caveman 4.0 released: cutting 65% of tokens with caveman-style output for coding agents

caveman 4.0 splits into six independently toggleable modules: Say less shortens agent replies (a React re-render answer drops from 63 to 20 tokens), Read less compresses logs, JSON, CSV, and test output before the model reads them (up to 98.9% per file, 33.2% per Claude Code session), Auto picks model and reasoning effort by question difficulty (54% cheaper than Opus 5.5 xhigh on 85 SWE tasks), and Blocks lets agents reuse scripts left in the repo (one test run falls from 397 to 76 tokens). Install is a single npx command, requires Node.js 22.13+, supports 30+ agents including Claude Code, Codex, Gemini CLI, and OpenCode, and keeps keys and logins local. For long-session Claude Code users, Read less and Blocks directly cut per-call context overhead; per-request GitHub Copilot users get no benefit from shorter output, only input compression. Auto requires a free account login and sends no data when logged out. The team also admits some data is not yet re-tested: the HTML compressor is not built, session figures come from an older August 2026 engine, and Auto's 85 tasks were only half run.