[bug] Incomplete spans exported to Phoenix after upgrading pydantic-ai from 0.8.1 to 1.0.0

[mridul-aiXbrain]

Describe the bug
After upgrading pydantic-ai from 0.8.1 to 1.0.x, spans exported to Phoenix via openinference-instrumentation-pydantic-ai become incomplete.

• With pydantic-ai==0.8.1 + openinference-instrumentation-pydantic-ai==0.1.5, spans look correct (instructions, user prompts, model info, tool calls, etc.).
• With pydantic-ai>=1.0.0 + the same instrumentation, spans degrade:
  • Some attributes are missing (instructions, tool call details, etc.).
  • Overall trace structure is much less informative compared to <1.0.0.

To Reproduce
pydantic-ai==1.0.10
openinference-instrumentation-pydantic-ai==0.1.5
arize-phoenix-otel==0.13.1
opentelemetry-sdk==1.37.0

self hosted phoenix v11.32.1

1. Run Phoenix locally with self-hosted setup
2. Connect Phoenix to a pydantic-ai Agent with OpenAIResponsesModel
3. Add a tool to the agent
4. Get agent response using agent.run_stream()
5. Verify the span in Phoenix UI

Expected behavior
with pydantic-ai >= 1.0.0, spans should not degrade.

Screenshots
with pydantic v0.8.1

with pydantic-ai >= v1.0.0

Desktop (please complete the following information):

• OS: Ubuntu
• Version 22.04
• Python 3.11.7

[dosubot]

Hi @mridul-aiXbrain! I'm Dosu and I’m helping the openinference team.

This issue is caused by changes in how pydantic-ai ≥1.0.0 structures span attributes, which the instrumentation relies on to extract details like instructions and tool calls. The openinference-instrumentation-pydantic-ai package was updated in version 0.1.5 to support the new tracing format, but you must update your agent setup to use InstrumentationSettings(version=2) from pydantic_ai.models.instrumented instead of instrument=True for compatibility with pydantic-ai ≥1.0.0. Without this, the instrumentation can't extract the expected fields, resulting in incomplete spans. See the updated usage example in the README.

If you have already set InstrumentationSettings(version=2) and still see missing attributes, double-check that your agent/tool configuration matches the new tracing format and that all expected keys are present in the span attributes. The extraction logic is tightly coupled to these keys—any mismatch or missing data from pydantic-ai will lead to degraded traces in Phoenix.

To reply, just mention @dosu.

---

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

[mridul-aiXbrain]

@dosu
i already have the setup that matches the README that you suggested, but the issue of degraded traces still persists.

def setup_tracing(service_name: str) -> TracerProvider:
    if not sentry_sdk.Hub.current.client:
        if cached_environ().sentry_dsn:
            sentry_sdk.init(
                dsn=cached_environ().sentry_dsn,
                environment="dev",
                release="0.1.0",
                send_default_pii=False,
                integrations=[
                    StarletteIntegration(),
                    FastApiIntegration(),
                    LoggingIntegration(
                        level=logging.INFO, event_level=logging.ERROR
                    ),
                ],
                instrumenter="otel",
                traces_sample_rate=None,
            )

    current_provider = trace.get_tracer_provider()
    if isinstance(current_provider, TracerProvider):
        return current_provider

    phoenix_api_key = cached_environ().phoenix_api_key

    headers = (
        {"Authorization": f"Bearer {phoenix_api_key}"}
        if phoenix_api_key
        else {}
    )
    endpoint = cached_environ().phoenix_collector_endpoint

    resource = Resource.create(
        {
            SERVICE_NAME: service_name,
            # add more resource attrs here if needed
        }
    )

    trace_provider = TracerProvider(resource=resource)
    exporter = OTLPSpanExporter(endpoint=endpoint, headers=headers)

    trace_provider.add_span_processor(OpenInferenceSpanProcessor())
    trace_provider.add_span_processor(SimpleSpanProcessor(exporter))
    trace_provider.add_span_processor(SentrySpanProcessor())

    trace.set_tracer_provider(trace_provider)
    return trace_provider

trace_provider = setup_tracing(service_name="demo")
model = OpenAIResponsesModel(
            model_name="gpt-4o",
            provider=OpenAIProvider(
                openai_client=AsyncOpenAI(
                    api_key="fake key"
                    project="project",
                ),
            ),
        )

agent = Agent(
            model,
            deps_type=[AgentDeps],
            system_prompt=("Respond to prompt to best of your ability"),
            instructions=AGENT_INSTRUCTIONS,
            tools=[get_weather],
            instrument=InstrumentationSettings(tracer_provider=trace_provider, version=2),
        )

agent.run_stream(prompt, deps=deps, message_history=message_history)

[dosubot]

Your setup matches the current recommendations, but there are a few known issues and troubleshooting steps that may help:

• The extraction logic in openinference-instrumentation-pydantic-ai for pydantic-ai ≥1.0.0 expects certain keys (like instructions, tools, tool call details, etc.) in the span attributes. If these are missing or structured differently, the resulting spans will be incomplete, and the processor will not fail loudly—just silently degrade the trace detail. There’s no comprehensive debug output by default, so missing fields can be hard to spot without inspecting the raw span payloads or adding custom logging to the extraction logic [source].

• A known quirk: the input.value attribute is set to the first user message in the message_history, not the current prompt. To ensure the correct input is captured, append the current prompt as the most recent user message in message_history before calling agent.run_stream [source].

• If you’re seeing JSON metadata but not explicit Inputs/Outputs in Phoenix, check the raw span payloads for input.value and output.value at the top level. If they’re missing, the issue is upstream in how the data is structured or extracted; if present but not showing in the UI, the issue is likely with Phoenix ingestion or display [source].

• There are no environment variables or config switches that would hide these fields by default; their presence depends on the upstream data and extraction logic.

If you’ve confirmed your agent/tool config matches the new tracing format and you’re still missing attributes, try adding debug prints or logging in the OpenInferenceSpanProcessor or extraction code to surface what’s actually being processed. This can help pinpoint if the expected keys are present and correctly structured before export.

Let me know if inspecting the raw span attributes or tweaking the message history resolves the issue, or if you spot any patterns in what’s missing!

To reply, just mention @dosu.

---

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

[caroger]

@mridul-aiXbrain , thanks for filing the issue and providing the code snippet. Our team will take a closer look at it and get it back to you.