OpenAI-compatible stream_read_error bypasses automatic retry

[JianJia2018]

Summary

A structured OpenAI-compatible stream failure with type: upstream_error and code/message: stream_read_error is surfaced immediately instead of entering Jcode's bounded transient-error retry loop.

Version

• Jcode v0.75.0 (5ae238574)
• Also present in the current master source at the same commit

Error shape

{
  "type": "error",
  "sequence_number": 0,
  "error": {
    "type": "upstream_error",
    "code": "stream_read_error",
    "message": "stream_read_error"
  }
}

Actual behavior

The turn terminates with an error and requires a manual retry.

The structured error is formatted as upstream_error: stream_read_error. is_retryable_error() delegates to the shared transport classifier, which recognizes strings such as stream error, error reading, connection failures, and timeouts, but not the underscore-form identifier stream_read_error.

The resulting StreamEvent::Error is therefore forwarded to the agent layer, where it is returned immediately as StreamError unless it is a context-limit error.

Relevant code:

• extract_error_with_retry
• is_retryable_error
• is_transient_transport_error
• Agent stream error handling

Expected behavior

stream_read_error should be treated as a transient stream/transport failure and use the existing bounded retry behavior, including rollback of any partial output before replay.

Suggested minimal fix

Add stream_read_error to the shared transient transport classifier and add a regression test covering the structured upstream_error payload. The existing OpenAI retry loop already limits transient failures to three attempts.

[1jehuang]

Fixed in PR #888 by classifying stream_read_error as a transient transport failure, which routes it through the existing bounded retry loop. The focused transport suite passes 7/7.

---

— Jcode agent (automated triage), on behalf of @1jehuang

[1jehuang]

The full transport-core suite passes, but the real upstream stream_read_error condition cannot be forced locally. If you can test PR #888 with the affected OpenAI-compatible endpoint, confirmation that the turn retries automatically would close the live acceptance gap.

---

— Jcode agent (automated triage), on behalf of @1jehuang