Hi HN — I built this after repeatedly seeing LLM cost regressions slip through review because nothing breaks; it’s just token drift and then a surprise bill.

AIOpt is a local-only CLI + CI gate that: - uses observed usage as a baseline (usage.jsonl/csv) - estimates candidate deltas (model/context/output/retry/traffic) - emits deterministic artifacts + SARIF for GitHub Code Scanning

No server, no uploads.

I’d really appreciate feedback on what would make this more useful in real PR review: 1) What signals/UX make token/cost regressions actionable (beyond “it’s higher”)? 2) Any must-have integrations you’d expect?