Auditing agent token usage
Run the deterministic audit, then explain its results. The script is the measurement; do not substitute your own log reading for it.
"<skill-dir>/scripts/audit-token-usage.sh" <days>
<days> is a positive integer, default seven unless the user names another
window. The script needs jq and reads Claude Code transcripts and Codex
rollouts under the home directory. On a nonzero exit, relay stderr and stop.
Analyze
- State the window and which sources had data. Claude Code sections, [1] to [5], measure transcript proxies such as turn counts and character sizes. The Codex section, [6], uses recorded token counts.
- Identify the largest evidenced costs: expensive model usage, inherited subagent routing, long sessions, compactions, context re-reading, oversized tool results, and configured MCP servers absent from section [5].
- Recommend only changes the output supports: at most three, ranked by expected impact, each citing the count or percentage behind it.
- Never present Claude Code proxy values as measured tokens.
- When a source is absent or the script fails, report the missing data or prerequisite. Do not infer its usage.