Our servers are currently overloaded. Please try again later.

[Cavan-leo]

[ldj18273656778-cmd]

I have also met this problem.

[m00tox]

I reproduced this on the first-party openai-codex route. It is not retry-budget exhaustion: the error is flattened to PI_AI_ERROR, so the normal retry policy never schedules even one llm/retry event.

I added the event evidence and a narrow upstream classification proposal to the existing technical report: #3407 (comment)

The important extension is that first-party Codex emits Our servers are currently overloaded without an HTTP status and without a server_error token. The upstream match therefore needs explicit currently/temporarily overloaded wording in addition to server_error, mapping it to SERVER while leaving unmatched PI_AI_ERROR terminal.