Codex server_is_overloaded stream errors are not retried

[ItsWendell]

Description

OpenAI Codex subscription streams can return transient overload errors in this shape:

{"type":"error","sequence_number":2,"error":{"type":"service_unavailable_error","code":"server_is_overloaded","message":"Our servers are currently overloaded. Please try again later.","param":null}}

OpenCode should treat this as a retryable provider overload. Today this nested stream error shape can miss retry classification because the overload code/type are nested under error.

Related broader reports: #16214 and #21893.

Plugins

None

OpenCode version

Latest dev

Steps to reproduce

1. Use an OpenAI Codex subscription model.
2. Hit a transient streamed overload response with error.code: server_is_overloaded.
3. Observe that the error can terminate instead of entering the existing retry path.

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Some transient stream/rate-limit errors bypass retry and terminate sessions #21893: Covers the broader pattern of transient stream errors (including wrapped rate-limit/concurrency-limit failures) bypassing the retry path — the same code paths in session/retry.ts and provider/error.ts are involved. The server_is_overloaded / service_unavailable_error shape from Codex may be a specific sub-case of what that issue tracks.

If this server_is_overloaded error shape is not yet handled by any fix or PR against #21893, please note that distinction in the existing issue or confirm this is a separate gap.