.Net: Fix Gemini streaming token usage metrics

[MohamedOthman1]

Motivation and Context

Fixes #13382.

Gemini streaming responses include cumulative usage metadata. The current connector records that metadata for every chunk, so a streamed response can inflate token counters by the number of chunks in the stream.

Description

This update keeps streaming chunks flowing as before, but suppresses usage logging while each chunk is parsed. Once the stream finishes normally, the connector records usage a single time from the latest chunk that actually included token usage metadata.

That keeps the non-streaming path unchanged and avoids losing metrics if Gemini ever sends a final chunk without usageMetadata. If the stream is cancelled or fails before normal completion, usage is not emitted for the partial response.

The added regression test uses repeated token counts across multiple stream chunks, followed by a final chunk without usage metadata, and verifies each Google token counter records the expected value once.

Pre-submit Notes

• Added focused regression coverage for repeated Gemini streaming usage metadata.
• Kept the change scoped to the Google Gemini connector and its unit tests.
• Checked the patch for whitespace issues with git diff --check.

[github-actions]

Automated Code Review

Reviewers: 4 | Confidence: 92% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Design Approach

---

Automated review by MohamedOthman1's agents

[Copilot]

Pull request overview

Fixes inflated OpenTelemetry token usage metrics for the Gemini streaming connector by ensuring accumulated usage metadata is emitted only once per streamed response (instead of once per chunk), aligning streaming behavior with Gemini’s accumulated usage reporting.

Changes:

• Update Gemini streaming response processing to suppress per-chunk usage logging and log usage once after streaming completes, using the last chunk that contains usage metadata.
• Refactor ProcessChatResponse to optionally skip usage logging so it can be reused for streaming without side effects.
• Add a unit test that validates prompt/completion/total token counters are each emitted exactly once for a multi-chunk stream (including a final chunk without usage metadata).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
dotnet/src/Connectors/Connectors.Google/Core/Gemini/Clients/GeminiChatCompletionClient.cs	Prevents duplicate token metrics in streaming by deferring usage logging until stream completion.
dotnet/src/Connectors/Connectors.Google.UnitTests/Core/Gemini/Clients/GeminiChatStreamingTests.cs	Adds regression coverage to ensure streaming usage metrics are emitted once even when the final chunk lacks usage metadata.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[MohamedOthman1]

@microsoft-github-policy-service agree