test(responses): cover message-item dedup suppression and text merge

Extend coverage for the #41109 duplicate message-item suppression
(_parse_event_chunk / _consume_pending_message_buffer) and the
response_format/verbosity text merge (_merge_text):
- a verbatim repeat of already-streamed text in a later message item is
  dropped when the item completes
- new content in a later message item is flushed on the terminal event
- unrelated items completing do not clear the pending message buffer
- _merge_text combines existing text params with the new update

Signed-off-by: rain <1504569896@qq.com>
This commit is contained in:
rain 2026-09-20 22:51:19 +00:00 • committed by Rainmemery
parent 2262bd0ace
commit 4e62076053

View file

@ -4523,3 +4523,105 @@ def test_chunk_parser_falls_back_to_static_translation_for_unsupported_chunks():
iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
with pytest.raises(ValueError):
iterator.chunk_parser("not a structured chunk")
# ------------------------------------------------------------------------
# #41109/#41117: duplicate message-item suppression and text merging
# ------------------------------------------------------------------------
def test_merge_text_merges_into_existing_text():
from litellm.completion_extras.litellm_responses_transformation.transformation import (
LiteLLMResponsesTransformationHandler,
)
merged = LiteLLMResponsesTransformationHandler._merge_text(
{"text": {"format": {"type": "text"}}}, {"verbosity": "high"}
)
assert merged == {"format": {"type": "text"}, "verbosity": "high"}
merged2 = LiteLLMResponsesTransformationHandler._merge_text({}, {"verbosity": "low"})
assert merged2 == {"verbosity": "low"}
def test_map_optional_params_merges_response_format_and_verbosity_into_text():
from litellm.completion_extras.litellm_responses_transformation.transformation import (
LiteLLMResponsesTransformationHandler,
)
handler = LiteLLMResponsesTransformationHandler()
request: dict = {}
handler._map_optional_params_to_responses_api_request(
{"response_format": {"type": "json_object"}, "verbosity": "high"}, request
)
assert request["text"]["format"]["type"] == "json_object"
assert request["text"]["verbosity"] == "high"
def test_duplicate_message_item_verbatim_repeat_is_dropped():
from litellm.completion_extras.litellm_responses_transformation.transformation import (
OpenAiResponsesToChatCompletionStreamIterator,
)
iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
# first message item streams its text
iterator.chunk_parser({"type": "response.output_text.delta", "delta": "Hello"})
# a later message item starts (gpt-5.4+ multi-message turn, #37299)
iterator.chunk_parser(
{"type": "response.output_item.added", "item": {"type": "message", "id": "msg_2"}}
)
# deltas are buffered while the item is pending -> empty chunk
buffered = iterator.chunk_parser({"type": "response.output_text.delta", "delta": "Hello"})
assert buffered.choices[0].delta.content == ""
# item completes with only a verbatim repeat of the already-streamed text -> dropped
done = iterator.chunk_parser(
{"type": "response.output_item.done", "item": {"type": "message", "id": "msg_2", "content": []}}
)
assert done.choices[0].delta.content != "Hello"
def test_duplicate_message_item_new_content_is_flushed():
from litellm.completion_extras.litellm_responses_transformation.transformation import (
OpenAiResponsesToChatCompletionStreamIterator,
)
iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
iterator.chunk_parser({"type": "response.output_text.delta", "delta": "Hello"})
iterator.chunk_parser(
{"type": "response.output_item.added", "item": {"type": "message", "id": "msg_3"}}
)
iterator.chunk_parser({"type": "response.output_text.delta", "delta": " World"})
# item completes with new content: buffered text is flushed on the terminal event
done = iterator.chunk_parser(
{"type": "response.output_item.done", "item": {"type": "message", "id": "msg_3", "content": []}}
)
assert done.choices[0].delta.content == " World"
def test_pending_message_buffer_ignores_unrelated_items_and_flushes_on_completed():
from litellm.completion_extras.litellm_responses_transformation.transformation import (
OpenAiResponsesToChatCompletionStreamIterator,
)
iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
iterator.chunk_parser({"type": "response.output_text.delta", "delta": "A"})
iterator.chunk_parser(
{"type": "response.output_item.added", "item": {"type": "message", "id": "m"}}
)
iterator.chunk_parser({"type": "response.output_text.delta", "delta": "B"})
# an unrelated item completing leaves the message buffer alone
iterator.chunk_parser(
{"type": "response.output_item.done", "item": {"type": "function_call", "id": "fc_1"}}
)
# the terminal event flushes the buffered new content
done = iterator.chunk_parser({"type": "response.completed"})
assert done.choices[0].delta.content == "B"