What problem does it solve? It replaces unreliable AI-estimated token counts with real numbers read directly from the session log, so you can see whether terse-output modes actually save tokens after accounting for per-turn rule overhead. ## Core Features & Use Cases - Real Token Receipts: Reads the current Claude Code JSONL session log on disk to report actual input and output token counts instead of model-generated estimates. - Net Savings Accounting: Computes estimated savings versus a non-caveman baseline, subtracts per-turn rule overhead (default 1,250 tokens/turn, configurable via CAVEMAN_RULE_OVERHEAD_TOKENS), and honestly reports net-negative sessions. - Lifetime Savings Badge: Writes a lifetime-savings suffix file consumed by the statusline badge for a glanceable gross-savings figure. - Use Case: After a long coding session, run the command to check whether the output-token savings exceeded the input-token cost of injected rules, and decide whether to keep the mode enabled for this workload. ## Quick Start Ask the assistant to run /caveman-stats to show this session's real token usage, estimated savings, and net overhead.
