imladris/hw/straylight/default.nix

{ inputs, pkgs, lib, ... }:
let
  pkgsUnstable = inputs.nixpkgs-unstable.legacyPackages.${pkgs.system};
  # Pin to b9828; nixpkgs-unstable lags too far behind for the features we need.
  # To update: nix-prefetch-url --unpack https://github.com/ggml-org/llama.cpp/archive/refs/tags/bNNNN.tar.gz
  # then: nix run nixpkgs#prefetch-npm-deps -- <unpacked>/tools/ui/package-lock.json
  llama-cpp-b9828 = (pkgsUnstable.llama-cpp.override { vulkanSupport = true; }).overrideAttrs (_: {
    version = "9828";
    src = pkgsUnstable.fetchFromGitHub {
      owner = "ggml-org";
      repo = "llama.cpp";
      tag = "b9828";
      hash = "sha256-CRj+smOs5T80fT5OqHZGbko7/PwChhYsc8MGZZkE/dQ=";
    };
    npmDepsHash = "sha256-X1DZgmhS/zHTqDT5zq0kywwntthcJ9vRXeqyO3zz6UU=";
  });
in
{
  imports = [
    ./hardware-configuration.nix
    ../../configs/desktop.nix
    ../../configs/qemu.nix
    ../../configs/mcpkg.nix
    ../../configs/mcp.nix # MCP agent + mcp user (straylight is becoming the core host)
  ];

  config = {
    # llama.cpp server with Vulkan backend for gfx1151 (Strix Halo).
    # Model is downloaded from HuggingFace on first start and cached in StateDirectory.
    # NOTE: BIOS "UMA Frame Buffer Size" must be ≥32GB for a 23GB model to fit on-GPU.
    users.users.llama-server = {
      isSystemUser = true;
      group = "llama-server";
      home = "/var/lib/llama-server";
    };
    users.groups.llama-server = {};

    systemd.services.llama-server = {
      description = "llama.cpp inference server";
      wantedBy = [ "multi-user.target" ];
      after = [ "network-online.target" ];
      wants = [ "network-online.target" ];
      environment = {
        HOME = "/var/lib/llama-server";
        HF_HOME = "/var/lib/llama-server/huggingface";
      };
      serviceConfig = {
        ExecStart = ''
          ${llama-cpp-b9828}/bin/llama-server \
            --jinja \
            -hf deepreinforce-ai/Ornith-1.0-35B-GGUF:Q4_K_M \
            --host 127.0.0.1 \
            --port 11434 \
            -c 262144 \
            --no-mmap \
            -fa on \
            --cache-type-k q8_0 \
            --cache-type-v q8_0 \
            -ngl all
        '';
        User = "llama-server";
        Group = "llama-server";
        StateDirectory = "llama-server";
        WorkingDirectory = "/var/lib/llama-server";
        SupplementaryGroups = [ "render" "video" ];
        Restart = "on-failure";
        RestartSec = "10s";
        TimeoutStartSec = "600";
      };
    };
    # straylight is the unikernel host. The shared mcp.nix locks the agent
    # down with PrivateDevices=true, which hides /dev/kvm and /dev/net/tun.
    # Relax that here (only on straylight) so the agent can boot Nanos
    # unikernel VMs under QEMU/KVM and (Phase 2) manage TAP devices.
    systemd.services.mcp-agent.serviceConfig = {
      PrivateDevices = lib.mkForce false;
      DeviceAllow = [ "/dev/kvm rw" "/dev/net/tun rw" ];
      SupplementaryGroups = [ "kvm" ];
      AmbientCapabilities = [ "CAP_NET_ADMIN" ];
      # The agent launches each unikernel as a daemonized QEMU process in its
      # own cgroup. With the default KillMode=control-group, restarting the
      # agent would SIGKILL every running VM. KillMode=process kills only the
      # agent's main process on stop/restart, so VMs survive an agent upgrade.
      # (If a VM does die, the agent's startup Recover restarts it.)
      KillMode = lib.mkForce "process";
    };
    # Let the mcp user reach /dev/kvm directly as well.
    users.users.mcp.extraGroups = [ "kvm" ];

    # Isolated host-only bridge for unikernel VMs (Phase 2). Each unikernel
    # gets a TAP on this bridge and a 10.99.0.0/24 static IP. The bridge has
    # NO uplink and NO NAT, and the firewall drops any VM traffic leaving the
    # bridge, so a VM can reach only the host gateway (10.99.0.1) -- mediation
    # is enforced by network topology, not convention.
    networking.bridges.mcp-br0.interfaces = [ ];
    networking.interfaces.mcp-br0.ipv4.addresses = [
      { address = "10.99.0.1"; prefixLength = 24; }
    ];
    # The host accepts traffic from VMs (so mc-proxy on the gateway can serve
    # them); the FORWARD drop prevents VMs from routing anywhere off-bridge.
    networking.firewall.trustedInterfaces = [ "mcp-br0" ];
    networking.firewall.extraCommands = ''
      iptables -D FORWARD -i mcp-br0 ! -o mcp-br0 -j DROP 2>/dev/null || true
      iptables -A FORWARD -i mcp-br0 ! -o mcp-br0 -j DROP
    '';
    networking.firewall.extraStopCommands = ''
      iptables -D FORWARD -i mcp-br0 ! -o mcp-br0 -j DROP 2>/dev/null || true
    '';

    # Allow rootless containers (podman) to bind low ports (53 for MCNS,
    # 443/8443/9443 for mc-proxy) as straylight takes over the core role.
    boot.kernel.sysctl."net.ipv4.ip_unprivileged_port_start" = 53;

    services.open-webui = {
      enable = true;
      host = "0.0.0.0";
      port = 8080;
      environment = {
        SCARF_NO_ANALYTICS = "True";
        DO_NOT_TRACK = "True";
        ANONYMIZED_TELEMETRY = "False";
        # Point at the local llama.cpp server (OpenAI-compatible API).
        OPENAI_API_BASE_URLS = "http://127.0.0.1:11434/v1";
        OPENAI_API_KEYS = "none";
        ENABLE_OLLAMA_API = "False";
      };
    };

    # Open ports: DNS (53), mc-proxy (443/8443/9443), agent (9444), master (9555), open-webui (8080).
    networking.firewall.allowedTCPPorts = [ 53 443 8080 8443 9443 9444 9555 ];
    networking.firewall.allowedUDPPorts = [ 53 ];

    # DNS: MCNS for internal zones, public resolvers as fallback.
    networking.nameservers = [
      "192.168.88.181"
      "100.95.252.120"
      "1.1.1.1"
      "8.8.8.8"
    ];
    services.resolved.domains = [
      "~mcp.metacircular.net"
    ];
  };
}