rope
aee59ff0b2
Profiled where the ~6s actually went, and it was not where I assumed. Haiku defaults to extended thinking, so it spent 225 output tokens and 1.9s deliberating before emitting an 80-character answer about the height of a tower. MAX_THINKING_TOKENS=0: 43 output tokens, time-to-first-text 2657ms -> 1122ms, wall clock 5.7s -> ~2.4s, answer unchanged. Two things measured NOT to be the bottleneck, recorded in the comment so nobody optimises them later: process startup (0.12s boot plus 28ms to fire the request, so a warm/persistent process saves nothing) and prompt size (184 input tokens, no bloat to trim). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
174 lines
8.4 KiB
Nix
174 lines
8.4 KiB
Nix
{ config, pkgs, lib, ... }:
|
|
{
|
|
config = lib.mkIf (config.networking.hostName == "FredOS-Mediaserver") {
|
|
# Create symlink from home to storage
|
|
systemd.tmpfiles.rules = [
|
|
"L+ /home/fred/storage - - - - /mnt/storage"
|
|
];
|
|
|
|
# Basic system packages
|
|
environment.systemPackages = with pkgs; [
|
|
mergerfs
|
|
wget
|
|
util-linux
|
|
javaPackages.compiler.temurin-bin.jre-25
|
|
unzip
|
|
screen
|
|
yt-dlp
|
|
ghostty.terminfo
|
|
usbutils
|
|
lm_sensors
|
|
(pkgs.writeShellScriptBin "transcode-hevc" ''
|
|
export PATH="${pkgs.jellyfin-ffmpeg}/bin:${pkgs.coreutils}/bin:${pkgs.findutils}/bin:${pkgs.gnugrep}/bin:${pkgs.gawk}/bin:${pkgs.bc}/bin:${pkgs.curl}/bin:$PATH"
|
|
exec ${pkgs.bash}/bin/bash ${../scripts/transcode-hevc.sh} "$@"
|
|
'')
|
|
(pkgs.writeShellScriptBin "record-update" ''
|
|
export PATH="${pkgs.nvd}/bin:${pkgs.coreutils}/bin:${pkgs.gnugrep}/bin:${pkgs.gnused}/bin:$PATH"
|
|
exec ${pkgs.bash}/bin/bash ${../scripts/record-update.sh} "$@"
|
|
'')
|
|
# Stats stream for the quickshell server monitor on the desktops:
|
|
# `ssh mediaserver qs-stats` = top's 2s batch stream, interleaved with
|
|
# one @STAT line per tick (hottest coretemp across both sockets, WAN
|
|
# byte rates from eno1). Everything rides one SSH connection.
|
|
(pkgs.writeShellScriptBin "qs-stats" ''
|
|
# Single writer: the loop re-emits top's lines itself and injects a
|
|
# @STAT line at each frame header. top writing the pipe directly in
|
|
# parallel raced the injected lines (pipe writes aren't line-atomic)
|
|
# and spliced @STAT mid-frame, where the client parser never saw it.
|
|
C=${pkgs.coreutils}/bin
|
|
export LC_ALL=C
|
|
prev_rx=""
|
|
${pkgs.procps}/bin/top -b -d 2 -w 512 | while IFS= read -r line; do
|
|
printf '%s\n' "$line"
|
|
case $line in
|
|
"top - "*)
|
|
t=0
|
|
for h in /sys/class/hwmon/*; do
|
|
[ "$($C/cat "$h/name" 2>/dev/null)" = coretemp ] || continue
|
|
for f in "$h"/temp*_input; do
|
|
v=$($C/cat "$f" 2>/dev/null || echo 0)
|
|
[ "$v" -gt "$t" ] && t=$v
|
|
done
|
|
done
|
|
read -r rx tx < <(${pkgs.gawk}/bin/awk '$1 == "eno1:" {print $2, $10}' /proc/net/dev)
|
|
if [ -n "$prev_rx" ]; then
|
|
printf '@STAT temp=%s rxbps=%s txbps=%s\n' "$((t / 1000))" "$(( (rx - prev_rx) / 2 ))" "$(( (tx - prev_tx) / 2 ))"
|
|
fi
|
|
prev_rx=$rx
|
|
prev_tx=$tx
|
|
;;
|
|
esac
|
|
done
|
|
'')
|
|
# Instant-answer backend for the quickshell launcher on the desktops:
|
|
# `printf '%s' "question" | ssh mediaserver qs-ask`.
|
|
#
|
|
# Runs through Claude Code so it draws on fred's existing subscription —
|
|
# no API key, no separate billing. The trade is latency: Claude Code boots
|
|
# a Node process per question, so this takes ~6s against the raw API's
|
|
# ~1s. The launcher therefore shows its Wikipedia answer first and lets
|
|
# this one supersede it when it lands.
|
|
#
|
|
# Reads the question from stdin — passing it as an ssh argv element would
|
|
# send it through the remote shell for a second round of word splitting.
|
|
#
|
|
# Every flag here is either isolation or trimming the invocation:
|
|
# --safe-mode no CLAUDE.md, skills, hooks, plugins, MCP or
|
|
# custom agents. Critical for correctness, not
|
|
# just speed: fred's ~/.claude/CLAUDE.md tells
|
|
# Claude to answer in caveman mode, which would
|
|
# otherwise leak into launcher answers. Auth
|
|
# keeps working normally, which is what makes
|
|
# the subscription usable here.
|
|
# --no-session-persistence writes no transcript, so launcher questions
|
|
# never appear in /resume or any project history
|
|
# --tools "" no tools at all: nothing to load, and it
|
|
# cannot touch the filesystem or network
|
|
# --disable-slash-commands no skill resolution
|
|
# --strict-mcp-config ignore every MCP config (none supplied)
|
|
# --system-prompt replaces the full Claude Code system prompt
|
|
# with one line, rather than appending to it
|
|
# --permission-mode dontAsk never block waiting on a prompt nobody sees
|
|
# --model claude-haiku-4-5 smallest model — a one-line fact needs no more
|
|
#
|
|
# NOT --bare, which looks like the right flag and is a trap: it skips even
|
|
# more, but its auth is "strictly ANTHROPIC_API_KEY or apiKeyHelper (OAuth
|
|
# and keychain are never read)" — i.e. it cannot use the subscription.
|
|
#
|
|
# STDOUT is the launcher's channel: the answer, or nothing at all if the
|
|
# call fails or the model doesn't know — silence is what the launcher
|
|
# reads as "keep the Wikipedia answer". Every reason for that silence goes
|
|
# to STDERR instead, which the launcher discards and a human running
|
|
# `qs-ask` by hand can read.
|
|
(pkgs.writeShellScriptBin "qs-ask" ''
|
|
q=$(${pkgs.coreutils}/bin/cat)
|
|
[ -z "$q" ] && { echo "qs-ask: no question on stdin" >&2; exit 0; }
|
|
|
|
# Claude Code derives its project state (transcripts, auto-memory) from
|
|
# the working directory, so run from a dedicated empty one. Combined with
|
|
# --no-session-persistence and --safe-mode, a launcher question leaves no
|
|
# trace in the real projects' history or memory.
|
|
work="$HOME/.cache/qs-ask"
|
|
${pkgs.coreutils}/bin/mkdir -p "$work" || { echo "qs-ask: cannot create $work" >&2; exit 0; }
|
|
cd "$work" || { echo "qs-ask: cannot enter $work" >&2; exit 0; }
|
|
|
|
# MAX_THINKING_TOKENS=0 is the single biggest win here, worth ~2.4x.
|
|
# Measured: Haiku defaults to extended thinking, and it spent 225 output
|
|
# tokens and 1.9s deliberating before emitting an 80-character answer
|
|
# about the height of a tower. With thinking off: 43 output tokens,
|
|
# time-to-first-text 2657ms -> 1122ms, wall clock 5.7s -> 2.4s, same
|
|
# answer. A one-line lookup has nothing to reason about.
|
|
#
|
|
# For the record, two things that are NOT the bottleneck, both measured:
|
|
# process startup (0.12s boot + 28ms to fire the request, so keeping a
|
|
# warm process saves nothing) and prompt size (184 input tokens).
|
|
ans=$(printf '%s' "$q" | ${pkgs.coreutils}/bin/env MAX_THINKING_TOKENS=0 \
|
|
${pkgs.claude-code}/bin/claude -p \
|
|
--safe-mode \
|
|
--no-session-persistence \
|
|
--tools "" \
|
|
--disable-slash-commands \
|
|
--strict-mcp-config \
|
|
--permission-mode dontAsk \
|
|
--output-format text \
|
|
--model claude-haiku-4-5 \
|
|
--system-prompt 'Answer the question in one or two short sentences, under 240 characters. Lead with the specific fact asked for, including units. No preamble, no caveats, no markdown, no follow-up offers. If you do not know, or the answer depends on live data you do not have, reply with exactly: UNKNOWN' \
|
|
2>/dev/null) \
|
|
|| { echo "qs-ask: claude exited non-zero (auth expired? run 'claude' once to log in)" >&2; exit 0; }
|
|
|
|
ans=$(printf '%s' "$ans" | ${pkgs.coreutils}/bin/tr -d '\r' | ${pkgs.gnused}/bin/sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//')
|
|
|
|
case "$ans" in
|
|
"") echo "qs-ask: empty answer" >&2; exit 0 ;;
|
|
UNKNOWN*) echo "qs-ask: model replied UNKNOWN" >&2; exit 0 ;;
|
|
esac
|
|
printf '%s' "$ans"
|
|
'')
|
|
];
|
|
|
|
# Basic networking
|
|
networking.useDHCP = lib.mkForce false;
|
|
|
|
# Allow fred to act as a remote Nix builder (trusted users can import
|
|
# unsigned store paths sent by the build client).
|
|
nix.settings.trusted-users = [ "root" "fred" ];
|
|
|
|
# Automatic daily system updates
|
|
system.autoUpgrade = {
|
|
enable = true;
|
|
flake = "git+https://forg.gregersen.it/rope/nixos";
|
|
dates = "05:15";
|
|
allowReboot = true;
|
|
};
|
|
|
|
# WAN exposure is controlled by nftables in services/router.nix +
|
|
# ports.toml (networking.firewall is disabled on this host).
|
|
services.openssh = {
|
|
enable = true;
|
|
settings = {
|
|
PermitRootLogin = "no";
|
|
PasswordAuthentication = false;
|
|
};
|
|
};
|
|
};
|
|
}
|