site-mariage/_byan/core/model-selector.yaml

# Model Selector - Intelligent AI Model Selection
# Based on complexity analysis and task requirements
# Inspired by BYAN v2 workers.md concept

version: "1.0.0"
description: "Auto-select optimal AI model based on task complexity"

# Complexity Levels
complexity_levels:
  simple:
    score_range: [0, 30]
    description: "Simple tasks - bash commands, file operations, detection"
    recommended_model: "gpt-5-mini"
    fallback_model: "gpt-5-mini"
    cost_tier: "FREE"
    use_cases:
      - Platform detection
      - File copy/installation
      - Dependency validation
      - Simple bash commands
    
  medium:
    score_range: [31, 60]
    description: "Medium tasks - code analysis, refactoring, documentation"
    recommended_model: "claude-haiku-4.5"
    fallback_model: "gpt-5-mini"
    cost_tier: "LOW"
    use_cases:
      - Code analysis
      - Refactoring suggestions
      - Test generation
      - Documentation writing
    
  complex:
    score_range: [61, 85]
    description: "Complex tasks - agent creation, architecture, workflow design"
    recommended_model: "claude-sonnet-4.5"
    fallback_model: "gpt-5.1-codex"
    cost_tier: "MEDIUM"
    use_cases:
      - Agent creation
      - Architecture design
      - Workflow orchestration
      - Business logic
    
  expert:
    score_range: [86, 100]
    description: "Expert tasks - critical review, security audit, optimization"
    recommended_model: "claude-opus-4.6"
    fallback_model: "claude-sonnet-4.5"
    cost_tier: "HIGH"
    use_cases:
      - Security audits
      - Critical code review
      - Performance optimization
      - Production deployment

# Complexity Calculation Factors
calculation_factors:
  
  task_type:
    description: "Type of task being performed"
    weights:
      detect: 5
      install: 10
      copy: 5
      validate: 10
      analyze: 40
      refactor: 45
      create: 70
      design: 75
      review: 60
      audit: 90
      optimize: 85
      document: 35
      test: 40
  
  context_size:
    description: "Amount of context/code being processed"
    weights:
      tiny: 0        # < 50 lines
      small: 5       # 50-100 lines
      medium: 20     # 100-1000 lines
      large: 40      # 1000-5000 lines
      huge: 60       # > 5000 lines
  
  reasoning_depth:
    description: "Depth of reasoning required"
    weights:
      shallow: 0     # Simple if/then logic
      medium: 20     # Some analysis required
      deep: 40       # Complex reasoning chains
      expert: 60     # Multi-step reasoning with validation
  
  quality_requirement:
    description: "Quality vs speed tradeoff"
    weights:
      fast: 0        # Speed priority, acceptable quality
      balanced: 10   # Balance speed and quality
      high: 20       # Quality priority
      critical: 30   # Maximum quality, production-critical

# Model Capabilities Matrix
models:
  gpt-5-mini:
    provider: "OpenAI"
    tier: "free"
    strengths: ["speed", "cost", "simple-tasks"]
    weaknesses: ["complex-reasoning", "large-context"]
    max_tokens: 128000
    typical_latency: "2-5s"
    cost_per_1k_tokens: 0.0
    
  claude-haiku-4.5:
    provider: "Anthropic"
    tier: "low-cost"
    strengths: ["balanced", "speed", "medium-tasks"]
    weaknesses: ["very-complex-reasoning"]
    max_tokens: 200000
    typical_latency: "3-7s"
    cost_per_1k_tokens: 0.00025
    
  claude-sonnet-4.5:
    provider: "Anthropic"
    tier: "standard"
    strengths: ["reasoning", "code-generation", "analysis"]
    weaknesses: ["cost"]
    max_tokens: 200000
    typical_latency: "5-15s"
    cost_per_1k_tokens: 0.003
    
  claude-opus-4.6:
    provider: "Anthropic"
    tier: "premium"
    strengths: ["expert-reasoning", "critical-tasks", "accuracy"]
    weaknesses: ["cost", "latency"]
    max_tokens: 200000
    typical_latency: "10-30s"
    cost_per_1k_tokens: 0.015
    
  gpt-5.1-codex:
    provider: "OpenAI"
    tier: "standard"
    strengths: ["code-generation", "technical-accuracy"]
    weaknesses: ["general-reasoning"]
    max_tokens: 128000
    typical_latency: "4-10s"
    cost_per_1k_tokens: 0.002

# Usage Examples
examples:
  yanstaller_detect:
    task_type: "detect"
    context_size: "small"
    reasoning_depth: "shallow"
    quality_requirement: "fast"
    calculated_score: 10
    selected_model: "gpt-5-mini"
    rationale: "Simple platform detection, no reasoning needed"
    
  byan_interview:
    task_type: "create"
    context_size: "medium"
    reasoning_depth: "deep"
    quality_requirement: "critical"
    calculated_score: 130
    selected_model: "claude-sonnet-4.5"
    rationale: "Complex agent creation requires deep reasoning"
    
  code_review_critical:
    task_type: "audit"
    context_size: "large"
    reasoning_depth: "expert"
    quality_requirement: "critical"
    calculated_score: 190
    selected_model: "claude-opus-4.6"
    rationale: "Critical security audit needs maximum quality"
    
  quick_refactor:
    task_type: "refactor"
    context_size: "medium"
    reasoning_depth: "medium"
    quality_requirement: "balanced"
    calculated_score: 75
    selected_model: "claude-sonnet-4.5"
    rationale: "Medium complexity refactoring"

# Override Rules
overrides:
  user_preference:
    description: "User can override with --model flag"
    priority: "highest"
    
  cost_limit:
    description: "If user has cost limit, downgrade model"
    priority: "high"
    
  platform_availability:
    description: "If preferred model unavailable, use fallback"
    priority: "medium"

# Logging and Metrics
logging:
  enabled: true
  log_level: "info"
  log_file: "{project-root}/_byan-output/model-selector.log"
  metrics:
    - model_selected
    - calculated_score
    - actual_tokens_used
    - actual_cost
    - execution_time

# Integration
integration:
  workflow_frontmatter: true
  agent_activation: true
  cli_flag: "--model"
  env_variable: "BYAN_MODEL"