The files that run up the bill

Token dashboards usually stop at the model name. That is the wrong grain. The invoice is a pile of reads: the same src/ tree pulled into context, the same generated fixture, the same 4,000-line types file the agent “just needed to check.” Graph Explorer is a tree of those files, sized by cache-aware cost — not by line count, not by how often a linter touched them.

Claude Code already writes which paths showed up in a turn. TokensLab rolls those touches up from file to folder to project so you can see concentration. A fat node is not a moral failure. It is a prompt you can cut in half tomorrow morning.

Why size-by-cost beats size-by-tokens

Raw token counts lie the same way a naive invoice lies. A cache read of two million tokens is not two million tokens at the input rate; it is 0.1×. A 1-hour cache write is 2×. If the graph used a single multiplier, the noisiest folder would always look like the most expensive one, and you would “optimize” the wrong place.

The graph uses the same five-class pricing as the dashboard. A file that is only ever cache-read looks smaller than a file that keeps forcing cache creation. That is the distinction that actually moves the bill.

What to do with a fat node

• Stop dumping the whole package into the system prompt. Point at the module.
• Keep generated or vendored trees out of the default context pack. Skills Explorer will show you if a skill is still dragging them in.
• If one project dominates all-time spend, open it from the dashboard. The briefing is the same dollars, scoped.

This is the gain the landing page is selling: not another chart of “tokens this week,” but a map of which files you paid to keep in the model’s working set. Install TokensLab, open Graph, and look at the heaviest branch. That branch is the prompt.