Merge pull request 'Dixie' (#1) from dixie into master

Reviewed-on: #1

configs/desktop-packages-full.nix

@@ -16,6 +16,7 @@
     clang-tools
     elinks
     freeglut
+    gh
     glfw
     glm
     inkscape
@@ -25,6 +26,7 @@
     libGL
     libGL.dev
     mesa
+    nodejs
     nyxt
     opencode
     qmk

configs/desktop-packages-light.nix

@@ -12,6 +12,7 @@ let system = pkgs.stdenv.hostPlatform.system; in
     bitwarden-cli
     bitwarden-desktop
     bluez
+    cliamp
     cmus
     deja-dup
     direnv

configs/opencode.nix

@@ -17,7 +17,7 @@ let
   cfg = {
     "$schema" = "https://opencode.ai/config.json";
     enabled_providers = [ "llama.cpp" "zai-coding-plan" ];
-    model = "llama.cpp/laguna-s-2.1";
+    model = "llama.cpp/ornith-1.5-35b-a3b";
     small_model = "llama.cpp/ornith-1.5-9b-uncensored";
     provider = {
       "zai-coding-plan" = {
@@ -52,7 +52,6 @@ let
           chunkTimeout = 7200000;
         };
         models = {
-          "laguna-s-2.1" = text "Laguna S 2.1";
           "qwen3.6-35b-a3b-abliterated" = text "Qwen3.6 35B A3B";
           "gemma4-26b-a4b-abliterated" = text "Gemma 4 26B A4B";
           "ornith-1.0-35b" = text "Ornith 1.0 35B";

flake.nix

@@ -67,6 +67,7 @@
     lib = nixpkgs.lib;
   in {
     nixosConfigurations = lib.genAttrs [
+      "dixie"
       "imladris"
       "ono-sendai"
       "orion"

hw/dixie/authorized_keys.pub

@@ -0,0 +1,4 @@
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIDXu7ME91tLfU+Q3fNFe8VyIouVvoktoYy8xVNpObxuv kyle@titan.local
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIGea83yMIdCi0QUUPgmhRgIrii7lS1dYxZ6LSxSsDOph kyle@europa
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIHukZAxnlMSoTlRtRF14mLbC0x+rb1zlxnSLVLg+6Xbh kyle@vade
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAILw318jrtFJw1Cn3WnOk17Gkr8bo7NOACQwMQ7S4hvoL kyle@straylight

hw/dixie/default.nix

@@ -0,0 +1,186 @@
+{ inputs, pkgs, lib, config, ... }:
+let
+  # Unstable instantiated with unfree allowed: CUDA is unfree, and legacyPackages does not
+  # inherit the host's nixpkgs.config (straylight's Vulkan build never hit this).
+  pkgsUnstable = import inputs.nixpkgs-unstable {
+    system = pkgs.stdenv.hostPlatform.system;
+    config = { allowUnfree = true; cudaSupport = true; };
+  };
+  # llama.cpp with the CUDA backend, from unstable to match straylight's build (same model
+  # files, same router semantics). Built once on straylight and shipped with dixie-push
+  # --closure, or pulled from the cache when Dixie's link allows.
+  llama-cpp = pkgsUnstable.llama-cpp.override { cudaSupport = true; };
+  modelsDir = "/var/lib/llama-server/models";
+
+  # Dixie is the fleet's helper tier: everything that is latency-sensitive, small, and was
+  # starving on straylight's CPU (2026-09-21). One RTX 3060 (12 GB) today; card 2 joins
+  # when its fans are replaced.
+  #
+  # Router: the three models that must never be swapped out — approval guardian / Honcho
+  # deriver (9B), Honcho dialectic (4B, non-thinking), Honcho embeddings — resident together:
+  # 5.6 + 2.5 + 0.7 GB weights, ~3 GB left for KV. Vision does NOT go here: 5 GB + 1.2 GB
+  # mmproj does not fit beside them, so it gets its own on-demand process below.
+  llamaModelsIni = pkgs.writeText "llama-models.ini" ''
+    version = 1
+    [*]
+    jinja = true
+    flash-attn = on
+    cache-type-k = q8_0
+    cache-type-v = q8_0
+    n-gpu-layers = 999
+    ; three resident models share 12 GB: keep contexts modest
+    ctx-size = 16384
+    parallel = 2
+    cache-ram = 512
+
+    ; Ornith 1.5 9B, refusal-ablated. Hermes auxiliary model (approval guardian, title,
+    ; web_extract) for all agents and Honcho's deriver/summary/dream model.
+    [ornith-1.5-9b-uncensored]
+    model = ${modelsDir}/ornith-1.5-9b-uncensored/Ornith-1.5-9B-uncensored.Q4_K_M.gguf
+    ; Runaway guard carried over from straylight: helper tasks need short answers.
+    reasoning-budget = 2048
+    n-predict = 4096
+
+    ; Honcho dialectic: non-thinking instruct model so answers are short and fast.
+    [honcho-dialectic]
+    model = ${modelsDir}/qwen3-4b-instruct-2507/Qwen3-4B-Instruct-2507-Q4_K_M.gguf
+    n-predict = 1024
+
+    ; Honcho embeddings (1024-dim; EMBEDDING_VECTOR_DIMENSIONS=1024 on rift).
+    [honcho-embed]
+    model = ${modelsDir}/qwen3-embedding-0.6b/Qwen3-Embedding-0.6B-Q8_0.gguf
+    embedding = true
+    pooling = cls
+    ubatch-size = 8192
+    ctx-size = 8192
+    parallel = 4
+  '';
+
+  # Vision (auxiliary.vision for all agents): its own server on :11433, loaded on first
+  # request and dropped again after an idle hour so the router's three keep their VRAM.
+  visionModelsIni = pkgs.writeText "llama-vision.ini" ''
+    version = 1
+    [*]
+    jinja = true
+    flash-attn = on
+    n-gpu-layers = 999
+    ctx-size = 16384
+    parallel = 1
+    sleep-idle-seconds = 3600
+
+    [qwen3-vl-8b-abliterated]
+    model = ${modelsDir}/qwen3-vl-8b-abliterated/Qwen3-VL-8B-Instruct-abliterated-v2.0.Q4_K_M.gguf
+    mmproj = ${modelsDir}/qwen3-vl-8b-abliterated/Qwen3-VL-8B-Instruct-abliterated-v2.0.mmproj-f16.gguf
+  '';
+
+  llamaService = { description, port, preset, modelsMax }: {
+    inherit description;
+    wantedBy = [ "multi-user.target" ];
+    after = [ "network-online.target" "var-lib-llama\\x2dserver-models.mount" ];
+    wants = [ "network-online.target" ];
+    environment.HOME = "/var/lib/llama-server";
+    serviceConfig = {
+      ExecStart = ''
+        ${llama-cpp}/bin/llama-server \
+          --host 0.0.0.0 \
+          --port ${toString port} \
+          --models-preset ${preset} \
+          --models-max ${toString modelsMax} \
+          --timeout 7200
+      '';
+      User = "llama-server";
+      Group = "llama-server";
+      StateDirectory = "llama-server";
+      WorkingDirectory = "/var/lib/llama-server";
+      SupplementaryGroups = [ "video" ];
+      Restart = "on-failure";
+      RestartSec = "10s";
+      TimeoutStartSec = "600";
+    };
+  };
+in
+{
+  imports = [
+    ./hardware-configuration.nix
+    ./disk-config.nix
+  ];
+
+  config = {
+    # Headless. The base configuration.nix has no desktop; nothing GUI is added here.
+    boot.loader.systemd-boot.enable = true;
+    boot.loader.efi.canTouchEfiVariables = true;
+    # 6.12 LTS: mt7921u (the AX9L USB Wi-Fi) is in-tree from 5.18; avoids the 6.17/6.18
+    # mt7921 monitor-mode regression for good measure.
+    boot.kernelPackages = pkgs.linuxPackages_6_12;
+
+    # NVIDIA, compute only: no X, no modesetting needed for CUDA.
+    hardware.graphics.enable = true;
+    services.xserver.videoDrivers = [ "nvidia" ];
+    hardware.nvidia = {
+      open = false;
+      modesetting.enable = false;
+      nvidiaPersistenced = true;
+      package = config.boot.kernelPackages.nvidiaPackages.production;
+    };
+    hardware.enableRedistributableFirmware = true;
+
+    # Power cap for the 3060(s): thermal hour on 2026-09-21 held 75 °C at 140 W, case closed.
+    systemd.services.nvidia-power-cap = {
+      description = "Cap GPU power (thermals in the Aurora chassis)";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "nvidia-persistenced.service" ];
+      serviceConfig = {
+        Type = "oneshot";
+        RemainAfterExit = true;
+        ExecStart = "${config.hardware.nvidia.package.bin}/bin/nvidia-smi -pl 140";
+      };
+    };
+
+    # Headless and Wi-Fi only: there is no console to add a key from after install, so the
+    # keys go in with the system. Same set the live ISO used (~/src/dixie-iso).
+    users.users.kyle.openssh.authorizedKeys.keyFiles = [ ./authorized_keys.pub ];
+    users.users.root.openssh.authorizedKeys.keyFiles = [ ./authorized_keys.pub ];
+
+    users.users.llama-server = { isSystemUser = true; group = "llama-server"; home = "/var/lib/llama-server"; };
+    users.groups.llama-server = { };
+
+    # Model weights live on the USB cache stick (ext4, label dixie-cache), populated from
+    # straylight with dixie-push. Both servers wait for the mount.
+    fileSystems."/mnt/cache" = {
+      device = "/dev/disk/by-label/dixie-cache";
+      fsType = "ext4";
+      options = [ "noatime" "nofail" "x-systemd.device-timeout=30s" ];
+    };
+    fileSystems."/var/lib/llama-server/models" = {
+      device = "/mnt/cache/models";
+      fsType = "none";
+      options = [ "bind" "nofail" "x-systemd.requires=/mnt/cache" ];
+    };
+
+    systemd.services.llama-server = llamaService {
+      description = "llama.cpp router: helper tier (9B guardian/deriver, dialectic, embeddings)";
+      port = 11434; preset = llamaModelsIni; modelsMax = 3;
+    };
+    systemd.services.llama-vision = llamaService {
+      description = "llama.cpp vision server (qwen3-vl-8b, on demand)";
+      port = 11433; preset = visionModelsIni; modelsMax = 1;
+    };
+
+    # Wi-Fi only host. NetworkManager comes from the base config; the AX9L (mt7921u) becomes
+    # the primary once plugged in and the internal CNVi radio stays as fallback. Connection
+    # profiles are copied from the live-USB session at install time (see the install notes),
+    # not stored in this repo.
+    networking.networkmanager.wifi.backend = "wpa_supplicant";
+
+    # Reachable only over the tailnet: 11434 (router) and 11433 (vision) on tailscale0, and
+    # SSH from the LAN (base config opens 22) for bring-up.
+    networking.firewall.interfaces.tailscale0.allowedTCPPorts = [ 11433 11434 ];
+
+    environment.systemPackages = with pkgs; [
+      llama-cpp
+      pciutils usbutils lm_sensors dmidecode nvme-cli smartmontools ethtool iw wpa_supplicant
+      htop tmux git jq curl rsync nvtopPackages.nvidia
+    ];
+
+  };
+}

hw/dixie/disk-config.nix

@@ -0,0 +1,38 @@
+## (dixie) disko config
+## Alienware Aurora R11: Samsung PM9A1 1 TB NVMe is the system disk. The 6 TB SATA HDD and the
+## USB cache stick are left alone here (the stick is mounted by label in default.nix).
+## Unencrypted on purpose: headless box that must come back from a power cut with nobody
+## at the console to unlock it.
+
+{ lib, ... }:
+
+{
+  disko.devices.disk.main = {
+    type = "disk";
+    device = "/dev/disk/by-id/nvme-eui.36355630521205220025384500000001";
+    content = {
+      type = "gpt";
+      partitions = {
+        ESP = {
+          size = "1G";
+          type = "EF00";
+          content = {
+            type = "filesystem";
+            format = "vfat";
+            mountpoint = "/boot";
+            mountOptions = [ "umask=0077" ];
+          };
+        };
+        root = {
+          size = "100%";
+          content = {
+            type = "filesystem";
+            format = "ext4";
+            mountpoint = "/";
+            mountOptions = [ "noatime" ];
+          };
+        };
+      };
+    };
+  };
+}

hw/dixie/hardware-configuration.nix

@@ -0,0 +1,18 @@
+# Do not modify this file!  It was generated by ‘nixos-generate-config’
+# and may be overwritten by future invocations.  Please make changes
+# to /etc/nixos/configuration.nix instead.
+{ config, lib, pkgs, modulesPath, ... }:
+
+{
+  imports =
+    [ (modulesPath + "/installer/scan/not-detected.nix")
+    ];
+
+  boot.initrd.availableKernelModules = [ "xhci_pci" "ahci" "nvme" "usbhid" "usb_storage" "sd_mod" ];
+  boot.initrd.kernelModules = [ ];
+  boot.kernelModules = [ "kvm-intel" ];
+  boot.extraModulePackages = [ ];
+
+  nixpkgs.hostPlatform = lib.mkDefault "x86_64-linux";
+  hardware.cpu.intel.updateMicrocode = lib.mkDefault config.hardware.enableRedistributableFirmware;
+}

hw/straylight/default.nix

@@ -25,6 +25,12 @@ let
     parallel = 2
     ; Unload a model's weights and KV cache after six idle hours; the next request reloads it.
     sleep-idle-seconds = 21600
+    ; Host-RAM prompt cache. This limit is PER RESIDENT MODEL, not global: each model's
+    ; child llama-server gets its own. The llama.cpp default is 8192 MiB, so with
+    ; --models-max 3 the worst case was 3 x 8 = 24 GiB. Helper, VL and comparison models
+    ; rarely reuse long prefixes, so they get 2 GiB; the model that does gets more in its
+    ; own section. Worst case is now 16 + 2 + 2 = 20 GiB.
+    cache-ram = 2048
 
     ; Qwen3.6-35B-A3B, refusal-ablated (HauhauCS "Aggressive"). Benchmark candidate A.
     [qwen3.6-35b-a3b-abliterated]
@@ -42,14 +48,17 @@ let
     ; DFlash GGUF is on disk (laguna-s-2.1-DFlash-BF16.gguf). nixpkgs llama.cpp 0.4.0
     ; rejects it at load ("expected 76, got 69" tensors); Poolside's file needs their
     ; laguna-branch loader. Leave unwired so Laguna still starts.
-    [laguna-s-2.1]
+    ; DISABLED 2026-09-19: resident beside Ornith 35B it left ~15 GB free, and a third
-    model = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/Laguna-S-2.1-UD-Q4_K_S-00001-of-00003.gguf
+    ; model load then hit the 104 GiB TTM cap and OOM-killed the desktop session.
+    ; Weights stay on disk; uncomment to re-enable.
+    ; [laguna-s-2.1]
+    ; model = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/Laguna-S-2.1-UD-Q4_K_S-00001-of-00003.gguf
     ; Two slots sharing one unified 131072-token KV pool: a second request (subagent,
     ; another client) runs alongside instead of queueing, and either one may still use
     ; the full 131072 as long as their combined context fits.
-    parallel = 2
+    ; parallel = 2
-    kv-unified = true
+    ; kv-unified = true
-    ctx-size = 131072
+    ; ctx-size = 131072
 
     ; Previous default model (a Qwen 3.5 derivative), kept for comparison runs. It was
     ; fetched by the old -hf flag into the HF cache; dedup hides the cache's own entry.
@@ -62,6 +71,13 @@ let
     [ornith-1.5-9b-uncensored]
     hf-repo = mradermacher/Ornith-1.5-9B-uncensored-GGUF:Q4_K_M
     dedup-cache-models = true
+    ; Runaway guard (2026-09-21): as Hermes' auxiliary model (approval, title, web
+    ; extract) it got requests with no max_tokens, never converged, and one generation
+    ; ran to 46k tokens (~70 min at 11 t/s) after the client's 60 s timeout, starving the
+    ; 35B. Helper tasks need short answers: cap thinking and total output server-side.
+    ; A client that sends its own smaller max_tokens still wins.
+    reasoning-budget = 2048
+    n-predict = 4096
 
     ; Qwen3-VL-8B Instruct, refusal-ablated (prithivMLmods v2 / mradermacher).
     ; Dedicated VLM in the same size class as Ornith 1.5 9B. Explicit mmproj-url
@@ -82,6 +98,23 @@ let
     top-k = 20
     reasoning-budget = 8192
     dedup-cache-models = true
+    ; Prompt-cache thrash (measured 2026-09-20, 30 h of logs): opencode, Hermes and its
+    ; delegation subagents shared 2 slots and an 8 GiB host cache whose entries are
+    ; 1-2 GiB each. 235 evictions; 138 prompt re-evaluations over 20k tokens costing
+    ; ~196 minutes; one 102,922-token prompt re-processed from scratch (~3.4 min).
+    ;
+    ; Four slots over ONE unified 262144-token KV pool: same GPU memory as before, but
+    ; four conversations keep their prefix live instead of two, and any one of them may
+    ; still use the whole pool when the others are idle. Cost: when all four decode at
+    ; once, each gets a smaller share of ~30 t/s.
+    parallel = 4
+    kv-unified = true
+    ; 16 GiB holds about ten evicted conversations instead of five.
+    ; IF LAGUNA IS RE-ENABLED this must drop to 4096 (Laguna 8192, helpers 0) and
+    ; --models-max to 2: Laguna ~73 GiB + this model ~26 + a 9B helper ~9 + desktop and
+    ; services ~20 + the honcho CPU lanes ~9 is ~137 GiB on a 125 GiB machine before
+    ; any cache at all. That is the 2026-09-19 OOM.
+    cache-ram = 16384
 
     ; Qwen 3.8 27B dense, refusal-ablated. Preferred uncensored coding agent.
     ; orcarouter/Qwen3.8-27B-Uncensored-GGUF is gated; huihui Q4_K_L keeps
@@ -224,7 +257,7 @@ in
             --host 0.0.0.0 \
             --port 11434 \
             --models-preset ${llamaModelsIni} \
-            --models-max 2 \
+            --models-max 3 \
             --timeout 7200
         '';
         User = "llama-server";
@@ -326,12 +359,14 @@ in
     # Pinned to the upstream flake (v0.9.0); nixpkgs-unstable only has 0.7.1.
     environment.systemPackages = [
       inputs.herdr.packages.x86_64-linux.default
-      gh
       llamaModelsCmd
       llamaUnloadCmd
-      nodejs
       whisperTranscribe
       whisperCpp
+      # tirith: pre-exec command security scanner used by Hermes (resolved from
+      # PATH before its self-downloaded ~/.hermes/bin/tirith). From unstable:
+      # 26.05 only has 0.3.3, Hermes ships 0.4.x.
+      pkgsUnstable.tirith
     ];
 
     services.open-webui = {
@@ -396,6 +431,42 @@ in
       };
     };
 
+    # Hermes Desktop (`hermes desktop`) builds and launches an unpackaged
+    # Electron binary from ~/.hermes/hermes-agent/apps/desktop. It runs through
+    # nix-ld, so the Electron/Chromium runtime libraries have to be listed here.
+    programs.nix-ld.libraries = with pkgs; [
+      glib
+      nspr
+      nss
+      dbus
+      at-spi2-atk
+      at-spi2-core
+      cups
+      cairo
+      gtk3
+      pango
+      gdk-pixbuf
+      libx11
+      libxcomposite
+      libxdamage
+      libxext
+      libxfixes
+      libxrandr
+      libxcb
+      libxkbcommon
+      libgbm
+      libdrm
+      libglvnd
+      mesa
+      vulkan-loader
+      expat
+      systemd # libudev
+      alsa-lib
+      libpulseaudio
+      libnotify
+      libsecret
+    ];
+
     # Headless Chromium exposing a CDP endpoint for Hermes browser tools.
     # Loopback only; browser.cdp_url in ~/.hermes/config.yaml points at it.
     # Ungoogled-chromium (already in desktop-packages-light.nix), no ad-block