[BUG]: INPUT_VALUE not set when using message_history (Pydantic AI)

[fyhuang]

Tracer Version(s)

4.3.2

Python Version(s)

3.14.2

Pip Version(s)

uv 0.9.8

Bug Report

The PydanticAI LLMObs integration extracts user input solely from the user_prompt argument (first positional arg to Agent.run()/Agent.iter()). When messages are passed via message_history instead, INPUT_VALUE on the agent span is None.

This affects PydanticAI's VercelAIAdapter (and any other caller that uses message_history instead of user_prompt). The adapter converts incoming messages into PydanticAI ModelMessage objects and passes them as message_history — it never sets user_prompt.

Reproduction Code

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "pydantic-ai[openai]>=1.29.0",
#     "ddtrace>=4.3.0",
# ]
# ///
"""Repro: ddtrace LLMObs doesn't tag user input when using VercelAIAdapter.

VercelAIAdapter passes all messages via `message_history`, never setting
`user_prompt`. ddtrace's PydanticAI integration only reads `user_prompt`,
so INPUT_VALUE on the agent span is always None.

Requires: OPENAI_API_KEY, DD_API_KEY, DD_SITE

Example: uv run repro_ddtrae_pydantic_input.py
"""

import asyncio
import os

os.environ.setdefault("DD_LLMOBS_AGENTLESS_ENABLED", "1")
os.environ.setdefault("DD_LLMOBS_ML_APP", "ddtrace-pydantic-input-bug-repro")

from ddtrace.llmobs import LLMObs

LLMObs.enable()

# Diagnostic hook: print INPUT_VALUE after ddtrace tags the agent span
from ddtrace.llmobs._constants import INPUT_VALUE
from ddtrace.llmobs._integrations.pydantic_ai import PydanticAIIntegration

_original = PydanticAIIntegration._llmobs_set_tags_agent


def _diagnostic(self, span, args, kwargs, response):
    _original(self, span, args, kwargs, response)
    print(f"INPUT_VALUE on agent span: {span._get_ctx_item(INPUT_VALUE)!r}")


PydanticAIIntegration._llmobs_set_tags_agent = _diagnostic

from pydantic_ai import Agent
from pydantic_ai.ui.vercel_ai._adapter import VercelAIAdapter
from pydantic_ai.ui.vercel_ai.request_types import SubmitMessage, TextUIPart, UIMessage

agent = Agent("openai:gpt-4o-mini", instructions="Reply in one sentence.")


async def main():
    # Simulate a Vercel AI SDK request (as a frontend would send).
    # VercelAIAdapter converts this into message_history, never setting user_prompt.
    # Expected: INPUT_VALUE = "What is 2 + 2?"
    # Actual:   INPUT_VALUE = None
    request = SubmitMessage(
        trigger="submit-message",
        id="msg-1",
        messages=[
            UIMessage(
                id="msg-1",
                role="user",
                parts=[TextUIPart(type="text", text="What is 2 + 2?")],
            ),
        ],
    )
    adapter = VercelAIAdapter(agent, request)
    async for event in adapter.run_stream():
        pass
    LLMObs.flush()


asyncio.run(main())

Error Logs

No response

Libraries in Use

• pydantic-ai: 1.29.0

Operating System

Darwin Kernel Version 25.2.0

CC @gleb-fractional