Return signature on anthropic streaming + migrate to signature field instead of signature_delta [MINOR bump] (#9021)

* Fix missing signature_delta in thinking blocks when streaming from Claude 3.7 (#8797)

Co-authored-by: Krish Dholakia <krrishdholakia@gmail.com>

* test: update test to enforce signature found

* feat(refactor-signature-param-to-be-'signature'-instead-of-'signature_delta'): keeps it in sync with anthropic

* fix: fix linting error

---------

Co-authored-by: Martin Krasser <krasserm@googlemail.com>

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -2983,7 +2983,7 @@ class BedrockConverseMessagesProcessor:
         reasoning_content_blocks: List[BedrockContentBlock] = []
         for thinking_block in thinking_blocks:
             reasoning_text = thinking_block.get("thinking")
-            reasoning_signature = thinking_block.get("signature_delta")
+            reasoning_signature = thinking_block.get("signature")
             text_block = BedrockConverseReasoningTextBlock(
                 text=reasoning_text or "",
             )

litellm/llms/anthropic/chat/handler.py

@@ -527,6 +527,7 @@ class ModelResponseIterator:
         provider_specific_fields = {}
         content_block = ContentBlockDelta(**chunk)  # type: ignore
         thinking_blocks: List[ChatCompletionThinkingBlock] = []
+
         self.content_blocks.append(content_block)
         if "text" in content_block["delta"]:
             text = content_block["delta"]["text"]
@@ -544,13 +545,13 @@ class ModelResponseIterator:
             provider_specific_fields["citation"] = content_block["delta"]["citation"]
         elif (
             "thinking" in content_block["delta"]
-            or "signature_delta" == content_block["delta"]
+            or "signature" in content_block["delta"]
         ):
             thinking_blocks = [
                 ChatCompletionThinkingBlock(
                     type="thinking",
-                    thinking=content_block["delta"].get("thinking"),
+                    thinking=content_block["delta"].get("thinking") or "",
-                    signature_delta=content_block["delta"].get("signature"),
+                    signature=content_block["delta"].get("signature"),
                 )
             ]
             provider_specific_fields["thinking_blocks"] = thinking_blocks

litellm/llms/bedrock/chat/converse_transformation.py

@@ -272,7 +272,7 @@ class AmazonConverseConfig(BaseConfig):
                 optional_params["temperature"] = value
             if param == "top_p":
                 optional_params["topP"] = value
-            if param == "tools":
+            if param == "tools" and isinstance(value, list):
                 optional_params = self._add_tools_to_optional_params(
                     optional_params=optional_params, tools=value
                 )
@@ -598,7 +598,7 @@ class AmazonConverseConfig(BaseConfig):
                 if _text is not None:
                     _thinking_block["thinking"] = _text
                 if _signature is not None:
-                    _thinking_block["signature_delta"] = _signature
+                    _thinking_block["signature"] = _signature
                 thinking_blocks_list.append(_thinking_block)
         return thinking_blocks_list
 

litellm/types/llms/openai.py

@@ -360,7 +360,7 @@ class ChatCompletionCachedContent(TypedDict):
 class ChatCompletionThinkingBlock(TypedDict, total=False):
     type: Required[Literal["thinking"]]
     thinking: str
-    signature_delta: str
+    signature: str
     cache_control: Optional[Union[dict, ChatCompletionCachedContent]]
 
 

tests/llm_translation/test_anthropic_completion.py

@@ -1188,18 +1188,20 @@ def test_anthropic_thinking_output(model):
     assert isinstance(resp.choices[0].message.thinking_blocks, list)
     assert len(resp.choices[0].message.thinking_blocks) > 0
 
+    assert resp.choices[0].message.thinking_blocks[0]["signature"] is not None
+
 
 @pytest.mark.parametrize(
     "model",
     [
         "anthropic/claude-3-7-sonnet-20250219",
-        "bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0",
+        # "bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0",
     ],
 )
 def test_anthropic_thinking_output_stream(model):
     # litellm.set_verbose = True
     try:
-        litellm._turn_on_debug()
+        # litellm._turn_on_debug()
         resp = litellm.completion(
             model=model,
             messages=[{"role": "user", "content": "Tell me a joke."}],
@@ -1209,6 +1211,7 @@ def test_anthropic_thinking_output_stream(model):
         )
 
         reasoning_content_exists = False
+        signature_block_exists = False
         for chunk in resp:
             print(f"chunk 2: {chunk}")
             if (
@@ -1220,8 +1223,11 @@ def test_anthropic_thinking_output_stream(model):
                 and isinstance(chunk.choices[0].delta.reasoning_content, str)
             ):
                 reasoning_content_exists = True
-                break
+                print(chunk.choices[0].delta.thinking_blocks[0])
+                if chunk.choices[0].delta.thinking_blocks[0].get("signature"):
+                    signature_block_exists = True
         assert reasoning_content_exists
+        assert signature_block_exists
     except litellm.Timeout:
         pytest.skip("Model is timing out")
 

tests/local_testing/test_streaming.py

@@ -4084,6 +4084,7 @@ def test_reasoning_content_completion(model):
         )
 
         reasoning_content_exists = False
+        signature_delta_exists = False
         for chunk in resp:
             print(f"chunk 2: {chunk}")
             if (
@@ -4118,3 +4119,4 @@ def test_is_delta_empty():
             audio=None,
         )
     )
+