budget.yaml

# CrashLens Policy Configuration
# Enforce cost controls and usage policies for LLM logs

metadata:
  name: "Budget Control Policy"
  version: "1.0"
  description: "Prevent cost overruns and enforce model usage policies"

rules:
  - id: "no-gpt4-in-retries"
    description: "Prevent expensive GPT-4 usage in retry scenarios"
    match:
      model: "gpt-4"
      retry_count: ">2"
    action: "fail"
    severity: "high"
    suggestion: "Use GPT-3.5-turbo for retries or reduce fallback steps to save costs"

  - id: "token-limit-exceeded"
    description: "Warn when token usage is very high"
    match:
      usage.total_tokens: ">10000"
    action: "warn"
    severity: "medium"
    suggestion: "Consider breaking down large prompts or using more efficient models"

  - id: "expensive-model-overuse"
    description: "Block overuse of expensive models"
    match:
      model: ["gpt-4", "gpt-4-32k", "claude-3-opus"]
      usage.prompt_tokens: ">5000"
    action: "fail"
    severity: "high"
    suggestion: "Use GPT-3.5-turbo for large prompts or implement prompt compression"

  - id: "development-model-restriction"
    description: "Restrict expensive models in development environment"
    match:
      model: "regex:gpt-4.*"
      environment: "development"
    action: "block"
    severity: "critical"
    suggestion: "Use GPT-3.5-turbo in development. GPT-4 is reserved for production"

  - id: "excessive-retries"
    description: "Prevent retry loops that waste tokens"
    match:
      retry_count: ">=5"
    action: "fail"
    severity: "high"
    suggestion: "Implement exponential backoff or circuit breaker pattern"

  - id: "unauthorized-model"
    description: "Block usage of unauthorized models"
    match:
      model: "not in:['gpt-3.5-turbo', 'gpt-4', 'claude-3-haiku']"
    action: "block"
    severity: "critical"
    suggestion: "Only approved models are allowed. Contact admin for model authorization"