Summary

classify_error() walks __cause__ / __context__ to extract a nested status code, but _extract_error_body() only inspects the top-level exception. When an SDK/API error is wrapped, Hermes can keep the nested 402 status code but lose the nested body message that distinguishes transient rate limits from billing failures.

Affected files

• agent/error_classifier.py:263-266
• agent/error_classifier.py:774-788

Why this is a bug

At classification time:

• status_code = _extract_status_code(error) walks the cause chain
• body = _extract_error_body(error) does not walk the cause chain

So a wrapped error like this:

• outer exception: Exception("outer")
• __cause__: provider/SDK exception with status_code=402 and body message "Usage limit reached, try again in 5 minutes"

gets classified using:

• status code: 402 (nested cause found)
• message/body: outer / {} (nested body lost)

Minimal reproduction

Wrap a mock API error with:

• status_code = 402
• body = {"error": {"message": "Usage limit reached, try again in 5 minutes"}}

inside an outer exception and pass the outer exception to classify_error().

Expected

Classify as a transient/rate-limit condition, because the nested body explicitly says to retry later.

Actual

Classifies as billing, because the nested body is ignored and only the status code survives.

Suggested investigation

Make _extract_error_body() traverse __cause__ / __context__ the same way _extract_status_code() already does.

A regression test should cover wrapped 402/400 cases where only the nested body contains the decisive message.