pydantic_ai integration: auto-populate LLMObs session_id from conversation_id

[mlorthiois]

Package Name

pydantic_ai

Package Version(s)

No response

Describe the goal of the feature

The pydantic_ai LLMObs integration (ddtrace/contrib/internal/pydantic_ai/patch.py) traces Agent.iter and
Agent.run_stream, both of which receive an optional conversation_id: str | None keyword argument (see
pydantic_ai/agent/__init__.py, Agent.iter signature). This conversation_id is exactly the semantic equivalent of LLMObs' session_id: it groups multiple agent runs that belong to the same end-user conversation/thread.

Today, traced_agent_iter and traced_agent_run_stream never read this argument, so it's silently dropped.

The only way to get session correlation in LLM Observability is to wrap the call in an explicit
LLMObs.workflow(session_id=...) / LLMObs.agent(session_id=...) context manager, duplicating a value the
integration already has in kwargs at interception time:

  def traced_agent_run_stream(func, instance, args, kwargs):
      integration = pydantic_ai._datadog_integration
      integration._run_stream_active = True
      span = integration.trace(
          _agent_resource(instance),
          span_name="pydantic_ai.agent",
          submit_to_llmobs=True,
          model=getattr(instance, "model", None),
          kind="agent",
      )
      ...

A solution: When conversation_id is present in kwargs for Agent.iter / Agent.run_stream, pass it through to _annotate_llmobs_span_data as session_id, e.g. in PydanticAIIntegration._llmobs_set_tags_agent :

  conversation_id = kwargs.get("conversation_id")
  _annotate_llmobs_span_data(
      span,
      name=agent_name or "PydanticAI Agent",
      input_value=user_prompt,
      output_value=result,
      session_id=conversation_id,
  )

Is your feature request related to a problem?

No response

Describe alternatives you've considered

No response

Additional context

No response