token-charts

Everything behind Fifty-five billion tokens: the scripts that walk the Claude Code transcripts, the aggregated data they produce, and the Vega-Lite specs that turn it into the charts. Each spec reads its sibling .data.json, so it will render as-is in the Vega editor.

The scripts read and write in the working directory, and the intermediates are large, so run them somewhere else: copy the .py files to a scratch directory, run them in the order below, and copy the resulting .vl.json and .data.json back here.

Scripts

• scan_daily.py — walk every transcript in ~/claude-logs and ~/zed-logs, deduplicate on the assistant message id, emit one row per API response
• build_data.py — aggregate those rows by day, by model and by hour
• build_charts.py — the calendar heatmap, the per-model ridgeline, the cumulative panels and the diurnal profile
• build_cost_charts.py — the per-model and per-category cost charts
• build_project_chart.py — the per-project ridgeline, public repositories only
• scope.py — what the charts count (measured responses, priced models) and what each token costs
• chart_style.py — the config, colours and plot widths all the charts share, so they render to one width and one type size
• render_svgs.py — render each spec to the static SVG the post embeds, linking the labels of the public repositories

Specs and data

• daily-heatmap.vl.json · data
• rise-and-fall.vl.json · data
• by-project.vl.json · data
• cumulative.vl.json · data
• diurnal.vl.json · data
• by-model.vl.json · data
• where-the-tokens-go.vl.json · data