Codex Desktop for Windows Repeatedly Exits and Relaunches During Reasoning Summaries (Workaround)

Bug Report

Title: Codex desktop for Windows repeatedly exits and relaunches during streamed reasoning summaries

Severity: Critical / application-blocking

Environment

  • Windows 11 Pro x64, build 26200
  • Codex Microsoft Store package: OpenAI.Codex_26.721.11231.0_x64
  • Executable: ChatGPT.exe
  • Chromium/Electron version reported by Crashpad: 150.0.7871.128
  • Model: gpt-5.6-sol
  • Reasoning effort: high
  • Memory: 64 GB DDR5
  • Multi-monitor configuration
  • Time zone: Europe/Berlin, CEST (UTC+2)

Problem

The Codex desktop application repeatedly disappears and relaunches while an active, tool-using turn is streaming reasoning-summary updates. During the worst periods, this happens approximately every one to two minutes.

The current reproducible exits are not conventional access-violation crashes:

  • The main ChatGPT.exe process, codex.exe, GPU process, renderers, utilities, and Crashpad handler all terminate almost simultaneously.
  • Every monitored process reports exit code 0.
  • Windows records the AppX container being destroyed and recreated.
  • The application relaunches approximately five to eight seconds later.
  • No Windows Error Reporting or Crashpad dump is generated for these reproducible exits.

Reproduction Steps

  1. Start the Codex desktop application on Windows.
  2. Open a workspace and an existing task.
  3. Use a reasoning-capable model with reasoning effort set to high.
  4. Start a longer, multi-step task involving several tool calls.
  5. Allow multiple reasoning-summary updates to stream into the UI.
  6. Continue generating reasoning updates and UI status changes.

The application eventually closes completely and relaunches.

Controlled Reproduction

A controlled test began at 2026-07-30 07:50:26 CEST.

Within the initial phase, the desktop log recorded:

  • 6 reasoning-summary events
  • 12 ResizeObserver errors

Immediately before the exit:

2026-07-30T05:50:47.128Z
Reasoning summary item completed
rendererWindowAppearance=primary

2026-07-30T05:50:47.128Z
Reasoning summary item completed
rendererWindowAppearance=avatarOverlay
rendererWindowVisible=false

2026-07-30T05:50:47.161Z
[desktop-notifications][global-error]
ResizeObserver loop completed with undelivered notifications.

2026-07-30T05:50:47.304Z
[desktop-notifications][global-error]
ResizeObserver loop completed with undelivered notifications.

2026-07-30T05:50:47.382Z
[desktop-notifications][global-error]
ResizeObserver loop completed with undelivered notifications.

2026-07-30T05:50:49.585Z
[desktop-notifications][global-error]
ResizeObserver loop completed with undelivered notifications.

At approximately 07:50:50 CEST, all application processes terminated with exit code 0. Windows recorded AppX container destruction at the same time. A new container and main process were created at 07:50:55.

The same clean-exit pattern was reproduced again at:

  • 07:53:17
  • 07:55:23

Relevant Feature-Override Evidence

Each new turn logs:

Reasoning summary turn-start config resolved
concurrentReasoningSummariesFeatureOverrideEnabled=true
summary=detailed

Reasoning-summary events are delivered to both:

rendererWindowAppearance=primary

and the invisible overlay renderer:

rendererWindowAppearance=avatarOverlay
rendererWindowVisible=false

This duplicated rendering may be relevant.

Configuration Workaround Is Overridden

The following documented configuration was applied:

model_reasoning_summary = "none"

However, subsequent turns still resolved to:

summary=detailed
concurrentReasoningSummariesFeatureOverrideEnabled=true

A second setting was then applied:

model_supports_reasoning_summaries = false

The frontend continued to create reasoning-summary items, including some containing text, and continued logging ResizeObserver errors.

The desktop feature override therefore appears to take precedence over the user’s global Codex configuration, preventing the affected feature from being fully disabled.

Troubleshooting Already Completed

  • MyASUS offline WinRT RAM and SSD diagnostics passed.
  • No WHEA, disk, or filesystem hardware errors were found.
  • Application memory consumption was not near the system limit.
  • The Codex SQLite log database passed PRAGMA quick_check.
  • A corrupted local process_manager/chat_processes.json file was backed up and recreated as valid JSON. This removed its parsing warning but did not stop the exits.
  • The application logs contain no quit, shutdown, relaunch, fatal, OOM, or render-process-gone entry immediately before termination.
  • No Crashpad dump is created for the reproducible clean exits.

Expected Behavior

The application should remain open while reasoning summaries and tool activity are streamed. A renderer or layout error must not terminate and relaunch the entire desktop application.

The settings below should also be honored by the desktop application:

model_reasoning_summary = "none"
model_supports_reasoning_summaries = false

Actual Behavior

Reasoning summaries remain forced to detailed by the concurrent_reasoning_summaries feature override. Repeated ResizeObserver errors occur, followed by a coordinated exit of the entire application process tree with exit code 0.

Requested Investigation

Please investigate:

  1. Why concurrent_reasoning_summaries forces summary=detailed over explicit user configuration.
  2. Why summary events are sent to both the primary renderer and an invisible avatarOverlay renderer.
  3. Whether the repeated ResizeObserver failures can trigger the application-wide clean exit.
  4. Why no shutdown reason or relaunch event is written before the AppX container is destroyed.
  5. Whether a temporary client-side kill switch can be provided for concurrent reasoning summaries.

Available diagnostic attachments include desktop logs, AppModel-Runtime events, and a timestamped JSONL process-exit trace.

Edit - After 54 Min. :

After applying model_reasoning_summary = "none" and model_supports_reasoning_summaries = false, application uptime improved from approximately 1–2 minutes to 53 minutes and 34 seconds. However, the failure still occurred. The final logged event was an empty reasoning-summary item delivered to both the primary renderer and the invisible avatar overlay. Approximately 1.4 seconds later, the entire process tree exited with status code 0. This suggests that the workaround reduces the failure frequency but does not eliminate the underlying issue.

Welcome to the forum!

You are welcome to discuss Codex issues on the forum, however the official place to report and track them is the GitHub issue tracker for OpenAI Codex.

I had ChatGPT look for the closest related issue and it identified:

The OP of the issue has several replies trying to locate the root cause and look for workarounds.

If that issue matches your problem, consider adding a :+1: reaction on the GitHub issue page itself, not just on the forum topic, as reactions on GitHub help the developers gauge impact and prioritize work.

If it is not the same issue, try searching the existing GitHub issues for a closer match. If you cannot find one, then open a new issue with as much detail as possible.