openai / openai/codex

Windows Desktop terminates when reopening tasks with very large tracked evidence and review state

Open
#38,622 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug code-review performance session windows-os
Dominant language
Rust
Stars
125k
Forks
19.5k
PR merge metrics
PR metrics pending

Description

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

The exact package version was not preserved in the sanitized incident record. The failures were observed on Windows Codex/ChatGPT Desktop on July 25, 2026.

What subscription do you have?

Not relevant to the local renderer/review-state failure.

What platform is your computer?

Windows x64.

What issue are you seeing?

Codex Desktop repeatedly terminated while opening or continuing a long-running task associated with a Git workspace containing a very large tracked generated-evidence surface. Opening a fork of the completed task history reproduced the instability, while moving to a blank task with a compact handoff avoided the hazardous cached history.

The repository conditions measured before cleanup were:

  • 2,065 tracked files under the evidence tree
  • 189,973,786 tracked working-tree bytes in that tree
  • 1,488 structured log/capture files
  • 340 binary files
  • nine bundle files totaling 83,462,101 bytes
  • approximately 603 MiB of loose Git objects
  • 124 additional untracked files during the triggering campaign
  • generated review commands containing 111 explicit artifact paths
  • repeated review/diff responses around 1.5 MiB

Crashpad captured a roughly 49 MiB minidump during one occurrence. Desktop logs around the failures included repeated ResizeObserver and item-state errors before the main process terminated.

The raw dump, logs, repository name, local paths, prompts, and task content are intentionally not attached because they may contain private project and conversation data.

What steps can reproduce the bug?

The original failure was workload-sensitive rather than reduced to a minimal fixture:

  1. Open a Git workspace containing thousands of tracked generated evidence files and roughly 180 MiB of evidence content.
  2. Run a long qualification task that causes large Git review/diff state and references many artifact paths.
  3. Reopen or continue that long-running task in Codex Desktop.
  4. The desktop application may terminate while loading or rendering the task/review state.
  5. Forking the completed task history can preserve the same instability.

A repository-side mitigation removed 1,945 raw generated files (163.35 MiB) from the current Git snapshot, retained only compact human-reviewable evidence, moved future raw payloads to an ignored generated-data directory, and used a blank task with a compact handoff.

What is the expected behavior?

Codex Desktop should degrade gracefully when repository or cached task/review state is unusually large. It should bound or paginate review processing, avoid embedding huge path lists or diff responses into renderer state, and surface a recoverable warning instead of terminating the application. Opening a historical task or a fork should not repeatedly crash the desktop client.

Additional information

This report is about robustness even when the workspace is poorly shaped. Keeping generated evidence out of Git substantially reduces the trigger, but oversized or inconsistent task/review state should still be handled safely by the desktop application.

Possibly related: #33996 reports ResizeObserver loops and Windows UI freezes, but the distinguishing trigger here is large tracked generated evidence plus cached review/task state.

Sanitized aggregate diagnostics or narrowly selected log excerpts can be provided if maintainers specify what is useful. The raw Crashpad bundle will not be uploaded publicly.

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

No source files, tests, or code entry points are named. Start by reviewing the available sanitized diagnostics and desktop logs, then reproduce with a large tracked-evidence workspace and reopened task state; done means oversized review state no longer terminates the Windows desktop client and instead remains recoverable.

Written by the indexing model from the issue text.

Assessment

Tech stack
git, rust
Domain
desktop
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.