fix(server-utils): Deduplicate Google GenAI streaming tool calls

[zkasuran]

The streaming instrumentation for @google/genai records every tool call twice.

handleCandidateContent in packages/server-utils/src/ai/google-genai/streaming.ts pushes tool calls to the span from two sources on the same chunk:

• chunk.functionCalls, spread straight into state.toolCalls
• part.functionCall, mapped again while iterating candidate.content.parts

chunk.functionCalls is a getter on the @google/genai response that is itself derived from those same candidate parts (it filters parts for functionCall then maps them). So a single real tool call lands in gen_ai.response.tool_calls as two entries. The two do not even share a schema: the getter keeps the SDK-native { id, name, args } while the parts loop emits { type, id, name, arguments }. One value ends up keyed args, the other arguments.

The non-streaming path (addResponseAttributes) reads tool calls from response.functionCalls and records one entry per call, so streaming and non-streaming disagreed on both count and shape.

Fix: take streaming tool calls only from chunk.functionCalls, the same accessor the non-streaming path uses, then drop the second push. Each call is now recorded once, in one shape. The streaming output matches the non-streaming output.

I kept chunk.functionCalls rather than the parts loop so both code paths share one source of truth and emit the SDK-native shape. This mirrors the OpenAI and Anthropic integrations, whose streaming paths reconstruct the exact tool-call shape their non-streaming paths produce. The deprecated gen_ai.response.tool_calls example shows { name, arguments }, but no provider integration normalizes to that literally (Anthropic keeps input, OpenAI nests under function), so consistency between a provider's streaming and non-streaming output was the stronger property to preserve here.

Root cause: chunk.functionCalls and part.functionCall are two views of the same data. Both were being written to the span.

Verified against a real gemini-3.6-flash streaming response that returns one controlLight tool call (response id IuF-auTiDpW2g8UPnNDKsAI). The identical response was replayed through the instrumentation before and after the change.

before (gen_ai.response.tool_calls):

[{"id":"call_2079699","args":{"colorTemperature":"warm","brightness":30},"name":"controlLight"},{"type":"function","id":"call_2079699","name":"controlLight","arguments":{"colorTemperature":"warm","brightness":30}}]

after:

[{"id":"call_2079699","args":{"colorTemperature":"warm","brightness":30},"name":"controlLight"}]

A unit test in packages/server-utils/test/ai/lib/tracing/google-genai-streaming.test.ts covers the single-call case, multiple calls across chunks and the recordOutputs: false case, then asserts the non-streaming path still records one entry in the same shape. It fails on develop (two entries) and passes with this change.

• If you've added code that should be tested, please add tests.
• Ensure your code lints and the test suite passes (yarn lint) & (yarn test).
• Link an issue if there is one related to your pull request. If no issue is linked, one will be auto-generated and linked.

---

AI assistance (Claude, Anthropic) was used in developing this change. The design, review and verification were done by the author. Verified locally before submitting: the new and existing @sentry/server-utils unit tests (377 passing), oxlint, oxfmt --check and the TypeScript type-check all pass, plus a real gemini-3.6-flash streaming run captured before and after.