Anthropic integration unexpectedly drops `message_stop` events from streamed responses

[john-sungjin]

It seems that the integration modifies the Anthropic stream iterators to ignore events of type message_stop:

sentry-python/sentry_sdk/integrations/anthropic.py

Lines 177 to 211 in 3b28649

	def new_iterator():
	# type: () -> Iterator[MessageStreamEvent]
	input_tokens = 0
	output_tokens = 0
	content_blocks = [] # type: list[str]
	for event in old_iterator:
	input_tokens, output_tokens, content_blocks = _collect_ai_data(
	event, input_tokens, output_tokens, content_blocks
	)
	if event.type != "message_stop":
	yield event
	_add_ai_data_to_span(
	span, integration, input_tokens, output_tokens, content_blocks
	)
	span.__exit__(None, None, None)
	async def new_iterator_async():
	# type: () -> AsyncIterator[MessageStreamEvent]
	input_tokens = 0
	output_tokens = 0
	content_blocks = [] # type: list[str]
	async for event in old_iterator:
	input_tokens, output_tokens, content_blocks = _collect_ai_data(
	event, input_tokens, output_tokens, content_blocks
	)
	if event.type != "message_stop":
	yield event
	_add_ai_data_to_span(
	span, integration, input_tokens, output_tokens, content_blocks
	)
	span.__exit__(None, None, None)

Is this intentional? This is a meaningful event (as indicated in the Anthropic documentation) and I wouldn't expect Sentry to modify this behavior.

[antonpirker]

Hey @john-sungjin , this seems in deed odd. I have asked the author of the integration internally. I will get back to you.

[antonpirker]

We have a PR with a fix merged! This will be released soon!