Codex Windows app exits when a shell tool launches interactive llama-cli.exe

Open
#34,543 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
48/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Quiet
Tech stack
powershell, rust
Domain
desktop, devtools

Research direction

Start by running the provided PowerShell reproduction with llama-cli.exe, then inspect the shell tool's interactive child-process path and the AppServerConnection stdio transport mentioned in the log. Done means an interactive child process no longer exits the Codex desktop app: it is safely isolated, rejected with a normal tool error, or terminated without stopping the app-server.

Written by the indexing model from the issue text.

Description

app bug tool-calls windows-os
What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.715.8383.0

The desktop log reports app-server version 0.145.0-alpha.18.

What subscription do you have?

ChatGPT subscription (exact tier not relevant to this local process crash).

What platform is your computer?

Microsoft Windows 11, version 10.0.26200.8894, x64.

What issue are you seeing?

The Codex Windows desktop app exits/restarts when a shell tool directly launches an interactive native console executable. I reproduced this twice with llama.cpp's llama-cli.exe in conversation mode.

In both cases, the active Codex tool call changed to aborted because the entire desktop app disappeared. After the app restarted, the thread could be resumed.

The relevant pre-restart desktop log ends abruptly with:

2026-07-21T13:30:40.420Z info [AppServerConnection] Stopping app-server transport connectionId=7 transport=stdio

There is no corresponding application crash entry in Windows Application Event Log and no useful exception/stack trace in the Codex desktop log.

The same GGUF and llama.cpp build load and generate successfully when llama-server.exe is launched as a hidden detached process with stdout/stderr redirected, then called over HTTP. This points to the interactive console/TTY child-process path rather than the model or executable itself.

What steps can reproduce the bug?
  1. Open the Codex Windows app and start a local thread in a PowerShell workspace.
  2. Have Codex execute this through its shell tool (paths can be adjusted to any local llama.cpp build and GGUF):
& '.\llama-cpp-cuda-13\llama-cli.exe' `
  -m '.\models\Astrea\R8-Chat-9B\Astrea-R8-Chat-9B-Q8_0.gguf' `
  -ngl all -c 4096 -cnv `
  -p 'Write exactly one vivid sentence about a lighthouse in a winter storm.' `
  -n 64 --temp 0.2 --top-p 0.95 --no-display-prompt
  1. The Codex desktop app exits during the tool call.
  2. Reopen Codex and resume the thread; the tool call is recorded as aborted.
  3. Repeating the same tool call crashes the app again.

Reproduction rate in this session: 2/2.

What is the expected behavior?

An interactive or TTY-oriented child process should not be able to terminate the Codex desktop app. Codex should either:

  • run it safely under isolated console handles / a PTY,
  • reject unsupported interactive execution with a normal tool error, or
  • terminate only the child process while keeping the desktop app and app-server alive.
Additional information
  • Codex package: OpenAI.Codex_26.715.8383.0_x64
  • Desktop executable: ChatGPT.exe
  • Shell: PowerShell
  • The child was a locally built Windows x64 llama.cpp CUDA executable.
  • Safe workaround: Start-Process -WindowStyle Hidden with -RedirectStandardOutput and -RedirectStandardError, then communicate with the detached llama-server.exe over localhost HTTP.
  • I did not attach the full desktop log because it contains unrelated local connection details. The final relevant line is included above.
Dominant language
Rust
Stars
125k
Forks
19.5k
Avg merge
1m
Merged PRs (30d)
1k

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.

More from openai/codex

All issues in openai/codex

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.