Stream disconnected before completion & possible response routing issue (receiving unrelated Polish text) #6398

What version of Codex is running?

codex-cli 0.56.0

What subscription do you have?

Plus

Which model were you using?

gpt-5-codex-mini

What platform is your computer?

Microsoft Windows NT 10.0.21996.0 x64

What issue are you seeing?

After running normally, Codex CLI disconnects the stream before completion and sometimes reconnects to an unrelated output source.
The error appears as:
stream disconnected before completion: An error occurred while processing your request.
You can retry your request, or contact us through our help center at help.openai.com
if the error persists. Please include the request ID 9d75d3d7-1e04-4af0-bbad-9971bf718589 in your message.
In a later attempt (same environment, no prompt change), I received a completely unrelated Polish message:
Przykro mi, ale nie mogę zapewnić doładowania kilometrów. Proszę skontaktować się bezpośrednio z wypożyczalnią w
Starachowicach – pracownicy na pewno pomogą.
(Translation: “I’m sorry, but I can’t provide additional kilometers. Please contact the rental company in Starachowice.”)

This response has no overlap with my input prompt or any previous context, suggesting a possible response routing or stream cross-talk between users/sessions.

What steps can reproduce the bug?

Uploaded thread: 019a6443-b683-7e41-b890-70301f56ee34

What is the expected behavior?

The stream should stay open until a response.completed event.

The model’s output should match the user prompt and remain coherent.

Reconnects (if any) should attach to the correct session or response context.

Additional information

Symptoms strongly indicate response routing / shard mix-up:

Random foreign-language output unrelated to the active prompt right after a disconnect and auto-reconnect.

Possible loss of session stickiness or reconnection to a different backend node.

Potential cross-tenant token leakage due to reconnection without affinity.

What to check internally:

Whether reconnects preserve the original response / session / thread IDs.

Whether the load balancer enforces stream affinity during retries.

If concurrent connections are multiplexed over HTTP/2 without unique per-stream isolation.