feat: add thinking config for anthropic llm

Author: sebastienc

src/google/adk/models/anthropic_llm.py

@@ -62,6 +62,13 @@ class _ToolUseAccumulator:
   args_json: str
 
 
+class _ThinkingAccumulator(BaseModel):
+  """Accumulates streamed thinking content block data."""
+
+  thinking: str = ""
+  signature: str = ""
+
+
 class ClaudeRequest(BaseModel):
   system_instruction: str
   messages: Iterable[anthropic_types.MessageParam]
@@ -108,7 +115,24 @@ def part_to_message_block(
     anthropic_types.DocumentBlockParam,
     anthropic_types.ToolUseBlockParam,
     anthropic_types.ToolResultBlockParam,
+    anthropic_types.ThinkingBlockParam,
+    anthropic_types.RedactedThinkingBlockParam,
 ]:
+  if part.thought:
+    signature_str = (
+        part.thought_signature.decode("utf-8") if part.thought_signature else ""
+    )
+    if part.text:
+      return anthropic_types.ThinkingBlockParam(
+          type="thinking",
+          thinking=part.text,
+          signature=signature_str,
+      )
+    else:
+      return anthropic_types.RedactedThinkingBlockParam(
+          type="redacted_thinking",
+          data=signature_str,
+      )
   if part.text:
     return anthropic_types.TextBlockParam(text=part.text, type="text")
   elif part.function_call:
@@ -229,6 +253,18 @@ def content_block_to_part(
     )
     part.function_call.id = content_block.id
     return part
+  if isinstance(content_block, anthropic_types.ThinkingBlock):
+    return types.Part(
+        text=content_block.thinking,
+        thought=True,
+        thought_signature=content_block.signature.encode("utf-8"),
+    )
+  if isinstance(content_block, anthropic_types.RedactedThinkingBlock):
+    return types.Part(
+        text="",
+        thought=True,
+        thought_signature=content_block.data.encode("utf-8"),
+    )
   raise NotImplementedError("Not supported yet.")
 
 
@@ -349,6 +385,26 @@ def function_declaration_to_tool_param(
   )
 
 
+def _build_thinking_param(
+    thinking_config: Optional[types.ThinkingConfig],
+    max_tokens: int,
+) -> Union[anthropic_types.ThinkingConfigEnabledParam, NotGiven]:
+  """Converts ADK ThinkingConfig to Anthropic ThinkingConfigEnabledParam.
+
+  Returns NOT_GIVEN if thinking is not configured or budget is 0.
+  Clamps budget_tokens to max_tokens - 1 to satisfy the API constraint.
+  """
+  if thinking_config is None:
+    return NOT_GIVEN
+  budget = thinking_config.thinking_budget
+  if not budget:
+    return NOT_GIVEN
+  return anthropic_types.ThinkingConfigEnabledParam(
+      type="enabled",
+      budget_tokens=min(budget, max_tokens - 1),
+  )
+
+
 class AnthropicLlm(BaseLlm):
   """Integration with Claude models via the Anthropic API.
 
@@ -401,6 +457,10 @@ async def generate_content_async(
         if llm_request.tools_dict
         else NOT_GIVEN
     )
+    thinking = _build_thinking_param(
+        llm_request.config.thinking_config if llm_request.config else None,
+        self.max_tokens,
+    )
 
     if not stream:
       message = await self._anthropic_client.messages.create(
@@ -410,11 +470,12 @@ async def generate_content_async(
           tools=tools,
           tool_choice=tool_choice,
           max_tokens=self.max_tokens,
+          thinking=thinking,
       )
       yield message_to_generate_content_response(message)
     else:
       async for response in self._generate_content_streaming(
-          llm_request, messages, tools, tool_choice
+          llm_request, messages, tools, tool_choice, thinking
       ):
         yield response
 
@@ -424,6 +485,9 @@ async def _generate_content_streaming(
       messages: list[anthropic_types.MessageParam],
       tools: Union[Iterable[anthropic_types.ToolUnionParam], NotGiven],
       tool_choice: Union[anthropic_types.ToolChoiceParam, NotGiven],
+      thinking: Union[
+          anthropic_types.ThinkingConfigEnabledParam, NotGiven
+      ] = NOT_GIVEN,
   ) -> AsyncGenerator[LlmResponse, None]:
     """Handles streaming responses from Anthropic models.
 
@@ -439,12 +503,15 @@ async def _generate_content_streaming(
         tool_choice=tool_choice,
         max_tokens=self.max_tokens,
         stream=True,
+        thinking=thinking,
     )
 
     # Track content blocks being built during streaming.
     # Each entry maps a block index to its accumulated state.
     text_blocks: dict[int, str] = {}
     tool_use_blocks: dict[int, _ToolUseAccumulator] = {}
+    thinking_blocks: dict[int, _ThinkingAccumulator] = {}
+    redacted_thinking_blocks: dict[int, str] = {}
     input_tokens = 0
     output_tokens = 0
 
@@ -463,6 +530,10 @@ async def _generate_content_streaming(
               name=block.name,
               args_json="",
           )
+        elif isinstance(block, anthropic_types.ThinkingBlock):
+          thinking_blocks[event.index] = _ThinkingAccumulator()
+        elif isinstance(block, anthropic_types.RedactedThinkingBlock):
+          redacted_thinking_blocks[event.index] = block.data
 
       elif event.type == "content_block_delta":
         delta = event.delta
@@ -479,16 +550,44 @@ async def _generate_content_streaming(
         elif isinstance(delta, anthropic_types.InputJSONDelta):
           if event.index in tool_use_blocks:
             tool_use_blocks[event.index].args_json += delta.partial_json
+        elif isinstance(delta, anthropic_types.ThinkingDelta):
+          if event.index in thinking_blocks:
+            thinking_blocks[event.index].thinking += delta.thinking
+        elif isinstance(delta, anthropic_types.SignatureDelta):
+          if event.index in thinking_blocks:
+            thinking_blocks[event.index].signature = delta.signature
 
       elif event.type == "message_delta":
         output_tokens = event.usage.output_tokens
 
     # Build the final aggregated response with all content.
     all_parts: list[types.Part] = []
     all_indices = sorted(
-        set(list(text_blocks.keys()) + list(tool_use_blocks.keys()))
+        set(
+            list(text_blocks.keys())
+            + list(tool_use_blocks.keys())
+            + list(thinking_blocks.keys())
+            + list(redacted_thinking_blocks.keys())
+        )
     )
     for idx in all_indices:
+      if idx in thinking_blocks:
+        acc = thinking_blocks[idx]
+        all_parts.append(
+            types.Part(
+                text=acc.thinking,
+                thought=True,
+                thought_signature=acc.signature.encode("utf-8"),
+            )
+        )
+      if idx in redacted_thinking_blocks:
+        all_parts.append(
+            types.Part(
+                text="",
+                thought=True,
+                thought_signature=redacted_thinking_blocks[idx].encode("utf-8"),
+            )
+        )
       if idx in text_blocks:
         all_parts.append(types.Part.from_text(text=text_blocks[idx]))
       if idx in tool_use_blocks: