Selected model is at capacity (server_overloaded) at 0% weekly usage on Pro Lite — Sol and Astra

What version of Codex CLI is running?

• PATH CLI: codex-cli 0.150.1
• Codex Desktop: 26.911.61220
• Desktop-bundled session cli_version: 0.155.0-alpha.2.6

What subscription do you have?

ChatGPT Pro Lite (plan_type: prolite), ChatGPT sign-in (no API key).

Which model were you using?

gpt-5.6-sol (high) and gpt-6-astra. Both fail.

What platform is your computer?

Darwin 25.6.0 arm64 arm (macOS Apple Silicon). Timezone Asia/Shanghai.

What terminal emulator and version are you using (if applicable)?

Codex Desktop (originator Codex Desktop / source: vscode). Also reproduced from CLI.

Codex doctor report

Auth/config healthy. ChatGPT inference URL reachable. No stored API key. Model in config: gpt-5.6-sol.

codex doctor --summary (PATH CLI 0.150.1): auth OK, config OK, ChatGPT tokens present, API key absent. WebSocket/CDN warnings in this diagnostic run are from a headless TERM=dumb environment and are not the failure mode in Desktop sessions below (those fail in ~3–8s with server_overloaded).

What issue are you seeing?

Turns fail immediately with:

Selected model is at capacity. Please try a different model.

Local rollout task_complete payload:

{
  "last_agent_message": null,
  "error": {
    "message": "Selected model is at capacity. Please try a different model.",
    "codex_error_info": "server_overloaded"
  }
}

Same turns report remaining quota, not a usage limit:

{
  "plan_type": "prolite",
  "rate_limit_reached_type": null,
  "primary": { "used_percent": 0.0, "window_minutes": 10080 },
  "credits": { "has_credits": false, "balance": "0" }
}

Counted 67 server_overloaded capacity failures on 2026-09-17 (UTC) in local rollouts. Latest: 2026-09-17T12:55:06.525Z.

This is not UsageLimitReached. Switching from Sol to Astra in the same thread still fails.

What steps can reproduce the bug?

1. Sign in with ChatGPT (codex login status → Logged in using ChatGPT).
2. Open Codex Desktop (or CLI) with gpt-5.6-sol.
3. Send a normal prompt in a new or existing thread.
4. Turn ends in a few seconds with the capacity message and codex_error_info: server_overloaded.
5. Switch to gpt-6-astra in the same thread and retry → same error.

Example thread / turns (Desktop, 2026-09-17 UTC):

• Thread: 01a0af61-1b42-7471-8b84-fa0387ba0d2f
• Turns: 01a0af61-1d31-7783-88ee-65b9f4ccb41d (12:39:36Z, Sol), 01a0af68-fbcd-71c3-a99e-3575f921bd28 (12:48:11Z, Sol), 01a0af6f-51fc-7f53-9876-8de3508173ad (12:55:06Z, Astra)

Happy to run /feedback on these thread IDs.

What is the expected behavior?

If weekly Codex allowance is used_percent: 0.0 and rate_limit_reached_type is null, the selected model should serve the turn, or the client should distinguish quota vs capacity vs account-routing and retry/pause instead of terminating the turn.

CodexErr::ServerOverloaded is currently non-retryable in codex-rs/protocol/src/error.rs. Automatic capacity retries in #31058 are not shipped.

Additional information

Related open reports: #43337 #39643 #41810. Official earlier explanation (#17014, #19583): this string is model capacity, not account rate-limit. That does not match a full-day failure at 0% usage across two models while ChatGPT auth remains healthy.

status.openai.com showed fully operational at report time (2026-09-17 ~13:20 UTC). Last Codex/Work incident (https://status.openai.com/incidents/01M2EWYR55J47M2BPG9WC76VEG) resolved 2026-09-14.

Not asking for a local-config workaround. Please correlate backend telemetry for the thread IDs above (admission / routing / capacity pool vs quota).

Start in codex-rs/protocol/src/error.rs, where CodexErr::ServerOverloaded is described as non-retryable, and review the automatic capacity retry work in #31058. Use the reported thread and turn IDs to compare capacity, routing, and quota behavior. Done means the client behavior distinguishes these failure modes and handles the reported capacity case as specified.