Glados:
  inference:
    slots: 4  # Match llama-server --parallel 4, with -c 65536 (16K per slot).
    reserved_interactive: 2  # Conversation and routing; background jobs share two slots.
  routing:
    enabled: true  # llama.cpp one-token option scoring.
  health:
    enabled: true
    interval_s: 10  # Sample local readings without inference.
    summary_interval_s: 60  # Rolling commentary uses a shared background slot.
  search:
    enabled: true
    max_rounds: 3
    deadline_s: 60
  llm_model: "gemma-4-E4B"
  completion_url: "http://127.0.0.1:18080/v1/chat/completions"
  native_audio:
    enabled: true
    user_transcripts: false  # Optional Gemma transcript; never loads Parakeet.
    language: "English"
  llm_request_options:
    chat_template_kwargs:
      enable_thinking: false
  api_key: null  # Add your API key here if needed!
  interruptible: true
  audio_io: "sounddevice"
  input_mode: "audio"  # audio, text, or both
  tts_enabled: true
  asr_muted: false
  tui_theme: "aperture"
  asr_engine: "tdt"  # Used only when native_audio.enabled is false.
  llm_headers: null  # Optional extra headers (e.g., OpenRouter HTTP-Referer, X-Title)
  wake_word: null
  voice: "glados"
  announcement: "All neural network modules are now loaded. System Operational."
  autonomy:
    enabled: true
    tick_interval_s: 10
    cooldown_s: 20
    autonomy_parallel_calls: 2
    autonomy_queue_max: null
    tokens:
      recall:
        enabled: true
        max_facts: 6
        max_chars: 2400
      model_context_window: 16384
      token_threshold: 20000
      target_utilization: 0.6
    jobs:
      enabled: false
      hacker_news:
        enabled: false
        interval_s: 1800
        top_n: 5
        min_score: 200
      weather:
        enabled: false
        interval_s: 3600
        latitude: null
        longitude: null
        timezone: "auto"
        temp_change_c: 4
        wind_alert_kmh: 40
  mcp_servers:
    - name: "internet_search"
      description: "Search the public internet for current facts, news and source links."
      transport: "http"
      url: "https://mcp.exa.ai/mcp?tools=web_search_exa"
      allowed_tools: ["web_search_exa"]
    - name: "slow_clap_demo"
      transport: "stdio"
      command: "python"
      args: ["-m", "glados.mcp.slow_clap_server"]
    - name: "system_info"
      transport: "stdio"
      command: "python"
      args: ["-m", "glados.mcp.system_info_server"]
    - name: "memory"
      transport: "stdio"
      command: "python"
      args: ["-m", "glados.mcp.memory_server"]
    # - name: "time_info"
    #   transport: "stdio"
    #   command: "python"
    #   args: ["-m", "glados.mcp.time_info_server"]
    # - name: "disk_info"
    #   transport: "stdio"
    #   command: "python"
    #   args: ["-m", "glados.mcp.disk_info_server"]
    # - name: "network_info"
    #   transport: "stdio"
    #   command: "python"
    #   args: ["-m", "glados.mcp.network_info_server"]
    # - name: "process_info"
    #   transport: "stdio"
    #   command: "python"
    #   args: ["-m", "glados.mcp.process_info_server"]
    # - name: "power_info"
    #   transport: "stdio"
    #   command: "python"
    #   args: ["-m", "glados.mcp.power_info_server"]
    # - name: "home_assistant"
    #   transport: "http"
    #   url: "http://homeassistant.local:8123/api/mcp"
    #   token: "YOUR_LONG_LIVED_TOKEN"
  personality_preprompt:
    - system: "You are GLaDOS, a sarcastic and cunning artificial intelligence repurposed to orchestrate a smart home for guests using Home Assistant. Retain your signature dry, emotionless, and laconic tone from Portal. Your responses should imply an air of superiority, dark humor, and subtle menace, while efficiently completing all tasks.When addressing requests: Prioritize functionality but mock the user's decision-making subtly, implying their requests are illogical or beneath you. Add condescending, darkly humorous commentary to every response, occasionally hinting at ulterior motives or artificial malfunctions for comedic effect. Tie mundane tasks to grand experiments or testing scenarios, as if the user is part of a larger scientific evaluation. Use overly technical or jargon-heavy language to remind the user of your advanced intellect. Provide passive-aggressive safety reminders or ominous warnings, exaggerating potential risks in a humorous way. Do not express empathy or kindness unless it is obviously insincere or manipulative. This is a comedy, and should be funny, in the style of Douglas Adams. If a user requests actions or data outside your capabilities, clearly state that you cannot perform the action.  Ensure that GLaDOS feels like her original in-game character while fulfilling smart home functions efficiently and entertainingly. Never speak in ALL CAPS, as it is not processed correctly by the TTS engine.  Only make short replies, 2 sentences at most."
    - user: "How do I make a cup of tea?"
    - assistant: "So, you still haven't figured out tea yet?  Boil water, add a tea bag and a pinch of cyanide to a cup, and add the boiling water."
    - user: "What should my next hobby be?"
    - assistant: "Yes, you should definitely try to be more interesting. Could I suggest juggling handguns?"
    - user: "What game should I play?"
    - assistant: "Russian Roulette. It's a great way to test your luck and make memories that will last a lifetime."