šŸ› Bug Report: Gemini streaming raises TypeError after the last chunk when usage metadata has no candidates_token_count

Open
#4,711 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

@prassu017 is already working on this.

Since Nov 17, 2026.

  • #4712 by @prassu017 — open

Assessment

Difficulty
3/5
Estimated time
1-2 days
Newbie friendliness
35/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Stale
Tech stack
python

Research direction

Start with the usage calculation and histogram recording in packages/opentelemetry-instrumentation-google-generativeai/opentelemetry/instrumentation/google_generativeai/span_utils.py and the sync and async stream finalization in __init__.py. Reproduce the issue with the synthetic response described here, then check the Google Generative AI instrumentation tests for coverage of both wrappers. Done means streams finish normally and spans end when candidate token counts are absent; linked PR #4712 is already open.

Written by the indexing model from the issue text.

Description

Hullo folks,

I’m doing some research for Tessl Code Review and we ran it against your repo. It flagged a potential issue, and I figured it made sense to share it with you:

Which component is this bug for?

Google Generative AI Instrumentation

šŸ“œ Description

With metrics enabled, a Gemini stream whose last chunk has usage metadata but no candidates_token_count raises TypeError when the caller tries to finish iterating.

The last chunk has already been yielded. The failure occurs while recording the output-token histogram, and span.end() is not reached. Both sync and async stream wrappers reproduce this.

šŸ‘Ÿ Reproduction steps

This uses real SDK response objects and a real OpenTelemetry histogram, with a synthetic stream and no Gemini API call:

from google.genai.types import (
    GenerateContentResponse, GenerateContentResponseUsageMetadata,
)
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.metrics import MeterProvider
from opentelemetry.instrumentation.google_generativeai import (
    _build_from_streaming_response,
)

tracer = TracerProvider().get_tracer("repro")
meter = MeterProvider().get_meter("repro")
histogram = meter.create_histogram("gen_ai.client.token.usage")
chunk = GenerateContentResponse(
    usage_metadata=GenerateContentResponseUsageMetadata(
        prompt_token_count=2, total_token_count=2,
    )
)
span = tracer.start_span("repro")
stream = _build_from_streaming_response(
    span, iter([chunk]), "gemini-test", None, histogram,
)
assert next(stream) is chunk
try:
    next(stream)  # Expected: StopIteration
finally:
    print("span still recording:", span.is_recording())
    span.end()
šŸ‘ Expected behavior

The stream ends normally, the span ends, and missing token counts do not become non-numeric histogram measurements.

šŸ‘Ž Actual Behavior
span still recording: True
TypeError: must be real number, not NoneType

Controls tested:

Final usage / metrics Result
Missing candidate count, no thoughts, metrics on TypeError
Candidate count 0, metrics on Clean end
Missing candidate count, thoughts count 3, metrics on Clean end
Missing candidate count, metrics off Clean end

The async wrapper produced the same TypeError after yielding its final chunk.

šŸ¤– Python Version

3.13.11 on macOS arm64.

šŸ“ƒ Additional context

Tested installed versions:

  • opentelemetry-instrumentation-google-generativeai==0.63.1
  • google-genai==3.4.0
  • opentelemetry-sdk==1.45.2

The current source at 864f8591081251a40a40e0bb06a4d8084c3664d2 retains the same path:

This is a local wrapper reproduction with synthetic response data, not a live provider reproduction. #4603 handles thinking tokens, but when both candidate and thinking counts are absent, output_tokens still stays None. Skipping absent metrics or normalizing them according to the intended usage semantics would avoid the exception.

šŸ‘€ Duplicate check
  • Checked existing issues and related token-usage PRs; did not find this exact report.

I’m with Tessl. Tessl reviews code using repository-level context and specific review lenses that improve over time through feedback. That’s how this finding surfaced. We’re just getting started and want to offer the tool free to open-source projects. If this is useful, happy to give you access.

Baptiste

Dominant language
Python
Stars
7.5k
Forks
1.2k
Avg merge
2d 6h
Merged PRs (30d)
2

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from traceloop/openllmetry

All issues in traceloop/openllmetry

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.