measure
Ground truth for everything else: nothing in rtok ships until this plugin can show a before/after number (decision D3).
| Surfaces | rtok stats, rtok bench, proxy usage capture |
| Spec | the spec (replaces) column of the catalogue in plan.md §1 |
| Default | on |
What it does
- Parses Claude Code transcripts (
~/.claude/projects/**/*.jsonl): tool calls, tool results,usageblocks, turn index. - Reads
usagerows written by the proxy. - Reports per-tool result sizes, Bash by command family, MCP server groups, cache hit rate, and context-token-turns (tokens × turns they stay in context).
- Saves and compares baselines (
rtok stats --save-baseline,--compare). - Optional calibration of the estimator against
count_tokens(--calibrate).
Config
None beyond [estimator] (chars per token per class), which --calibrate rewrites.
Tasks
See roadmap.md § measure. Checks in plan.md.
T1.1 JSONL parser · T1.2 rtok stats · T1.3 baseline · T1.5 calibration · T5.5 cache health · T9.1 rtok bench · T11.6 per-API usage.
Status
Manifest only. First task: T1.1.