Metadata (session_id / user_id / tags) not attached to parent trace with LangChain ReAct agent; only appears on child span

[eschullert]

Metadata (session_id / user_id / tags) not attached to parent trace with LangChain ReAct agent; only appears on child span

Summary
When instrumenting a simple LangChain chain and a built-in ReAct agent with LangfuseCallbackHandler, metadata passed via LangChain config.metadata (e.g., langfuse_session_id, langfuse_user_id, langfuse_tags, and custom fields) is not visible on the parent trace in Langfuse. Instead, I only see it on the second span/child span. The parent trace shows only OpenTelemetry resource attributes.

---

Environment

• Langfuse: Cloud
• Langfuse Python SDK: 3.2.4
• Python: 3.12.7

---

Minimal repro

agent = create_react_agent(
    model=model,
    tools=[get_weather],
    prompt=PROMPT
)

thread_id = "example-thread-1"
langfuse_handler = LangfuseCallbackHandler()

res = agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]},
    config={
        "run_name": "Weather Agent",
        "callbacks": [langfuse_handler],
        "configurable": {"thread_id": thread_id},
        "metadata": {
            "langfuse_session_id": thread_id,
            "langfuse_user_id": "user_456",
            "langfuse_tags": ["production", "humor-bot"],
            "custom_field": "additional metadata"
        }
    }
)

---

Expected behavior

• The parent trace in Langfuse should display:

  • session_id = example-thread-1
  • user_id = user_456
  • tags = ["production", "humor-bot"]
  • metadata.custom_field = "additional metadata"

---

Actual behavior

• Parent trace metadata only shows OpenTelemetry resource attributes; no session_id/user_id/tags/custom fields:

{
  "resourceAttributes": {
    "telemetry.sdk.language": "python",
    "telemetry.sdk.name": "opentelemetry",
    "telemetry.sdk.version": "1.36.0",
    "service.name": "unknown_service"
  },
  "scope": {
    "name": "langfuse-sdk",
    "version": "3.2.4",
    "attributes": { "public_key": "REDACTED" }
  }
}

• The second span has the metadata (session/thread/tags/custom):

{
  "custom_field": "additional metadata",
  "thread_id": "example-thread-1",
  "resourceAttributes": {
    "telemetry.sdk.language": "python",
    "telemetry.sdk.name": "opentelemetry",
    "telemetry.sdk.version": "1.36.0",
    "service.name": "unknown_service"
  },
  "scope": {
    "name": "langfuse-sdk",
    "version": "3.2.4",
    "attributes": { "public_key": "REDACTED" }
  }
}

---

What I’ve tried

• Passing metadata via config.metadata as shown above (results: only present on child span).
• I tried the solution in this discussion

[inkeep-ai-assistant[bot]]

When using the Langfuse Python SDK (v3) with LangChain, metadata such as langfuse_session_id, langfuse_user_id, and langfuse_tags can be attached to traces by passing them through the LangChain metadata argument. The callback handler is designed to map these fields to Langfuse observations, and they should be set on the trace if provided(1).

However, the documentation states these fields are handled as follows:

• langfuse_session_id: Sets the session ID for the trace
• langfuse_user_id: Sets the user ID for the trace
• langfuse_tags: Sets tags for the trace (list of strings)

If you observe these fields only on a child span and not on the parent trace, this may be because the parent trace is created outside of the Langfuse callback’s lifecycle or due to the order of how OpenTelemetry spans are set up in the LangChain agent's orchestration. The SDK expects the metadata to be available when the top-level trace is created, but if the parent trace is created before metadata is passed, it will not be updated retroactively(1).

Actionable steps:

• Make sure the metadata (with langfuse_session_id, etc.) is available at the very first call that triggers trace creation.
• If the orchestration creates a root span/trace before your config.metadata is applied, the metadata won’t be attached to the parent trace by default.
• There is no documented way in the SDK to retroactively update the parent trace’s metadata if it was not present at creation(1).

If you need to guarantee these fields are on the parent trace, you may need to explicitly create the trace and set the metadata using the SDK before invoking the agent, ensuring all subsequent spans and generations are children of that trace. There is no documented workaround in the SDK to patch or update the parent trace’s metadata after it is created(1).

📚 Sources:

1. Python SDK (v3)

---

Have another question? Just tag @inkeep.

[eschullert]

I'm only using the callback handler, so opening the first trace with the metadata should be done by the handler itself

[jannikmaierhoefer]

Hi @eschullert, could you please share the link to the trace where this is the case?

[jannikmaierhoefer]

Thanks for sharing. If I open a trace, I can see the session user ID and tags attached to the top-level trace (see screenshot). Is this not the case for you?

[thomasgtaylor]

@jannikmaierhoefer I'm also running into this problem! Great to see there's a discussion around it.

While the session + user id shows up on the parent trace, if you look at the Metadata section it doesn't have the custom field attached to it. If you click on the 2nd "Weather Agent", it does.

Top-most parent trace:

Second trace:

This prevents being able to search by metadata in the traces view.

[eschullert]

@jannikmaierhoefer it is as @thomasgtaylor said. The top-most parent trace include the Session Id and the User ID, but all of the custom metadata fields are not there.

[jannikmaierhoefer]

Thank you @eschullert and @thomasgtaylor, I agree it would be better if the metadata attached to the config of the langgraph invocation would be attached to the top-level span. Looping in @hassiebp to check if this is possible.

The second option to add attributes (including metadata) to a trace is by using an enclosing span. This should populate the metadata to the top-level trace (documented here):

from langfuse import get_client
from langfuse.langchain import CallbackHandler
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
 
langfuse = get_client()
langfuse_handler = CallbackHandler()
 
llm = ChatOpenAI(model_name="gpt-4o")
prompt = ChatPromptTemplate.from_template("Tell me a joke about {topic}")
chain = prompt | llm
 
# Set trace attributes dynamically via enclosing span
with langfuse.start_as_current_span(name="dynamic-langchain-trace") as span:
    span.update_trace(
        user_id="random-user",
        session_id="random-session",
        tags=["random-tag-1", "random-tag-2"],
        input={"animal": "dog"},
        metadata={"foo": "bar"}
    )
 
    response = chain.invoke({"topic": "cats"}, config={"callbacks": [langfuse_handler]})
 
    span.update_trace(output={"response": response.content})

Please let me know if this works in your case.

[eschullert]

This is working for me. Thank you.

[thomasgtaylor]

@jannikmaierhoefer Thank you for your response. While this does work, it creates another top-level trace and still omits the metadata on the 2nd node.

For context again, here's how it looked before:

Here's what it looks like now (notice the metadata is there, therefor we can filter for it now)

The 2nd span in the trace still does not contain it:

While this temporarily fixes it, it's certainly not ideal to create a duplicative span just for the metadata information.

[jannikmaierhoefer]

Thanks @eschullert, agree that this is not ideal. We have logged this issue and will check if metadata can be passed directly via the invocation as session and user id!

[MaryKuznet]

I had a similar problem and setting the update_trace flag helped me. I think this is the simplest solution.

langfuse_handler = CallbackHandler(update_trace=True)

[jannikmaierhoefer]

Thanks for sharing this @MaryKuznet!