caveman-stats

Run the /caveman-stats command during the session. The mode-tracker hook reads the JSONL session log on disk and returns actual input and output token counts plus estimated savings versus a baseline.

Savings are computed by the hook script, not the model. It reads real token counts from the session log and compares output tokens against an estimated non-caveman baseline to derive the saved amount and percentage.

No. The hook returns decision block with the formatted stats as the reason, so the numbers come entirely from the session log file. The model takes no action when the skill fires.

Each run writes a lifetime-savings suffix file that the statusline badge reads to display cumulative savings, such as a compact formatted token count next to the badge icon.

Model self-estimates of token counts are unreliable. This skill bypasses estimation entirely by parsing the actual JSONL session log, so the figures reflect true recorded usage.