›_ Codex Issues
Codex Issues / #24414

Codex IDE extension often stays on "Thinking" for a long time even for simple prompts

Open 💬 7 comments Opened May 25, 2026 by Yong-yuan-X
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the IDE extension are you using?

ChatGPT/Codex VS Code extension: 26.513.21555

What subscription do you have?

ChatGPT Plus

Which IDE are you using?

VS Code

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Codex IDE extension stays on "Thinking" for a long time

Over the past two days, the Codex IDE extension has often stayed on the "Thinking" state for a long time, even for very simple prompts. It can take around 5 minutes even for simple tasks.

My network connection is working normally, and other websites and services work as expected. This issue happens repeatedly in VS Code. Sometimes Codex does not return a response for a long time and does not show any clear error message. Other extensions do not seem to have this issue.

<img width="552" height="169" alt="Image" src="https://github.com/user-attachments/assets/a8721367-4c75-44c9-80d0-ce569166a086" />

What steps can reproduce the bug?

  1. Open VS Code.
  2. Open any repository.
  3. Open the Codex IDE extension panel.
  4. Send a simple prompt, for example:
  • "Please explain this small function."
  • "Check this file."
  • "Review this small change."
  1. Observe that Codex often stays on "Thinking" for a long time, sometimes around 5 minutes, without returning a response or showing a clear error message.

This does not happen every single time, but it has been happening repeatedly over the past two days.

What is the expected behavior?

Codex should start responding within a reasonable time, or show a clear timeout/error message if something goes wrong, instead of staying on the "Thinking" state indefinitely.

Additional information

​- This issue started happening around the past two days.

  • It happens even with simple prompts.
  • Screenshot attached.
  • VS Code version: 1.121.0.
  • ChatGPT/Codex VS Code extension: 26.513.21555.
  • Subscription: ChatGPT Plus.
  • IDE: VS Code.
  • Platform: Windows 11 x64.

View original on GitHub ↗

7 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #23807
  • #24287

Powered by Codex Action

Yong-yuan-X · 1 month ago

I reviewed the potential duplicate issues suggested by the bot.

#23807 looks possibly related because it also mentions ~300s / 5-minute stalls. However, it seems to mainly be about Codex CLI, and it uses a custom provider endpoint. Its logs also show stream disconnected - retrying sampling request. My report is about the VS Code IDE extension UI staying on "Thinking" even for simple prompts.

#24287 has a similar "stuck in Thinking" symptom, but it seems to be about the Codex Desktop app on macOS, and involves Stop/cancel not working and session state desync issues. My environment is the VS Code extension on Windows.

So these issues may be related, but I do not think they are exact duplicates.

Edelweiss-F · 1 month ago

same problem

MisterRound · 1 month ago

Widespread at this point, see my issue https://github.com/openai/codex/issues/24260

AndrewXiGua · 1 month ago

I believe this is a widespread issue. Even when I ask a simple question like '1+1=?', the model takes several minutes of thinking before delivering an answer.

Black-Boom · 1 month ago

same problem

squanch111 · 5 days ago

I am seeing a closely related issue on macOS, with a more specific post-tool continuation failure.

Environment

  • macOS 15.7.4, Apple Silicon
  • VS Code OpenAI/Codex extension: 26.707.31428 (darwin-arm64)
  • Codex CLI: 0.144.0-alpha.4
  • Model: gpt-5.6-sol
  • Approval policy: on-request
  • Sandbox: workspace-write with restricted network

Observed behavior

This has happened repeatedly after network-, MCP-, keychain-, or shell-backed tool calls:

  • A tool call completes, fails, or times out.
  • The terminal/tool output already shows completion or an error.
  • The Codex UI remains on Thinking.
  • No assistant continuation, status summary, or final response is produced.
  • The user must send another message or manually interrupt the turn to regain control.

Examples include:

  • gh auth status waiting on the macOS keychain or GitHub API
  • Web/document fetches returning DNS or TLS errors
  • MCP/web calls returning an ongoing session that requires polling
  • Commands that have already completed while the turn remains open

This is intermittent but recurrent. A user instruction such as “stop or notify me if an operation takes longer than 60 seconds” does not reliably prevent the silent wait.

Expected behavior

  • Tool and network calls should have bounded timeouts.
  • After success, failure, or timeout, Codex should always regain control and produce a final status response.
  • The UI should distinguish between “tool still running”, “waiting for model continuation”, and “composing final response”.
  • Long-running turns should provide a heartbeat and an obvious cancel option.

This appears related to #24434, which describes tool results returning successfully but the assistant continuation never resuming. The difference here is that it reproduces in the VS Code extension on macOS and includes web, MCP, and keychain-backed GitHub CLI operations.

I can provide redacted logs or codex doctor --json output if that would help.