langchain - 💡(How to fix) Fix Duplicate tool_calls in content_blocks with OpenAI Responses API streaming [1 comments, 2 participants]

"""
Minimal reproduction: content_blocks includes duplicate tool_calls
when using OpenAI Responses API streaming.

No API key needed - demonstrates the core issue.
"""

from langchain_core.messages import AIMessage, ToolMessage


def main():
    # Simulates what OpenAI Responses API streaming produces:
    # - content: accumulates fc_ prefixed blocks (partial, empty args)
    # - tool_calls: final call_ prefixed blocks (full args)
    
    simulated_content = [
        {"type": "text", "text": "Let me check the weather."},
        # fc_ blocks accumulated during streaming (partial, no args)
        {"type": "tool_call", "id": "fc_12345", "name": "get_weather", "args": {}},
        {"type": "tool_call", "id": "fc_67890", "name": "get_time", "args": {}},
    ]
    
    # Final tool_calls with call_ prefix and full args
    simulated_tool_calls = [
        {"id": "call_abc123", "name": "get_weather", "args": {"city": "Tokyo"}},
        {"id": "call_def456", "name": "get_time", "args": {"timezone": "JST"}},
    ]
    
    msg = AIMessage(content=simulated_content, tool_calls=simulated_tool_calls)
    
    print(f"content: {len(msg.content)} blocks")
    print(f"content_blocks: {len(msg.content_blocks)} blocks")
    print(f"tool_calls property: {len(msg.tool_calls)} entries")
    
    print("\nTool calls in content_blocks:")
    for block in msg.content_blocks:
        if isinstance(block, dict) and block.get("type") == "tool_call":
            print(f"  id={block.get('id')}, args_empty={not bool(block.get('args'))}")
    
    # Serialize using content_blocks (recommended pattern)
    serialized = {"type": msg.type, "content_blocks": list(msg.content_blocks)}
    
    # Deserialize
    reloaded = AIMessage(content_blocks=serialized["content_blocks"])
    
    print(f"\nAfter reload - tool_calls: {len(reloaded.tool_calls)} entries")
    for tc in reloaded.tool_calls:
        print(f"  id={tc.get('id')}, has_args={bool(tc.get('args'))}")
    
    # The bug: fc_ blocks have no matching ToolMessage responses
    tool_responses = [
        ToolMessage(content="Sunny", tool_call_id="call_abc123"),
        ToolMessage(content="14:30", tool_call_id="call_def456"),
    ]
    response_ids = {tm.tool_call_id for tm in tool_responses}
    missing = [tc for tc in reloaded.tool_calls if tc.get("id") not in response_ids]
    
    print(f"\nMissing responses for: {[tc.get('id') for tc in missing]}")
    print("API error: 'No tool output found for function call fc_12345'")


if __name__ == "__main__":
    main()

---

content: 3 blocks
content_blocks: 5 blocks
tool_calls property: 2 entries

Tool calls in content_blocks:
  id=fc_12345, args_empty=True
  id=fc_67890, args_empty=True
  id=call_abc123, args_empty=False
  id=call_def456, args_empty=False

After reload - tool_calls: 4 entries
  id=fc_12345, has_args=False
  id=fc_67890, has_args=False
  id=call_abc123, has_args=True
  id=call_def456, has_args=True

Missing responses for: ['fc_12345', 'fc_67890']
API error: 'No tool output found for function call fc_12345'

---

# Current code dedupes by ID only
content_tool_call_ids = {
    block.get("id")
    for block in self.content
    if isinstance(block, dict) and block.get("type") == "tool_call"
}
for tool_call in self.tool_calls:
    if (id_ := tool_call.get("id")) and id_ not in content_tool_call_ids:
        blocks.append(tool_call_block)  # fc_xxx != call_xxx, so BOTH added

---

# In content_blocks property, after getting blocks from super()
blocks = [
    block for block in blocks
    if not (
        isinstance(block, dict) and
        block.get("type") == "tool_call" and
        not block.get("args")
    )
]

---

System Information
------------------
> OS:  Darwin
> OS Version:  Darwin Kernel Version 25.2.0
> Python Version:  3.13.11

Package Information
-------------------
> langchain_core: 1.3.1
> langchain: 1.2.15
> langchain_openai: 1.2.0
> langsmith: 0.7.35
> openai: 2.32.0