Fix cost and token tracking between provider styles

[mrubens]

Before:

After:

This also fixes the real-time stats and telemetry sent to cloud.

---

Important

Refactor cost and token tracking across provider handlers to ensure consistent and accurate reporting for Anthropic and OpenAI protocols.

• Cost Calculation:
  • Refactor calculateApiCostAnthropic and calculateApiCostOpenAI in cost.ts to return ApiCostResult with totalInputTokens, totalOutputTokens, and totalCost.
  • Update cost calculation logic in calculateApiCostInternal() to handle both Anthropic and OpenAI protocols.
• Provider Handlers:
  • Update AnthropicHandler, CerebrasHandler, DeepInfraHandler, GroqHandler, LiteLLMHandler, OpenAiNativeHandler, and RequestyHandler to use the new cost calculation functions.
  • Ensure totalCost is extracted and used consistently across handlers.
• Token Usage Tracking:
  • Modify getApiMetrics() in getApiMetrics.ts to calculate contextTokens using tokensIn and tokensOut.
  • Update Task.ts to handle cost and token tracking using the new ApiCostResult structure.
• Testing:
  • Update tests in cost.spec.ts to validate new cost calculation logic for both Anthropic and OpenAI.

^{This description was created by for 6be9762. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Review Summary

✅ No issues found - This PR successfully fixes cost and token tracking between provider styles.

Changes Reviewed

Core Changes:

• Modified calculateApiCostAnthropic() and calculateApiCostOpenAI() to return ApiCostResult object with totalInputTokens, totalOutputTokens, and totalCost
• Updated all provider implementations to use the new return structure
• Fixed token tracking in Task.ts to use provider-aware total tokens
• Updated getApiMetrics() to use total input tokens correctly

Provider Updates:

• anthropic.ts
• cerebras.ts
• deepinfra.ts
• groq.ts
• lite-llm.ts
• openai-native.ts
• requesty.ts

Test Coverage:

• All tests in cost.spec.ts updated to verify new return structure

Verification

✅ Type safety maintained across all changes
✅ Correct token calculation for both Anthropic (excludes cache from base input) and OpenAI (includes cache in base input)
✅ All call sites updated consistently
✅ Test coverage updated appropriately
✅ Telemetry and cloud stats will now receive accurate total token counts

View Job

[roomote]

Review TODOs

• Relax ParsedApiReqStartedTextType fields to optional in getApiMetrics.ts to match actual payloads; current implementation uses typeof guards implying keys can be absent.

Follow Along on Roo Code Cloud

[roomote]

ParsedApiReqStartedTextType marks tokensIn/tokensOut/cacheWrites/cacheReads as required, but this block guards for missing fields with typeof checks. api_req_started payloads may not always include these keys (e.g., before usage aggregation), so the type should make them optional to match runtime reality and prevent misleading typings. Suggested change: make these fields optional in getApiMetrics.ts.