Merge d5268a349e into 295d79fc7d

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -1439,16 +1439,17 @@ def _field_of(value: object, name: str) -> object:
     return getattr(value, name, None)
 
 
-def _delta_reasoning_items(delta: object) -> Sequence[object]:
-    items: Final = _field_of(delta, "reasoning_items")
+def _delta_reasoning_items(delta: object, field: str = "reasoning_items") -> Sequence[object]:
+    items: Final = _field_of(delta, field)
     return cast(Sequence[object], items) if isinstance(items, list) else ()  # cast-ok: items forwarded verbatim
 
 
-def _chunk_reasoning_items(chunk: object) -> Sequence[object]:
+def _chunk_reasoning_items(chunk: object, field: str = "reasoning_items") -> Sequence[object]:
     choices: Final = _field_of(chunk, "choices")
     if not isinstance(choices, list) or not choices:
         return ()
-    return _delta_reasoning_items(_field_of(cast(Sequence[object], choices)[0], "delta"))  # cast-ok: untyped chunk
+    delta: Final = _field_of(cast(Sequence[object], choices)[0], "delta")  # cast-ok: untyped chunk
+    return _delta_reasoning_items(delta, field)
 
 
 _NON_TEXT_DELTA_FIELDS: Final = (
@@ -1456,6 +1457,7 @@ _NON_TEXT_DELTA_FIELDS: Final = (
     "function_call",
     "reasoning_content",
     "reasoning_items",
+    "reasoning_details",
     "thinking_blocks",
     "annotations",
     "audio",
@@ -1760,6 +1762,19 @@ def stream_chunk_builder(
                 _choice = response.choices[0]
                 _choice.message.provider_specific_fields = combined_provider_fields
 
+        # Keep every streamed reasoning_details entry in arrival order, as the
+        # non-streaming response does. Do not merge by index: some providers
+        # send whole entries that all have index 0.
+        reasoning_details: Final = tuple(
+            chain.from_iterable(_chunk_reasoning_items(chunk, "reasoning_details") for chunk in typed_chunks)
+        )
+        if reasoning_details:
+            message: Final = response.choices[0].message
+            message.provider_specific_fields = {
+                **(message.provider_specific_fields or {}),
+                "reasoning_details": list(reasoning_details),
+            }
+
         completion_output = get_content_from_model_response(response)
 
         reasoning_tokens: Final = processor.count_reasoning_tokens(response)

tests/unit/litellm_core_utils/test_streaming_chunk_builder_utils.py

@@ -1103,6 +1103,34 @@ def test_stream_chunk_builder_tolerates_trailing_chunk_without_choices():
     assert response.choices[0].message.content == "Hello world"
 
 
+def test_stream_chunk_builder_keeps_reasoning_details_from_every_chunk():
+    """A streamed `reasoning_details` list must rebuild into the same
+    `provider_specific_fields["reasoning_details"]` that a non-streaming
+    response carries: every entry, in arrival order, not merged by `index`
+    (Vercel AI Gateway sends whole entries that all have `index: 0`)."""
+    text_entry: Final = {"type": "reasoning.text", "text": "step one", "index": 0}
+    encrypted_entry: Final = {"type": "reasoning.encrypted", "data": "ENC-TWO", "index": 0}
+    chunks: Final = [
+        ModelResponseStream(
+            model="m",
+            choices=[
+                StreamingChoices(index=0, delta=Delta(role="assistant", content="Hi", reasoning_details=[text_entry]))
+            ],
+        ),
+        ModelResponseStream(
+            model="m", choices=[StreamingChoices(index=0, delta=Delta(reasoning_details=[encrypted_entry]))]
+        ),
+        ModelResponseStream(model="m", choices=[StreamingChoices(index=0, delta=Delta(content=" there"))]),
+    ]
+
+    response = stream_chunk_builder(chunks=chunks)
+
+    assert response is not None
+    message = response.choices[0].message
+    assert message.content == "Hi there"
+    assert message.provider_specific_fields == {"reasoning_details": [text_entry, encrypted_entry]}
+
+
 def test_anthropic_speed_and_geo_survive_stream_assembly():
     """Anthropic prices fast mode and non-global regions with a multiplier read off
     ``usage.speed`` / ``usage.inference_geo``. Dropping them while reassembling a stream