Skip to content

Codex iOS remote task on Windows 26.527 ignores xhigh and starts gpt-5.5 medium #25583

Open
Open
Codex iOS remote task on Windows 26.527 ignores xhigh and starts gpt-5.5 medium#25583
Labels
appIssues related to the Codex desktop appapp-serverIssues involving app server protocol or interfacesbugSomething isn't workingiOSIssues related to the Codex iOS appremotewindows-osIssues related to Codex on Windows systems

Description

@xiaobing318
xiaobing318
opened on Jun 1, 2026

What version of the Codex App are you using (From “About Codex” dialog)?

Codex 26.527.31326

What subscription do you have?

PRO

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Summary

When I start a Codex task from the ChatGPT iOS app on my iPhone and explicitly select gpt-5.5 with xhigh reasoning, the corresponding session on my Windows Codex app consistently starts/runs as gpt-5.5 medium.

This appears to happen specifically through the mobile remote-task / remote-control path. My expectation is that the model and reasoning effort selected on iOS should be propagated to the Windows Codex app-server/session for the initial turn.

What I am seeing

  1. I open ChatGPT on iOS.

  2. I use the Codex integration to send/start a task for an existing project on my Windows machine.

  3. Before sending the task, I manually select:

    • model: gpt-5.5
    • reasoning effort: xhigh

  4. The task is successfully delivered to the Windows Codex app.

  5. When I inspect the corresponding session on the Windows machine, the session shows/runs as:

    • gpt-5.5 medium

This happens repeatedly. I have not seen the mobile-started task preserve xhigh on the Windows side.

Actual behavior

The Windows Codex session is created as:

gpt-5.5 medium

This suggests that the mobile-selected xhigh reasoning effort may not be passed to the Windows app-server/session creation path, or it is being overwritten by a default / persisted / collaboration-mode setting.

What steps can reproduce the bug?

Steps to reproduce

  1. Install and sign in to Codex for Windows.

  2. Use a Windows Codex app version with Windows remote-control support, for example 26.527.31326.

  3. Connect the Windows Codex host to ChatGPT mobile.

  4. On iPhone, open ChatGPT iOS 1.2026.139.

  5. Open Codex from ChatGPT iOS.

  6. Select a project hosted on the Windows machine.

  7. Start a new Codex task from the iOS app.

  8. Before submitting the task, select:

    • gpt-5.5
    • xhigh

  9. Submit the task.

  10. On the Windows machine, open the corresponding Codex session.

  11. Check the model/reasoning shown in the session, /status, or the local session JSONL.

What is the expected behavior?

Expected behavior

The Windows Codex session created from ChatGPT iOS should use the selected model and reasoning effort:

gpt-5.5 xhigh

The initial turn should receive the explicit xhigh reasoning setting selected on iOS.

Additional information

No response

Activity

  1. xiaobing318
    added
    appIssues related to the Codex desktop app
    on Jun 1, 2026
  2. github-actions
    added
    bugSomething isn't working
    iOSIssues related to the Codex iOS app
    windows-osIssues related to Codex on Windows systems
    remote
    app-serverIssues involving app server protocol or interfaces
    on Jun 1, 2026

  3. github-actions commented on Jun 1, 2026

    @github-actions
    github-actionsbot
    on Jun 1, 2026 – with GitHub Actions
    Contributor

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

    Powered by Codex Action

  4. nerdeveloper commented on Jul 14, 2026

    @nerdeveloper
    nerdeveloper
    on Jul 14, 2026

    Confirming this on an iOS-controlled Codex thread backed by a macOS host, with a more specific Plan mode trigger.

    Environment

    • Client: ChatGPT iOS Codex remote UI
    • Host: macOS 14.8.7 (arm64)
    • Host app: ChatGPT/Codex Desktop 26.707.62119 (build 5211)
    • Embedded Codex/app-server: 0.144.2
    • Model: gpt-5.6-sol

    Reproduction

    1. In the same iOS-controlled remote thread, leave collaboration mode on Default.
    2. Select Max reasoning and send a prompt.
    3. Select Ultra reasoning and send another prompt.
    4. Switch the iOS composer to Plan mode.
    5. Select a higher reasoning effort and send several prompts.
    6. Inspect the host-side session turn_context metadata.

    Observed

    The normal-mode selections propagated correctly:

    default  xhigh
default  max
default  ultra

    As soon as Plan mode was enabled, the iOS selector reverted to Medium and every tested Plan-mode turn was recorded as:

    plan  medium
plan  medium
plan  medium
plan  medium

    The effective effort remained Medium across repeated attempts. This is therefore not only a stale UI label: the host-side turn metadata confirms that Plan mode is actually sent/executed with medium.

    Expected

    An explicit reasoning-effort selection in the iOS composer should be preserved when switching to Plan mode and propagated to the host session. Plan mode should not silently replace Max/Ultra/xhigh with Medium.

  5. alloutflo commented on Oct 16, 2026

    @alloutflo
    alloutflo
    on Oct 16, 2026

    Adding a current macOS-host reproduction of the Plan-mode behavior described in the July 14 comment.

    Observed on October 16, 2026:

    • Controller: ChatGPT iPhone Remote UI; mobile app/iOS versions not captured.
    • Host: Mac mini, macOS 27.0.1.
    • Installed ChatGPT/Codex desktop app: 26.1007.71326 (build 20398).
    • Recorded embedded CLI version: 0.162.0-alpha.17.2.
    • Model: gpt-6-astra.
    • No explicit plan_mode_reasoning_effort entry in the inspected user config.

    Reported interaction: select Astra Ultra, enable Plan mode, then submit a short test prompt. The iPhone reasoning control changes to Medium after submission. User-provided screenshots document the composer/control transition, but the expanded picker labels are not visible, so the pre-submit Ultra selection is user-reported rather than independently established by those images.

    The matching host-side turn_context records were inspected directly:

    Local time (UTC+02:00) Prompt Model Recorded effort Collaboration mode
    11:38:48 This is a test gpt-6-astra medium plan
    11:38:58 Lets go gpt-6-astra medium plan
    11:39:16 Create a picture of a tree gpt-6-astra medium plan
    11:39:26 Implement the proposed plan. gpt-6-astra medium default

    This establishes that the local recorded settings were Medium for the initial request and subsequent requests, including the transition back to Default; this is not merely an inference from the composer icon. These are local turn metadata, not independent server-side inference telemetry.

    Expected: preserve the explicitly selected reasoning effort across the Plan-mode transition and submission, or clearly display any override before sending. The silent change makes it difficult to know which reasoning setting the initial and follow-up requests will use.

    The mobile-versus-desktop A/B test and exact root cause remain unconfirmed. Could maintainers investigate the iOS Remote Plan-mode settings propagation path? This appears closely aligned with the July macOS report, now observed with Astra and a newer host runtime.

    No private conversation content, raw session logs, or account/session identifiers are attached.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appapp-serverIssues involving app server protocol or interfacesbugSomething isn't workingiOSIssues related to the Codex iOS appremotewindows-osIssues related to Codex on Windows systems

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions