The new _wrap_synchronous_message_iterator (line 393-452) and _wrap_asynchronous_message_iterator (line 455-513) call _collect_ai_data but only unpack 3 values (model, usage, content_blocks). However, _collect_ai_data (line 149) returns a 4-tuple including response_id. This will raise ValueError: too many values to unpack when the code executes. The existing implementations at lines 516+ correctly unpack all 4 values.

Also found at:

• sentry_sdk/integrations/anthropic.py:393-452
• sentry_sdk/integrations/anthropic.py:442-452

In _wrap_message_stream_manager_enter, span.__enter__() is called at line 912 but there's no try/except/finally block to ensure the span is properly closed if an exception occurs in subsequent operations (lines 914-932). If _set_common_input_data() or the stream._iterator assignment raises an exception, the span will remain open, causing a potential resource leak. Other wrapper functions in this file (e.g., _sentry_patched_create_sync) use a finally block to close the span on error.

The new test_stream_messages test does not assert that SPANDATA.GEN_AI_SYSTEM == "anthropic" is set on the span, unlike the similar test_streaming_create_message test (line 306). The implementation sets this via _set_common_input_data, but the test doesn't verify it, which could mask regressions.

Also found at:

• tests/integrations/anthropic/test_anthropic.py:439