normalize_v1_streamed_tool_calls: The Reliable Stream Fix

What it does

When you use LangChain’s OpenAI integration with tool calling in streaming mode, the model can emit tool call chunks in a non-deterministic order or with inconsistent formatting. The normalize_v1_streamed_tool_calls fix standardizes how these streamed events are normalized, ensuring tool arguments are correctly reconstructed even when the model emits partial or reordered chunks. Concretely, it patches the streaming path so AIMessageChunk.tool_calls is always a consistent list of ToolCall objects with the correct name, args, and id fields.

Why it matters

Streaming tool calls are critical in real-time agentic pipelines, such as a customer-service bot that streams a tool use while the user is still watching. If streamed tool calls are malformed or duplicated, downstream components that consume tool_calls directly (like a ReAct agent or a tool-executing loop) can crash or silently skip tool execution. This fix eliminates the most common source of streaming tool-call bugs in LangChain’s OpenAI integration.

Example

$ Accumulate streamed tool calls

Received tool_call chunk: ToolCall(name='get_weather', args='{"city": "Tokyo"}', id='...')

Total tool calls received: 1

Common flags

AIMessageChunk.tool_calls

Normalized list of ToolCall objects on streaming chunks

PydanticStructurableMixin

Enables Pydantic models as tool schemas, benefiting from normalization