What problem does it solve? Kiro cache-estimation coefficients (Kmodel) drift over time as pricing, multipliers, and usage patterns change, causing inaccurate conservative cache estimates. This Skill provides a reproducible offline process to recompute per-model Kmodel values from real usage data without touching production configuration. ## Core Features & Use Cases - Filtered Usage Sampling: Queries the LanceDB table llm_gateway_usage_events for successful Kiro requests from the last 30 days, excluding failed requests and missing credit data. - Statistical Coefficient Computation: Normalizes model aliases, applies sample-quality filters, and computes p50, p80, and p90 ratios per model, recommending p80 as the conservative Kmodel. - Safe Recommendation Output: Produces a report with sample counts, percentiles, and filter drop counts for manual review before updating /admin/kiro-gateway. - Use Case: When Kiro credit usage trends drift from current estimates, run this calibration to get data-backed per-model Kmodel recommendations, then manually apply them in the admin panel. ## Quick Start Recalibrate the Kiro Kmodel coefficients using the last 30 days of successful usage events and show me the recommended values per model.
