phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-08-15 14:08:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat(responses): add MCP argument streaming and content part events (#3136)

Browse source 
# What does this PR do?

Adds content part streaming events to the OpenAI-compatible Responses API to support more granular streaming of response content. This introduces:

1. New schema types for content parts: `OpenAIResponseContentPart` with variants for text output and refusals

2. New streaming event types:
   - `OpenAIResponseObjectStreamResponseContentPartAdded` for when content parts begin
   - `OpenAIResponseObjectStreamResponseContentPartDone` for when content parts complete

3. Implementation in the reference provider to emit these events during streaming responses. Also emits MCP arguments just like function call ones.


## Test Plan

Updated existing streaming tests to verify content part events are properly emitted
This commit is contained in:
Ashwin Bharambe 2025-08-13 16:34:26 -07:00 • committed by GitHub
parent 8638537d14
commit e1e161553c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 480 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

77
tests/integration/non_ci/responses/test_responses.py
View file

@ -590,9 +590,17 @@ def test_response_streaming_multi_turn_tool_execution(compat_client, text_model_
# Verify tool call streaming events are present
chunk_types = [chunk.type for chunk in chunks]
# Should have function call arguments delta events for tool calls
delta_events = [chunk for chunk in chunks if chunk.type == "response.function_call_arguments.delta"]
done_events = [chunk for chunk in chunks if chunk.type == "response.function_call_arguments.done"]
# Should have function call or MCP arguments delta/done events for tool calls
delta_events = [
chunk
for chunk in chunks
if chunk.type in ["response.function_call_arguments.delta", "response.mcp_call.arguments.delta"]
]
done_events = [
chunk
for chunk in chunks
if chunk.type in ["response.function_call_arguments.done", "response.mcp_call.arguments.done"]
]
# Should have output item events for tool calls
item_added_events = [chunk for chunk in chunks if chunk.type == "response.output_item.added"]
@ -606,8 +614,12 @@ def test_response_streaming_multi_turn_tool_execution(compat_client, text_model_
assert len(chunks) > 10, f"Expected rich streaming with many events, got only {len(chunks)} chunks"
# Since this test involves MCP tool calls, we should see streaming events
assert len(delta_events) > 0, f"Expected function_call_arguments.delta events, got chunk types: {chunk_types}"
assert len(done_events) > 0, f"Expected function_call_arguments.done events, got chunk types: {chunk_types}"
assert len(delta_events) > 0, (
f"Expected function_call_arguments.delta or mcp_call.arguments.delta events, got chunk types: {chunk_types}"
)
assert len(done_events) > 0, (
f"Expected function_call_arguments.done or mcp_call.arguments.done events, got chunk types: {chunk_types}"
)
# Should have output item events for function calls
assert len(item_added_events) > 0, f"Expected response.output_item.added events, got chunk types: {chunk_types}"
@ -670,22 +682,32 @@ def test_response_streaming_multi_turn_tool_execution(compat_client, text_model_
assert isinstance(done_event.output_index, int), "Output index should be integer"
assert done_event.output_index >= 0, "Output index should be non-negative"
# Group function call argument events by item_id (these should have proper tracking)
function_call_events_by_item_id = {}
# Group function call and MCP argument events by item_id (these should have proper tracking)
argument_events_by_item_id = {}
for chunk in chunks:
if hasattr(chunk, "item_id") and chunk.type in [
"response.function_call_arguments.delta",
"response.function_call_arguments.done",
"response.mcp_call.arguments.delta",
"response.mcp_call.arguments.done",
]:
item_id = chunk.item_id
if item_id not in function_call_events_by_item_id:
function_call_events_by_item_id[item_id] = []
function_call_events_by_item_id[item_id].append(chunk)
if item_id not in argument_events_by_item_id:
argument_events_by_item_id[item_id] = []
argument_events_by_item_id[item_id].append(chunk)
for item_id, related_events in function_call_events_by_item_id.items():
# Should have at least one delta and one done event for a complete function call
delta_events = [e for e in related_events if e.type == "response.function_call_arguments.delta"]
done_events = [e for e in related_events if e.type == "response.function_call_arguments.done"]
for item_id, related_events in argument_events_by_item_id.items():
# Should have at least one delta and one done event for a complete tool call
delta_events = [
e
for e in related_events
if e.type in ["response.function_call_arguments.delta", "response.mcp_call.arguments.delta"]
]
done_events = [
e
for e in related_events
if e.type in ["response.function_call_arguments.done", "response.mcp_call.arguments.done"]
]
assert len(delta_events) > 0, f"Item {item_id} should have at least one delta event"
assert len(done_events) == 1, f"Item {item_id} should have exactly one done event"
@ -694,6 +716,33 @@ def test_response_streaming_multi_turn_tool_execution(compat_client, text_model_
for event in related_events:
assert event.item_id == item_id, f"Event should have consistent item_id {item_id}, got {event.item_id}"
# Verify content part events if they exist (for text streaming)
content_part_added_events = [chunk for chunk in chunks if chunk.type == "response.content_part.added"]
content_part_done_events = [chunk for chunk in chunks if chunk.type == "response.content_part.done"]
# Content part events should be paired (if any exist)
if len(content_part_added_events) > 0:
assert len(content_part_done_events) > 0, (
"Should have content_part.done events if content_part.added events exist"
)
# Verify content part event structure
for added_event in content_part_added_events:
assert hasattr(added_event, "response_id"), "Content part added event should have response_id"
assert hasattr(added_event, "item_id"), "Content part added event should have item_id"
assert hasattr(added_event, "part"), "Content part added event should have part"
# TODO: enable this after the client types are updated
# assert added_event.part.type == "output_text", "Content part should be an output_text"
for done_event in content_part_done_events:
assert hasattr(done_event, "response_id"), "Content part done event should have response_id"
assert hasattr(done_event, "item_id"), "Content part done event should have item_id"
assert hasattr(done_event, "part"), "Content part done event should have part"
# TODO: enable this after the client types are updated
# assert len(done_event.part.text) > 0, "Content part should have text when done"
# Basic pairing check: each output_item.added should be followed by some activity
# (but we can't enforce strict 1:1 pairing due to the complexity of multi-turn scenarios)
assert len(item_added_events) > 0, "Should have at least one output_item.added event"

36
tests/unit/providers/agents/meta_reference/test_openai_responses.py
View file

@ -136,9 +136,12 @@ async def test_create_openai_response_with_string_input(openai_responses_impl, m
input=input_text,
model=model,
temperature=0.1,
stream=True, # Enable streaming to test content part events
)
# Verify
# For streaming response, collect all chunks
chunks = [chunk async for chunk in result]
mock_inference_api.openai_chat_completion.assert_called_once_with(
model=model,
messages=[OpenAIUserMessageParam(role="user", content="What is the capital of Ireland?", name=None)],
@ -147,11 +150,32 @@ async def test_create_openai_response_with_string_input(openai_responses_impl, m
stream=True,
temperature=0.1,
)
# Should have content part events for text streaming
# Expected: response.created, content_part.added, output_text.delta, content_part.done, response.completed
assert len(chunks) >= 4
assert chunks[0].type == "response.created"
# Check for content part events
content_part_added_events = [c for c in chunks if c.type == "response.content_part.added"]
content_part_done_events = [c for c in chunks if c.type == "response.content_part.done"]
text_delta_events = [c for c in chunks if c.type == "response.output_text.delta"]
assert len(content_part_added_events) >= 1, "Should have content_part.added event for text"
assert len(content_part_done_events) >= 1, "Should have content_part.done event for text"
assert len(text_delta_events) >= 1, "Should have text delta events"
# Verify final event is completion
assert chunks[-1].type == "response.completed"
# When streaming, the final response is in the last chunk
final_response = chunks[-1].response
assert final_response.model == model
assert len(final_response.output) == 1
assert isinstance(final_response.output[0], OpenAIResponseMessage)
openai_responses_impl.responses_store.store_response_object.assert_called_once()
assert result.model == model
assert len(result.output) == 1
assert isinstance(result.output[0], OpenAIResponseMessage)
assert result.output[0].content[0].text == "Dublin"
assert final_response.output[0].content[0].text == "Dublin"
async def test_create_openai_response_with_string_input_with_tools(openai_responses_impl, mock_inference_api):
@ -272,6 +296,8 @@ async def test_create_openai_response_with_tool_call_type_none(openai_responses_
# Check that we got the content from our mocked tool execution result
chunks = [chunk async for chunk in result]
# Verify event types
# Should have: response.created, output_item.added, function_call_arguments.delta,
# function_call_arguments.done, output_item.done, response.completed
assert len(chunks) == 6