cost-monitor

name	cost-monitor
description	Token/cost visibility for agentic engineering — per session, per day, per model lane. Use when someone asks about token usage, AI spend, "which lane is the money going to", or as a weekly habit once multi-model routing is live. Read-only reporting; changes nothing.
source	ccusage (the de-facto local usage analyzer); aios-workspace cost-monitor skill (team-level rollup variant)
am_pattern	E1 (cost axis), B6

You are producing a cost picture of the engineer's agent usage. Cost discipline is an AM axis for a reason: lanes and loops multiply spend quietly, and the routing table's economics (models/routing.yaml) are claims until someone measures them.

Step 1 — local usage (Claude Code)

ccusage reads the local session logs; no account access needed:

npx ccusage@latest            # daily table: tokens, cost (API-equivalent), models
npx ccusage@latest --json     # machine-readable, for trend tracking
npx ccusage@latest session    # per-session breakdown — find the expensive ones

Notes to carry into the report: figures are API-equivalent estimates from logs (a subscription user's invoice differs); other runtimes (opencode, Codex) have their own usage surfaces — include them if the team runs lanes through them.

Step 2 — read it against the routing table

Answer four questions, with numbers:

1. Lane mix — what share of tokens ran on the frontier lane vs bulk/utility? If ~everything is frontier, the routing table isn't being used (or isn't needed yet — say which).
2. Expensive sessions — the top 3 sessions by cost: were they expensive because the task was big, or because of thrash (retries, context bloat, a loop that should have been stopped)? Thrash findings go to modules/context-monitor territory.
3. Trend — spend per week over the last month; flag a >50% jump with its cause.
4. Unit economics — rough cost per merged change this period. This is the number that makes lane routing worth it (or not) — and remember E1: a frontier model you don't babysit is often cheaper end-to-end than a cheap model you correct twice.

Step 3 — report

One short table (day/week totals, lane mix, top sessions) + 2–3 sentences of "so what": the one behavior change that would save the most money without dropping a rung on the verification ladder. Never recommend saving money by weakening a review gate — MG6 exists precisely because cheap-lane output without frontier review is the most expensive thing in this whole system.

Team rollup (optional)

With modules/aios-cli + a Team Brain, each engineer pushes their metrics (aios analyze --since billing --push) and spend lands on one shared surface instead of in N terminals. Same habit, team-visible.