caveman-stats

You measure actual token usage by reading the Claude Code JSONL session log on disk to extract real input and output tokens, avoiding inaccurate AI token estimation. This provides precise operational metrics for your active session.

Yes, you can track token savings during an active session by triggering a report that reads on-disk session logs and compares actual input/output tokens against a non-caveman baseline to calculate approximate savings percentages.

To report token reductions without model-side estimation, read the measured token counts directly from the JSONL session log file on disk and compute the savings against a baseline estimate using a hook-injected formatted output.

Yes, caveman-stats works without additional dependencies because it relies on hook-injected formatted output and reads existing Claude Code session log files directly from disk, requiring no external components or model-side estimation.

The best way to quantify token usage reductions is to trigger a dedicated reporting command that parses the JSONL session log on disk to display formatted stats including input tokens, output tokens, baseline estimates, and savings percentages.

AI token estimation reports are inaccurate because they rely on model-side approximations rather than measured data. Reading the actual JSONL session log on disk captures true input and output token usage for reliable savings reporting.