[sernote]

Orca-Code-Review — push 1

Severity	Count
P0	0
P1	0
P2	0
P3	0

✅ no blocking findings

Native Anthropic responses exposed LiteLLM's inclusive prompt total as ordinary input_tokens and omitted the cache read/write breakdown. Native Gemini responses also dropped the available cached-token count. Both blocking responses and terminal stream usage were affected.

Project known cache counters into each native protocol: subtract known read/write components once for Anthropic, and retain Gemini's inclusive prompt count while adding cachedContentTokenCount. Preserve explicit zero, omit unknown counters, and reject inconsistent or out-of-bounds breakdowns. The documented fallback retains the legacy inclusive total when a precise ordinary-input count cannot be established.

Validation: 842 tests passed in the offline Python 3.12 suite. Independent review reran 130 targeted tests plus 12 additional controls using actual LiteLLM usage normalizers and stream parsers. Native SDK terminal-counter behavior was checked against official SDK source; native SDK packages were not installed. Ruff and git diff --check passed. No live provider or billing measurements.

Discovered during a repository audit using audit-prompt-caching, then independently reproduced and reviewed.

[orcacode-review]

🐳 OrcaCode Review

✅ No findings — nothing to flag in this PR. Great work!

_{OrcaCode Review — Route Smarter. Ship Safer. Spend Less.
Engine-reported: 198 calls · 7.7M tokens · 94% cached}

❤️ Share · Install OrcaCode Review

Free on GitHub — the review runs on your own OrcaRouter key. If it helped, a shout-out goes a long way.

Share: X · Reddit · LinkedIn
Follow: X · Discord · LinkedIn · OrcaRouter