wrap_gemini: streamed responses are traced without their function calls or finish_reason #43

Tracing Method

• SDK/Client

Language

• Python

Host

• Hosted (https://api.smith.langchain.com)

Other System Info

• langsmith: main (14bb087), 0.14.11
• google-genai 2.31.0, Python 3.12, Linux

Description

With wrap_gemini, a streamed generate_content_stream call that ends in a function call is traced as plain text: the run output has no tool_calls and no finish_reason. The same response without streaming is traced with both.

_reduce_generate_content_chunks builds the streamed output only from chunk.text (L383-L384). In google-genai, .text returns only the text parts, so function_call parts never reach the run. The non-streamed path, _process_generate_content_response, does convert them (L276).

When the model only calls a tool and sends no text, the streamed run output is {"content": ""}.

Reproduction

No network or API key needed: real GenerateContentResponse objects, shaped like the chunks Gemini streams, go through the two output processors wrap_gemini uses.

from google.genai import types
from langsmith.wrappers import _gemini

text_chunk = {"candidates": [{"content": {"role": "model", "parts": [{"text": "Checking the weather. "}]}, "index": 0}]}
call_chunk = {
    "candidates": [{"content": {"role": "model", "parts": [
        {"function_call": {"name": "get_weather", "args": {"city": "Paris"}}}]},
        "finish_reason": "STOP", "index": 0}],
    "usage_metadata": {"prompt_token_count": 5, "candidates_token_count": 7, "total_token_count": 12},
}
chunks = [types.GenerateContentResponse.model_validate(c) for c in (text_chunk, call_chunk)]
full = types.GenerateContentResponse.model_validate({
    **call_chunk,
    "candidates": [{**call_chunk["candidates"][0], "content": {"role": "model", "parts": [
        text_chunk["candidates"][0]["content"]["parts"][0],
        call_chunk["candidates"][0]["content"]["parts"][0]]}}],
})

streamed = _gemini._reduce_generate_content_chunks(chunks)
non_streamed = _gemini._process_generate_content_response(full)
print("streamed:    ", {k: v for k, v in streamed.items() if k != "usage_metadata"})
print("non-streamed:", {k: v for k, v in non_streamed.items() if k != "usage_metadata"})

Output:

streamed:     {'content': 'Checking the weather. '}
non-streamed: {'content': 'Checking the weather. ', 'role': 'assistant', 'finish_reason': <FinishReason.STOP: 'STOP'>, 'tool_calls': [{'id': 'call_0', 'type': 'function', 'index': 0, 'function': {'name': 'get_weather', 'arguments': '{"city":"Paris"}'}}]}

The result is the same with chunks taken from a real generate_content_stream call (served by a local fake Gemini endpoint) instead of hand-built ones.

Expected behavior

The streamed run output matches the non-streamed one: role, finish_reason and tool_calls are present.

Proposed fix: in _reduce_generate_content_chunks, collect the parts of every chunk (text, function_call, thought) and the last non-empty finish_reason, then build the output with the same conversion _process_generate_content_response uses, e.g. by passing it a response whose candidate holds the merged parts.

Happy to open a PR with this and a test if that's useful.

This report was drafted with Claude Code, and the reproduction above was run on main to confirm the output.