[openai] Responses API drops function_call and function_call_output from gen_ai.input.messages

[sfc-gh-zeningchen]

Describe your environment

OS: Linux
Python version: Python 3.14.7
Package version: opentelemetry-instrumentation-genai-openai 1.2b0.dev
GenAI library and version: openai 2.47.0

What happened?

On the Responses API path, get_input_messages skips any input item without a string role. Responses input arrives as a flat item list, so function_call and function_call_output items carry no role and never reach gen_ai.input.messages.

A model call that follows a tool call therefore records only the original user message, and no tool_call_response part is emitted anywhere. The recorded conversation is not just incomplete, it misleads: the span shows the model answering with no visible tool result, and the recorded input token count contradicts the recorded message list.

Steps to Reproduce

Run a two-turn tool loop with the plain openai SDK, no agent framework, with content capture enabled:

first = client.responses.create(
model="gpt-4o-mini",
input=[{"role": "user", "content": "What's the weather in Seattle?"}],
tools=[weather_tool],
)

history = [{"role": "user", "content": "What's the weather in Seattle?"}]
history.extend(first.output) # the function_call item
history.append({
"type": "function_call_output",
"call_id": first.output[0].call_id,
"output": "Sunny, 18C",
})

client.responses.create(model="gpt-4o-mini", input=history, tools=[weather_tool])

Then compare the second request body against the second span's gen_ai.input.messages.

Expected Result

Three messages, matching the three items sent on the wire: the user turn, an assistant message holding a tool_call part, and a tool message holding a tool_call_response part whose id matches the tool_call.

This is the shape the semantic conventions show for a tool loop, and opentelemetry-instrumentation-genai-anthropic already records it that way given an equivalent history.

Actual Result

One message. Both tool items are dropped:

wire request #2: input items=3 types=['user', 'function_call', 'function_call_output']
span : gen_ai.input.messages = 1 msg
[{"role":"user","parts":[{"content":"What's the weather in Seattle?","type":"text"}]}]
gen_ai.usage.input_tokens = 129

129 input tokens against a single short user message.

Additional context

No response

Would you like to implement a fix?

Yes

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}