[Escalation] Codex Desktop repeatedly disconnects on long-running sessions (5/5 retries); request engineering ownership review #52867

Formal reliability complaint and request for engineering escalation

Codex Desktop is repeatedly failing during a long-running development project. The user has already tried troubleshooting, and the disruption is preventing normal project completion. Please escalate this to engineering management for review of issue ownership, remediation priority, and the effectiveness of the current fix process.

Reproduction context

• Windows Codex Desktop (exact version not yet collected)
• ASUS Merlin router with Shadowsocks proxy
• Other Codex conversations on the same connection work normally
• One existing long-running development conversation is disproportionately affected; keeping that conversation and its history intact is a requirement

Observed failures

• stream disconnected before completion: Transport error: network error: error decoding response body
• Occasional unexpected TLS close / EOF
• Automatic retries progress to 5/5 and may still fail
• At other times the session resumes and continues work, then disconnects again at unpredictable points during normal execution, not only during visible context compaction

Impact

Repeated failures are blocking substantial, long-running software development. Asking the user to start a new conversation sacrifices project continuity and is not an adequate remedy. Repeated manual retries also risk duplicated or partially completed steps.

Requested action

1. Investigate streaming transport, session resumption, and long-context processing for existing conversations, including how failures are recovered after all retries are exhausted.
2. Preserve the original conversation and project state; provide a supported in-place recovery solution.
3. Escalate recurring reports to engineering management, review ownership/staffing or bring in additional expertise if needed, and communicate a concrete progress update.
4. Provide a way to capture useful, privacy-safe diagnostic logs and correlation identifiers without requiring users to delete conversation history.

This report does not assume the root cause is proven: the long session is correlated with the failures, while router/network and server transport also remain possibilities.

Please avoid closing this solely with generic network-troubleshooting advice without reviewing the session-specific failure pattern.