Problem

Closes #14271

json.JSONDecodeError inherits from ValueError. When the OpenAI SDK fails to parse a provider response (empty body, truncated JSON, garbled bytes), it raises json.JSONDecodeError. The is_local_validation_error check in run_agent.py catches it as isinstance(api_error, ValueError) and triggers a non-retryable abort — but the error is transient (provider-side), not a programming bug.

The error classifier (agent/error_classifier.py) already correctly returns FailoverReason.unknown, retryable=True for these errors. The bug is the inline isinstance check that overrides the classifier.

Secondary bug found

UnicodeDecodeError (also a ValueError subclass) from garbled provider responses gets the same wrong treatment. The existing code only excluded UnicodeEncodeError but a garbled response body can also cause UnicodeDecodeError.

Fix

Updated is_local_validation_error predicate in run_agent.py:

Before (upstream had only):

is_local_validation_error = (
    isinstance(api_error, (ValueError, TypeError))
    and not isinstance(api_error, UnicodeEncodeError)
)

After (this PR, merged with upstream ssl.SSLError fix from #14367):

is_local_validation_error = (
    isinstance(api_error, (ValueError, TypeError))
    and not isinstance(api_error, (UnicodeError, json.JSONDecodeError))
    and not isinstance(api_error, ssl.SSLError)
)

Three exclusions from the ValueError catch:

• UnicodeError (parent class) covers UnicodeEncodeError, UnicodeDecodeError, and UnicodeTranslateError — all indicate provider-side transport issues
• json.JSONDecodeError — malformed/truncated provider response body
• ssl.SSLError — TLS transport failure (inherits ValueError via Python MRO, cherry-picked from upstream ssl.SSLCertVerificationError misclassified as local validation error #14367 / commit 4e27e49)

Testing

9 tests in tests/test_json_decode_error_misclassification.py:

• 5 RED tests (would fail on old code): json.JSONDecodeError, UnicodeDecodeError, UnicodeTranslateError must NOT be classified as local validation errors; classifier must return retryable=True
• 4 GREEN tests (sanity): plain ValueError, TypeError, UnicodeEncodeError, and ssl.SSLError still correctly excluded

COUPLING note in both production and test file — update both if changing the predicate.

Rebase notes

Rebased onto origin/main (commit 4fade39). Conflict in run_agent.py resolved by merging our UnicodeError broadening with upstream's ssl.SSLError exclusion — both are complementary fixes for the same class of bug (ValueError subclasses from provider transport, not local code).

Impact

Low risk, high value. The change only affects the is_local_validation_error boolean — the classifier pipeline and all other recovery paths are untouched. Genuine ValueError/TypeError from coding bugs still trigger non-retryable abort. Only provider-originated ValueError subclasses now correctly fall through to the retryable recovery path.