VS Code IDE extension repeatedly disconnects and retries streaming requests after GPT-5.6 rollout

Summary

After selecting GPT-5.6 in the Codex VS Code extension, Codex has become substantially slower and frequently reconnects/retries streaming requests. The same workspace, WSL environment, and proxy setup were stable before GPT-5.6.

This is not limited to WSL: a colleague reports the same slow responses and repeated reconnects in VS Code on native Windows.

Environment

• Subscription: ChatGPT Pro
• IDE: Visual Studio Code
• Extension: openai.chatgpt-26.707.71524-linux-x64 (primary reproduction)
• Primary platform: Windows 10 x64 with Ubuntu WSL2 (Linux 6.18.33.2-microsoft-standard-WSL2 x86_64)
• Additional affected environment: native Windows VS Code (colleague; exact extension version not recorded)

Observed behavior

The extension log shows a normal sampling request repeatedly losing its stream:

codex_core::responses_retry: stream disconnected - retrying sampling request (1/5 ...)
...
codex_core::responses_retry: stream disconnected - retrying sampling request (5/5 ...)
codex_core::client: falling back to HTTP

During this loop, the Codex app-server sustained roughly 82% CPU and the VS Code extension host roughly 43% CPU, making the IDE unresponsive.

Short proxy connectivity checks were stable, so this does not look like a simple proxy reachability failure. The Codex desktop app has not shown the same recurring behavior in this setup.

Steps to reproduce

1. Use Codex in the VS Code IDE extension with GPT-5.6 selected.
2. Send a normal coding request, including short requests such as inspecting or reviewing the current workspace.
3. Observe that the extension can stay at "thinking" for a long time, then logs repeated stream-disconnect retries and reconnects before falling back to HTTP.

The issue is intermittent but frequent enough to make the IDE extension unreliable for normal project work.

Expected behavior

Streaming should remain connected or recover without repeated long retries. A normal request should not drive the Codex app-server and VS Code extension host to sustained high CPU.

Additional information

The timing strongly correlates with the GPT-5.6 rollout, but this report does not assume whether the regression is in the model route, the IDE extension, or the backend streaming path. No project files, account data, or raw logs are attached.