[Bug]: 100% stream disconnected before completion on specific account after app-server debugging

Bug Description

Our Codex CLI (@openai/codex v0.154.0 on Windows 11) is experiencing 100% persistent stream disconnections on account 1324940197@qq.com across all models (gpt-6-astra, gpt-5.6-terra) for over 24 hours.

Every single prompt—even starting a clean thread and typing hi or test—immediately fails after 5 WebSocket reconnection attempts and HTTPS fallback with:
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 <id> in your message.

Triggering Scenario

Before this issue began, we were developing and testing local client integration with the app-server daemon / WebSocket RPC endpoints. It appears these debugging requests triggered backend rate limiting, anti-abuse protection, or hung stream state locking on this specific user identity.

Critical Control Group Evidence

• Same physical machine, same network, same IP, same Codex binary: Switching authentication to another account (chaos.realms.ai@gmail.com) works flawlessly with zero stream errors.
• Switching back to 1324940197@qq.com fails 100% on every single invocation.
• codex doctor confirms local databases, config, proxy, and WebSocket handshakes (HTTP 101 Switching Protocols) are all healthy. The failure happens exclusively during backend token stream generation.

Request IDs Across Sessions (Last 24h)

• 3f5a3d8b-26f6-405e-a70a-3b27bde5e939
• 321df09e-c5df-4651-8d9c-0dfa64f35b4e
• 74089af2-87bf-43fc-8dd0-95dd76cee5ee
• b73c3e93-0f25-4010-923b-d6413c5a9302
• ef159bd8-e40f-458b-b9aa-699f5b4f121c
• 8859fc81-f66d-4aa2-bc19-31c341c0a14e

Could the team please check why stream generation aborts for this account, and advise if this account-level state can be reset?

Start with the app-server daemon and its WebSocket RPC endpoints, using codex doctor and the provided request IDs while comparing the failing account with the working account. Done means identifying a repository-owned cause and reproducible fix; if the failure is account-level backend state, document that boundary and the required reset path.