fix(gemini): correct token usage accounting

[liugy789]

AgentScope-Java Version

2.0.3-SNAPSHOT

Description

Fixes #3033.

GeminiResponseParser previously excluded tool-use prompt tokens from input usage and subtracted thinking tokens from output usage. This caused inaccurate usage, budget, and cost accounting for responses that use server-side tools or thinking.

This change:

• classifies promptTokenCount + toolUsePromptTokenCount as input tokens;
• classifies candidatesTokenCount + thoughtsTokenCount as output tokens;
• falls back to totalTokenCount - inputTokens when candidate count is absent;
• preserves reported thinking tokens when both candidate and total counts are absent;
• adds parser-level regression coverage for thinking, server-side tool use, the missing-candidate fallback, and the thinking-only fallback.

No public API, dependency, or user-facing documentation changes are required. Existing cached-token handling is preserved.

Validation:

• GeminiResponseParserTest passes with 15 tests.
• The Gemini module and its dependencies pass.
• Spotless and git diff --check pass.
• Repository-wide GitHub CI is running for the latest commit.

Checklist

• Code has been formatted with Spotless.
• All affected tests are passing.
• Javadoc comments are complete and follow project conventions. No public API was added or changed.
• Related documentation has been updated. No documentation is required for this internal bug fix.
• Code is ready for review.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[oss-maintainer]

Summary

Corrects Gemini usage accounting by billing toolUsePromptTokenCount as input and thoughtsTokenCount as output, with sensible fallbacks when candidate counts are missing. Well covered by four new parser-level tests. Main ask is on documentation of the behaviour flip: previously output deliberately excluded thinking tokens, so consumers that added them back will now double-count.

---

Automated review by github-manager-bot

[oss-maintainer]

Classifying toolUsePromptTokenCount as input looks right and matches how the other providers bill server-side tool turns. Please confirm the same assumption holds for cachedContentTokenCount: in Gemini the cached count is a subset of promptTokenCount, so inputTokens stays correct, but a comment here would prevent a future "fix" from subtracting it twice.

[oss-maintainer]

This flips the sign of the previous intentional behaviour (thinking tokens used to be excluded from output). Two consequences worth a explicit note: (1) stored historical usage for Gemini sessions is not comparable across this boundary, and (2) any user-side cost/budget math that added thinkingTokens back on top of outputTokens now double-counts. Could you call this out in the changelog/release note, and check the OpenAI/DashScope parsers so all providers use the same convention?