[bug][openai-agents] LLM span GENERATED section shows {"role": "assistant"} with no content

[syedmohsinbukhari]

Bug Report

Describe the bug

Two related issues when instrumenting the OpenAI Agents SDK (Responses API path) with OpenAIAgentsInstrumentor:

Issue 1 — LLM span output missing message content

The [LLM] span's output shows only [{"role": "assistant"}] with no message content, despite the model returning a valid text response.
Root cause (from source inspection of _processor.py): _get_attributes_from_message correctly emits llm.output_messages.N.message.contents.I.message_content.text for the response text. However, OUTPUT_VALUE is set via response.model_dump_json() which serialises the full Response object. Downstream consumers (e.g. LangWatch UI) that read from a flat message.content field find nothing, because output messages use the nested multi-part message.contents.* form rather than a flat message.content string — unlike input messages which do emit a flat message.content.

Issue 2 — Reasoning items silently dropped

When agents are configured with ModelSettings(reasoning=Reasoning(...)), the reasoning summary returned by the model never appears in the trace. In _get_attributes_from_response_output, "reasoning" items are explicitly listed in the unhandled branch with a TODO comment and produce zero span attributes:

elif item.type in (
    "file_search_call", "web_search_call", "computer_call",
    "reasoning",  # silently skipped
    ...
):
    ...  # TODO

To Reproduce

1. Instrument any OpenAI Agents SDK app with OpenAIAgentsInstrumentor
2. Configure an agent with ModelSettings(reasoning=Reasoning(effort="low", summary="detailed"))
3. Run a simple text reply (no tools)
4. Inspect the LLM span output in any OTel-compatible UI (LangWatch, etc.)
5. The GENERATED / output section shows [{"role": "assistant"}] — role only, no content
6. No reasoning summary appears anywhere in the trace

Expected behavior

• The LLM span output should include the response text, e.g. [{"role": "assistant", "content": "Hello! How can I help?"}]
• Reasoning summaries produced by the model should appear as span attributes
  Actual behavior
• [{"role": "assistant"}] — role only, no content
• Reasoning items are silently dropped with no attributes emitted

Desktop

• OS: MacOS
• Version: 26.3.1

[dosubot]

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

Thanks for the detailed bug report! Your root cause analysis is correct.

The issue is in _processor.py — output messages always use the nested MESSAGE_CONTENTS.{i}.* structure, even for simple text:

for i, item in enumerate(obj.content):
    if isinstance(item, ResponseOutputText):
        yield f"{prefix}{MESSAGE_CONTENTS}.{i}.{MESSAGE_CONTENT_TYPE}", "text"
        yield f"{prefix}{MESSAGE_CONTENTS}.{i}.{MESSAGE_CONTENT_TEXT}", item.text

Meanwhile, input messages correctly check content type and emit flat MESSAGE_CONTENT for strings:

if isinstance(content, str):
    yield f"{prefix}{MESSAGE_CONTENT}", content  # Flat structure

Per the OpenInference semantic conventions, MESSAGE_CONTENT should be used for simple string content, while MESSAGE_CONTENTS is for multimodal/structured arrays. Other instrumentations (OpenAI, Anthropic, LiteLLM) all follow this pattern.

The fix would be to check if output content is a single text item and emit MESSAGE_CONTENT directly in that case, falling back to MESSAGE_CONTENTS for multimodal content.

To reply, just mention @dosu.

---

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

   

[syedmohsinbukhari]

Thanks @dosu — that matches my reading of the code. The fix makes sense: emit flat MESSAGE_CONTENT when the output is a single ResponseOutputText, fall back to MESSAGE_CONTENTS.* for multimodal. Would you like me to open a PR with this fix, or is the team picking it up?

[dosubot]

A PR from you would be welcome! Community contributions are great, and since you've already done the root cause analysis, you're well-positioned to implement the fix. Just make sure to follow the existing patterns from the other instrumentations (like the OpenAI one) for consistency.

If the maintainers have other plans or want to handle it differently, they can chime in here.

To reply, just mention @dosu.

---

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

   

[syedmohsinbukhari-chaos]

@dosu I will pick it up from my company account then. So I will go ahead and create the PR.