feat: Support Anthropic extended thinking and interleaved thinking

Anthropic's API requires thinking blocks to be the first content in assistant
messages when reasoning is enabled and tool calls are present.
ref: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking

This change:
- Stores thinking text in ResponseReasoningItem.content field
- Stores signature in encrypted_content field for metadata
- Reconstructs thinking blocks when converting items->messages
- Handles both streaming and non-streaming cases
- Only affects Anthropic models with reasoning enabled
- Adds regression test to prevent future breakage

Fixes compatibility with Claude models when using tools with interleaved thinking.

Author: serialx

src/agents/models/chatcmpl_converter.py

@@ -39,7 +39,7 @@
     ResponseReasoningItemParam,
 )
 from openai.types.responses.response_input_param import FunctionCallOutput, ItemReference, Message
-from openai.types.responses.response_reasoning_item import Summary
+from openai.types.responses.response_reasoning_item import Content, Summary
 
 from ..agent_output import AgentOutputSchemaBase
 from ..exceptions import AgentsException, UserError
@@ -93,24 +93,38 @@ def convert_response_format(
     def message_to_output_items(cls, message: ChatCompletionMessage) -> list[TResponseOutputItem]:
         items: list[TResponseOutputItem] = []
 
-        # Handle reasoning content if available
+        # Check if message is agents.extentions.models.litellm_model.InternalChatCompletionMessage
+        # We can't actually import it here because litellm is an optional dependency
+        # So we use hasattr to check for reasoning_content and thinking_blocks
         if hasattr(message, "reasoning_content") and message.reasoning_content:
             reasoning_item = ResponseReasoningItem(
                 id=FAKE_RESPONSES_ID,
                 summary=[Summary(text=message.reasoning_content, type="summary_text")],
                 type="reasoning",
             )
 
-            # Store full thinking blocks for Anthropic compatibility
+            # Store thinking blocks for Anthropic compatibility
             if hasattr(message, "thinking_blocks") and message.thinking_blocks:
-                # Store thinking blocks in the reasoning item's content
-                # Convert thinking blocks to Content objects
-                from openai.types.responses.response_reasoning_item import Content
-
-                reasoning_item.content = [
-                    Content(text=str(block.get("thinking", "")), type="reasoning_text")
-                    for block in message.thinking_blocks
-                ]
+                # Store thinking text in content and signature in encrypted_content
+                reasoning_item.content = []
+                signature = None
+                for block in message.thinking_blocks:
+                    if isinstance(block, dict):
+                        thinking_text = block.get("thinking", "")
+                        if thinking_text:
+                            reasoning_item.content.append(
+                                Content(text=thinking_text, type="reasoning_text")
+                            )
+                        # Store the signature if present
+                        if block.get("signature"):
+                            signature = block.get("signature")
+
+                # Store only the last signature in encrypted_content
+                # If there are multiple thinking blocks, this should be a problem.
+                # In practice, there should only be one signature for the entire reasoning step.
+                # Tested with: claude-sonnet-4-20250514
+                if signature:
+                    reasoning_item.encrypted_content = signature
 
             items.append(reasoning_item)
 
@@ -325,6 +339,7 @@ def items_to_messages(
 
         result: list[ChatCompletionMessageParam] = []
         current_assistant_msg: ChatCompletionAssistantMessageParam | None = None
+        pending_thinking_blocks: list[dict[str, str]] | None = None
 
         def flush_assistant_message() -> None:
             nonlocal current_assistant_msg
@@ -336,10 +351,17 @@ def flush_assistant_message() -> None:
                 current_assistant_msg = None
 
         def ensure_assistant_message() -> ChatCompletionAssistantMessageParam:
-            nonlocal current_assistant_msg
+            nonlocal current_assistant_msg, pending_thinking_blocks
             if current_assistant_msg is None:
                 current_assistant_msg = ChatCompletionAssistantMessageParam(role="assistant")
                 current_assistant_msg["tool_calls"] = []
+
+            # If we have pending thinking blocks, use them as the content
+            # This is required for Anthropic API tool calls with interleaved thinking
+            if pending_thinking_blocks:
+                current_assistant_msg["content"] = pending_thinking_blocks  # type: ignore
+                pending_thinking_blocks = None  # Clear after using
+
             return current_assistant_msg
 
         for item in items:
@@ -483,9 +505,28 @@ def ensure_assistant_message() -> ChatCompletionAssistantMessageParam:
                     f"Encountered an item_reference, which is not supported: {item_ref}"
                 )
 
-            # 7) reasoning message => not handled
-            elif cls.maybe_reasoning_message(item):
-                pass
+            # 7) reasoning message => extract thinking blocks if present
+            elif reasoning_item := cls.maybe_reasoning_message(item):
+                # Reconstruct thinking blocks from content (text) and encrypted_content (signature)
+                content_items = reasoning_item.get("content", [])
+                signature = reasoning_item.get("encrypted_content")
+
+                if content_items:
+                    # Reconstruct thinking blocks from content and signature
+                    pending_thinking_blocks = []
+                    for content_item in content_items:
+                        if (
+                            isinstance(content_item, dict)
+                            and content_item.get("type") == "reasoning_text"
+                        ):
+                            thinking_block = {
+                                "type": "thinking",
+                                "thinking": content_item.get("text", ""),
+                            }
+                            # Add signature if available
+                            if signature:
+                                thinking_block["signature"] = signature
+                            pending_thinking_blocks.append(thinking_block)
 
             # 8) If we haven't recognized it => fail or ignore
             else:

src/agents/models/chatcmpl_stream_handler.py

@@ -62,6 +62,9 @@ class StreamingState:
     # Fields for real-time function call streaming
     function_call_streaming: dict[int, bool] = field(default_factory=dict)
     function_call_output_idx: dict[int, int] = field(default_factory=dict)
+    # Store accumulated thinking text and signature for Anthropic compatibility
+    thinking_text: str = ""
+    thinking_signature: str | None = None
 
 
 class SequenceNumber:
@@ -101,6 +104,19 @@ async def handle_stream(
 
             delta = chunk.choices[0].delta
 
+            # Handle thinking blocks from Anthropic (for preserving signatures)
+            if hasattr(delta, "thinking_blocks") and delta.thinking_blocks:
+                for block in delta.thinking_blocks:
+                    if isinstance(block, dict):
+                        # Accumulate thinking text
+                        thinking_text = block.get("thinking", "")
+                        if thinking_text:
+                            state.thinking_text += thinking_text
+                        # Store signature if present
+                        signature = block.get("signature")
+                        if signature:
+                            state.thinking_signature = signature
+
             # Handle reasoning content for reasoning summaries
             if hasattr(delta, "reasoning_content"):
                 reasoning_content = delta.reasoning_content
@@ -527,7 +543,19 @@ async def handle_stream(
 
         # include Reasoning item if it exists
         if state.reasoning_content_index_and_output:
-            outputs.append(state.reasoning_content_index_and_output[1])
+            reasoning_item = state.reasoning_content_index_and_output[1]
+            # Store thinking text in content and signature in encrypted_content
+            if state.thinking_text:
+                # Add thinking text as a Content object
+                if not reasoning_item.content:
+                    reasoning_item.content = []
+                reasoning_item.content.append(
+                    Content(text=state.thinking_text, type="reasoning_text")
+                )
+            # Store signature in encrypted_content
+            if state.thinking_signature:
+                reasoning_item.encrypted_content = state.thinking_signature
+            outputs.append(reasoning_item)
 
         # include text or refusal content if they exist
         if state.text_content_index_and_output or state.refusal_content_index_and_output:

tests/test_anthropic_thinking_blocks.py

@@ -10,7 +10,10 @@
 
 from __future__ import annotations
 
-from typing import Any
+from typing import Any, cast
+
+from openai.types.chat import ChatCompletionMessageToolCall
+from openai.types.chat.chat_completion_message_tool_call import Function
 
 from agents.extensions.models.litellm_model import InternalChatCompletionMessage
 from agents.models.chatcmpl_converter import Converter
@@ -99,3 +102,115 @@ def test_reasoning_items_preserved_in_message_conversion():
         thinking_block = reasoning_item.content[0]
         assert thinking_block.type == "reasoning_text"
         assert thinking_block.text == "I need to call the weather function for Paris"
+
+
+def test_anthropic_thinking_blocks_with_tool_calls():
+    """
+    Test for models with extended thinking and interleaved thinking with tool calls.
+
+    This test verifies the Anthropic's API's requirements for thinking blocks
+    to be the first content in assistant messages when reasoning is enabled and tool
+    calls are present.
+    """
+    # Create a message with reasoning, thinking blocks and tool calls
+    message = InternalChatCompletionMessage(
+        role="assistant",
+        content="I'll check the weather for you.",
+        reasoning_content="The user wants weather information, I need to call the weather function",
+        thinking_blocks=[
+            {
+                "type": "thinking",
+                "thinking": (
+                    "The user is asking about weather. "
+                    "Let me use the weather tool to get this information."
+                ),
+                "signature": "TestSignature123",
+            }
+        ],
+        tool_calls=[
+            ChatCompletionMessageToolCall(
+                id="call_123",
+                type="function",
+                function=Function(name="get_weather", arguments='{"city": "Tokyo"}'),
+            )
+        ],
+    )
+
+    # Step 1: Convert message to output items
+    output_items = Converter.message_to_output_items(message)
+
+    # Verify reasoning item exists and contains thinking blocks
+    reasoning_items = [
+        item for item in output_items if hasattr(item, "type") and item.type == "reasoning"
+    ]
+    assert len(reasoning_items) == 1, "Should have exactly one reasoning item"
+
+    reasoning_item = reasoning_items[0]
+
+    # Verify thinking text is stored in content
+    assert hasattr(reasoning_item, "content") and reasoning_item.content, (
+        "Reasoning item should have content"
+    )
+    assert reasoning_item.content[0].type == "reasoning_text", (
+        "Content should be reasoning_text type"
+    )
+
+    # Verify signature is stored in encrypted_content
+    assert hasattr(reasoning_item, "encrypted_content"), (
+        "Reasoning item should have encrypted_content"
+    )
+    assert reasoning_item.encrypted_content == "TestSignature123", "Signature should be preserved"
+
+    # Verify tool calls are present
+    tool_call_items = [
+        item for item in output_items if hasattr(item, "type") and item.type == "function_call"
+    ]
+    assert len(tool_call_items) == 1, "Should have exactly one tool call"
+
+    # Step 2: Convert output items back to messages
+    # Convert items to dicts for the converter (simulating serialization/deserialization)
+    items_as_dicts: list[dict[str, Any]] = []
+    for item in output_items:
+        if hasattr(item, "model_dump"):
+            items_as_dicts.append(item.model_dump())
+        else:
+            items_as_dicts.append(cast(dict[str, Any], item))
+
+    messages = Converter.items_to_messages(items_as_dicts)  # type: ignore[arg-type]
+
+    # Find the assistant message with tool calls
+    assistant_messages = [
+        msg for msg in messages if msg.get("role") == "assistant" and msg.get("tool_calls")
+    ]
+    assert len(assistant_messages) == 1, "Should have exactly one assistant message with tool calls"
+
+    assistant_msg = assistant_messages[0]
+
+    # Content must start with thinking blocks, not text
+    content = assistant_msg.get("content")
+    assert content is not None, "Assistant message should have content"
+
+    assert isinstance(content, list) and len(content) > 0, (
+        "Assistant message content should be a non-empty list"
+    )
+
+    first_content = content[0]
+    assert first_content.get("type") == "thinking", (
+        f"First content must be 'thinking' type for Anthropic compatibility, "
+        f"but got '{first_content.get('type')}'"
+    )
+    expected_thinking = (
+        "The user is asking about weather. Let me use the weather tool to get this information."
+    )
+    assert first_content.get("thinking") == expected_thinking, (
+        "Thinking content should be preserved"
+    )
+    # Signature should also be preserved
+    assert first_content.get("signature") == "TestSignature123", (
+        "Signature should be preserved in thinking block"
+    )
+
+    # Verify tool calls are preserved
+    tool_calls = assistant_msg.get("tool_calls", [])
+    assert len(cast(list[Any], tool_calls)) == 1, "Tool calls should be preserved"
+    assert cast(list[Any], tool_calls)[0]["function"]["name"] == "get_weather"