Codex Desktop: let users disable native structured-question UI and preserve an explicit transcript decision protocol

Summary

Codex Desktop currently lets a model/harness-native structured-question control override an explicit user-defined decision protocol in the chat. That is an unacceptable loss of user control.

A user may deliberately require a formal, durable, Markdown-based decision process in AGENTS.md and a skill. For example, the protocol can require:

• exactly one decision at a time;
• a fully stated question;
• stable A–Z options with clearly scoped meanings;
• an explicit prior-decision recap;
• a durable log/audit trail in the transcript.

The native picker may be convenient for users who want it, but it must not silently replace a user's explicitly requested process. In the reported occurrence, the picker also flashed and disappeared, leaving neither a usable native control nor the mandated transcript fallback. That makes the decision process impossible to run.

This is distinct from related issues

• #43057 and #43803 concern a question UI that disappears. This issue concerns the authority boundary: even a perfectly rendered native picker must not override an explicitly required transcript protocol.
• #43821 requests an app-server configuration opt-out for embedding hosts. This issue requests a user-controlled Codex Desktop product setting and a per-chat priority override, suitable for ordinary users and local policy files.

Required behavior

Provide both of the following, with the user's explicit instruction taking precedence over the native control:

1. Per-chat priority override. A documented, reliable way for a user to state that native structured-question UI is disabled for the current chat and that questions must be rendered as transcript Markdown. This must be honored by the model and harness.
2. Persistent desktop setting. A user-facing Codex Desktop preference to disable native structured-question UI globally. When enabled, the structured-question tool must not be exposed to the model/harness, or it must deterministically return an unsupported result that causes a transcript fallback.

The setting needs a clear scope and visible state. It must not rely on a prompt that the harness can ignore after exposing the native tool.

Acceptance criteria

1. A user can turn off native structured-question UI in Codex Desktop settings.
2. In that state, the model cannot silently invoke a UI-only selection control.
3. When a skill or AGENTS.md requires a Markdown protocol, the resulting question is present, durable, copyable, and answerable in the conversation transcript.
4. A per-chat explicit override works even if the global setting remains enabled.
5. The UI state survives app restart and is verifiable by the user.
6. If a client cannot honor the requested transcript-only mode, it reports that limitation clearly instead of claiming a question was successfully presented.

Why this matters

Structured decision processes are not merely prompts to be prettified. They can be operating procedures with audit, governance, accessibility, documentation, and retention requirements. A harness that unilaterally substitutes an ephemeral proprietary control for the user's chosen protocol is not assisting the workflow; it is taking control of it.

The correct relationship is straightforward: native UI is an optional enhancement, never an undisclosed replacement for an explicit user-defined formal process.

Environment and occurrence

• Surface: ChatGPT Desktop / Codex harness on macOS
• ChatGPT Desktop / Codex harness: 26.903.71938, build 8576
• Bundled Codex CLI: 0.153.4
• Operating system: macOS 15.7.9 (build 24G830)
• Hardware: Mac Studio with Apple M2 Ultra
• Architecture: Apple Silicon (arm64)
• Trigger: a $grill-me skill plus AGENTS.md requiring lettered, transcript-based decisions
• Observed behavior: native picker briefly appears then disappears; the required Markdown fallback is absent
• No tab switch or chat switch was involved in this occurrence

The version and hardware values were read directly from the affected Mac on 2026-09-10. No private task content, screenshots, or logs are included.

Start with the Codex Desktop settings entry point and the structured-question tool exposure path, then trace how per-chat instructions reach the harness. Done means a persistent, visible global setting and a reliable per-chat override prevent native UI-only questions and preserve a durable Markdown transcript fallback, with limitations reported clearly.