openai / openai/codex

[Windows][26.825.6671.0] Desktop crashes or hides its window during three concurrent image-generation tasks

Open
#41,959 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug imagen performance windows-os
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

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

OpenAI.Codex 26.825.6671.0 (Microsoft Store MSIX)

The packaged ChatGPT.exe reports Chromium version 151.0.7922.174.

What subscription do you have?

ChatGPT paid subscription. The exact tier was not verified; the failure is in the local Windows desktop client.

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

Additional hardware information:

  • 16 GB RAM
  • NVIDIA GeForce RTX 3050
  • NVIDIA driver 32.0.15.9186
What issue are you seeing?

Codex Desktop for Windows becomes unstable when three task threads invoke image generation concurrently. I observed two full browser/main-process crashes and a third reproduction where the foreground window disappeared while the original main process and background image-generation tasks continued running.

The failure has happened repeatedly. Crashpad recorded 10 crashes during the same morning: 9 with ptype=browser and 1 with ptype=renderer.

Two reproductions have a particularly close timing relationship to concurrent image-generation activity (all times below are local time, UTC+08:00):

  • Run 1: three threads issued interleaved image-generation calls from 09:38:14 through 09:46:02. Crashpad recorded a browser-process crash at 09:46:10, eight seconds after the last call.
  • Run 2: three threads issued calls from 09:52:32 through 09:55:05. Crashpad recorded another browser-process crash at 09:55:26, 21 seconds after the last call.
  • Run 3: three threads began at approximately 10:03 and continued generating images through at least 10:15. The foreground Codex window disappeared, but the original ChatGPT.exe browser process remained alive and the background image-generation calls continued.

In Run 3, the three session rollout files grew to approximately 44.8 MB, 42.5 MB, and 47.1 MB (about 134.4 MB total). About 70 records exceeded 1 MB, with individual records up to roughly 2.7 MB. One renderer process reached approximately 1.04 GB working set.

This suggests that concurrent image results and/or large image payloads are stressing a Codex Desktop browser/renderer/window lifecycle path. Resource load may be an amplifier, but the evidence does not support ordinary computer under-specification as the root cause.

During some restart loops, the desktop log contained:

[window-manager] Failed to load bundle at errorCode=ERR_FAILED errorMessage="ERR_FAILED (-2) loading 'app://-/index.html'"

The renderer also repeatedly emitted:

[desktop-notifications][global-error] ResizeObserver loop completed with undelivered notifications

These are included as accompanying/recovery symptoms, not asserted as the root cause.

What steps can reproduce the bug?
  1. Launch Codex Desktop on Windows.
  2. Open three separate Codex task threads.
  3. Assign an image-generation workload to each thread.
  4. Let all three threads invoke the image-generation tool concurrently.
  5. Continue queuing or delegating additional images while all three tasks are active.
  6. Observe that the Codex foreground window closes/disappears, or the browser/main process crashes and the app restarts.

The problem is much easier to reproduce with three simultaneous image-generation tasks than with ordinary text or single-task work.

What is the expected behavior?

Codex should keep the desktop window and browser/main process stable while multiple background tasks are active. If image-generation concurrency or payload size exceeds a safe limit, the client should queue, throttle, virtualize, or release those payloads instead of closing the foreground window or crashing the application.

A renderer or task failure should also be isolated so other tasks and the desktop host remain usable.

Additional information

Checks performed around the reproductions:

  • No Windows resource-exhaustion or out-of-memory event was recorded.
  • No WHEA hardware error, NVIDIA display-driver reset, or LiveKernel GPU report was recorded.
  • A pagefile of approximately 21 GB was available, with substantial free virtual memory.
  • The Codex process group used approximately 2.1 GB at the time of inspection, rather than exhausting all system memory.
  • No full native Codex dump was retained; only Crashpad sidecar metadata is currently available.
  • Raw rollout files and full desktop logs are not attached because they contain local paths, prompts, and conversation data. Additional sanitized timestamps can be provided, and a full ChatGPT.exe dump can be configured for a future reproduction if maintainers request it.

Current workaround: allow only one active image-generation task at a time.

Potentially related but not exact duplicates:

  • #41581 — long/tool-heavy Windows sessions can close the Desktop app, but it does not identify concurrent image generation as a repeatable trigger.
  • #40349 — browser/main crashes after restoring oversized in-app-browser tasks, with a different trigger path.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Reproduce the failure with three concurrent image-generation tasks, then inspect the Crashpad metadata, desktop log, and three session rollout files from the run. A fix is complete when the Windows desktop window and browser/main process remain usable under this workload, or task and payload limits isolate failures without closing or crashing the app.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
desktop
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.