usage

Report usage from the team dashboard. Arguments, possibly empty: $ARGUMENTS

1. Read ~/.usage-analyzer.json for url and the user's own name. If that file is missing, tell them to run the setup skill first and stop.

2. Fetch the window they asked for, defaulting to 30 days:

   python3 - <<'PY'
   import json, urllib.request
   from datetime import date, timedelta
   from pathlib import Path
   cfg = json.loads((Path.home() / ".usage-analyzer.json").read_text())
   days = 30  # replace with what the user asked for
   since = (date.today() - timedelta(days=days)).isoformat()
   with urllib.request.urlopen(cfg["url"].rstrip("/") + f"/api/stats?since={since}", timeout=30) as r:
       print(json.dumps(json.load(r)["by_user"], indent=2))
   PY

3. Answer in a short table, biggest consumer first. One row per person, merging their claude and codex rows unless the user asked to split by tool. Derive:
   • total tokens = input + output + cache read + cache write
   • tokens/prompt = total tokens / prompts
   • cache hit = cache read / (cache read + input + cache write), as a percent
   • prompts/session, tools/prompt
4. Add at most two lines of reading, and only where the numbers support it: cache hit under 40% means context is rebuilt every turn, usually a session that should have been /cleared. Very high tokens/prompt with few tool calls means a big context is being carried for little work.
5. End with the dashboard URL, since it has the sortable version.

Do not guess numbers for someone whose row is absent — they have simply not synced.