Caveman 2.1.0 labels which of its savings claims are real

JuliusBrussee/caveman hit 98,651 stars and 5,710 forks four months after its April 4 creation, cutting v2.1.0 on August 16 at 19:22 UTC, per GitHub.

The 65% figure the project advertises covers output tokens only. Input tokens are where a long agent session spends money, especially one re-reading the same files across 40 turns.

The pinned Claude Code benchmark for the Caveman 2 proxy measures 33.2% fewer input tokens, passing 18 exact-answer checks.

The release notes tag every local number inferred and every benchmark number benchmark_counterfactual. A repo with 98k stars had every reason to print the big number and stop. This one told readers which numbers are measured and which are simulated.

The new caveman learn feature reports context-depth: it buckets each session's peak context-window share and counts how often you crossed 30% and 50%. That signal shows where your harness is straining, not just what got trimmed.

Licensing splits by component. The skill, CLI, and SDKs are MIT. The engine and proxy are BSL-1.1, converting to Apache 2.0 by June 2030.

Running the skill locally, MIT covers you. Embedding the engine in something you sell means reading the BSL terms first.

A 33.2% reduction on a benchmark isn't a guarantee on your codebase. The compression drops material the proxy judges recoverable. Test it against your own repo before trusting it in an unattended run.