openai / openai/codex

[Windows][Desktop 26.915.31029] Codex repeatedly auto-compacts context, reconnects, times out, and re-executes trivial tool calls

Open
#46,423 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug connectivity context performance 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)?

ChatGPT Desktop version: 26.915.31029

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Codex Desktop on Windows is extremely slow and intermittently fails during even trivial local shell commands.

Minimal reproduction:

Workspace:
E:\Codex_Test

The workspace is completely empty, and the test was performed in a brand-new Codex conversation.

Prompt / command:

Write-Output "CODEX_OK"

Expected behavior:

The command should execute once and return:

CODEX_OK

within a few seconds.

Actual behavior:

The latest test took approximately 3 minutes 55 seconds.

During the turn, Codex displayed:

Context automatically compacted

Reconnecting /2

request timed out

The same command was then executed multiple times:

Write-Output "CODEX_OK"

Each shell execution successfully returned:

CODEX_OK

but Codex repeated the tool call before eventually completing the response.

Earlier, very small commands such as:

Get-Location

also intermittently returned:

Output exceeded the available model context and was truncated

even though the expected output was only one line.

This issue occurs in:

  • my original research workspace,
  • a new conversation,
  • and a completely empty test workspace.

Troubleshooting already performed:

  • updated ChatGPT Desktop,
  • restarted the app,
  • restarted Windows,
  • reset the Windows app,
  • fully uninstalled ChatGPT Desktop,
  • reinstalled it from the Microsoft Store.

After Windows App Reset, the same trivial command still took approximately 1 minute 21 seconds. After a clean reinstall, it took approximately 3 minutes 55 seconds and showed repeated context compaction, reconnecting, timeout, and repeated tool execution.

ChatGPT Desktop version:
26.915.31029
Release date shown in the app: September 18, 2026.

The local PowerShell command itself appears to execute successfully. The problem seems to occur around Codex tool-result handling, request/reconnection, or context management rather than local file access or workspace size.

What steps can reproduce the bug?

Steps to reproduce:

  1. Open ChatGPT Desktop on Windows.
  2. Open Codex.
  3. Open a completely empty local folder as the workspace.
  4. Start a brand-new Codex conversation.
  5. Send the following prompt:
Run only the following PowerShell command:

Write-Output "CODEX_OK"
  1. Observe the execution.

Expected result:

CODEX_OK

should be returned within a few seconds and the command should execute only once.

Actual result:

In my latest reproduction, the turn took approximately 3 minutes 55 seconds.

During the turn, Codex displayed:

Context automatically compacted
Reconnecting /2
request timed out

The same PowerShell command was then executed successfully multiple times before the response eventually completed.

The issue can also be reproduced with another trivial command:

Get-Location

On some attempts, Codex instead reports:

Output exceeded the available model context and was truncated

even though the expected command output is only one line.

The issue reproduces in:

  • a completely empty workspace,
  • a brand-new conversation,
  • and after resetting and fully reinstalling ChatGPT Desktop.

Session ID: Not available / not shown in the Desktop UI.

Token limit usage: Not available / not shown in the Desktop UI.

Context window usage: Not available / not shown in the Desktop UI. The conversation was brand new with only the minimal reproduction prompt, but Codex still displayed "Context automatically compacted."

What is the expected behavior?

Codex Desktop should execute a trivial local shell command once and return the result within a few seconds.

For example, when running:

Write-Output "CODEX_OK"

the expected result is:

CODEX_OK

The command should not trigger automatic context compaction, reconnect attempts, request timeouts, or repeated execution of the same tool call.

A new conversation in an empty workspace should also complete this simple command without noticeable delay.

Additional information

The issue is reproducible in a completely empty local workspace and a brand-new conversation, so it does not appear to depend on project size, repository contents, or conversation history.

The local shell command itself succeeds when it eventually runs. The abnormal behavior appears to happen around tool-result handling, reconnection, timeout, or context compaction.

I am also concerned about repeated execution of the same successful tool call after a reconnect. This could be unsafe for non-idempotent operations such as appending to files, moving files, or running scripts that modify outputs.

I can provide additional logs or diagnostic information if needed.

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 by reproducing the issue in Codex Desktop on Windows with an empty workspace and a new conversation, using Write-Output "CODEX_OK" and Get-Location. Compare the observed execution with the expected single result, focusing on context compaction, reconnects, timeouts, and repeated tool calls; completion means the command runs once and returns within a few seconds.

Written by the indexing model from the issue text.

Assessment

Tech stack
powershell
Domain
desktop, devtools
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.