Caveman 4.0 trims coding agent tokens with terse output, compressed input
Caveman, an open-source skill and proxy for coding agents that topped GitHub Trending and Hacker News, released version 4.0: it makes agent replies terse (a sample 63-token answer becomes 20 tokens) and shrinks what agents read — logs, JSON, CSV, YAML and test output — for a claimed 33.2% cut in per-session input tokens. It adds Auto, which picks the model and effort per request, Blocks code mode that keeps agent-written scripts for reuse (one test run drops from 397 to 76 tokens), and prompt-cache warming while the user pauses. The project page says Adobe Research cited caveman-style output, JetBrains A/B tested it on 86 coding tasks with no measurable quality loss, and Elasticsearch Labs remade caveman mode for MCP; the Auto cost comparison — 54% cheaper than Claude Opus 5.5 at xhigh effort on 85 SWE-Interact/SWE-Together tasks — is self-reported, half-complete, and the score gap is stated to be within ±7 points.
If the compression and routing numbers hold up, everyday coding agents could become meaningfully cheaper and faster without changing the user's prompt.