Codex CLI repeatedly fails with server_overloaded during long-running tasks despite available Pro quota #53318

What version of Codex CLI is running?

codex-cli 0.162.1

What subscription do you have?

ChatGPT Pro 100

Which model were you using?

gpt-6.1-sol，gpt-6-Astra

What platform is your computer?

Ubuntu Linux (x86_64), running Codex CLI v0.162.1 locally.

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

Ubuntu Linux terminal. The exact terminal emulator and version have not yet been verified.

Codex doctor report

### Codex Doctor Report (Sanitized)

**Environment**
- OS: Ubuntu 22.04 (Jammy), x86_64
- System locale: zh_CN.UTF-8
- Terminal: VTE 6800
- Disk space: 365.3 GiB available

**Codex Installation**
- Codex CLI version: 0.162.1
- Installation: standalone (Linux x86_64 musl)
- Installation consistency: PASS
- Runtime: PASS
- Configuration parsing: PASS
- Codex version: up to date

**Authentication**
- Authentication mode: ChatGPT
- Authentication configured: YES
- API Key authentication: NO
- Actual credentials: REDACTED

**Local State**
- State database integrity: PASS
- Session/thread inventory consistency: PASS
- Rollout scan errors: 0
- Local configuration: loaded successfully
- MCP servers configured: 0

**Network Connectivity**
- Proxy environment variables: none
- WebSocket handshake: SUCCESS (HTTP 101 Switching Protocols)
- ChatGPT inference endpoint: reachable
- HTTP reachability result: 405
- Connection timeout: 15 seconds

**Background Server**
- App-server status: running
- App-server version: 0.162.1
- Server mode: ephemeral

**Overall Doctor Result**

All reported diagnostic categories passed. No installation, configuration, local database integrity, or basic connectivity failures were detected.

However, long-running Codex tasks repeatedly fail with `server_overloaded`, despite successful connectivity checks and available subscription usage.

The diagnostic connectivity test does not rule out intermittent backend model overload.

What issue are you seeing?

Codex CLI repeatedly fails with server_overloaded / Selected model is at capacity. Please try a different model. during long-running, multi-step tasks, despite having sufficient Pro subscription usage available.

Observed behavior:

• Short requests, such as simply replying "OK", usually succeed.
• Longer tasks involving multiple reasoning steps and tool calls frequently fail after approximately 10 minutes.
• The issue occurs with both GPT-6 Astra and GPT-6.1 Sol.
• Switching models, changing working directories, starting new sessions, and forking existing sessions do not reliably prevent the error.
• Updating Codex CLI to v0.162.1 did not resolve the issue.
• After a failure, the task stops and requires manual intervention to resume.

Diagnostic evidence:

I analyzed my local Codex JSONL session logs and identified:

• 77 server_overloaded error records.
• Approximately 25 failure clusters across 4 affected sessions, after grouping nearby errors.
• Errors spanning September 7 through October 12, 2026.
• One forked session generated 36 error records within approximately one second, potentially representing a single underlying failure.

The issue existed before my recent upgrade from ChatGPT Plus to Pro 100.

My subscription and usage limits appear normal, with approximately 99% of weekly usage remaining.

This appears to be a recurring capacity or request-handling issue affecting long-running tasks rather than a usage-quota exhaustion problem. The underlying cause has not been confirmed.

What steps can reproduce the bug?

Uploaded thread: 01a12efc-9bdb-76a2-bf23-c4666301650a

What is the expected behavior?

Codex should be able to complete long-running, multi-step tasks when subscription usage limits are available and the selected model can be served.

If a model temporarily becomes overloaded, Codex should ideally:

• Respect server-provided Retry-After values when available.
• Apply appropriate retry and exponential backoff mechanisms.
• Preserve the current task state and completed tool operations.
• Allow execution to resume reliably without repeating completed work.
• Provide a clearer recovery path rather than repeatedly terminating the task.

Temporary backend overload should not consistently interrupt multi-step workflows or require frequent manual restarts.

Additional information

Environment:

• Operating system: Ubuntu Linux
• Codex CLI version: 0.162.1
• Codex execution: Local CLI
• Subscription: ChatGPT Pro 100
• Weekly usage remaining: approximately 99%
• Affected models: GPT-6 Astra and GPT-6.1 Sol
• Observed reasoning levels: Medium and Xhigh

Affected session IDs:

Original session:
019f63d4-6e57-7663-8864-d401103e8e2b

Forked session:
01a12e16-6fb1-7173-908f-f457575da6ec

Additional affected sessions:
01a12e25-7556-7492-b1c5-7108cb11b380
01a12efc-9bdb-76a2-bf23-c4666301650a

Diagnostic session uploaded through Codex feedback:

Thread ID:
01a12efc-9bdb-76a2-bf23-c4666301650a

The Codex CLI confirmed:

Feedback uploaded.

Log analysis:

• 77 server_overloaded records
• Approximately 25 failure clusters
• 4 affected sessions
• Failure history from September 7 to October 12, 2026

The failure clusters are estimates based on grouping errors occurring within five minutes in the same session, not exact counts of independent failed requests.

HTTP 503 responses and Retry-After headers have not yet been independently verified from local debug logs.

I have also contacted OpenAI Support, which recommended submitting the affected session through Codex's in-product feedback mechanism for backend investigation.

Please investigate whether repeated backend capacity failures, request routing, or retry behavior are responsible for these interruptions.