From 20d4a0d5dccb64f20f680ed7adf8417d0b31dcbd Mon Sep 17 00:00:00 2001
From: Muspi Merol <me@promplate.dev>
Date: Mon, 31 Aug 2026 01:44:32 +0800
Subject: [PATCH 1/2] fix(proxy): keep reasoning_items on the streaming
 fast-serialization path

---
 litellm/proxy/proxy_server.py                 |  1 +
 .../proxy_server/test_streaming_helpers.py    | 24 +++++++++++++++++++
 2 files changed, 25 insertions(+)

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index db6294a364d..9e519839ce9 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -9626,6 +9626,7 @@ def _fast_serialize_simple_model_response_stream(
         "annotations",
         "reasoning_content",
         "thinking_blocks",
+        "reasoning_items",
         "provider_specific_fields",
         "refusal",
     )
diff --git a/tests/unit/proxy/proxy_server/test_streaming_helpers.py b/tests/unit/proxy/proxy_server/test_streaming_helpers.py
index 69fa195e9d6..4ae4361cc1a 100644
--- a/tests/unit/proxy/proxy_server/test_streaming_helpers.py
+++ b/tests/unit/proxy/proxy_server/test_streaming_helpers.py
@@ -550,6 +550,16 @@ def test_fast_serialize_simple_model_response_stream_with_usage_returns_none_inv
     assert _fast_serialize_simple_model_response_stream(chunk) is None
 
 
+def test_fast_serialize_simple_model_response_stream_with_reasoning_items_returns_none():
+    """``reasoning_items`` carries the provider's encrypted reasoning state, which
+    the fast path cannot emit — it must bail so the slow path serializes it."""
+    chunk = _simple_chunk(content="")
+    chunk.choices[0].delta.reasoning_items = [
+        {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
+    ]
+    assert _fast_serialize_simple_model_response_stream(chunk) is None
+
+
 # ---------------------------------------------------------------------------
 # _serialize_streaming_chunk
 # ---------------------------------------------------------------------------
@@ -573,6 +583,20 @@ def test_serialize_streaming_chunk_simple_uses_fast_path_bytes():
     }
 
 
+def test_serialize_streaming_chunk_preserves_reasoning_items():
+    """End-to-end pin for the bug: a chunk whose only payload beside text is
+    ``reasoning_items`` must still reach the wire with those items. The fast
+    path emits ``role``/``content`` only, so it has to decline this chunk."""
+    chunk = _simple_chunk(content="")
+    chunk.choices[0].delta.reasoning_items = [
+        {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
+    ]
+    payload = json.loads(_serialize_streaming_chunk(chunk))
+    assert payload["choices"][0]["delta"]["reasoning_items"] == [
+        {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
+    ]
+
+
 def test_serialize_streaming_chunk_invalid_input_raises_attribute_error():
     """The helper is typed as ``BaseModel`` — handing it a plain dict trips
     the attribute-access path (no ``model_dump_json``)."""

From 297d7d50afce82feb42d0a198730a55b405df1cd Mon Sep 17 00:00:00 2001
From: Muspi Merol <me@promplate.dev>
Date: Tue, 20 Oct 2026 15:57:35 +0800
Subject: [PATCH 2/2] test(proxy): assert the complete reasoning items stream
 payload

---
 .../proxy_server/test_streaming_helpers.py    | 26 ++++++++++++-------
 1 file changed, 17 insertions(+), 9 deletions(-)

diff --git a/tests/unit/proxy/proxy_server/test_streaming_helpers.py b/tests/unit/proxy/proxy_server/test_streaming_helpers.py
index 4ae4361cc1a..165f3a14cbe 100644
--- a/tests/unit/proxy/proxy_server/test_streaming_helpers.py
+++ b/tests/unit/proxy/proxy_server/test_streaming_helpers.py
@@ -551,8 +551,6 @@ def test_fast_serialize_simple_model_response_stream_with_usage_returns_none_inv
 
 
 def test_fast_serialize_simple_model_response_stream_with_reasoning_items_returns_none():
-    """``reasoning_items`` carries the provider's encrypted reasoning state, which
-    the fast path cannot emit — it must bail so the slow path serializes it."""
     chunk = _simple_chunk(content="")
     chunk.choices[0].delta.reasoning_items = [
         {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
@@ -584,17 +582,27 @@ def test_serialize_streaming_chunk_simple_uses_fast_path_bytes():
 
 
 def test_serialize_streaming_chunk_preserves_reasoning_items():
-    """End-to-end pin for the bug: a chunk whose only payload beside text is
-    ``reasoning_items`` must still reach the wire with those items. The fast
-    path emits ``role``/``content`` only, so it has to decline this chunk."""
     chunk = _simple_chunk(content="")
     chunk.choices[0].delta.reasoning_items = [
         {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
     ]
-    payload = json.loads(_serialize_streaming_chunk(chunk))
-    assert payload["choices"][0]["delta"]["reasoning_items"] == [
-        {"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}
-    ]
+    payload: Final = json.loads(_serialize_streaming_chunk(chunk))
+    assert payload == {
+        "id": "chatcmpl-test",
+        "created": 0,
+        "model": "gpt-4",
+        "object": "chat.completion.chunk",
+        "choices": [
+            {
+                "index": 0,
+                "delta": {
+                    "role": "assistant",
+                    "content": "",
+                    "reasoning_items": [{"id": "rs_abc", "type": "reasoning", "encrypted_content": "ENCRYPTED-BLOB"}],
+                },
+            }
+        ],
+    }
 
 
 def test_serialize_streaming_chunk_invalid_input_raises_attribute_error():