Merge 34f5c5c8bf into 8d6b8dab3d

litellm/llms/anthropic/pass_through/adapters/transformation.py

@@ -1692,6 +1692,9 @@ class LiteLLMAnthropicMessagesAdapter:
                         assert isinstance(thinking, str)
                         assert isinstance(signature, str)
 
+                        # Without a signature the same text arrives again as a
+                        # thinking_delta, so an SSE accumulator seeding from this
+                        # start body would show it twice.
                         return "thinking", ChatCompletionThinkingBlock(
                             type="thinking", thinking=thinking if signature else "", signature=signature
                         )

tests/unit/llms/anthropic/pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -288,6 +288,7 @@ def test_translate_streaming_openai_chunk_to_anthropic_thinking_content_block():
     ) = LiteLLMAnthropicMessagesAdapter().translate_streaming_openai_chunk_to_anthropic_content_block(choices=choices)
 
     assert block_type == "thinking"
+    # Unsigned: the text is delivered by the thinking_delta, so the start body stays empty.
     assert content_block_start == {
         "type": "thinking",
         "thinking": "",

tests/unit/llms/anthropic/pass_through/adapters/test_streaming_iterator_first_delta.py

@@ -948,6 +948,27 @@ def test_mixed_thinking_start_body_is_empty_and_thinking_not_doubled_sync():
     assert _text_deltas(events) == ["Answer."]
 
 
+def test_pure_thinking_start_body_is_empty_and_thinking_not_doubled_sync():
+    chunks = [
+        _thinking_chunk("Hello"),
+        _thinking_chunk(" world"),
+        _make_chunk(Delta(content="answer")),
+        _make_chunk(Delta(content=None), finish_reason="stop"),
+    ]
+    wrapper = AnthropicStreamWrapper(completion_stream=iter(chunks), model="claude-x")
+    events = _drain_sync(wrapper)
+
+    accumulated = ""
+    for event in events:
+        if event.get("type") == "content_block_start" and event["content_block"].get("type") == "thinking":
+            assert not event["content_block"].get("thinking"), event["content_block"]
+            accumulated += event["content_block"].get("thinking") or ""
+        if event.get("type") == "content_block_delta" and event["delta"].get("type") == "thinking_delta":
+            accumulated += event["delta"]["thinking"]
+    assert accumulated == "Hello world"
+    assert _text_deltas(events) == ["answer"]
+
+
 def test_mixed_chunk_with_tool_argument_continuation_is_not_split_sync():
     """Streaming providers send a tool call's name only on its first chunk;
     later chunks carry argument fragments with ``name=None``. Splitting a