codecrucible-synth/config/unified-model-config.yaml

Production-Ready AI Development Platform with Multi-Voice Synthesis, Smithery MCP Integration, Enterprise Security, and Zero-Timeout Reliability

# Unified Model Configuration - Single Source of Truth
# Created: August 21, 2025
# Purpose: Resolves configuration conflicts identified in audit report

llm:
  # Default provider selection
  default_provider: "ollama"
  fallback_chain: ["ollama", "lm-studio", "mock"]
  
  providers:
    ollama:
      endpoint: "http://localhost:11434"
      timeout:
        connection: 5000     # 5s for connection establishment
        response: 30000      # 30s for response generation
        status_check: 3000   # 3s for health checks
      models:
        # Preferred models in priority order
        preferred: 
          - "qwen2.5-coder:7b"
          - "qwen2.5-coder:3b"
          - "deepseek-coder:8b"
        # Fallback models if preferred not available
        fallback:
          - "llama3.2:latest"
          - "gemma:latest"
        # Model-specific settings
        settings:
          "qwen2.5-coder:7b":
            temperature: 0.1
            max_tokens: 128000
            context_window: 128000
          "deepseek-coder:8b":
            temperature: 0.2
            max_tokens: 128000
            context_window: 128000
      # Task routing preferences
      optimal_for:
        - "analysis"
        - "planning"
        - "complex"
        - "multi-file"
        - "architecture"
        - "debugging"
    
    lm-studio:
      endpoint: "http://localhost:1234"
      timeout:
        connection: 3000     # 3s for connection
        response: 15000      # 15s for response
        status_check: 2000   # 2s for health checks
      models:
        preferred:
          - "codellama-7b-instruct"
          - "gemma-2b-it"
          - "qwen/qwen2.5-coder-14b"
        fallback:
          - "gpt-3.5-turbo"  # OpenAI compatible fallback
        settings:
          "codellama-7b-instruct":
            temperature: 0.7
            max_tokens: 128000
            context_window: 128000
      optimal_for:
        - "template"
        - "edit"
        - "format"
        - "boilerplate"
        - "quick-fix"
        - "streaming"

  # Routing strategy configuration
  routing:
    strategy: "hybrid"  # Options: hybrid, simple, complex
    
    # Task complexity mapping
    task_complexity:
      simple:
        - "format"
        - "template"
        - "boilerplate"
        - "edit"
        - "rename"
      complex:
        - "analysis"
        - "architecture"
        - "planning"
        - "debug"
        - "security"
        - "multi-file"
    
    # Confidence thresholds for routing decisions
    confidence_thresholds:
      high: 0.9       # Use fast path
      medium: 0.7     # Standard routing
      low: 0.5        # Require validation
      escalation: 0.3 # Escalate to human
    
    # Performance targets
    performance_targets:
      simple_task_response: 1000    # 1s for simple tasks
      complex_task_response: 30000  # 30s for complex tasks
      streaming_latency: 100        # 100ms for first token

  # Monitoring and metrics
  monitoring:
    enabled: true
    metrics_port: 3001
    collect_performance: true
    collect_errors: true
    
    # Alert thresholds
    alerts:
      error_rate: 0.05         # Alert if >5% errors
      response_time_p99: 5000  # Alert if p99 > 5s
      availability: 0.95       # Alert if <95% available

  # Security settings
  security:
    validate_inputs: true
    sanitize_outputs: true
    max_prompt_length: 128000
    max_response_length: 128000
    blocked_patterns:
      - "password"
      - "secret"
      - "api_key"
      - "private_key"

  # Caching configuration
  caching:
    enabled: true
    ttl: 3600               # 1 hour cache TTL
    max_cache_size: 1000    # Max cached responses
    cache_strategy: "lru"   # Least recently used

  # Experimental features
  experimental:
    dual_agent_review: true     # Enable sequential review system
    auto_retry_on_timeout: true # Automatic retry with backoff
    adaptive_timeouts: true     # Adjust timeouts based on task
    connection_pooling: true    # HTTP connection pooling
    circuit_breaker: false      # Not yet implemented