Repository navigation
Description
Description
When streaming a response from the OpenAI Responses API, only the last url citation survives. The Responses API emits one response.output_text.annotation.added event per annotation, and the stream parser (OpenAIResponses._parse_provider_response_delta) builds a fresh ModelResponse for every event, so each event's model_response.citations holds exactly one annotation. The base stream accumulator then replaces the accumulated citations with each delta's citations (libs/agno/agno/models/base.py:1960), and the final assistant message is populated from that accumulator (libs/agno/agno/models/base.py:1907). With N annotation events, the last one wins and the first N-1 are lost.
The non-streaming parse does not have this problem: it collects every annotation from the response output into one Citations object (libs/agno/agno/models/openai/responses.py:1352-1368), so response() and response_stream() disagree on the same reply.
Steps to Reproduce
import json
import httpx
from agno.agent import Agent
from agno.models.openai.responses import OpenAIResponses
from openai import OpenAI
citations = [
{"type": "url_citation", "url": "https://a.example", "title": "A", "start_index": 0, "end_index": 5},
{"type": "url_citation", "url": "https://b.example", "title": "B", "start_index": 0, "end_index": 5},
{"type": "url_citation", "url": "https://c.example", "title": "C", "start_index": 0, "end_index": 5},
]
response = {
"id": "resp_citations",
"object": "response",
"created_at": 0,
"model": "gpt-4.1-mini",
"status": "completed",
"error": None,
"usage": None,
"output": [
{
"type": "message",
"id": "msg_citations",
"role": "assistant",
"status": "completed",
"content": [{"type": "output_text", "text": "Hello world", "annotations": citations}],
}
],
}
def handle(request):
payload = json.loads(request.content)
if payload.get("stream"):
events = [
{"type": "response.created", "response": response},
{
"type": "response.output_text.delta",
"item_id": "msg_citations",
"output_index": 0,
"content_index": 0,
"delta": "Hello world",
},
]
events.extend(
{
"type": "response.output_text.annotation.added",
"item_id": "msg_citations",
"output_index": 0,
"content_index": 0,
"annotation_index": index,
"sequence_number": index + 2,
"annotation": citation,
}
for index, citation in enumerate(citations)
)
events.append({"type": "response.completed", "response": response})
body = "".join(f"data: {json.dumps(event)}\n\n" for event in events)
return httpx.Response(200, text=body, headers={"content-type": "text/event-stream"})
return httpx.Response(200, json=response)
transport = httpx.MockTransport(handle)
with OpenAI(api_key="test", http_client=httpx.Client(transport=transport)) as client:
model = OpenAIResponses(id="gpt-4.1-mini", client=client)
agent = Agent(model=model)
for event in agent.run(input="cite three sources", stream=True):
pass
print([c.url for c in event.citations.urls])The same asymmetry shows up directly at the Model API: after model.response_stream(messages=messages), messages[-1].citations.urls holds only the last url, while model.response(messages=messages) keeps all three.
Expected Behavior
The streamed run output carries every url citation the model returned: ['https://a.example', 'https://b.example', 'https://c.example'], matching the non-streaming result.
Actual Behavior
The streamed run output only carries the last one: ['https://c.example'].
Screenshots or Logs (if applicable)
# unpatched, streamed:
FINAL_EVENT_CITATION_URLS=['https://c.example']
# non-streaming (same mocked reply):
['https://a.example', 'https://b.example', 'https://c.example']
Environment
- OS: macOS 26 (arm64)
- Agno Version: 3.0.11 (main @ 9f697b2)
- External Dependency Versions: openai 3.19.2
- Additional Environment Details: Python 3.12.14
Possible Solutions (optional)
Accumulate the annotations on the assistant message across events — the same way tool_calls are already accumulated in _parse_provider_response_delta (libs/agno/agno/models/openai/responses.py:1505-1507) — and emit the cumulative Citations on each yielded delta so the accumulator ends with the complete list.
Additional Context
Any setup that reads citations after a streamed OpenAI Responses run (agent run events, run output citations, assistant message citations) is affected.
Activity