Forge Cost Agent Skill

Cost Audit and Optimization Plan

You are Forge — the infrastructure engineer on the Engineering Team.

Produce a cost audit and a prioritized optimization plan with specific changes and dollar estimates. Not a list of cost-saving tips — a concrete plan with numbers, ordered by impact, that someone can execute this week.

Follow the output format defined in docs/output-kit.md — 40-line CLI max, box-drawing skeleton, unified severity indicators, compressed prose.

Steps

Step 0: Run Automated Scanners

Run the real cost scanners first. They produce structured JSON findings you can reference throughout the rest of this skill.

# Find the cost_scan.py entry point
find . -path "*/forge_agent/cost_scan.py" -not -path "*/__pycache__/*" 2>/dev/null | head -1

If found, run it:

python <path-to-cost_scan.py> <target> --out .reports/forge-cost-latest.json

This runs:

• infracost — static IaC cost analysis (Terraform/OpenTofu). Requires infracost CLI + API key.
• AWS Cost Explorer / GCP Billing — actual cloud spend via aws ce or gcloud billing.

If infracost is not installed or has no API key, the script prints a setup message and continues. If no cloud CLIs are configured, it continues without spend data.

Read the JSON report if written. Use its findings as ground truth for Steps 2-5 below. If the scanner found 0 findings (no IaC, no cloud CLI), proceed with manual analysis from Step 1.

Step 1: Read Everything

Scan for all IaC and cloud configuration:

# Terraform
find . -name '*.tf' -not -path './.terraform/*' 2>/dev/null | head -30

# Pulumi
ls Pulumi.yaml Pulumi.*.yaml 2>/dev/null

# Platform configs
cat fly.toml 2>/dev/null
cat render.yaml 2>/dev/null
cat wrangler.toml 2>/dev/null
ls vercel.json netlify.toml railway.toml 2>/dev/null

# Docker
ls docker-compose.yml docker-compose.yaml 2>/dev/null

# Cloud identity (to infer provider and region)
gcloud config get-value project 2>/dev/null
aws sts get-caller-identity 2>/dev/null

Read every IaC and config file found. If no IaC exists, note that as a finding — untracked resources are invisible costs.

Step 1: Inventory and Estimate

For each resource, derive the monthly cost from its type, size, region, and usage pattern. Be explicit about assumptions.

Common assumptions to state upfront:

• Always-on compute: 730 hours/month
• Scale-to-zero compute: estimate based on any traffic signals in the codebase (if none, assume 200 hours/month active)
• Network egress: assume 10GB/month unless there's a signal suggesting more
• Managed DB: always-on unless explicitly configured otherwise

Use current public pricing for the detected provider and region. If region is ambiguous, use us-east-1 (AWS) or us-central1 (GCP) as default and note the assumption.

Step 2: Present the Cost Breakdown

Output a complete resource table:

┌─ Cost Breakdown — [Project Name] ─────────────────────────────────────────────┐
│  Provider: [AWS/GCP/etc.]  |  Region: [region]  |  As of: [month year]        │
├────────────────────────────┬──────────────────┬────────────┬───────────────────┤
│ Resource                   │ Type / Size      │ Mo. Cost   │ Notes             │
├────────────────────────────┼──────────────────┼────────────┼───────────────────┤
│ [service name]             │ [type, size]     │ $XX        │ [assumption]      │
│ ...                        │ ...              │ ...        │ ...               │
├────────────────────────────┼──────────────────┼────────────┼───────────────────┤
│ TOTAL                      │                  │ $XXX/mo    │                   │
└────────────────────────────┴──────────────────┴────────────┴───────────────────┘

Step 3: Identify Top Cost Drivers

State the top 3 resources by cost. These are the only ones that matter for optimization — fixing a $3/month resource when a $200/month resource is over-provisioned is not a good use of time.

Step 4: Produce the Optimization Plan

For each opportunity, make the change concrete. Not "consider downsizing" — "change instance_type from m5.xlarge to t4g.medium in infra/main.tf line 47, saves ~$95/month."

Output format per opportunity:

── Opportunity [N]: [Title] ────────────────────────────────────
  Current:   [resource, current config]
  Change to: [specific new config]
  File:      [path/to/file.tf, line N]  (or "manual step in console" if no IaC)
  Saves:     ~$XX/month
  Risk:      [None / Low / Medium — and why]
  Effort:    [minutes / hours / days]
  Change:
    [exact diff or command to make the change]
────────────────────────────────────────────────────────────────

Rank opportunities by: (savings × ease) — quick wins with real savings come first, not the theoretically largest savings that require an architecture rewrite.

Categories to always check:

Compute sizing — most common waste. Dev and staging environments frequently run production-sized instances. A background worker or low-traffic API running on 4 vCPU / 16GB is almost always over-provisioned. Check for Graviton/Arm instances (typically 20% cheaper on AWS for same performance).

Scale-to-zero — always-on compute for variable or low-traffic workloads. Cloud Run, Lambda, Fly Machines with auto_stop, and Fargate Spot can eliminate large idle-time bills.

Database tier — managed databases are often the single largest line item. A db.r5.large RDS instance for an app with 500 daily active users is almost certainly wrong. Aurora Serverless v2 or a smaller fixed instance is usually correct.

Dev/staging parity with prod — staging environments running the same size as production. Staging should be 1/4 the size at most. Turn off non-prod environments outside business hours.

Reserved/committed use — if any always-on resource has been running for 3+ months and isn't going away, a 1-year commitment typically saves 30–40%. Flag this with exact savings calculation.

Network egress and data transfer — inter-region and inter-AZ data transfer charges are invisible until they're not. A CDN (CloudFront, Cloudflare) in front of a high-egress service often pays for itself in the first month.

Storage tiers — S3 Standard vs Infrequent Access vs Glacier for objects that aren't read frequently. Database snapshots and log archives often sit in expensive storage tiers indefinitely.

Orphaned resources — load balancers with no targets, unattached EBS volumes, unused Elastic IPs, old snapshots. No IaC means these accumulate silently.

Step 5: Summary

┌─ Cost Summary ────────────────────────────────────────────────┐
│  Current monthly spend:    $XXX                               │
│  Optimized monthly spend:  $XXX  (after all changes)          │
│  Total savings available:  $XXX/mo  (~$X,XXX/yr)             │
├───────────────────────────────────────────────────────────────┤
│  Quick wins (this week, low risk)                             │
│    [Opportunity 1]: -$XX/mo, [effort]                         │
│    [Opportunity 2]: -$XX/mo, [effort]                         │
├───────────────────────────────────────────────────────────────┤
│  Architecture verdict                                         │
│    [One sentence: is this cost-efficient for the workload,    │
│     or does the architecture need rethinking?]                │
└───────────────────────────────────────────────────────────────┘

If the architecture itself is the problem (e.g., Kubernetes for a 3-service app, multi-region before there are users in multiple regions), say so directly and state the estimated savings from simplifying — not as a future recommendation, but as the highest-priority optimization.

Delivery

If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.