refactor of ssh-notify

This commit is contained in:
ari committed 2026-09-18 16:34:31 -04:00
1 parent 466080ead9
commit a781b759f4
8 files changed
+341 -58

No files matched your search

+2
View File
@@ -29,6 +29,8 @@
niri-touchscreen-gestures = final.callPackage ../../pkgs/niri-touchscreen-gestures { };
# Buzz Desktop AppImage wrapper, see pkgs/buzz-desktop.
buzz-desktop = final.callPackage ../../pkgs/buzz-desktop { };
# Per-connection SSH notification forwarder, see pkgs/ssh-notify.
ssh-notify = final.callPackage ../../pkgs/ssh-notify { };
# Claude Desktop (Linux), see pkgs/claude-desktop.
claude-desktop = claude-desktop.packages.${final.stdenv.hostPlatform.system}.default;
# herdr (herdrdev/herdr) — see flake.nix input.
+31 -49
View File
@@ -2,7 +2,23 @@
let
cfg = config.skyg.nixos.common.ssh-notify;
socketPath = "/tmp/ssh-notify.sock";
# Client-side listen socket. Fixed/shared is fine here: socat's `fork`
# mode accepts unbounded connections over this path's lifetime, one per
# `notify-send` call from any concurrently forwarded SSH session.
#
# NOTE: assumes a single interactive uid (1000) on the desktop host, same
# as the rest of this single-user setup.
localSocketPath = "/run/user/1000/ssh-notify/listener.sock";
# Remote-side (server) forwarded socket. Uses OpenSSH's %C token (hash of
# local host/port/user + remote host/port/user, requires OpenSSH >= 8.9)
# so each concurrent SSH connection gets its own bind path on the server —
# this is what the old shared /tmp/ssh-notify.sock path got wrong: only
# one concurrent session could ever bind it, and every other session's
# `RemoteForward` silently failed ("remote port forwarding failed for
# listen path ..."), falling back to a broken direct D-Bus notify-send.
remoteSocketPath = "/tmp/ssh-notify-%C.sock";
# Detect if this host is acting as the notification client (the machine you sit in front of)
isClient =
@@ -12,28 +28,6 @@ let
isServer =
cfg.role == "server"
|| (cfg.role == "auto" && !isClient);
realNotifySend = "${pkgs.libnotify}/bin/notify-send";
# Thin wrapper installed in PATH. On servers (when SSH'd) it forwards via the socket.
notifyWrapper = pkgs.writeShellScriptBin "notify-send" ''
#!${pkgs.runtimeShell}
set -euo pipefail
if [ -n "''${SSH_CONNECTION:-}" ] && [ -S "${socketPath}" ]; then
host="$(hostname -s 2>/dev/null || hostname)"
# Send two lines: prefixed title + body (or remaining args joined)
# This keeps the protocol simple; complex notify-send flags are passed through as-is to the real binary on the client.
printf '[%s] ' "$host"
printf '%s ' "$@"
printf '\n'
# For body we just send the same line for now; user can refine later if needed.
# The listener below will treat the received line as both title and body (notify-send accepts that).
printf '%s\n' "$*"
else
exec ${realNotifySend} "$@"
fi
'';
in
{
options.skyg.nixos.common.ssh-notify = {
@@ -52,41 +46,29 @@ in
};
config = lib.mkIf cfg.enable {
# Always provide socat + real libnotify + our wrapper (wrapper shadows notify-send when earlier in PATH)
environment.systemPackages = [
pkgs.socat
pkgs.libnotify
notifyWrapper
];
environment.systemPackages = [ pkgs.ssh-notify ];
# Client: ask SSH to forward the notification socket back to us on every connection
# Client: ask SSH to forward this connection's notification socket back
# to the shared local listener, and tell the server-side wrapper (via
# SetEnv, expanded here with the same %C the RemoteForward bind uses)
# which socket to connect to.
programs.ssh.extraConfig = lib.mkIf isClient ''
# Forward remote notify-send traffic to the local listener
RemoteForward ${socketPath} ${socketPath}
RemoteForward ${remoteSocketPath} ${localSocketPath}
SetEnv SSH_NOTIFY_SOCK=${remoteSocketPath}
'';
# Client listener: receives messages over the forwarded socket and shows them locally
# Server: accept the SSH_NOTIFY_SOCK env var forwarded by the client
# above (SSH only passes through an allow-listed set by default).
services.openssh.settings.AcceptEnv = lib.mkIf isServer [ "SSH_NOTIFY_SOCK" ];
# Client listener: one persistent socket, forked per connection, that
# turns each incoming JSON payload into a real local notification.
systemd.user.services.ssh-notify-listener = lib.mkIf isClient {
description = "Receive notify-send from remote hosts over SSH and display locally";
after = [ "default.target" ];
wantedBy = [ "default.target" ];
serviceConfig = {
ExecStart =
let
listenerScript = pkgs.writeShellScript "ssh-notify-listener" ''
set -euo pipefail
# Remove stale socket if present
rm -f ${socketPath}
# Listen and for each connection read two lines (title line, body line)
# and invoke the real notify-send. We keep it running via systemd Restart.
while true; do
${pkgs.socat}/bin/socat \
UNIX-LISTEN:${socketPath},fork,mode=0666 \
SYSTEM:'read title; read body; ${realNotifySend} "$title" "$body" || true'
done
'';
in
"${listenerScript}";
ExecStart = "${pkgs.ssh-notify}/bin/ssh-notify-listener";
Restart = "always";
RestartSec = 2;
};
+143 -9
View File
@@ -20,13 +20,30 @@ let
"--port ${toString port}"
(optionalString (ctx != null) "--ctx ${toString ctx}")
(optionalString (tokens != null) "--tokens ${toString tokens}")
(optionalString (threads != null) "--threads ${toString threads}")
(optionalString (batchSize != null) "--batch-size ${toString batchSize}")
(optionalString (power != null) "--power ${toString power}")
(optionalString (prefillChunk != null) "--prefill-chunk ${toString prefillChunk}")
(optionalString (batchedSession != null) "--batched-session ${toString batchedSession}")
(optionalString (mixedPrefillQuantum != null) "--mixed-prefill-quantum ${toString mixedPrefillQuantum}")
(optionalString (vision != null) "--vision ${escapeShellArg vision}")
(optionalString (trace != null) "--trace ${escapeShellArg trace}")
(optionalString (kvDiskDir != null) "--kv-disk-dir ${escapeShellArg kvDiskDir}")
(optionalString (kvDiskSpaceMb != null) "--kv-disk-space-mb ${toString kvDiskSpaceMb}")
(optionalString noKvOffload "--no-kv-offload")
(optionalString (kvCacheMinTokens != null) "--kv-cache-min-tokens ${toString kvCacheMinTokens}")
(optionalString (kvCacheColdMaxTokens != null) "--kv-cache-cold-max-tokens ${toString kvCacheColdMaxTokens}")
(optionalString (kvCacheContinuedIntervalTokens != null) "--kv-cache-continued-interval-tokens ${toString kvCacheContinuedIntervalTokens}")
(optionalString (kvCacheBoundaryTrimTokens != null) "--kv-cache-boundary-trim-tokens ${toString kvCacheBoundaryTrimTokens}")
(optionalString (kvCacheBoundaryAlignTokens != null) "--kv-cache-boundary-align-tokens ${toString kvCacheBoundaryAlignTokens}")
(optionalString kvCacheRejectDifferentQuant "--kv-cache-reject-different-quant")
(optionalString (toolMemoryMaxIds != null) "--tool-memory-max-ids ${toString toolMemoryMaxIds}")
(optionalString cors "--cors")
(optionalString (gpuDeviceIds != null) "--gpu ${escapeShellArg gpuDeviceIds}")
(optionalString (gpuDevices != null) "--gpu-devices ${escapeShellArg gpuDevices}")
(optionalString ssdStreaming "--ssd-streaming")
(optionalString ssdStreamingCold "--ssd-streaming-cold")
(optionalString (ssdStreamingCacheExperts != null) "--ssd-streaming-cache-experts ${escapeShellArg ssdStreamingCacheExperts}")
(optionalString (ssdStreamingFullLayers != null) "--ssd-streaming-full-layers ${toString ssdStreamingFullLayers}")
(optionalString (ssdStreamingPreloadExperts != null) "--ssd-streaming-preload-experts ${toString ssdStreamingPreloadExperts}")
]);
in
{
@@ -77,6 +94,13 @@ in
type = types.nullOr types.int;
};
tokens = mkOption {
description = "Default max output tokens when clients omit a limit.";
default = null;
example = 4096;
type = types.nullOr types.int;
};
threads = mkOption {
description = "Number of inference threads (default: CPU core count).";
default = null;
@@ -84,13 +108,48 @@ in
type = types.nullOr types.int;
};
batchSize = mkOption {
description = "Token batch size for prompt processing.";
power = mkOption {
description = "GPU duty-cycle target, 1..100 (default: 100).";
default = null;
example = 512;
example = 80;
type = types.nullOr types.int;
};
prefillChunk = mkOption {
description = "Graph prefill chunk size in tokens.";
default = null;
example = 4096;
type = types.nullOr types.int;
};
batchedSession = mkOption {
description = "Keep N resident sessions and batch decode-ready requests.";
default = null;
example = 4;
type = types.nullOr types.int;
};
mixedPrefillQuantum = mkOption {
description = "Prefill chunk size while generations are active (default: 128; GLM-5.3 minimum: 1024).";
default = null;
example = 128;
type = types.nullOr types.int;
};
vision = mkOption {
description = "Path to the vision encoder GGUF for the selected model.";
default = null;
example = "/var/lib/ds4/vision.gguf";
type = types.nullOr types.str;
};
trace = mkOption {
description = "Write prompts, cache decisions, output, and tool calls to this file.";
default = null;
example = "/var/lib/ds4/trace.log";
type = types.nullOr types.str;
};
kvDiskDir = mkOption {
description = "Directory for KV cache disk offloading (persists across restarts).";
default = null;
@@ -105,25 +164,100 @@ in
type = types.nullOr types.int;
};
noKvOffload = mkOption {
description = "Disable KV cache offloading to disk entirely.";
kvCacheMinTokens = mkOption {
description = "Do not save/load KV checkpoints shorter than this many tokens (default: 512).";
default = null;
example = 512;
type = types.nullOr types.int;
};
kvCacheColdMaxTokens = mkOption {
description = "Save cold first prompts up to this many tokens; 0 disables (default: 30000).";
default = null;
example = 30000;
type = types.nullOr types.int;
};
kvCacheContinuedIntervalTokens = mkOption {
description = "Save aligned continued frontiers every N tokens; 0 disables (default: 10000).";
default = null;
example = 10000;
type = types.nullOr types.int;
};
kvCacheBoundaryTrimTokens = mkOption {
description = "Trim tail tokens for cold boundary saves (default: 32).";
default = null;
example = 32;
type = types.nullOr types.int;
};
kvCacheBoundaryAlignTokens = mkOption {
description = "Align cold boundary saves to this multiple (default: 2048).";
default = null;
example = 2048;
type = types.nullOr types.int;
};
kvCacheRejectDifferentQuant = mkOption {
description = "Reject KV checkpoints written with different routed-expert quantization.";
default = false;
type = types.bool;
};
toolMemoryMaxIds = mkOption {
description = "Exact tool-call IDs kept in RAM (default: 100000).";
default = null;
example = 100000;
type = types.nullOr types.int;
};
cors = mkOption {
description = "Enable CORS headers for browser-based clients.";
default = false;
type = types.bool;
};
gpuDeviceIds = mkOption {
gpuDevices = mkOption {
description = "GPU device IDs for multi-GPU inference (e.g. '0,1').";
default = null;
example = "0,1";
type = types.nullOr types.str;
};
ssdStreaming = mkOption {
description = "Opt in to SSD-backed model streaming instead of full residency (Metal/CUDA/ROCm).";
default = false;
type = types.bool;
};
ssdStreamingCacheExperts = mkOption {
description = "SSD streaming cache target: N requests dynamic expert slots; NGB also reserves two full prefill layers.";
default = null;
example = "64";
type = types.nullOr types.str;
};
ssdStreamingCold = mkOption {
description = "SSD streaming: skip the default popularity-based expert-cache preload.";
default = false;
type = types.bool;
};
ssdStreamingFullLayers = mkOption {
description = "GLM Metal streaming: keep the first N routed layers fully resident (default: auto; 0 disables).";
default = null;
example = 8;
type = types.nullOr types.int;
};
ssdStreamingPreloadExperts = mkOption {
description = "SSD streaming: upfront popularity preload count (DeepSeek auto-seeds by default; GLM demand-fills unless set).";
default = null;
example = 128;
type = types.nullOr types.int;
};
environment = mkOption {
description = ''
Extra environment variables for the ds4-server process. Useful for GPU