bug(langfuse): content tracing fails Agent runs when a tool returns None, a string, or an image

Describe the bug

With HAYSTACK_CONTENT_TRACING_ENABLED=true and LangfuseConnector, LangfuseSpan.set_content_tag can raise inside the traced run, so the whole Agent or Pipeline run fails. With tracing off, the same run succeeds.

Since Haystack 3.0 (deepset-ai/haystack#11892), the Agent creates one span per tool call and passes the raw tool arguments and result to set_content_tag. The Langfuse tracer assumes that every *.input value containing "messages" is a dict of ChatMessages, and that every *.output value containing "replies" is a dict. Values that don't fit this break it:

To Reproduce

This needs only haystack-ai and langfuse-haystack. Any Langfuse keys work: the error doesn't depend on spans actually being exported. A scripted chat generator stands in for a real one so that no LLM API key is needed. OpenAIChatGenerator behaves the same way.

import os

os.environ["HAYSTACK_CONTENT_TRACING_ENABLED"] = "true"
os.environ.setdefault("LANGFUSE_PUBLIC_KEY", "pk-lf-...")
os.environ.setdefault("LANGFUSE_SECRET_KEY", "sk-lf-...")

from haystack import component
from haystack.components.agents import Agent
from haystack.dataclasses import ChatMessage, ToolCall
from haystack.tools import tool

from haystack_integrations.components.connectors.langfuse import LangfuseConnector


@component
class ScriptedChatGenerator:
    """Calls the tool once, then answers."""

    def __init__(self):
        self.calls = 0

    @component.output_types(replies=list[ChatMessage])
    def run(self, messages: list[ChatMessage], tools=None, **kwargs):
        self.calls += 1
        if self.calls == 1:
            return {"replies": [ChatMessage.from_assistant(tool_calls=[ToolCall(tool_name="notify", arguments={"to": "bob"}, id="call_1")])]}
        return {"replies": [ChatMessage.from_assistant("Done")]}


@tool
def notify(to: str) -> None:
    """Send a notification to a user."""


LangfuseConnector("repro").warm_up()
agent = Agent(chat_generator=ScriptedChatGenerator(), tools=[notify])
print(agent.run(messages=[ChatMessage.from_user("Tell Bob")])["last_message"].text)
# -> TypeError: argument of type 'NoneType' is not iterable

Expected behavior

Tracing never makes a run fail. Values that aren't chat messages are recorded with tracing_utils.coerce_tag_value, like other tags.

Describe your environment (please complete the following information):

• OS: macOS
• Haystack version: haystack-ai 3.1.1
• Integration version: langfuse-haystack 6.0.0 (also main); langfuse 4.15.2