| name | caveman-stats |
| version | 1.0.0 |
| targets | ["claude-code"] |
| type | skill |
| description | Use when the user invokes /caveman-stats or asks for caveman token-usage stats. Shows real token usage and estimated savings for the current session, read directly from the Claude Code session log — no AI estimation. Output is injected by the `caveman-mode-tracker` hook (vendored alongside this skill), which intercepts the slash command and returns the formatted stats as a blocked-decision reason. |
| category | {"primary":"economy"} |
| license | {"upstream":"MIT","source":"https://github.com/JuliusBrussee/caveman@63a91ec","path":"skills/caveman-stats/SKILL.md"} |
Caveman Stats
This skill is delivered by the hooks/caveman-mode-tracker hook, which intercepts /caveman-stats on UserPromptSubmit and runs the bundled caveman-stats.js against the active Claude Code session transcript. The hook returns decision: "block" with the formatted stats as the reason — the user sees real token counts and estimated savings immediately, and the model never has to compute or estimate anything.
Numbers come from the JSONL session log on disk:
output_tokens and cache_read_input_tokens summed across every assistant turn.- Estimated savings derived from the mean per-task ratio recorded in
caveman-stats.js (COMPRESSION table — currently only full mode has benchmark data; other modes report no estimate).
- Lifetime aggregation across sessions via
/caveman-stats --all or /caveman-stats --since 7d, backed by ~/.claude/.caveman-history.jsonl.
If the hook isn't registered in the active settings, the slash command falls through to the model — surface that fact and point the user at the wardrobe hooks/caveman-mode-tracker component (or the upstream installer at https://github.com/JuliusBrussee/caveman).
