Integrating LangFuse with OpenAI Agent SDK drops reasoning summaries?

[lshalon]

Describe your question

Integration Pattern

1. OpenAIAgentsInstrumentor().instrument() — auto-instruments all agent runs via OTel
2. @observe() decorator on the top-level run_domino_chain() function — creates a
   Langfuse-native trace
3. langfuse.get_prompt("domino-model/manager-instruction", label="production").compile()
   — versioned prompt management, 10+ prompts managed in LangFuse
4. langfuse.openai.OpenAI — instrumented OpenAI client for auxiliary LLM calls
   (context window reduction/summarization within a custom SessionABC implementation)
5. langfuse.get_trace_url() — trace URLs stored in databases for linking business
   records back to LangFuse traces

Issue: Reasoning Content Text Not Captured (Primary)

What happens: GPT-5.4 with Reasoning(effort="high") produces reasoning items in the conversation history:

{
    "id": "rs_00b3fcaccac6020c...",
    "summary": [
        {
            "text": "I'm thinking about setting the year to null...",
            "type": "summary_text"
        }
    ],
    "type": "reasoning"
}

Our custom session correctly handles these (they're tracked as context-providing messages for summarization decisions). But LangFuse traces do not surface the reasoning text in any structured or visible way.

What we need: The reasoning summary text (which OpenAI returns when Reasoning(summary="detailed") is set) should appear in the LangFuse trace UI — either as a dedicated field on the generation observation, a nested span, or at minimum in the observation metadata. This is critical for debugging agent decisions.

Langfuse Cloud or Self-Hosted?

Langfuse Cloud

If Self-Hosted

No response

If Langfuse Cloud

https://cloud.langfuse.com/project/cm0pekufh000m4hgzdv9svzrq/traces?filter=type%3BstringOptions%3B%3Bany+of%3BAGENT&peek=ad4e3d82ccb3aff0&observation=94fb3287ecb7df5e&traceId=f21084576bb848ae2a887270fa4d4fa9&timestamp=2026-03-28T18%3A59%3A11.719Z

SDK and integration versions

Stack

• openai-agents v0.6.1 + openinference-instrumentation-openai-agents v1.3.1
• GPT-5.4 with Reasoning(effort="high") as the primary model
• langfuse Python SDK for prompt management + @observe() decorator
• LangFuse Cloud (EU)

Pre-Submission Checklist

• I have checked for existing issues/discussions and consulted Langfuse AI.

[inkeep-ai-assistant[bot]]

Reasoning summaries from OpenAI models are captured by Langfuse, but there was a known issue with how they were displayed in the platform's "formatted" view. (1)

The reasoning summary data does get stored in the observation output. When viewing a generation span in JSON mode, you can see the full reasoning structure including the summary text. (1) The issue was specifically with the "formatted" view rendering—it applied markdown parsing to structured JSON, causing poor readability and bolding artifacts inside JSON blobs. (1)

This formatting issue has since been fixed. A user confirmed in the GitHub issue that "it looks like this has been fixed" when checking the cloud platform. (1)

To verify your reasoning data is being captured:

1. Navigate to your trace in Langfuse
2. Select the generation observation
3. Switch to "JSON" view mode instead of "formatted" view
4. Look for the output field containing the reasoning structure with summary array

If you're not seeing reasoning content at all (even in JSON view), the issue may be with how the OpenAI Agents Instrumentor is serializing the response. Some attributes may be stored in the metadata object rather than being mapped to the Langfuse data model directly. (2)

If the mapping or integration doesn't work as expected, you can raise an issue on GitHub. (2)

📚 Sources:

1. OpenAI, generation spans: reasoning summaries badly formatted in platform
2. Parallel Integration

---

Have another question? Just tag @inkeep.