cost-tracker

cost-tracker — Spend Anomaly Detection per PR

Why: MicroSaaS margin is tight. A feature that works but 5×'s Claude API cost isn't shippable without conscious tradeoff. This skill surfaces that tradeoff before the PR merges.

Invoked by: /smo-score (automatic during Engineering hat Q8 if PR touches Claude/OpenAI/DB paths) Consumes: 7-day rolling spend from Anthropic console, Supabase dashboard, Vercel analytics Produces: docs/cost-reports/YYYY-MM-DD-{pr}.md + annotation on Engineering hat score

---

What it checks

1. Claude API token spend

• Compare this PR's branch deploy vs main (7-day avg)
• Metric: tokens/request, cache hit rate, request count/minute
• Flag if: >2× token-per-request OR cache hit <30% on paths that were >70%

2. OpenAI spend (if project uses it)

• Same as Claude — comparative

3. Database cost signals (Supabase)

• Read rows/request (N+1 query detection)
• Write rows/request (bulk insert bloat)
• Egress bytes (large JSON payloads)
• Flag if: any >3× delta

4. Frontend bundle size (if UI PR)

• Compare .next/analyze output vs main
• Flag if: >50KB net add without justification in PR description

5. Server cost (if deploy PR)

• PM2 memory footprint (pm2 list --json)
• CPU % at steady state
• Flag if: memory >1.5× OR CPU sustained >70% (headroom lost)

---

Signal, not gate

This skill does NOT block the PR. It annotates Engineering hat Q8 in /smo-score with a cost flag:

Q8 (deps + cost): 7 → cost-tracker flags 3.2× Claude token spend on /api/draft
Action required: document tradeoff OR optimize before ship

Dev must either:

• Justify in PR description (e.g., "new feature uses longer context, accepted tradeoff") — Q8 returns to 9+ with comment
• Optimize (batch calls, cache responses, reduce context window) — Q8 returns to 10

---

Integration with /smo-retro

Monthly, /smo-retro aggregates cost-tracker flags across projects. Pattern surfaces:

• "3 PRs last month doubled SSE's Claude spend. Time to evaluate alternate model or prompt compression."

This becomes a lesson in smorch-brain/canonical/lessons.md (cross-project).

---

Data sources + how to pull

Anthropic console

Export usage CSV from console.anthropic.com/billing/usage manually OR via API:

curl -H "x-api-key: $ANTHROPIC_API_KEY" \
  "https://api.anthropic.com/v1/organizations/{org}/usage?start=2026-04-12&end=2026-04-19"

Store 7-day rolling window in ~/.claude/cost-baseline/claude-7d.csv.

Supabase

SELECT date, SUM(db_size_bytes) as size, SUM(egress_bytes) as egress, SUM(db_operations) as ops
FROM supabase_usage
WHERE date > NOW() - INTERVAL '7 days'
GROUP BY date ORDER BY date;

Vercel

vercel inspect {deployment-url} --token $VERCEL_TOKEN → JSON with function invocations + bandwidth.

Server (PM2)

Run on the target server via /smo-deploy --dry-run or SSH:

pm2 list --json | jq '.[] | {name, monit}'

---

Configuration

Per-project in .smorch/project.json:

{
  "cost_tracking": {
    "claude_multiplier_threshold": 2.0,
    "supabase_ops_threshold": 3.0,
    "bundle_size_kb_threshold": 50,
    "exclude_paths": ["/api/seed"]
  }
}

Omitted = defaults above.

---

What this skill does NOT do

• It does not STOP a deploy. Just flags.
• It does not manage budget alerts (those live in Anthropic/Stripe/Supabase dashboards).
• It does not predict future cost — only compares PR-vs-baseline.
• It does not audit third-party pricing changes. (That's a monthly review in /smo-retro.)

---

First run setup

mkdir -p ~/.claude/cost-baseline
# Pull 7 days of historical data once:
bash ~/.claude/plugins/smorch-dev/skills/cost-tracker/scripts/seed-baseline.sh
# Subsequent runs auto-refresh

Seed script (to be created) pulls from all configured sources and stores baseline CSVs.