Summary

• Normalize usage in Roo provider so Anthropic-protocol models report non-cached input tokens while non-Anthropic models keep total prompt_tokens.

Changes

• Roo provider: determine protocol and normalize input tokens:
  • Anthropic models: inputTokens = prompt_tokens − cache_write − cache_read
  • Non-Anthropic models: inputTokens = prompt_tokens
  • Implementation: roo.ts, protocol via getApiProtocol()
• Core (Task): revert prior heuristic to keep provider-agnostic logic:
  • Cost calc sites unchanged: Task.ts, Task.ts

Why

• Roo backend usage reports prompt_tokens in OpenAI-like shape. For Anthropic-prefixed models, treating prompt_tokens as “non-cached” would double count cached tokens when totals are derived. Normalizing at the provider ensures consistent semantics and prevents double counting.

Verification

• Unit tests pass (vitest).
• Manual: run a Roo Anthropic model; confirm api_req_started tokensIn equals non-cached input (not including cacheWrites/cacheReads), and contextTokens = tokensIn + tokensOut once.

References

• Provider change: roo.ts
• Protocol detection: getApiProtocol()
• Core call sites: Task.ts, Task.ts