Power profiles: Balanced, Gaming, Eco

modules/nixos/hardware/power.nix (the daemon, the GPU half, the automatic holds), modules/nixos/gaming/steam.nix (GameMode's hold), modules/home/services/power-profile.nix (the user-side effects), modules/home/desktop/hypr/lua/rules.lua (the game-window trigger). Three operating profiles with ONE owner of the state: power-profiles-daemon (ppd). Everything else FOLLOWS it.

Automatic by default (19/10/2026). I never pick a profile: Balanced is the base, and two HOLDS move it, both released by ppd when their unit stops, when it crashes, or when I pick a profile by hand (a manual pick always wins).

• Away → Eco. hypridle's 5 min listener starts the user unit power-idle-eco (powerprofilesctl launch --profile power-saver --reason Idle --appid idle), and coming back stops it. It is PartOf=hypridle.service, so the Sunshine guard, which stops hypridle while streaming, ends it too. SignalSift stays on balanced under THIS Eco: the free PC is when its heavy jobs should run. A manual Eco still pauses it.
• A game → Gaming, by two independent triggers: GameMode (gamemoderun, or a Bottles bottle with GameMode on) starts gamemode-performance; and Hyprland itself starts game-performance while any game WINDOW is mapped (rules.lua: *.exe, steam_app_N, gamescope, the confine rule's set, minus battle.net.exe, a launcher left open for hours). So a Steam game without the launch option still gets Gaming.
• Never Eco over a game. ppd lets a power-saver hold win over a performance one, so both game units stop power-idle-eco before holding, and power-idle-eco has an ExecCondition that skips it while any performance hold exists (a controller game with no keyboard input goes idle in hypridle's eyes).

TESTED live 19/10/2026 by starting the units by hand: power-idle-eco gave power-saver, both GTs power_saving at 400 MHz and SignalSift still balanced; game-performance on top stopped it, raised performance, put the GPU back on base at 1200 MHz and SignalSift on gaming; a second power-idle-eco under the game was skipped by its ExecCondition; stopping the game hold returned to balanced with no hold left. The real triggers (5 min away, a game window) are the same units started by hypridle and Hyprland.

The GPU half runs as root (power-profile-gpu, the same gdbus monitor pattern as the user sync): /sys/class/drm/card0/device/tile0/gt{0,1}/freq0/power_profile and min_freq. MEASURED 18/10/2026, both GTs ship power_profile base with a floor of 1200 MHz while RPn is 400; the driver's floor is recorded once in /var/lib/power-profile-gpu/ and restored outside Eco. The board's power1_cap (210 W) is not touched: a cap only bites under load, and Eco is for light use. The monitors are not touched either (see the LG experiment below).

The watcher is NOT ordered after ppd. ppd's upstream unit carries After=multi-user.target, so after = power-profiles-daemon.service on a unit wanted by multi-user.target closes a cycle, and systemd broke it by deleting this job: the GPU half silently never ran at boot (seen 25/10/2026). gdbus monitor already reacts to ppd taking its bus name, so the ordering bought nothing.

Kept in EVERY profile: the network, the tunnel, Caddy and every published service, Syncthing, Dropbox, the backups and their timers, Sunshine, audio. No profile suspends the machine, stops a service, or touches the monitors' power state.

What the CPU offers (MEASURED 18/10/2026)Copy Anchor Link

The i5-11400 runs intel_pstate in active mode with HWP: governors performance and powersave, governor powersave, EPP balance_performance, the five EPP values available (default performance balance_performance balance_power power), turbo on, cpuidle on intel_idle with C1/C2/C3 (ACPI). No /sys/firmware/acpi/platform_profile, so ppd drives intel_pstate only. ASPM is at the kernel's default policy and stays there (see below).

With HWP the governor is not the lever: powersave already lets the hardware boost to the top, and EPP is the hint that decides how eagerly it does. That is exactly what ppd writes.

Why power-profiles-daemonCopy Anchor Link

• It is the upstream, maintained owner of this exact knob, with a D-Bus API and polkit rules that let the ACTIVE session switch profiles without root (switch-profile and hold-profile are allow_active yes in 0.30), so neither Quickshell nor a script needs privileges.
• Holds: a client can hold a profile while it runs. ppd drops the hold when the holder exits, when it crashes, or when the user switches profile by hand, and then returns to the profile the user had picked. That is the whole "leave the game and go back to Eco, not to Balanced" rule, without a state file of our own.
• Quickshell 0.3.2 binds it natively (Quickshell.Services.UPower, PowerProfiles).
• The NixOS module asserts against TLP and auto-cpufreq; neither is on, and no other module here touches EPP or the governor (powerManagement.cpuFreqGovernor is null).

The profile survives reboots: ppd keeps it in /var/lib/power-profiles-daemon/state.ini.

GameMode: a hold, not the governorCopy Anchor Link

MEASURED 18/10/2026: GameMode had never boosted anything. Every game start logged Requesting update of governor policy to performance, then pkexec ... cpugovctl set performance: Not authorized and ERROR: Failed to update cpu governor policy, and the split-lock helper failed the same way. The gamemode polkit rule only authorizes members of the gamemode group, and the user was not in it.

The fix is two parts:

1. The user joins gamemode, which is what the NixOS module creates the group for. The helpers stop failing; split-lock mitigation goes off during a game, GameMode's default.
2. The governor stays powersave (general.desiredgov). The performance governor would pin HWP and make EPP unwritable, a second owner fighting ppd. Instead GameMode's custom.start starts the user unit gamemode-performance, which runs powerprofilesctl launch --profile performance around a sleep infinity, and custom.end stops it.

Why this shape: GameMode runs start once for the first game and end once after the last, so concurrent games share one hold. The unit BindsTo=gamemoded.service, so a crashed GameMode ends the hold instead of leaving the CPU on performance. A manual pick during a game wins: ppd releases every hold when the user switches.

The user-side effects: power-profile-syncCopy Anchor Link

A user unit (default.target, so it works over SSH too) runs gdbus monitor on ppd's object and calls power-profile-apply on every ActiveProfile change and every time ppd (re)appears on the bus, which covers its own start. Event-driven, no polling. The apply, serialized by flock:

• SignalSift: signalsift profile set balanced|gaming|eco, the app's own interface. gaming keeps the API and light collection and stops transcription and inference; eco stops all background work. The app kills a running heavy job with its process group and keeps what it already transcribed (its docs/benchmarks/whisper-arc-b580.md), so nothing is lost. The profile lives only until the daemon restarts, so signalsift.service gets an ExecStartPost=-power-profile-apply --signalsift-only that re-applies it, retrying for 10 s while the daemon comes up.
• Ollama: in Gaming and Eco, every model in /api/ps gets keep_alive: 0, which frees its VRAM as soon as any request in flight ends. It never blocks the API: Duo's next run simply loads the model again. Its VRAM no longer competes with a game.

The outcome of each apply lands in $XDG_RUNTIME_DIR/power-profile.json (profile, SignalSift result, Ollama result, time), which the NixOS Center reads, and in the journal:

journalctl --user -u power-profile-sync -b
cat "$XDG_RUNTIME_DIR/power-profile.json"
power-profile-apply                          # re-apply by hand, idempotent

TESTED on 18/10/2026 before any switch, with powerprofilesctl and signalsift stubbed and the real Ollama API: power-saver gave eco (ok) and unloaded 0; a refusing SignalSift was retried 10 times and recorded as failed: error: connection refused; --signalsift-only kept the previous Ollama result.

TESTED live after the switch, the same evening: powerprofilesctl set power-saver moved EPP to power and the sync recorded eco (ok); starting gamemode-performance from the user manager was allowed by polkit, showed the hold (performance, GameMode) and EPP performance with SignalSift on gaming; stopping it returned to power-saver, NOT to balanced, with EPP power and SignalSift back on eco.

A switch does NOT reload a running gamemoded. It reads /etc/gamemode.ini at its own start, and NixOS restarts no user unit on activation. MEASURED on 18/10/2026: the instance from before the switch still had the old config, so gamemoderun raised the GOVERNOR to performance (EPP performance with ppd still reporting balanced and no hold), the exact second owner this setup removes. And it was allowed: polkit reads the group from /etc/group, so the new gamemode membership counts for the helpers before any re-login. After systemctl --user restart gamemoded (no game running), gamemoderun sleep 8 gave the hold (performance, GameMode), governor still powersave, SignalSift gaming, and on exit balanced again, with no polkit error. So after a switch that touches programs.gamemode: restart gamemoded with no game open, or log in again.

The LG at 60 Hz in Eco: an experiment, not wiredCopy Anchor Link

The LG offers 1920x1080 at 60 Hz (hyprctl monitors), and a lower refresh lowers the GPU's work for that output. It is NOT automated, because the repo's history with this panel is DisplayPort link loss, a panel DDC cannot wake, and black screens under Sunshine (see monitors and away), and because Hyprland 0.55 killed hyprctl keyword, so a runtime mode change needs hyprctl eval on hl.monitor, untested here. The experiment, by hand, with nothing streaming:

1. Read the board power for 60 s at 143.98 Hz (the command in the baseline).
2. Switch the LG to 60 Hz through the Lua API, keeping transform=1 and its position, and read the board power again for 60 s.
3. Switch back and check hyprctl monitors shows 143.98 Hz, the same position and transform.

Only a measured saving with a clean way back earns it a place in the Eco apply.

The power log: what is measured, and what is not (19/10/2026)Copy Anchor Link

modules/nixos/hardware/power-log.nix. A root timer every minute appends one line to /var/lib/power-log/YYYY-MM.csv. It reads CUMULATIVE energy counters, so each line is the exact average over its interval and no burst between samples is lost. A first sample, a new boot or a gap over 10 min writes nothing (there is no interval to report).

Checked on 19/10/2026 that this board has no other source: no PSU telemetry on USB, no UPS, no /sys/class/power_supply. The board, RAM, SSD, fans, the PSU's own loss and the monitors have no sensor. So the macro view needs ONE outlet meter with a local API (a smart plug with energy metering, read over the LAN); its reading becomes wall_w, and the unmeasured remainder is wall_w - measured_w. Until then the log is the micro view and a floor, and says so.

power-report        # the last 24 h: averages, kWh, by profile, by hour of the day
power-report 168    # the last week

First readings, 19/10/2026 around noon, the desktop IN USE (Zen, VS Code, Discord, Spotify and an agent's terminal): CPU package 38 to 48 W at only 4 to 10% busy, the cores alone 25 to 33 W, the GPU board 43 to 48 W. The cores sat at 4.2 GHz and spent 57% of their idle time in C1, against 18% each in the two deep states, under ~20 800 interrupts/s (14 000 of them local timers, ~1 200 per core): apps waking the CPU, not a BIOS limit, since CPU C-states and Package C State Limit both read Auto (255) in CpuSetup and the deep states are in use. An A/B on the same use: three minutes of Eco gave a CPU package of 33 to 35 W (cores 19 to 21 W) at 7.7 to 8.3% busy, against 43 W (cores 29 W) for Balanced at 7%, so about 8 to 9 W or ~20% off the CPU, and nothing off the GPU, as expected from a profile that only moves EPP. The load drifted between minutes, so it is an indication, not a benchmark. The profile column is read at the END of each interval.

A trap for anyone measuring by hand: Claude Code's spinner and streaming output repaint the terminal, which wakes the GPU at the monitor's refresh rate. MEASURED 19/10/2026: render idle 0% and 50 W on the board while an agent ran its commands, 50% and 43 W a few minutes later. Idle readings belong to this log, taken while nobody is at the keyboard.

RGBCopy Anchor Link

Not part of any profile. The LEDs are already driven by OpenRGB profiles and the away mode, and the gain is too small to justify a second writer of their state.

RejectedCopy Anchor Link

• TLP, auto-cpufreq, tuned: a laptop's battery policy, or a second owner of the same EPP.
• Fixing GameMode's governor change alone: it would work, then fight ppd's EPP writes.
• A profile daemon of our own: ppd already owns the state, the D-Bus API, the polkit gate and the hold semantics.
• ASPM overrides (pcie_aspm=force, pcie_aspm.policy=powersupersave): the B580 is known to idle high without ASPM L1 on some boards, but forcing it globally risks link instability on every PCIe device, and nothing here can read the link state without root (lspci -vv is not installed). Measure first, see the baseline.

Checking itCopy Anchor Link

powerprofilesctl                     # the active profile and every hold, with its reason
powerprofilesctl list-holds          # during a game: performance, reason GameMode
cat /sys/devices/system/cpu/cpufreq/policy0/energy_performance_preference
journalctl --user -u gamemoded -b | grep -iE 'govern|Not authorized'   # must be quiet
systemctl --user status gamemode-performance

RollbackCopy Anchor Link

• Everything: delete ./power.nix from modules/nixos/hardware/default.nix and the custom.*/desiredgov lines in steam.nix; EPP goes back to the kernel default, and power-profile.nix turns itself off (it is gated on ppd).
• Only the side effects: delete ./power-profile.nix from modules/home/services/default.nix. SignalSift then stays on its config's profile and Ollama on its own keep-alive.
• Only the automatic mode: delete the power-idle-eco listener in lockscreen.nix and the game-window block in rules.lua; GameMode's hold stays.
• Only the GPU half: delete systemd.services.power-profile-gpu in power.nix, then echo base | sudo tee /sys/class/drm/card0/device/tile0/gt*/freq0/power_profile and write the floors saved in /var/lib/power-profile-gpu/ back to min_freq (or reboot).
• Stuck on performance or power-saver (should not happen): powerprofilesctl list-holds names the unit; systemctl --user stop gamemode-performance game-performance power-idle-eco, or pick a profile by hand, which drops every hold.