[Windows][Codex Desktop] Long-running Local workflows intermittently fail with server_overloaded #52833

What version of the Codex App are you using (From “About Codex” dialog)?

Codex Desktop version: 26.1007.21434

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26300.0 x64

What issue are you seeing?

Codex Desktop on Windows intermittently fails during long-running Local workflows with:

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

The local session logs record:

codex_error_info: server_overloaded

The issue does not appear to be a complete model availability problem. Short tasks can succeed, and the same models can work normally in ChatGPT Web Work.

The failure pattern appears related to longer multi-step coding workflows involving extended reasoning and multiple tool calls.

Environment:

• Codex Desktop Windows app
• Version: 26.1007.21434
• Execution mode: Local
• OS: Windows 11 Pro
• Timezone: UTC+08:00

Affected models observed:

• gpt-6.1-sol
• gpt-6-astra
• gpt-6-sol
• gpt-6-luna
• gpt-5.6-sol

Example failure:

Model:

gpt-5.6-sol

Time:

2026-10-10 15:47:10 UTC+08:00

Duration:

898.8 seconds

Result:

codex_error_info: server_overloaded

A local audit of Codex session logs found:

• Completion events: 101
• Errors: 46
• Capacity errors: 46
• Parsing/reading issues: 0

The UI only shows the capacity message and does not provide Retry-After information.

What steps can reproduce the bug?

1. Open Codex Desktop on Windows.
2. Use Local execution mode.
3. Select a supported model such as GPT-6.1 Sol or GPT-6 Astra.
4. Start a multi-step coding task requiring:
   • extended reasoning,
   • multiple tool calls,
   • file inspection/modification,
   • longer execution time.
5. After running successfully for some time, the task may terminate with:

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

Local JSONL session logs record:

codex_error_info: server_overloaded

Representative Turn IDs:

01a124a1-8499-7bf0-a604-c8331d32bca4
01a124a2-17ac-7200-b924-18ed9f078619
01a124a4-cf5e-7902-870f-ae7cf40fa086
01a124a5-0902-77c1-9c79-f8ca4bb2214e
01a124a5-33a8-7080-9cf1-f14ee0b01ec7
01a124ac-9bfc-7d02-8400-601b939ae9ea
01a124ad-5ded-7ba0-a77f-80a4b95b4496
01a124b9-fe1e-7292-9cc8-a9dd9a3f68cb

Additional observations:

• Simple Local tasks with GPT-6.1 Sol can complete successfully.
• ChatGPT Web Work with GPT-6.1 Sol works normally.
• Failures are mainly observed during longer Local workflows.

The in-app feedback command is also unavailable:
typing / opens the command/Skills menu, but there is no "Submit feedback" option.

What is the expected behavior?

Long-running Codex Desktop Local workflows should either:

1. complete successfully, or
2. gracefully recover from temporary capacity limitations by retrying or providing a clear wait/retry mechanism.

Currently, tasks may fail after substantial progress with only:

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

without visible recovery guidance.

Additional information

Related information:

• OpenAI Support confirmed the local logs are consistent with:
  server_overloaded / model at capacity

• I tested:

  • Codex Desktop Local + GPT-6.1 Sol (short task): successful
  • ChatGPT Web Work + GPT-6.1 Sol: successful

This suggests the issue may be related to long-running Local Codex workflow handling, request lifecycle management, or capacity recovery behavior rather than complete model unavailability.

Sanitized session excerpts can be provided if required.

Local session ID:
01a11be1-7194-7e20-9f3c-b3a3757d14d1