Network 503 api proxy

Codex unexpected status 503 Service Unavailable

A 503 from chatgpt.com/backend-api/codex/responses with biscuit_baker_service_me_circuit_open points to temporary Codex backend degradation. Check OpenAI Status before changing local settings.

Snapshot

This 503 is usually server-side, not a bad API key. Pause repeated retries and troubleshoot locally only if OpenAI reports no incident.

#codex unexpected status 503 service unavailable #biscuit_baker_service_me_circuit_open #codex auth error 503 #error running remote compact task #codex service temporarily unavailable #codex websocket 503 #codex falling back from websockets to https

Codex reports:

Error running remote compact task: unexpected status 503 Service Unavailable: Service Unavailable, url: https://chatgpt.com/backend-api/codex/responses, cf-ray: ..., auth error: 503, auth error code: biscuit_baker_service_me_circuit_open

What It Means

chatgpt.com/backend-api/codex/responses is the official ChatGPT-authenticated Codex backend. HTTP 503 and circuit_open mean that backend is temporarily unavailable; this is not the usual invalid-credential error.

Was This an OpenAI Outage?

Yes. OpenAI reported elevated error rates affecting APIs, ChatGPT, and Codex. GitHub issues #35322 and #35323 recorded the same 503 and WebSocket fallback during that incident.

How To Fix It

  1. Check OpenAI Status and stop repeated retries during an active incident.
  2. Save current work, then restart Codex after recovery.
  3. If status is clear and the 503 persists, test another network and update Codex.

Do not delete ~/.codex, revoke credentials, or reinstall Codex during a confirmed outage.

Related errors

Claude Code / Codex 503 No available accounts Do not rotate keys first. This error usually points to an unavailable account pool, provider, group, or model route. Test a minimal request, confirm the model and group, then decide whether to wait, reduce context, change model, or switch provider. Codex: Selected model is at capacity. Error: Selected model is at capacity. Please try a different model. Treat it as model-level capacity, admission, or streaming interruption first. It can happen with ChatGPT-authenticated Codex CLI, proxy account pools, or Sub2API upstream overloaded passthrough. Codex warning: Your conversation has multiple possible cybersecurity risk flags Warning: Your conversation has multiple possible cybersecurity risk flags. Treat it as an OpenAI/Codex safety-routing signal first, then check whether an API proxy, sensitive keywords, long code context, or opaque model routing is making it worse. Error 429 Too Many Requests for AI API proxies Error: exceeded retry limit, last status: 429 Too Many Requests, request id: <request-id>. With an AI API proxy, repeated 429s usually mean the upstream account pool is rate limited, cooling down, or out of usable quota. Ask the provider to switch account, route, or model first. Error 401 Unauthorized for AI API proxies A 401 is an authentication failure. Check which key is loaded, which Base URL receives the request, whether the header format is correct, and whether the key is still active for that provider.

FAQ

Common questions

Does auth error: 503 mean my ChatGPT login expired?

Usually not. Invalid credentials normally return 401, 403, or a sign-in prompt. A 503 with circuit_open indicates backend unavailability.

Why does Codex fall back from WebSockets to HTTPS and still fail?

Both transports reach the same backend, so HTTPS can return the same 503 after the WebSocket fails.

Can a failed remote compact task damage my thread?

It means context compaction failed; it does not show that the repository or local thread data was corrupted.