Auto-select the gaming GPU + cross-vendor CPU-power tooling

MangoHud logs one GPU. On a hybrid box (discrete NVIDIA/AMD beside an
AMD/Intel iGPU) its default often latches onto the idle iGPU, so every
gpu_* column reads ~0 while the discrete card does the rendering. Detect
the actual gaming GPU — any vendor, single- or multi-GPU — and pin it via
gpu_list. Selection: single GPU -> that one; discrete NVIDIA preferred;
otherwise the card with the most VRAM. Override with MANGOHUD_LOGGER_GPU.
Verified against MangoHud's own enumeration (gpu_list=0 -> renderD128,
RTX 5060 Ti) and 7 synthetic layouts incl. NVIDIA as the second node.

Also make CPU telemetry work across AMD and Intel. cpu_power was logging
0 because the RAPL energy counters are root-only since CVE-2020-8694
(same intel-rapl powercap path on both vendors). Add:
  - doctor: read-only GPU/CPU telemetry readiness report
  - fix-cpu-power: install a udev rule making RAPL counters user-readable
  - enable-time warnings pointing at the fix
CPU temp already works via hwmon (k10temp/zenpower/coretemp).

Default no-arg behavior (toggle on/off) is unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -21,8 +21,9 @@ won't trample anything you wrote by hand):
 
 1. **`~/.config/MangoHud/MangoHud.conf`** — appends a marker-bracketed block
    with `output_folder`, `autostart_log=1`, `log_duration=0` (= until you
-   quit), `log_interval=100`, `toggle_logging=Shift_L+F2`, and a
+   quit), `log_interval=100`, `toggle_logging=Shift_L+F2`, a `no_display=0`
-   `no_display=0` override. (Also serves as the on/off state marker.)
+   override, and a `gpu_list=<index>` pin (see [GPU selection](#gpu-selection)).
+   (Also serves as the on/off state marker.)
 
 2. **`~/.config/environment.d/95-mangohud-logger.conf`** — contains:
    - `MANGOHUD=1` so the Vulkan loader auto-injects MangoHud into every game.
@@ -39,6 +40,40 @@ won't trample anything you wrote by hand):
 
 When you toggle it off, the script removes everything it wrote.
 
+## GPU selection
+
+MangoHud logs stats for **one** GPU. On a hybrid machine — a discrete
+NVIDIA/AMD card alongside an AMD/Intel iGPU (common on modern AMD desktops,
+laptops, and handhelds) — MangoHud's default frequently latches onto the *idle*
+iGPU, so every `gpu_*` column in the CSV reads ~0 (0 % load, idle clocks, idle
+power) even while the discrete card does all the rendering.
+
+On enable, the script auto-detects the GPU that's actually rendering games and
+pins MangoHud to it with `gpu_list=<index>`. It's **vendor-agnostic**:
+
+- **Single GPU** (any vendor) → that GPU.
+- **Discrete NVIDIA** present → the NVIDIA card.
+- **Two AMD/Intel GPUs** → the one with the most VRAM (the discrete card;
+  iGPUs report little or none).
+
+It works by mapping DRM render nodes (`renderD128`, `renderD129`, …) to
+MangoHud's `gpu_list` index order. The chosen GPU is printed on enable:
+
+```
+:: gpu logged:   gpu_list=0 → renderD128 (NVIDIA), 2 GPU(s) detected
+```
+
+If the guess is wrong, override it:
+
+```bash
+MANGOHUD_LOGGER_GPU=amd    ./mangohud-logger.sh   # or: nvidia | intel
+MANGOHUD_LOGGER_GPU=1      ./mangohud-logger.sh   # or an explicit gpu_list index
+```
+
+> **NVIDIA note:** MangoHud reads NVIDIA load/power/clocks via NVML. If NVML
+> (`libnvidia-ml`, shipped with `nvidia-utils`) is missing, the script warns you
+> — without it the NVIDIA `gpu_*` columns stay at 0.
+
 ## Session-flavor detection
 
 Game Mode is launched differently on different distros, and the fix differs
@@ -140,13 +175,30 @@ mangohud-logger
 ## Usage
 
 ```bash
-./mangohud-logger.sh
+./mangohud-logger.sh                # toggle logging on/off (default)
+./mangohud-logger.sh doctor         # report GPU/CPU telemetry readiness (no changes)
+./mangohud-logger.sh fix-cpu-power  # make RAPL CPU-power counters readable (sudo)
+./mangohud-logger.sh help           # usage
 ```
 
-No flags, no subcommands. First run enables; second run disables. After the
+Run with no argument to toggle: first run enables, second disables. After the
 first enable, **log out and back in** (Game Mode users: fully restart into Game
 Mode) so the env vars become effective. You only need to do that once.
 
+Before your first benchmark, run `doctor` — it tells you exactly which columns
+will hold real data and what (if anything) needs fixing:
+
+```
+:: GPUs (DRM render nodes):
+     renderD128  (NVIDIA)
+     renderD129  (AMD)
+✓ will log: gpu_list=0 → renderD128 (NVIDIA), 2 GPU(s) detected
+✓   NVML present → NVIDIA load/power/clocks will log
+:: CPU: AMD
+✓   temp source: hwmon 'k10temp' → cpu_temp will log
+!   power (RAPL): root-only → cpu_power = 0.  Fix:  mangohud-logger.sh fix-cpu-power
+```
+
 In-game keys (set by the script's MangoHud block):
 
 | Key | Action |
@@ -157,6 +209,29 @@ In-game keys (set by the script's MangoHud block):
 CSV files appear in `~/Downloads/mango-logs/`. Drop them into
 <https://flightlessmango.com/> for graphs, or open in a spreadsheet.
 
+## CPU power (`cpu_power` column reads 0?)
+
+MangoHud reads CPU package power from the RAPL powercap energy counters. Since
+CVE-2020-8694 those counters are **root-only** (mode `0400`), so user-mode
+MangoHud can't read them and `cpu_power` logs as `0`. This is the same on **AMD
+and Intel** — the powercap interface is named `intel-rapl` on both (it's the
+framework name; on AMD `intel_rapl_msr` feeds it from the Zen RAPL MSRs).
+
+`fix-cpu-power` installs a udev rule
+(`/etc/udev/rules.d/60-mangohud-logger-rapl.rules`) that makes `energy_uj`
+user-readable, and applies it immediately:
+
+```bash
+./mangohud-logger.sh fix-cpu-power
+```
+
+> **Security note:** this re-exposes the low-severity PLATYPUS power
+> side-channel. Fine for a personal gaming/benchmark box; skip it on shared or
+> multi-user machines. Remove the rule file to revert.
+
+CPU **temperature** needs no fix — it comes from hwmon (`k10temp`/`zenpower` on
+AMD, `coretemp` on Intel) and works out of the box.
+
 ## Compatibility
 
 - **SteamOS Game Mode** (`/usr/lib/steamos/gamescope-session`) — logs via

mangohud-logger.sh

@@ -10,11 +10,20 @@
 #      → appends a marker block with output_folder, autostart_log, etc.
 #      → keeps your existing keys; overrides `no_display` while logging
 #        (autostart_log rides on the render hook that no_display disables).
+#      → pins gpu_list to the GPU actually rendering games (see below), so the
+#        gpu_* columns log real data instead of an idle second GPU.
 #   2. ~/.config/environment.d/95-mangohud-logger.conf
 #      → contains MANGOHUD=1 so the Vulkan layer auto-loads into every
 #        game (gamescope sessions, Steam, native Vulkan apps).
 #        Picked up by systemd-user at next session login.
 #
+# GPU selection (vendor-agnostic):
+#   MangoHud logs stats for ONE GPU. On a hybrid box (an AMD/Intel iGPU beside a
+#   discrete NVIDIA/AMD card) its default often picks the idle iGPU, so every
+#   gpu_* column reads ~0. We auto-detect the render node that's the actual
+#   gaming GPU — any vendor — and pin `gpu_list` to its index. Override with
+#   MANGOHUD_LOGGER_GPU=<index> or =nvidia|amd|intel if the guess is wrong.
+#
 # Logs land in the user's Downloads folder under "mango-logs"
 # (honours XDG_DOWNLOAD_DIR from ~/.config/user-dirs.dirs, else ~/Downloads).
 
@@ -75,15 +84,269 @@ info() { printf '\033[36m::\033[0m %s\n' "$*"; }
 ok()   { printf '\033[32m✓\033[0m %s\n' "$*"; }
 warn() { printf '\033[33m!\033[0m %s\n' "$*"; }
 
-if ! command -v mangohud >/dev/null 2>&1; then
+# PCI vendor IDs (as reported by /sys/class/drm/*/device/vendor).
-  err "mangohud is not installed — install it first (e.g. 'sudo pacman -S mangohud lib32-mangohud')."
+readonly VENDOR_NVIDIA=0x10de
-  exit 1
+readonly VENDOR_AMD=0x1002
-fi
+readonly VENDOR_INTEL=0x8086
 
-mkdir -p "$CONFIG_DIR"
+# select_gpu — figure out which GPU MangoHud should log, vendor-agnostically.
-[[ -f "$CONFIG_FILE" ]] || : > "$CONFIG_FILE"
+#
+# MangoHud enumerates DRM render nodes (renderD128, renderD129, …) in ascending
+# order and `gpu_list=<n>` selects the n-th of them (verified: gpu_list=0 maps to
+# renderD128). We pick the node that's the real gaming GPU and return its index.
+#
+# Selection order:
+#   1. MANGOHUD_LOGGER_GPU override — a numeric index, or nvidia|amd|intel.
+#   2. Single GPU → index 0 (covers every single-card machine, any vendor).
+#   3. Multi-GPU → prefer a discrete NVIDIA; else the card with the most VRAM
+#      (discrete AMD/Intel report mem_info_vram_total; iGPUs report far less/none).
+#
+# Sets globals: GPU_INDEX (may be ""), GPU_VENDOR, GPU_DESC.
+select_gpu() {
+  GPU_INDEX=""; GPU_VENDOR=""; GPU_DESC=""
 
-if grep -qF "$MARKER_BEGIN" "$CONFIG_FILE"; then
+  local nodes=()
+  mapfile -t nodes < <(
+    for n in /sys/class/drm/renderD*; do
+      [[ -e "$n/device/vendor" ]] && basename "$n"
+    done | sort   # ascending == MangoHud's gpu_list index order
+  )
+
+  local count=${#nodes[@]}
+  if (( count == 0 )); then
+    GPU_DESC="no DRM render node found (software renderer?) — gpu_list left unset"
+    return
+  fi
+
+  # Index-aligned vendor + VRAM for each node.
+  local vendors=() vrams=() i
+  for i in "${!nodes[@]}"; do
+    vendors[i]=$(cat "/sys/class/drm/${nodes[i]}/device/vendor" 2>/dev/null || echo unknown)
+    vrams[i]=$(cat "/sys/class/drm/${nodes[i]}/device/mem_info_vram_total" 2>/dev/null || echo 0)
+  done
+
+  if [[ -n "${MANGOHUD_LOGGER_GPU:-}" ]]; then
+    local ov="${MANGOHUD_LOGGER_GPU,,}"
+    if [[ "$ov" =~ ^[0-9]+$ ]]; then
+      GPU_INDEX="$ov"
+    else
+      local want=""
+      case "$ov" in
+        nvidia) want=$VENDOR_NVIDIA ;;
+        amd)    want=$VENDOR_AMD ;;
+        intel)  want=$VENDOR_INTEL ;;
+      esac
+      for i in "${!nodes[@]}"; do
+        [[ "${vendors[i]}" == "$want" ]] && { GPU_INDEX="$i"; break; }
+      done
+      [[ -z "$GPU_INDEX" ]] && GPU_INDEX=0   # requested vendor absent → first GPU
+    fi
+  elif (( count == 1 )); then
+    GPU_INDEX=0
+  else
+    # Prefer discrete NVIDIA (NVIDIA never populates mem_info_vram_total via DRM
+    # sysfs, so the VRAM check below can't see it — hence this explicit pass).
+    for i in "${!nodes[@]}"; do
+      [[ "${vendors[i]}" == "$VENDOR_NVIDIA" ]] && { GPU_INDEX="$i"; break; }
+    done
+    if [[ -z "$GPU_INDEX" ]]; then
+      local best=0 bestvram=-1
+      for i in "${!nodes[@]}"; do
+        if (( ${vrams[i]:-0} > bestvram )); then bestvram=${vrams[i]:-0}; best=$i; fi
+      done
+      GPU_INDEX="$best"
+    fi
+  fi
+
+  # Describe the choice (guard against an out-of-range manual override).
+  local node="${nodes[$GPU_INDEX]:-?}" vsel="${vendors[$GPU_INDEX]:-unknown}" name
+  case "$vsel" in
+    "$VENDOR_NVIDIA") name=NVIDIA ;;
+    "$VENDOR_AMD")    name=AMD ;;
+    "$VENDOR_INTEL")  name=Intel ;;
+    *)                name="vendor $vsel" ;;
+  esac
+  GPU_VENDOR="$vsel"
+  GPU_DESC="gpu_list=$GPU_INDEX → $node ($name), $count GPU(s) detected"
+}
+
+# libnvidia-ml (NVML) is what lets MangoHud read NVIDIA load/power/clocks.
+have_nvml() {
+  ldconfig -p 2>/dev/null | grep -q 'libnvidia-ml' \
+    || [[ -e /usr/lib/libnvidia-ml.so.1 || -e /usr/lib64/libnvidia-ml.so.1 ]]
+}
+
+# ── CPU telemetry (vendor-neutral: works on AMD *and* Intel) ────────────────
+#
+# MangoHud reads CPU temperature from hwmon and CPU package power from the RAPL
+# powercap energy counters. Both paths are the same regardless of CPU vendor:
+#   • temp  → hwmon driver: k10temp/zenpower (AMD), coretemp (Intel)
+#   • power → /sys/class/powercap/intel-rapl:* — despite the "intel" name this is
+#     the powercap *framework*; on AMD it's fed by intel_rapl_msr from the Zen
+#     RAPL MSRs. Since CVE-2020-8694 the energy_uj files are root-only (0400),
+#     so user-mode MangoHud can't read them and cpu_power logs as 0.
+
+readonly RAPL_UDEV_FILE=/etc/udev/rules.d/60-mangohud-logger-rapl.rules
+
+# cpu_vendor → AMD | Intel | <raw string>
+cpu_vendor() {
+  local v; v=$(awk -F': ' '/^vendor_id/{print $2; exit}' /proc/cpuinfo 2>/dev/null)
+  case "$v" in
+    AuthenticAMD) echo AMD ;;
+    GenuineIntel) echo Intel ;;
+    *)            echo "${v:-unknown}" ;;
+  esac
+}
+
+# First RAPL package energy counter (enough to gate on), else empty.
+rapl_energy_file() { compgen -G '/sys/class/powercap/*rapl*/energy_uj' 2>/dev/null | head -1; }
+
+# rapl_state → readable | denied | absent
+rapl_state() {
+  local f; f=$(rapl_energy_file)
+  [[ -z "$f" ]] && { echo absent; return; }
+  if cat "$f" >/dev/null 2>&1; then echo readable; else echo denied; fi
+}
+
+# cpu_temp_source → hwmon 'name' providing CPU temp (any vendor), else empty.
+cpu_temp_source() {
+  local h n
+  for h in /sys/class/hwmon/hwmon*; do
+    n=$(cat "$h/name" 2>/dev/null) || continue
+    case "$n" in
+      k10temp|zenpower|zenpower3|coretemp|k8temp) echo "$n"; return ;;
+    esac
+  done
+  echo ""
+}
+
+# do_fix_cpu_power — install a udev rule making RAPL energy counters readable so
+# MangoHud can log cpu_power. Vendor-neutral (same RAPL path on AMD and Intel).
+do_fix_cpu_power() {
+  local vend; vend=$(cpu_vendor)
+  info "CPU power ($vend) comes from RAPL energy counters, root-only since CVE-2020-8694."
+  info "This installs a udev rule so they're user-readable and MangoHud can log cpu_power."
+  warn "Security note: re-exposes the low-severity PLATYPUS power side-channel."
+  warn "Fine for a personal gaming/benchmark box; skip on shared/multi-user machines."
+  echo
+
+  case "$(rapl_state)" in
+    absent)
+      err "no RAPL powercap interface found."
+      info "try:  sudo modprobe intel_rapl_msr   (then re-run)"
+      return 1 ;;
+    readable)
+      ok "RAPL energy counters are already user-readable — nothing to do."
+      return 0 ;;
+  esac
+
+  local chmod_bin; chmod_bin=$(command -v chmod || echo /usr/bin/chmod)
+  # KERNEL match covers the package (intel-rapl:0) and its subzones (intel-rapl:0:0).
+  # %S%p resolves to the node's /sys path; we chmod its energy_uj attribute.
+  local rule="SUBSYSTEM==\"powercap\", KERNEL==\"intel-rapl:*\", RUN+=\"$chmod_bin 0444 %S%p/energy_uj\""
+
+  info "writing $RAPL_UDEV_FILE (sudo)…"
+  if ! printf '%s\n' \
+        "# Installed by mangohud-logger: make RAPL energy_uj user-readable for CPU-power logging." \
+        "# Remove this file to revert. See: mangohud-logger help" \
+        "$rule" | sudo tee "$RAPL_UDEV_FILE" >/dev/null; then
+    err "failed to write $RAPL_UDEV_FILE"; return 1
+  fi
+
+  sudo udevadm control --reload-rules 2>/dev/null || true
+  sudo udevadm trigger --subsystem-match=powercap 2>/dev/null || true
+  # Apply to already-present nodes now (udev 'add' won't re-fire for existing ones).
+  sudo "$chmod_bin" -f 0444 /sys/class/powercap/*rapl*/energy_uj 2>/dev/null || true
+
+  if [[ "$(rapl_state)" == readable ]]; then
+    ok "done — cpu_power will now be logged (persists across reboots)."
+  else
+    warn "rule installed but counters still not readable; a reboot should apply it."
+  fi
+}
+
+# do_doctor — report cross-vendor telemetry readiness; makes no changes.
+do_doctor() {
+  info "mangohud-logger doctor — telemetry readiness (read-only)"
+  echo
+  if command -v mangohud >/dev/null 2>&1; then
+    ok "mangohud: $(mangohud --version 2>/dev/null | head -1)"
+  else
+    err "mangohud: NOT installed"
+  fi
+  info "session flavor: $GS_FLAVOR"
+  echo
+
+  # GPUs — list every render node, then the one that will be logged.
+  info "GPUs (DRM render nodes):"
+  local n vend name
+  for n in $(compgen -G '/sys/class/drm/renderD*' 2>/dev/null | sort); do
+    [[ -e "$n/device/vendor" ]] || continue
+    vend=$(cat "$n/device/vendor" 2>/dev/null)
+    case "$vend" in
+      "$VENDOR_NVIDIA") name=NVIDIA ;; "$VENDOR_AMD") name=AMD ;;
+      "$VENDOR_INTEL")  name=Intel  ;; *) name="vendor $vend" ;;
+    esac
+    printf '     %s  (%s)\n' "$(basename "$n")" "$name"
+  done
+  select_gpu
+  if [[ -n "$GPU_INDEX" ]]; then ok "will log: $GPU_DESC"
+  else warn "will log: $GPU_DESC"; fi
+  if [[ "$GPU_VENDOR" == "$VENDOR_NVIDIA" ]]; then
+    if have_nvml; then ok "  NVML present → NVIDIA load/power/clocks will log"
+    else err "  NVML missing → NVIDIA gpu_* columns will be 0 (install nvidia-utils)"; fi
+  fi
+  echo
+
+  # CPU — temp + power, vendor-neutral.
+  info "CPU: $(cpu_vendor)"
+  local ts; ts=$(cpu_temp_source)
+  if [[ -n "$ts" ]]; then ok "  temp source: hwmon '$ts' → cpu_temp will log"
+  else warn "  temp: no known sensor (k10temp/zenpower/coretemp) → cpu_temp may be 0"; fi
+  case "$(rapl_state)" in
+    readable) ok   "  power (RAPL): readable → cpu_power will log" ;;
+    denied)   warn "  power (RAPL): root-only → cpu_power = 0.  Fix:  $(basename "$0") fix-cpu-power" ;;
+    absent)   warn "  power (RAPL): no powercap interface → cpu_power = 0 (modprobe intel_rapl_msr)" ;;
+  esac
+  echo
+
+  if [[ -f "$CONFIG_FILE" ]] && grep -qF "$MARKER_BEGIN" "$CONFIG_FILE"; then
+    ok "logging is currently ENABLED (log folder: $LOG_DIR)"
+  else
+    info "logging is currently disabled — run '$(basename "$0")' to enable"
+  fi
+}
+
+usage() {
+  cat <<EOF
+mangohud-logger — toggle MangoHud CSV logging, with cross-vendor telemetry setup
+
+USAGE:
+  $(basename "$0")                  Toggle logging on/off (default action)
+  $(basename "$0") doctor           Report GPU/CPU telemetry readiness (no changes)
+  $(basename "$0") fix-cpu-power    Make RAPL CPU-power counters user-readable (sudo)
+  $(basename "$0") help             Show this help
+
+ENVIRONMENT:
+  MANGOHUD_LOGGER_GPU=<index|nvidia|amd|intel>
+                                    Force which GPU MangoHud logs (default: auto —
+                                    prefers the discrete gaming GPU, any vendor)
+
+Works across AMD/Intel CPUs and NVIDIA/AMD/Intel GPUs, single- or multi-GPU.
+EOF
+}
+
+# do_toggle — the original enable/disable behavior (default action).
+do_toggle() {
+  if ! command -v mangohud >/dev/null 2>&1; then
+    err "mangohud is not installed — install it first (e.g. 'sudo pacman -S mangohud lib32-mangohud')."
+    exit 1
+  fi
+
+  mkdir -p "$CONFIG_DIR"
+  [[ -f "$CONFIG_FILE" ]] || : > "$CONFIG_FILE"
+
+  if grep -qF "$MARKER_BEGIN" "$CONFIG_FILE"; then
   # Currently enabled → strip the block and remove the env file.
   tmp=$(mktemp)
   awk -v b="$MARKER_BEGIN" -v e="$MARKER_END" '
@@ -121,10 +384,13 @@ else
   # Currently disabled → append the config block + write the env file.
   # Check for pre-existing manual log keys *before* we write, so the warning is accurate.
   manual_log_keys=0
-  if grep -qE '^\s*(output_folder|autostart_log|log_duration|log_interval|toggle_logging)\s*=' "$CONFIG_FILE"; then
+  if grep -qE '^\s*(output_folder|autostart_log|log_duration|log_interval|toggle_logging|gpu_list)\s*=' "$CONFIG_FILE"; then
     manual_log_keys=1
   fi
 
+  # Work out which GPU to log (any vendor, single- or multi-GPU) before writing.
+  select_gpu
+
   mkdir -p "$LOG_DIR"
   {
     [[ -s "$CONFIG_FILE" ]] && printf '\n'
@@ -138,6 +404,9 @@ else
     # Override any earlier `no_display` — MangoHud's autostart_log rides on the
     # render hook, which `no_display` disables. Without this, logs never start.
     printf 'no_display=0\n'
+    # Pin the GPU MangoHud logs to the actual gaming GPU (see select_gpu). Omitted
+    # only when no render node was found, so MangoHud keeps its own default.
+    [[ -n "$GPU_INDEX" ]] && printf 'gpu_list=%s\n' "$GPU_INDEX"
     printf '%s\n' "$MARKER_END"
   } >> "$CONFIG_FILE"
 
@@ -147,6 +416,7 @@ else
   # but it never sets MANGOHUD_CONFIG, so ours (from environment.d) wins for every game.
   # NOTE: relies on output_folder having no commas/spaces (true for the default path).
   MH_INLINE="output_folder=$LOG_DIR,autostart_log=1,log_duration=0,log_interval=100,toggle_logging=Shift_L+F2,no_display=0"
+  [[ -n "$GPU_INDEX" ]] && MH_INLINE+=",gpu_list=$GPU_INDEX"
 
   mkdir -p "$ENV_DIR"
   cat > "$ENV_FILE" <<EOF
@@ -208,11 +478,32 @@ EOF
     warn "pre-existing log-related keys found in config — they may override ours."
   fi
 
+  # NVIDIA GPU stats come from NVML; without it every gpu_* column logs as 0.
+  if [[ "$GPU_VENDOR" == "$VENDOR_NVIDIA" ]] && ! have_nvml; then
+    warn "selected an NVIDIA GPU but libnvidia-ml (NVML) not found — GPU load/power"
+    warn "won't be logged. Install the NVIDIA userspace (e.g. 'nvidia-utils')."
+  fi
+
+  # CPU power/temp readiness (same checks on AMD and Intel).
+  rapl=$(rapl_state)
+  if [[ "$rapl" == denied ]]; then
+    warn "cpu_power won't log — RAPL counters are root-only. Enable it with:"
+    warn "    $(basename "$0") fix-cpu-power"
+  elif [[ "$rapl" == absent ]]; then
+    warn "cpu_power won't log — no RAPL powercap interface (try: sudo modprobe intel_rapl_msr)."
+  fi
+  [[ -z "$(cpu_temp_source)" ]] && warn "no known CPU temp sensor found — cpu_temp may read 0."
+
   ok "MangoHud logging enabled"
   info "log folder:   $LOG_DIR"
   info "toggle key:   Shift+F2 (during a game)"
   info "config file:  $CONFIG_FILE"
   info "env file:     $ENV_FILE  (MANGOHUD=1)"
+  info "gpu logged:   ${GPU_DESC:-MangoHud default}"
+  if [[ -n "$GPU_INDEX" ]]; then
+    info "              (override with MANGOHUD_LOGGER_GPU=<index|nvidia|amd|intel>)"
+  fi
+  info "cpu ($( cpu_vendor )):    temp=$( [[ -n "$(cpu_temp_source)" ]] && echo ok || echo none )  power(RAPL)=$rapl"
   info "session type: $GS_FLAVOR  (gamescope override: $gs_action)"
   info "note: while logging is on, the HUD becomes visible during games"
   info "      (autostart_log needs the render hook that no_display disables)"
@@ -231,3 +522,13 @@ EOF
     warn "log out and back in (or restart your gamescope session) so the env var takes effect."
   fi
 fi
+}
+
+# ── Command dispatch ────────────────────────────────────────────────────────
+case "${1:-toggle}" in
+  toggle|"")                     do_toggle ;;
+  doctor|check|--check|-c)       do_doctor ;;
+  fix-cpu-power|fix|--fix-cpu-power) do_fix_cpu_power ;;
+  -h|--help|help)                usage ;;
+  *) err "unknown command: $1"; echo; usage; exit 1 ;;
+esac