Windows desktop: recurring response delays, HTTP 503 retries, and conversation-state errors #53528

Summary

Codex on Windows repeatedly became slow generating responses and executing tools on October 12, 2026. Local logs show an earlier app-server queue stall, later model-service failures, and repeated desktop conversation-state errors. These may be separate problems; this report does not establish a single root cause.

Environment

• Windows desktop app
• Observed Windows app package version: 26.1007.2314.0
• App-server client version observed in logs: 26.1007.21434
• Models observed during short-response tests: GPT-6 Astra and GPT-6.1 Sol, medium reasoning
• All times below are US Eastern (UTC−04:00)

Observed behavior and evidence

1. Before restarting around 9:27 p.m., internal requests repeatedly failed with App server request expired while queued, often after approximately 60 seconds. Requests included chat history, thread lists, configuration, and plugins.
2. Restart cleared those queue-expiry errors. However, short responses still took 9.5–18.3 seconds to finish, while local turn/start request acceptance took only 48–299 ms. The first response-message item event arrived about 5–8 seconds after submission. This is an item-event measurement, not a precise first-token or screen-render measurement.
3. At 10:29 p.m. (02:29 UTC October 13), a request retried after stream disconnected before completion: websocket closed by server before response.completed.
4. At 10:31:56 and 10:33:29 p.m. (02:31:56 and 02:33:29 UTC), requests returned unexpected status 503 Service Unavailable and automatically retried. These failures occurred in other active chats within the same desktop session, rather than the diagnostic chat.
5. Desktop logs repeatedly recorded Conversation state not found, No turns for conversation, and item events for unknown conversations. Their contribution to response latency is unproven.

Short-response measurements

These turns recorded zero reasoning-output tokens. Input context was approximately 78,000–97,000 tokens, mostly cached; the Sol speed check had 83,021 input tokens with no cache hit. Context overhead is a possible contributor, not a demonstrated cause of the service failures.

Diagnostics and workarounds tried

• Restarted Codex; the earlier queued-request failures stopped.
• Removed a Netlify plugin after repeated authentication failures; slow responses recurred.
• Latest resource snapshot during reported slowness: CPU 14%, free RAM 6.1 GB, disk queue 0. Earlier checks also showed ample disk space and usage allowance.
• No renewed app-server queue expirations were observed during the later slowdown.
• OpenAI's status endpoint reported all systems operational at the time of the 503 investigation.

Expected behavior

Reliable response streams and responsive short replies, without repeated queued-request expirations, 503 retries, or conversation-state errors.

Reproduction limitations

The slowdown recurred across active chats, but no deterministic minimal reproduction has been established. HTTP 503 errors and server-closed streams are directly recorded; their relationship to the desktop-state errors remains unknown.

OpenAI's web support assistant directed this report to this issue tracker. No raw logs, credentials, account identifiers, personal paths, or email addresses are attached.