[bug] openinference-instrumentation-openai-agents: streaming responses missing output, model, and token usage

[polae]

Description

When using openinference-instrumentation-openai-agents (v1.4.0) with Langfuse, streaming responses from agents do not capture output data, model name, or token usage. Non-streaming/short responses work correctly.

Environment

• openinference-instrumentation-openai-agents: 1.4.0
• opentelemetry-sdk: 1.39.1
• Langfuse (via OpenTelemetry exporter)
• Python 3.13
• OpenAI Agents SDK with Runner.run_streamed()

Steps to Reproduce

1. Create an agent that produces a long streaming text response:

from agents import Agent, Runner

agent = Agent(
    name="MyAgent",
    instructions="Write a detailed story...",
    model="gpt-4o",
)

result = Runner.run_streamed(agent, input_items, context=context)
async for event in result.stream_events():
    # Process streaming events
    pass

2. Check the resulting GENERATION observation in Langfuse

Expected Behavior

The GENERATION observation should contain:

• model: The model name (e.g., "gpt-4o")
• output: The complete response content
• usage: Token counts (input, output, total)
• calculatedTotalCost: Computed cost

Actual Behavior

For streaming responses (long text output, ~7 seconds):

Name: response
Type: GENERATION
Model: None
Tokens: input=0, output=0, total=0
Cost: 0
Has Output: False
modelParameters: {}
usageDetails: {}

For non-streaming/short responses (structured output, ~1 second):

Name: response
Type: GENERATION
Model: gpt-5.1-2025-11-13
Tokens: input=623, output=27, total=650
Cost: 0.00104875
Has Output: True
modelParameters: {full data}
usageDetails: {detailed usage}

Analysis

The key difference is response duration/streaming behavior:

• Working: Agent with output_type= (structured output) that completes quickly (~941ms)
• Broken: Agent streaming long text narrative (~6919ms)

Both agents use identical code patterns (Runner.run_streamed() + async iteration over stream_events()), but the instrumentation only captures data for the fast/structured completions.

Related Issues

• [bug] Agno OpenInference instrumentation: output.value missing in streaming mode ⇒ "undefined" in Langfuse #2481: "[bug] Agno OpenInference instrumentation: output.value missing in streaming mode"
• [BUG] Anthropic streaming wrapper returns wrong object, breaking get_final_message() #2467: "[BUG] Anthropic streaming wrapper returns wrong object"

This appears to be part of a broader pattern where streaming mode causes telemetry data loss across multiple integrations.

Workaround

Currently investigating manual capture of usage data from result.raw_responses after streaming completes.

[dosubot]

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

This is a known limitation with streaming in openinference-instrumentation-openai-agents: output, model, and token usage are often missing because the OpenAI Agents SDK does not populate usage data or output fields by default during streaming. The instrumentation only captures these fields if the final response object includes them when the span ends. For token usage, you must explicitly request it by passing stream_options={"include_usage": True} in your streaming API call; otherwise, usage data will not be included and telemetry tools like Langfuse or Phoenix will show zero tokens and cost (source, source, source).

For missing output, the root cause is that the instrumentation only sets output.value if RunOutput.content is non-None, but in streaming mode, content can be incomplete or None while the real output is available via other fields. A workaround is to explicitly pass yield_run_output=True when running your agent in streaming mode—this ensures the final output is yielded and traced (source). There are open PRs to improve this behavior, but until they are merged and released, these workarounds are necessary.

If you need to capture usage or output after streaming completes, you can manually extract this data from result.raw_responses or by collecting the streamed events and aggregating the output and usage yourself.

To reply, just mention @dosu.

---

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