Token Audit

Analyse all Claude Code session files to show where tokens go, how well the cache is working, and which sessions are most expensive.

Installation

# The audit script lives at ~/.claude/skills/token-audit/audit.py
# Requires Python 3.10+ (stdlib only)

Usage

python3 ~/.claude/skills/token-audit/audit.py              # last 30 days
python3 ~/.claude/skills/token-audit/audit.py --days 7     # last 7 days
python3 ~/.claude/skills/token-audit/audit.py --today      # today only
python3 ~/.claude/skills/token-audit/audit.py --session ID # single session detail
python3 ~/.claude/skills/token-audit/audit.py --json       # machine-readable

What It Reports

Token Audit — Last 30 days
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Sessions:         247   Turns:            18,420
  Input tokens:     13K   Cache write:      19.6M
  Output tokens:   3.8M   Cache read:     1042.0M

  Cache hit rate: 98.2%  ███████████████████░
  Net cost est:   $483.20  Cache saved:   $9059.18

  ✓ Cache hit rate healthy (≥ 60%).

  Top sessions by cost          turns    cost   hit%
  ─────────────────────────── ─────  ──────  ─────
  730d5a18  2026-05-03         1762  $104.00    98%
  ...

Cache Hit Rate Targets

Rate	Status	Action
< 40%	Cold starts dominating	Use context-pack before new tasks; avoid /clear mid-session
40–60%	Improvable	Compact at task boundaries, not mid-task
≥ 60%	Healthy	Continue current strategy

Cost Notes

• Prices use Anthropic 2025/2026 API rates per model (Haiku / Sonnet / Opus)
• Mixed-model sessions priced per-turn for accuracy
• Values are estimates — actual billing may differ by plan

Session End Summary

Wire the companion hook for a per-session summary at shutdown:

"SessionEnd": [{
  "matcher": "*",
  "hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/session-token-stop.sh" }]
}]