Caveman 2.1.0 Ships a Context-Depth Report and Honest-Numbers Labelling — 98,651 Stars for a Claude Code Skill Whose Proxy Benchmarks 33.2% Fewer Input Tokens
JuliusBrussee/caveman cut v2.1.0 on Aug 16 at 19:22 UTC (GitHub API: 98,651 stars, 5,710 forks, created April 4, 2026), adding `caveman learn` context-depth reporting that buckets each session's peak context-window share with over-30%/over-50% counts, plus a two-card savings view splitting proxy-measured actual savings from replay counterfactuals. The headline 65% figure is output tokens only; the pinned Claude Code benchmark for the newer Caveman 2 proxy measures 33.2% fewer provider-reported input tokens while passing 18 exact-answer checks. Notably the release notes label every local number `inferred` and every benchmark number `benchmark_counterfactual` — unusually disciplined for a viral repo, and the reason to trust it. Licensing is split: MIT for the skill/CLI/SDKs, BSL-1.1 for the engine and proxy, converting to Apache 2.0 by June 2030.
↳ Follow the thread