[bug] pydantic-ai: output.value not set for plain text (str) agent responses

[ritiztambi]

Description

The OpenInferenceSpanProcessor for PydanticAI does not set output.value when the agent returns plain text (output_type=str, the default). It only sets output.value when a final_result tool call is present (structured output mode).

This causes downstream consumers that rely on output.value (e.g. Langfuse OTLP ingestion) to show null for the output field on LLM generation spans.

Steps to reproduce

1. Create a PydanticAI agent with default output_type=str
2. Instrument with OpenInferenceSpanProcessor
3. Run the agent — LLM responds with plain assistant text
4. Inspect the exported span attributes

Expected: output.value is set to the assistant's text response
Actual: output.value is missing; only llm.output_messages.0.message.content is set

Root cause

In semantic_conventions.py, _extract_from_gen_ai_messages() processes gen_ai.output.messages:

• For text parts (line ~824): yields llm.output_messages.*.message.content but does NOT set output_value
• For tool call parts with name final_result (line ~842): sets output_value ✅

So output.value is only emitted for structured output agents. Plain text agents (the default) get no output.value.

Note that input.value IS correctly set from the last user text part in gen_ai.input.messages — the output side is just missing the equivalent logic.

Environment

• openinference-instrumentation-pydantic-ai==0.1.12
• pydantic-ai==1.70.0

[dosubot]

Hi @ritiztambi! I'm Dosu and I’m helping the openinference team.

Good find — I can confirm this bug. Looking at _find_llm_output_value(), the issue is that when MESSAGE_CONTENT exists on the output message, the function checks if it's a string and returns it if so, but if it's not a string, it returns None immediately — which also blocks the fallback to final_result tool calls. More importantly for your specific case (plain text str agents), the function should be returning the string content but doesn't seem to be reaching this code path at all if MESSAGE_CONTENT isn't present on the output message dict.

The structure of the function is:

def _find_llm_output_value(output_messages):
    for message in output_messages:
        if MessageAttributes.MESSAGE_CONTENT in message:
            content = message[MessageAttributes.MESSAGE_CONTENT]
            if isinstance(content, str):
                return content
            return None  # early return blocks tool call fallback

        if MessageAttributes.MESSAGE_TOOL_CALLS in message:
            # ... check for final_result

For comparison, _find_llm_input_value() has the same early-return pattern but works for inputs because user messages reliably have string content.

There are a couple of related issues that have been filed around output.value extraction [1] [2], and some recent PRs have addressed adjacent problems [3] [4], but none specifically fix the plain-text output case you're describing.

A minimal fix would be to remove the early return None on line 555 so that the function continues to check tool calls when MESSAGE_CONTENT is non-string, and — more relevant to your case — to ensure the output message dict actually contains MESSAGE_CONTENT for plain text responses. It may be worth checking whether the upstream message extraction (in _extract_output_messages()) is populating MESSAGE_CONTENT correctly for assistant text parts.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}