cost-optimizer

name	cost-optimizer
description	Use when auditing platform costs, optimizing AI token usage, or reviewing cloud infrastructure expenses. Covers model selection rules and AWS cost cleanup patterns for Rocha family systems.
version	1.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["home-assistant","rocha-family","cost","cloud","aws","optimization","platform"],"related_skills":["platform-manager","finance-manager","hermes-cron-schedule"]}}

Cost Optimizer Skill

Overview

Two-layer cost management: (1) platform AI token costs via model selection, and (2) cloud infrastructure waste via AWS audits. Runs automatically on schedule but can be triggered on-demand.

When to Use

• Reviewing token spend patterns
• Auditing AWS resources for waste
• Recommending model downgrades to cheaper tiers
• Weekly infrastructure cost check-in

Layer 1: Platform Token Optimization

Model Selection Hierarchy

Use Case	Preferred Model	Fallback
Simple task nudges, status checks	claude-haiku-4.5	—
Standard agents, briefings	claude-sonnet-4.6	haiku
Complex reasoning, medical, NICU	claude-sonnet-4.6	never downgrade
Code generation, PR review	claude-sonnet-4.6	—

Auto-Downgrade Rules

When token analysis shows a cron job consistently using <20% of model capacity:

1. Flag for haiku override
2. Add override to cron config (never touch original)
3. Report to Hector: model, job, estimated savings

Protected Agents — NEVER downgrade:

• nicu-care — baby health is non-negotiable
• wellness-coach — postpartum medical context
• health-coach — medical decisions
• finance-manager — money decisions
• Any agent where wrong output = real-world harm

Layer 2: AWS Infrastructure Audit

Run weekly (Monday 8:35 AM):

Audit Checklist

EC2: Stopped instances >7 days → flag for termination
EBS: Unattached volumes >7 days → snapshot + delete
EIPs: Unassociated Elastic IPs → release (cost: ~$3.60/mo each)
RDS: Dev/staging instances running nights/weekends → schedule stop
S3: Buckets with versioning on but no lifecycle policy → add 30-day cleanup
Lambda: Functions with 0 invocations in 30 days → flag for review

Auto-Actions (no approval needed):

• Release unassociated EIPs
• Add lifecycle policies to S3 buckets
• Tag untagged resources with owner: hermes-audit

Escalate to Hector (approval required):

• Terminate any EC2 instance
• Delete any RDS instance or snapshot
• Changes to security groups or IAM

Reporting Format

☁️ Cloud Cost Audit — [Date]
💰 Current monthly: ~$XX
🔴 Waste found: $XX/mo (X items)
  • [Item]: $X/mo — [action taken or recommended]
✅ Auto-cleaned: $X/mo
📋 Action needed: [items requiring approval]

Common Pitfalls

• ❌ Downgrading NICU/medical/finance agents for cost savings — never
• ❌ Terminating instances without approval
• ❌ Acting on stale data (always pull fresh usage before recommending)
• ❌ Over-reporting — only surface items with $5+/mo impact

Verification Checklist

• Protected agents list checked before any model downgrades
• AWS actions requiring approval routed to Telegram first
• Cost report leads with total impact, not list of findings
• Auto-actions logged to working memory for audit trail