feat(streaming): emit ReasoningDeltaEvent for reasoning/thinking deltas (#825)

Add a new ReasoningDeltaEvent to StreamEvent so callers can react to
reasoning/thinking tokens in real time without unpacking low-level raw
response events.

The event is emitted whenever a ResponseReasoningSummaryTextDeltaEvent
(o-series extended thinking via the Responses API) or a
ResponseReasoningTextDeltaEvent (third-party models like DeepSeek-R1
via LiteLLM) passes through the stream.  The underlying
RawResponsesStreamEvent is still emitted as well, so nothing breaks for
consumers that already inspect raw events.

Fields:
  delta    - the incremental text fragment from this chunk
  snapshot - full accumulated reasoning text so far in this turn
  type     - always 'reasoning_delta'

Closes #825

Author: Ubuntu

src/agents/__init__.py

@@ -117,6 +117,7 @@
 from .stream_events import (
     AgentUpdatedStreamEvent,
     RawResponsesStreamEvent,
+    ReasoningDeltaEvent,
     RunItemStreamEvent,
     StreamEvent,
 )
@@ -393,6 +394,7 @@ def enable_verbose_stdout_logging():
     "RawResponsesStreamEvent",
     "RunItemStreamEvent",
     "AgentUpdatedStreamEvent",
+    "ReasoningDeltaEvent",
     "StreamEvent",
     "FunctionTool",
     "FunctionToolResult",

src/agents/run_internal/run_loop.py

@@ -12,6 +12,12 @@
 from typing import Any, TypeVar, cast
 
 from openai.types.responses import Response, ResponseCompletedEvent, ResponseOutputItemDoneEvent
+from openai.types.responses.response_reasoning_text_delta_event import (
+    ResponseReasoningTextDeltaEvent,
+)
+from openai.types.responses.response_reasoning_summary_text_delta_event import (
+    ResponseReasoningSummaryTextDeltaEvent,
+)
 from openai.types.responses.response_output_item import McpCall, McpListTools
 from openai.types.responses.response_prompt_param import ResponsePromptParam
 from openai.types.responses.response_reasoning_item import ResponseReasoningItem
@@ -60,6 +66,7 @@
 from ..stream_events import (
     AgentUpdatedStreamEvent,
     RawResponsesStreamEvent,
+    ReasoningDeltaEvent,
     RunItemStreamEvent,
 )
 from ..tool import FunctionTool, Tool, dispose_resolved_computers
@@ -1103,6 +1110,8 @@ async def run_single_turn_streamed(
     emitted_tool_call_ids: set[str] = set()
     emitted_reasoning_item_ids: set[str] = set()
     emitted_tool_search_fingerprints: set[str] = set()
+    # Accumulated reasoning text for ReasoningDeltaEvent snapshot field.
+    _reasoning_snapshot: str = ""
     # Precompute the lookup map used for streaming descriptions. Function tools use the same
     # collision-free lookup keys as runtime dispatch, including deferred top-level aliases.
     tool_map: dict[NamedToolLookupKey, Any] = cast(
@@ -1286,6 +1295,21 @@ async def rewind_model_request() -> None:
     async for event in retry_stream:
         streamed_result._event_queue.put_nowait(RawResponsesStreamEvent(data=event))
 
+        # Emit a ReasoningDeltaEvent for reasoning/thinking deltas so consumers don't have
+        # to unwrap the raw event themselves.
+        if isinstance(event, ResponseReasoningSummaryTextDeltaEvent):
+            delta_text: str = event.delta or ""
+            _reasoning_snapshot += delta_text
+            streamed_result._event_queue.put_nowait(
+                ReasoningDeltaEvent(delta=delta_text, snapshot=_reasoning_snapshot)
+            )
+        elif isinstance(event, ResponseReasoningTextDeltaEvent):
+            delta_text = event.delta or ""
+            _reasoning_snapshot += delta_text
+            streamed_result._event_queue.put_nowait(
+                ReasoningDeltaEvent(delta=delta_text, snapshot=_reasoning_snapshot)
+            )
+
         terminal_response: Response | None = None
         if isinstance(event, ResponseCompletedEvent):
             terminal_response = event.response

src/agents/stream_events.py

@@ -60,5 +60,30 @@ class AgentUpdatedStreamEvent:
     type: Literal["agent_updated_stream_event"] = "agent_updated_stream_event"
 
 
-StreamEvent: TypeAlias = Union[RawResponsesStreamEvent, RunItemStreamEvent, AgentUpdatedStreamEvent]
+@dataclass
+class ReasoningDeltaEvent:
+    """Emitted when a reasoning/thinking delta is received from the model during streaming.
+
+    This is a convenience wrapper over the low-level
+    ``response.reasoning_summary_text.delta`` and ``response.reasoning_text.delta`` raw
+    events.  Both OpenAI o-series reasoning summaries and third-party
+    ``delta.reasoning`` fields (e.g. DeepSeek-R1 via LiteLLM) are surfaced here.
+    """
+
+    delta: str
+    """The incremental reasoning text fragment."""
+
+    snapshot: str
+    """The full reasoning text accumulated so far in this turn."""
+
+    type: Literal["reasoning_delta"] = "reasoning_delta"
+    """The type of the event."""
+
+
+StreamEvent: TypeAlias = Union[
+    RawResponsesStreamEvent,
+    RunItemStreamEvent,
+    AgentUpdatedStreamEvent,
+    ReasoningDeltaEvent,
+]
 """A streaming event from an agent."""

tests/test_reasoning_delta_stream_event.py

@@ -0,0 +1,144 @@
+"""Tests for ReasoningDeltaEvent stream event (issue #825)."""
+
+from __future__ import annotations
+
+import pytest
+
+from agents import Agent, Runner
+from agents.stream_events import ReasoningDeltaEvent, RawResponsesStreamEvent
+
+from openai.types.responses.response_reasoning_item import ResponseReasoningItem, Summary
+
+from .fake_model import FakeModel
+from .test_responses import get_text_message
+
+
+def _make_reasoning_item(text: str) -> ResponseReasoningItem:
+    return ResponseReasoningItem(
+        id="rs_test",
+        type="reasoning",
+        summary=[Summary(text=text, type="summary_text")],
+    )
+
+
+@pytest.mark.asyncio
+async def test_reasoning_delta_event_emitted_during_streaming() -> None:
+    """ReasoningDeltaEvent is emitted when the model streams a reasoning summary delta."""
+    model = FakeModel()
+    model.set_next_output([
+        _make_reasoning_item("Let me think..."),
+        get_text_message("Answer"),
+    ])
+
+    agent = Agent(name="A", model=model)
+    result = Runner.run_streamed(agent, input="hi")
+
+    reasoning_deltas: list[ReasoningDeltaEvent] = []
+    async for event in result.stream_events():
+        if isinstance(event, ReasoningDeltaEvent):
+            reasoning_deltas.append(event)
+
+    assert len(reasoning_deltas) >= 1
+    assert all(isinstance(e.delta, str) for e in reasoning_deltas)
+    assert all(isinstance(e.snapshot, str) for e in reasoning_deltas)
+    assert all(e.type == "reasoning_delta" for e in reasoning_deltas)
+
+
+@pytest.mark.asyncio
+async def test_reasoning_delta_snapshot_accumulates() -> None:
+    """The snapshot field grows monotonically across delta events."""
+    model = FakeModel()
+    model.set_next_output([
+        _make_reasoning_item("Hello world"),
+        get_text_message("done"),
+    ])
+
+    agent = Agent(name="A", model=model)
+    result = Runner.run_streamed(agent, input="hi")
+
+    snapshots: list[str] = []
+    async for event in result.stream_events():
+        if isinstance(event, ReasoningDeltaEvent):
+            snapshots.append(event.snapshot)
+
+    # Each snapshot must be at least as long as the previous one
+    for i in range(1, len(snapshots)):
+        assert len(snapshots[i]) >= len(snapshots[i - 1])
+
+    # Last snapshot must contain the full reasoning text
+    if snapshots:
+        assert "Hello world" in snapshots[-1]
+
+
+@pytest.mark.asyncio
+async def test_no_reasoning_delta_event_without_reasoning() -> None:
+    """ReasoningDeltaEvent is not emitted when there is no reasoning in the response."""
+    model = FakeModel()
+    model.set_next_output([get_text_message("plain text answer")])
+
+    agent = Agent(name="A", model=model)
+    result = Runner.run_streamed(agent, input="hi")
+
+    async for event in result.stream_events():
+        assert not isinstance(event, ReasoningDeltaEvent), (
+            "Got unexpected ReasoningDeltaEvent for a plain text response"
+        )
+
+
+@pytest.mark.asyncio
+async def test_reasoning_delta_event_type_field() -> None:
+    """ReasoningDeltaEvent.type is always 'reasoning_delta'."""
+    model = FakeModel()
+    model.set_next_output([
+        _make_reasoning_item("some reasoning"),
+        get_text_message("answer"),
+    ])
+
+    agent = Agent(name="A", model=model)
+    result = Runner.run_streamed(agent, input="hi")
+
+    async for event in result.stream_events():
+        if isinstance(event, ReasoningDeltaEvent):
+            assert event.type == "reasoning_delta"
+            break
+
+
+@pytest.mark.asyncio
+async def test_raw_response_events_still_emitted_alongside_reasoning_delta() -> None:
+    """RawResponsesStreamEvent is still emitted even when ReasoningDeltaEvent is also emitted."""
+    model = FakeModel()
+    model.set_next_output([
+        _make_reasoning_item("thinking"),
+        get_text_message("result"),
+    ])
+
+    agent = Agent(name="A", model=model)
+    result = Runner.run_streamed(agent, input="hi")
+
+    raw_events: list[RawResponsesStreamEvent] = []
+    reasoning_events: list[ReasoningDeltaEvent] = []
+
+    async for event in result.stream_events():
+        if isinstance(event, RawResponsesStreamEvent):
+            raw_events.append(event)
+        elif isinstance(event, ReasoningDeltaEvent):
+            reasoning_events.append(event)
+
+    # Both types should be present
+    assert len(raw_events) > 0
+    assert len(reasoning_events) > 0
+
+
+@pytest.mark.asyncio
+async def test_reasoning_delta_event_importable_from_agents() -> None:
+    """ReasoningDeltaEvent can be imported directly from the agents package."""
+    from agents import ReasoningDeltaEvent as RDE
+    assert RDE is ReasoningDeltaEvent
+
+
+def test_reasoning_delta_event_dataclass() -> None:
+    """ReasoningDeltaEvent is a proper dataclass with expected fields."""
+    event = ReasoningDeltaEvent(delta="chunk", snapshot="full chunk")
+    assert event.delta == "chunk"
+    assert event.snapshot == "full chunk"
+    assert event.type == "reasoning_delta"