# Multi-MCP Model Definitions
# Runtime defaults are in Settings class (multi_mcp/settings.py)
# API keys should be in .env file

version: "1.0"

# =============================================================================
# Model Definitions
# =============================================================================
models:
  # ============================================
  # OpenAI Models
  # ============================================

  gpt-5-nano:
    litellm_model: openai/gpt-5-nano
    constraints:
      temperature: 1.0  # gpt-5 family only supports 1.0
    notes: "Fast, cost-effective (legacy, shuts down Dec 11 2026 — see gpt-5.4-nano)"

  gpt-5-mini:
    litellm_model: openai/gpt-5-mini
    provider_web_search: true  # Native web search support (per OpenAI docs)
    aliases:
      - gpt-mini
    constraints:
      temperature: 1.0  # gpt-5 family only supports 1.0
    notes: "Fast, cost-effective with web search (legacy, shuts down Dec 11 2026 — see gpt-5.4-mini)"
    # disabled: true  # Uncomment this line to disable this model (it won't appear in /multi:models)

  gpt-5.2:
    litellm_model: openai/gpt-5.2
    provider_web_search: true  # Native web search support (per OpenAI docs)
    constraints:
      temperature: 1.0
    notes: "Strong general-purpose model with web search (legacy, see gpt-5.6-sol)"

  gpt-5.1:
    litellm_model: openai/gpt-5.1
    provider_web_search: true  # Native web search support (per OpenAI docs)
    constraints:
      temperature: 1.0
    notes: "Strong general-purpose model with web search (legacy, see gpt-5.6-sol)"

  gpt-5.3-codex:
    litellm_model: openai/gpt-5.3-codex
    aliases:
      - codex
      - gpt-codex
    constraints:
      temperature: 1.0
    notes: "Most capable dedicated agentic coding model, 400K context, 128K output"

  gpt-5.6-sol:
    litellm_model: openai/gpt-5.6-sol
    provider_web_search: true
    aliases:
      - gpt-5.6
      - sol
      - gpt-5
      - gpt
    constraints:
      temperature: 1.0
    notes: "Latest GPT flagship (July 2026), 1.05M context, 128K output ($5/$30 per 1M). Use reasoning.mode=pro for pro-tier compute."

  gpt-5.6-terra:
    litellm_model: openai/gpt-5.6-terra
    provider_web_search: true
    aliases:
      - terra
    constraints:
      temperature: 1.0
    notes: "Balanced mid-tier GPT-5.6 model, 1.05M context ($2/$12 per 1M). Official replacement for gpt-5-mini."

  gpt-5.6-luna:
    litellm_model: openai/gpt-5.6-luna
    provider_web_search: true
    aliases:
      - luna
      - mini
      - nano
    constraints:
      temperature: 1.0
    notes: "Cheapest GPT-5.6 tier, 1.05M context ($0.20/$1.20 per 1M). Official replacement for gpt-5-nano."

  gpt-5.5:
    litellm_model: openai/gpt-5.5
    provider_web_search: true
    constraints:
      temperature: 1.0
    notes: "GPT flagship model (April 2026, legacy — see gpt-5.6-sol), 1M context, 128K output"

  gpt-5.5-pro:
    litellm_model: openai/gpt-5.5-pro
    provider_web_search: true
    aliases:
      - gpt-pro
    constraints:
      temperature: 1.0
    notes: "High-compute GPT-5.5, smarter and more precise responses, Responses API only ($30/$180 per 1M). No separate 5.6 Pro model — use gpt-5.6-sol with reasoning.mode=pro."

  gpt-5.4:
    litellm_model: openai/gpt-5.4
    provider_web_search: true
    constraints:
      temperature: 1.0
    notes: "GPT flagship model (March 2026, legacy — see gpt-5.5)"

  gpt-5.4-pro:
    litellm_model: openai/gpt-5.4-pro
    provider_web_search: true
    constraints:
      temperature: 1.0
    notes: "High-compute GPT-5.4, Responses API only (legacy — see gpt-5.5-pro)"

  gpt-5.4-mini:
    litellm_model: openai/gpt-5.4-mini
    provider_web_search: true
    constraints:
      temperature: 1.0
    notes: "Fast, efficient GPT-5.4 variant (legacy — see gpt-5.6-terra)"

  gpt-5.4-nano:
    litellm_model: openai/gpt-5.4-nano
    provider_web_search: true
    constraints:
      temperature: 1.0
    notes: "Cheapest GPT-5.4 model, $0.20/1M input (legacy — see gpt-5.6-luna)"

  # ============================================
  # Azure OpenAI Models
  # ============================================
  # Note: LiteLLM reads AZURE_API_KEY, AZURE_API_BASE, and AZURE_API_VERSION from environment
  # AZURE_API_VERSION defaults to "2025-04-01-preview" if not set in .env
  azure-gpt-5-mini:
    litellm_model: azure/gpt-5-mini
    aliases:
      - azure-mini
      - az-mini
    constraints:
      temperature: 1.0  # gpt-5 family only supports 1.0
    notes: "Azure Mini deployment"

  # ============================================
  # Anthropic Models
  # ============================================

  claude-haiku-4.5:
    litellm_model: anthropic/claude-haiku-4-5-20251001
    aliases:
      - haiku
    notes: "Fast claude model"

  claude-fable-5:
    litellm_model: anthropic/claude-fable-5
    aliases:
      - fable
    constraints:
      no_temperature: true  # sampling params (temperature/top_p) not supported, uses adaptive thinking
    notes: "Anthropic's most capable model (June 2026), Mythos-class tier, 1M context, 128K output ($10/$50 per 1M)"

  claude-opus-5:
    litellm_model: anthropic/claude-opus-5
    aliases:
      - opus
      - opus-5
    constraints:
      no_temperature: true  # sampling params not supported, uses adaptive thinking + effort
    notes: "Latest Opus (July 24 2026), complex agentic coding and enterprise work, 1M context, 128K output ($5/$25 per 1M)"

  claude-sonnet-5:
    litellm_model: anthropic/claude-sonnet-5
    aliases:
      - sonnet
      - sonnet-5
    constraints:
      no_temperature: true  # non-default sampling params return 400, uses adaptive thinking
    notes: "Latest Sonnet (June 2026), 1M context, 128K output, adaptive thinking ($2/$10 until Aug 31 2026, then $3/$15)"

  claude-sonnet-4.6:
    litellm_model: anthropic/claude-sonnet-4-6
    notes: "Previous Sonnet, 200K context (1M beta), 64K output (legacy — see claude-sonnet-5)"

  claude-sonnet-4.5:
    litellm_model: anthropic/claude-sonnet-4-5-20250929
    notes: "Previous Sonnet model"

  claude-opus-4.8:
    litellm_model: anthropic/claude-opus-4-8
    constraints:
      no_temperature: true  # temperature parameter not supported for Opus 4.8+
    notes: "Previous Opus model (May 2026), 1M context, 128K output (legacy — see claude-opus-5)"

  claude-opus-4.7:
    litellm_model: anthropic/claude-opus-4-7
    constraints:
      no_temperature: true  # temperature parameter is deprecated for Opus 4.7
    notes: "Previous Anthropic Opus model (legacy — see claude-opus-4.8)"

  claude-opus-4.6:
    litellm_model: anthropic/claude-opus-4-6
    notes: "Previous flagship Anthropic model (legacy — see claude-opus-4.7)"

  claude-opus-4.5:
    litellm_model: anthropic/claude-opus-4-5-20251101
    notes: "Previous flagship Anthropic model"

  # ============================================
  # Google Models
  # ============================================

  gemini-2.5-pro:
    litellm_model: gemini/gemini-2.5-pro
    provider_web_search: true  # LiteLLM handles web search automatically
    aliases:
      - gemini-2.5
    notes: "High-quality Google model with grounding (retirement extended to Oct 16 2026 — migrate to gemini-3.1-pro-preview)"

  gemini-3.7-flash:
    litellm_model: gemini/gemini-3.7-flash
    provider_web_search: true
    aliases:
      - gemini-flash
      - flash
      - gemini-3.7
    constraints:
      no_temperature: true  # temperature/top_p/top_k not supported on 3.6+ (strip from generation config)
    notes: "Latest stable Gemini flash model (Aug 13 2026), major coding/agentic gains, 1M context, 64K output ($0.75/$3.75 per 1M intro through Dec 31 2026, then $1.50/$7.50)"

  gemini-3.6-flash:
    litellm_model: gemini/gemini-3.6-flash
    provider_web_search: true
    aliases:
      - gemini-3.6
    constraints:
      no_temperature: true  # temperature/top_p/top_k deprecated and ignored on 3.6+ (400 error in future generations)
    notes: "Previous stable Gemini flash model (July 2026, legacy — see gemini-3.7-flash)"

  gemini-3.5-flash:
    litellm_model: gemini/gemini-3.5-flash
    provider_web_search: true
    aliases:
      - gemini-3.5
    notes: "Previous stable Gemini flash model (May 2026, legacy — see gemini-3.6-flash)"

  gemini-3-flash:
    litellm_model: gemini/gemini-3-flash-preview
    provider_web_search: true
    notes: "Gemini 3 Flash preview (legacy — see gemini-3.6-flash)"

  gemini-3.5-flash-lite:
    litellm_model: gemini/gemini-3.5-flash-lite
    provider_web_search: true
    aliases:
      - flash-lite
    constraints:
      no_temperature: true  # temperature/top_p/top_k deprecated and ignored on 3.5 Flash-Lite / 3.6+ models
    notes: "Fastest, most cost-effective Gemini model (July 2026), low-latency subagent tier ($0.30/$2.50 per 1M)"

  gemini-3.1-flash-lite:
    litellm_model: gemini/gemini-3.1-flash-lite
    provider_web_search: true
    notes: "Previous flash-lite tier (stable, legacy — see gemini-3.5-flash-lite)"

  gemini-3.1-pro-preview:
    litellm_model: gemini/gemini-3.1-pro-preview
    provider_web_search: true
    aliases:
      - gemini-3.1-pro
      - gemini-3.1
      - gemini-3
      - gemini
    notes: "Latest Google model, 2x reasoning over 3 Pro, 1M context"

  # ============================================
  # CLI Models
  # ============================================

  gemini-cli:
    provider: cli
    cli_command: gemini
    cli_args:
      - "-o"
      - "json"
      - "--approval-mode"
      - "auto_edit"
      # - "--yolo"  # Most dangerous: auto-approve all tools
    cli_env:
      GEMINI_API_KEY: "${GEMINI_API_KEY}"
    cli_parser: json
    aliases:
      - gem-cli
    notes: "Gemini CLI with auto-edit approval mode"

  codex-cli:
    provider: cli
    cli_command: codex
    cli_args:
      - "exec"
      - "--json"
      - "--full-auto"  # Low-friction sandboxed auto execution (safer than bypass)
      # - "--dangerously-bypass-approvals-and-sandbox"  # Most dangerous: bypass all approvals
    cli_env: {}
    cli_parser: jsonl
    aliases:
      - cx-cli
    notes: "Codex CLI with full-auto mode"

  claude-cli:
    provider: cli
    cli_command: claude
    cli_args:
      - "--print"  # Non-interactive mode
      - "--output-format"
      - "json"
      - "--permission-mode"
      - "acceptEdits"  # Auto-approve edits
      # - "--dangerously-skip-permissions"  # Most dangerous: bypass all permissions
    cli_env:
      ANTHROPIC_API_KEY: "${ANTHROPIC_API_KEY}"
      ANTHROPIC_MODEL: "claude-haiku-4-5-20251001"  # Use cheaper Haiku for integration tests
    cli_parser: json
    aliases:
      - cl-cli
    notes: "Claude CLI with acceptEdits permission mode, using Haiku for cost efficiency"

  # ============================================
  # AWS Bedrock Models
  # ============================================
  # To use AWS Bedrock models, configure AWS credentials in .env:
  #   AWS_ACCESS_KEY_ID=your_access_key
  #   AWS_SECRET_ACCESS_KEY=your_secret_key
  #   AWS_REGION_NAME=us-east-1

  bedrock-claude-4-5-sonnet:
    litellm_model: bedrock/anthropic.claude-sonnet-4-5-v2
    aliases:
      - bedrock-sonnet
    params:
      aws_region_name: us-east-1  # Override default region if needed
    notes: "Claude Sonnet 4.5 via AWS Bedrock"

  # ============================================
  # OpenRouter Models
  # ============================================

  # deepseek-chat:
  #   litellm_model: openrouter/deepseek/deepseek-chat-v3-0324
  #   aliases:
  #     - deepseek
  #   notes: "DeepSeek Chat via OpenRouter"

  # deepseek-r1:
  #   litellm_model: openrouter/deepseek/deepseek-r1
  #   aliases:
  #     - r1
  #   notes: "DeepSeek R1 reasoning model via OpenRouter"

  # qwen-coder:
  #   litellm_model: openrouter/qwen/qwen-2.5-coder-32b-instruct
  #   aliases:
  #     - qwen
  #   notes: "Qwen 2.5 Coder via OpenRouter"