What problem does it solve? It removes guesswork from measuring token consumption by reading real input/output token counts directly from the session JSONL log instead of relying on model self-estimates, and it exposes when rule-injection overhead makes a terse-output workflow net-negative. ## Core Features & Use Cases - Real token accounting: Reads the current Claude Code session log on disk and reports actual input and output token counts per session. - Savings and net estimation: Compares output against an estimated non-caveman baseline, then subtracts per-turn rule overhead (default 1,250 tokens/turn, configurable via CAVEMAN_RULE_OVERHEAD_TOKENS) to show an honest net figure. - Statusline badge support: Writes a lifetime-savings suffix file consumed by a statusline badge showing cumulative gross savings. - Use Case: After a long coding session, run /caveman-stats to see whether the injected brevity rules actually saved tokens for that workload or cost more in input overhead than they returned in output savings. ## Quick Start Type /caveman-stats in the current session to display real token usage, estimated savings, rule overhead, and net result.
