扩展思考 + 工具调用
在扩展思考模式下结合工具使用
🌐 查看英文原文 | 源码 Notebook
扩展思考与工具调用
目录
This notebook demonstrates how to use Claude 3.7 Sonnet’s extended thinking feature with tools. The extended thinking feature allows you to see Claude’s step-by-step thinking before it provides a final answer, providing transparency into how it decides which tools to use and how it interprets tool results.
When using extended thinking with tool use, the model will show its thinking before making tool requests, but not repeat the thinking process after receiving tool results. Claude will not output another thinking block until after the next non-tool_result user turn. For more information on extended thinking, see our documentation.
环境配置
First, let’s install the necessary packages and set up our environment.
%pip install anthropicimport anthropic
import os
import json
# Global variables for model and token budgets
MODEL_NAME = "claude-sonnet-4-6"
MAX_TOKENS = 4000
THINKING_BUDGET_TOKENS = 2000
# Set your API key as an environment variable or directly
# os.environ["ANTHROPIC_API_KEY"] = "your_api_key_here"
# Initialize the client
client = anthropic.Anthropic()
# Helper functions
def print_thinking_response(response):
"""Pretty print a message response with thinking blocks."""
print("\n==== FULL RESPONSE ====")
for block in response.content:
if block.type == "thinking":
print("\n🧠 THINKING BLOCK:")
# Show truncated thinking for readability
print(block.thinking[:500] + "..." if len(block.thinking) > 500 else block.thinking)
print(f"\n[Signature available: {bool(getattr(block, 'signature', None))}]")
if hasattr(block, 'signature') and block.signature:
print(f"[Signature (first 50 chars): {block.signature[:50]}...]")
elif block.type == "redacted_thinking":
print("\n🔒 REDACTED THINKING BLOCK:")
print(f"[Data length: {len(block.data) if hasattr(block, 'data') else 'N/A'}]")
elif block.type == "text":
print("\n✓ FINAL ANSWER:")
print(block.text)
print("\n==== END RESPONSE ====")
def count_tokens(messages, tools=None):
"""Count tokens for a given message list with optional tools."""
if tools:
response = client.messages.count_tokens(
model=MODEL_NAME,
messages=messages,
tools=tools
)
else:
response = client.messages.count_tokens(
model=MODEL_NAME,
messages=messages
)
return response.input_tokens思考与单次工具调用
此示例演示了如何将思考与单次工具调用相结合，使用一个模拟天气工具。
def tool_use_with_thinking():
# Define a weather tool
tools = [
{
"name": "weather",
"description": "Get current weather information for a location.",
"input_schema": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The location to get weather for."
}
},
"required": ["location"]
}
}
]
def weather(location):
# Mock weather data
weather_data = {
"New York": {"temperature": 72, "condition": "Sunny"},
"London": {"temperature": 62, "condition": "Cloudy"},
"Tokyo": {"temperature": 80, "condition": "Partly cloudy"},
"Paris": {"temperature": 65, "condition": "Rainy"},
"Sydney": {"temperature": 85, "condition": "Clear"},
"Berlin": {"temperature": 60, "condition": "Foggy"},
}
return weather_data.get(location, {"error": f"No weather data available for {location}"})
# Initial request with tool use and thinking
response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=[{
"role": "user",
"content": "What's the weather like in Paris today?"
}]
)
# Detailed diagnostic output of initial response
print("\n=== INITIAL RESPONSE ===")
print(f"Response ID: {response.id}")
print(f"Stop reason: {response.stop_reason}")
print(f"Model: {response.model}")
print(f"Content blocks: {len(response.content)} blocks")
for i, block in enumerate(response.content):
print(f"\nBlock {i+1}: Type = {block.type}")
if block.type == "thinking":
print(f"Thinking content: {block.thinking[:150]}...")
print(f"Signature available: {bool(getattr(block, 'signature', None))}")
elif block.type == "text":
print(f"Text content: {block.text}")
elif block.type == "tool_use":
print(f"Tool: {block.name}")
print(f"Tool input: {block.input}")
print(f"Tool ID: {block.id}")
print("=== END INITIAL RESPONSE ===\n")
# Extract thinking blocks to include in the conversation history
assistant_blocks = []
for block in response.content:
if block.type in ["thinking", "redacted_thinking", "tool_use"]:
assistant_blocks.append(block)
# Handle tool use if required
full_conversation = [{
"role": "user",
"content": "What's the weather like in Paris today?"
}]
if response.stop_reason == "tool_use":
# Add entire assistant response with thinking blocks and tool use
full_conversation.append({
"role": "assistant",
"content": assistant_blocks
})
# Find the tool_use block
tool_use_block = next((block for block in response.content if block.type == "tool_use"), None)
if tool_use_block:
# Execute the tool
print(f"\n=== EXECUTING TOOL ===")
print(f"Tool name: {tool_use_block.name}")
print(f"Location to check: {tool_use_block.input['location']}")
tool_result = weather(tool_use_block.input["location"])
print(f"Result: {tool_result}")
print("=== TOOL EXECUTION COMPLETE ===\n")
# Add tool result to conversation
full_conversation.append({
"role": "user",
"content": [{
"type": "tool_result",
"tool_use_id": tool_use_block.id,
"content": json.dumps(tool_result)
}]
})
# Continue the conversation with the same thinking configuration
print("\n=== SENDING FOLLOW-UP REQUEST WITH TOOL RESULT ===")
response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=full_conversation
)
print(f"Follow-up response received. Stop reason: {response.stop_reason}")
print_thinking_response(response)
# Run the example
tool_use_with_thinking()
=== INITIAL RESPONSE ===
Response ID: msg_01NhR4vE9nVh2sHs5fXbzji8
Stop reason: tool_use
Model: claude-sonnet-4-6
Content blocks: 3 blocks
Block 1: Type = thinking
Thinking content: The user is asking about the current weather in Paris. I can use the `weather` function to get this information.
The `weather` function requires a "l...
Signature available: True
Block 2: Type = text
Text content: I'll check the current weather in Paris for you.
Block 3: Type = tool_use
Tool: weather
Tool input: {'location': 'Paris'}
Tool ID: toolu_01WaeSyitUGJFaaPe68cJuEv
=== END INITIAL RESPONSE ===
=== EXECUTING TOOL ===
Tool name: weather
Location to check: Paris
Result: {'temperature': 65, 'condition': 'Rainy'}
=== TOOL EXECUTION COMPLETE ===
=== SENDING FOLLOW-UP REQUEST WITH TOOL RESULT ===
Follow-up response received. Stop reason: end_turn
==== FULL RESPONSE ====
✓ FINAL ANSWER:
Currently in Paris, it's 65°F (18°C) and rainy. You might want to bring an umbrella if you're heading out!
==== END RESPONSE ====思考与多次工具调用
此示例演示了如何处理多次工具调用（例如模拟新闻和天气服务），同时观察思考过程。
def multiple_tool_calls_with_thinking():
# Define tools
tools = [
{
"name": "weather",
"description": "Get current weather information for a location.",
"input_schema": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The location to get weather for."
}
},
"required": ["location"]
}
},
{
"name": "news",
"description": "Get latest news headlines for a topic.",
"input_schema": {
"type": "object",
"properties": {
"topic": {
"type": "string",
"description": "The topic to get news about."
}
},
"required": ["topic"]
}
}
]
def weather(location):
# Mock weather data
weather_data = {
"New York": {"temperature": 72, "condition": "Sunny"},
"London": {"temperature": 62, "condition": "Cloudy"},
"Tokyo": {"temperature": 80, "condition": "Partly cloudy"},
"Paris": {"temperature": 65, "condition": "Rainy"},
"Sydney": {"temperature": 85, "condition": "Clear"},
"Berlin": {"temperature": 60, "condition": "Foggy"},
}
return weather_data.get(location, {"error": f"No weather data available for {location}"})
def news(topic):
# Mock news data
news_data = {
"technology": [
"New AI breakthrough announced by research lab",
"Tech company releases latest smartphone model",
"Quantum computing reaches milestone achievement"
],
"sports": [
"Local team wins championship game",
"Star player signs record-breaking contract",
"Olympic committee announces host city for 2036"
],
"weather": [
"Storm system developing in the Atlantic",
"Record temperatures recorded across Europe",
"Climate scientists release new research findings"
]
}
return {"headlines": news_data.get(topic.lower(), ["No news available for this topic"])}
# Initial request
response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=[{
"role": "user",
"content": "What's the weather in London, and can you also tell me the latest news about technology?"
}]
)
# Print detailed information about initial response
print("\n=== INITIAL RESPONSE ===")
print(f"Response ID: {response.id}")
print(f"Stop reason: {response.stop_reason}")
print(f"Model: {response.model}")
print(f"Content blocks: {len(response.content)} blocks")
# Print each content block
for i, block in enumerate(response.content):
print(f"\nBlock {i+1}: Type = {block.type}")
if block.type == "thinking":
print(f"Thinking content: {block.thinking[:150]}...")
print(f"Signature available: {bool(getattr(block, 'signature', None))}")
elif block.type == "text":
print(f"Text content: {block.text}")
elif block.type == "tool_use":
print(f"Tool: {block.name}")
print(f"Tool input: {block.input}")
print(f"Tool ID: {block.id}")
print("=== END INITIAL RESPONSE ===\n")
# Handle potentially multiple tool calls
full_conversation = [{
"role": "user",
"content": "What's the weather in London, and can you also tell me the latest news about technology?"
}]
# Track iteration count for multi-turn tool use
iteration = 0
while response.stop_reason == "tool_use":
iteration += 1
print(f"\n=== TOOL USE ITERATION {iteration} ===")
# Extract thinking blocks and tool use to include in conversation history
assistant_blocks = []
for block in response.content:
if block.type in ["thinking", "redacted_thinking", "tool_use"]:
assistant_blocks.append(block)
# Add assistant response with thinking blocks and tool use
full_conversation.append({
"role": "assistant",
"content": assistant_blocks
})
# Find the tool_use block
tool_use_block = next((block for block in response.content if block.type == "tool_use"), None)
if tool_use_block:
print(f"\n=== EXECUTING TOOL ===")
print(f"Tool name: {tool_use_block.name}")
# Execute the appropriate tool
if tool_use_block.name == "weather":
print(f"Location to check: {tool_use_block.input['location']}")
tool_result = weather(tool_use_block.input["location"])
elif tool_use_block.name == "news":
print(f"Topic to check: {tool_use_block.input['topic']}")
tool_result = news(tool_use_block.input["topic"])
else:
tool_result = {"error": "Unknown tool"}
print(f"Result: {tool_result}")
print("=== TOOL EXECUTION COMPLETE ===\n")
# Add tool result to conversation
full_conversation.append({
"role": "user",
"content": [{
"type": "tool_result",
"tool_use_id": tool_use_block.id,
"content": json.dumps(tool_result)
}]
})
# Continue the conversation
print("\n=== SENDING FOLLOW-UP REQUEST WITH TOOL RESULT ===")
response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=full_conversation
)
# Print follow-up response details
print(f"\n=== FOLLOW-UP RESPONSE (ITERATION {iteration}) ===")
print(f"Response ID: {response.id}")
print(f"Stop reason: {response.stop_reason}")
print(f"Content blocks: {len(response.content)} blocks")
for i, block in enumerate(response.content):
print(f"\nBlock {i+1}: Type = {block.type}")
if block.type == "thinking":
print(f"Thinking content preview: {block.thinking[:100]}...")
elif block.type == "text":
print(f"Text content preview: {block.text[:100]}...")
elif block.type == "tool_use":
print(f"Tool: {block.name}")
print(f"Tool input preview: {str(block.input)[:100]}")
print(f"=== END FOLLOW-UP RESPONSE (ITERATION {iteration}) ===\n")
if response.stop_reason != "tool_use":
print("\n=== FINAL RESPONSE ===")
print_thinking_response(response)
print("=== END FINAL RESPONSE ===")
else:
print("No tool_use block found in response.")
break
# Run the example
multiple_tool_calls_with_thinking()
=== INITIAL RESPONSE ===
Response ID: msg_01VwqpBMARVoTP1H8Ytvmvsb
Stop reason: tool_use
Model: claude-sonnet-4-6
Content blocks: 3 blocks
Block 1: Type = thinking
Thinking content: The user is asking for two pieces of information:
1. The weather in London
2. The latest news about technology
Let me check what tools I have availab...
Signature available: True
Block 2: Type = text
Text content: I'll get that information for you right away.
Block 3: Type = tool_use
Tool: weather
Tool input: {'location': 'London'}
Tool ID: toolu_016xHQWMR4JsKtWvH9nbsZyA
=== END INITIAL RESPONSE ===
=== TOOL USE ITERATION 1 ===
=== EXECUTING TOOL ===
Tool name: weather
Location to check: London
Result: {'temperature': 62, 'condition': 'Cloudy'}
=== TOOL EXECUTION COMPLETE ===
=== SENDING FOLLOW-UP REQUEST WITH TOOL RESULT ===
=== FOLLOW-UP RESPONSE (ITERATION 1) ===
Response ID: msg_01EhR96Z2Z2t5EDhuWeodUod
Stop reason: tool_use
Content blocks: 1 blocks
Block 1: Type = tool_use
Tool: news
Tool input preview: {'topic': 'technology'}
=== END FOLLOW-UP RESPONSE (ITERATION 1) ===
=== TOOL USE ITERATION 2 ===
=== EXECUTING TOOL ===
Tool name: news
Topic to check: technology
Result: {'headlines': ['New AI breakthrough announced by research lab', 'Tech company releases latest smartphone model', 'Quantum computing reaches milestone achievement']}
=== TOOL EXECUTION COMPLETE ===
=== SENDING FOLLOW-UP REQUEST WITH TOOL RESULT ===
=== FOLLOW-UP RESPONSE (ITERATION 2) ===
Response ID: msg_01WUEfC4UxPFaJaktjVDMJEN
Stop reason: end_turn
Content blocks: 1 blocks
Block 1: Type = text
Text content preview: Here's the information you requested:
## Weather in London
Currently, it's 62°F and cloudy in Londo...
=== END FOLLOW-UP RESPONSE (ITERATION 2) ===
=== FINAL RESPONSE ===
==== FULL RESPONSE ====
✓ FINAL ANSWER:
Here's the information you requested:
## Weather in London
Currently, it's 62°F and cloudy in London.
## Latest Technology News Headlines
- New AI breakthrough announced by research lab
- Tech company releases latest smartphone model
- Quantum computing reaches milestone achievement
==== END RESPONSE ====
=== END FINAL RESPONSE ===保留思考块
在使用扩展思考和工具时，请确保：
保留思考块签名：每个思考块包含一个验证对话上下文的加密签名。在将思考块传回给 Claude 时必须包含这些签名。
避免修改先前的上下文：如果在使用工具结果提交新请求时修改了任何先前内容（包括思考块），API 将拒绝请求。
同时处理 thinking 和 redacted_thinking 块：两种类型的块都必须在对话历史中保留，即使已编辑块的内容对人类不可读。
有关不使用工具的扩展思考的更多详细信息，请参阅主”扩展思考”notebook。
def thinking_block_preservation_example():
# Define a simple weather tool
tools = [
{
"name": "weather",
"description": "Get current weather information for a location.",
"input_schema": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The location to get weather for."
}
},
"required": ["location"]
}
}
]
def weather(location):
# Mock weather data
weather_data = {
"New York": {"temperature": 72, "condition": "Sunny"},
"London": {"temperature": 62, "condition": "Cloudy"},
"Tokyo": {"temperature": 80, "condition": "Partly cloudy"},
"Paris": {"temperature": 65, "condition": "Rainy"},
"Sydney": {"temperature": 85, "condition": "Clear"},
"Berlin": {"temperature": 60, "condition": "Foggy"},
}
return weather_data.get(location, {"error": f"No weather data available for {location}"})
# Initial request with tool use and thinking
response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=[{
"role": "user",
"content": "What's the weather like in Berlin right now?"
}]
)
# Extract blocks from response
thinking_blocks = [b for b in response.content if b.type == "thinking"]
tool_use_blocks = [b for b in response.content if b.type == "tool_use"]
print("\n=== INITIAL RESPONSE ===")
print(f"Response contains:")
print(f"- {len(thinking_blocks)} thinking blocks")
print(f"- {len(tool_use_blocks)} tool use blocks")
# Check if tool use was triggered
if tool_use_blocks:
tool_block = tool_use_blocks[0]
print(f"\nTool called: {tool_block.name}")
print(f"Location to check: {tool_block.input['location']}")
# Execute the tool
tool_result = weather(tool_block.input["location"])
print(f"Tool result: {tool_result}")
# First, let's try WITHOUT including the thinking block
print("\n=== TEST 1: WITHOUT thinking block ===")
try:
# Notice we're only including the tool_use block, not the thinking block
partial_blocks = tool_use_blocks
incomplete_response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=[
{"role": "user", "content": "What's the weather like in Berlin right now?"},
{"role": "assistant", "content": partial_blocks},
{"role": "user", "content": [{
"type": "tool_result",
"tool_use_id": tool_block.id,
"content": json.dumps(tool_result)
}]}
]
)
print("SUCCESS: Response received without thinking block (not expected)")
except Exception as e:
print(f"ERROR: {e}")
print("This demonstrates that thinking blocks must be preserved")
# Now try WITH the thinking block included (correct approach)
print("\n=== TEST 2: WITH thinking block (correct approach) ===")
try:
# Include all blocks from the response
complete_blocks = thinking_blocks + tool_use_blocks
complete_response = client.messages.create(
model=MODEL_NAME,
max_tokens=MAX_TOKENS,
thinking={
"type": "enabled",
"budget_tokens": THINKING_BUDGET_TOKENS
},
tools=tools,
messages=[
{"role": "user", "content": "What's the weather like in Berlin right now?"},
{"role": "assistant", "content": complete_blocks},
{"role": "user", "content": [{
"type": "tool_result",
"tool_use_id": tool_block.id,
"content": json.dumps(tool_result)
}]}
]
)
print("SUCCESS: Response received with thinking blocks included")
# Check if second response has thinking blocks
second_thinking = [b for b in complete_response.content if b.type == "thinking"]
second_text = [b for b in complete_response.content if b.type == "text"]
print(f"\nSecond response contains:")
print(f"- {len(second_thinking)} thinking blocks")
print(f"- {len(second_text)} text blocks")
if second_text:
print(f"\nFinal answer: {second_text[0].text}")
print("\nNote: The second response after tool use doesn't contain thinking blocks.")
print("This is expected behavior - thinking is shown before tool use but not after receiving tool results.")
except Exception as e:
print(f"ERROR: {e}")
# Uncomment to run the example
thinking_block_preservation_example()
=== INITIAL RESPONSE ===
Response contains:
- 1 thinking blocks
- 1 tool use blocks
Tool called: weather
Location to check: Berlin
Tool result: {'temperature': 60, 'condition': 'Foggy'}
=== TEST 1: WITHOUT thinking block ===
ERROR: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.1.content.0.type: Expected `thinking` or `redacted_thinking`, but found `tool_use`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking'}}
This demonstrates that thinking blocks must be preserved
=== TEST 2: WITH thinking block (correct approach) ===
SUCCESS: Response received with thinking blocks included
Second response contains:
- 0 thinking blocks
- 1 text blocks
Final answer: Currently in Berlin, it's foggy with a temperature of 60°F (about 15.5°C).
Note: The second response after tool use doesn't contain thinking blocks.
This is expected behavior - thinking is shown before tool use but not after receiving tool results.结论
本 notebook 展示了如何将 Claude 的扩展思考功能与工具调用相结合。主要优势包括：
- 在使用工具时透明地查看 Claude 的思考过程
- 了解 Claude 如何决定何时使用工具与使用内部知识
- 更好地理解涉及多次工具调用的多步骤任务
- 深入了解 Claude 如何解释工具结果并将其融入响应中
在使用扩展思考与工具时，请注意： - 为复杂任务设置适当的思考预算（最少 1,024 个 token） - 在传递工具结果时始终保留思考块及其签名 - 在对话历史中包含普通思考块和已编辑思考块 - 确保系统提示、工具和思考配置在调用之间保持一致 - 预期工具结果回合不会包含额外的思考块 - 工具调用与思考结合会增加 token 使用量和响应时间
Tip
中文读者提示：本章节的代码和输出为英文原文。如需理解具体实现细节，可参考上方中文导读和代码注释。如有疑问，欢迎在 GitHub Issues 讨论。