Pin compose example to v0.1.10

.gitignore

@@ -4,3 +4,6 @@
 /compose.yml
 /signing/
 
+/gpu-turnstile.exe.old
+/gpu-turnstile.exe.new
+/gpu-turnstile.exe~

README.md

@@ -2,9 +2,10 @@
 
 GPU arbitration proxy for Ollama + ComfyUI. One consumer GPU is shared by an
 LLM server (Ollama) and an image generator (ComfyUI); gpu-turnstile sits in
-front of both and guarantees the GPU is always in exactly one of three states:
+front of both and guarantees the GPU is always in exactly one of four states:
-`idle`, `llm` (N ≥ 1 Ollama requests in flight), or `image` (exactly one
+`idle`, `llm` (N ≥ 1 Ollama requests in flight), `image` (exactly one
-ComfyUI job, Ollama models unloaded). See [SPEC.md](SPEC.md) for the full
+ComfyUI job, Ollama models unloaded), or `external` (a foreign process such
+as a game holds the GPU). See [SPEC.md](SPEC.md) for the full
 design.
 
 gpu-turnstile listens on the ports the services normally use; the actual
@@ -32,8 +33,9 @@ Open WebUI / n8n ────► :8188  ───┘
 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. Future consumers (e.g. local game
+the Ollama unload/warm steps are skipped. A third, URL-less consumer —
-detection) plug into the same lock the same way.
+detection of foreign GPU holders such as games — is enabled by `GAME_PROCS`
+and/or `GPU_FOREIGN_VRAM_MB` (see below).
 
 ## Configuration
 
@@ -45,8 +47,8 @@ override file values. Invalid values fail at startup.
 
 | Var | Default | Meaning |
 |---|---|---|
-| `LISTEN_OLLAMA` | `:11434` | Ollama-facing listener |
+| `LISTEN_OLLAMA` | `:11434` | Listener for Ollama-compatible clients |
-| `LISTEN_COMFY` | `:8188` | ComfyUI-facing listener |
+| `LISTEN_COMFY` | `:8188` | Listener for ComfyUI clients |
 | `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 |
@@ -56,12 +58,21 @@ override file values. Invalid values fail at startup.
 | `LLM_BUSY_STATUS` | `503` | HTTP status for rejected LLM requests in reject mode (400–599, e.g. 429) |
 | `BUSY_RETRY_AFTER` | `30` | Seconds sent as `Retry-After` on busy responses (both modes) |
 | `WARM_MODEL` | _(empty)_ | Model to reload after an image job (off by default) |
+| `COMFY_CMD` | _(derived from `COMFY_DIR`; both empty = unmanaged)_ | Supervise ComfyUI: start on demand, stop when idle to free VRAM. Requires `COMFY_URL` |
+| `COMFY_DIR` | _(empty)_ | Standard venv install root: set alone to supervise ComfyUI with the derived command (`.venv` + `main.py`); also the working directory for `COMFY_CMD` |
+| `COMFY_IDLE_TIMEOUT` | `5m` | Stop the managed ComfyUI after this long idle |
+| `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 |
+| `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 |
 | `LOG_FILE` | _(empty)_ | Append logs to this file instead of stderr |
 | `UNLOAD_POLL_INTERVAL` | `500ms` | `/api/ps` poll interval while unloading |
 | `HISTORY_POLL_INTERVAL` | `1s` | `/history/<id>` poll interval while a job runs |
-| `PROBE_TIMEOUT` | `5s` | Startup probe of both upstreams |
+| `PROBE_TIMEOUT` | `5s` | Startup probe of both upstreams (also per-probe health check timeout) |
+| `HEALTH_INTERVAL` | `30s` | Periodic upstream probe; down/recovered changes are logged |
 | `FREE_TIMEOUT` | `30s` | `POST /free` call after an image job |
 | `WARM_TIMEOUT` | `2m` | Warm-model reload after an image job |
 | `SHUTDOWN_TIMEOUT` | `10s` | Graceful shutdown on SIGINT/SIGTERM |
@@ -76,7 +87,7 @@ override file values. Invalid values fail at startup.
 
 ## Observability
 
-- `GET /healthz` (both listeners): `{"state":"idle|llm|image","llm_inflight":N,"image_pending":B}`
+- `GET /healthz` (both listeners): `{"state":"idle|llm|image|external","llm_inflight":N,"image_pending":B}`
 - `GET /metrics` (both listeners): Prometheus text format — `gpu_turnstile_state`,
   `gpu_turnstile_llm_inflight`, `gpu_turnstile_image_pending`,
   `gpu_turnstile_image_jobs_total`, `gpu_turnstile_lock_wait_seconds`
@@ -88,6 +99,62 @@ override file values. Invalid values fail at startup.
   class) in text mode, which renders in `docker compose logs` on Windows
   Terminal. Set `NO_COLOR` to disable colors.
 
+## Managed ComfyUI (`COMFY_CMD` / `COMFY_DIR`)
+
+Don't want ComfyUI running 24/7 (it holds VRAM even when idle — and the
+Desktop app kills its server when you close it)? gpu-turnstile can supervise
+it: the first request starts it, it stops again after `COMFY_IDLE_TIMEOUT`
+(default 5m) without work, freeing the GPU for games or the LLM.
+
+The easy way — point `COMFY_DIR` at a standard venv install (a folder with
+`.venv` and `main.py`, or `.venv` and `ComfyUI\main.py`) and the launch
+command is derived from it, including `--port` from `COMFY_URL`:
+
+```
+COMFY_URL=http://127.0.0.1:8189
+COMFY_DIR=C:\ComfyUI
+```
+
+For other layouts, spell the command out yourself (run it once manually to
+confirm it works):
+
+```
+COMFY_URL=http://127.0.0.1:8189
+COMFY_CMD="C:\ComfyUI\.venv\Scripts\python.exe" main.py --port 8189
+COMFY_DIR=C:\ComfyUI
+```
+
+`POST /prompt` waits for the server to answer before taking the GPU lock
+(LLM traffic keeps flowing while torch loads); crashes are logged and the
+next request respawns. Shutting gpu-turnstile down stops the child too.
+
+Running the ComfyUI Desktop app alongside is safe: if something already
+answers on the port, gpu-turnstile just uses it instead of spawning
+(and never kills it — it only ever stops its own child). If the managed
+instance already holds the port when you open the desktop app, the
+desktop's server is the one that fails to bind.
+
+## Game detection
+
+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
+detection paths, each optional, polled every `GAME_POLL_INTERVAL` (15s):
+
+```
+GAME_PROCS=cyberpunk2077.exe,bg3.exe     # the reliable way on Windows
+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
+game exits, requests resume automatically.
+
 ## Build and run
 
 ```sh

SPEC.md

@@ -10,11 +10,12 @@ becomes very slow; on Linux it would OOM instead.
 ## Goal
 
 A single Go binary that sits in front of **both** services and guarantees that
-at any moment the GPU is in exactly one of three states:
+at any moment the GPU is in exactly one of four states:
 
 - `idle` — nothing in flight
 - `llm` — N ≥ 1 Ollama inference requests in flight (concurrency allowed)
 - `image` — exactly one ComfyUI job in flight, Ollama models unloaded
+- `external` — a foreign process (e.g. a game) holds the GPU; new work waits
 
 Clients (LiteLLM, Open WebUI, n8n) point at gpu-turnstile instead of at the
 services. gpu-turnstile is transparent for everything that does not touch the
@@ -53,9 +54,10 @@ consumer gets no listener, no startup probe, and no lock participation:
   blocks since no image jobs can arrive.
 - **Only `COMFY_URL`**: image jobs are tracked and ComfyUI's VRAM is freed
   afterwards, but the Ollama unload and warm-reload steps are skipped.
-- Future consumers (e.g. detecting a local game holding VRAM) plug into the
+- **Game detection** is a third, optional consumer without a URL: enabled by
-  same lock the same way: enabled by their config knob, excluded when
+  `GAME_PROCS` and/or `GPU_FOREIGN_VRAM_MB` it watches for foreign processes
-  absent.
+  holding the GPU (see below) and plugs into the same lock the same way —
+  excluded when both knobs are unset.
 
 ### Lock semantics
 
@@ -75,6 +77,11 @@ are LLM requests and the single "writer" is an image job):
   requests start), waits until n == 0, sets state := `image`. Released after
   the ComfyUI job finished and models were freed.
 - Concurrent image jobs queue FIFO behind each other.
+- **External hold**: game detection calls `SetExternal(holder)` while a
+  foreign process holds the GPU. New LLM and image grants block (same busy
+  handling as above) until `ClearExternal()`; in-flight work is not
+  preempted, it drains. Snapshot reports state `external` once nothing else
+  is in flight.
 - All waits are context-aware: a client that disconnects while waiting is
   removed from the queue.
 
@@ -121,6 +128,73 @@ state is `idle`, send `POST /api/generate {"model":WARM_MODEL,"keep_alive":-1}`
 with empty prompt to reload the chat model so the next chat doesn't pay the
 load time. Off by default.
 
+### Managed ComfyUI (`COMFY_CMD` / `COMFY_DIR`)
+
+When `COMFY_CMD` is set, gpu-turnstile runs ComfyUI as a supervised child
+process instead of expecting an always-on server. Setting only `COMFY_DIR`
+enables the same management with the launch command derived from the
+standard venv layout under it (`.venv\Scripts\python.exe` on Windows,
+`.venv/bin/python` on Linux; `ComfyUI\main.py`, or a flat `main.py` when
+that is what exists; `--port` from the `COMFY_URL` port). Missing layout
+files are flagged in the startup log.
+
+- **Start on demand**: any ComfyUI request spawns it (double quotes in the
+  command line group arguments with spaces; `COMFY_DIR` sets the working
+  directory). `POST /prompt` additionally waits for readiness
+  (`/system_stats`) for up to `COMFY_START_TIMEOUT` *before* taking the GPU
+  lock, so LLM traffic flows while torch loads. Other requests are bridged
+  by the normal retry backoff. Spawn failure or a readiness timeout
+  answers 502.
+- **Idle stop**: after `COMFY_IDLE_TIMEOUT` without requests or finished
+  jobs — and only while the GPU lock is idle — the process tree is killed,
+  freeing the VRAM ComfyUI holds. The next request restarts it.
+- **Crash**: an unexpected exit is logged; the next request respawns.
+  gpu-turnstile's own shutdown stops the child too.
+- **Coexistence**: before spawning, the URL is probed — if another server
+  already answers (e.g. the ComfyUI desktop app), it is used as-is and no
+  child is spawned; the idle watcher and shutdown only ever stop the
+  supervisor's own process, never the external one. The other direction —
+  starting the desktop app while the managed instance holds the port —
+  makes the *desktop* server fail to bind; gpu-turnstile is unaffected.
+- Its stdout/stderr is forwarded to the log at INFO. The health check
+  skips the intentionally-stopped/starting states; a failed probe while
+  the process is alive and was previously ready is logged as DOWN.
+
+## 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
+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 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.
+
+While a holder is detected, gpu-turnstile:
+
+1. takes the external hold on the lock (`SetExternal`), so new LLM/image
+   requests wait or are rejected per `LLM_BUSY_MODE`; busy responses name the
+   holder,
+2. once in-flight work has drained, **frees VRAM for the foreign process**:
+   the managed ComfyUI is stopped (an external server on its port is never
+   touched) and Ollama's resident models are unloaded,
+3. refuses to spawn the managed ComfyUI; ComfyUI requests that would need a
+   spawn are answered 503 + `Retry-After` (a running desktop instance keeps
+   being proxied),
+4. logs the transitions at WARN ("GPU held by an external process …" /
+   "released the GPU; resuming") and reports state `external` in `/healthz`
+   and `/metrics`.
+
+When the holder disappears, the hold is lifted and queued requests proceed.
+
 ## Configuration (env)
 
 Configuration comes from environment variables and/or an `.env`-style
@@ -131,8 +205,8 @@ override file values. A missing file is fine; a malformed one is fatal.
 
 | Var | Default | Meaning |
 |---|---|---|
-| `LISTEN_OLLAMA` | `:11434` | Ollama-facing listener |
+| `LISTEN_OLLAMA` | `:11434` | listener for Ollama-compatible clients |
-| `LISTEN_COMFY` | `:8188` | ComfyUI-facing listener |
+| `LISTEN_COMFY` | `:8188` | listener for ComfyUI clients |
 | `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 |
@@ -142,12 +216,21 @@ override file values. A missing file is fine; a malformed one is fatal.
 | `LLM_BUSY_STATUS` | `503` | HTTP status for rejected LLM requests in reject mode (400–599, e.g. 429) |
 | `BUSY_RETRY_AFTER` | `30` | seconds sent as `Retry-After` on busy responses (both modes) |
 | `WARM_MODEL` | `` | optional model to reload after an image job |
+| `COMFY_CMD` | _(derived from `COMFY_DIR`; both empty = unmanaged)_ | spawn and supervise ComfyUI on demand: first request starts it, idle stop after `COMFY_IDLE_TIMEOUT` frees its VRAM. Requires `COMFY_URL` |
+| `COMFY_DIR` | `` | standard venv install root: set alone to supervise ComfyUI with the derived launch command (`.venv` + `main.py`, `--port` from `COMFY_URL`); also the working directory for `COMFY_CMD` |
+| `COMFY_IDLE_TIMEOUT` | `5m` | stop the managed ComfyUI after this long without requests or jobs |
+| `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 |
+| `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 |
 | `LOG_FILE` | `` | append logs to this file instead of stderr (useful as a service) |
 | `UNLOAD_POLL_INTERVAL` | `500ms` | `/api/ps` poll interval while unloading |
 | `HISTORY_POLL_INTERVAL` | `1s` | `/history/<id>` poll interval while a job runs |
-| `PROBE_TIMEOUT` | `5s` | startup probe of both upstreams |
+| `PROBE_TIMEOUT` | `5s` | startup probe of both upstreams (also the per-probe health check timeout) |
+| `HEALTH_INTERVAL` | `30s` | periodic probe of enabled upstreams; status changes (down/recovered) are logged |
 | `FREE_TIMEOUT` | `30s` | `POST /free` call after an image job |
 | `WARM_TIMEOUT` | `2m` | warm-model reload after an image job |
 | `SHUTDOWN_TIMEOUT` | `10s` | graceful shutdown on SIGINT/SIGTERM |
@@ -159,7 +242,7 @@ override file values. A missing file is fine; a malformed one is fatal.
 | `UPDATE_REPO` | `https://git.rambossek.at/PUBLIC/gpu-turnstile` | repository to check for releases |
 | `UPDATE_ASSET` | `gpu-turnstile.exe` | release asset to download |
 | `APP_VER` | `stable` | version to run: `dev` disables updates, `stable` tracks the latest release, or an exact `vX.Y.Z` pin (up- or downgraded to) |
-| `CFG_VER` | _(installer-managed)_ | config format reference written by `--install-service`; missing = the file is replaced with a fresh sample (backup `.bak`) |
+| `CFG_VER` | _(installer-managed)_ | config format reference written by `--install-service` (always a concrete `vX.Y.Z`; a dev build stamps `v0.0.0`); missing = the file is replaced with a fresh sample (backup `.bak`). New settings are appended (commented out) at install and at every startup after an update changed the version |
 
 Startup fails fast on unparsable values and when neither consumer URL is
 set. Enabled upstreams are probed once at start (`/api/version`,

cmd/gpu-turnstile/main.go

@@ -15,17 +15,20 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"syscall"
 	"time"
 
 	"gpu-turnstile/internal/comfy"
 	"gpu-turnstile/internal/config"
+	"gpu-turnstile/internal/game"
 	"gpu-turnstile/internal/lock"
 	"gpu-turnstile/internal/metrics"
 	"gpu-turnstile/internal/ollama"
 	"gpu-turnstile/internal/proxy"
 	"gpu-turnstile/internal/service"
+	"gpu-turnstile/internal/supervise"
 	"gpu-turnstile/internal/update"
 )
 
@@ -170,6 +173,7 @@ func main() {
 	}
 	log, logOut, logCloser := newLogger(cfg)
 	defer logCloser.Close()
+	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 {
@@ -186,6 +190,27 @@ func main() {
 	}
 }
 
+// syncEnvFile upgrades an installer-written config file after an update:
+// settings added since its CFG_VER are appended (commented out) and CFG_VER
+// is bumped. Files not written by the installer (no CFG_VER), up-to-date
+// files and dev builds are left untouched; a write failure is logged, not
+// fatal.
+func syncEnvFile(path, logFile string, log *slog.Logger) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return // no config file; nothing to upgrade
+	}
+	synced, changed := config.SyncSample(string(data), version, logFile)
+	if !changed {
+		return
+	}
+	if err := os.WriteFile(path, []byte(synced), 0o644); err != nil {
+		log.Warn("could not append new settings to the config file", "path", path, "err", err)
+		return
+	}
+	log.Warn("config file updated: new settings appended", "path", path, "version", version)
+}
+
 // defaultConfigPath returns gpu-turnstile.env next to the executable.
 func defaultConfigPath() string {
 	exe, err := os.Executable()
@@ -410,7 +435,26 @@ func orDisabled(url string) string {
 	return url
 }
 
+// managedComfyCommand resolves how ComfyUI is launched when it is managed:
+// COMFY_CMD verbatim, or the standard venv layout under COMFY_DIR. Empty
+// when neither is set (unmanaged).
+func managedComfyCommand(cfg config.Config) string {
+	if cfg.ComfyCmd != "" {
+		return cfg.ComfyCmd
+	}
+	if cfg.ComfyDir != "" {
+		return supervise.DefaultComfyCommand(runtime.GOOS, cfg.ComfyDir, cfg.ComfyURL)
+	}
+	return ""
+}
+
 func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Writer, isService bool) 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)
+	// so its VRAM is freed.
+	comfyCmdLine := managedComfyCommand(cfg)
+
 	// The startup line carries the version and every setting and is emitted
 	// at WARN so it is visible even with the default (quiet) log level.
 	log.Log(ctx, slog.LevelWarn, "starting gpu-turnstile",
@@ -428,6 +472,7 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		"unload_poll_interval", cfg.UnloadPollInterval,
 		"history_poll_interval", cfg.HistoryPollInterval,
 		"probe_timeout", cfg.ProbeTimeout,
+		"health_interval", cfg.HealthInterval,
 		"free_timeout", cfg.FreeTimeout,
 		"warm_timeout", cfg.WarmTimeout,
 		"shutdown_timeout", cfg.ShutdownTimeout,
@@ -435,6 +480,14 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		"backoff_max", cfg.BackoffMax,
 		"prompt_capture_limit", cfg.PromptCaptureLimit,
 		"warm_model", cfg.WarmModel,
+		"comfy_cmd", orDisabled(comfyCmdLine),
+		"comfy_dir", cfg.ComfyDir,
+		"comfy_idle_timeout", cfg.ComfyIdleTimeout,
+		"comfy_start_timeout", cfg.ComfyStartTimeout,
+		"game_procs", cfg.GameProcs,
+		"gpu_foreign_vram_mb", cfg.GPUForeignVRAMMB,
+		"gpu_ignore_procs", cfg.GPUIgnoreProcs,
+		"game_poll_interval", cfg.GamePollInterval,
 		"auto_update", cfg.AutoUpdate,
 		"update_interval", cfg.UpdateInterval,
 		"update_repo", cfg.UpdateRepo,
@@ -461,12 +514,38 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		}
 	}
 
+	var comfySup *supervise.Process
+	if comfyCmdLine != "" {
+		if cfg.ComfyCmd == "" {
+			// Derived from COMFY_DIR: flag a wrong-looking layout early,
+			// while the operator is still watching the startup log.
+			python, script := supervise.ComfyLayout(runtime.GOOS, cfg.ComfyDir)
+			for _, p := range []string{python, script} {
+				if _, err := os.Stat(p); err != nil {
+					log.Warn("COMFY_DIR: file not found; ComfyUI requests will fail until it exists", "path", p)
+				}
+			}
+		}
+		var err error
+		comfySup, err = supervise.New("comfy", comfyCmdLine, cfg.ComfyDir, comfyClient.Probe, cfg.ComfyStartTimeout, log)
+		if err != nil {
+			return err
+		}
+		defer comfySup.Stop()
+		gpuIdle := func() bool {
+			state, _, pending := lk.Snapshot()
+			return state == lock.StateIdle && !pending
+		}
+		go comfySup.WatchIdle(ctx, cfg.ComfyIdleTimeout, gpuIdle)
+	}
+
 	srv, err := proxy.New(proxy.Config{
 		OllamaURL:           cfg.OllamaURL,
 		ComfyURL:            cfg.ComfyURL,
 		Lock:                lk,
 		Ollama:              ollamaClient,
 		Comfy:               comfyClient,
+		ComfySup:            comfySup,
 		Metrics:             metrics.New(),
 		Log:                 log,
 		LogColor:            !cfg.LogJSON && cfg.LogFile == "" && os.Getenv("NO_COLOR") == "",
@@ -490,19 +569,52 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		return err
 	}
 
-	// Probe the enabled upstreams once; failure is logged, not fatal.
+	// Probe the enabled upstreams once; failure is logged, not fatal. A
-	probeCtx, probeCancel := context.WithTimeout(ctx, cfg.ProbeTimeout)
+	// managed ComfyUI is intentionally down at startup — the first request
+	// starts it — so neither the startup probe nor the health check treats
+	// that as an outage.
+	probes := map[string]func(context.Context) error{}
 	if ollamaClient != nil {
-		if err := ollamaClient.Probe(probeCtx); err != nil {
+		probes["ollama"] = ollamaClient.Probe
-			log.Warn("ollama probe failed", "url", cfg.OllamaURL, "err", err)
-		}
 	}
 	if comfyClient != nil {
-		if err := comfyClient.Probe(probeCtx); err != nil {
+		if comfySup == nil {
-			log.Warn("comfy probe failed", "url", cfg.ComfyURL, "err", err)
+			probes["comfy"] = comfyClient.Probe
+		} else {
+			// Managed upstream: an idle-stopped or still-starting server is
+			// not an outage, so it is skipped until it has answered once
+			// (Ready resets on every spawn/stop). After that, a failed
+			// probe while the process lives is a real "DOWN".
+			probes["comfy"] = func(ctx context.Context) error {
+				if !comfySup.Ready() {
+					if comfySup.Running() {
+						if err := comfyClient.Probe(ctx); err == nil {
+							comfySup.MarkReady()
+						}
+					}
+					return errManagedDown
+				}
+				return comfyClient.Probe(ctx)
+			}
+		}
+	}
+	probeCtx, probeCancel := context.WithTimeout(ctx, cfg.ProbeTimeout)
+	for name, probe := range probes {
+		if err := probe(probeCtx); 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)
+	}
+
+	// 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)
+	}
 
 	// Bind the listeners up front so a port conflict fails fast and the
 	// readiness notification below really means "accepting connections".
@@ -560,6 +672,128 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 	return nil
 }
 
+// errManagedDown marks a managed upstream that is intentionally stopped
+// (idle); health checks skip it instead of logging an outage.
+var errManagedDown = errors.New("managed upstream intentionally stopped")
+
+// 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) {
+	ticker := time.NewTicker(cfg.GamePollInterval)
+	defer ticker.Stop()
+	held, freed := false, false
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+		holders, err := det.Check(ctx)
+		if err != nil && ctx.Err() == nil {
+			log.Warn("game detection failed", "err", err)
+		}
+		switch {
+		case len(holders) > 0 && !held:
+			held = true
+			lk.SetExternal(summarizeHolders(holders))
+			log.Warn("GPU held by an external process; new LLM/image requests wait",
+				"holders", summarizeHolders(holders))
+		case len(holders) == 0 && held:
+			held, freed = false, false
+			lk.ClearExternal()
+			log.Warn("external process released the GPU; resuming")
+		}
+		if held && !freed {
+			if state, _, _ := lk.Snapshot(); state != lock.StateLLM && state != lock.StateImage {
+				freed = true
+				freeVRAM(ctx, cfg.UnloadTimeout, log, ollamaClient, comfySup)
+			}
+		}
+	}
+}
+
+// summarizeHolders joins holder descriptions for logs and busy responses,
+// capping the list so a process name matching dozens of PIDs (system
+// services) does not flood the log.
+func summarizeHolders(holders []string) string {
+	const max = 5
+	if len(holders) > max {
+		return strings.Join(holders[:max], "; ") + fmt.Sprintf("; +%d more", len(holders)-max)
+	}
+	return strings.Join(holders, "; ")
+}
+
+// freeVRAM stops the managed ComfyUI (never an external server on its port)
+// and unloads Ollama's resident models so the foreign process gets the GPU
+// memory.
+func freeVRAM(ctx context.Context, unloadTimeout time.Duration, log *slog.Logger, ollamaClient *ollama.Client, comfySup *supervise.Process) {
+	if comfySup != nil && comfySup.Running() {
+		log.Warn("stopping the managed ComfyUI to free VRAM")
+		comfySup.Stop()
+	}
+	if ollamaClient == nil {
+		return
+	}
+	uctx, cancel := context.WithTimeout(ctx, unloadTimeout)
+	defer cancel()
+	if models, err := ollamaClient.LoadedModels(uctx); err != nil || len(models) == 0 {
+		return // nothing resident (or ollama unreachable); nothing to free
+	}
+	elapsed, err := ollamaClient.UnloadAll(uctx)
+	if err != nil {
+		log.Warn("ollama unload incomplete; continuing", "err", err)
+		return
+	}
+	log.Warn("ollama models unloaded to free VRAM", "seconds", elapsed.Seconds())
+}
+
+// healthLoop probes the enabled upstreams every interval and logs status
+// 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) {
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	up := map[string]bool{}
+	managed := map[string]bool{}
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+		for name, probe := range probes {
+			pctx, cancel := context.WithTimeout(ctx, probeTimeout)
+			err := probe(pctx)
+			cancel()
+			if errors.Is(err, errManagedDown) {
+				managed[name] = true // intentionally stopped; not an outage
+				continue
+			}
+			now := err == nil
+			if managed[name] {
+				// First real probe after an idle stop only re-baselines —
+				// an on-demand start is not a "recovery".
+				managed[name] = false
+				up[name] = now
+				continue
+			}
+			was, seen := up[name]
+			if seen && now != was {
+				if now {
+					log.Warn(name + " upstream recovered")
+				} else {
+					log.Warn(name+" upstream is DOWN", "err", err)
+				}
+			}
+			up[name] = now
+		}
+	}
+}
+
 // updateLoop checks for signed updates on startup and every UPDATE_INTERVAL.
 // In service mode a staged update is applied by exiting with exitCodeUpdate
 // once the GPU lock is idle; the service recovery configuration restarts the

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:v0.1.7
+    image: git.rambossek.at/public/gpu-turnstile:v0.1.10
     restart: unless-stopped
     environment:
       # Each consumer is enabled by setting its URL; leave one unset to

internal/config/config.go

@@ -26,6 +26,7 @@ type Config struct {
 	UnloadPollInterval  time.Duration
 	HistoryPollInterval time.Duration
 	ProbeTimeout        time.Duration
+	HealthInterval      time.Duration
 	FreeTimeout         time.Duration
 	WarmTimeout         time.Duration
 	ShutdownTimeout     time.Duration
@@ -50,6 +51,30 @@ type Config struct {
 	LLMBusyStatus  int
 	BusyRetryAfter int
 
+	// ComfyCmd spawns and supervises a ComfyUI server on demand. When
+	// ComfyCmd is empty but ComfyDir is set, management is enabled with the
+	// standard venv layout under ComfyDir (.venv + main.py or
+	// ComfyUI/main.py; --port from the COMFY_URL port) — ComfyCmd is the
+	// override for other layouts and doubles as the working directory when
+	// 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
+
+	// 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
+
 	WarmModel string
 	LogLevel  slog.Level
 	LogJSON   bool
@@ -70,6 +95,7 @@ func Defaults() Config {
 		UnloadPollInterval:  500 * time.Millisecond,
 		HistoryPollInterval: time.Second,
 		ProbeTimeout:        5 * time.Second,
+		HealthInterval:      30 * time.Second,
 		FreeTimeout:         30 * time.Second,
 		WarmTimeout:         2 * time.Minute,
 		ShutdownTimeout:     10 * time.Second,
@@ -87,6 +113,14 @@ func Defaults() Config {
 		LLMBusyStatus:  503,
 		BusyRetryAfter: 30,
 
+		ComfyIdleTimeout:  5 * time.Minute,
+		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"},
+		GamePollInterval: 15 * time.Second,
+
 		LogLevel: slog.LevelWarn,
 	}
 }
@@ -122,6 +156,17 @@ func ParseEnvFile(r io.Reader) (map[string]string, error) {
 	return values, scanner.Err()
 }
 
+// splitList parses a comma-separated setting into trimmed, non-empty items.
+func splitList(v string) []string {
+	var out []string
+	for _, item := range strings.Split(v, ",") {
+		if item = strings.TrimSpace(item); item != "" {
+			out = append(out, item)
+		}
+	}
+	return out
+}
+
 func envDuration(getenv func(string) string, name string, dst *time.Duration) error {
 	v := getenv(name)
 	if v == "" {
@@ -148,6 +193,8 @@ func Load(getenv func(string) string) (Config, error) {
 		{"OLLAMA_URL", &cfg.OllamaURL},
 		{"COMFY_URL", &cfg.ComfyURL},
 		{"WARM_MODEL", &cfg.WarmModel},
+		{"COMFY_CMD", &cfg.ComfyCmd},
+		{"COMFY_DIR", &cfg.ComfyDir},
 		{"UPDATE_REPO", &cfg.UpdateRepo},
 		{"UPDATE_ASSET", &cfg.UpdateAsset},
 		{"LOG_FILE", &cfg.LogFile},
@@ -166,17 +213,34 @@ func Load(getenv func(string) string) (Config, error) {
 		{"UNLOAD_POLL_INTERVAL", &cfg.UnloadPollInterval},
 		{"HISTORY_POLL_INTERVAL", &cfg.HistoryPollInterval},
 		{"PROBE_TIMEOUT", &cfg.ProbeTimeout},
+		{"HEALTH_INTERVAL", &cfg.HealthInterval},
 		{"FREE_TIMEOUT", &cfg.FreeTimeout},
 		{"WARM_TIMEOUT", &cfg.WarmTimeout},
 		{"SHUTDOWN_TIMEOUT", &cfg.ShutdownTimeout},
 		{"BACKOFF_INITIAL", &cfg.BackoffInitial},
 		{"BACKOFF_MAX", &cfg.BackoffMax},
+		{"COMFY_IDLE_TIMEOUT", &cfg.ComfyIdleTimeout},
+		{"COMFY_START_TIMEOUT", &cfg.ComfyStartTimeout},
 		{"UPDATE_INTERVAL", &cfg.UpdateInterval},
+		{"GAME_POLL_INTERVAL", &cfg.GamePollInterval},
 	} {
 		if err := envDuration(getenv, e.name, e.dst); err != nil {
 			return cfg, err
 		}
 	}
+	if v := getenv("GAME_PROCS"); v != "" {
+		cfg.GameProcs = splitList(v)
+	}
+	if v := getenv("GPU_IGNORE_PROCS"); v != "" {
+		cfg.GPUIgnoreProcs = splitList(v)
+	}
+	if v := getenv("GPU_FOREIGN_VRAM_MB"); v != "" {
+		n, err := strconv.Atoi(v)
+		if err != nil || n < 0 {
+			return cfg, fmt.Errorf("GPU_FOREIGN_VRAM_MB: must be a non-negative integer (MiB, 0 = disabled)")
+		}
+		cfg.GPUForeignVRAMMB = n
+	}
 	if v := getenv("PROMPT_CAPTURE_LIMIT"); v != "" {
 		n, err := strconv.ParseInt(v, 10, 64)
 		if err != nil || n < 0 {
@@ -241,6 +305,12 @@ func Load(getenv func(string) string) (Config, error) {
 	default:
 		return cfg, fmt.Errorf("LOG_FORMAT: must be \"text\" or \"json\"")
 	}
+	if cfg.ComfyCmd != "" && cfg.ComfyURL == "" {
+		return cfg, fmt.Errorf("COMFY_CMD requires COMFY_URL to be set (the proxy needs somewhere to forward)")
+	}
+	if cfg.ComfyCmd == "" && cfg.ComfyDir != "" && cfg.ComfyURL == "" {
+		return cfg, fmt.Errorf("COMFY_DIR without COMFY_CMD requires COMFY_URL to be set (it enables the managed ComfyUI)")
+	}
 	if cfg.OllamaURL == "" && cfg.ComfyURL == "" {
 		return cfg, ErrNoConsumer
 	}

internal/config/config_test.go

@@ -74,6 +74,29 @@ func TestAppVersion(t *testing.T) {
 	}
 }
 
+func TestComfyCmdRequiresURL(t *testing.T) {
+	_, err := Load(func(k string) string {
+		if k == "COMFY_CMD" {
+			return "python main.py"
+		}
+		return ""
+	})
+	if err == nil || !strings.Contains(err.Error(), "COMFY_CMD requires COMFY_URL") {
+		t.Fatalf("err = %v, want COMFY_CMD/COMFY_URL validation error", err)
+	}
+	if _, err := Load(func(k string) string {
+		switch k {
+		case "COMFY_CMD":
+			return "python main.py"
+		case "COMFY_URL":
+			return "http://127.0.0.1:8188"
+		}
+		return ""
+	}); err != nil {
+		t.Fatalf("COMFY_CMD with COMFY_URL must load: %v", err)
+	}
+}
+
 func TestParseEnvFile(t *testing.T) {
 	input := `# comment
 OLLAMA_URL=http://host:11435
@@ -134,6 +157,8 @@ func TestLoadErrors(t *testing.T) {
 		{"LLM_BUSY_MODE", "bogus"},
 		{"LLM_BUSY_STATUS", "200"},
 		{"BUSY_RETRY_AFTER", "0"},
+		{"GPU_FOREIGN_VRAM_MB", "-1"},
+		{"GAME_POLL_INTERVAL", "bogus"},
 	} {
 		_, err := Load(func(k string) string {
 			if k == tc.key {
@@ -149,3 +174,76 @@ func TestLoadErrors(t *testing.T) {
 		}
 	}
 }
+
+func TestGameDetectionSettings(t *testing.T) {
+	cfg, err := Load(func(k string) string {
+		switch k {
+		case "OLLAMA_URL":
+			return "http://127.0.0.1:11435"
+		case "GAME_PROCS":
+			return " cyberpunk2077.exe, hl2.exe ,, "
+		case "GPU_FOREIGN_VRAM_MB":
+			return "1024"
+		case "GPU_IGNORE_PROCS":
+			return "ollama, my-trainer"
+		}
+		return ""
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(cfg.GameProcs) != 2 || cfg.GameProcs[0] != "cyberpunk2077.exe" || cfg.GameProcs[1] != "hl2.exe" {
+		t.Fatalf("GameProcs = %v", cfg.GameProcs)
+	}
+	if cfg.GPUForeignVRAMMB != 1024 {
+		t.Fatalf("GPUForeignVRAMMB = %d", cfg.GPUForeignVRAMMB)
+	}
+	if len(cfg.GPUIgnoreProcs) != 2 || cfg.GPUIgnoreProcs[1] != "my-trainer" {
+		t.Fatalf("GPUIgnoreProcs = %v", cfg.GPUIgnoreProcs)
+	}
+	if cfg.GamePollInterval != 15*time.Second {
+		t.Fatalf("GamePollInterval = %v, want 15s default", cfg.GamePollInterval)
+	}
+
+	// Defaults: both detection paths off, ignore list covers our consumers.
+	cfg, err = Load(func(k string) string {
+		if k == "OLLAMA_URL" {
+			return "http://127.0.0.1:11435"
+		}
+		return ""
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(cfg.GameProcs) != 0 || cfg.GPUForeignVRAMMB != 0 {
+		t.Fatalf("detection must be off by default: %v %d", cfg.GameProcs, cfg.GPUForeignVRAMMB)
+	}
+	if len(cfg.GPUIgnoreProcs) == 0 {
+		t.Fatal("GPUIgnoreProcs default must not be empty")
+	}
+}
+
+func TestComfyDirOnlyEnablesManaged(t *testing.T) {
+	// COMFY_DIR without COMFY_CMD and without COMFY_URL is a mistake.
+	_, err := Load(func(k string) string {
+		if k == "COMFY_DIR" {
+			return `C:\ComfyUI`
+		}
+		return ""
+	})
+	if err == nil || !strings.Contains(err.Error(), "COMFY_DIR") {
+		t.Fatalf("err = %v, want COMFY_DIR/COMFY_URL validation error", err)
+	}
+	// With COMFY_URL it loads — the launch command is derived from the dir.
+	if _, err := Load(func(k string) string {
+		switch k {
+		case "COMFY_DIR":
+			return `C:\ComfyUI`
+		case "COMFY_URL":
+			return "http://127.0.0.1:8189"
+		}
+		return ""
+	}); err != nil {
+		t.Fatalf("COMFY_DIR with COMFY_URL must load: %v", err)
+	}
+}

internal/config/sample.go

@@ -23,11 +23,19 @@ const appVerComment = `Version to run: "dev" disables updates, "stable" tracks t
 // CFG_VER and APP_VER are not entries — they head the file, always active.
 func sampleEntries(logFile string) []sampleEntry {
 	return []sampleEntry{
-		{"LISTEN_OLLAMA", ":11434", "Ollama-facing listener address", false},
+		{"LISTEN_OLLAMA", ":11434", "Listen address for Ollama-compatible clients (gpu-turnstile poses as Ollama here)", false},
-		{"LISTEN_COMFY", ":8188", "ComfyUI-facing listener address", false},
+		{"LISTEN_COMFY", ":8188", "Listen address for ComfyUI clients (gpu-turnstile poses as ComfyUI here)", 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},
+		{"COMFY_CMD", `"C:\ComfyUI\.venv\Scripts\python.exe" ComfyUI\main.py --port 8188`, "Spawn and supervise ComfyUI on demand: the first request starts it, it stops after COMFY_IDLE_TIMEOUT to free VRAM (default: derived from COMFY_DIR; both empty = unmanaged)", false},
+		{"COMFY_DIR", `C:\ComfyUI`, "Root of a standard ComfyUI venv install (.venv + main.py): setting it alone supervises ComfyUI with the derived launch command; also the working directory for COMFY_CMD", false},
+		{"COMFY_IDLE_TIMEOUT", "5m", "Stop the managed ComfyUI after this long without requests or jobs (frees VRAM)", false},
+		{"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},
+		{"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},
 		{"LLM_WAIT_TIMEOUT", "10m", "Max time an LLM request waits for the GPU before being answered 503 (wait mode)", false},
@@ -39,7 +47,8 @@ func sampleEntries(logFile string) []sampleEntry {
 		{"LOG_FILE", logFile, "Append logs to this file instead of stderr (a Windows service has no console)", logFile != ""},
 		{"UNLOAD_POLL_INTERVAL", "500ms", "/api/ps poll interval while unloading", false},
 		{"HISTORY_POLL_INTERVAL", "1s", "/history/<id> poll interval while a job runs", false},
-		{"PROBE_TIMEOUT", "5s", "Startup probe timeout for the enabled upstreams", false},
+		{"PROBE_TIMEOUT", "5s", "Probe timeout for the startup probe and the periodic health check", false},
+		{"HEALTH_INTERVAL", "30s", "How often enabled upstreams are probed; status changes are logged", false},
 		{"FREE_TIMEOUT", "30s", "Timeout for the POST /free call after an image job", false},
 		{"WARM_TIMEOUT", "2m", "Timeout for the warm-model reload after an image job", false},
 		{"SHUTDOWN_TIMEOUT", "10s", "Graceful shutdown timeout on SIGINT/SIGTERM", false},
@@ -57,10 +66,18 @@ func sampleEntries(logFile string) []sampleEntry {
 // comment line. Everything is commented out — so all defaults apply —
 // except the CFG_VER/APP_VER header and LOG_FILE when logFile is non-empty
 // (a Windows service has no console). CFG_VER records the version that
-// wrote the file so later installs can upgrade it.
+// wrote the file so installs — and startups after an update — can upgrade
+// it.
 func SampleEnv(version, logFile string) string {
+	// CFG_VER is always a concrete vX.Y.Z — never "dev". A dev build
+	// stamps v0.0.0, which sorts older than any release, so the next
+	// release install upgrades the file and stamps a proper version.
+	cfgVer := version
+	if cfgVer == "" || cfgVer == "dev" {
+		cfgVer = "v0.0.0"
+	}
 	var b strings.Builder
-	fmt.Fprintf(&b, "CFG_VER=%s\n", version)
+	fmt.Fprintf(&b, "CFG_VER=%s\n", cfgVer)
 	b.WriteString("# Config format reference, written by the installer — do not edit.\n")
 	b.WriteString("# The installer uses it to append newly added settings on updates.\n\n")
 	b.WriteString("# gpu-turnstile configuration\n")

internal/config/sample_test.go

@@ -122,6 +122,19 @@ func TestSyncSampleAppendsMissingAppVer(t *testing.T) {
 	}
 }
 
+func TestSampleEnvDevStampsConcreteVersion(t *testing.T) {
+	// A dev build must never write CFG_VER=dev: it stamps v0.0.0, and the
+	// next release install upgrades the file to a proper version.
+	dev := SampleEnv("dev", sampleLogPath)
+	if !strings.HasPrefix(dev, "CFG_VER=v0.0.0\n") {
+		t.Errorf("dev sample first line: %q", strings.SplitN(dev, "\n", 2)[0])
+	}
+	out, changed := SyncSample(dev, "v0.1.7", sampleLogPath)
+	if !changed || !strings.HasPrefix(out, "CFG_VER=v0.1.7\n") {
+		t.Errorf("dev-stamped file was not upgraded to v0.1.7 (changed=%v)", changed)
+	}
+}
+
 func TestCompareVersions(t *testing.T) {
 	cases := []struct {
 		a, b string

internal/game/game.go

@@ -0,0 +1,159 @@
+// 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.
+package game
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+// Process is one running OS process.
+type Process struct {
+	PID  int
+	Name string
+}
+
+// computeApp is one process holding GPU memory, as reported by nvidia-smi.
+type computeApp struct {
+	PID    int
+	UsedMB int
+}
+
+// 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.
+type Detector struct {
+	procs    map[string]bool // normalized names from GAME_PROCS
+	vramMB   int             // foreign VRAM 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
+}
+
+// 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 {
+	if log == nil {
+		log = slog.Default()
+	}
+	return &Detector{
+		procs:  nameSet(procs),
+		vramMB: vramMB,
+		ignore: nameSet(ignore),
+		log:    log,
+	}
+}
+
+// normName lowercases a process name and strips a trailing ".exe" so the
+// watch and ignore lists match on Windows and Linux spellings alike.
+func normName(s string) string {
+	return strings.TrimSuffix(strings.ToLower(strings.TrimSpace(s)), ".exe")
+}
+
+func nameSet(names []string) map[string]bool {
+	set := make(map[string]bool, len(names))
+	for _, n := range names {
+		if n = normName(n); n != "" {
+			set[n] = true
+		}
+	}
+	return set
+}
+
+// Check looks once for foreign GPU holders and returns a human-readable
+// 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).
+func (d *Detector) Check(ctx context.Context) ([]string, error) {
+	ps, psErr := processes()
+	if d.vramMB <= 0 || d.noNvidia {
+		return d.detect(ps, nil), psErr
+	}
+	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
+}
+
+// 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 {
+	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 {
+		names := make(map[int]string, len(ps))
+		for _, p := range ps {
+			names[p.PID] = p.Name
+		}
+		for _, a := range apps {
+			name := names[a.PID]
+			if d.ignore[normName(name)] || a.UsedMB < d.vramMB {
+				continue
+			}
+			if name == "" {
+				name = "unknown process"
+			}
+			holders = append(holders, fmt.Sprintf("%s (pid %d) using %d MiB VRAM", name, a.PID, a.UsedMB))
+		}
+	}
+	return holders
+}
+
+// queryComputeApps runs nvidia-smi and parses the per-process VRAM list.
+// Note: under Windows' WDDM driver, nvidia-smi only sees compute
+// allocations, so graphics-only games may not appear there — GAME_PROCS is
+// the reliable path on Windows; on Linux both work.
+func queryComputeApps(ctx context.Context) ([]computeApp, error) {
+	out, err := exec.CommandContext(ctx, "nvidia-smi",
+		"--query-compute-apps=pid,used_memory", "--format=csv,noheader,nounits").Output()
+	if err != nil {
+		return nil, err
+	}
+	return parseComputeApps(string(out))
+}
+
+// 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) {
+	var apps []computeApp
+	for _, line := range strings.Split(out, "\n") {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		pidStr, memStr, ok := strings.Cut(line, ",")
+		if !ok {
+			return nil, fmt.Errorf("nvidia-smi: unexpected line %q", line)
+		}
+		pid, err := strconv.Atoi(strings.TrimSpace(pidStr))
+		if err != nil {
+			return nil, fmt.Errorf("nvidia-smi: unexpected pid in %q", line)
+		}
+		mem, err := strconv.Atoi(strings.TrimSpace(memStr))
+		if err != nil {
+			continue // "N/A" and friends: unsupported under WDDM
+		}
+		apps = append(apps, computeApp{PID: pid, UsedMB: mem})
+	}
+	return apps, nil
+}

internal/game/game_test.go

@@ -0,0 +1,96 @@
+package game
+
+import (
+	"runtime"
+	"slices"
+	"testing"
+)
+
+func TestNormName(t *testing.T) {
+	for in, want := range map[string]string{
+		"Cyberpunk2077.exe": "cyberpunk2077",
+		"ollama":            "ollama",
+		"OLLAMA APP.EXE":    "ollama app",
+		" python.exe ":      "python",
+		"hl2":               "hl2",
+	} {
+		if got := normName(in); got != want {
+			t.Errorf("normName(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestParseComputeApps(t *testing.T) {
+	apps, err := parseComputeApps("1234, 512\n  42 , 8192 \n\n")
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := []computeApp{{PID: 1234, UsedMB: 512}, {PID: 42, UsedMB: 8192}}
+	if !slices.Equal(apps, want) {
+		t.Errorf("got %+v, want %+v", apps, want)
+	}
+
+	// WDDM reports "N/A" for memory; those rows are skipped, not fatal.
+	apps, err = parseComputeApps("1234, N/A\n42, 1024\n")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !slices.Equal(apps, []computeApp{{PID: 42, UsedMB: 1024}}) {
+		t.Errorf("got %+v", apps)
+	}
+
+	if _, err := parseComputeApps("garbage\n"); err == nil {
+		t.Error("expected an error for a malformed line")
+	}
+}
+
+func TestDetect(t *testing.T) {
+	d := New([]string{"Cyberpunk2077.exe", "hl2"}, 1024,
+		[]string{"ollama", "python", "pythonw"}, nil)
+	ps := []Process{
+		{PID: 10, Name: "ollama.exe"},
+		{PID: 20, Name: "python.exe"},
+		{PID: 30, Name: "cyberpunk2077.exe"},
+	}
+	apps := []computeApp{
+		{PID: 10, UsedMB: 8192}, // ignored: ollama
+		{PID: 20, UsedMB: 4096}, // ignored: python (ComfyUI)
+		{PID: 40, UsedMB: 2048}, // foreign, above threshold
+		{PID: 50, UsedMB: 100},  // foreign but below threshold
+	}
+	holders := d.detect(ps, apps)
+	if len(holders) != 2 {
+		t.Fatalf("got %v, want 2 holders", holders)
+	}
+	if holders[0] != "cyberpunk2077.exe (pid 30)" {
+		t.Errorf("holders[0] = %q", holders[0])
+	}
+	if holders[1] != "unknown process (pid 40) using 2048 MiB VRAM" {
+		t.Errorf("holders[1] = %q", holders[1])
+	}
+}
+
+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 {
+		t.Errorf("got %v, want none", got)
+	}
+}
+
+func TestProcessesLive(t *testing.T) {
+	if runtime.GOOS != "windows" && runtime.GOOS != "linux" {
+		t.Skip("no process listing on this platform")
+	}
+	ps, err := processes()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(ps) == 0 {
+		t.Fatal("no processes listed")
+	}
+	for _, p := range ps {
+		if p.Name == "" {
+			t.Errorf("pid %d has an empty name", p.PID)
+		}
+	}
+}

internal/game/procs_linux.go

@@ -0,0 +1,30 @@
+//go:build linux
+
+package game
+
+import (
+	"os"
+	"strconv"
+	"strings"
+)
+
+// processes lists the running processes from /proc/<pid>/comm.
+func processes() ([]Process, error) {
+	entries, err := os.ReadDir("/proc")
+	if err != nil {
+		return nil, err
+	}
+	var ps []Process
+	for _, e := range entries {
+		pid, err := strconv.Atoi(e.Name())
+		if err != nil {
+			continue
+		}
+		comm, err := os.ReadFile("/proc/" + e.Name() + "/comm")
+		if err != nil {
+			continue // process vanished mid-walk
+		}
+		ps = append(ps, Process{PID: pid, Name: strings.TrimSpace(string(comm))})
+	}
+	return ps, nil
+}

internal/game/procs_other.go

@@ -0,0 +1,7 @@
+//go:build !windows && !linux
+
+package game
+
+// processes is unsupported on this platform; the process watch list never
+// matches and the nvidia-smi path reports PIDs without names.
+func processes() ([]Process, error) { return nil, nil }

internal/game/procs_windows.go

@@ -0,0 +1,35 @@
+//go:build windows
+
+package game
+
+import (
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// processes lists the running processes via the Toolhelp32 snapshot API.
+func processes() ([]Process, error) {
+	h, err := windows.CreateToolhelp32Snapshot(windows.TH32CS_SNAPPROCESS, 0)
+	if err != nil {
+		return nil, err
+	}
+	defer windows.CloseHandle(h) //nolint:errcheck // best effort
+
+	var entry windows.ProcessEntry32
+	entry.Size = uint32(unsafe.Sizeof(entry))
+	if err := windows.Process32First(h, &entry); err != nil {
+		return nil, err
+	}
+	var ps []Process
+	for {
+		ps = append(ps, Process{
+			PID:  int(entry.ProcessID),
+			Name: windows.UTF16ToString(entry.ExeFile[:]),
+		})
+		if err := windows.Process32Next(h, &entry); err != nil {
+			break // ERROR_NO_MORE_FILES ends the walk
+		}
+	}
+	return ps, nil
+}

internal/lock/lock.go

@@ -1,6 +1,8 @@
 // Package lock implements the two-mode GPU arbitration lock: any number of
 // concurrent LLM requests ("readers") or exactly one image job ("writer"),
-// with image jobs taking priority over newly arriving LLM requests.
+// with image jobs taking priority over newly arriving LLM requests. An
+// external hold (SetExternal) blocks new grants of both kinds while a
+// foreign process — e.g. a game — holds the GPU; in-flight work drains.
 package lock
 
 import (
@@ -13,9 +15,10 @@ import (
 type State string
 
 const (
-	StateIdle  State = "idle"
+	StateIdle     State = "idle"
-	StateLLM   State = "llm"
+	StateLLM      State = "llm"
-	StateImage State = "image"
+	StateImage    State = "image"
+	StateExternal State = "external"
 )
 
 type imageWaiter struct{ id uint64 }
@@ -30,6 +33,7 @@ type Lock struct {
 	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
 
 	log *slog.Logger
 }
@@ -52,12 +56,41 @@ func (l *Lock) logTransition(msg string, args ...any) {
 	}
 }
 
+// SetExternal records that a process outside gpu-turnstile's control (a
+// game, another ML job) holds the GPU: new LLM and image grants block until
+// ClearExternal. In-flight work is not preempted. holder describes the
+// process for logs and busy responses.
+func (l *Lock) SetExternal(holder string) {
+	l.mu.Lock()
+	l.external = holder
+	l.broadcast()
+	l.mu.Unlock()
+	l.logTransition("lock transition", "state", StateExternal, "holder", holder)
+}
+
+// ClearExternal lifts the external hold; waiting LLM and image requests
+// proceed.
+func (l *Lock) ClearExternal() {
+	l.mu.Lock()
+	l.external = ""
+	l.broadcast()
+	l.mu.Unlock()
+	l.logTransition("lock transition", "state", StateIdle)
+}
+
+// External returns the current external holder, or "" when none.
+func (l *Lock) External() string {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	return l.external
+}
+
 // AcquireLLM blocks until no image job is active or pending, then registers
 // one in-flight LLM request. Returns ctx.Err() if the context is cancelled
 // while waiting; no state is changed in that case.
 func (l *Lock) AcquireLLM(ctx context.Context) error {
 	l.mu.Lock()
-	for l.imageActive || len(l.imageQ) > 0 {
+	for l.imageActive || len(l.imageQ) > 0 || l.external != "" {
 		ch := l.change
 		l.mu.Unlock()
 		select {
@@ -76,10 +109,10 @@ func (l *Lock) AcquireLLM(ctx context.Context) error {
 
 // TryAcquireLLM acquires one in-flight LLM slot without waiting and
 // reports whether it succeeded. It fails when an image job is active or
-// pending.
+// pending or an external hold is set.
 func (l *Lock) TryAcquireLLM() bool {
 	l.mu.Lock()
-	if l.imageActive || len(l.imageQ) > 0 {
+	if l.imageActive || len(l.imageQ) > 0 || l.external != "" {
 		l.mu.Unlock()
 		return false
 	}
@@ -120,7 +153,7 @@ func (l *Lock) AcquireImage(ctx context.Context) error {
 
 	for {
 		l.mu.Lock()
-		if l.imageQ[0].id == w.id && l.n == 0 && !l.imageActive {
+		if l.imageQ[0].id == w.id && l.n == 0 && !l.imageActive && l.external == "" {
 			l.imageQ = l.imageQ[1:]
 			l.imageActive = true
 			l.mu.Unlock()
@@ -166,6 +199,8 @@ func (l *Lock) Snapshot() (state State, llmInflight int, imagePending bool) {
 		state = StateImage
 	case l.n > 0:
 		state = StateLLM
+	case l.external != "":
+		state = StateExternal
 	default:
 		state = StateIdle
 	}

internal/lock/lock_test.go

@@ -226,3 +226,70 @@ func TestRace(t *testing.T) {
 		t.Fatalf("leaked lock state: state=%s n=%d pending=%v", state, n, pending)
 	}
 }
+
+func TestExternalHoldBlocksBoth(t *testing.T) {
+	lk := New(nil)
+	ctx := context.Background()
+
+	lk.SetExternal("game.exe (pid 42)")
+	if lk.TryAcquireLLM() {
+		t.Fatal("TryAcquireLLM succeeded during external hold")
+	}
+	if got := lk.External(); got != "game.exe (pid 42)" {
+		t.Fatalf("External() = %q", got)
+	}
+	if state, _, _ := lk.Snapshot(); state != StateExternal {
+		t.Fatalf("state=%s, want external", state)
+	}
+
+	llmAcquired := make(chan struct{})
+	go func() {
+		if err := lk.AcquireLLM(ctx); err != nil {
+			t.Error(err)
+		}
+		close(llmAcquired)
+	}()
+	assertBlocked(t, llmAcquired, "LLM acquire during external hold")
+
+	imageAcquired := make(chan struct{})
+	go func() {
+		if err := lk.AcquireImage(ctx); err != nil {
+			t.Error(err)
+		}
+		close(imageAcquired)
+	}()
+	assertBlocked(t, imageAcquired, "image acquire during external hold")
+
+	lk.ClearExternal()
+	// The queued image job wins over the LLM waiter (image priority).
+	waitFor(t, imageAcquired, "image acquire after ClearExternal")
+	assertBlocked(t, llmAcquired, "LLM acquire while image active")
+	lk.ReleaseImage()
+	waitFor(t, llmAcquired, "LLM acquire after image release")
+	lk.ReleaseLLM()
+}
+
+func TestExternalHoldDoesNotPreempt(t *testing.T) {
+	lk := New(nil)
+	ctx := context.Background()
+	if err := lk.AcquireLLM(ctx); err != nil {
+		t.Fatal(err)
+	}
+
+	lk.SetExternal("game.exe")
+	// In-flight LLM work keeps the llm state; the hold blocks new grants.
+	if state, _, _ := lk.Snapshot(); state != StateLLM {
+		t.Fatalf("state=%s, want llm while work in flight", state)
+	}
+	if lk.TryAcquireLLM() {
+		t.Fatal("TryAcquireLLM succeeded during external hold")
+	}
+	lk.ReleaseLLM()
+	if state, _, _ := lk.Snapshot(); state != StateExternal {
+		t.Fatalf("state=%s, want external after drain", state)
+	}
+	lk.ClearExternal()
+	if state, _, _ := lk.Snapshot(); state != StateIdle {
+		t.Fatalf("state=%s, want idle", state)
+	}
+}

internal/metrics/metrics.go

@@ -89,7 +89,7 @@ func (m *Metrics) Render(w io.Writer, state string, llmInflight int, imagePendin
 	fmt.Fprint(w, `# HELP gpu_turnstile_state Current GPU state (1 for the active state).
 # TYPE gpu_turnstile_state gauge
 `)
-	for _, s := range []string{"idle", "llm", "image"} {
+	for _, s := range []string{"idle", "llm", "image", "external"} {
 		v := 0
 		if s == state {
 			v = 1

internal/proxy/proxy.go

@@ -26,6 +26,7 @@ import (
 	"gpu-turnstile/internal/lock"
 	"gpu-turnstile/internal/metrics"
 	"gpu-turnstile/internal/ollama"
+	"gpu-turnstile/internal/supervise"
 )
 
 // defaultCaptureLimit bounds how much of a /prompt response body is
@@ -44,6 +45,11 @@ type Config struct {
 	Metrics *metrics.Metrics
 	Log     *slog.Logger
 
+	// ComfySup, when non-nil, is the managed ComfyUI process: any ComfyUI
+	// request starts it on demand, /prompt additionally waits for
+	// readiness before taking the GPU lock. nil = unmanaged upstream.
+	ComfySup *supervise.Process
+
 	// LogColor enables ANSI colors in per-request log lines. Ignored when
 	// the log level is above INFO (request lines are not emitted at all).
 	LogColor bool
@@ -438,7 +444,7 @@ func isLLMRequest(r *http.Request) bool {
 	return r.Method == http.MethodPost && llmPaths[r.URL.Path]
 }
 
-// OllamaHandler serves the Ollama-facing listener.
+// OllamaHandler serves the listener for Ollama-compatible clients.
 func (s *Server) OllamaHandler() http.Handler {
 	return s.logRequests("ollama", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		switch r.URL.Path {
@@ -458,10 +464,14 @@ func (s *Server) OllamaHandler() http.Handler {
 		if s.busyMode == "reject" {
 			if !s.cfg.Lock.TryAcquireLLM() {
 				s.cfg.Metrics.ObserveLockWait("llm", time.Since(start).Seconds())
+				msg := "GPU busy: image job active or queued"
+				if holder := s.cfg.Lock.External(); holder != "" {
+					msg = "GPU busy: " + holder
+				}
 				s.log.Info("llm request rejected; GPU busy",
 					"path", r.URL.Path, "status", s.busyStatus)
 				w.Header().Set("Retry-After", strconv.Itoa(s.busyRetryAfter))
-				http.Error(w, "GPU busy: image job active or queued", s.busyStatus)
+				http.Error(w, msg, s.busyStatus)
 				return
 			}
 			s.cfg.Metrics.ObserveLockWait("llm", time.Since(start).Seconds())
@@ -485,7 +495,7 @@ func (s *Server) OllamaHandler() http.Handler {
 	}))
 }
 
-// ComfyHandler serves the ComfyUI-facing listener.
+// ComfyHandler serves the listener for ComfyUI clients.
 func (s *Server) ComfyHandler() http.Handler {
 	return s.logRequests("comfy", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		switch r.URL.Path {
@@ -500,6 +510,21 @@ func (s *Server) ComfyHandler() http.Handler {
 			s.handlePrompt(w, r)
 			return
 		}
+		if s.cfg.ComfySup != nil {
+			// Any other ComfyUI request also wakes the managed server; the
+			// retry backoff bridges the time it needs to come up. While a
+			// foreign process holds the GPU we refuse to spawn it — the
+			// request gets the busy answer instead of fighting for VRAM.
+			if holder := s.cfg.Lock.External(); holder != "" && !s.cfg.ComfySup.Running() {
+				w.Header().Set("Retry-After", strconv.Itoa(s.busyRetryAfter))
+				http.Error(w, "GPU busy: "+holder, http.StatusServiceUnavailable)
+				return
+			}
+			if err := s.cfg.ComfySup.EnsureRunning(); err != nil {
+				http.Error(w, fmt.Sprintf("cannot start ComfyUI: %v", err), http.StatusBadGateway)
+				return
+			}
+		}
 		s.comfyProxy.ServeHTTP(w, r)
 	}))
 }
@@ -539,6 +564,23 @@ func (w *captureWriter) Unwrap() http.ResponseWriter { return w.ResponseWriter }
 func (s *Server) handlePrompt(w http.ResponseWriter, r *http.Request) {
 	log := s.log.With("op", "image")
 
+	// Normally the managed server is brought up *before* taking the GPU
+	// lock: torch can take a minute to load, and LLM traffic should keep
+	// flowing in the meantime. While a foreign process holds the GPU, LLM
+	// traffic is blocked anyway and a fresh ComfyUI would fight it for
+	// VRAM — so the lock comes first in that case.
+	comfyFirst := s.cfg.ComfySup != nil && s.cfg.Lock.External() == ""
+	if comfyFirst {
+		if err := s.cfg.ComfySup.EnsureRunning(); err != nil {
+			http.Error(w, fmt.Sprintf("cannot start ComfyUI: %v", err), http.StatusBadGateway)
+			return
+		}
+		if err := s.cfg.ComfySup.WaitReady(r.Context()); err != nil {
+			http.Error(w, fmt.Sprintf("ComfyUI did not become ready: %v", err), http.StatusBadGateway)
+			return
+		}
+	}
+
 	start := time.Now()
 	if err := s.cfg.Lock.AcquireImage(r.Context()); err != nil {
 		if errors.Is(err, context.DeadlineExceeded) {
@@ -549,6 +591,19 @@ func (s *Server) handlePrompt(w http.ResponseWriter, r *http.Request) {
 	s.cfg.Metrics.ObserveLockWait("image", time.Since(start).Seconds())
 	log.Info("image lock acquired")
 
+	if s.cfg.ComfySup != nil && !comfyFirst {
+		if err := s.cfg.ComfySup.EnsureRunning(); err != nil {
+			s.cfg.Lock.ReleaseImage()
+			http.Error(w, fmt.Sprintf("cannot start ComfyUI: %v", err), http.StatusBadGateway)
+			return
+		}
+		if err := s.cfg.ComfySup.WaitReady(r.Context()); err != nil {
+			s.cfg.Lock.ReleaseImage()
+			http.Error(w, fmt.Sprintf("ComfyUI did not become ready: %v", err), http.StatusBadGateway)
+			return
+		}
+	}
+
 	if s.cfg.Ollama != nil {
 		uctx, ucancel := context.WithTimeout(r.Context(), s.cfg.UnloadTimeout)
 		elapsed, uerr := s.cfg.Ollama.UnloadAll(uctx)
@@ -608,6 +663,10 @@ func (s *Server) finishImageJob(promptID string) {
 
 	s.cfg.Lock.ReleaseImage()
 	log.Info("image lock released")
+	if s.cfg.ComfySup != nil {
+		// The idle clock starts when the job ends, not when it began.
+		s.cfg.ComfySup.NoteActivity()
+	}
 
 	if s.cfg.Ollama != nil && s.cfg.WarmModel != "" {
 		if state, _, _ := s.cfg.Lock.Snapshot(); state == lock.StateIdle {

internal/proxy/proxy_test.go

@@ -46,8 +46,8 @@ type fakes struct {
 	rec       *recorder
 	ollama    *httptest.Server
 	comfy     *httptest.Server
-	server    *httptest.Server // Ollama-facing gpu-turnstile listener
+	server    *httptest.Server // gpu-turnstile listener for Ollama-compatible clients
-	comfySrv  *httptest.Server // ComfyUI-facing gpu-turnstile listener
+	comfySrv  *httptest.Server // gpu-turnstile listener for ComfyUI clients
 	freeCh    chan struct{}
 	chatCh    chan struct{}
 	historyMu sync.Mutex

internal/supervise/supervise.go

@@ -0,0 +1,314 @@
+// Package supervise runs an upstream server (ComfyUI) as a managed child
+// process: started on demand when a request needs it, stopped after an
+// idle timeout so the GPU memory it holds is freed, and stopped with the
+// parent. Crashes are logged; the next request respawns it.
+package supervise
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/url"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"sync"
+	"time"
+)
+
+// ComfyLayout returns the interpreter and script path of a standard ComfyUI
+// venv install rooted at dir for the given GOOS: .venv\Scripts\python.exe
+// on Windows, .venv/bin/python elsewhere. The script is main.py — either in
+// a ComfyUI subdirectory or directly under dir, whichever exists (the
+// subdirectory form wins ties and is the default when neither exists yet,
+// so the caller's missing-file warning points at the documented layout).
+func ComfyLayout(goos, dir string) (python, script string) {
+	if goos == "windows" {
+		python = filepath.Join(dir, ".venv", "Scripts", "python.exe")
+	} else {
+		python = filepath.Join(dir, ".venv", "bin", "python")
+	}
+	script = filepath.Join(dir, "ComfyUI", "main.py")
+	if _, err := os.Stat(script); err != nil {
+		if _, err := os.Stat(filepath.Join(dir, "main.py")); err == nil {
+			script = filepath.Join(dir, "main.py")
+		}
+	}
+	return python, script
+}
+
+// DefaultComfyCommand builds the launch command for the standard venv
+// layout (see ComfyLayout): the script is passed relative to dir so dir
+// stays the working directory, and --port is taken from comfyURL when the
+// URL carries one.
+func DefaultComfyCommand(goos, dir, comfyURL string) string {
+	python, script := ComfyLayout(goos, dir)
+	rel, err := filepath.Rel(dir, script)
+	if err != nil {
+		rel = script
+	}
+	cmd := `"` + python + `" ` + rel
+	if u, err := url.Parse(comfyURL); err == nil && u.Port() != "" {
+		cmd += " --port " + u.Port()
+	}
+	return cmd
+}
+
+// Process is one managed child process.
+type Process struct {
+	name         string
+	argv         []string
+	dir          string
+	probe        func(context.Context) error
+	startTimeout time.Duration
+	log          *slog.Logger
+
+	mu           sync.Mutex
+	cmd          *exec.Cmd
+	stopping     bool
+	ready        bool
+	external     bool // someone else serves the port; not our process
+	lastActivity time.Time
+}
+
+// New parses cmdLine (double quotes group arguments containing spaces) and
+// prepares a managed process. probe reports whether the server answers
+// (e.g. the comfy client's Probe); startTimeout bounds WaitReady. dir is
+// the child's working directory; empty inherits ours.
+func New(name, cmdLine, dir string, probe func(context.Context) error, startTimeout time.Duration, log *slog.Logger) (*Process, error) {
+	argv, err := splitCommandLine(cmdLine)
+	if err != nil {
+		return nil, fmt.Errorf("%s command: %w", name, err)
+	}
+	if len(argv) == 0 {
+		return nil, fmt.Errorf("%s command is empty", name)
+	}
+	if log == nil {
+		log = slog.Default()
+	}
+	if startTimeout <= 0 {
+		startTimeout = 2 * time.Minute
+	}
+	return &Process{name: name, argv: argv, dir: dir, probe: probe, startTimeout: startTimeout, log: log}, nil
+}
+
+// splitCommandLine splits a command line on whitespace, treating
+// double-quoted sections as one argument (quotes removed). Backslashes are
+// literal — this matches Windows paths.
+func splitCommandLine(s string) ([]string, error) {
+	var argv []string
+	var cur strings.Builder
+	inQuote := false
+	have := false
+	flush := func() {
+		if have {
+			argv = append(argv, cur.String())
+			cur.Reset()
+			have = false
+		}
+	}
+	for _, r := range s {
+		switch {
+		case r == '"':
+			inQuote = !inQuote
+			have = true
+		case (r == ' ' || r == '\t') && !inQuote:
+			flush()
+		default:
+			cur.WriteRune(r)
+			have = true
+		}
+	}
+	if inQuote {
+		return nil, fmt.Errorf("unterminated quote in %q", s)
+	}
+	flush()
+	return argv, nil
+}
+
+// Running reports whether the child process is currently alive.
+func (p *Process) Running() bool {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.cmd != nil
+}
+
+// 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 {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.ready
+}
+
+// MarkReady records that the server answered.
+func (p *Process) MarkReady() {
+	p.mu.Lock()
+	p.ready = true
+	p.mu.Unlock()
+}
+
+// NoteActivity resets the idle clock; called for every request served.
+func (p *Process) NoteActivity() {
+	p.mu.Lock()
+	p.lastActivity = time.Now()
+	p.mu.Unlock()
+}
+
+// EnsureRunning starts the child if it is not running. It returns as soon
+// as the process is spawned; readiness is WaitReady's job (and the proxy's
+// retry backoff bridges the gap for plain proxied requests). When the URL
+// already answers — e.g. the ComfyUI desktop app grabbed the port — no
+// child is spawned: the external server is used as-is, and the idle
+// watcher never touches it (it only kills its own child).
+func (p *Process) EnsureRunning() error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.lastActivity = time.Now()
+	if p.cmd != nil {
+		return nil
+	}
+	pctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
+	err := p.probe(pctx)
+	cancel()
+	if err == nil {
+		p.ready = true
+		if !p.external {
+			p.external = true
+			p.log.Info(p.name + " is already served externally; not spawning a managed instance")
+		}
+		return nil
+	}
+	p.external = false
+	cmd := exec.Command(p.argv[0], p.argv[1:]...)
+	cmd.Dir = p.dir
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return err
+	}
+	stderr, err := cmd.StderrPipe()
+	if err != nil {
+		return err
+	}
+	if err := cmd.Start(); err != nil {
+		return fmt.Errorf("start %s: %w", p.name, err)
+	}
+	p.cmd = cmd
+	p.stopping = false
+	p.ready = false
+	go p.pipeLog(stdout)
+	go p.pipeLog(stderr)
+	go func() {
+		err := cmd.Wait()
+		p.mu.Lock()
+		p.cmd = nil
+		p.ready = false
+		intentional := p.stopping
+		p.mu.Unlock()
+		if intentional {
+			p.log.Info(p.name + " stopped")
+		} else {
+			p.log.Warn(p.name+" exited unexpectedly; the next request restarts it", "err", err)
+		}
+	}()
+	p.log.Info(p.name+" starting", "pid", cmd.Process.Pid, "cmd", strings.Join(p.argv, " "))
+	return nil
+}
+
+// WaitReady blocks until the probe succeeds, ctx ends, or the start
+// timeout passes.
+func (p *Process) WaitReady(ctx context.Context) error {
+	ctx, cancel := context.WithTimeout(ctx, p.startTimeout)
+	defer cancel()
+	for {
+		pctx, pcancel := context.WithTimeout(ctx, 5*time.Second)
+		err := p.probe(pctx)
+		pcancel()
+		if err == nil {
+			p.NoteActivity()
+			p.MarkReady()
+			return nil
+		}
+		select {
+		case <-ctx.Done():
+			return fmt.Errorf("%s did not become ready: %w", p.name, ctx.Err())
+		case <-time.After(500 * time.Millisecond):
+		}
+	}
+}
+
+// Stop kills the child process (the whole tree on Windows) if running.
+func (p *Process) Stop() {
+	p.mu.Lock()
+	cmd := p.cmd
+	if cmd == nil {
+		p.mu.Unlock()
+		return
+	}
+	p.stopping = true
+	p.ready = false
+	p.mu.Unlock()
+	stopTree(cmd)
+}
+
+// WatchIdle stops the child after idleTimeout without activity, but only
+// when gpuIdle reports the GPU lock is free (no active or pending work).
+// Returns when ctx ends.
+func (p *Process) WatchIdle(ctx context.Context, idleTimeout time.Duration, gpuIdle func() bool) {
+	ticker := time.NewTicker(5 * time.Second)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+		p.mu.Lock()
+		idleFor := time.Since(p.lastActivity)
+		running := p.cmd != nil
+		p.mu.Unlock()
+		if running && idleFor > idleTimeout && gpuIdle() {
+			p.log.Info(p.name+" idle; stopping to free the GPU", "idle_for", idleFor.Round(time.Second))
+			p.Stop()
+		}
+	}
+}
+
+// pipeLog forwards one child output stream to the log at INFO, line by
+// line, prefixed with the process name.
+func (p *Process) pipeLog(r io.Reader) {
+	buf := make([]byte, 4096)
+	var line string
+	for {
+		n, err := r.Read(buf)
+		line += string(buf[:n])
+		for {
+			i := strings.IndexByte(line, '\n')
+			if i < 0 {
+				break
+			}
+			p.log.Info(p.name + ": " + strings.TrimRight(line[:i], "\r"))
+			line = line[i+1:]
+		}
+		if err != nil {
+			if strings.TrimSpace(line) != "" {
+				p.log.Info(p.name + ": " + line)
+			}
+			return
+		}
+	}
+}
+
+// stopTree kills cmd's process, including its children on Windows (python
+// launchers tend to spawn some). The Wait goroutine reaps it.
+func stopTree(cmd *exec.Cmd) {
+	if runtime.GOOS == "windows" {
+		exec.Command("taskkill", "/T", "/F", "/PID",
+			fmt.Sprint(cmd.Process.Pid)).Run() //nolint:errcheck // best effort
+		return
+	}
+	cmd.Process.Kill() //nolint:errcheck // best effort
+}

internal/supervise/supervise_test.go

@@ -0,0 +1,241 @@
+package supervise
+
+import (
+	"context"
+	"errors"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestSplitCommandLine(t *testing.T) {
+	cases := []struct {
+		in   string
+		want []string
+	}{
+		{`python main.py --port 8188`, []string{"python", "main.py", "--port", "8188"}},
+		{`"C:\Program Files\py\python.exe" main.py`, []string{`C:\Program Files\py\python.exe`, "main.py"}},
+		{`  spaced   out  `, []string{"spaced", "out"}},
+		{`a "b c" d`, []string{"a", "b c", "d"}},
+		{"", nil},
+	}
+	for _, c := range cases {
+		got, err := splitCommandLine(c.in)
+		if err != nil {
+			t.Errorf("splitCommandLine(%q): %v", c.in, err)
+			continue
+		}
+		if len(got) != len(c.want) {
+			t.Errorf("splitCommandLine(%q) = %v, want %v", c.in, got, c.want)
+			continue
+		}
+		for i := range got {
+			if got[i] != c.want[i] {
+				t.Errorf("splitCommandLine(%q) = %v, want %v", c.in, got, c.want)
+				break
+			}
+		}
+	}
+	if _, err := splitCommandLine(`"unterminated`); err == nil {
+		t.Error("expected error for unterminated quote")
+	}
+}
+
+// TestHelperProcess is the child executed by the lifecycle tests: it just
+// sleeps. The env marker is set only around the spawn, so in the normal
+// test run this returns immediately.
+func TestHelperProcess(t *testing.T) {
+	if os.Getenv("GO_HELPER_PROCESS") != "1" {
+		return
+	}
+	time.Sleep(30 * time.Second)
+	os.Exit(0)
+}
+
+func newHelper(t *testing.T, name string) *Process {
+	t.Helper()
+	// The probe always fails: nothing external serves the port, so
+	// EnsureRunning spawns the helper child.
+	p, err := New(name, `"`+os.Args[0]+`" -test.run=TestHelperProcess`, "",
+		func(context.Context) error { return errors.New("nothing there") }, 5*time.Second, slog.Default())
+	if err != nil {
+		t.Fatal(err)
+	}
+	return p
+}
+
+// startHelper spawns the child with the marker set; exec.Command inherits
+// the environment at spawn time, so it can be unset right after.
+func startHelper(t *testing.T, p *Process) {
+	t.Helper()
+	os.Setenv("GO_HELPER_PROCESS", "1")
+	defer os.Unsetenv("GO_HELPER_PROCESS")
+	if err := p.EnsureRunning(); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func waitStopped(t *testing.T, p *Process, timeout time.Duration) {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for p.Running() && time.Now().Before(deadline) {
+		time.Sleep(50 * time.Millisecond)
+	}
+	if p.Running() {
+		t.Fatal("process still running")
+	}
+}
+
+func TestEnsureRunningAndStop(t *testing.T) {
+	p := newHelper(t, "helper")
+	if p.Running() {
+		t.Fatal("Running before start")
+	}
+	startHelper(t, p)
+	if !p.Running() {
+		t.Fatal("not Running after EnsureRunning")
+	}
+	if err := p.EnsureRunning(); err != nil {
+		t.Fatal("second EnsureRunning must be a no-op")
+	}
+	p.Stop()
+	waitStopped(t, p, 5*time.Second)
+}
+
+func TestEnsureRunningPrefersExternalServer(t *testing.T) {
+	// The port is already served (e.g. the ComfyUI desktop app): no child
+	// is spawned, the supervisor reports ready, and Stop is a no-op.
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer srv.Close()
+	probe := func(ctx context.Context) error {
+		req, _ := http.NewRequestWithContext(ctx, http.MethodGet, srv.URL, nil)
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			return err
+		}
+		resp.Body.Close()
+		return nil
+	}
+	p, err := New("external", `"`+os.Args[0]+`"`, "", probe, 5*time.Second, slog.Default())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := p.EnsureRunning(); err != nil {
+		t.Fatal(err)
+	}
+	if p.Running() {
+		t.Fatal("spawned a child even though the port is already served")
+	}
+	if !p.Ready() {
+		t.Fatal("external server should count as ready")
+	}
+	p.Stop() // must not touch the external server
+}
+
+func TestWaitReady(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer srv.Close()
+	probe := func(ctx context.Context) error {
+		req, _ := http.NewRequestWithContext(ctx, http.MethodGet, srv.URL, nil)
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			return err
+		}
+		resp.Body.Close()
+		return nil
+	}
+	p, err := New("ready", `"`+os.Args[0]+`"`, "", probe, 5*time.Second, slog.Default())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := p.WaitReady(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+
+	failing, err := New("failing", `"`+os.Args[0]+`"`, "",
+		func(context.Context) error { return errors.New("no") }, 500*time.Millisecond, slog.Default())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := failing.WaitReady(context.Background()); err == nil {
+		t.Fatal("expected timeout error from WaitReady")
+	}
+}
+
+func TestWatchIdleStops(t *testing.T) {
+	p := newHelper(t, "idle")
+	startHelper(t, p)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go p.WatchIdle(ctx, 200*time.Millisecond, func() bool { return true })
+	waitStopped(t, p, 10*time.Second)
+}
+
+func TestWatchIdleRespectsBusyGPU(t *testing.T) {
+	p := newHelper(t, "busy")
+	startHelper(t, p)
+	defer p.Stop()
+	ctx, cancel := context.WithCancel(context.Background())
+	go p.WatchIdle(ctx, 100*time.Millisecond, func() bool { return false })
+	time.Sleep(600 * time.Millisecond)
+	cancel()
+	if !p.Running() {
+		t.Fatal("process was stopped while the GPU was busy")
+	}
+}
+
+func TestComfyLayoutAndDefaultCommand(t *testing.T) {
+	// Nested layout (ComfyUI/main.py under dir) wins.
+	dir := t.TempDir()
+	nested := filepath.Join(dir, "ComfyUI", "main.py")
+	if err := os.MkdirAll(filepath.Dir(nested), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(nested, []byte("x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	python, script := ComfyLayout("windows", dir)
+	if want := filepath.Join(dir, ".venv", "Scripts", "python.exe"); python != want {
+		t.Errorf("python = %s, want %s", python, want)
+	}
+	if script != nested {
+		t.Errorf("script = %s, want %s", script, nested)
+	}
+	cmd := DefaultComfyCommand("windows", dir, "http://127.0.0.1:8189")
+	want := `"` + filepath.Join(dir, ".venv", "Scripts", "python.exe") + `" ` + filepath.Join("ComfyUI", "main.py") + " --port 8189"
+	if cmd != want {
+		t.Errorf("cmd = %q, want %q", cmd, want)
+	}
+
+	// Flat layout (main.py directly under dir) is found too.
+	flat := t.TempDir()
+	if err := os.WriteFile(filepath.Join(flat, "main.py"), []byte("x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if _, script := ComfyLayout("linux", flat); script != filepath.Join(flat, "main.py") {
+		t.Errorf("flat script = %s", script)
+	}
+	cmd = DefaultComfyCommand("linux", flat, "http://comfy.internal")
+	if strings.Contains(cmd, "--port") {
+		t.Errorf("cmd = %q, want no --port for a port-less URL", cmd)
+	}
+	if !strings.HasSuffix(cmd, `" main.py`) {
+		t.Errorf("cmd = %q, want quoted python + relative main.py", cmd)
+	}
+
+	// Neither exists yet: default to the documented nested form so the
+	// startup warning points there.
+	empty := t.TempDir()
+	if _, script := ComfyLayout("windows", empty); script != filepath.Join(empty, "ComfyUI", "main.py") {
+		t.Errorf("missing-layout script = %s", script)
+	}
+}