diff --git a/src/google/adk/models/lite_llm.py b/src/google/adk/models/lite_llm.py
index 2e4cf65f32..c996971618 100644
--- a/src/google/adk/models/lite_llm.py
+++ b/src/google/adk/models/lite_llm.py
@@ -543,9 +543,11 @@ def _convert_reasoning_value_to_parts(reasoning_value: Any) -> List[types.Part]:
           continue
         if block_type == "thinking":
           thinking_text = block.get("thinking", "")
-          if thinking_text:
+          signature = block.get("signature")
+          # Anthropic streams a signature in a final chunk with empty text.
+          # Preserve signature-only blocks so the signature survives aggregation.
+          if thinking_text or signature:
             part = types.Part(text=thinking_text, thought=True)
-            signature = block.get("signature")
             if signature:
               decoded_signature = _decode_thought_signature(signature)
               part.thought_signature = decoded_signature or str(
@@ -565,6 +567,43 @@ def _convert_reasoning_value_to_parts(reasoning_value: Any) -> List[types.Part]:
   ]
 
 
+def _aggregate_streaming_thought_parts(
+    thought_parts: Iterable[types.Part],
+) -> List[types.Part]:
+  """Aggregates fragmented streaming thought parts into clean individual parts.
+
+  During streaming, Anthropic splits a thinking block across many deltas:
+  text-only chunks followed by a signature-only chunk at block_stop. This helper
+  joins the text chunks and attaches the signature, producing clean individual
+  thought parts for session history and outbound requests.
+  """
+  parts_list = list(thought_parts)
+  if not parts_list:
+    return []
+  aggregated: List[types.Part] = []
+  current_texts: List[str] = []
+  for part in parts_list:
+    if part.text:
+      current_texts.append(part.text)
+    if part.thought_signature:
+      aggregated.append(
+          types.Part(
+              text="".join(current_texts),
+              thought=True,
+              thought_signature=part.thought_signature,
+          )
+      )
+      current_texts = []
+  if current_texts:
+    aggregated.append(
+        types.Part(
+            text="".join(current_texts),
+            thought=True,
+        )
+    )
+  return aggregated
+
+
 def _extract_reasoning_value(message: Message | Delta | None) -> Any:
   """Fetches the reasoning payload from a LiteLLM message.
 
@@ -1023,17 +1062,24 @@ async def _content_to_message_param(
     # For Anthropic models, rebuild thinking_blocks with signatures so that
     # thinking is preserved across tool call boundaries. Without this,
     # Anthropic silently drops thinking after the first turn.
+    #
+    # Streaming splits one Anthropic thinking block across many deltas:
+    # text-only chunks followed by a signature-only chunk at block_stop.
+    # Aggregate them back into one thinking block for outbound.
     if model and _is_anthropic_model(model) and reasoning_parts:
+      aggregated_parts = _aggregate_streaming_thought_parts(reasoning_parts)
       thinking_blocks = []
-      for part in reasoning_parts:
+      for part in aggregated_parts:
         if part.text and part.thought_signature:
           sig = part.thought_signature
           if isinstance(sig, bytes):
-            sig = base64.b64encode(sig).decode("utf-8")
+            sig_str = base64.b64encode(sig).decode("utf-8")
+          else:
+            sig_str = str(sig)
           thinking_blocks.append({
               "type": "thinking",
               "thinking": part.text,
-              "signature": sig,
+              "signature": sig_str,
           })
       if thinking_blocks:
         msg = ChatCompletionAssistantMessage(
@@ -1877,6 +1923,11 @@ def _has_meaningful_signal(message: Message | Delta | None) -> bool:
         or message.get("function_call")
         or message.get("reasoning_content")
         or message.get("reasoning")
+        # Anthropic streams the thinking block's signature in a final delta
+        # where content/reasoning_content are empty and only thinking_blocks
+        # carries the signature. Without this, the delta is discarded before
+        # _extract_reasoning_value can preserve the signature.
+        or message.get("thinking_blocks")
     )
 
   if isinstance(response, ModelResponseStream):
@@ -2717,7 +2768,10 @@ def _finalize_tool_call_response(
         )
         mapped_finish_reason = _map_finish_reason(finish_reason)
         llm_response.finish_reason = mapped_finish_reason
-        if mapped_finish_reason != types.FinishReason.STOP:
+        if (
+            mapped_finish_reason is not None
+            and mapped_finish_reason != types.FinishReason.STOP
+        ):
           llm_response.error_code = mapped_finish_reason
           llm_response.error_message = _finish_reason_to_error_message(
               mapped_finish_reason
@@ -2738,7 +2792,10 @@ def _finalize_text_response(
         )
         mapped_finish_reason = _map_finish_reason(finish_reason)
         llm_response.finish_reason = mapped_finish_reason
-        if mapped_finish_reason != types.FinishReason.STOP:
+        if (
+            mapped_finish_reason is not None
+            and mapped_finish_reason != types.FinishReason.STOP
+        ):
           llm_response.error_code = mapped_finish_reason
           llm_response.error_message = _finish_reason_to_error_message(
               mapped_finish_reason
diff --git a/tests/unittests/models/test_litellm.py b/tests/unittests/models/test_litellm.py
index 6f4341f55f..b887f598f2 100644
--- a/tests/unittests/models/test_litellm.py
+++ b/tests/unittests/models/test_litellm.py
@@ -27,6 +27,7 @@
 from unittest.mock import patch
 import warnings
 
+from google.adk.models.lite_llm import _aggregate_streaming_thought_parts
 from google.adk.models.lite_llm import _append_fallback_user_content_if_missing
 from google.adk.models.lite_llm import _content_to_message_param
 from google.adk.models.lite_llm import _convert_reasoning_value_to_parts
@@ -5396,15 +5397,48 @@ def test_convert_reasoning_value_to_parts_skips_redacted_blocks():
   assert parts[0].text == "visible"
 
 
-def test_convert_reasoning_value_to_parts_skips_empty_thinking():
-  """Blocks with empty thinking text are excluded."""
+def test_convert_reasoning_value_to_parts_preserves_signature_only_blocks():
+  """Signature-only blocks (empty text) are preserved for streaming aggregation.
+
+  Anthropic emits the block_stop signature as a delta with empty thinking text.
+  Dropping it would lose the signature, breaking multi-turn thinking continuity.
+  Blocks with neither text nor signature are still skipped.
+  """
   thinking_blocks = [
       {"type": "thinking", "thinking": "", "signature": "c2lnMQ=="},
       {"type": "thinking", "thinking": "real thought", "signature": "c2lnMg=="},
+      {
+          "type": "thinking",
+          "thinking": "",
+          "signature": "",
+      },  # fully empty: drop
   ]
   parts = _convert_reasoning_value_to_parts(thinking_blocks)
-  assert len(parts) == 1
-  assert parts[0].text == "real thought"
+  assert len(parts) == 2
+  assert parts[0].text == ""
+  assert parts[0].thought is True
+  assert parts[0].thought_signature == b"sig1"
+  assert parts[1].text == "real thought"
+  assert parts[1].thought_signature == b"sig2"
+
+
+def test_aggregate_streaming_thought_parts():
+  """Tests aggregating fragmented streaming thought parts and multiple blocks."""
+  parts = [
+      types.Part(text="First block ", thought=True),
+      types.Part(text="text.", thought=True),
+      types.Part(text="", thought=True, thought_signature=b"sig1"),
+      types.Part(text="Second block", thought=True, thought_signature=b"sig2"),
+      types.Part(text="Trailing without sig", thought=True),
+  ]
+  aggregated = _aggregate_streaming_thought_parts(parts)
+  assert len(aggregated) == 3
+  assert aggregated[0].text == "First block text."
+  assert aggregated[0].thought_signature == b"sig1"
+  assert aggregated[1].text == "Second block"
+  assert aggregated[1].thought_signature == b"sig2"
+  assert aggregated[2].text == "Trailing without sig"
+  assert aggregated[2].thought_signature is None
 
 
 def test_convert_reasoning_value_to_parts_flat_string_unchanged():
@@ -5478,6 +5512,33 @@ async def test_content_to_message_param_anthropic_model_round_trip_preserves_sig
   assert result.get("reasoning_content") is None
 
 
+@pytest.mark.asyncio
+async def test_content_to_message_param_anthropic_split_thinking_and_signature():
+  """Combines separate thinking and signature parts into a single thinking_block."""
+  content = types.Content(
+      role="model",
+      parts=[
+          types.Part(text="deep thought", thought=True),
+          types.Part(
+              text="", thought=True, thought_signature=b"sig_round_trip"
+          ),
+          types.Part(text="Hello!"),
+      ],
+  )
+  result = await _content_to_message_param(
+      content, model="anthropic/claude-4-sonnet"
+  )
+  assert result["role"] == "assistant"
+  assert "thinking_blocks" in result
+  assert result.get("reasoning_content") is None
+  blocks = result["thinking_blocks"]
+  assert len(blocks) == 1
+  assert blocks[0]["type"] == "thinking"
+  assert blocks[0]["thinking"] == "deep thought"
+  assert blocks[0]["signature"] == "c2lnX3JvdW5kX3RyaXA="
+  assert result["content"] == "Hello!"
+
+
 @pytest.mark.asyncio
 async def test_content_to_message_param_non_anthropic_uses_reasoning_content():
   """For non-Anthropic models, reasoning_content is used as before."""
@@ -5577,7 +5638,7 @@ def test_convert_reasoning_value_to_parts_empty_thinking_does_not_fall_through()
           "type": "thinking",
           "thinking": "",
           "text": "leaked",
-          "signature": "c2ln",
+          "signature": "",
       },
   ]
   parts = _convert_reasoning_value_to_parts(thinking_blocks)
@@ -5678,6 +5739,76 @@ async def test_content_to_message_param_anthropic_provider_embeds_thinking_block
   assert result.get("reasoning_content") is None
 
 
+@pytest.mark.asyncio
+async def test_content_to_message_param_anthropic_aggregates_streaming_split_thinking():
+  """Streaming splits one Anthropic thinking block across many parts:
+  text-only chunks followed by a signature-only chunk at block_stop.
+  _content_to_message_param must re-join them into one thinking_block.
+  """
+  content = types.Content(
+      role="model",
+      parts=[
+          # Text-only chunks from streaming deltas (no signature)
+          types.Part(text="The user wants ", thought=True),
+          types.Part(text="GST research ", thought=True),
+          types.Part(text="on secondment.", thought=True),
+          # Final signature-only chunk (empty text, signature carries the whole block)
+          types.Part(
+              text="", thought=True, thought_signature=b"ErEDClsIDBACGAIfull"
+          ),
+          # Non-thought response content
+          types.Part.from_function_call(name="create_plan", args={"q": "test"}),
+      ],
+  )
+  result = await _content_to_message_param(
+      content, model="anthropic/claude-4-sonnet"
+  )
+  # One aggregated thinking block with combined text and the block's signature
+  blocks = result["thinking_blocks"]
+  assert len(blocks) == 1
+  assert blocks[0]["type"] == "thinking"
+  assert blocks[0]["thinking"] == "The user wants GST research on secondment."
+  assert blocks[0]["signature"] == "RXJFRENsc0lEQkFDR0FJZnVsbA=="
+  # Legacy reasoning_content is not set when the Anthropic branch takes
+  assert result.get("reasoning_content") is None
+
+
+def test_model_response_to_chunk_preserves_signature_only_delta():
+  """Anthropic streams a final thinking delta where content and
+  reasoning_content are empty but thinking_blocks carries the signature.
+  _has_meaningful_signal must recognize thinking_blocks as signal so the
+  signature survives into a ReasoningChunk.
+  """
+  stream = ModelResponseStream(
+      id="x",
+      created=0,
+      model="claude",
+      choices=[
+          StreamingChoices(
+              index=0,
+              delta=Delta(
+                  role=None,
+                  content="",
+                  reasoning_content="",
+                  thinking_blocks=[{
+                      "type": "thinking",
+                      "thinking": "",
+                      "signature": "SignatureOnlyChunk",
+                  }],
+              ),
+          )
+      ],
+  )
+  chunks = list(_model_response_to_chunk(stream))
+  reasoning_chunks = [c for c, _ in chunks if isinstance(c, ReasoningChunk)]
+  assert len(reasoning_chunks) == 1
+  parts = reasoning_chunks[0].parts
+  assert len(parts) == 1
+  assert parts[0].text == ""
+  assert parts[0].thought is True
+  assert parts[0].thought_signature == b"SignatureOnlyChunk"
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     "log_level,should_call",