modes. | User identifies the attached page as Codex Web; full route absent. | Codex is part of the ChatGPT desktop app, retaining a developer experience. | Screenshot observation; July 9 changelog. Exact screenshot route NOT_RUN. |
| What agent configuration applies? | Web Work does not load local Codex config. | Published cloud setup defines task environment; full personality/global-guidance inheritance remains unresolved. | Local Codex agents share CLI/IDE configuration layers. | Developer settings and cloud environments. Matched effective profiles NOT_RUN. |
| Which style inputs apply? | Web Work has its own personalization controls. | Do not infer exact inheritance from the word Codex or Cloud. | Local personality and personal/project instructions can contribute. | Personalization; user-reported difference. Causal trace NOT_RUN. |
| Which memory is used? | ChatGPT memory follows its product controls. | No claim that all local memories are copied to cloud tasks. | Local Codex memory is separately controlled and stored. | Memories. Exact session retrieval NOT_RUN. |
| Where is the same chat found? | Work history has its own continuity and access rules. | Cloud tasks retain independent saved state. | Host-associated Codex chats; desktop/IDE can share eligible project chats. | Local access, cloud environments, developer settings. Same-ID reopening NOT_RUN. |
| Does a shared link continue it? | Depends on the underlying object and permissions. | A task link is not an automatic executor transfer. | A shared local-thread link is a read-only snapshot. | Feature digest. Actual link trial NOT_RUN. |
| Does Work Sync combine histories? | Eligible Work continuity can include local access. | Codex configuration/history behavior remains separate. | Work and Codex are not combined by Work Sync. | Local computer access and Work overview. Account eligibility NOT_RUN. |
| Can the target change while running? | Work cannot switch local execution to cloud mid-turn. | Codex Cloud is not a remote-handoff destination. | Supported host handoff transfers chat/Git state and interrupts an active response. | Local access and remote connections. No unsupported transfer attempted. |
| Are all cloud resources equal? | Work has its own cloud capabilities and policy. | Codex Cloud has published setup and capability limits. | Local/remote host has its own files, browser, config and credentials. | Cloud environments, remote connections, Work overview. Source-review comparison only. |
| What do Dots add? | A persistent agent's responsibility can outlive a task. | Dot coding delegation is a separate task/conversation with selected context. | A dot can use permitted computer resources; choosing another computer does not relocate an existing task. | Dots and task/memory docs. End-to-end journey NOT_RUN. |
| Is a returned artifact the same as shared context? | Page and source permissions differ. | Delegation context is selected rather than all parent history. | Private chats/memory are not disclosed by Page sharing. | Space and Dot tasks. Permission fixture NOT_RUN. |

## Three priority findings

### C1 — Continuation is conditional, but discovery should explain the condition

**Established:** separate history, host-access and snapshot contracts exist. **Reported:** difficulty finding the same work across clients. **Inference:** the user cannot see a sufficient common continuation contract. **Proposed:** a permission-filtered Activity index with stable identity, location, owner and an explicit Continue action. **Next check:** protocol T1.

### C2 — Different inputs can explain different behavior; the particular cause remains unknown

**Established:** local configuration and memory differ from web Work inputs. **Reported:** styles and settings change across surfaces. **Inference:** effective-source visibility is inconsistent. **Proposed:** common field-level profile provenance, extending local /status and /debug-config. **Next check:** protocol T2. Identical wording is not a valid acceptance test.

### C3 — Execution and recovery need a visible resource contract

**Established:** Work, Codex Cloud and connected hosts have different capabilities and transfer rules. **Inference:** similar labels can cause unsupported resource expectations. **Proposed:** execution manifest, preflight and receipt-based recovery. **Next check:** protocol T3. No unsafe-retry defect is asserted.

## Actual investigation status

| Check | Evidence state | Execution status | Evidence / limit |
|---|---|---|---|
| Official product-contract review | SOURCE_REVIEWED | REVIEWED | Official pages identified in the source-capture manifest; no private backend inference represented as fact. |
| Supplied web screenshot | SOURCE_REVIEWED | REVIEWED | Cloud, environment, approval and model controls visible; full route/product heading absent. |
| User experience explanation | HYPOTHESIS | NOT_RUN | Reported settings, style and chat-saving differences have not been independently reproduced or causally traced. |
| Direct desktop UI inspection | AWAITING_LOCAL_RUN | NOT_RUN | Computer Use rejected access to the application for safety reasons; no workaround attempted. |
| Matched web/desktop session comparison | AWAITING_LOCAL_RUN | NOT_RUN | Desktop UI unavailable; exact routes and profile traces not collected. |
| Host-disconnection/recovery trial | AWAITING_LOCAL_RUN | NOT_RUN | No isolated supported fixture executed; no real network or permission changes. |
| Proposed OpenAI architecture | SPECIFIED | NOT_RUN | Architecture document exists; implementation and release metrics remain untested. |

## Primary sources

1. [Developer settings](https://learn.chatgpt.com/docs/developer-settings)
2. [Personalization](https://learn.chatgpt.com/docs/personalize)
3. [Memories](https://learn.chatgpt.com/docs/customization/memories)
4. [Codex Cloud environments](https://learn.chatgpt.com/docs/environments/cloud-environments)
5. [Remote connections](https://learn.chatgpt.com/docs/remote-connections)
6. [Local computer access](https://learn.chatgpt.com/docs/enterprise/cloud-local-access)
7. [ChatGPT Work overview](https://learn.chatgpt.com/docs/enterprise/chatgpt-work-overview)
8. [Meet Dots](https://learn.chatgpt.com/docs/dots)
9. [Dot tasks and memory](https://learn.chatgpt.com/docs/dots/tasks-and-memory)
10. [Space](https://learn.chatgpt.com/docs/space)
11. [Changelog](https://learn.chatgpt.com/docs/changelog)
12. [Feature digest](https://learn.chatgpt.com/docs/whats-new)

Keep this record versioned. Promote a reported issue to observed only after a dated trial; preserve contrary results, account limitations and older source contracts.

Revision 1.0.1: provenance classes and execution statuses are separate from canonical evidence states. Exact source-capture hashes are listed in [the source manifest](openai-source-captures.json).