bug: Issue rendering tool calls made with Langchain 1.0 #11850
Description
Describe the bug
Summary
After upgrading to LangChain 1.0, tools are no longer serialized in the expected LangGraph / OpenAI “function tool” format. As a result, downstream consumers (e.g. LangGraph adapters / tracing UIs) fail to recognize tools correctly unless custom patching is applied.
This appears to be caused by a change in how tools are represented internally and passed through callbacks.
Expected behavior
Tools should be serialized in a standard function-tool format that downstream systems can reliably detect and render, for example:
{
"type": "function",
"function": {
"name": "example_tool",
"description": "Example tool description",
"parameters": {
"type": "object",
"properties": {
"query": { "type": "string" }
},
"required": ["query"]
}
}
}
This format is relied upon by LangGraph-style adapters and tracing/observability tools to:
- Detect tool definitions
- Associate tool calls with tool results
- Render tools distinctly from normal messages
Actual behavior in Langchain 1.0
Actual behavior in LangChain 1.0
With LangChain 1.0, tools (e.g. StructuredTool) are now serialized roughly as:
{
"name": "example_tool",
"description": "Example tool description",
"args_schema": { ... },
"return_direct": false,
"verbose": false,
...
}
Notably:
- The type: "function" field is missing
- The function container is missing
- Tool schemas are exposed via args_schema instead of parameters
When these objects are forwarded as-is through callbacks (e.g. via invocation_params["tools"]), downstream systems do not recognize them as tools and instead treat them as generic JSON payloads.
Solution
I have temporarily patched the handler like so
class PatchedLangfuseCallbackHandler(CallbackHandler):
def on_llm_start(self, *args, **kwargs):
# 1) Shallow-copy kwargs so we don't mutate the original dict
patched_kwargs = dict(kwargs)
# 2) Copy invocation_params if present
invocation_params = patched_kwargs.get("invocation_params")
if isinstance(invocation_params, dict):
invocation_params = dict(invocation_params) # copy nested dict
patched_kwargs["invocation_params"] = invocation_params
tools = invocation_params.get("tools")
if isinstance(tools, list):
invocation_params["tools"] = [
self._serialize_tool(t) for t in tools
]
# 3) Call upstream logic with patched kwargs
return super().on_llm_start(*args, **patched_kwargs)
def on_chat_model_start(self, *args, **kwargs):
# 1) Shallow-copy kwargs so we don't mutate the original dict
patched_kwargs = dict(kwargs)
# 2) Copy invocation_params if present
invocation_params = patched_kwargs.get("invocation_params")
if isinstance(invocation_params, dict):
invocation_params = dict(invocation_params) # copy nested dict
patched_kwargs["invocation_params"] = invocation_params
tools = invocation_params.get("tools")
if isinstance(tools, list):
invocation_params["tools"] = [
self._serialize_tool(t) for t in tools
]
# 3) Call upstream logic with patched kwargs
return super().on_llm_start(*args, **patched_kwargs)
def _serialize_tool(self, tool: Any) -> Dict[str, Any]:
from langchain_core.tools import StructuredTool
"""
Output the exact "LangGraph tool definition" schema that your adapter expects:
{
"type": "function",
"function": { "name": ..., "description": ..., "parameters": ... }
}
"""
# Case 1: StructuredTool
if StructuredTool is not None and isinstance(tool, StructuredTool):
return self._serialize_structured_tool(tool)
# Case 2: APITool
if isinstance(tool, APITool):
return self._serialize_api_tool(tool)
# Last resort for unknown tool object
return {
"type": "function",
"function": {
"name": getattr(tool, "name", tool.__class__.__name__),
"description": getattr(tool, "description", None) or None,
"parameters": getattr(tool, "parameters", None) or {"type": "object", "properties": {}},
},
}
def _serialize_structured_tool(self, tool: StructuredTool) -> Dict[str, Any]:
name = getattr(tool, "name", "")
description = getattr(tool, "description", None)
parameters = tool.args
return {
"type": "function",
"function": {
"name": name,
**({"description": description} if description else {}),
"parameters": parameters,
},
}
Steps to reproduce
Use langchain-1.0 and langgraph with tool calling. When tool calling the invocation_parms should use Structured tools.
Langfuse Cloud or self-hosted?
Self-hosted
If self-hosted, what version are you running?
3.148
SDK and integration versions
3.11
Additional information
No response
Are you interested in contributing a fix for this bug?
No