fix: extract thoughtsTokenCount for Gemini thinking models

[colegottdank]

Summary

• Add reasoning_tokens extraction from Gemini usageMetadata.thoughtsTokenCount in worker's getDetailedUsage()
• Add reasoningTokens to GoogleBodyProcessor output in Jawn
• Fix promptTokens calculation when no caching is present (was returning undefined instead of raw promptTokenCount)

Problem

Gemini thinking models (2.5 Flash, 3.x Pro) return thoughtsTokenCount separately from candidatesTokenCount:

"usageMetadata": {
  "promptTokenCount": 15,
  "candidatesTokenCount": 359,
  "totalTokenCount": 1035,
  "thoughtsTokenCount": 661  // This was not being extracted!
}

This caused:

1. Token count discrepancies (input + output didn't match total)
2. Incorrect cost calculations (thinking tokens weren't tracked separately)
3. Missing reasoning_tokens in the detailed usage breakdown

Test plan

• Added test for Gemini thinking model response with thoughtsTokenCount
• Existing GoogleBodyProcessor tests pass
• Full packages test suite passes (578 tests)

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

3 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
helicone	Skipped		Feb 24, 2026 7:18am
helicone-bifrost	Skipped		Feb 24, 2026 7:18am
helicone-eu	Skipped		Feb 24, 2026 7:18am

[supabase]

This pull request has been ignored for the connected project bolqqmqbrciybnypvklh because there are no changes detected in supbase/supabase directory. You can change this behaviour in Project Integrations Settings ↗︎.

---

Preview Branches by Supabase.
Learn more about Supabase Branching ↗︎.

[supabase]

This pull request has been ignored for the connected project lmahfbbnchpworytrrqk because there are no changes detected in .supabase/supabase directory. You can change this behaviour in Project Integrations Settings ↗︎.

---

Preview Branches by Supabase.
Learn more about Supabase Branching ↗︎.

[claude]

Claude finished @colegottdank's task —— View job

---

PR Review: Gemini Thinking Tokens Extraction

Reviewing changes for:

• Code quality and implementation correctness
• Test coverage and edge cases
• Performance and security considerations
• API consistency and breaking changes
• Provide final assessment and recommendations

Analysis

Changes Overview:

1. Worker (DBLoggable.ts): Added reasoning_tokens extraction from Gemini usageMetadata.thoughtsTokenCount
2. Jawn (GoogleBodyProcessor.ts): Added reasoningTokens field and fixed promptTokens calculation
3. Tests: Added comprehensive test for Gemini thinking model scenario

Detailed Review

✅ Code Quality - Excellent

• Correct Implementation: Properly extracts thoughtsTokenCount and maps it consistently
• Important Bug Fix: Fixed critical issue where promptTokens was undefined when no caching present (line 76-78 in googleBodyProcessor.ts)
• Type Safety: Uses proper optional chaining and null coalescing operators
• Consistency: Changes align with existing codebase patterns and type definitions

✅ Test Coverage - Comprehensive

• Realistic Test Data: Uses actual Gemini 2.5 Flash response structure with authentic token counts
• Proper Assertions: Verifies all token calculations including the new reasoning tokens extraction
• Edge Case Coverage: Tests the conditional reasoningTokens field behavior (undefined when 0, value when > 0)
• Regression Prevention: Ensures existing functionality remains intact

✅ Performance & Security - No Issues

• Minimal Performance Impact: Only adds simple field extraction with no computational overhead
• No Security Concerns: Changes involve only data parsing/mapping without external dependencies
• Memory Efficient: Uses existing patterns for optional field handling

✅ API Consistency - Backward Compatible

• Non-Breaking: All changes are additive (new optional fields)
• Type Consistency: reasoningTokens field properly typed across all layers (Worker → Jawn → Web)
• Naming Conventions: Follows existing camelCase patterns (reasoningTokens vs reasoning_tokens)

Score: 9/10 - Excellent implementation, ready to merge

Minor Suggestions:

1. Documentation: Consider adding a comment explaining the Gemini thinking model token structure for future maintainers
2. Constants: The conditional logic thoughtsTokenCount > 0 ? thoughtsTokenCount : undefined could use a helper function, but this matches existing patterns in the codebase

Suggestions Summary:

• ✅ No critical fixes needed - Implementation is solid and well-tested
• ✅ No breaking changes - All changes are backward compatible
• ✅ Good test coverage - New test properly validates the thinking token extraction
• 💡 Optional enhancement: Add inline documentation for the thinking model token logic

Recommendation: Merge with confidence. This PR correctly fixes a significant gap in Gemini thinking model token tracking and includes proper testing.

---

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.