Pin compose example to v0.3.2

The status snapshot carries the UI address; the monitor renders it as a
dim 'UI: http://…' line below the GPU line.

README.md

@@ -34,8 +34,8 @@ Each consumer is enabled by setting its URL (`OLLAMA_URL`, `COMFY_URL`) and
 disabled by leaving it empty — at least one is required. With only Ollama
 the proxy is a pass-through (no image jobs can arrive); with only ComfyUI
 the Ollama unload/warm steps are skipped. A third, URL-less consumer —
-detection of foreign GPU holders such as games — is enabled by `GAME_PROCS`
-and/or `GPU_FOREIGN_VRAM_MB` (see below).
+detection of foreign GPU holders such as games — is enabled by `GAME_PROCS`,
+`GPU_FOREIGN_VRAM_MB` and/or `GPU_FOREIGN_UTIL_PCT` (see below).
 
 ## Configuration
 
@@ -49,6 +49,8 @@ override file values. Invalid values fail at startup.
 |---|---|---|
 | `LISTEN_OLLAMA` | `:11434` | Listener for Ollama-compatible clients |
 | `LISTEN_COMFY` | `:8188` | Listener for ComfyUI clients |
+| `LISTEN_UI` | _(empty = disabled)_ | Web UI listener (`host:port`) mirroring `--monitor` (live status, update/reload buttons). Keep it on `127.0.0.1` unless you set auth |
+| `UI_USER` / `UI_PASS` | _(empty = no auth)_ | HTTP basic auth for the web UI; must be set together |
 | `OLLAMA_URL` | _(empty = disabled)_ | Ollama upstream; set to enable the Ollama consumer |
 | `COMFY_URL` | _(empty = disabled)_ | ComfyUI upstream; set to enable the ComfyUI consumer |
 | `UNLOAD_TIMEOUT` | `60s` | Wait for Ollama to unload before an image job |
@@ -64,7 +66,8 @@ override file values. Invalid values fail at startup.
 | `COMFY_START_TIMEOUT` | `2m` | Max wait for the managed ComfyUI to come up |
 | `GAME_PROCS` | _(empty = disabled)_ | Process names (comma-separated); while any runs, the GPU counts as held: requests wait, Ollama unloads, managed ComfyUI stops |
 | `GPU_FOREIGN_VRAM_MB` | `0` (disabled) | Also treat the GPU as held when a non-ignored process uses more VRAM than this (needs nvidia-smi) |
-| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw` | Process names never counted as foreign GPU users |
+| `GPU_FOREIGN_UTIL_PCT` | `0` (disabled) | Also treat the GPU as held when a non-ignored process uses more than this percent of the GPU 3D engine (Windows PDH counters — what Task Manager shows; catches games without an exe list) |
+| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw,dwm` | Process names never counted as foreign GPU users |
 | `GAME_POLL_INTERVAL` | `15s` | How often game/VRAM detection runs (don't go below ~10s — nvidia-smi polls keep the GPU awake) |
 | `LOGLEVEL` | `warn` | `info` logs every request (colored arrows in text mode), `debug` adds lock transitions. `LOG_LEVEL` works as an alias |
 | `LOG_FORMAT` | `text` | `json` for structured JSON logs |
@@ -148,20 +151,25 @@ adds a `BindPaths=` to the systemd unit on Linux. Re-run it after changing
 Want to game on the same GPU without Ollama/ComfyUI squatting on the VRAM?
 gpu-turnstile can watch for foreign GPU holders and, while one is active,
 make LLM/image requests wait (or 503, per `LLM_BUSY_MODE`), unload Ollama's
-models and stop the managed ComfyUI so the game gets the memory. Two
+models and stop the managed ComfyUI so the game gets the memory. Three
 detection paths, each optional, polled every `GAME_POLL_INTERVAL` (15s):
 
 ```
-GAME_PROCS=cyberpunk2077.exe,bg3.exe     # the reliable way on Windows
+GPU_FOREIGN_UTIL_PCT=30                  # zero-config on Windows: any non-ignored
+                                         # process using >30% of the GPU 3D engine
+GAME_PROCS=cyberpunk2077.exe,bg3.exe     # explicit exe watch list
 GPU_FOREIGN_VRAM_MB=1024                 # catch-all via nvidia-smi
 ```
 
-`GAME_PROCS` matches running process names (case-insensitive, `.exe`
-optional). `GPU_FOREIGN_VRAM_MB` asks nvidia-smi which processes hold GPU
-memory and treats anything not in `GPU_IGNORE_PROCS` above the threshold as
-foreign — handy as a catch-all, but note that under Windows' WDDM driver
-graphics-only games may not show up in nvidia-smi's per-process list, so
-name your games in `GAME_PROCS` there; on Linux both paths work. When the
+`GPU_FOREIGN_UTIL_PCT` reads the same per-process GPU engine counters as
+Task Manager (PDH), which cover graphics work under Windows' WDDM driver —
+so games show up without naming them, and the offender is named in the log
+and monitor. Windows-only. `GAME_PROCS` matches running process names
+(case-insensitive, `.exe` optional). `GPU_FOREIGN_VRAM_MB` asks nvidia-smi
+which processes hold GPU memory and treats anything not in
+`GPU_IGNORE_PROCS` above the threshold as foreign — handy as a catch-all on
+Linux, but under WDDM graphics-only games may not show up in nvidia-smi's
+per-process list, so on Windows prefer `GPU_FOREIGN_UTIL_PCT`. When the
 game exits, requests resume automatically.
 
 ## Build and run

SPEC.md

@@ -55,9 +55,9 @@ consumer gets no listener, no startup probe, and no lock participation:
 - **Only `COMFY_URL`**: image jobs are tracked and ComfyUI's VRAM is freed
   afterwards, but the Ollama unload and warm-reload steps are skipped.
 - **Game detection** is a third, optional consumer without a URL: enabled by
-  `GAME_PROCS` and/or `GPU_FOREIGN_VRAM_MB` it watches for foreign processes
-  holding the GPU (see below) and plugs into the same lock the same way —
-  excluded when both knobs are unset.
+  `GAME_PROCS`, `GPU_FOREIGN_VRAM_MB` and/or `GPU_FOREIGN_UTIL_PCT` it
+  watches for foreign processes holding the GPU (see below) and plugs into
+  the same lock the same way — excluded when all knobs are unset.
 
 ### Lock semantics
 
@@ -171,20 +171,26 @@ files are flagged in the startup log.
 ## Game detection (foreign GPU holders)
 
 Games and other foreign GPU users sit outside the URL-based consumer model —
-nothing proxies through gpu-turnstile for them. Two independent detection
-paths, polled every `GAME_POLL_INTERVAL` (default 15 s); either one being
+nothing proxies through gpu-turnstile for them. Three independent detection
+paths, polled every `GAME_POLL_INTERVAL` (default 15 s); any one being
 configured enables the feature:
 
 - **Process watch list** (`GAME_PROCS`, comma-separated, case-insensitive,
   `.exe` optional): while any listed process runs, the GPU counts as held.
-  This is the reliable path on Windows.
+- **Foreign 3D-engine utilization** (`GPU_FOREIGN_UTIL_PCT`, Windows only):
+  per-process GPU engine counters from PDH — the same data Task Manager
+  shows — cover graphics work under WDDM, so any process not in
+  `GPU_IGNORE_PROCS` using more than the threshold percent of the 3D engine
+  counts as a foreign holder, with no exe list needed. Values are summed per
+  process across engines; a missing/failing PDH counter disables the path
+  (logged once).
 - **Foreign VRAM threshold** (`GPU_FOREIGN_VRAM_MB`): `nvidia-smi
   --query-compute-apps` lists per-process GPU memory; any process not in
-  `GPU_IGNORE_PROCS` (default: Ollama and python — ComfyUI runs under python)
-  holding more than the threshold counts as a foreign holder. Needs
-  nvidia-smi on the PATH (absent: logged once, path disabled) and works best
-  on Linux — under Windows' WDDM driver, graphics-only games may not appear
-  in the per-process list.
+  `GPU_IGNORE_PROCS` (default: Ollama and python — ComfyUI runs under python
+  — plus dwm, the desktop compositor) holding more than the threshold counts
+  as a foreign holder. Needs nvidia-smi on the PATH (absent: logged once,
+  path disabled) and works best on Linux — under Windows' WDDM driver,
+  graphics-only games may not appear in the per-process list.
 
 While a holder is detected, gpu-turnstile:
 
@@ -215,6 +221,8 @@ override file values. A missing file is fine; a malformed one is fatal.
 |---|---|---|
 | `LISTEN_OLLAMA` | `:11434` | listener for Ollama-compatible clients |
 | `LISTEN_COMFY` | `:8188` | listener for ComfyUI clients |
+| `LISTEN_UI` | _(empty = disabled)_ | web UI listener (`host:port`) mirroring `--monitor` (live status, update/reload buttons); keep it on `127.0.0.1` unless auth is set |
+| `UI_USER` / `UI_PASS` | _(empty = no auth)_ | HTTP basic auth for the web UI; must be set together |
 | `OLLAMA_URL` | _(empty = disabled)_ | Ollama upstream; set to enable the Ollama consumer |
 | `COMFY_URL` | _(empty = disabled)_ | ComfyUI upstream; set to enable the ComfyUI consumer |
 | `UNLOAD_TIMEOUT` | `60s` | wait for Ollama to unload |
@@ -230,7 +238,8 @@ override file values. A missing file is fine; a malformed one is fatal.
 | `COMFY_START_TIMEOUT` | `2m` | how long a request waits for the managed ComfyUI to come up |
 | `GAME_PROCS` | _(empty = disabled)_ | comma-separated process names (case-insensitive, `.exe` optional); while any runs, the GPU counts as held by it: requests wait, Ollama unloads, the managed ComfyUI stops |
 | `GPU_FOREIGN_VRAM_MB` | `0` (disabled) | also treat the GPU as held when a process not in `GPU_IGNORE_PROCS` uses more VRAM than this; needs nvidia-smi |
-| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw` | process names never counted as foreign GPU users |
+| `GPU_FOREIGN_UTIL_PCT` | `0` (disabled) | also treat the GPU as held when a process not in `GPU_IGNORE_PROCS` uses more than this percent of the GPU 3D engine (Windows PDH counters, as shown by Task Manager; catches games without an exe list) |
+| `GPU_IGNORE_PROCS` | `ollama,ollama app,ollama_llama_server,python,pythonw,dwm` | process names never counted as foreign GPU users |
 | `GAME_POLL_INTERVAL` | `15s` | how often game/VRAM detection runs (nvidia-smi polls keep the GPU awake; don't go below ~10s) |
 | `LOGLEVEL` | `warn` | `info` logs every request (colored arrows in text mode), `debug` adds lock transitions. `LOG_LEVEL` is accepted as an alias |
 | `LOG_FORMAT` | `text` | `json` for structured JSON logs |

cmd/gpu-turnstile/main.go

@@ -5,6 +5,7 @@ package main
 import (
 	"bufio"
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -15,6 +16,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"reflect"
 	"runtime"
 	"strings"
 	"sync"
@@ -38,7 +40,8 @@ import (
 var version = "dev"
 
 // exitCodeUpdate tells the service recovery configuration to restart the
-// process: a signed update has been staged and the GPU lock is idle.
+// process: a signed update has been staged or a changed config reload was
+// requested, and the GPU lock is idle.
 const exitCodeUpdate = 3
 
 // exitCodeStaged is returned by an elevated --force-update child when it
@@ -58,7 +61,7 @@ func stdoutIsTerminal() bool {
 // --install-service / --remove-service switches, --no-copy, -h/--help,
 // -v/--version, --force-update and the hidden --elevated-child marker from
 // args.
-func parseFlags(args []string) (configPath string, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, elevatedChild bool, rest []string) {
+func parseFlags(args []string) (configPath string, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, reloadEnv, monitor, elevatedChild bool, rest []string) {
 	rest = args[:0]
 	for i := 0; i < len(args); i++ {
 		switch {
@@ -67,9 +70,9 @@ func parseFlags(args []string) (configPath string, install, remove, noCopy, help
 			i++
 		case strings.HasPrefix(args[i], "-config="):
 			configPath = strings.TrimPrefix(args[i], "-config=")
-		case args[i] == "--install-service" || args[i] == "-install-service":
+		case args[i] == "--install-service" || args[i] == "-install-service" || args[i] == "-i":
 			install = true
-		case args[i] == "--remove-service" || args[i] == "-remove-service":
+		case args[i] == "--remove-service" || args[i] == "-remove-service" || args[i] == "-r":
 			remove = true
 		case args[i] == "--no-copy" || args[i] == "-no-copy":
 			noCopy = true
@@ -81,13 +84,17 @@ func parseFlags(args []string) (configPath string, install, remove, noCopy, help
 			forceUpdate = true
 		case args[i] == "--update-now" || args[i] == "-update-now":
 			updateNow = true
+		case args[i] == "--reload-env" || args[i] == "-reload-env":
+			reloadEnv = true
+		case args[i] == "--monitor" || args[i] == "-monitor" || args[i] == "-m":
+			monitor = true
 		case args[i] == "--elevated-child":
 			elevatedChild = true
 		default:
 			rest = append(rest, args[i])
 		}
 	}
-	return configPath, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, elevatedChild, rest
+	return configPath, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, reloadEnv, monitor, elevatedChild, rest
 }
 
 // versionLine is printed at the top of every help and error screen.
@@ -97,14 +104,20 @@ const usageText = `GPU arbitration proxy for Ollama + ComfyUI
 
 Usage:
   gpu-turnstile -config <path>                                run the proxy
-  gpu-turnstile --install-service [--no-copy] [-config path]  install + start as a service
-  gpu-turnstile --remove-service                              stop + uninstall the service
+  gpu-turnstile -i | --install-service [--no-copy] [-config path]
+                                                              install + start as a service
+  gpu-turnstile -r | --remove-service                         stop + uninstall the service
   gpu-turnstile -v | --version                                print just the version
   gpu-turnstile --force-update                                check for a signed update now,
                                                               apply it and restart the service
                                                               (no admin needed when the service runs)
   gpu-turnstile --update-now                                  like --force-update, but only
                                                               through the running service
+  gpu-turnstile --reload-env                                  make the service re-read and
+                                                              validate its config file, then
+                                                              restart onto it if it changed
+  gpu-turnstile -m | --monitor                                live status view (downstreams,
+                                                              GPU lock, queue); Ctrl+C quits
   gpu-turnstile -h | --help                                   this help
 
 Options:
@@ -135,12 +148,12 @@ func fatalUsage(format string, args ...any) {
 }
 
 func main() {
-	configPath, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, elevatedChild, args := parseFlags(os.Args[1:])
+	configPath, install, remove, noCopy, help, showVersion, forceUpdate, updateNow, reloadEnv, monitor, elevatedChild, args := parseFlags(os.Args[1:])
 	if showVersion {
 		fmt.Println(version)
 		return
 	}
-	bare := configPath == "" && !install && !remove && !forceUpdate && !updateNow && !elevatedChild && len(args) == 0
+	bare := configPath == "" && !install && !remove && !forceUpdate && !updateNow && !reloadEnv && !monitor && !elevatedChild && len(args) == 0
 	if help || (bare && stdoutIsTerminal()) {
 		// Bare invocation in a terminal (e.g. double-clicked on Windows)
 		// shows the help instead of starting a proxy window with no visible
@@ -165,6 +178,10 @@ func main() {
 		fatalUsage("error: --force-update cannot be combined with --install-service/--remove-service")
 	case updateNow && (install || remove || forceUpdate):
 		fatalUsage("error: --update-now cannot be combined with other commands")
+	case reloadEnv && (install || remove || forceUpdate || updateNow):
+		fatalUsage("error: --reload-env cannot be combined with other commands")
+	case monitor && (install || remove || forceUpdate || updateNow || reloadEnv):
+		fatalUsage("error: --monitor cannot be combined with other commands")
 	case install:
 		os.Exit(serviceCommand(configPath, true, noCopy, elevatedChild))
 	case remove:
@@ -173,6 +190,10 @@ func main() {
 		os.Exit(forceUpdateCommand(configPath, elevatedChild))
 	case updateNow:
 		os.Exit(updateNowCommand())
+	case reloadEnv:
+		os.Exit(reloadEnvCommand())
+	case monitor:
+		os.Exit(monitorCommand())
 	}
 	if len(args) > 0 {
 		fatalUsage("error: unknown arguments: %s", strings.Join(args, " "))
@@ -192,7 +213,7 @@ func main() {
 	syncEnvFile(resolveConfigPath(configPath), cfg.LogFile, log)
 
 	if service.IsService() {
-		if err := service.Run(func(ctx context.Context) error { return run(ctx, cfg, log, logOut, true) }); err != nil {
+		if err := service.Run(func(ctx context.Context) error { return run(ctx, cfg, log, logOut, true, configPath) }); err != nil {
 			log.Error("service failed", "err", err)
 			os.Exit(1)
 		}
@@ -200,7 +221,7 @@ func main() {
 	}
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
 	defer stop()
-	if err := run(ctx, cfg, log, logOut, false); err != nil {
+	if err := run(ctx, cfg, log, logOut, false, configPath); err != nil {
 		log.Error("listener failed", "err", err)
 		os.Exit(1)
 	}
@@ -352,6 +373,10 @@ func forceUpdateCommand(configPath string, elevatedChild bool) int {
 		fmt.Fprintf(os.Stderr, "gpu-turnstile: cannot locate executable: %v\n", err)
 		return 1
 	}
+	// One-shot CLI: the updater logs to stderr, never to the service's
+	// LOG_FILE — that file is ACL'd to the service account, and a CLI run
+	// has nothing worth persisting there.
+	cfg.LogFile = ""
 	log, _, logCloser := newLogger(cfg)
 	defer logCloser.Close()
 	if cfg.AppVersion == "dev" {
@@ -437,6 +462,18 @@ func updateNowCommand() int {
 	return printControlReply(reply)
 }
 
+// reloadEnvCommand asks the running service, over the control channel, to
+// re-read and validate its config file. The client sends no path — the
+// service only ever re-reads its own configured file.
+func reloadEnvCommand() int {
+	reply, err := control.Ask(control.CmdReloadEnv)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "%s\n\ngpu-turnstile: no running service to ask\n", versionLine())
+		return 1
+	}
+	return printControlReply(reply)
+}
+
 // reportElevatedUpdate prints the parent's summary of an elevated
 // --force-update child: exitCodeStaged means the child staged a new binary,
 // 0 means it found nothing to do. to is the tag the parent's own check
@@ -517,7 +554,7 @@ func managedComfyCommand(cfg config.Config) string {
 	return ""
 }
 
-func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Writer, isService bool) error {
+func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Writer, isService bool, configPath string) error {
 	// ComfyUI can run as a managed child — COMFY_CMD verbatim, or the
 	// standard venv layout derived from COMFY_DIR alone: started on demand
 	// by the proxy, stopped after COMFY_IDLE_TIMEOUT idle (and on shutdown)
@@ -530,6 +567,8 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		"version", version,
 		"listen_ollama", cfg.ListenOllama,
 		"listen_comfy", cfg.ListenComfy,
+		"listen_ui", orDisabled(cfg.ListenUI),
+		"ui_auth", cfg.UIUser != "",
 		"ollama_url", orDisabled(cfg.OllamaURL),
 		"comfy_url", orDisabled(cfg.ComfyURL),
 		"unload_timeout", cfg.UnloadTimeout,
@@ -555,6 +594,7 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		"comfy_start_timeout", cfg.ComfyStartTimeout,
 		"game_procs", cfg.GameProcs,
 		"gpu_foreign_vram_mb", cfg.GPUForeignVRAMMB,
+		"gpu_foreign_util_pct", cfg.GPUForeignUtilPct,
 		"gpu_ignore_procs", cfg.GPUIgnoreProcs,
 		"game_poll_interval", cfg.GamePollInterval,
 		"auto_update", cfg.AutoUpdate,
@@ -567,6 +607,8 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 	)
 
 	lk := lock.New(log)
+	health := newHealthTracker()
+	started := time.Now()
 	// Each consumer is enabled by setting its URL; a disabled consumer gets
 	// no client, no listener and no probe.
 	var ollamaClient *ollama.Client
@@ -676,20 +718,25 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 	}
 	probeCtx, probeCancel := context.WithTimeout(ctx, cfg.ProbeTimeout)
 	for name, probe := range probes {
-		if err := probe(probeCtx); err != nil && !errors.Is(err, errManagedDown) {
+		err := probe(probeCtx)
+		health.set(name, err == nil)
+		if err != nil && !errors.Is(err, errManagedDown) {
 			log.Warn(name+" probe failed", "err", err)
 		}
 	}
 	probeCancel()
 	if cfg.HealthInterval > 0 {
-		go healthLoop(ctx, cfg.HealthInterval, cfg.ProbeTimeout, log, probes)
+		go healthLoop(ctx, cfg.HealthInterval, cfg.ProbeTimeout, log, probes, health)
 	}
 
-	// Foreign GPU holders (games, other ML jobs) — enabled by GAME_PROCS
-	// and/or GPU_FOREIGN_VRAM_MB — hold the lock externally while they run.
-	if len(cfg.GameProcs) > 0 || cfg.GPUForeignVRAMMB > 0 {
-		det := game.New(cfg.GameProcs, cfg.GPUForeignVRAMMB, cfg.GPUIgnoreProcs, log)
-		go gameLoop(ctx, cfg, log, det, lk, ollamaClient, comfySup)
+	// Foreign GPU holders (games, other ML jobs) — enabled by GAME_PROCS,
+	// GPU_FOREIGN_VRAM_MB and/or GPU_FOREIGN_UTIL_PCT — hold the lock
+	// externally while they run. gw collects the VRAM reading and the last
+	// check result for the status channel.
+	gw := &gpuWatch{enabled: len(cfg.GameProcs) > 0 || cfg.GPUForeignVRAMMB > 0 || cfg.GPUForeignUtilPct > 0}
+	if gw.enabled {
+		det := game.New(cfg.GameProcs, cfg.GPUForeignVRAMMB, cfg.GPUForeignUtilPct, cfg.GPUIgnoreProcs, log)
+		go gameLoop(ctx, cfg, log, det, lk, ollamaClient, comfySup, gw)
 	}
 
 	// Bind the listeners up front so a port conflict fails fast and the
@@ -727,38 +774,65 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 	service.NotifyReady()
 	service.StartWatchdog(ctx)
 
+	var u *update.Updater
+	var exePath string
+	// restartWhenIdle exits with exitCodeUpdate once the GPU lock is idle;
+	// the service recovery configuration brings the process back. Shared by
+	// staged updates and config reloads; the once guard makes repeated
+	// triggers idempotent.
+	var restartOnce sync.Once
+	restartWhenIdle := func(reason string) {
+		log.Warn("restarting once the GPU is idle", "reason", reason)
+		restartOnce.Do(func() {
+			go func() {
+				if waitForIdle(ctx, lk, 24*time.Hour) {
+					log.Warn("restarting now", "reason", reason)
+					os.Exit(exitCodeUpdate)
+				}
+			}()
+		})
+	}
+	applyStaged := func(to string) {}
 	if cfg.AutoUpdate {
-		exePath, err := os.Executable()
-		if err != nil {
+		if p, err := os.Executable(); err != nil {
 			log.Warn("auto-update disabled: cannot locate executable", "err", err)
 		} else {
-			u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
-			// applyStaged is shared by the hourly loop and the control
-			// channel; the once guard keeps a second trigger from
-			// double-waiting on the GPU lock.
-			var once sync.Once
-			applyStaged := func(to string) {
+			exePath = p
+			u = &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
+			applyStaged = func(to string) {
 				if !isService {
 					log.Warn("auto-update: new binary staged; restart gpu-turnstile to apply", "version", to)
 					return
 				}
-				log.Warn("auto-update: staged; restarting once the GPU is idle", "version", to)
-				once.Do(func() {
-					go func() {
-						if waitForIdle(ctx, lk, 24*time.Hour) {
-							log.Warn("auto-update: restarting to apply update")
-							os.Exit(exitCodeUpdate)
-						}
-					}()
-				})
+				restartWhenIdle("update to " + to)
 			}
 			go updateLoop(ctx, cfg.UpdateInterval, log, u, exePath, applyStaged)
-			if isService {
-				serveControl(ctx, log, u, exePath, applyStaged)
-			}
 		}
 	}
 
+	// The command handler backs both the local control channel (served
+	// whenever running as a service) and the optional web UI (LISTEN_UI),
+	// independent of AUTO_UPDATE.
+	handler := controlHandler(ctx, u, exePath, applyStaged,
+		statusProvider(cfg, lk, comfySup, health, started, gw, ollamaClient),
+		reloadHandler(cfg, configPath, restartWhenIdle))
+	if isService {
+		serveControl(ctx, log, handler)
+	}
+	if cfg.ListenUI != "" {
+		ln, err := net.Listen("tcp", cfg.ListenUI)
+		if err != nil {
+			return fmt.Errorf("listen ui on %s: %w", cfg.ListenUI, err)
+		}
+		if host, _, err := net.SplitHostPort(cfg.ListenUI); err == nil && !isLoopbackHost(host) && cfg.UIUser == "" {
+			log.Warn("web UI is reachable from other machines WITHOUT auth; set UI_USER/UI_PASS or bind to 127.0.0.1", "addr", cfg.ListenUI)
+		}
+		uiSrv := &http.Server{Addr: cfg.ListenUI, Handler: uiHandler(handler, cfg.UIUser, cfg.UIPass)}
+		servers = append(servers, uiSrv)
+		log.Warn("listening", "consumer", "ui", "addr", cfg.ListenUI)
+		go func() { errCh <- uiSrv.Serve(ln) }()
+	}
+
 	select {
 	case err := <-errCh:
 		if err != nil && !errors.Is(err, http.ErrServerClosed) {
@@ -780,12 +854,52 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 // (idle); health checks skip it instead of logging an outage.
 var errManagedDown = errors.New("managed upstream intentionally stopped")
 
+// gpuWatch records the latest VRAM reading and game-detector finding from
+// the game detector's poll loop, for the status channel. Enabled is false
+// when game detection is not configured (no polling happens then); Known is
+// false until the first successful nvidia-smi reading.
+type gpuWatch struct {
+	mu      sync.Mutex
+	enabled bool
+	usedMB  int
+	total   int
+	tempC   int
+	fanPct  int
+	known   bool
+	foreign string // last Check result: external holders, "" when none
+	at      time.Time
+}
+
+func (g *gpuWatch) setVRAM(st game.GPUStats) {
+	g.mu.Lock()
+	g.usedMB, g.total = st.UsedMB, st.TotalMB
+	g.tempC, g.fanPct, g.known = st.TempC, st.FanPct, true
+	g.mu.Unlock()
+}
+
+func (g *gpuWatch) setCheck(foreign string) {
+	g.mu.Lock()
+	g.foreign, g.at = foreign, time.Now()
+	g.mu.Unlock()
+}
+
+func (g *gpuWatch) get() (st game.GPUStats, known bool, foreign string, ageS int64) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	ageS = -1
+	if !g.at.IsZero() {
+		ageS = int64(time.Since(g.at).Seconds())
+	}
+	return game.GPUStats{UsedMB: g.usedMB, TotalMB: g.total, TempC: g.tempC, FanPct: g.fanPct},
+		g.known, g.foreign, ageS
+}
+
 // gameLoop polls for foreign GPU holders (a game, another ML job). While one
 // is detected it holds the lock externally so new LLM and image requests
 // wait (or are rejected per LLM_BUSY_MODE), and — once in-flight work has
 // drained — frees VRAM for it: the managed ComfyUI is stopped and Ollama's
 // resident models are unloaded.
-func gameLoop(ctx context.Context, cfg config.Config, log *slog.Logger, det *game.Detector, lk *lock.Lock, ollamaClient *ollama.Client, comfySup *supervise.Process) {
+func gameLoop(ctx context.Context, cfg config.Config, log *slog.Logger, det *game.Detector, lk *lock.Lock, ollamaClient *ollama.Client, comfySup *supervise.Process, gw *gpuWatch) {
 	ticker := time.NewTicker(cfg.GamePollInterval)
 	defer ticker.Stop()
 	held, freed := false, false
@@ -799,6 +913,10 @@ func gameLoop(ctx context.Context, cfg config.Config, log *slog.Logger, det *gam
 		if err != nil && ctx.Err() == nil {
 			log.Warn("game detection failed", "err", err)
 		}
+		gw.setCheck(summarizeHolders(holders))
+		if st, verr := game.QueryGPUStats(ctx); verr == nil {
+			gw.setVRAM(st)
+		}
 		switch {
 		case len(holders) > 0 && !held:
 			held = true
@@ -858,7 +976,8 @@ func freeVRAM(ctx context.Context, unloadTimeout time.Duration, log *slog.Logger
 // transitions — "is DOWN" when a previously healthy upstream stops
 // answering, "recovered" when it comes back. The first round only
 // establishes the baseline; the startup probe already reported that state.
-func healthLoop(ctx context.Context, interval, probeTimeout time.Duration, log *slog.Logger, probes map[string]func(context.Context) error) {
+// Every result goes into the tracker for the status channel.
+func healthLoop(ctx context.Context, interval, probeTimeout time.Duration, log *slog.Logger, probes map[string]func(context.Context) error, tracker *healthTracker) {
 	ticker := time.NewTicker(interval)
 	defer ticker.Stop()
 	up := map[string]bool{}
@@ -875,9 +994,11 @@ func healthLoop(ctx context.Context, interval, probeTimeout time.Duration, log *
 			cancel()
 			if errors.Is(err, errManagedDown) {
 				managed[name] = true // intentionally stopped; not an outage
+				tracker.set(name, false)
 				continue
 			}
 			now := err == nil
+			tracker.set(name, now)
 			if managed[name] {
 				// First real probe after an idle stop only re-baselines —
 				// an on-demand start is not a "recovery".
@@ -919,18 +1040,39 @@ func updateLoop(ctx context.Context, interval time.Duration, log *slog.Logger, u
 	}
 }
 
-// serveControl opens the local control channel (named pipe on Windows,
-// unix socket on Linux) so unprivileged local users can trigger an update
-// check via --force-update without admin rights. The check payload is
-// signature-verified regardless of who asks; triggers are rate-limited to
-// one per minute so the channel cannot be used to spam restarts.
-func serveControl(ctx context.Context, log *slog.Logger, u *update.Updater, exePath string, applyStaged func(to string)) {
+// controlHandler builds the command handler shared by the control channel
+// (named pipe / unix socket) and the web UI's action endpoints: status for
+// --monitor / the UI, an update check (--force-update/--update-now) and a
+// config reload (--reload-env), all safe for unprivileged local users. The
+// update payload is signature-verified regardless of who asks; update
+// triggers are rate-limited to one per minute and reloads to one per two
+// seconds so neither can be used to spam restarts or disk churn. u is nil
+// when AUTO_UPDATE=false.
+func controlHandler(ctx context.Context, u *update.Updater, exePath string, applyStaged func(to string), status func() string, reload func() string) control.Handler {
 	var mu sync.Mutex
-	var lastTrigger time.Time
-	h := func(cmd string) string {
-		if cmd != control.CmdUpdateNow {
+	var lastTrigger, lastReload time.Time
+	return func(cmd string) string {
+		switch cmd {
+		case control.CmdStatus:
+			return "OK " + status()
+		case control.CmdReloadEnv:
+			// reload re-reads the config file from disk; a short limiter
+			// keeps a local flood from turning into disk churn.
+			mu.Lock()
+			if wait := 2*time.Second - time.Since(lastReload); wait > 0 {
+				mu.Unlock()
+				return fmt.Sprintf("ERR rate limited: retry in %ds", int(wait.Seconds())+1)
+			}
+			lastReload = time.Now()
+			mu.Unlock()
+			return reload()
+		case control.CmdUpdateNow:
+		default:
 			return "ERR unknown command: " + cmd
 		}
+		if u == nil {
+			return "ERR auto-update is disabled on this instance"
+		}
 		mu.Lock()
 		if wait := time.Minute - time.Since(lastTrigger); wait > 0 {
 			mu.Unlock()
@@ -950,11 +1092,202 @@ func serveControl(ctx context.Context, log *slog.Logger, u *update.Updater, exeP
 		applyStaged(to)
 		return "OK updated from " + version + " to " + to + "; the service restarts once the GPU is idle"
 	}
+}
+
+// serveControl opens the local control channel (named pipe on Windows,
+// unix socket on Linux) for unprivileged local users (--monitor,
+// --force-update/--update-now, --reload-env).
+func serveControl(ctx context.Context, log *slog.Logger, h control.Handler) {
 	if err := control.Serve(ctx, h, log); err != nil {
 		log.Warn("control channel disabled", "err", err)
 	}
 }
 
+// healthTracker records the latest probe result per upstream for the
+// status channel.
+type healthTracker struct {
+	mu sync.Mutex
+	up map[string]bool
+}
+
+func newHealthTracker() *healthTracker {
+	return &healthTracker{up: map[string]bool{}}
+}
+
+func (h *healthTracker) set(name string, up bool) {
+	h.mu.Lock()
+	h.up[name] = up
+	h.mu.Unlock()
+}
+
+func (h *healthTracker) get(name string) bool {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return h.up[name]
+}
+
+// statusDownstream/statusLock/statusSnapshot are the JSON the control
+// channel serves on CmdStatus; the monitor mode renders them.
+type statusDownstream struct {
+	Name    string `json:"name"`
+	URL     string `json:"url"`
+	Up      bool   `json:"up"`
+	Managed string `json:"managed,omitempty"`
+	// Models lists Ollama's loaded models with their VRAM footprint. Null
+	// when unknown (query failed / not applicable); [] means none loaded —
+	// deliberately no omitempty so the two stay distinguishable.
+	Models []statusModel `json:"models"`
+}
+
+// statusModel is one loaded Ollama model.
+type statusModel struct {
+	Name   string `json:"name"`
+	VRAMMB int64  `json:"vram_mb"` // 0 = resident in RAM, not VRAM
+}
+
+type statusLock struct {
+	State       string `json:"state"`
+	Detail      string `json:"detail,omitempty"`
+	LLMInflight int    `json:"llm_inflight"`
+	LLMWaiting  int    `json:"llm_waiting"`
+	ImageQueue  int    `json:"image_queue"`
+	External    string `json:"external,omitempty"`
+	SinceS      int64  `json:"since_s"`
+}
+
+type statusSnapshot struct {
+	Version     string             `json:"version"`
+	UptimeS     int64              `json:"uptime_s"`
+	Downstreams []statusDownstream `json:"downstreams"`
+	Lock        statusLock         `json:"lock"`
+	// GPU carries the latest VRAM reading and detector finding; Enabled is
+	// false when game detection (and with it nvidia-smi polling) is not
+	// configured.
+	GPU statusGPU `json:"gpu"`
+	// UI is the web UI's URL when LISTEN_UI is active, empty otherwise.
+	UI string `json:"ui,omitempty"`
+	// MonitorNote is set client-side (never over the wire) when the
+	// monitor's own binary differs from the service's version.
+	MonitorNote string `json:"-"`
+}
+
+type statusGPU struct {
+	// Enabled reports whether game detection is configured (and with it
+	// VRAM polling); when false the other fields carry no information.
+	Enabled bool `json:"enabled"`
+	UsedMB  int  `json:"used_mb"`
+	TotalMB int  `json:"total_mb"`
+	// TempC/FanPct are -1 when unknown (never sampled or nvidia-smi
+	// reported N/A).
+	TempC  int  `json:"temp_c"`
+	FanPct int  `json:"fan_pct"`
+	Known  bool `json:"known"`
+	// Foreign is the last detector finding (external GPU holders), empty
+	// when the last check found none.
+	Foreign string `json:"foreign,omitempty"`
+	// AgeS is how long ago the last check ran; -1 before the first check.
+	AgeS int64 `json:"age_s"`
+}
+
+// reloadHandler re-reads and validates the service's config file for
+// CmdReloadEnv. An invalid config is reported and the service keeps running
+// untouched; a valid, changed config triggers a GPU-idle-gated restart onto
+// it (same mechanism as staged updates); unchanged is a no-op.
+func reloadHandler(current config.Config, configPath string, restartWhenIdle func(reason string)) func() string {
+	return func() string {
+		ncfg, err := loadMergedConfig(configPath)
+		if err != nil {
+			return "ERR config invalid: " + err.Error()
+		}
+		changed := diffConfig(current, ncfg)
+		if len(changed) == 0 {
+			return "OK config unchanged"
+		}
+		restartWhenIdle("config reload (" + strings.Join(changed, ", ") + ")")
+		return "OK config valid; restarting once the GPU is idle (changed: " + strings.Join(changed, ", ") + ")"
+	}
+}
+
+// diffConfig lists the env names of settings whose values differ between
+// two configs (from each field's env tag, so users recognize them).
+func diffConfig(a, b config.Config) []string {
+	va, vb := reflect.ValueOf(a), reflect.ValueOf(b)
+	t := va.Type()
+	var out []string
+	for i := 0; i < t.NumField(); i++ {
+		if !reflect.DeepEqual(va.Field(i).Interface(), vb.Field(i).Interface()) {
+			name := t.Field(i).Tag.Get("env")
+			if name == "" {
+				name = t.Field(i).Name
+			}
+			out = append(out, name)
+		}
+	}
+	return out
+}
+
+// statusProvider assembles the one-line JSON snapshot for CmdStatus. The
+// loaded-model query to Ollama gets a short timeout so a wedged upstream
+// cannot stall the status channel for long.
+func statusProvider(cfg config.Config, lk *lock.Lock, comfySup *supervise.Process, health *healthTracker, started time.Time, gw *gpuWatch, ollamaClient *ollama.Client) func() string {
+	return func() string {
+		snap := statusSnapshot{
+			Version: version,
+			UptimeS: int64(time.Since(started).Seconds()),
+			UI:      uiURL(cfg.ListenUI),
+		}
+		st, known, foreign, ageS := gw.get()
+		snap.GPU = statusGPU{
+			Enabled: gw.enabled,
+			UsedMB:  st.UsedMB, TotalMB: st.TotalMB, Known: known,
+			TempC: -1, FanPct: -1,
+			Foreign: foreign, AgeS: ageS,
+		}
+		if known {
+			snap.GPU.TempC, snap.GPU.FanPct = st.TempC, st.FanPct
+		}
+		if cfg.OllamaURL != "" {
+			d := statusDownstream{
+				Name: "ollama", URL: cfg.OllamaURL, Up: health.get("ollama"),
+			}
+			if ollamaClient != nil {
+				mctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+				models, err := ollamaClient.LoadedModelDetails(mctx)
+				cancel()
+				if err == nil {
+					d.Models = make([]statusModel, 0, len(models))
+					for _, m := range models {
+						d.Models = append(d.Models, statusModel{Name: m.Name, VRAMMB: m.SizeVRAM / (1024 * 1024)})
+					}
+				}
+			}
+			snap.Downstreams = append(snap.Downstreams, d)
+		}
+		if cfg.ComfyURL != "" {
+			d := statusDownstream{Name: "comfy", URL: cfg.ComfyURL, Up: health.get("comfy")}
+			if comfySup != nil {
+				d.Managed = comfySup.Status()
+			}
+			snap.Downstreams = append(snap.Downstreams, d)
+		}
+		lst := lk.Status()
+		snap.Lock = statusLock{
+			State:       string(lst.State),
+			Detail:      lst.Detail,
+			LLMInflight: lst.LLMInflight,
+			LLMWaiting:  lst.LLMWaiting,
+			ImageQueue:  lst.ImageQueue,
+			External:    lst.External,
+			SinceS:      int64(time.Since(lst.Since).Seconds()),
+		}
+		b, err := json.Marshal(snap)
+		if err != nil {
+			return `{"version":"` + version + `"}`
+		}
+		return string(b)
+	}
+}
+
 // waitForIdle polls the lock until no LLM or image work is active or
 // pending, max at most. Returns false on timeout or cancellation.
 func waitForIdle(ctx context.Context, lk *lock.Lock, max time.Duration) bool {

cmd/gpu-turnstile/monitor.go

@@ -0,0 +1,345 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+	"time"
+
+	"gpu-turnstile/internal/control"
+)
+
+// ANSI colors for the monitor frame.
+const (
+	cReset  = "\x1b[0m"
+	cDim    = "\x1b[2m"
+	cRed    = "\x1b[31m"
+	cGreen  = "\x1b[32m"
+	cYellow = "\x1b[33m"
+	cCyan   = "\x1b[36m"
+)
+
+// hotkeysLine is the monitor's footer.
+const hotkeysLine = " " + cDim + "q quit · u update now · r reload config" + cReset + "\x1b[K\n"
+
+// monitorCommand renders a live status view of the running service,
+// refreshed every second from the control channel. When the service
+// reports a different version and the executable on disk changed (the
+// updater replaced it), the monitor restarts itself onto the new binary.
+// Hotkeys: q quits, u triggers an update check on the service, r asks the
+// service to reload its config file.
+func monitorCommand() int {
+	if !stdoutIsTerminal() {
+		fmt.Fprintln(os.Stderr, "gpu-turnstile: --monitor needs an interactive terminal")
+		return 1
+	}
+	enableVirtualTerminal()
+	restore := enableRawKeys()
+	defer func() {
+		if restore != nil {
+			restore()
+		}
+	}()
+	fmt.Print("\x1b[2J") // clear once; frames then redraw in place
+	defer fmt.Print(cReset + "\n")
+	exe, _ := os.Executable()
+	var exeStamp time.Time
+	if st, err := os.Stat(exe); err == nil {
+		exeStamp = st.ModTime()
+	}
+
+	keys := make(chan byte, 8)
+	go readKeys(keys)
+	ticker := time.NewTicker(time.Second)
+	defer ticker.Stop()
+
+	var note string
+	var noteAt time.Time
+	noteCh := make(chan string, 1)
+	askPending := false
+
+	// ask sends a one-shot command to the service and reports the reply in
+	// the note line. Only one ask runs at a time.
+	ask := func(cmd, busy, label string) {
+		if askPending {
+			return
+		}
+		askPending = true
+		note, noteAt = busy, time.Now()
+		go func() {
+			reply, err := control.Ask(cmd)
+			if err != nil {
+				noteCh <- label + ": no answer from the service"
+				return
+			}
+			msg := strings.TrimPrefix(reply, "OK ")
+			msg = strings.TrimPrefix(msg, "ERR ")
+			noteCh <- label + ": " + msg
+		}()
+	}
+
+	poll := func() string {
+		frame := renderWaiting()
+		if reply, err := control.Ask(control.CmdStatus); err == nil {
+			if msg, ok := strings.CutPrefix(reply, "OK "); ok {
+				var snap statusSnapshot
+				if json.Unmarshal([]byte(msg), &snap) == nil {
+					if snap.Version != "" && snap.Version != version {
+						if exeChanged(exe, exeStamp) {
+							fmt.Print("\x1b[2J\x1b[H")
+							fmt.Printf("gpu-turnstile: service updated to %s — restarting the monitor\n", snap.Version)
+							restartSelf(exe, "--monitor")
+							return "" // re-execed; this process exits below
+						}
+						snap.MonitorNote = fmt.Sprintf("note: the service runs %s, this monitor is %s", snap.Version, version)
+					}
+					if note != "" {
+						snap.MonitorNote = note
+					}
+					frame = renderMonitor(snap, termWidth())
+				}
+			}
+		}
+		return frame
+	}
+
+	for {
+		frame := poll()
+		if frame == "" {
+			return 0 // restartSelf fired
+		}
+		fmt.Print("\x1b[H" + frame + "\x1b[J") // home, frame, clear below
+		select {
+		case <-ticker.C:
+			if note != "" && time.Since(noteAt) > 15*time.Second {
+				note = ""
+			}
+		case k, ok := <-keys:
+			if !ok {
+				keys = nil
+				continue
+			}
+			switch k {
+			case 'q', 'Q', 3: // q or Ctrl+C (raw mode delivers it as a byte)
+				return 0
+			case 'u', 'U':
+				ask(control.CmdUpdateNow, "checking for updates…", "update")
+			case 'r', 'R':
+				ask(control.CmdReloadEnv, "reloading config…", "reload")
+			}
+		case n := <-noteCh:
+			askPending = false
+			note, noteAt = n, time.Now()
+		}
+	}
+}
+
+// readKeys reads single keypresses from stdin (raw mode was enabled by the
+// caller) and delivers them until stdin fails.
+func readKeys(keys chan<- byte) {
+	defer close(keys)
+	buf := make([]byte, 1)
+	for {
+		n, err := os.Stdin.Read(buf)
+		if n > 0 {
+			keys <- buf[0]
+		}
+		if err != nil {
+			return
+		}
+	}
+}
+
+// exeChanged reports whether the executable on disk was replaced since the
+// recorded stamp (the updater swaps it via rename, which changes ModTime).
+func exeChanged(exe string, stamp time.Time) bool {
+	if exe == "" || stamp.IsZero() {
+		return false
+	}
+	st, err := os.Stat(exe)
+	return err == nil && !st.ModTime().Equal(stamp)
+}
+
+// restartSelf starts a fresh copy of this executable with the given args on
+// the same console; the caller exits right after.
+func restartSelf(exe string, args ...string) {
+	cmd := exec.Command(exe, args...)
+	cmd.Stdin, cmd.Stdout, cmd.Stderr = os.Stdin, os.Stdout, os.Stderr
+	cmd.Start() //nolint:errcheck // best effort: on failure we just exit
+}
+
+func renderWaiting() string {
+	return cDim + " gpu-turnstile — waiting for a running service…" + cReset + "\x1b[K\n\x1b[K\n" + hotkeysLine
+}
+
+// renderMonitor draws one full frame. Each line ends with \x1b[K (clear to
+// end of line) so shrinking content leaves no residue.
+func renderMonitor(snap statusSnapshot, width int) string {
+	if width < 40 {
+		width = 80
+	}
+	var b strings.Builder
+
+	left := " gpu-turnstile"
+	if snap.UptimeS > 0 {
+		left += "  " + cDim + "up " + fmtDur(snap.UptimeS) + cReset
+	}
+	right := snap.Version
+	pad := width - printableLen(" gpu-turnstile  up "+fmtDur(snap.UptimeS)) - len(right) - 1
+	if snap.UptimeS == 0 {
+		pad = width - len(" gpu-turnstile") - len(right) - 1
+	}
+	if pad < 1 {
+		pad = 1
+	}
+	b.WriteString(cDim + left + strings.Repeat(" ", pad) + right + cReset + "\x1b[K\n")
+	b.WriteString(cDim + " " + strings.Repeat("─", width-2) + cReset + "\x1b[K\n")
+
+	for _, d := range snap.Downstreams {
+		busy := (d.Name == "ollama" && snap.Lock.State == "llm" && snap.Lock.LLMInflight > 0) ||
+			(d.Name == "comfy" && snap.Lock.State == "image")
+		b.WriteString(renderDownstream(d, busy) + "\x1b[K\n")
+	}
+	b.WriteString("\x1b[K\n")
+	b.WriteString(renderLock(snap.Lock) + "\x1b[K\n")
+	if snap.Lock.ImageQueue > 0 {
+		b.WriteString(fmt.Sprintf(" Queue: %s%d image job(s) waiting%s\x1b[K\n",
+			cYellow, snap.Lock.ImageQueue, cReset))
+	}
+	if snap.GPU.Enabled {
+		b.WriteString(renderGPU(snap.GPU) + "\x1b[K\n")
+	}
+	if snap.UI != "" {
+		b.WriteString(" " + cDim + "UI: " + snap.UI + cReset + "\x1b[K\n")
+	}
+	if snap.MonitorNote != "" {
+		b.WriteString(" " + cYellow + snap.MonitorNote + cReset + "\x1b[K\n")
+	}
+	b.WriteString("\x1b[K\n")
+	b.WriteString(hotkeysLine)
+	return b.String()
+}
+
+// renderGPU renders the GPU line: VRAM usage (when nvidia-smi answered),
+// the game detector's last finding, and how long ago it ran.
+func renderGPU(g statusGPU) string {
+	s := " GPU: "
+	if g.Known {
+		s += renderVRAM(g.UsedMB, g.TotalMB)
+		if g.TempC >= 0 {
+			s += fmt.Sprintf(" · %d°C", g.TempC)
+		}
+		if g.FanPct >= 0 {
+			s += fmt.Sprintf(" · fan %d%%", g.FanPct)
+		}
+	} else {
+		s += cDim + "VRAM unknown (nvidia-smi not answering)" + cReset
+	}
+	if g.Foreign != "" {
+		s += " · external: " + cRed + g.Foreign + cReset
+	} else {
+		s += cDim + " · no external process" + cReset
+	}
+	if g.AgeS >= 0 {
+		s += cDim + fmt.Sprintf(" (checked %s ago)", fmtDur(g.AgeS)) + cReset
+	}
+	return s
+}
+
+// formatMB renders MiB as "4.2 GiB" (or MiB below 1 GiB).
+func formatMB(mb int64) string {
+	if mb >= 1024 {
+		return fmt.Sprintf("%.1f GiB", float64(mb)/1024)
+	}
+	return fmt.Sprintf("%d MiB", mb)
+}
+
+// renderVRAM renders "4.2 / 16.0 GiB used" (or MiB below 1 GiB).
+func renderVRAM(used, total int) string {
+	if total > 0 {
+		return formatMB(int64(used)) + " / " + formatMB(int64(total)) + " used"
+	}
+	return formatMB(int64(used)) + " used"
+}
+
+// printableLen counts characters without ANSI escapes (ASCII-only content).
+func printableLen(s string) int { return len(s) }
+
+func renderDownstream(d statusDownstream, busy bool) string {
+	url := cDim + d.URL + cReset
+	detail := ""
+	if d.Models != nil {
+		if len(d.Models) == 0 {
+			detail = cDim + " · no models loaded" + cReset
+		} else {
+			parts := make([]string, 0, len(d.Models))
+			for _, m := range d.Models {
+				if m.VRAMMB > 0 {
+					parts = append(parts, fmt.Sprintf("%s (%s VRAM)", m.Name, formatMB(m.VRAMMB)))
+				} else {
+					parts = append(parts, m.Name+" (in RAM)")
+				}
+			}
+			detail = " · " + strings.Join(parts, ", ")
+		}
+	}
+	if busy {
+		detail += " · " + cCyan + "busy" + cReset
+	}
+	switch d.Managed {
+	case "stopped":
+		return fmt.Sprintf(" %s○%s %-8s %sstopped (managed — starts on demand)%s  %s",
+			cDim, cReset, d.Name, cDim, cReset, url)
+	case "starting":
+		return fmt.Sprintf(" %s◌%s %-8s %sstarting…%s  %s",
+			cYellow, cReset, d.Name, cYellow, cReset, url)
+	}
+	suffix := ""
+	if d.Managed == "external" {
+		suffix = " (external)"
+	}
+	if d.Up {
+		return fmt.Sprintf(" %s●%s %-8s %sUP%s%s%s   %s", cGreen, cReset, d.Name, cGreen, cReset, suffix, detail, url)
+	}
+	return fmt.Sprintf(" %s●%s %-8s %sDOWN%s %s", cRed, cReset, d.Name, cRed, cReset, url)
+}
+
+func renderLock(l statusLock) string {
+	dur := cDim + "(" + fmtDur(l.SinceS) + ")" + cReset
+	switch l.State {
+	case "idle":
+		return fmt.Sprintf(" Lock: %sidle%s %s", cGreen, cReset, dur)
+	case "llm":
+		s := fmt.Sprintf(" Lock: %sLLM%s — %d in flight", cCyan, cReset, l.LLMInflight)
+		if l.LLMWaiting > 0 {
+			s += fmt.Sprintf(", %d waiting", l.LLMWaiting)
+		}
+		if l.Detail != "" {
+			s += " — " + l.Detail
+		}
+		return s + " " + dur
+	case "image":
+		s := fmt.Sprintf(" Lock: %sIMAGE%s", cYellow, cReset)
+		if l.Detail != "" {
+			s += " — " + l.Detail
+		}
+		return s + " " + dur
+	case "external":
+		return fmt.Sprintf(" Lock: %sEXTERNAL%s — %s %s", cRed, cReset, l.External, dur)
+	}
+	return " Lock: unknown"
+}
+
+// fmtDur renders seconds as a compact duration ("1m32s", "2h07m").
+func fmtDur(s int64) string {
+	if s < 0 {
+		s = 0
+	}
+	d := time.Duration(s) * time.Second
+	if d >= time.Hour {
+		return fmt.Sprintf("%dh%02dm", int(d.Hours()), int(d.Minutes())%60)
+	}
+	return d.Round(time.Second).String()
+}

cmd/gpu-turnstile/monitor_test.go

@@ -0,0 +1,110 @@
+package main
+
+import (
+	"gpu-turnstile/internal/config"
+	"reflect"
+	"strings"
+	"testing"
+)
+
+func TestRenderMonitor(t *testing.T) {
+	snap := statusSnapshot{
+		Version: "v0.2.3",
+		UptimeS: 3725,
+		Downstreams: []statusDownstream{
+			{Name: "ollama", URL: "http://127.0.0.1:11435", Up: true},
+			{Name: "comfy", URL: "http://127.0.0.1:8189", Managed: "stopped"},
+		},
+		Lock: statusLock{State: "llm", LLMInflight: 2, LLMWaiting: 1, Detail: "ollama: POST /api/generate", SinceS: 95},
+	}
+	frame := renderMonitor(snap, 80)
+	for _, want := range []string{"v0.2.3", "1h02m", "ollama", "UP", "comfy", "stopped", "LLM", "2 in flight", "1 waiting", "1m35s"} {
+		if !strings.Contains(frame, want) {
+			t.Errorf("frame missing %q:\n%s", want, frame)
+		}
+	}
+
+	snap.Lock = statusLock{State: "external", External: "cyberpunk2077.exe (pid 1234)", SinceS: 3, ImageQueue: 2}
+	frame = renderMonitor(snap, 40) // narrow: falls back to 80
+	for _, want := range []string{"EXTERNAL", "cyberpunk2077.exe", "2 image job(s) waiting"} {
+		if !strings.Contains(frame, want) {
+			t.Errorf("frame missing %q:\n%s", want, frame)
+		}
+	}
+
+	snap.GPU = statusGPU{Enabled: true, Known: true, UsedMB: 4300, TotalMB: 16384, TempC: 55, FanPct: 42, Foreign: "cyberpunk2077.exe (pid 1234)", AgeS: 12}
+	frame = renderMonitor(snap, 80)
+	for _, want := range []string{"GPU:", "4.2 GiB / 16.0 GiB used", "55°C", "fan 42%", "external:", "checked 12s ago"} {
+		if !strings.Contains(frame, want) {
+			t.Errorf("frame missing %q:\n%s", want, frame)
+		}
+	}
+
+	snap.GPU = statusGPU{Enabled: true, AgeS: 4}
+	frame = renderMonitor(snap, 80)
+	for _, want := range []string{"VRAM unknown", "no external process"} {
+		if !strings.Contains(frame, want) {
+			t.Errorf("frame missing %q:\n%s", want, frame)
+		}
+	}
+
+	snap.Downstreams[0].Models = []statusModel{{Name: "llama3.1:8b", VRAMMB: 4900}, {Name: "embed", VRAMMB: 0}}
+	snap.Lock = statusLock{State: "llm", LLMInflight: 1}
+	frame = renderMonitor(snap, 80)
+	for _, want := range []string{"llama3.1:8b (4.8 GiB VRAM)", "embed (in RAM)", "busy"} {
+		if !strings.Contains(frame, want) {
+			t.Errorf("frame missing %q:\n%s", want, frame)
+		}
+	}
+
+	snap.Downstreams[0].Models = []statusModel{}
+	snap.Lock = statusLock{State: "idle"}
+	frame = renderMonitor(snap, 80)
+	if !strings.Contains(frame, "no models loaded") {
+		t.Errorf("frame missing %q:\n%s", "no models loaded", frame)
+	}
+	if strings.Contains(frame, "busy") {
+		t.Errorf("idle lock still shows busy:\n%s", frame)
+	}
+
+	snap.UI = "http://127.0.0.1:7860"
+	frame = renderMonitor(snap, 80)
+	if !strings.Contains(frame, "UI: http://127.0.0.1:7860") {
+		t.Errorf("frame missing UI line:\n%s", frame)
+	}
+}
+
+func TestFmtDur(t *testing.T) {
+	cases := map[int64]string{0: "0s", 5: "5s", 95: "1m35s", 3725: "1h02m", -3: "0s"}
+	for in, want := range cases {
+		if got := fmtDur(in); got != want {
+			t.Errorf("fmtDur(%d) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestDiffConfig(t *testing.T) {
+	a := config.Defaults()
+	b := a
+	if got := diffConfig(a, b); len(got) != 0 {
+		t.Fatalf("identical configs: got %v", got)
+	}
+	b.LogLevel = -4
+	b.GameProcs = []string{"game.exe"}
+	got := diffConfig(a, b)
+	if len(got) != 2 || got[0] != "GAME_PROCS" || got[1] != "LOGLEVEL" {
+		t.Fatalf("got %v, want [GAME_PROCS LOGLEVEL]", got)
+	}
+}
+
+// Every Config field must carry an env tag so user-facing output (the
+// reload diff) can name the setting the user would actually change.
+func TestConfigFieldsHaveEnvTags(t *testing.T) {
+	typ := reflect.TypeOf(config.Config{})
+	for i := 0; i < typ.NumField(); i++ {
+		f := typ.Field(i)
+		if f.Tag.Get("env") == "" {
+			t.Errorf("config.Config.%s has no env tag", f.Name)
+		}
+	}
+}

cmd/gpu-turnstile/monitor_unix.go

@@ -0,0 +1,39 @@
+//go:build !windows
+
+package main
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// enableVirtualTerminal is a no-op: unix terminals speak ANSI natively.
+func enableVirtualTerminal() {}
+
+// termWidth is the terminal width in columns, 80 when unknown.
+func termWidth() int {
+	ws, err := unix.IoctlGetWinsize(int(os.Stdout.Fd()), unix.TIOCGWINSZ)
+	if err != nil || ws.Col == 0 {
+		return 80
+	}
+	return int(ws.Col)
+}
+
+// enableRawKeys switches the terminal to per-keypress mode (ICANON and ECHO
+// off) and returns the restore function, nil when stdin is not a terminal.
+func enableRawKeys() func() {
+	fd := int(os.Stdin.Fd())
+	term, err := unix.IoctlGetTermios(fd, unix.TCGETS)
+	if err != nil {
+		return nil
+	}
+	raw := *term
+	raw.Lflag &^= unix.ICANON | unix.ECHO
+	raw.Cc[unix.VMIN] = 1
+	raw.Cc[unix.VTIME] = 0
+	if err := unix.IoctlSetTermios(fd, unix.TCSETS, &raw); err != nil {
+		return nil
+	}
+	return func() { unix.IoctlSetTermios(fd, unix.TCSETS, term) } //nolint:errcheck
+}

cmd/gpu-turnstile/monitor_windows.go

@@ -0,0 +1,49 @@
+//go:build windows
+
+package main
+
+import (
+	"os"
+
+	"golang.org/x/sys/windows"
+)
+
+// enableVirtualTerminal asks the console to honor ANSI escapes (Windows 10+);
+// mintty/Git Bash already does, so errors are ignored.
+func enableVirtualTerminal() {
+	h := windows.Handle(os.Stdout.Fd())
+	var mode uint32
+	if err := windows.GetConsoleMode(h, &mode); err != nil {
+		return
+	}
+	windows.SetConsoleMode(h, mode|windows.ENABLE_VIRTUAL_TERMINAL_PROCESSING) //nolint:errcheck
+}
+
+// termWidth is the console window width in columns, 80 when unknown.
+func termWidth() int {
+	var info windows.ConsoleScreenBufferInfo
+	if err := windows.GetConsoleScreenBufferInfo(windows.Handle(os.Stdout.Fd()), &info); err != nil {
+		return 80
+	}
+	return int(info.Window.Right-info.Window.Left) + 1
+}
+
+// enableRawKeys puts the console's stdin into per-keypress mode (no line
+// buffering, no echo) and returns the restore function. When stdin is not a
+// real console (mintty/Git Bash pipes) it returns nil: ptys already deliver
+// keystrokes immediately.
+func enableRawKeys() func() {
+	h := windows.Handle(os.Stdin.Fd())
+	var mode uint32
+	if err := windows.GetConsoleMode(h, &mode); err != nil {
+		return nil
+	}
+	const (
+		enableLineInput = 0x0002
+		enableEchoInput = 0x0004
+	)
+	if err := windows.SetConsoleMode(h, mode&^(enableLineInput|enableEchoInput)); err != nil {
+		return nil
+	}
+	return func() { windows.SetConsoleMode(h, mode) } //nolint:errcheck
+}

cmd/gpu-turnstile/webui.go

@@ -0,0 +1,242 @@
+package main
+
+import (
+	"crypto/subtle"
+	"fmt"
+	"net"
+	"net/http"
+	"strings"
+
+	"gpu-turnstile/internal/control"
+)
+
+// uiHandler serves the web UI (LISTEN_UI): a single page mirroring
+// --monitor, its status JSON (the same snapshot the control channel
+// serves) and action endpoints for the update/reload buttons. user/pass
+// (UI_USER/UI_PASS) enable HTTP basic auth; both empty means open access,
+// which is meant for localhost only.
+func uiHandler(h control.Handler, user, pass string) http.Handler {
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/" {
+			http.NotFound(w, r)
+			return
+		}
+		w.Header().Set("Content-Type", "text/html; charset=utf-8")
+		fmt.Fprint(w, uiPage)
+	})
+	mux.HandleFunc("GET /api/status", func(w http.ResponseWriter, r *http.Request) {
+		msg := strings.TrimPrefix(h(control.CmdStatus), "OK ")
+		w.Header().Set("Content-Type", "application/json")
+		if strings.HasPrefix(msg, "ERR ") {
+			w.WriteHeader(http.StatusInternalServerError)
+			msg = strings.TrimPrefix(msg, "ERR ")
+		}
+		fmt.Fprint(w, msg)
+	})
+	mux.HandleFunc("POST /api/action/{cmd}", func(w http.ResponseWriter, r *http.Request) {
+		cmd := r.PathValue("cmd")
+		if cmd != control.CmdUpdateNow && cmd != control.CmdReloadEnv {
+			http.NotFound(w, r)
+			return
+		}
+		msg := strings.TrimPrefix(h(cmd), "OK ")
+		w.Header().Set("Content-Type", "text/plain; charset=utf-8")
+		if strings.HasPrefix(msg, "ERR ") {
+			w.WriteHeader(http.StatusConflict)
+			msg = strings.TrimPrefix(msg, "ERR ")
+		}
+		fmt.Fprint(w, msg)
+	})
+	if user != "" {
+		return basicAuth(mux, user, pass)
+	}
+	return mux
+}
+
+// basicAuth wraps next with HTTP basic auth (constant-time compares).
+func basicAuth(next http.Handler, user, pass string) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		u, p, ok := r.BasicAuth()
+		if !ok ||
+			subtle.ConstantTimeCompare([]byte(u), []byte(user)) != 1 ||
+			subtle.ConstantTimeCompare([]byte(p), []byte(pass)) != 1 {
+			w.Header().Set("WWW-Authenticate", `Basic realm="gpu-turnstile"`)
+			http.Error(w, "unauthorized", http.StatusUnauthorized)
+			return
+		}
+		next.ServeHTTP(w, r)
+	})
+}
+
+// isLoopbackHost reports whether a listen host is localhost-only. An empty
+// host (":7860") binds all interfaces and is therefore not loopback.
+func isLoopbackHost(host string) bool {
+	if host == "localhost" {
+		return true
+	}
+	ip := net.ParseIP(host)
+	return ip != nil && ip.IsLoopback()
+}
+
+// uiURL renders a LISTEN_UI address as a browsable URL ("" when disabled).
+// A wildcard host reads as 127.0.0.1 — that's where the browser is usually
+// running.
+func uiURL(addr string) string {
+	if addr == "" {
+		return ""
+	}
+	if strings.HasPrefix(addr, ":") {
+		addr = "127.0.0.1" + addr
+	}
+	return "http://" + addr
+}
+
+// uiPage is the whole web UI: a terminal-styled page that polls /api/status
+// once a second and renders the same lines as --monitor. The buttons (and
+// the u/r keys) hit the action endpoints, like the monitor's hotkeys.
+const uiPage = `<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>gpu-turnstile</title>
+<style>
+  body { background:#111; color:#ddd; font:14px/1.6 ui-monospace,Consolas,"Cascadia Mono",monospace;
+         margin:0 auto; padding:14px 18px; max-width:860px; }
+  header { display:flex; justify-content:space-between; border-bottom:1px solid #333;
+           padding-bottom:4px; margin-bottom:10px; }
+  footer { border-top:1px solid #333; margin-top:10px; padding-top:8px;
+           display:flex; gap:8px; align-items:center; }
+  button { background:#1d1d1d; color:#bbb; border:1px solid #444; border-radius:4px;
+           padding:2px 10px; font:inherit; cursor:pointer; }
+  button:hover { background:#2c2c2c; color:#eee; }
+  .dim { color:#777; } .up { color:#4caf50; } .down { color:#ef5350; }
+  .llm { color:#26c6da; } .warn { color:#fbc02d; }
+  #note { color:#fbc02d; min-height:1.6em; }
+</style>
+</head>
+<body>
+<header><span>gpu-turnstile <span id="up" class="dim"></span></span><span id="ver" class="dim"></span></header>
+<div id="downstreams"></div>
+<div id="lock"></div>
+<div id="gpu"></div>
+<div id="note"></div>
+<footer>
+  <button id="bu">u update now</button>
+  <button id="br">r reload config</button>
+  <span class="dim">refreshes every second</span>
+</footer>
+<script>
+const $ = id => document.getElementById(id);
+const el = (cls, text) => { const e = document.createElement("span"); if (cls) e.className = cls; e.textContent = text; return e; };
+const line = (...parts) => { const d = document.createElement("div"); d.append(...parts); return d; };
+const fmtMB = mb => mb >= 1024 ? (mb / 1024).toFixed(1) + " GiB" : mb + " MiB";
+const fmtDur = s => {
+  s = Math.max(0, Math.floor(s));
+  if (s >= 3600) return Math.floor(s / 3600) + "h" + String(Math.floor(s / 60) % 60).padStart(2, "0") + "m";
+  if (s >= 60) return Math.floor(s / 60) + "m" + String(s % 60).padStart(2, "0") + "s";
+  return s + "s";
+};
+
+function renderDownstream(d, lock) {
+  const busy = (d.name === "ollama" && lock.state === "llm" && lock.llm_inflight > 0) ||
+               (d.name === "comfy" && lock.state === "image");
+  if (d.managed === "stopped")
+    return line(el("dim", "○ " + d.name + " stopped (managed — starts on demand)  " + d.url));
+  if (d.managed === "starting")
+    return line(el("warn", "◌ " + d.name + " starting…  "), el("dim", d.url));
+  const parts = [el(d.up ? "up" : "down", "● " + d.name + " " + (d.up ? "UP" : "DOWN"))];
+  if (d.managed === "external") parts.push(el("", " (external)"));
+  if (d.models) {
+    if (d.models.length === 0) {
+      parts.push(el("dim", " · no models loaded"));
+    } else {
+      parts.push(el("", " · " + d.models.map(m =>
+        m.vram_mb > 0 ? m.name + " (" + fmtMB(m.vram_mb) + " VRAM)" : m.name + " (in RAM)").join(", ")));
+    }
+  }
+  if (busy) parts.push(el("llm", " · busy"));
+  parts.push(el("dim", "   " + d.url));
+  return line(...parts);
+}
+
+function renderLock(l) {
+  const dur = el("dim", " (" + fmtDur(l.since_s || 0) + ")");
+  switch (l.state) {
+    case "idle": return line(el("up", "Lock: idle"), dur);
+    case "llm": {
+      let s = "Lock: LLM — " + l.llm_inflight + " in flight";
+      if (l.llm_waiting > 0) s += ", " + l.llm_waiting + " waiting";
+      if (l.detail) s += " — " + l.detail;
+      return line(el("llm", s), dur);
+    }
+    case "image": return line(el("warn", "Lock: IMAGE" + (l.detail ? " — " + l.detail : "")), dur);
+    case "external": return line(el("down", "Lock: EXTERNAL — " + (l.external || "")), dur);
+  }
+  return line(el("", "Lock: unknown"));
+}
+
+function renderGPU(g) {
+  const parts = [el("", "GPU: ")];
+  if (g.known) {
+    let s = fmtMB(g.used_mb) + " / " + fmtMB(g.total_mb) + " used";
+    if (g.temp_c >= 0) s += " · " + g.temp_c + "°C";
+    if (g.fan_pct >= 0) s += " · fan " + g.fan_pct + "%";
+    parts.push(el("", s));
+  } else {
+    parts.push(el("dim", "VRAM unknown (nvidia-smi not answering)"));
+  }
+  parts.push(g.foreign ? el("down", " · external: " + g.foreign) : el("dim", " · no external process"));
+  if (g.age_s >= 0) parts.push(el("dim", " (checked " + fmtDur(g.age_s) + " ago)"));
+  return line(...parts);
+}
+
+async function poll() {
+  let snap;
+  try {
+    const r = await fetch("/api/status");
+    if (!r.ok) throw new Error();
+    snap = await r.json();
+  } catch {
+    $("downstreams").replaceChildren(line(el("dim", "gpu-turnstile — waiting for a running service…")));
+    $("lock").replaceChildren(); $("gpu").replaceChildren();
+    return;
+  }
+  $("ver").textContent = snap.version || "";
+  $("up").textContent = snap.uptime_s > 0 ? "up " + fmtDur(snap.uptime_s) : "";
+  const ds = $("downstreams"); ds.replaceChildren();
+  for (const d of snap.downstreams || []) ds.appendChild(renderDownstream(d, snap.lock || {}));
+  const lk = $("lock"); lk.replaceChildren(renderLock(snap.lock || {}));
+  if ((snap.lock || {}).image_queue > 0)
+    lk.appendChild(line(el("warn", "Queue: " + snap.lock.image_queue + " image job(s) waiting")));
+  const g = $("gpu"); g.replaceChildren();
+  if (snap.gpu && snap.gpu.enabled) g.appendChild(renderGPU(snap.gpu));
+}
+
+let noteTimer;
+async function act(cmd) {
+  const note = $("note");
+  note.textContent = "…";
+  try {
+    const r = await fetch("/api/action/" + cmd, { method: "POST" });
+    note.textContent = await r.text();
+  } catch {
+    note.textContent = "no answer from the service";
+  }
+  clearTimeout(noteTimer);
+  noteTimer = setTimeout(() => { note.textContent = ""; }, 15000);
+}
+
+$("bu").onclick = () => act("update-now");
+$("br").onclick = () => act("reload-env");
+document.addEventListener("keydown", e => {
+  if (e.key === "u") act("update-now");
+  if (e.key === "r") act("reload-env");
+});
+poll();
+setInterval(poll, 1000);
+</script>
+</body>
+</html>
+`

cmd/gpu-turnstile/webui_test.go

@@ -0,0 +1,185 @@
+package main
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gpu-turnstile/internal/control"
+)
+
+func testUIHandler() *httptest.Server {
+	h := uiHandler(func(cmd string) string {
+		switch cmd {
+		case control.CmdStatus:
+			return `OK {"version":"v9.9.9","uptime_s":5}`
+		case control.CmdUpdateNow:
+			return "OK v9.9.9 is up to date"
+		default:
+			return "ERR unknown command: " + cmd
+		}
+	}, "", "")
+	return httptest.NewServer(h)
+}
+
+func TestUIPage(t *testing.T) {
+	srv := testUIHandler()
+	defer srv.Close()
+	resp, err := srv.Client().Get(srv.URL + "/")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body := make([]byte, 1024*64)
+	n, _ := resp.Body.Read(body)
+	if resp.StatusCode != 200 {
+		t.Fatalf("status %d", resp.StatusCode)
+	}
+	if ct := resp.Header.Get("Content-Type"); !strings.Contains(ct, "text/html") {
+		t.Errorf("Content-Type %q", ct)
+	}
+	if !strings.Contains(string(body[:n]), "gpu-turnstile") {
+		t.Error("page does not mention gpu-turnstile")
+	}
+}
+
+func TestUIStatus(t *testing.T) {
+	srv := testUIHandler()
+	defer srv.Close()
+	resp, err := srv.Client().Get(srv.URL + "/api/status")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body := make([]byte, 4096)
+	n, _ := resp.Body.Read(body)
+	if resp.StatusCode != 200 {
+		t.Fatalf("status %d", resp.StatusCode)
+	}
+	if ct := resp.Header.Get("Content-Type"); !strings.Contains(ct, "application/json") {
+		t.Errorf("Content-Type %q", ct)
+	}
+	if got := strings.TrimSpace(string(body[:n])); got != `{"version":"v9.9.9","uptime_s":5}` {
+		t.Errorf("body %q", got)
+	}
+}
+
+func TestUIActions(t *testing.T) {
+	srv := testUIHandler()
+	defer srv.Close()
+
+	resp, err := srv.Client().Post(srv.URL+"/api/action/update-now", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body := make([]byte, 4096)
+	n, _ := resp.Body.Read(body)
+	if resp.StatusCode != 200 || string(body[:n]) != "v9.9.9 is up to date" {
+		t.Errorf("update-now: %d %q", resp.StatusCode, body[:n])
+	}
+
+	// The handler answers ERR for reload-env: mapped to 409 with the
+	// message stripped.
+	resp2, err := srv.Client().Post(srv.URL+"/api/action/reload-env", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp2.Body.Close()
+	n, _ = resp2.Body.Read(body)
+	if resp2.StatusCode != 409 || !strings.HasPrefix(string(body[:n]), "unknown command") {
+		t.Errorf("reload-env: %d %q", resp2.StatusCode, body[:n])
+	}
+
+	// Unknown actions are not routed to the handler.
+	resp3, err := srv.Client().Post(srv.URL+"/api/action/reboot", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp3.Body.Close()
+	if resp3.StatusCode != 404 {
+		t.Errorf("bogus action: %d, want 404", resp3.StatusCode)
+	}
+
+	// GET on an action path falls through to the root handler's guard,
+	// which only serves "/" itself.
+	resp4, err := srv.Client().Get(srv.URL + "/api/action/update-now")
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp4.Body.Close()
+	if resp4.StatusCode != 404 {
+		t.Errorf("GET action: %d, want 404", resp4.StatusCode)
+	}
+}
+
+func TestUIBasicAuth(t *testing.T) {
+	h := uiHandler(func(cmd string) string { return `OK {"version":"v9.9.9"}` }, "admin", "s3cret")
+	srv := httptest.NewServer(h)
+	defer srv.Close()
+
+	// No credentials: 401 with the auth challenge on page and API alike.
+	for _, path := range []string{"/", "/api/status"} {
+		resp, err := srv.Client().Get(srv.URL + path)
+		if err != nil {
+			t.Fatal(err)
+		}
+		resp.Body.Close()
+		if resp.StatusCode != 401 {
+			t.Errorf("GET %s without auth: %d, want 401", path, resp.StatusCode)
+		}
+		if resp.Header.Get("WWW-Authenticate") == "" {
+			t.Errorf("GET %s: no WWW-Authenticate header", path)
+		}
+	}
+
+	// Wrong credentials: 401. Right credentials: 200.
+	req, _ := http.NewRequest("GET", srv.URL+"/api/status", nil)
+	req.SetBasicAuth("admin", "wrong")
+	resp, err := srv.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	resp.Body.Close()
+	if resp.StatusCode != 401 {
+		t.Errorf("wrong password: %d, want 401", resp.StatusCode)
+	}
+
+	req, _ = http.NewRequest("GET", srv.URL+"/api/status", nil)
+	req.SetBasicAuth("admin", "s3cret")
+	resp, err = srv.Client().Do(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	body := make([]byte, 4096)
+	n, _ := resp.Body.Read(body)
+	if resp.StatusCode != 200 || strings.TrimSpace(string(body[:n])) != `{"version":"v9.9.9"}` {
+		t.Errorf("with auth: %d %q", resp.StatusCode, body[:n])
+	}
+}
+
+func TestIsLoopbackHost(t *testing.T) {
+	for host, want := range map[string]bool{
+		"127.0.0.1": true, "localhost": true, "::1": true,
+		"": false, "0.0.0.0": false, "192.168.1.10": false, "example.com": false,
+	} {
+		if got := isLoopbackHost(host); got != want {
+			t.Errorf("isLoopbackHost(%q) = %v, want %v", host, got, want)
+		}
+	}
+}
+
+func TestUIURL(t *testing.T) {
+	for addr, want := range map[string]string{
+		"":                 "",
+		"127.0.0.1:7860":   "http://127.0.0.1:7860",
+		":7860":            "http://127.0.0.1:7860",
+		"192.168.1.5:9000": "http://192.168.1.5:9000",
+	} {
+		if got := uiURL(addr); got != want {
+			t.Errorf("uiURL(%q) = %q, want %q", addr, got, want)
+		}
+	}
+}

compose.yaml.example

@@ -6,7 +6,7 @@
 # ComfyUI --listen 0.0.0.0 --port 8189).
 services:
   gpu-turnstile:
-    image: git.rambossek.at/public/gpu-turnstile:v0.2.3
+    image: git.rambossek.at/public/gpu-turnstile:v0.3.2
     restart: unless-stopped
     environment:
       # Each consumer is enabled by setting its URL; leave one unset to

internal/config/config.go

@@ -13,43 +13,48 @@ import (
 	"time"
 )
 
-// Config holds every gpu-turnstile setting.
+// Config holds every gpu-turnstile setting. Each field's env tag names the
+// environment variable / config-file key that sets it; user-facing output
+// (e.g. the reload diff) uses those names, never the Go field names.
 type Config struct {
-	ListenOllama   string
-	ListenComfy    string
-	OllamaURL      string
-	ComfyURL       string
-	UnloadTimeout  time.Duration
-	JobTimeout     time.Duration
-	LLMWaitTimeout time.Duration
+	ListenOllama   string        `env:"LISTEN_OLLAMA"`
+	ListenComfy    string        `env:"LISTEN_COMFY"`
+	ListenUI       string        `env:"LISTEN_UI"`
+	UIUser         string        `env:"UI_USER"`
+	UIPass         string        `env:"UI_PASS"`
+	OllamaURL      string        `env:"OLLAMA_URL"`
+	ComfyURL       string        `env:"COMFY_URL"`
+	UnloadTimeout  time.Duration `env:"UNLOAD_TIMEOUT"`
+	JobTimeout     time.Duration `env:"JOB_TIMEOUT"`
+	LLMWaitTimeout time.Duration `env:"LLM_WAIT_TIMEOUT"`
 
-	UnloadPollInterval  time.Duration
-	HistoryPollInterval time.Duration
-	ProbeTimeout        time.Duration
-	HealthInterval      time.Duration
-	FreeTimeout         time.Duration
-	WarmTimeout         time.Duration
-	ShutdownTimeout     time.Duration
-	BackoffInitial      time.Duration
-	BackoffMax          time.Duration
-	PromptCaptureLimit  int64
+	UnloadPollInterval  time.Duration `env:"UNLOAD_POLL_INTERVAL"`
+	HistoryPollInterval time.Duration `env:"HISTORY_POLL_INTERVAL"`
+	ProbeTimeout        time.Duration `env:"PROBE_TIMEOUT"`
+	HealthInterval      time.Duration `env:"HEALTH_INTERVAL"`
+	FreeTimeout         time.Duration `env:"FREE_TIMEOUT"`
+	WarmTimeout         time.Duration `env:"WARM_TIMEOUT"`
+	ShutdownTimeout     time.Duration `env:"SHUTDOWN_TIMEOUT"`
+	BackoffInitial      time.Duration `env:"BACKOFF_INITIAL"`
+	BackoffMax          time.Duration `env:"BACKOFF_MAX"`
+	PromptCaptureLimit  int64         `env:"PROMPT_CAPTURE_LIMIT"`
 
-	AutoUpdate     bool
-	UpdateInterval time.Duration
-	UpdateRepo     string
-	UpdateAsset    string
+	AutoUpdate     bool          `env:"AUTO_UPDATE"`
+	UpdateInterval time.Duration `env:"UPDATE_INTERVAL"`
+	UpdateRepo     string        `env:"UPDATE_REPO"`
+	UpdateAsset    string        `env:"UPDATE_ASSET"`
 
 	// AppVersion is the version the user wants to run: "dev" disables
 	// updates, "stable" tracks the latest release, anything else is an
 	// exact vX.Y.Z release to pin. From APP_VER; defaults to "stable".
-	AppVersion string
+	AppVersion string `env:"APP_VER"`
 
 	// LLMBusyMode is "wait" (hold requests until the lock is free or
 	// LLMWaitTimeout expires) or "reject" (immediately answer with
 	// LLMBusyStatus + Retry-After when an image job is active or pending).
-	LLMBusyMode    string
-	LLMBusyStatus  int
-	BusyRetryAfter int
+	LLMBusyMode    string `env:"LLM_BUSY_MODE"`
+	LLMBusyStatus  int    `env:"LLM_BUSY_STATUS"`
+	BusyRetryAfter int    `env:"BUSY_RETRY_AFTER"`
 
 	// ComfyCmd spawns and supervises a ComfyUI server on demand. When
 	// ComfyCmd is empty but ComfyDir is set, management is enabled with the
@@ -59,26 +64,29 @@ type Config struct {
 	// set explicitly. The managed server is stopped after ComfyIdleTimeout
 	// without requests, freeing its VRAM; ComfyStartTimeout bounds how long
 	// a request waits for it to come up.
-	ComfyCmd          string
-	ComfyDir          string
-	ComfyIdleTimeout  time.Duration
-	ComfyStartTimeout time.Duration
+	ComfyCmd          string        `env:"COMFY_CMD"`
+	ComfyDir          string        `env:"COMFY_DIR"`
+	ComfyIdleTimeout  time.Duration `env:"COMFY_IDLE_TIMEOUT"`
+	ComfyStartTimeout time.Duration `env:"COMFY_START_TIMEOUT"`
 
 	// GameProcs (GAME_PROCS) is a watch list of process names; while any of
 	// them runs, the GPU is treated as held by a foreign process. The
 	// nvidia-smi path (GPUForeignVRAMMB, GPU_FOREIGN_VRAM_MB) does the same
 	// when a process not in GPUIgnoreProcs (GPU_IGNORE_PROCS) holds more than
-	// that many MiB of VRAM. GamePollInterval (GAME_POLL_INTERVAL) is how
-	// often both checks run.
-	GameProcs        []string
-	GPUForeignVRAMMB int
-	GPUIgnoreProcs   []string
-	GamePollInterval time.Duration
+	// that many MiB of VRAM, and the PDH path (GPUForeignUtilPct,
+	// GPU_FOREIGN_UTIL_PCT) when such a process uses more than that many
+	// percent of the GPU 3D engine (Windows only). GamePollInterval
+	// (GAME_POLL_INTERVAL) is how often all checks run.
+	GameProcs         []string      `env:"GAME_PROCS"`
+	GPUForeignVRAMMB  int           `env:"GPU_FOREIGN_VRAM_MB"`
+	GPUForeignUtilPct int           `env:"GPU_FOREIGN_UTIL_PCT"`
+	GPUIgnoreProcs    []string      `env:"GPU_IGNORE_PROCS"`
+	GamePollInterval  time.Duration `env:"GAME_POLL_INTERVAL"`
 
-	WarmModel string
-	LogLevel  slog.Level
-	LogJSON   bool
-	LogFile   string
+	WarmModel string     `env:"WARM_MODEL"`
+	LogLevel  slog.Level `env:"LOGLEVEL"`
+	LogJSON   bool       `env:"LOG_FORMAT"`
+	LogFile   string     `env:"LOG_FILE"`
 }
 
 // Defaults returns the configuration used when neither the environment nor
@@ -117,8 +125,9 @@ func Defaults() Config {
 		ComfyStartTimeout: 2 * time.Minute,
 
 		// ComfyUI runs under python; excluding it (and Ollama) by name keeps
-		// our own consumers from tripping the foreign-VRAM check.
-		GPUIgnoreProcs:   []string{"ollama", "ollama app", "ollama_llama_server", "python", "pythonw"},
+		// our own consumers from tripping the foreign-VRAM check. dwm is the
+		// desktop compositor — it always shows some 3D-engine usage.
+		GPUIgnoreProcs:   []string{"ollama", "ollama app", "ollama_llama_server", "python", "pythonw", "dwm"},
 		GamePollInterval: 15 * time.Second,
 
 		LogLevel: slog.LevelWarn,
@@ -190,6 +199,9 @@ func Load(getenv func(string) string) (Config, error) {
 	}{
 		{"LISTEN_OLLAMA", &cfg.ListenOllama},
 		{"LISTEN_COMFY", &cfg.ListenComfy},
+		{"LISTEN_UI", &cfg.ListenUI},
+		{"UI_USER", &cfg.UIUser},
+		{"UI_PASS", &cfg.UIPass},
 		{"OLLAMA_URL", &cfg.OllamaURL},
 		{"COMFY_URL", &cfg.ComfyURL},
 		{"WARM_MODEL", &cfg.WarmModel},
@@ -241,6 +253,13 @@ func Load(getenv func(string) string) (Config, error) {
 		}
 		cfg.GPUForeignVRAMMB = n
 	}
+	if v := getenv("GPU_FOREIGN_UTIL_PCT"); v != "" {
+		n, err := strconv.Atoi(v)
+		if err != nil || n < 0 || n > 100 {
+			return cfg, fmt.Errorf("GPU_FOREIGN_UTIL_PCT: must be an integer in 0-100 (percent of the GPU 3D engine, 0 = disabled)")
+		}
+		cfg.GPUForeignUtilPct = n
+	}
 	if v := getenv("PROMPT_CAPTURE_LIMIT"); v != "" {
 		n, err := strconv.ParseInt(v, 10, 64)
 		if err != nil || n < 0 {
@@ -305,6 +324,12 @@ func Load(getenv func(string) string) (Config, error) {
 	default:
 		return cfg, fmt.Errorf("LOG_FORMAT: must be \"text\" or \"json\"")
 	}
+	if (cfg.UIUser == "") != (cfg.UIPass == "") {
+		return cfg, fmt.Errorf("UI_USER and UI_PASS must be set together (both empty = no auth)")
+	}
+	if cfg.UIUser != "" && cfg.ListenUI == "" {
+		return cfg, fmt.Errorf("UI_USER/UI_PASS have no effect without LISTEN_UI")
+	}
 	if cfg.ComfyCmd != "" && cfg.ComfyURL == "" {
 		return cfg, fmt.Errorf("COMFY_CMD requires COMFY_URL to be set (the proxy needs somewhere to forward)")
 	}

internal/config/config_test.go

@@ -97,6 +97,30 @@ func TestComfyCmdRequiresURL(t *testing.T) {
 	}
 }
 
+func TestUIAuthPairing(t *testing.T) {
+	env := func(set map[string]string) func(string) string {
+		return func(k string) string { return set[k] }
+	}
+	// Only one of UI_USER/UI_PASS: error.
+	_, err := Load(env(map[string]string{"OLLAMA_URL": "http://x", "LISTEN_UI": "127.0.0.1:7860", "UI_USER": "admin"}))
+	if err == nil || !strings.Contains(err.Error(), "UI_USER and UI_PASS must be set together") {
+		t.Fatalf("err = %v, want pairing error", err)
+	}
+	// Auth without LISTEN_UI: error.
+	_, err = Load(env(map[string]string{"OLLAMA_URL": "http://x", "UI_USER": "admin", "UI_PASS": "x"}))
+	if err == nil || !strings.Contains(err.Error(), "no effect without LISTEN_UI") {
+		t.Fatalf("err = %v, want LISTEN_UI error", err)
+	}
+	// Both with LISTEN_UI: loads.
+	cfg, err := Load(env(map[string]string{"OLLAMA_URL": "http://x", "LISTEN_UI": "127.0.0.1:7860", "UI_USER": "admin", "UI_PASS": "x"}))
+	if err != nil {
+		t.Fatalf("auth pair with LISTEN_UI must load: %v", err)
+	}
+	if cfg.UIUser != "admin" || cfg.UIPass != "x" {
+		t.Errorf("got %q/%q", cfg.UIUser, cfg.UIPass)
+	}
+}
+
 func TestParseEnvFile(t *testing.T) {
 	input := `# comment
 OLLAMA_URL=http://host:11435

internal/config/sample.go

@@ -25,6 +25,9 @@ func sampleEntries(logFile string) []sampleEntry {
 	return []sampleEntry{
 		{"LISTEN_OLLAMA", ":11434", "Listen address for Ollama-compatible clients (gpu-turnstile poses as Ollama here)", false},
 		{"LISTEN_COMFY", ":8188", "Listen address for ComfyUI clients (gpu-turnstile poses as ComfyUI here)", false},
+		{"LISTEN_UI", "127.0.0.1:7860", "Web UI listen address (host:port): live status like --monitor, with update/reload buttons (default: empty = disabled; bind 127.0.0.1 unless you set auth)", false},
+		{"UI_USER", "", "HTTP basic auth for the web UI (UI_USER and UI_PASS must be set together; both empty = no auth)", false},
+		{"UI_PASS", "", "See UI_USER", false},
 		{"OLLAMA_URL", "http://127.0.0.1:11434", "Ollama upstream URL; setting it enables the Ollama consumer (default: empty = disabled)", false},
 		{"COMFY_URL", "http://127.0.0.1:8188", "ComfyUI upstream URL; setting it enables the ComfyUI consumer (default: empty = disabled)", false},
 		{"WARM_MODEL", "", "Optional model to reload after an image job (default: empty = none)", false},
@@ -34,7 +37,8 @@ func sampleEntries(logFile string) []sampleEntry {
 		{"COMFY_START_TIMEOUT", "2m", "How long a request waits for the managed ComfyUI to come up", false},
 		{"GAME_PROCS", "cyberpunk2077.exe,hl2.exe", "While a listed process runs, the GPU counts as held by it: requests wait, Ollama unloads, managed ComfyUI stops (default: empty = disabled)", false},
 		{"GPU_FOREIGN_VRAM_MB", "1024", "Also treat the GPU as held when a process not in GPU_IGNORE_PROCS uses more VRAM than this (needs nvidia-smi; 0/empty = disabled)", false},
-		{"GPU_IGNORE_PROCS", "ollama,ollama app,ollama_llama_server,python,pythonw", "Process names never counted as foreign GPU users (ComfyUI runs under python)", false},
+		{"GPU_FOREIGN_UTIL_PCT", "30", "Also treat the GPU as held when a process not in GPU_IGNORE_PROCS uses more than this percent of the GPU 3D engine (Windows Task-Manager counters; catches games without an exe list; 0/empty = disabled)", false},
+		{"GPU_IGNORE_PROCS", "ollama,ollama app,ollama_llama_server,python,pythonw,dwm", "Process names never counted as foreign GPU users (ComfyUI runs under python; dwm is the desktop compositor)", false},
 		{"GAME_POLL_INTERVAL", "15s", "How often game/VRAM detection runs (nvidia-smi polls keep the GPU awake; don't go below ~10s)", false},
 		{"UNLOAD_TIMEOUT", "60s", "How long to wait for Ollama to unload a model", false},
 		{"JOB_TIMEOUT", "15m", "Maximum time to wait for a ComfyUI job", false},

internal/control/control.go

@@ -7,6 +7,13 @@
 // triggers, so the worst a local user can cause is a cheap, throttled
 // check and a GPU-idle-gated restart onto a signed binary.
 //
+// Abuse hardening: the command read is capped (4 KiB), each connection is
+// force-closed after connTimeout so a stalled client cannot pin a goroutine
+// (or a Windows pipe instance) forever, and concurrently served connections
+// are capped at maxConns — beyond that, connections are closed on arrival.
+// On Windows the pipe's ACL additionally denies network logons, so the
+// channel cannot be reached from another machine.
+//
 // Protocol: the client writes one command line, the server answers with
 // one reply line ("OK ..." or "ERR ...") and hangs up.
 package control
@@ -17,12 +24,20 @@ import (
 	"fmt"
 	"io"
 	"strings"
+	"time"
 )
 
 // CmdUpdateNow asks the service to check for, stage and (once the GPU is
 // idle) restart onto a signed update immediately.
 const CmdUpdateNow = "update-now"
 
+// CmdStatus asks for a one-line JSON status snapshot (monitor mode).
+const CmdStatus = "status"
+
+// CmdReloadEnv asks the service to re-read and validate its config file,
+// and to restart onto it (once the GPU is idle) when it changed.
+const CmdReloadEnv = "reload-env"
+
 // ErrUnavailable means no running service offers the control channel.
 var ErrUnavailable = errors.New("control channel unavailable")
 
@@ -30,9 +45,50 @@ var ErrUnavailable = errors.New("control channel unavailable")
 // line. It must start with "OK " or "ERR ".
 type Handler func(cmd string) string
 
+// connTimeout bounds one connection's lifetime: a client that stops
+// mid-command or never reads the reply would otherwise pin its goroutine
+// (and on Windows one of the pipe instances) indefinitely. A var so tests
+// can shrink it.
+var connTimeout = 10 * time.Second
+
+// maxConns caps concurrently served connections; beyond it, new
+// connections are closed on arrival. Bound on the goroutines a local
+// flood can pile up.
+const maxConns = 32
+
+var connSem = make(chan struct{}, maxConns)
+
+// serve dispatches connection handling under the concurrency cap. It
+// returns false when the cap is reached — the caller must then close the
+// connection itself.
+func serve(c io.ReadWriteCloser, h Handler) bool {
+	select {
+	case connSem <- struct{}{}:
+		go func() {
+			defer func() { <-connSem }()
+			serveConn(c, h)
+		}()
+		return true
+	default:
+		return false
+	}
+}
+
+// forceCloser is implemented by connections that can be torn down
+// abortively, unblocking pending reads and writes (Windows pipe:
+// DisconnectNamedPipe; unix socket: a deadline in the past). The
+// connection watchdog uses it; normal closes still flush the reply.
+type forceCloser interface {
+	ForceClose() error
+}
+
 // serveConn runs the line protocol on one accepted connection.
 func serveConn(c io.ReadWriteCloser, h Handler) {
 	defer c.Close()
+	if fc, ok := c.(forceCloser); ok {
+		timer := time.AfterFunc(connTimeout, func() { fc.ForceClose() })
+		defer timer.Stop()
+	}
 	line, err := bufio.NewReader(io.LimitReader(c, 4096)).ReadString('\n')
 	cmd := strings.TrimSpace(line)
 	if cmd == "" {

internal/control/control_linux.go

@@ -37,12 +37,24 @@ func Serve(ctx context.Context, h Handler, log *slog.Logger) error {
 			if err != nil {
 				return // shutting down
 			}
-			go serveConn(c, h)
+			uc := unixConn{c}
+			if !serve(uc, h) {
+				uc.ForceClose()
+			}
 		}
 	}()
 	return nil
 }
 
+// unixConn adds an abortive ForceClose to net.Conn: a deadline in the
+// past fails pending and future I/O immediately.
+type unixConn struct{ net.Conn }
+
+func (c unixConn) ForceClose() error {
+	c.SetDeadline(time.Now().Add(-time.Second)) //nolint:errcheck // best effort
+	return c.Conn.Close()
+}
+
 // Ask sends one command to the running service and returns its reply.
 func Ask(cmd string) (string, error) {
 	c, err := net.DialTimeout("unix", sockPath, 2*time.Second)

internal/control/control_test.go

@@ -5,6 +5,7 @@ import (
 	"net"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestRoundTrip(t *testing.T) {
@@ -44,3 +45,59 @@ func TestEmptyReplyIsUnavailable(t *testing.T) {
 		t.Fatalf("err = %v, want ErrUnavailable", err)
 	}
 }
+
+func TestServeCap(t *testing.T) {
+	for i := 0; i < maxConns; i++ {
+		connSem <- struct{}{}
+	}
+	defer func() {
+		for i := 0; i < maxConns; i++ {
+			<-connSem
+		}
+	}()
+	server, client := net.Pipe()
+	defer server.Close()
+	defer client.Close()
+	if serve(server, func(string) string { return "OK" }) {
+		t.Fatal("serve accepted a connection beyond the cap")
+	}
+}
+
+// forcePipe records ForceClose calls for the watchdog test.
+type forcePipe struct {
+	net.Conn
+	forced chan struct{}
+}
+
+func (c forcePipe) ForceClose() error {
+	err := c.Conn.Close()
+	close(c.forced)
+	return err
+}
+
+func TestConnWatchdog(t *testing.T) {
+	old := connTimeout
+	connTimeout = 50 * time.Millisecond
+	defer func() { connTimeout = old }()
+
+	server, client := net.Pipe()
+	defer client.Close()
+	fc := forcePipe{Conn: server, forced: make(chan struct{})}
+	done := make(chan struct{})
+	go func() {
+		serveConn(fc, func(string) string { return "OK" })
+		close(done)
+	}()
+	// The client never sends anything; the watchdog must tear the
+	// connection down instead of blocking forever.
+	select {
+	case <-fc.forced:
+	case <-time.After(5 * time.Second):
+		t.Fatal("watchdog did not force-close the stalled connection")
+	}
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("serveConn still blocked after the force close")
+	}
+}

internal/control/control_windows.go

@@ -7,6 +7,8 @@ import (
 	"fmt"
 	"log/slog"
 	"os"
+	"syscall"
+	"time"
 	"unsafe"
 
 	"golang.org/x/sys/windows"
@@ -76,37 +78,90 @@ func Serve(ctx context.Context, h Handler, log *slog.Logger) error {
 				log.Warn("control channel stopped", "err", err)
 				return
 			}
-			go func() {
-				// Blocks until a client connects; on process exit the
-				// handle goes away with everything else.
-				if err := windows.ConnectNamedPipe(pipe, nil); err != nil {
-					windows.CloseHandle(pipe)
-					return
-				}
-				f := os.NewFile(uintptr(pipe), pipePath)
-				serveConn(f, h) // closes f, and with it the pipe handle
-			}()
+			// Blocks until a client connects — only then is the next
+			// instance created, so instances are not burned without
+			// clients. ERROR_PIPE_CONNECTED means the client raced us
+			// and connected before the call: that is a success. Process
+			// exit reaps the blocked call on shutdown.
+			if err := windows.ConnectNamedPipe(pipe, nil); err != nil && err != errnoPipeConnected {
+				windows.CloseHandle(pipe)
+				continue
+			}
+			conn := &pipeConn{f: os.NewFile(uintptr(pipe), pipePath), h: pipe}
+			if !serve(conn, h) {
+				conn.ForceClose()
+			}
 		}
 	}()
 	return nil
 }
 
+// errnoPipeConnected is ConnectNamedPipe's "the client connected before we
+// called" result, which means the connection is established.
+var errnoPipeConnected = syscall.Errno(535) // ERROR_PIPE_CONNECTED
+
+// errnoPipeBusy is CreateFile's "all pipe instances are busy" result. It
+// loses the race against another client that grabbed the instance
+// WaitNamedPipe just reported — the caller must wait and retry.
+var errnoPipeBusy = syscall.Errno(231) // ERROR_PIPE_BUSY
+
+// pipeConn adapts a pipe handle to io.ReadWriteCloser. Close flushes first
+// (FlushFileBuffers blocks until the client has read the reply) and then
+// disconnects — closing the bare handle right after writing can discard
+// unread reply bytes, which clients see as an empty, failed request.
+type pipeConn struct {
+	f *os.File
+	h windows.Handle
+}
+
+func (c *pipeConn) Read(p []byte) (int, error)  { return c.f.Read(p) }
+func (c *pipeConn) Write(p []byte) (int, error) { return c.f.Write(p) }
+
+func (c *pipeConn) Close() error {
+	windows.FlushFileBuffers(c.h)    //nolint:errcheck // best effort
+	windows.DisconnectNamedPipe(c.h) //nolint:errcheck // best effort
+	return c.f.Close()
+}
+
+// ForceClose aborts the connection without flushing: disconnecting
+// unblocks pending reads and writes at the cost of possibly discarding an
+// unread reply. Used by the connection watchdog; normal closes flush.
+func (c *pipeConn) ForceClose() error {
+	windows.DisconnectNamedPipe(c.h) //nolint:errcheck // best effort
+	return c.f.Close()
+}
+
 // Ask sends one command to the running service and returns its reply.
+//
+// The server keeps exactly one listening instance per connection, so
+// concurrent clients race for it: WaitNamedPipe can report an instance
+// that another client grabs before our CreateFile runs (ERROR_PIPE_BUSY).
+// Retry on that — with the monitor polling status every second, a single
+// attempt loses that race regularly.
 func Ask(cmd string) (string, error) {
 	name, err := windows.UTF16PtrFromString(pipePath)
 	if err != nil {
 		return "", err
 	}
-	if err := waitNamedPipe(name, 2000); err != nil {
-		return "", ErrUnavailable
+	deadline := time.Now().Add(5 * time.Second)
+	for {
+		if err := waitNamedPipe(name, 2000); err != nil {
+			return "", ErrUnavailable
+		}
+		handle, err := windows.CreateFile(name,
+			windows.GENERIC_READ|windows.GENERIC_WRITE, 0, nil,
+			windows.OPEN_EXISTING, 0, 0)
+		if err == errnoPipeBusy {
+			if time.Now().After(deadline) {
+				return "", ErrUnavailable
+			}
+			continue
+		}
+		if err != nil {
+			return "", ErrUnavailable
+		}
+		f := os.NewFile(uintptr(handle), pipePath)
+		defer f.Close()
+		return readReply(f, cmd)
 	}
-	handle, err := windows.CreateFile(name,
-		windows.GENERIC_READ|windows.GENERIC_WRITE, 0, nil,
-		windows.OPEN_EXISTING, 0, 0)
-	if err != nil {
-		return "", ErrUnavailable
-	}
-	f := os.NewFile(uintptr(handle), pipePath)
-	defer f.Close()
-	return readReply(f, cmd)
 }

internal/game/game.go

@@ -1,8 +1,10 @@
 // Package game detects processes outside gpu-turnstile's control that hold
 // the GPU — typically a game — so the proxy can block new GPU work and free
-// VRAM while they run. Two detection paths: an explicit process watch list
-// (GAME_PROCS) and a foreign-VRAM threshold via nvidia-smi
-// (GPU_FOREIGN_VRAM_MB) that catches anything not on the ignore list.
+// VRAM while they run. Three detection paths: an explicit process watch
+// list (GAME_PROCS), a foreign-VRAM threshold via nvidia-smi
+// (GPU_FOREIGN_VRAM_MB), and a per-process GPU 3D-engine utilization
+// threshold via Windows PDH counters (GPU_FOREIGN_UTIL_PCT). The latter two
+// catch anything not on the ignore list without naming individual games.
 package game
 
 import (
@@ -11,6 +13,7 @@ import (
 	"fmt"
 	"log/slog"
 	"os/exec"
+	"slices"
 	"strconv"
 	"strings"
 )
@@ -28,28 +31,33 @@ type computeApp struct {
 }
 
 // Detector checks whether a foreign process holds the GPU. The zero value
-// (no watch list, no threshold) never detects anything; main only starts the
-// poll loop when at least one path is configured.
+// (no watch list, no thresholds) never detects anything; main only starts
+// the poll loop when at least one path is configured.
 type Detector struct {
 	procs    map[string]bool // normalized names from GAME_PROCS
 	vramMB   int             // foreign VRAM threshold; 0 = disabled
+	utilPct  int             // foreign 3D-engine utilization threshold; 0 = disabled
 	ignore   map[string]bool // normalized names never counted as foreign
 	log      *slog.Logger
-	noNvidia bool // nvidia-smi was not found; VRAM path disabled for good
+	noNvidia bool              // nvidia-smi was not found; VRAM path disabled for good
+	sampler  *gpuEngineSampler // open PDH query, opened lazily on first Check
+	noPDH    bool              // engine counters unavailable; util path disabled for good
 }
 
 // New builds a Detector from the configured watch list, VRAM threshold in
-// MiB (0 disables the nvidia-smi path) and ignore list. Names are matched
-// case-insensitively, with or without a trailing ".exe".
-func New(procs []string, vramMB int, ignore []string, log *slog.Logger) *Detector {
+// MiB, 3D-engine utilization threshold in percent (both 0 = disabled) and
+// ignore list. Names are matched case-insensitively, with or without a
+// trailing ".exe".
+func New(procs []string, vramMB, utilPct int, ignore []string, log *slog.Logger) *Detector {
 	if log == nil {
 		log = slog.Default()
 	}
 	return &Detector{
-		procs:  nameSet(procs),
-		vramMB: vramMB,
-		ignore: nameSet(ignore),
-		log:    log,
+		procs:   nameSet(procs),
+		vramMB:  vramMB,
+		utilPct: utilPct,
+		ignore:  nameSet(ignore),
+		log:     log,
 	}
 }
 
@@ -73,34 +81,63 @@ func nameSet(names []string) map[string]bool {
 // description of each (empty when the GPU is free for gpu-turnstile's
 // consumers). A failing nvidia-smi call is returned as an error only when
 // the process list found nothing; a missing nvidia-smi binary disables the
-// VRAM path permanently (logged once).
+// VRAM path permanently (logged once), as do missing engine counters.
 func (d *Detector) Check(ctx context.Context) ([]string, error) {
 	ps, psErr := processes()
-	if d.vramMB <= 0 || d.noNvidia {
-		return d.detect(ps, nil), psErr
+	utils := d.engineUtil()
+	var apps []computeApp
+	if d.vramMB > 0 && !d.noNvidia {
+		var err error
+		apps, err = queryComputeApps(ctx)
+		if errors.Is(err, exec.ErrNotFound) {
+			d.noNvidia = true
+			d.log.Warn("GPU_FOREIGN_VRAM_MB is set but nvidia-smi was not found; VRAM detection disabled")
+		} else if err != nil {
+			return d.detect(ps, nil, utils), err
+		}
 	}
-	apps, err := queryComputeApps(ctx)
-	if errors.Is(err, exec.ErrNotFound) {
-		d.noNvidia = true
-		d.log.Warn("GPU_FOREIGN_VRAM_MB is set but nvidia-smi was not found; VRAM detection disabled")
-		return d.detect(ps, nil), nil
-	}
-	if err != nil {
-		return d.detect(ps, nil), err
-	}
-	return d.detect(ps, apps), nil
+	return d.detect(ps, apps, utils), psErr
 }
 
-// detect is the pure core of Check: given the process table and (optionally)
-// the nvidia-smi compute-apps list, it returns the foreign holders.
-func (d *Detector) detect(ps []Process, apps []computeApp) []string {
+// engineUtil samples per-process 3D-engine utilization via PDH. The first
+// call only primes the rate counters and returns nil. A failing open or
+// sample disables the path permanently (logged once).
+func (d *Detector) engineUtil() map[int]float64 {
+	if d.utilPct <= 0 || d.noPDH {
+		return nil
+	}
+	if d.sampler == nil {
+		s, err := openGPUEngineSampler()
+		if err != nil {
+			d.noPDH = true
+			d.log.Warn("GPU_FOREIGN_UTIL_PCT is set but per-process GPU counters are unavailable; engine detection disabled", "err", err)
+			return nil
+		}
+		d.sampler = s
+	}
+	utils, err := d.sampler.sample()
+	if err != nil {
+		if errors.Is(err, errNotPrimed) {
+			return nil
+		}
+		d.noPDH = true
+		d.log.Warn("per-process GPU counters failed; engine detection disabled", "err", err)
+		return nil
+	}
+	return utils
+}
+
+// detect is the pure core of Check: given the process table and
+// (optionally) the nvidia-smi compute-apps list and the PDH engine
+// utilization, it returns the foreign holders.
+func (d *Detector) detect(ps []Process, apps []computeApp, utils map[int]float64) []string {
 	var holders []string
 	for _, p := range ps {
 		if d.procs[normName(p.Name)] {
 			holders = append(holders, fmt.Sprintf("%s (pid %d)", p.Name, p.PID))
 		}
 	}
-	if d.vramMB > 0 && apps != nil {
+	if (d.vramMB > 0 && apps != nil) || (d.utilPct > 0 && utils != nil) {
 		names := make(map[int]string, len(ps))
 		for _, p := range ps {
 			names[p.PID] = p.Name
@@ -115,6 +152,23 @@ func (d *Detector) detect(ps []Process, apps []computeApp) []string {
 			}
 			holders = append(holders, fmt.Sprintf("%s (pid %d) using %d MiB VRAM", name, a.PID, a.UsedMB))
 		}
+		// Sorted for stable output (map iteration order is random).
+		pids := make([]int, 0, len(utils))
+		for pid := range utils {
+			pids = append(pids, pid)
+		}
+		slices.Sort(pids)
+		for _, pid := range pids {
+			util := utils[pid]
+			name := names[pid]
+			if d.ignore[normName(name)] || util < float64(d.utilPct) {
+				continue
+			}
+			if name == "" {
+				name = "unknown process"
+			}
+			holders = append(holders, fmt.Sprintf("%s (pid %d) using %.0f%% GPU", name, pid, util))
+		}
 	}
 	return holders
 }
@@ -132,6 +186,56 @@ func queryComputeApps(ctx context.Context) ([]computeApp, error) {
 	return parseComputeApps(string(out))
 }
 
+// GPUStats is one nvidia-smi reading of the whole card.
+type GPUStats struct {
+	UsedMB  int
+	TotalMB int
+	TempC   int // -1 when nvidia-smi reports N/A
+	FanPct  int // -1 when N/A (some cards don't expose the fan)
+}
+
+// QueryGPUStats returns VRAM usage, temperature and fan speed via
+// nvidia-smi. Unlike the per-process list this works under WDDM too.
+func QueryGPUStats(ctx context.Context) (GPUStats, error) {
+	out, err := exec.CommandContext(ctx, "nvidia-smi",
+		"--query-gpu=memory.used,memory.total,temperature.gpu,fan.speed", "--format=csv,noheader,nounits").Output()
+	if err != nil {
+		return GPUStats{}, err
+	}
+	return parseGPUStats(string(out))
+}
+
+// parseGPUStats parses one "used, total, temp, fan" CSV line (MiB, °C,
+// percent). The memory fields must be numeric; temperature and fan fall
+// back to -1 on "N/A" and friends.
+func parseGPUStats(out string) (GPUStats, error) {
+	fields := strings.Split(strings.TrimSpace(out), ",")
+	if len(fields) != 4 {
+		return GPUStats{}, fmt.Errorf("nvidia-smi: unexpected output %q", strings.TrimSpace(out))
+	}
+	num := func(s string) (int, error) {
+		return strconv.Atoi(strings.TrimSpace(s))
+	}
+	optional := func(s string) int {
+		n, err := num(s)
+		if err != nil {
+			return -1
+		}
+		return n
+	}
+	var st GPUStats
+	var err error
+	if st.UsedMB, err = num(fields[0]); err != nil {
+		return GPUStats{}, fmt.Errorf("nvidia-smi: unexpected used memory in %q", strings.TrimSpace(out))
+	}
+	if st.TotalMB, err = num(fields[1]); err != nil {
+		return GPUStats{}, fmt.Errorf("nvidia-smi: unexpected total memory in %q", strings.TrimSpace(out))
+	}
+	st.TempC = optional(fields[2])
+	st.FanPct = optional(fields[3])
+	return st, nil
+}
+
 // parseComputeApps parses "pid, used_memory" CSV lines (no header, MiB
 // units). Unsupported rows ("N/A" on WDDM) are skipped.
 func parseComputeApps(out string) ([]computeApp, error) {

internal/game/game_test.go

@@ -45,7 +45,7 @@ func TestParseComputeApps(t *testing.T) {
 }
 
 func TestDetect(t *testing.T) {
-	d := New([]string{"Cyberpunk2077.exe", "hl2"}, 1024,
+	d := New([]string{"Cyberpunk2077.exe", "hl2"}, 1024, 0,
 		[]string{"ollama", "python", "pythonw"}, nil)
 	ps := []Process{
 		{PID: 10, Name: "ollama.exe"},
@@ -58,7 +58,7 @@ func TestDetect(t *testing.T) {
 		{PID: 40, UsedMB: 2048}, // foreign, above threshold
 		{PID: 50, UsedMB: 100},  // foreign but below threshold
 	}
-	holders := d.detect(ps, apps)
+	holders := d.detect(ps, apps, nil)
 	if len(holders) != 2 {
 		t.Fatalf("got %v, want 2 holders", holders)
 	}
@@ -70,13 +70,77 @@ func TestDetect(t *testing.T) {
 	}
 }
 
+func TestDetectEngineUtil(t *testing.T) {
+	d := New(nil, 0, 30, []string{"dwm", "python"}, nil)
+	ps := []Process{
+		{PID: 10, Name: "dwm.exe"},
+		{PID: 20, Name: "game.exe"},
+		{PID: 40, Name: "browser.exe"},
+	}
+	utils := map[int]float64{
+		10: 45, // ignored: dwm
+		20: 61, // foreign, above threshold
+		30: 82, // foreign, unknown name
+		40: 5,  // below threshold
+	}
+	holders := d.detect(ps, nil, utils)
+	want := []string{
+		"game.exe (pid 20) using 61% GPU",
+		"unknown process (pid 30) using 82% GPU",
+	}
+	if !slices.Equal(holders, want) {
+		t.Errorf("got %v, want %v", holders, want)
+	}
+}
+
 func TestDetectNothingConfigured(t *testing.T) {
-	d := New(nil, 0, nil, nil)
-	if got := d.detect([]Process{{PID: 1, Name: "game.exe"}}, nil); len(got) != 0 {
+	d := New(nil, 0, 0, nil, nil)
+	if got := d.detect([]Process{{PID: 1, Name: "game.exe"}}, nil, nil); len(got) != 0 {
 		t.Errorf("got %v, want none", got)
 	}
 }
 
+func TestParseGPUEngineInstance(t *testing.T) {
+	pid, eng, ok := parseGPUEngineInstance("pid_1234_luid_0x00000000_0x00011A2B_phys_0_eng_0_engtype_3D")
+	if !ok || pid != 1234 || eng != "3D" {
+		t.Errorf("got %d, %q, %v", pid, eng, ok)
+	}
+	pid, eng, ok = parseGPUEngineInstance("pid_42_luid_0x0_0x0_phys_0_eng_1_engtype_Copy")
+	if !ok || pid != 42 || eng != "Copy" {
+		t.Errorf("got %d, %q, %v", pid, eng, ok)
+	}
+	for _, bad := range []string{"", "something", "pid_", "pid_x_luid", "pid_-1_luid_0"} {
+		if _, _, ok := parseGPUEngineInstance(bad); ok {
+			t.Errorf("%q parsed, want failure", bad)
+		}
+	}
+}
+
+func TestParseGPUStats(t *testing.T) {
+	st, err := parseGPUStats("4300, 16384, 55, 42\n")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if st.UsedMB != 4300 || st.TotalMB != 16384 || st.TempC != 55 || st.FanPct != 42 {
+		t.Errorf("got %+v", st)
+	}
+
+	// Cards that don't expose temperature/fan report N/A.
+	st, err = parseGPUStats("1024, 16384, N/A, N/A")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if st.TempC != -1 || st.FanPct != -1 {
+		t.Errorf("got %+v, want -1 for N/A fields", st)
+	}
+
+	for _, bad := range []string{"", "1, 2", "x, 16384, 55, 42", "1024, x, 55, 42", "1, 2, 3, 4, 5"} {
+		if _, err := parseGPUStats(bad); err == nil {
+			t.Errorf("%q parsed, want failure", bad)
+		}
+	}
+}
+
 func TestProcessesLive(t *testing.T) {
 	if runtime.GOOS != "windows" && runtime.GOOS != "linux" {
 		t.Skip("no process listing on this platform")

internal/game/pdh.go

@@ -0,0 +1,36 @@
+package game
+
+import (
+	"errors"
+	"strconv"
+	"strings"
+)
+
+// errNotPrimed marks the first PDH sample after opening a query: rate-based
+// counters (like engine utilization) need two collections before they
+// return meaningful values.
+var errNotPrimed = errors.New("GPU engine counter needs a second sample")
+
+// parseGPUEngineInstance splits a PDH "GPU Engine" instance name —
+// "pid_1234_luid_0x00000000_0x00011A2B_phys_0_eng_0_engtype_3D" — into PID
+// and engine type ("3D", "Copy", "VideoDecode", ...). engType is empty when
+// the name carries no engtype marker.
+func parseGPUEngineInstance(name string) (pid int, engType string, ok bool) {
+	rest, found := strings.CutPrefix(name, "pid_")
+	if !found {
+		return 0, "", false
+	}
+	digits, rest, found := strings.Cut(rest, "_")
+	if !found {
+		return 0, "", false
+	}
+	pid, err := strconv.Atoi(digits)
+	if err != nil || pid < 0 {
+		return 0, "", false
+	}
+	const marker = "engtype_"
+	if i := strings.LastIndex(rest, marker); i >= 0 {
+		engType = rest[i+len(marker):]
+	}
+	return pid, engType, true
+}

internal/game/pdh_other.go

@@ -0,0 +1,16 @@
+//go:build !windows
+
+package game
+
+import "errors"
+
+// errNoEngineCounters marks platforms without per-process GPU engine
+// counters (the PDH path is Windows-only).
+var errNoEngineCounters = errors.New("per-process GPU engine counters are only available on Windows")
+
+// gpuEngineSampler is a stub on non-Windows platforms.
+type gpuEngineSampler struct{}
+
+func openGPUEngineSampler() (*gpuEngineSampler, error) { return nil, errNoEngineCounters }
+
+func (s *gpuEngineSampler) sample() (map[int]float64, error) { return nil, errNoEngineCounters }

internal/game/pdh_windows.go

@@ -0,0 +1,111 @@
+//go:build windows
+
+package game
+
+import (
+	"fmt"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// Per-process GPU engine utilization via PDH — the same counters Task
+// Manager's "GPU engine" columns read. Unlike nvidia-smi's compute-apps
+// this covers graphics work under WDDM, so games show up. The counter is
+// added with PdhAddEnglishCounterW, which is independent of the Windows
+// display language.
+
+var (
+	pdhDLL                          = windows.NewLazySystemDLL("pdh.dll")
+	procPdhOpenQuery                = pdhDLL.NewProc("PdhOpenQueryW")
+	procPdhAddEnglishCounter        = pdhDLL.NewProc("PdhAddEnglishCounterW")
+	procPdhCollectQueryData         = pdhDLL.NewProc("PdhCollectQueryData")
+	procPdhGetFormattedCounterArray = pdhDLL.NewProc("PdhGetFormattedCounterArrayW")
+	procPdhCloseQuery               = pdhDLL.NewProc("PdhCloseQuery")
+)
+
+const (
+	pdhFmtDouble = 0x00000200 // PDH_FMT_DOUBLE
+	pdhMoreData  = 0x800007D2 // PDH_MORE_DATA
+)
+
+// pdhCountervalueItem mirrors PDH_FMT_COUNTERVALUE_ITEM (64-bit, double).
+type pdhCountervalueItem struct {
+	name    *uint16
+	cStatus uint32
+	_       uint32 // alignment padding
+	value   float64
+}
+
+// gpuEngineSampler holds an open PDH query on the wildcard GPU Engine
+// utilization counter. Keeping the query open across polls is what makes
+// the rate-based values meaningful; the sampler lives as long as the
+// process (PdhCloseQuery would only matter on unload).
+type gpuEngineSampler struct {
+	query   uintptr // PDH_HQUERY
+	counter uintptr // PDH_HCOUNTER
+	primed  bool
+}
+
+// openGPUEngineSampler opens a query on the per-process GPU engine
+// utilization counter (all instances).
+func openGPUEngineSampler() (*gpuEngineSampler, error) {
+	var q uintptr
+	if r, _, _ := procPdhOpenQuery.Call(0, 0, uintptr(unsafe.Pointer(&q))); r != 0 {
+		return nil, fmt.Errorf("PdhOpenQuery: status %#x", r)
+	}
+	path, err := windows.UTF16PtrFromString(`\GPU Engine(*)\Utilization Percentage`)
+	if err != nil {
+		procPdhCloseQuery.Call(q)
+		return nil, err
+	}
+	var c uintptr
+	if r, _, _ := procPdhAddEnglishCounter.Call(q, uintptr(unsafe.Pointer(path)), 0, uintptr(unsafe.Pointer(&c))); r != 0 {
+		procPdhCloseQuery.Call(q)
+		return nil, fmt.Errorf("PdhAddEnglishCounter: status %#x", r)
+	}
+	return &gpuEngineSampler{query: q, counter: c}, nil
+}
+
+// sample collects the counter once and returns per-PID 3D-engine
+// utilization in percent. The first call after open only primes the rate
+// calculation and returns errNotPrimed. Processes can drive several 3D
+// engines; their values are summed.
+func (s *gpuEngineSampler) sample() (map[int]float64, error) {
+	if r, _, _ := procPdhCollectQueryData.Call(s.query); r != 0 {
+		return nil, fmt.Errorf("PdhCollectQueryData: status %#x", r)
+	}
+	if !s.primed {
+		s.primed = true
+		return nil, errNotPrimed
+	}
+	var size, count uint32
+	r, _, _ := procPdhGetFormattedCounterArray.Call(s.counter, pdhFmtDouble,
+		uintptr(unsafe.Pointer(&size)), uintptr(unsafe.Pointer(&count)), 0)
+	if r == pdhMoreData && size == 0 {
+		return nil, nil // no GPU engine instances at all
+	}
+	if r != pdhMoreData {
+		return nil, fmt.Errorf("PdhGetFormattedCounterArray(size): status %#x", r)
+	}
+	buf := make([]byte, size)
+	r, _, _ = procPdhGetFormattedCounterArray.Call(s.counter, pdhFmtDouble,
+		uintptr(unsafe.Pointer(&size)), uintptr(unsafe.Pointer(&count)),
+		uintptr(unsafe.Pointer(&buf[0])))
+	if r != 0 {
+		return nil, fmt.Errorf("PdhGetFormattedCounterArray: status %#x", r)
+	}
+	items := unsafe.Slice((*pdhCountervalueItem)(unsafe.Pointer(&buf[0])), int(count))
+	out := make(map[int]float64)
+	for i := range items {
+		if items[i].cStatus != 0 || items[i].name == nil {
+			continue
+		}
+		pid, engType, ok := parseGPUEngineInstance(windows.UTF16PtrToString(items[i].name))
+		if !ok || engType != "3D" {
+			continue // only the 3D engine marks game-like work
+		}
+		out[pid] += items[i].value
+	}
+	return out, nil
+}

internal/game/pdh_windows_test.go

@@ -0,0 +1,33 @@
+//go:build windows
+
+package game
+
+import (
+	"errors"
+	"testing"
+	"time"
+)
+
+// TestGPUEngineSamplerLive opens the real PDH query and takes two samples;
+// the first only primes the rate counters. Skipped (not failed) when the
+// machine has no GPU counters.
+func TestGPUEngineSamplerLive(t *testing.T) {
+	s, err := openGPUEngineSampler()
+	if err != nil {
+		t.Skipf("no GPU engine counters: %v", err)
+	}
+	if _, err := s.sample(); !errors.Is(err, errNotPrimed) {
+		t.Fatalf("first sample: err = %v, want errNotPrimed", err)
+	}
+	time.Sleep(200 * time.Millisecond)
+	utils, err := s.sample()
+	if err != nil {
+		t.Fatalf("second sample: %v", err)
+	}
+	for pid, util := range utils {
+		if pid < 0 || util < 0 {
+			t.Errorf("pid %d: util %.2f", pid, util)
+		}
+	}
+	t.Logf("%d processes with 3D-engine usage", len(utils))
+}

internal/lock/lock.go

@@ -9,6 +9,7 @@ import (
 	"context"
 	"log/slog"
 	"sync"
+	"time"
 )
 
 // State is the current GPU occupancy state.
@@ -30,10 +31,13 @@ type Lock struct {
 	change chan struct{} // closed and replaced on every state change
 
 	n           int  // LLM requests in flight
+	llmWaiting  int  // LLM requests blocked waiting for the GPU
 	imageActive bool // an image job holds the GPU
 	imageQ      []imageWaiter
 	nextID      uint64
-	external    string // non-empty: a foreign process (e.g. a game) holds the GPU
+	external    string    // non-empty: a foreign process (e.g. a game) holds the GPU
+	detail      string    // what the current holder is doing (best effort)
+	since       time.Time // when the current state began
 
 	log *slog.Logger
 }
@@ -41,7 +45,7 @@ type Lock struct {
 // New returns a ready-to-use Lock. log may be nil; if set, every state
 // transition is logged at debug level.
 func New(log *slog.Logger) *Lock {
-	return &Lock{change: make(chan struct{}), log: log}
+	return &Lock{change: make(chan struct{}), log: log, since: time.Now()}
 }
 
 // broadcast wakes all waiters. Call with mu held.
@@ -63,6 +67,7 @@ func (l *Lock) logTransition(msg string, args ...any) {
 func (l *Lock) SetExternal(holder string) {
 	l.mu.Lock()
 	l.external = holder
+	l.since = time.Now()
 	l.broadcast()
 	l.mu.Unlock()
 	l.logTransition("lock transition", "state", StateExternal, "holder", holder)
@@ -73,6 +78,7 @@ func (l *Lock) SetExternal(holder string) {
 func (l *Lock) ClearExternal() {
 	l.mu.Lock()
 	l.external = ""
+	l.since = time.Now()
 	l.broadcast()
 	l.mu.Unlock()
 	l.logTransition("lock transition", "state", StateIdle)
@@ -90,17 +96,31 @@ func (l *Lock) External() string {
 // while waiting; no state is changed in that case.
 func (l *Lock) AcquireLLM(ctx context.Context) error {
 	l.mu.Lock()
+	waiting := false
 	for l.imageActive || len(l.imageQ) > 0 || l.external != "" {
+		if !waiting {
+			l.llmWaiting++
+			waiting = true
+		}
 		ch := l.change
 		l.mu.Unlock()
 		select {
 		case <-ctx.Done():
+			l.mu.Lock()
+			l.llmWaiting--
+			l.mu.Unlock()
 			return ctx.Err()
 		case <-ch:
 		}
 		l.mu.Lock()
 	}
+	if waiting {
+		l.llmWaiting--
+	}
 	l.n++
+	if l.n == 1 {
+		l.since = time.Now()
+	}
 	n := l.n
 	l.mu.Unlock()
 	l.logTransition("lock transition", "state", StateLLM, "llm_inflight", n)
@@ -129,6 +149,8 @@ func (l *Lock) ReleaseLLM() {
 	l.n--
 	n := l.n
 	if l.n == 0 {
+		l.since = time.Now()
+		l.detail = ""
 		l.broadcast()
 	}
 	l.mu.Unlock()
@@ -156,6 +178,7 @@ func (l *Lock) AcquireImage(ctx context.Context) error {
 		if l.imageQ[0].id == w.id && l.n == 0 && !l.imageActive && l.external == "" {
 			l.imageQ = l.imageQ[1:]
 			l.imageActive = true
+			l.since = time.Now()
 			l.mu.Unlock()
 			l.logTransition("lock transition", "state", StateImage)
 			return nil
@@ -184,6 +207,8 @@ func (l *Lock) AcquireImage(ctx context.Context) error {
 func (l *Lock) ReleaseImage() {
 	l.mu.Lock()
 	l.imageActive = false
+	l.since = time.Now()
+	l.detail = ""
 	l.broadcast()
 	l.mu.Unlock()
 	l.logTransition("lock transition", "state", StateIdle)
@@ -206,3 +231,51 @@ func (l *Lock) Snapshot() (state State, llmInflight int, imagePending bool) {
 	}
 	return state, l.n, l.imageActive || len(l.imageQ) > 0
 }
+
+// SetDetail records what the current holder is doing (e.g. the request
+// path), for status displays. Best effort: overwritten by each new holder,
+// cleared when the GPU goes idle.
+func (l *Lock) SetDetail(detail string) {
+	l.mu.Lock()
+	l.detail = detail
+	l.mu.Unlock()
+}
+
+// Status is a point-in-time view of the lock for monitoring.
+type Status struct {
+	State       State
+	Detail      string
+	LLMInflight int
+	LLMWaiting  int
+	ImageActive bool
+	ImageQueue  int
+	External    string
+	Since       time.Time
+}
+
+// Status reports the full lock state, including waiters and how long the
+// current state has held.
+func (l *Lock) Status() Status {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	s := Status{
+		Detail:      l.detail,
+		LLMInflight: l.n,
+		LLMWaiting:  l.llmWaiting,
+		ImageActive: l.imageActive,
+		ImageQueue:  len(l.imageQ),
+		External:    l.external,
+		Since:       l.since,
+	}
+	switch {
+	case l.imageActive:
+		s.State = StateImage
+	case l.n > 0:
+		s.State = StateLLM
+	case l.external != "":
+		s.State = StateExternal
+	default:
+		s.State = StateIdle
+	}
+	return s
+}

internal/ollama/client.go

@@ -60,13 +60,21 @@ func (c *Client) Probe(ctx context.Context) error {
 
 type psResponse struct {
 	Models []struct {
-		Name  string `json:"name"`
-		Model string `json:"model"`
+		Name     string `json:"name"`
+		Model    string `json:"model"`
+		SizeVRAM int64  `json:"size_vram"` // bytes resident in VRAM (0 = RAM-only)
 	} `json:"models"`
 }
 
-// LoadedModels returns the names of models currently held in memory.
-func (c *Client) LoadedModels(ctx context.Context) ([]string, error) {
+// LoadedModel is one model currently held in memory.
+type LoadedModel struct {
+	Name     string
+	SizeVRAM int64 // bytes resident in VRAM; 0 when the model sits in RAM
+}
+
+// LoadedModelDetails returns the models currently held in memory with
+// their VRAM footprint.
+func (c *Client) LoadedModelDetails(ctx context.Context) ([]LoadedModel, error) {
 	req, err := http.NewRequestWithContext(ctx, http.MethodGet, c.base+"/api/ps", nil)
 	if err != nil {
 		return nil, err
@@ -84,13 +92,26 @@ func (c *Client) LoadedModels(ctx context.Context) ([]string, error) {
 	if err := json.NewDecoder(resp.Body).Decode(&ps); err != nil {
 		return nil, err
 	}
-	models := make([]string, 0, len(ps.Models))
+	models := make([]LoadedModel, 0, len(ps.Models))
 	for _, m := range ps.Models {
-		if m.Name != "" {
-			models = append(models, m.Name)
-		} else {
-			models = append(models, m.Model)
+		name := m.Name
+		if name == "" {
+			name = m.Model
 		}
+		models = append(models, LoadedModel{Name: name, SizeVRAM: m.SizeVRAM})
+	}
+	return models, nil
+}
+
+// LoadedModels returns the names of models currently held in memory.
+func (c *Client) LoadedModels(ctx context.Context) ([]string, error) {
+	details, err := c.LoadedModelDetails(ctx)
+	if err != nil {
+		return nil, err
+	}
+	models := make([]string, 0, len(details))
+	for _, m := range details {
+		models = append(models, m.Name)
 	}
 	return models, nil
 }

internal/proxy/proxy.go

@@ -491,6 +491,7 @@ func (s *Server) OllamaHandler() http.Handler {
 			return
 		}
 		defer s.cfg.Lock.ReleaseLLM()
+		s.cfg.Lock.SetDetail("ollama: " + r.Method + " " + r.URL.Path)
 		s.ollamaProxy.ServeHTTP(w, r)
 	}))
 }
@@ -590,6 +591,7 @@ func (s *Server) handlePrompt(w http.ResponseWriter, r *http.Request) {
 	}
 	s.cfg.Metrics.ObserveLockWait("image", time.Since(start).Seconds())
 	log.Info("image lock acquired")
+	s.cfg.Lock.SetDetail("comfy: POST /prompt")
 
 	if s.cfg.ComfySup != nil && !comfyFirst {
 		if err := s.cfg.ComfySup.EnsureRunning(); err != nil {

internal/supervise/supervise.go

@@ -164,6 +164,24 @@ func (p *Process) Running() bool {
 	return p.cmd != nil
 }
 
+// Status describes the child for status displays: "external" when something
+// else serves the port, "running" once ready, "starting" while the child
+// boots, "stopped" otherwise.
+func (p *Process) Status() string {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	switch {
+	case p.external:
+		return "external"
+	case p.cmd == nil:
+		return "stopped"
+	case p.ready:
+		return "running"
+	default:
+		return "starting"
+	}
+}
+
 // Ready reports whether the server has answered a probe since its last
 // (re)start. Health checks use it to tell "starting up" from "outage".
 func (p *Process) Ready() bool {