openai / openai/codex

[CRITICAL][Windows] Recurring helper setup failures block both browser control and local execution

Open
#43,721 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app browser bug sandbox tool-calls 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)?

Not verified in this session. The local command runner fails before process creation, and the browser control runtime fails during initialization. No version number is inferred from other reports.

What subscription do you have?

Not provided in this report.

What platform is your computer?

Windows, Codex desktop app, local workspace, PowerShell.
Session configuration: workspace-write filesystem sandbox, restricted network access, managed automatic approval review.
Exact Windows build and architecture were not collected.

What issue are you seeing?

Reporter severity: CRITICAL — recurring failure blocks the requested workflow before useful work can start. Please prioritize investigation.

Both browser control and local command execution fail in the same task. The user reports that this is an ongoing, repeatedly recurring problem, not an isolated inconvenience. The evidence below documents the failures directly observed in this session on September 8, 2026; the longer recurrence history is user-reported.

The requested workflow was to read a user-provided local PDF, compare it with the user's professional profile, and add missing skills through the browser. Browser initialization never returned usable state, and local execution failed before the document could be inspected. No profile changes were made.

Browser control error, returned twice:

trusted Node process exited unexpectedly; kernel reset, rerun your request

Local execution error, returned on four exec_command attempts across the original task and preparation of this report:

exec_command failed: CreateProcess { message: "Rejected(\"Failed to create unified exec process: helper_unknown_error: setup refresh had errors\")" }

The failures also prevent ordinary local diagnostics from running. This leaves the user unable to complete the task or obtain an actionable explanation through the affected tools.

What steps can reproduce the bug?

Observed sequence in the affected session:

  1. Open a local Codex desktop task on Windows and request a workflow requiring a local file read and browser interaction.

  2. Attempt ordinary read-only PowerShell file reads through exec_command. Two independent calls both fail during helper setup, before returning file contents.

  3. Initialize browser discovery with the documented entry point:

    await cua.getState();
    

    Result: trusted Node process exits; the tool reports a kernel reset.

  4. Retry browser initialization using the documented URL entry point. The actual target was the user's profile; this example redacts the personal URL:

    let browser = await cua.getBrowser({ url: "https://www.linkedin.com/in/<redacted>/" });
    

    Result: identical trusted Node process error. The placeholder URL above is illustrative; it was not separately tested.

  5. Retry a read-only PowerShell file read using exec_command with login: false. It fails with the same setup refresh error.

  6. During the subsequent request to report the bug, another read-only exec_command call with login: false fails identically.

Observed counts: 2 failed browser initialization calls and 4 failed local execution calls. The browser discovery call fails without a target URL, so the observed browser failure precedes any profile interaction.

Personal paths, document content, and profile identity have been omitted. No diagnostic bundle or feedback ID was collected for this report.

What is the expected behavior?
  • Browser control should initialize and return browser/tab state.
  • The local command runner should start authorized read-only commands.
  • A runtime initialization failure should provide an actionable cause and a supported recovery path.
  • Recovery should preserve the current task and should not depend on repeatedly asking users to restart without evidence that a restart addresses the failure.
Additional information

Troubleshooting actually performed: retried browser initialization after the tool-reported kernel reset; retried local execution with login: false. Neither restored the affected capability.

The assistant initially suggested restarting Codex, then explicitly acknowledged that it had no evidence a restart was necessary or would fix this incident. A restart was not performed or validated in this session. Please do not treat that suggestion as a confirmed workaround.

The messages indicate failure during runtime/helper initialization, but the underlying cause and whether the two failures share a cause are unverified. There is no evidence here of a LinkedIn authentication failure, a malformed PDF, or an explicit approval-review denial.

GitHub connector requests still work, which makes it possible to submit this report despite the local execution and browser failures.

Related reports found before submission:

  • #43255 — Windows Computer Use initialization failure reporting both setup refresh and trusted Node process errors.
  • #42905 — Windows local command runner failure with the same helper error and trusted Node exit.
  • #43416 — Windows sandbox setup refresh failure blocking a browser smoke test.
  • #42513 — Repeated Windows sandbox setup refresh failure, including in a new task.
  • #39841 — Earlier workspace terminal startup failure with the same setup refresh message.

These reports show similar symptoms; identical root cause has not been established. This issue supplies another concrete recurrence with simultaneous browser and local execution failure and an explicit request for critical triage. Please link it to the owning incident/root-cause investigation and provide a durable fix, actionable diagnostics, and regression coverage.

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

Start with the exec_command helper setup refresh and the cua.getState() browser initialization entry points, then compare the related Windows reports listed in the issue. Done means identifying the failure cause, providing actionable diagnostics and recovery, and adding regression coverage for the recurring setup failures.

Written by the indexing model from the issue text.

Assessment

Tech stack
node.js, powershell, rust
Domain
devtools, operating-systems
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.