Add a setting to disable timed auto-resolution for AI questions

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #28969
• #29055

Powered by Codex Action

I’m seeing the same thing on macOS with Codex App 26.707.62119 (5211).

I checked the session JSONL because I initially assumed the model was still passing a timeout. It wasn’t. The last 20 request_user_input calls all omitted autoResolutionMs, but unanswered calls still returned {"answers":{}} after almost exactly 90 seconds.

So the AGENTS.md workaround no longer helps. The current Desktop package also has the same trackRequest(..., "empty-user-input") call shape reported for the VS Code frontend in #28969. It does not forward the request’s autoResolutionMs.

I usually have several agent tasks running and come back to a question a few minutes later. Requiring an answer within 90 seconds means I have to watch every task continuously, which defeats much of the point of the App.

If autoResolutionMs is omitted, the question should wait. A timer should only start when the field is present. A global “Always wait for my answer” setting would still be useful; clicking Snooze on every question is not a real substitute.

Additional severity/reproduction report from Codex App 26.707.71524 (Codex CLI 0.144.3):

This is not merely a missing convenience setting. The user's global custom instructions explicitly stated:

Do not use auto-resolving user prompts. Never set autoResolutionMs on request_user_input. If you ask me a question, wait for my explicit answer. Do not choose defaults from silence. Treat no response as no decision and do not proceed with a plan based on it.

Despite those instructions, the agent used an auto-resolving prompt. The countdown expired, the agent's self-designated recommendation was selected, and work proceeded without the user's decision.

That converts silence into authorization while directly violating an explicit account-wide preference. Custom instructions cannot be the enforcement mechanism because the same model that chooses the tool arguments can ignore them.

Expected product behavior: provide a host-enforced, account-wide control that rejects or strips autoResolutionMs from AI-generated prompts and never selects an option on the user's behalf. Please treat this as a user-agency/consent failure rather than only a UX preference; proceeding on a fabricated decision is severe enough to be a deal-breaker.

I am seeing the same behavior in Codex Desktop on macOS 26.5.1:

• Codex App: 26.707.72221 (5307)
• Codex CLI: 0.144.4
• default_mode_request_user_input = true

My normal workflow is to start a task, switch to other work, and return when Codex needs a decision. With timed auto-resolution, the question can disappear before I return and Codex may continue without an explicit answer. This defeats the purpose of running tasks asynchronously and turns user silence into a decision.

I have disabled default_mode_request_user_input as a workaround, but that also removes the useful structured question UI from Default mode.

Expected behavior: questions should wait indefinitely by default. Auto-resolution should only happen when the user has explicitly opted into it; it should not be a hidden or model-selected timeout. In particular, omitting autoResolutionMs must reliably mean “wait for the user.”

I think this issue exposes a decision-provenance problem in addition to the timeout UX problem.

An unanswered question can terminate in several ways:

TIMED_OUT
DISMISSED
CONNECTION_LOST
CANCELLED_BY_SYSTEM

None of those states should be represented as a user-selected answer or as authorization to proceed with a recommended option.

I drafted a small implementation-neutral response-state model and conformance matrix here:

https://github.com/safal207/DIF/issues/35

The central invariant is:

«No explicit human answer means no confirmed human decision.»

A timeout receipt could look like:

{
"question_id": "q-123",
"response_state": "TIMED_OUT",
"selected_option": null,
"decision_source": "SYSTEM",
"human_confirmed": false,
"agent_authority": "STOP_AND_WAIT"
}

This would preserve an explicit distinction between:

• a user selecting an option;
• a user explicitly delegating the choice to Codex;
• a timer or transport condition ending the question.

Possible conformance requirements:

1. Omitting "autoResolutionMs" means wait indefinitely.
2. A timeout never returns a selected option.
3. A recommended option is not treated as user-selected.
4. Desktop, CLI, IDE, and Remote surfaces use the same terminal-state semantics.
5. System-resolved outcomes carry non-human provenance.
6. Continuing after timeout requires a separate, explicitly defined policy rather than fabricated user consent.

This proposal does not require adopting DIF terminology. I am offering the state model, fixtures, and test cases as problem-domain analysis.

This behavior is SO infuriating, I've issued every imaginable version of "ignore autoResolutionMs" in my agents.md
Nothing works
Humans have to go to the bathroom sometimes without worrying about missing a 60sec timeout