GCP Quota & Spending Governance Skill

This skill provides an expert workflow for analyzing cloud spend velocity and establishing enforceable hard caps, rate limits, and automated circuit breakers across Google Cloud Platform and Vertex AI.

---

CRITICAL SAFETY MANDATE: READ-ONLY AUDIT & HUMAN EXECUTION

STRICT RULE FOR AGENT EXECUTION:

1. The AI agent MUST NOT execute any mutable quota changes, scaling overrides, or billing modifications directly on the user's infrastructure.
2. All discovery operations MUST be strictly read-only (list, describe, inspect_quotas.py).
3. Recommended quota adjustments and scaling limits MUST be formatted as copy-pasteable CLI commands for human review and execution.

---

When to Use

Use this skill when:

• Investigating runaway API spend (e.g., Vertex AI Claude models, Gemini API, Text-to-Speech).
• Setting hard request/min or token/min rate limits via gcloud beta quotas preferences.
• Capping Cloud Run microservice auto-scaling (--max-instances).
• Designing automated billing circuit breakers (Pub/Sub + Cloud Functions) to enforce hard dollar cutoffs.
• Inspecting existing quota overrides across projects using scripts/inspect_quotas.py.

---

Bundled Tools & References

1. Quota Inspector Script (scripts/inspect_quotas.py):
   • Read-only script that queries configured Cloud Run max instances and custom quota preferences.
   • Run: python3 scripts/inspect_quotas.py --project <PROJECT_ID>
2. Quota Capping Guide (references/quota-capping-guide.md):
   • Syntax and dimension mappings for Vertex AI and Gemini base models.
3. Automated Circuit Breaker Architecture (references/automated-circuit-breaker.md):
   • Reference implementation for event-driven hard spend caps.

---

Step-by-Step Workflow

Step 1: Inspect Current Scaling & Quotas

Run the read-only inspection script:

python3 scripts/inspect_quotas.py --project <PROJECT_ID> --service aiplatform.googleapis.com

Step 2: Determine Appropriate Cap Granularity

Based on the cost drivers identified:

• Layer 1 (Per-Model / Per-Token Cap): For frontier LLMs (Claude Sonnet, Opus, Gemini 1.5 Pro).
• Layer 2 (Regional API Request Cap): Total prediction requests/min in a specific region (e.g., us-central1).
• Layer 3 (Microservice Instance Cap): Setting --max-instances=2 on Cloud Run services.
• Layer 4 (Hard Dollar Spend Cap): Automated billing circuit breaker via Pub/Sub.

Step 3: Present Recommendations to Human Operator

Formulate copy-pasteable commands for the human operator to apply:

# Example: Cap online prediction requests to 50/min in us-central1
gcloud beta quotas preferences create \
  --service=aiplatform.googleapis.com \
  --project=<PROJECT_ID> \
  --quota-id=OnlinePredictionRequestsPerMinutePerProjectPerRegion \
  --dimensions=region=us-central1 \
  --preferred-value=50 \
  --allow-high-percentage-quota-decrease

# Example: Enforce Cloud Run max instances
gcloud run services update <SERVICE_NAME> \
  --project=<PROJECT_ID> \
  --region=<REGION> \
  --max-instances=2