Claude Code tells you the model, not the culprit: 106 lines of shell to name the Skill

5015 tool_use events across 39 sessions in a week, the sample summary the script prints in `--short` mode ^[10], works out at about 129 tool calls per session ^[1] and roughly 716 a day ^[2]. That is the traffic `/usage` compresses into one currency figure per model ^[1]. In the sample output reproduced in the dev.to writeup, two rows cover the whole period: 3,240 yen for Sonnet and 892 for Opus ^[2], so 78 percent of the money is attributed to a model name and nothing else ^[3].

The gap is a unit mismatch, and the script does not close it. `usage-breakdown.sh` counts invocations ^[3]; the built-in report is denominated in currency ^[1]. Its author is straight about this, noting in the code that real token totals would mean aggregating the `usage` object off API responses and that call count is standing in until then ^[7].

The time window is coarser than the flag suggests. Sessions are included or skipped by comparing each file's `mtime` against a cutoff, with 7d the default and 14d or 30d available by argument ^[9]. That is a file-level test, so a long session whose last write happens to land inside the window is counted in full, earlier events and all.

The counters are flat, too. One Skill call and one Agent call each add 1 to their own tally ^[11], and the Agent branch keys on `input.subagent_type` ^[6], so a subagent that runs its own conversation weighs the same as a skill that injects a prompt. The author's defence of call counts holds where the compared things are alike: a Skill called 100 times against one called once really is orders of magnitude apart in tokens ^[8]. Across categories it does not hold.

One reporting choice earns its keep. Skills named in `plugin:skill` form are counted under their namespace as well as individually, on the reasoning that you often only need to know the `superpowers` plugin is heavy rather than which of its skills ^[12]. Less comfortable: parse failures are swallowed by try/except ^[13], which stops one malformed line killing a run, at the price of undercounting without saying so.

Ranking one Skill against its own history at a fixed window is the reading this data supports. Anything wider produces a number you would not want to defend in a budget meeting.