Skip to content
Longterm Wiki has not been actively maintained since May 2026. Some data is still updated occasionally, but please don't rely on this site being current or accurate.
Longterm Wiki
Navigation
Updated 2026-05-02HistoryData
Page StatusDocumentation
Edited 5 months ago366 wordsUpdated quarterlyOverdue by 69 days
Content3/13
SummaryScheduleEntityEdit historyOverview
Tables1/ ~1Diagrams0Int. links0/ ~3Ext. links3/ ~2Footnotes0/ ~2References0/ ~1Quotes0Accuracy0
Issues1
StaleLast edited 159 days ago - may need review

Agent Cost Monitoring

We run ≈$8-9k/week of Claude Code list-price spend across slots a1-a10. Cost incidents — runaway patrol sessions, broken cache eviction, accidental Opus usage — used to surface manually weeks late by parsing JSONL transcripts. pnpm crux sys cost report is the on-demand terminal view that catches them in days, not weeks.

Quick reference

# Default: last 7 days, daily totals
pnpm crux sys cost report

# Rank slots by spend over last 14 days
pnpm crux sys cost report --by=session --since=14d

# 5-hour billing-block view (matches Claude's rate-limit windows)
pnpm crux sys cost report --by=blocks

# Per-model breakdown — useful for "is this Opus or Sonnet?"
pnpm crux sys cost report --by=session --breakdown

# Machine-readable for piping into another tool
pnpm crux sys cost report --by=session --json | jq '.sessions[0]'

Wrapper around ccusage, which parses the same JSONL transcripts (~/.claude/projects/*.jsonl) we used to grep manually. Cost numbers match list-price billing (Opus 4.7 = $5/$25 per Mtok input/output, Sonnet 4.6 = $3/$15, etc.).

What --by means

--byGroupingUseful for
daily (default)Per-day totalsSpotting day-over-day spikes
sessionPer project directory (= per slot for us)"Which slot is burning the most?"
weeklyPer ISO weekWeekly trend
monthlyPer calendar monthBudget review
blocksPer 5-hour billing windowMatches Claude's rate-limit semantics

Note: ccusage calls a slot a "session" because it groups by project directory; an actual Claude Code conversation has its own UUID JSONL file. For per-conversation breakdowns, raw ~/.claude/projects/<slot>/<uuid>.jsonl is the source — there's no first-class command for that yet.

Detecting an incident — playbook

When you suspect a cost spike (Discord ping, monthly bill review, gut feel):

  1. pnpm crux sys cost report --by=session --since=7d to rank slots.
  2. If one slot is way ahead, drill in: pnpm crux sys cost report --by=blocks --since=2d --breakdown.
  3. Check the slot's tmux window or crux sys sessions list --slot=N to see what's running.
  4. Common causes:
    • Opus on a refactor that should be Sonnet — kill, swap model, restart.
    • Re-read/re-summarize loop — kill, file a Linear ticket on the looping pattern.
    • A patrol session with a stuck loop — see docs/agent-rules/agent-planning-discipline.md.

What's deliberately out of scope

  • Live dashboards / alerts — we run Grafana in the production k8s cluster already. If we ever want live cost dashboards, the right move is to wire Claude Code's built-in OTel exporter (CLAUDE_CODE_ENABLE_TELEMETRY=1) to the existing prod Prometheus, not to run a parallel docker stack on each workstation.
  • Per-PR / per-Linear-ticket cost attribution — needs a join with agent_sessions. Separate ticket if it becomes useful.
  • Cost-by-user attribution — single-user setup, not worth the labels.

References