Pin compose example to v0.1.8

Game detection: foreign GPU holders take an external lock hold (GAME_PROCS, GPU_FOREIGN_VRAM_MB)
Supervisor probes before spawning: an external server on the port is used, never fought or killed
2026-09-21 17:19:16 +02:00 · 2026-09-21 17:16:56 +02:00 · 2026-09-21 16:06:28 +02:00 · 2026-09-21 13:48:07 +02:00 · 2026-09-21 13:14:33 +02:00 · 2026-09-21 13:14:18 +02:00
27 changed files with 2322 additions and 151 deletions
@@ -4,3 +4,6 @@
 /compose.yml
 /signing/

+/gpu-turnstile.exe.old
+/gpu-turnstile.exe.new
+/gpu-turnstile.exe~
@@ -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:
-`idle`, `llm` (N ≥ 1 Ollama requests in flight), or `image` (exactly one
-ComfyUI job, Ollama models unloaded). See [SPEC.md](SPEC.md) for the full
+front of both and guarantees the GPU is always in exactly one of four states:
+`idle`, `llm` (N ≥ 1 Ollama requests in flight), `image` (exactly one
+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
-detection) plug into the same lock the same way.
+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).

 ## Configuration

@@ -45,8 +47,8 @@ override file values. Invalid values fail at startup.

 | Var | Default | Meaning |
 |---|---|---|
-| `LISTEN_OLLAMA` | `:11434` | Ollama-facing listener |
-| `LISTEN_COMFY` | `:8188` | ComfyUI-facing listener |
+| `LISTEN_OLLAMA` | `:11434` | Listener for Ollama-compatible clients |
+| `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` | _(empty = unmanaged)_ | Supervise ComfyUI: start on demand, stop when idle to free VRAM. Requires `COMFY_URL` |
+| `COMFY_DIR` | _(empty)_ | 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` | `5s` | How often game/VRAM detection runs |
 | `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 |
@@ -72,10 +83,11 @@ override file values. Invalid values fail at startup.
 | `UPDATE_INTERVAL` | `6h` | Auto-update check interval |
 | `UPDATE_REPO` | `https://git.rambossek.at/PUBLIC/gpu-turnstile` | Repository checked for releases |
 | `UPDATE_ASSET` | `gpu-turnstile.exe` | Release asset to download |
+| `APP_VER` | `stable` | `dev` disables updates, `stable` tracks latest, or pin an exact `vX.Y.Z` |

 ## 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`
@@ -87,6 +99,52 @@ 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`)
+
+Don't want ComfyUI running 24/7 (it holds VRAM even when idle — and the
+Desktop app kills its server when you close it)? Point `COMFY_CMD` at a
+standalone launch command and gpu-turnstile supervises 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. Example for a Desktop
+install (run it once manually to confirm it works):
+
+```
+COMFY_URL=http://127.0.0.1:8188
+COMFY_CMD="C:\ComfyUI\.venv\Scripts\python.exe ComfyUI\main.py --port 8188"
+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` (5s):
+
+```
+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
@@ -112,12 +170,19 @@ gpu-turnstile.exe --remove-service
 ```

 Layout: `C:\Program Files\gpu-turnstile\` holds the exe and
-`gpu-turnstile.env`, logs go to `C:\ProgramData\gpu-turnstile\` (set
-`LOG_FILE` in the env file — there is no console). The service always runs
+`gpu-turnstile.env`, logs go to `C:\ProgramData\gpu-turnstile\`. If you
+install without a config, the installer writes a sample env file with every
+setting commented and explained — only `LOG_FILE` is active (a service has
+no console). Your own `LOG_FILE` setting is always kept. The service always
+runs
 as the virtual account `NT SERVICE\gpu-turnstile` (low-privilege,
 per-service, no password); the installer automatically grants it write
 access to the install and data directories — nothing else to do.

+Re-running `--install-service` is safe: it stops a running service,
+replaces the installed binary only if it changed, fixes the registration
+only where it drifted, and restarts the service only if it was running.
+
 ### Run natively on Linux (systemd)

 The same binary works on Linux. Install it as a systemd service as root:
@@ -144,7 +209,10 @@ containers and interactive shells.
 release on startup and every `UPDATE_INTERVAL`, verifies the Ed25519
 signature of the download against the public key embedded at build time,
 and — once the GPU lock is idle — restarts the service onto the new
-version. Disable with `AUTO_UPDATE=false`. Releases are signed by CI with
+version. `APP_VER` controls the target: `dev` disables updates, `stable`
+(the default) tracks the latest release, and an exact `vX.Y.Z` pins that
+release (even as a downgrade or to replace a dev build). Disable entirely
+with `AUTO_UPDATE=false`. Releases are signed by CI with
 OpenSSL; the matching public key lives in `internal/update/pubkey.go`
 (one-time setup: `openssl genpkey -algorithm ed25519 -out private.pem`,
 `openssl pkey -in private.pem -pubout -out public.pem`; private key goes
@@ -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
-  same lock the same way: enabled by their config knob, excluded when
-  absent.
+- **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.

 ### 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 +128,68 @@ 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`)
+
+When `COMFY_CMD` is set, gpu-turnstile runs ComfyUI as a supervised child
+process instead of expecting an always-on server:
+
+- **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 5 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 +200,8 @@ override file values. A missing file is fine; a malformed one is fatal.

 | Var | Default | Meaning |
 |---|---|---|
-| `LISTEN_OLLAMA` | `:11434` | Ollama-facing listener |
-| `LISTEN_COMFY` | `:8188` | ComfyUI-facing listener |
+| `LISTEN_OLLAMA` | `:11434` | listener for Ollama-compatible clients |
+| `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 +211,21 @@ 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` | _(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` | `` | 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` | `5s` | how often game/VRAM detection runs |
 | `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 |
@@ -158,6 +236,8 @@ override file values. A missing file is fine; a malformed one is fatal.
 | `UPDATE_INTERVAL` | `6h` | auto-update check interval |
 | `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` (always a concrete `vX.Y.Z`; a dev build stamps `v0.0.0`); missing = the file is replaced with a fresh sample (backup `.bak`) |

 Startup fails fast on unparsable values and when neither consumer URL is
 set. Enabled upstreams are probed once at start (`/api/version`,
@@ -176,12 +256,27 @@ prompt instead of failing — the command relaunches itself elevated, waits
 for the child, and mirrors its exit code. The legacy form
 `gpu-turnstile service install|remove` does the same thing.

+Re-running install on an already-registered service converges instead of
+failing: a running service is stopped first, the installed binary copy is
+refreshed only when the content differs, the registration (Windows service
+config / systemd unit) is updated only where it drifted, and the service is
+started again only if it was running before.
+
 By default install creates the canonical layout and copies the binary into
 it (Windows: `%ProgramFiles%\gpu-turnstile\`, plus
 `%ProgramData%\gpu-turnstile\` for logs; Linux: `/var/lib/gpu-turnstile/`
-with the config at `/etc/gpu-turnstile.env`). An existing config in the
-target location is never overwritten. `--no-copy` registers the current
-executable location as-is instead.
+with the config at `/etc/gpu-turnstile.env`). If there is no config at all,
+install writes a sample env file covering every setting — each with a
+comment line, everything commented out — except the `CFG_VER`/`APP_VER`
+header and `LOG_FILE`, which is active on Windows
+(`%ProgramData%\gpu-turnstile\gpu-turnstile.log`) since a service has no
+console; on Linux it stays commented because stderr goes to the journal.
+The first line is `CFG_VER=vX.Y.Z`, recording the installer version. When a
+later version's install finds an older `CFG_VER`, it appends every setting
+the file does not mention (commented or not) at the end and updates
+`CFG_VER`; a file without `CFG_VER` is invalid and gets replaced by a fresh
+sample, with the old content kept as `<file>.bak`. `--no-copy` registers
+the current executable location as-is and leaves the config untouched.

 ### Windows

@@ -189,8 +284,9 @@ executable location as-is instead.
  `%ProgramData%\gpu-turnstile\`, copies the exe and (if none exists there
  yet) the `gpu-turnstile.env` into the Program Files directory, and
  registers that copy as a Windows service; recovery actions restart it
-  5 s after any failure. Logs go to the ProgramData directory via
-  `LOG_FILE` since there is no console.
+  5 s after any failure. The install ensures the env file sets `LOG_FILE`
+  to `%ProgramData%\gpu-turnstile\gpu-turnstile.log` since there is no
+  console — an existing `LOG_FILE` setting is kept.
 - **Account**: the service always runs as the virtual account
  `NT SERVICE\gpu-turnstile` — a per-service low-privilege identity the
  SCM manages (no password, automatic logon-as-a-service right, no admin
@@ -238,19 +334,25 @@ executable location as-is instead.
  `RestartSec=5s` brings up the staged binary after the updater exits with
  code 3.
 - **Auto-update**: on startup and every `UPDATE_INTERVAL`, the binary
-  checks `UPDATE_REPO`'s latest release; if its tag is a newer `vX.Y.Z`,
-  it downloads `UPDATE_ASSET` plus its `.sig` (and `.sha256` when present)
-  and verifies an Ed25519 signature against the public key embedded in
+  consults `APP_VER`: `dev` disables updates; `stable` (the default)
+  fetches `UPDATE_REPO`'s latest release and applies it when its tag is a
+  newer `vX.Y.Z` (a `dev` binary cannot be compared and is replaced by the
+  latest release); a `vX.Y.Z` pin fetches that exact tag and stages it on
+  any difference, including downgrades. Applying means downloading
+  `UPDATE_ASSET` plus its `.sig` (and `.sha256` when present) and verifying
+  an Ed25519 signature against the public key embedded in
  `internal/update/pubkey.go`. A verified binary is swapped in next to the
  running exe (rename-aside, allowed on Windows), and once the GPU lock is
  idle the process exits with code 3 so the service recovery restarts it
  on the new version. Interactive runs only log "restart to apply".
-  `dev` builds and builds without an embedded public key never update.
- **`--force-update`** runs the same check immediately: it downloads,
-  verifies and stages a newer release, and if the service is running it
-  restarts it right away (otherwise the new version applies on next
-  start). On Windows it elevates via UAC only when the stage or restart
-  needs permissions the caller does not have.
+  Builds without an embedded public key never update.
+- **`--force-update`** runs the same check immediately, single-shot: one
+  attempt with a 30 s timeout, then exit — "up to date" (exit 0) or the
+  error (exit 1), no retries. When a newer release is found it downloads,
+  verifies and stages it, and if the service is running it restarts it
+  right away (otherwise the new version applies on next start). On Windows
+  it elevates via UAC only when the stage or restart needs permissions the
+  caller does not have.
 - **Signing setup (one time)**: `openssl genpkey -algorithm ed25519 -out
  private.pem`; `openssl pkey -in private.pem -pubout -out public.pem`.
  Private key → repo secret `RELEASE_SIGNING_KEY`; public key → committed into
@@ -334,7 +436,8 @@ are new.
  receives the first chunk before the last is sent (no buffering).
 - `internal/config`: env-file parsing, precedence, fail-fast values.
 - `internal/update`: fake Gitea releases API; staged update happy path,
-  tampered signature rejected, older versions and dev builds skipped.
+  tampered signature rejected, older versions skipped, APP_VER=dev and
+  pinned releases honored.

 ## Build and CI

@@ -21,11 +21,13 @@ import (

 	"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"
 )

@@ -299,6 +301,9 @@ func forceUpdateCommand(configPath string, elevatedChild bool) int {
 		defer waitForEnter()
 	}
 	cfg, err := loadMergedConfig(configPath)
+	if errors.Is(err, config.ErrNoConsumer) {
+		err = nil // update settings do not depend on a consumer URL
+	}
 	if err != nil {
 		fmt.Fprintf(os.Stderr, "%s\n\ngpu-turnstile: %v\n", versionLine(), err)
 		return 1
@@ -310,9 +315,15 @@ func forceUpdateCommand(configPath string, elevatedChild bool) int {
 	}
 	log, _, logCloser := newLogger(cfg)
 	defer logCloser.Close()
-	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Log: log}
+	if cfg.AppVersion == "dev" {
+		fmt.Printf("%s: APP_VER=dev, updates disabled\n", versionLine())
+		return 0
+	}
+	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}

-	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	// Single-shot: one attempt, fail fast when the server is unreachable
+	// instead of hanging in a TCP connect for minutes.
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
 	defer cancel()
 	staged, err := u.Check(ctx, exePath)
 	if err != nil && isPermission(err) && !service.Elevated() {
@@ -381,7 +392,7 @@ func serviceCommand(configPath string, install, noCopy, elevatedChild bool) int
 		if abs, absErr := filepath.Abs(path); absErr == nil {
 			path = abs
 		}
-		err = service.Install(path, !noCopy)
+		err = service.Install(path, !noCopy, version)
 	} else {
 		err = service.Remove()
 	}
@@ -419,6 +430,7 @@ 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,
@@ -426,6 +438,14 @@ 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", cfg.ComfyCmd,
+		"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,
@@ -452,12 +472,31 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 		}
 	}

+	// With COMFY_CMD set, ComfyUI runs as a managed child: started on
+	// demand by the proxy, stopped after COMFY_IDLE_TIMEOUT idle (and on
+	// shutdown) so its VRAM is freed.
+	var comfySup *supervise.Process
+	if cfg.ComfyCmd != "" {
+		var err error
+		comfySup, err = supervise.New("comfy", cfg.ComfyCmd, 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") == "",
@@ -481,19 +520,52 @@ 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.
-	probeCtx, probeCancel := context.WithTimeout(ctx, cfg.ProbeTimeout)
+	// Probe the enabled upstreams once; failure is logged, not fatal. A
+	// 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 {
-			log.Warn("ollama probe failed", "url", cfg.OllamaURL, "err", err)
-		}
+		probes["ollama"] = ollamaClient.Probe
 	}
 	if comfyClient != nil {
-		if err := comfyClient.Probe(probeCtx); err != nil {
-			log.Warn("comfy probe failed", "url", cfg.ComfyURL, "err", err)
+		if comfySup == nil {
+			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".
@@ -551,6 +623,128 @@ 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
@@ -561,7 +755,7 @@ func updateLoop(ctx context.Context, cfg config.Config, log *slog.Logger, lk *lo
 		log.Warn("auto-update disabled: cannot locate executable", "err", err)
 		return
 	}
-	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Log: log}
+	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
 	for {
 		staged, err := u.Check(ctx, exePath)
 		if err != nil && ctx.Err() == nil {
@@ -6,7 +6,7 @@
 # ComfyUI --listen 0.0.0.0 --port 8189).
 services:
  gpu-turnstile:
-    image: git.rambossek.at/public/gpu-turnstile:v0.1.6
+    image: git.rambossek.at/public/gpu-turnstile:v0.1.8
    restart: unless-stopped
    environment:
      # Each consumer is enabled by setting its URL; leave one unset to
@@ -4,6 +4,7 @@ package config

 import (
 	"bufio"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -25,6 +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
@@ -37,6 +39,11 @@ type Config struct {
 	UpdateRepo     string
 	UpdateAsset    string

+	// 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
+
 	// 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).
@@ -44,6 +51,27 @@ type Config struct {
 	LLMBusyStatus  int
 	BusyRetryAfter int

+	// ComfyCmd spawns and supervises a ComfyUI server on demand (empty =
+	// unmanaged, the current behavior). ComfyDir is its working directory.
+	// 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
@@ -64,6 +92,7 @@ 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,
@@ -75,15 +104,29 @@ func Defaults() Config {
 		UpdateInterval: 6 * time.Hour,
 		UpdateRepo:     "https://git.rambossek.at/PUBLIC/gpu-turnstile",
 		UpdateAsset:    "gpu-turnstile.exe",
+		AppVersion:     "stable",

 		LLMBusyMode:    "wait",
 		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: 5 * time.Second,
+
 		LogLevel: slog.LevelWarn,
 	}
 }

+// ErrNoConsumer is returned by Load when neither OLLAMA_URL nor COMFY_URL
+// is set. Commands that never talk to an upstream (--force-update) may
+// ignore it and proceed with the rest of the configuration.
+var ErrNoConsumer = errors.New("at least one of OLLAMA_URL or COMFY_URL must be set (each URL enables its consumer)")
+
 // ParseEnvFile parses a .env-style file: KEY=VALUE lines, blank lines and
 // #-comments are ignored, no quoting. A line without '=' is an error.
 func ParseEnvFile(r io.Reader) (map[string]string, error) {
@@ -110,6 +153,17 @@ 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 == "" {
@@ -136,6 +190,8 @@ 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},
@@ -154,17 +210,34 @@ 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 {
@@ -199,6 +272,17 @@ func Load(getenv func(string) string) (Config, error) {
 		}
 		cfg.BusyRetryAfter = n
 	}
+	if v := getenv("APP_VER"); v != "" {
+		switch {
+		case v == "dev" || v == "stable":
+			cfg.AppVersion = v
+		default:
+			if _, ok := parseVersion(v); !ok {
+				return cfg, fmt.Errorf("APP_VER: must be \"dev\", \"stable\" or a vX.Y.Z version")
+			}
+			cfg.AppVersion = "v" + strings.TrimPrefix(v, "v")
+		}
+	}
 	// LOGLEVEL is the canonical spelling; LOG_LEVEL is kept as an alias.
 	logLevelValue := getenv("LOGLEVEL")
 	if logLevelValue == "" {
@@ -218,8 +302,11 @@ 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.OllamaURL == "" && cfg.ComfyURL == "" {
-		return cfg, fmt.Errorf("at least one of OLLAMA_URL or COMFY_URL must be set (each URL enables its consumer)")
+		return cfg, ErrNoConsumer
 	}
 	return cfg, nil
 }
@@ -1,6 +1,7 @@
 package config

 import (
+	"errors"
 	"log/slog"
 	"strings"
 	"testing"
@@ -39,6 +40,61 @@ func TestLoadRequiresConsumer(t *testing.T) {
 	if err == nil || !strings.Contains(err.Error(), "OLLAMA_URL") {
 		t.Fatalf("err = %v, want missing-consumer error", err)
 	}
+	if !errors.Is(err, ErrNoConsumer) {
+		t.Fatalf("err = %v, want errors.Is(err, ErrNoConsumer)", err)
+	}
+}
+
+func TestAppVersion(t *testing.T) {
+	load := func(appVer string) (Config, error) {
+		return Load(func(k string) string {
+			switch k {
+			case "OLLAMA_URL":
+				return "http://127.0.0.1:11435"
+			case "APP_VER":
+				return appVer
+			}
+			return ""
+		})
+	}
+	cfg, err := load("")
+	if err != nil || cfg.AppVersion != "stable" {
+		t.Fatalf("default AppVersion = %q, err %v; want stable", cfg.AppVersion, err)
+	}
+	for _, v := range []string{"dev", "stable"} {
+		if cfg, err := load(v); err != nil || cfg.AppVersion != v {
+			t.Fatalf("APP_VER=%s: got %q, err %v", v, cfg.AppVersion, err)
+		}
+	}
+	if cfg, err := load("1.2.3"); err != nil || cfg.AppVersion != "v1.2.3" {
+		t.Fatalf("APP_VER=1.2.3: got %q, err %v; want normalized v1.2.3", cfg.AppVersion, err)
+	}
+	if _, err := load("nightly"); err == nil {
+		t.Fatal("APP_VER=nightly: want validation error")
+	}
+}
+
+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) {
@@ -101,6 +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 {
@@ -116,3 +174,51 @@ 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 != 5*time.Second {
+		t.Fatalf("GamePollInterval = %v, want 5s 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")
+	}
+}
@@ -0,0 +1,192 @@
+package config
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+)
+
+// sampleEntry is one line pair in the sample env file: a comment and the
+// (usually commented-out) KEY=VALUE line.
+type sampleEntry struct {
+	name    string
+	value   string
+	comment string
+	active  bool // rendered uncommented
+}
+
+// appVerComment documents APP_VER wherever it is rendered.
+const appVerComment = `Version to run: "dev" disables updates, "stable" tracks the latest release, or pin an exact release like v0.1.7`
+
+// sampleEntries lists every setting in sample order. logFile activates the
+// LOG_FILE line (Windows install); empty keeps it commented like the rest.
+// CFG_VER and APP_VER are not entries — they head the file, always active.
+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},
+		{"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: empty = unmanaged)", false},
+		{"COMFY_DIR", `C:\ComfyUI`, "Working directory for COMFY_CMD (default: empty = inherit)", 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", "5s", "How often game/VRAM detection runs", 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},
+		{"LLM_BUSY_MODE", "wait", `How blocked LLM requests are handled: "wait" holds them, "reject" fails them immediately`, false},
+		{"LLM_BUSY_STATUS", "503", "HTTP status for rejected LLM requests in reject mode (400-599, e.g. 429)", false},
+		{"BUSY_RETRY_AFTER", "30", "Seconds sent as Retry-After on busy responses (both modes)", false},
+		{"LOGLEVEL", "warn", `Log verbosity: debug, info, warn, error; "info" logs every request (LOG_LEVEL works too)`, false},
+		{"LOG_FORMAT", "text", `Log format: "text" or "json"`, false},
+		{"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", "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},
+		{"BACKOFF_INITIAL", "1s", "First retry wait when an upstream connection fails", false},
+		{"BACKOFF_MAX", "60s", "Cap for the exponential retry backoff", false},
+		{"PROMPT_CAPTURE_LIMIT", "65536", "Bytes of the /prompt response buffered to find prompt_id (pass-through unaffected)", false},
+		{"AUTO_UPDATE", "true", "Poll the releases API for signed updates", false},
+		{"UPDATE_INTERVAL", "6h", "Auto-update check interval", false},
+		{"UPDATE_REPO", "https://git.rambossek.at/PUBLIC/gpu-turnstile", "Repository to check for releases", false},
+		{"UPDATE_ASSET", "gpu-turnstile.exe", "Release asset to download", false},
+	}
+}
+
+// SampleEnv renders a sample .env file covering every setting, each with a
+// 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.
+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", 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")
+	b.WriteString("# KEY=VALUE lines; \"#\" starts a comment. Every setting below is at its\n")
+	b.WriteString("# default and commented out — remove the \"#\" to change it.\n")
+	b.WriteString("# At least one of OLLAMA_URL / COMFY_URL must be set for the proxy to start.\n\n")
+	writeEntry(&b, sampleEntry{"APP_VER", "stable", appVerComment, true})
+	for _, e := range sampleEntries(logFile) {
+		writeEntry(&b, e)
+	}
+	return b.String()
+}
+
+func writeEntry(b *strings.Builder, e sampleEntry) {
+	fmt.Fprintf(b, "# %s\n", e.comment)
+	if e.active {
+		fmt.Fprintf(b, "%s=%s\n\n", e.name, e.value)
+	} else {
+		fmt.Fprintf(b, "#%s=%s\n\n", e.name, e.value)
+	}
+}
+
+// SyncSample upgrades an installer-written env file: when its CFG_VER says
+// it was written by an older gpu-turnstile, every setting the file does not
+// mention — commented or not — is appended at the end, and CFG_VER is
+// updated to version. Files without CFG_VER (hand-written or foreign),
+// up-to-date files and "dev" builds are returned unchanged; changed reports
+// whether the returned content differs.
+func SyncSample(data, version, logFile string) (string, bool) {
+	if version == "" || version == "dev" {
+		return data, false
+	}
+	values, err := ParseEnvFile(strings.NewReader(data))
+	if err != nil || values["CFG_VER"] == "" {
+		return data, false // not installer-written; the caller decides
+	}
+	if compareVersions(values["CFG_VER"], version) >= 0 {
+		return data, false // same or newer
+	}
+
+	present := make(map[string]bool)
+	for _, line := range strings.Split(data, "\n") {
+		line = strings.TrimSpace(strings.TrimPrefix(strings.TrimSpace(line), "#"))
+		if name, _, ok := strings.Cut(line, "="); ok {
+			present[strings.TrimSpace(name)] = true
+		}
+	}
+
+	lines := strings.Split(data, "\n")
+	for i, l := range lines {
+		if strings.HasPrefix(strings.TrimSpace(l), "CFG_VER=") {
+			lines[i] = "CFG_VER=" + version
+			break
+		}
+	}
+	out := strings.Join(lines, "\n")
+	if !strings.HasSuffix(out, "\n") {
+		out += "\n"
+	}
+
+	var b strings.Builder
+	if !present["APP_VER"] {
+		fmt.Fprintf(&b, "\n# Added by gpu-turnstile %s:\n", version)
+		writeEntry(&b, sampleEntry{"APP_VER", "stable", appVerComment, true})
+	}
+	for _, e := range sampleEntries(logFile) {
+		if !present[e.name] {
+			fmt.Fprintf(&b, "\n# Added by gpu-turnstile %s:\n", version)
+			writeEntry(&b, e)
+		}
+	}
+	out += b.String()
+	return out, out != data
+}
+
+// compareVersions orders two vX.Y.Z version strings. Unparseable versions
+// (including "dev") sort before any release.
+func compareVersions(a, b string) int {
+	pa, oka := parseVersion(a)
+	pb, okb := parseVersion(b)
+	if oka != okb {
+		if oka {
+			return 1
+		}
+		return -1
+	}
+	for i := range pa {
+		if pa[i] != pb[i] {
+			if pa[i] > pb[i] {
+				return 1
+			}
+			return -1
+		}
+	}
+	return 0
+}
+
+func parseVersion(v string) ([3]int, bool) {
+	var out [3]int
+	v = strings.TrimPrefix(strings.TrimSpace(v), "v")
+	parts := strings.Split(v, ".")
+	if len(parts) != 3 {
+		return out, false
+	}
+	for i, p := range parts {
+		n, err := strconv.Atoi(p)
+		if err != nil || n < 0 {
+			return out, false
+		}
+		out[i] = n
+	}
+	return out, true
+}
@@ -0,0 +1,156 @@
+package config
+
+import (
+	"strings"
+	"testing"
+)
+
+const sampleLogPath = `C:\ProgramData\gpu-turnstile\gpu-turnstile.log`
+
+var allSettingNames = []string{
+	"LISTEN_OLLAMA", "LISTEN_COMFY", "OLLAMA_URL", "COMFY_URL",
+	"WARM_MODEL", "UNLOAD_TIMEOUT", "JOB_TIMEOUT", "LLM_WAIT_TIMEOUT",
+	"LLM_BUSY_MODE", "LLM_BUSY_STATUS", "BUSY_RETRY_AFTER",
+	"LOGLEVEL", "LOG_FORMAT", "LOG_FILE",
+	"UNLOAD_POLL_INTERVAL", "HISTORY_POLL_INTERVAL", "PROBE_TIMEOUT",
+	"FREE_TIMEOUT", "WARM_TIMEOUT", "SHUTDOWN_TIMEOUT",
+	"BACKOFF_INITIAL", "BACKOFF_MAX", "PROMPT_CAPTURE_LIMIT",
+	"AUTO_UPDATE", "UPDATE_INTERVAL", "UPDATE_REPO", "UPDATE_ASSET",
+}
+
+func TestSampleEnv(t *testing.T) {
+	sample := SampleEnv("v0.1.7", sampleLogPath)
+
+	if !strings.HasPrefix(sample, "CFG_VER=v0.1.7\n") {
+		t.Errorf("first line does not carry CFG_VER: %q", strings.SplitN(sample, "\n", 2)[0])
+	}
+
+	// Every setting known to Load must appear.
+	for _, name := range allSettingNames {
+		if !strings.Contains(sample, name+"=") {
+			t.Errorf("sample is missing %s", name)
+		}
+	}
+
+	// The sample must parse cleanly; active values are CFG_VER, APP_VER
+	// and LOG_FILE.
+	values, err := ParseEnvFile(strings.NewReader(sample))
+	if err != nil {
+		t.Fatalf("sample does not parse: %v", err)
+	}
+	want := map[string]string{"CFG_VER": "v0.1.7", "APP_VER": "stable", "LOG_FILE": sampleLogPath}
+	if len(values) != len(want) {
+		t.Fatalf("active values = %v, want %v", values, want)
+	}
+	for k, v := range want {
+		if values[k] != v {
+			t.Errorf("%s = %q, want %q", k, values[k], v)
+		}
+	}
+
+	// Without a log path LOG_FILE stays commented out.
+	values, err = ParseEnvFile(strings.NewReader(SampleEnv("v0.1.7", "")))
+	if err != nil {
+		t.Fatalf("sample without log path does not parse: %v", err)
+	}
+	if len(values) != 2 || values["LOG_FILE"] != "" {
+		t.Fatalf("active values = %v, want only CFG_VER and APP_VER", values)
+	}
+}
+
+func TestSyncSample(t *testing.T) {
+	old := SampleEnv("v0.1.6", sampleLogPath)
+	// Simulate a setting that did not exist yet when the file was written.
+	old = strings.Replace(old, "#UPDATE_ASSET=gpu-turnstile.exe\n", "", 1)
+
+	out, changed := SyncSample(old, "v0.1.7", sampleLogPath)
+	if !changed {
+		t.Fatal("older installer file was not upgraded")
+	}
+	if !strings.Contains(out, "\nCFG_VER=v0.1.7\n") && !strings.HasPrefix(out, "CFG_VER=v0.1.7\n") {
+		t.Error("CFG_VER was not updated to the new version")
+	}
+	if !strings.Contains(out, "#UPDATE_ASSET=gpu-turnstile.exe") {
+		t.Error("missing setting was not appended")
+	}
+	if !strings.Contains(out, "LOG_FILE="+sampleLogPath) {
+		t.Error("existing active LOG_FILE was lost")
+	}
+	if !strings.Contains(out, "Added by gpu-turnstile v0.1.7") {
+		t.Error("appended section is not attributed")
+	}
+	if _, err := ParseEnvFile(strings.NewReader(out)); err != nil {
+		t.Fatalf("upgraded file does not parse: %v", err)
+	}
+
+	// A file written by the same or a newer version is left alone.
+	if _, changed := SyncSample(SampleEnv("v0.1.7", sampleLogPath), "v0.1.7", sampleLogPath); changed {
+		t.Error("same-version file was modified")
+	}
+	if _, changed := SyncSample(SampleEnv("v0.2.0", sampleLogPath), "v0.1.7", sampleLogPath); changed {
+		t.Error("newer-version file was modified")
+	}
+
+	// Files without CFG_VER are not installer-written; the installer
+	// replaces them, SyncSample leaves them alone.
+	user := "OLLAMA_URL=http://host:11434\n"
+	if out, changed := SyncSample(user, "v0.1.7", sampleLogPath); changed || out != user {
+		t.Error("file without CFG_VER was modified")
+	}
+
+	// A dev build never upgrades.
+	if _, changed := SyncSample(old, "dev", sampleLogPath); changed {
+		t.Error("dev build modified the file")
+	}
+}
+
+func TestSyncSampleAppendsMissingAppVer(t *testing.T) {
+	old := SampleEnv("v0.1.6", sampleLogPath)
+	old = strings.Replace(old, "# "+appVerComment+"\n", "", 1)
+	old = strings.Replace(old, "APP_VER=stable\n", "", 1)
+
+	out, changed := SyncSample(old, "v0.1.7", sampleLogPath)
+	if !changed {
+		t.Fatal("file without APP_VER was not upgraded")
+	}
+	values, err := ParseEnvFile(strings.NewReader(out))
+	if err != nil {
+		t.Fatalf("upgraded file does not parse: %v", err)
+	}
+	if values["APP_VER"] != "stable" {
+		t.Fatalf("APP_VER = %q, want appended default \"stable\"", values["APP_VER"])
+	}
+}
+
+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
+		want int
+	}{
+		{"v0.1.6", "v0.1.7", -1},
+		{"v0.1.7", "v0.1.7", 0},
+		{"v1.0.0", "v0.9.9", 1},
+		{"0.1.7", "v0.1.6", 1},
+		{"dev", "v0.1.7", -1},
+		{"v0.1.7", "dev", 1},
+		{"dev", "dev", 0},
+	}
+	for _, c := range cases {
+		if got := compareVersions(c.a, c.b); got != c.want {
+			t.Errorf("compareVersions(%q, %q) = %d, want %d", c.a, c.b, got, c.want)
+		}
+	}
+}
@@ -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
+}
@@ -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)
+		}
+	}
+}
@@ -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
+}
@@ -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 }
@@ -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
+}
@@ -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"
-	StateLLM   State = "llm"
-	StateImage State = "image"
+	StateIdle     State = "idle"
+	StateLLM      State = "llm"
+	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
 	}
@@ -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)
+	}
+}
@@ -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
@@ -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 {
@@ -46,8 +46,8 @@ type fakes struct {
 	rec       *recorder
 	ollama    *httptest.Server
 	comfy     *httptest.Server
-	server    *httptest.Server // Ollama-facing gpu-turnstile listener
-	comfySrv  *httptest.Server // ComfyUI-facing gpu-turnstile listener
+	server    *httptest.Server // gpu-turnstile listener for Ollama-compatible clients
+	comfySrv  *httptest.Server // gpu-turnstile listener for ComfyUI clients
 	freeCh    chan struct{}
 	chatCh    chan struct{}
 	historyMu sync.Mutex
@@ -0,0 +1,47 @@
+// Shared env-file handling for the installers (Windows and Linux).
+package service
+
+import (
+	"fmt"
+	"os"
+	"strings"
+
+	"gpu-turnstile/internal/config"
+)
+
+// syncEnvFile ensures the installed env file at path is a current
+// installer-written sample. A missing file is created; a file without a
+// CFG_VER line (hand-written or from before versioning) is invalid and
+// replaced by a fresh sample after a .bak backup; a file written by an
+// older version gets newly added settings appended via config.SyncSample.
+// logPath activates the LOG_FILE line (Windows); empty leaves it commented.
+func syncEnvFile(path, version, logPath string) error {
+	data, err := os.ReadFile(path)
+	switch {
+	case os.IsNotExist(err):
+		return writeEnvFile(path, config.SampleEnv(version, logPath))
+	case err != nil:
+		return fmt.Errorf("read %s: %w", path, err)
+	}
+	values, perr := config.ParseEnvFile(strings.NewReader(string(data)))
+	if perr == nil && values["CFG_VER"] != "" {
+		synced, changed := config.SyncSample(string(data), version, logPath)
+		if !changed {
+			return nil
+		}
+		return writeEnvFile(path, synced)
+	}
+	// No readable CFG_VER: the file is invalid — keep a backup and start
+	// from a fresh sample.
+	if err := os.WriteFile(path+".bak", data, 0o644); err != nil {
+		return fmt.Errorf("back up %s: %w", path, err)
+	}
+	return writeEnvFile(path, config.SampleEnv(version, logPath))
+}
+
+func writeEnvFile(path, content string) error {
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		return fmt.Errorf("write %s: %w", path, err)
+	}
+	return nil
+}
@@ -6,6 +6,7 @@
 package service

 import (
+	"bytes"
 	"context"
 	"fmt"
 	"io"
@@ -126,12 +127,18 @@ func copyFile(src, dst string, mode os.FileMode) error {
 // copyBin=false the current executable location and config path are
 // registered as-is instead. Needs root.
 //
+// Re-running install converges an existing unit instead of failing: it is
+// stopped first if active, the installed binary is replaced only when the
+// content differs, the unit file is rewritten (followed by daemon-reload)
+// only when it changed, and the service is started again only if it was
+// active before.
+//
 // The binary lives in the StateDirectory rather than /usr/local/sbin on
 // purpose: replacing a running binary needs write access to its
 // *directory*, and granting the sandboxed service user write access to a
 // shared system directory would let a compromised service overwrite other
 // binaries. /var/lib/gpu-turnstile is exclusively ours.
-func Install(configPath string, copyBin bool) error {
+func Install(configPath string, copyBin bool, version string) error {
 	exe, err := os.Executable()
 	if err != nil {
 		return err
@@ -139,6 +146,16 @@ func Install(configPath string, copyBin bool) error {
 	if abs, absErr := filepath.Abs(exe); absErr == nil {
 		exe = abs
 	}
+	unit := Name + ".service"
+	_, statErr := os.Stat(unitPath)
+	fresh := os.IsNotExist(statErr)
+	wasRunning := exec.Command("systemctl", "is-active", "--quiet", unit).Run() == nil
+	if wasRunning {
+		if out, err := exec.Command("systemctl", "stop", unit).CombinedOutput(); err != nil {
+			return fmt.Errorf("systemctl stop (run as root): %w (%s)", err, out)
+		}
+	}
+
 	cfg := etcConfig
 	if copyBin {
 		if err := os.MkdirAll(stateDir, 0o755); err != nil {
@@ -146,33 +163,74 @@ func Install(configPath string, copyBin bool) error {
 		}
 		installedExe := filepath.Join(stateDir, Name)
 		if exe != installedExe {
-			if err := copyFile(exe, installedExe, 0o755); err != nil {
-				return fmt.Errorf("install binary to %s: %w", installedExe, err)
+			if same, _ := sameFileContent(exe, installedExe); !same {
+				if err := copyFile(exe, installedExe, 0o755); err != nil {
+					return fmt.Errorf("install binary to %s: %w", installedExe, err)
+				}
 			}
 		}
 		exe = installedExe
 		if _, err := os.Stat(etcConfig); os.IsNotExist(err) && configPath != "" {
-			// Missing config is not fatal: the service fails fast with a
-			// clear "no consumer URL" error until the user writes one.
 			copyFile(configPath, etcConfig, 0o644) //nolint:errcheck // best effort
 		}
+		// Missing configs get a fully commented sample (LOG_FILE stays
+		// commented — stderr goes to the journal on Linux);
+		// installer-written ones from older versions get new settings
+		// appended; anything without CFG_VER is invalid and gets replaced
+		// (backup kept as .bak).
+		if err := syncEnvFile(etcConfig, version, ""); err != nil {
+			return err
+		}
 	} else if configPath != "" {
 		if abs, absErr := filepath.Abs(configPath); absErr == nil {
 			cfg = abs
 		}
 	}
-	if err := os.WriteFile(unitPath, []byte(renderUnit(exe, cfg)), 0o644); err != nil {
-		return fmt.Errorf("write %s (run as root): %w", unitPath, err)
+	rendered := renderUnit(exe, cfg)
+	if old, _ := os.ReadFile(unitPath); string(old) != rendered {
+		if err := os.WriteFile(unitPath, []byte(rendered), 0o644); err != nil {
+			return fmt.Errorf("write %s (run as root): %w", unitPath, err)
+		}
+		if out, err := exec.Command("systemctl", "daemon-reload").CombinedOutput(); err != nil {
+			return fmt.Errorf("systemctl daemon-reload: %w (%s)", err, out)
+		}
 	}
-	if out, err := exec.Command("systemctl", "daemon-reload").CombinedOutput(); err != nil {
-		return fmt.Errorf("systemctl daemon-reload: %w (%s)", err, out)
+	if fresh {
+		if out, err := exec.Command("systemctl", "enable", "--now", unit).CombinedOutput(); err != nil {
+			return fmt.Errorf("systemctl enable --now: %w (%s)", err, out)
+		}
+		return nil
 	}
-	if out, err := exec.Command("systemctl", "enable", "--now", Name+".service").CombinedOutput(); err != nil {
-		return fmt.Errorf("systemctl enable --now: %w (%s)", err, out)
+	if exec.Command("systemctl", "is-enabled", "--quiet", unit).Run() != nil {
+		if out, err := exec.Command("systemctl", "enable", unit).CombinedOutput(); err != nil {
+			return fmt.Errorf("systemctl enable: %w (%s)", err, out)
+		}
+	}
+	if wasRunning {
+		if out, err := exec.Command("systemctl", "start", unit).CombinedOutput(); err != nil {
+			return fmt.Errorf("systemctl start: %w (%s)", err, out)
+		}
 	}
 	return nil
 }

+// sameFileContent reports whether two files hold identical bytes. A missing
+// destination is simply "different".
+func sameFileContent(a, b string) (bool, error) {
+	ba, err := os.ReadFile(a)
+	if err != nil {
+		return false, err
+	}
+	bb, err := os.ReadFile(b)
+	if os.IsNotExist(err) {
+		return false, nil
+	}
+	if err != nil {
+		return false, err
+	}
+	return bytes.Equal(ba, bb), nil
+}
+
 // Remove stops and disables the service and deletes the unit file and the
 // installed binary. The config file in /etc is left in place (user data).
 func Remove() error {
@@ -29,7 +29,7 @@ func Run(run func(ctx context.Context) error) error {
 }

 // Install is unsupported on non-Windows, non-Linux platforms.
-func Install(string, bool) error { return errUnsupported }
+func Install(string, bool, string) error { return errUnsupported }

 // Remove is unsupported on non-Windows, non-Linux platforms.
 func Remove() error { return errUnsupported }
@@ -8,6 +8,7 @@
 package service

 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
@@ -109,7 +110,13 @@ func (h *handler) Execute(_ []string, requests <-chan svc.ChangeRequest, status
 // config file if it lives elsewhere. The grants must come after
 // CreateService: the virtual account's SID only exists once the service is
 // registered.
-func Install(configPath string, copyBin bool) error {
+//
+// Re-running install on an existing service converges instead of failing:
+// the service is stopped first if running (so the binary can be replaced),
+// the installed binary is refreshed only when the content differs, the
+// registration is updated only where it drifted, and the service is
+// started again only if it was running before.
+func Install(configPath string, copyBin bool, version string) error {
 	exe, err := os.Executable()
 	if err != nil {
 		return err
@@ -123,62 +130,191 @@ func Install(configPath string, copyBin bool) error {
 		}
 	}

-	installDir, _ := installDirs()
-	if copyBin && !strings.EqualFold(filepath.Dir(exe), installDir) {
-		if err := os.MkdirAll(installDir, 0o755); err != nil {
-			return fmt.Errorf("create %s: %w", installDir, err)
-		}
-		installedExe := filepath.Join(installDir, "gpu-turnstile.exe")
-		if err := copyFile(exe, installedExe); err != nil {
-			return fmt.Errorf("copy binary to %s: %w", installedExe, err)
-		}
-		exe = installedExe
-		targetCfg := filepath.Join(installDir, "gpu-turnstile.env")
-		if configPath != "" && !strings.EqualFold(configPath, targetCfg) {
-			if _, statErr := os.Stat(targetCfg); os.IsNotExist(statErr) {
-				copyFile(configPath, targetCfg) //nolint:errcheck // best effort
-			}
-			configPath = targetCfg
-		}
-	}
-
 	m, err := mgr.Connect()
 	if err != nil {
 		return fmt.Errorf("connect to service manager (run as administrator): %w", err)
 	}
 	defer m.Disconnect()

-	binPath := fmt.Sprintf(`"%s" -config "%s"`, exe, configPath)
-	s, err := m.CreateService(Name, binPath, mgr.Config{
-		StartType:        mgr.StartAutomatic,
-		DisplayName:      "gpu-turnstile",
-		Description:      "GPU arbitration proxy for Ollama and ComfyUI",
-		ServiceStartName: virtualAccount,
-	})
-	if err != nil {
-		return fmt.Errorf("create service: %w", err)
+	// An existing service is converged, not an error. Stop it first so the
+	// binary copy can be replaced, and remember whether to start it again.
+	var s *mgr.Service
+	wasRunning := false
+	if existing, openErr := m.OpenService(Name); openErr == nil {
+		s = existing
+		defer s.Close()
+		if st, qErr := s.Query(); qErr == nil &&
+			(st.State == svc.Running || st.State == svc.StartPending) {
+			wasRunning = true
+			if err := stopAndWait(s); err != nil {
+				return err
+			}
+		}
 	}
-	defer s.Close()

-	restart := mgr.RecoveryAction{Type: mgr.ServiceRestart, Delay: 5 * time.Second}
-	if err := s.SetRecoveryActions([]mgr.RecoveryAction{restart, restart, restart}, 24*60*60); err != nil {
+	installDir, dataDir := installDirs()
+	if copyBin {
+		targetCfg := filepath.Join(installDir, "gpu-turnstile.env")
+		if !strings.EqualFold(filepath.Dir(exe), installDir) {
+			if err := os.MkdirAll(installDir, 0o755); err != nil {
+				return fmt.Errorf("create %s: %w", installDir, err)
+			}
+			installedExe := filepath.Join(installDir, "gpu-turnstile.exe")
+			if same, _ := sameFileContent(exe, installedExe); !same {
+				if err := copyFile(exe, installedExe); err != nil {
+					return fmt.Errorf("copy binary to %s: %w", installedExe, err)
+				}
+			}
+			exe = installedExe
+			if configPath != "" && !strings.EqualFold(configPath, targetCfg) {
+				if _, statErr := os.Stat(targetCfg); os.IsNotExist(statErr) {
+					copyFile(configPath, targetCfg) //nolint:errcheck // best effort
+				}
+			}
+		}
+		// A service has no console: without LOG_FILE the output vanishes, so
+		// the sample pre-wires it to ProgramData. Missing configs get a
+		// fresh sample; installer-written ones from older versions get new
+		// settings appended; anything without CFG_VER is invalid and gets
+		// replaced (backup kept as .bak).
+		logPath := filepath.Join(dataDir, "gpu-turnstile.log")
+		if err := syncEnvFile(targetCfg, version, logPath); err != nil {
+			return err
+		}
+		configPath = targetCfg
+	}
+	binPath := fmt.Sprintf(`"%s" -config "%s"`, exe, configPath)
+
+	if s == nil {
+		s, err = m.CreateService(Name, binPath, mgr.Config{
+			StartType:        mgr.StartAutomatic,
+			DisplayName:      "gpu-turnstile",
+			Description:      "GPU arbitration proxy for Ollama and ComfyUI",
+			ServiceStartName: virtualAccount,
+		})
+		if err != nil {
+			return fmt.Errorf("create service: %w", err)
+		}
+		defer s.Close()
+
+		if err := ensureRecovery(s); err != nil {
+			s.Delete() // roll back so a retry starts clean
+			return err
+		}
+		if err := grantAll(exe, configPath); err != nil {
+			s.Delete() // roll back so a retry starts clean
+			return err
+		}
+		// Best effort: start now instead of waiting for the next boot. A
+		// missing config (no consumer URLs) fails the start; the service stays
+		// registered and can be started once the config exists.
+		s.Start()
+		return nil
+	}
+
+	// Existing service: update the registration only where it drifted.
+	cur, err := s.Config()
+	if err != nil {
+		return fmt.Errorf("query service config: %w", err)
+	}
+	const displayName = "gpu-turnstile"
+	const description = "GPU arbitration proxy for Ollama and ComfyUI"
+	if cur.BinaryPathName != binPath ||
+		cur.StartType != mgr.StartAutomatic ||
+		cur.ServiceStartName != virtualAccount ||
+		cur.DisplayName != displayName ||
+		cur.Description != description {
+		upd := cur
+		upd.BinaryPathName = binPath
+		upd.StartType = mgr.StartAutomatic
+		upd.ServiceStartName = virtualAccount
+		upd.DisplayName = displayName
+		upd.Description = description
+		if err := s.UpdateConfig(upd); err != nil {
+			return fmt.Errorf("update service config: %w", err)
+		}
+	}
+	if err := ensureRecovery(s); err != nil {
+		return err
+	}
+	// Idempotent: re-assert the virtual account's ACLs (granting an
+	// existing ACE is a no-op).
+	if err := grantAll(exe, configPath); err != nil {
+		return err
+	}
+	if wasRunning {
+		if err := s.Start(); err != nil {
+			return fmt.Errorf("start service: %w", err)
+		}
+	}
+	return nil
+}
+
+// ensureRecovery makes sure the service restarts 5s after a failure — the
+// mechanism that also brings up a staged update. The current actions are
+// queried first so an up-to-date service is left untouched.
+func ensureRecovery(s *mgr.Service) error {
+	want := mgr.RecoveryAction{Type: mgr.ServiceRestart, Delay: 5 * time.Second}
+	actions, err := s.RecoveryActions()
+	onFailure, flagErr := s.RecoveryActionsOnNonCrashFailures()
+	if err == nil && flagErr == nil && onFailure && len(actions) == 3 {
+		ok := true
+		for _, a := range actions {
+			if a.Type != want.Type || a.Delay != want.Delay {
+				ok = false
+			}
+		}
+		if ok {
+			return nil
+		}
+	}
+	if err := s.SetRecoveryActions([]mgr.RecoveryAction{want, want, want}, 24*60*60); err != nil {
 		return fmt.Errorf("set recovery actions: %w", err)
 	}
 	if err := s.SetRecoveryActionsOnNonCrashFailures(true); err != nil {
 		return fmt.Errorf("set failure actions flag: %w", err)
 	}
-
-	if err := grantAll(exe, configPath); err != nil {
-		s.Delete() // roll back so a retry starts clean
-		return err
-	}
-	// Best effort: start now instead of waiting for the next boot. A
-	// missing config (no consumer URLs) fails the start; the service stays
-	// registered and can be started once the config exists.
-	s.Start()
 	return nil
 }

+// stopAndWait stops the service and waits up to 30s for the stopped state.
+func stopAndWait(s *mgr.Service) error {
+	if _, err := s.Control(svc.Stop); err != nil {
+		return fmt.Errorf("stop service: %w", err)
+	}
+	deadline := time.Now().Add(30 * time.Second)
+	for {
+		st, err := s.Query()
+		if err != nil {
+			return fmt.Errorf("query service: %w", err)
+		}
+		if st.State == svc.Stopped {
+			return nil
+		}
+		if time.Now().After(deadline) {
+			return fmt.Errorf("service did not stop within 30s")
+		}
+		time.Sleep(300 * time.Millisecond)
+	}
+}
+
+// sameFileContent reports whether two files hold identical bytes. A missing
+// destination is simply "different".
+func sameFileContent(a, b string) (bool, error) {
+	ba, err := os.ReadFile(a)
+	if err != nil {
+		return false, err
+	}
+	bb, err := os.ReadFile(b)
+	if os.IsNotExist(err) {
+		return false, nil
+	}
+	if err != nil {
+		return false, err
+	}
+	return bytes.Equal(ba, bb), nil
+}
+
 // copyFile copies src to dst (0755 on the new file).
 func copyFile(src, dst string) error {
 	in, err := os.Open(src)
@@ -374,22 +510,8 @@ func RestartIfRunning() (bool, error) {
 	if st.State != svc.Running && st.State != svc.StartPending {
 		return false, nil
 	}
-	if _, err := s.Control(svc.Stop); err != nil {
-		return false, fmt.Errorf("stop service: %w", err)
-	}
-	deadline := time.Now().Add(30 * time.Second)
-	for {
-		st, err = s.Query()
-		if err != nil {
-			return false, fmt.Errorf("query service: %w", err)
-		}
-		if st.State == svc.Stopped {
-			break
-		}
-		if time.Now().After(deadline) {
-			return false, fmt.Errorf("service did not stop within 30s")
-		}
-		time.Sleep(300 * time.Millisecond)
+	if err := stopAndWait(s); err != nil {
+		return false, err
 	}
 	if err := s.Start(); err != nil {
 		return false, fmt.Errorf("start service: %w", err)
@@ -0,0 +1,273 @@
+// 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"
+	"os/exec"
+	"runtime"
+	"strings"
+	"sync"
+	"time"
+)
+
+// 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
+}
@@ -0,0 +1,192 @@
+package supervise
+
+import (
+	"context"
+	"errors"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"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")
+	}
+}
@@ -26,11 +26,15 @@ import (
 // maxAssetSize bounds release asset downloads.
 const maxAssetSize = 512 << 20

-// Updater checks one Gitea repository for newer releases.
+// Updater checks one Gitea repository for releases.
 type Updater struct {
 	Repo    string // e.g. https://git.rambossek.at/PUBLIC/gpu-turnstile
 	Asset   string // e.g. gpu-turnstile.exe
-	Version string // current version, e.g. v0.1.2 ("dev" disables updates)
+	Version string // current binary version, e.g. v0.1.2 ("dev" cannot be compared)
+	// Desired is the APP_VER policy: "dev" disables updates, "stable" (or
+	// empty) tracks the latest release, anything else is an exact vX.Y.Z
+	// release tag to pin — staging it even when that means a downgrade.
+	Desired string
 	Log     *slog.Logger
 	Client  *http.Client
 }
@@ -167,15 +171,20 @@ func CleanupOld(exePath string) {
 	os.Remove(exePath + ".new")
 }

-// Check performs a single update check. staged is true when a newer,
+// Check performs a single update check. staged is true when a
 // signature-verified binary has been swapped into place at exePath; the
 // caller should then restart the process. A nil error with staged=false
-// means "no action" (up to date, disabled, or dev build); a non-nil error
-// means the check failed and the running binary is untouched.
+// means "no action" (up to date, APP_VER=dev, or no embedded public key);
+// a non-nil error means the check failed and the running binary is
+// untouched.
 func (u *Updater) Check(ctx context.Context, exePath string) (staged bool, err error) {
 	log := u.logger()
-	if u.Version == "" || u.Version == "dev" {
-		log.Debug("auto-update: dev build, skipping")
+	desired := u.Desired
+	if desired == "" {
+		desired = "stable"
+	}
+	if desired == "dev" {
+		log.Debug("auto-update: APP_VER=dev, skipping")
 		return false, nil
 	}
 	if publicKeyPEM == "" {
@@ -187,21 +196,37 @@ func (u *Updater) Check(ctx context.Context, exePath string) (staged bool, err e
 		return false, err
 	}

-	body, err := u.get(ctx, api+"/releases/latest")
+	pinned := desired != "stable"
+	endpoint := api + "/releases/latest"
+	if pinned {
+		endpoint = api + "/releases/tags/" + desired
+	}
+	body, err := u.get(ctx, endpoint)
 	if err != nil {
-		return false, fmt.Errorf("fetch latest release: %w", err)
+		return false, fmt.Errorf("fetch release: %w", err)
 	}
 	var rel release
 	if err := json.Unmarshal(body, &rel); err != nil {
 		return false, fmt.Errorf("parse release: %w", err)
 	}
-	newer, err := newerVersion(u.Version, rel.TagName)
-	if err != nil {
-		return false, err
-	}
-	if !newer {
-		log.Debug("auto-update: up to date", "version", u.Version, "latest", rel.TagName)
-		return false, nil
+	if pinned {
+		// Pin mode: any difference from the target tag means stage it —
+		// including downgrades and replacing a dev binary.
+		if u.Version == rel.TagName {
+			log.Debug("auto-update: already on pinned version", "version", u.Version)
+			return false, nil
+		}
+	} else if u.Version != "" && u.Version != "dev" {
+		// Stable mode: only strictly newer releases count; a dev binary
+		// cannot be compared and is always replaced by the latest release.
+		newer, err := newerVersion(u.Version, rel.TagName)
+		if err != nil {
+			return false, err
+		}
+		if !newer {
+			log.Debug("auto-update: up to date", "version", u.Version, "latest", rel.TagName)
+			return false, nil
+		}
 	}

 	urls := make(map[string]string, len(rel.Assets))
@@ -43,7 +43,7 @@ func newFakeGitea(t *testing.T, tag string, assetContent []byte) *fakeGitea {
 	sign := func() []byte { return ed25519.Sign(priv, f.asset) }

 	mux := http.NewServeMux()
-	mux.HandleFunc("/api/v1/repos/o/r/releases/latest", func(w http.ResponseWriter, r *http.Request) {
+	serveRelease := func(w http.ResponseWriter, r *http.Request) {
 		assets := []map[string]string{
 			{"name": "gpu-turnstile.exe", "browser_download_url": f.srv.URL + "/dl/exe"},
 			{"name": "gpu-turnstile.exe.sha256", "browser_download_url": f.srv.URL + "/dl/sha"},
@@ -52,7 +52,9 @@ func newFakeGitea(t *testing.T, tag string, assetContent []byte) *fakeGitea {
 			assets = append(assets, map[string]string{"name": "gpu-turnstile.exe.sig", "browser_download_url": f.srv.URL + "/dl/sig"})
 		}
 		json.NewEncoder(w).Encode(map[string]any{"tag_name": f.tag, "assets": assets})
-	})
+	}
+	mux.HandleFunc("/api/v1/repos/o/r/releases/latest", serveRelease)
+	mux.HandleFunc("/api/v1/repos/o/r/releases/tags/"+f.tag, serveRelease)
 	mux.HandleFunc("/dl/exe", func(w http.ResponseWriter, r *http.Request) { w.Write(f.asset) })
 	mux.HandleFunc("/dl/sig", func(w http.ResponseWriter, r *http.Request) {
 		sig := sign()
@@ -73,6 +75,12 @@ func (f *fakeGitea) updater(version string) *Updater {
 	return &Updater{Repo: f.srv.URL + "/o/r", Asset: "gpu-turnstile.exe", Version: version}
 }

+func (f *fakeGitea) updaterDesired(version, desired string) *Updater {
+	u := f.updater(version)
+	u.Desired = desired
+	return u
+}
+
 func fakeExe(t *testing.T) string {
 	t.Helper()
 	exe := filepath.Join(t.TempDir(), "gpu-turnstile.exe")
@@ -153,12 +161,61 @@ func TestCheckSkipsWithoutPublicKey(t *testing.T) {
 	}
 }

-func TestCheckSkipsDevBuild(t *testing.T) {
+func TestCheckDevBuildGetsStable(t *testing.T) {
+	// A dev binary cannot be compared; APP_VER=stable replaces it with the
+	// latest release.
 	f := newFakeGitea(t, "v9.9.9", []byte("new-binary"))
 	withPublicKey(t, f.pubPEM)
-	staged, err := f.updater("dev").Check(context.Background(), fakeExe(t))
+	exe := fakeExe(t)
+	staged, err := f.updater("dev").Check(context.Background(), exe)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !staged {
+		t.Fatal("expected dev binary to be replaced by the latest release")
+	}
+	content, _ := os.ReadFile(exe)
+	if string(content) != "new-binary" {
+		t.Fatalf("exe content = %q", content)
+	}
+}
+
+func TestCheckDesiredDevDisables(t *testing.T) {
+	f := newFakeGitea(t, "v9.9.9", []byte("new-binary"))
+	withPublicKey(t, f.pubPEM)
+	for _, version := range []string{"dev", "v0.1.2"} {
+		staged, err := f.updaterDesired(version, "dev").Check(context.Background(), fakeExe(t))
+		if err != nil || staged {
+			t.Fatalf("version %s: staged=%v err=%v, want no action with APP_VER=dev", version, staged, err)
+		}
+	}
+}
+
+func TestCheckPinned(t *testing.T) {
+	// Pin mode stages the exact tag — up or down — and skips when the
+	// binary already matches.
+	for _, version := range []string{"v0.1.2", "v9.9.9", "dev"} {
+		f := newFakeGitea(t, "v0.5.0", []byte("pinned-binary"))
+		withPublicKey(t, f.pubPEM)
+		exe := fakeExe(t)
+		staged, err := f.updaterDesired(version, "v0.5.0").Check(context.Background(), exe)
+		if err != nil {
+			t.Fatalf("version %s: %v", version, err)
+		}
+		if !staged {
+			t.Fatalf("version %s: expected pinned v0.5.0 to be staged", version)
+		}
+		content, _ := os.ReadFile(exe)
+		if string(content) != "pinned-binary" {
+			t.Fatalf("version %s: exe content = %q", version, content)
+		}
+	}
+
+	f := newFakeGitea(t, "v0.5.0", []byte("pinned-binary"))
+	withPublicKey(t, f.pubPEM)
+	staged, err := f.updaterDesired("v0.5.0", "v0.5.0").Check(context.Background(), fakeExe(t))
 	if err != nil || staged {
-		t.Fatalf("staged=%v err=%v, want no action for dev build", staged, err)
+		t.Fatalf("staged=%v err=%v, want no action when already on the pinned version", staged, err)
 	}
 }
Author	SHA1	Message	Date
mram	482731ed9e	Pin compose example to v0.1.8 ci / test (push) Successful in 14s Details ci / docker (push) Successful in 1m11s Details ci / release (push) Successful in 15s Details	2026-09-21 17:19:16 +02:00
mram	e4bdc92ece	Game detection: foreign GPU holders take an external lock hold (GAME_PROCS, GPU_FOREIGN_VRAM_MB)	2026-09-21 17:16:56 +02:00
mram	14120bf4a4	Supervisor probes before spawning: an external server on the port is used, never fought or killed	2026-09-21 16:06:28 +02:00
mram	d7566329ae	Managed ComfyUI: COMFY_CMD starts it on demand, idle stop frees VRAM (internal/supervise)	2026-09-21 13:48:07 +02:00
mram	e43ad02fc4	Ignore staged-update leftovers next to the dev exe	2026-09-21 13:14:33 +02:00
mram	232f5b61f2	Periodic upstream health checks (HEALTH_INTERVAL, default 30s); log down/recovered transitions	2026-09-21 13:14:18 +02:00
mram	5e7a042cad	CFG_VER is always a concrete version: dev builds stamp v0.0.0, never "dev"	2026-09-21 12:42:32 +02:00
mram	14c2e30478	Clarify listener wording: Ollama-compatible clients, not Ollama-facing	2026-09-21 12:38:03 +02:00
mram	dcdb2cc72e	Pin compose example to v0.1.7 ci / test (push) Successful in 13s Details ci / docker (push) Successful in 1m7s Details ci / release (push) Successful in 16s Details	2026-09-21 10:04:29 +02:00
mram	44a8e9fbdc	CFG_VER/APP_VER in the env file: invalid configs replaced, updates follow APP_VER (dev/stable/pin)	2026-09-21 09:28:29 +02:00
mram	272a3a467d	force-update: single-shot with a 30s fail-fast timeout	2026-09-21 09:09:27 +02:00
mram	cedf28a809	force-update does not require consumer URLs (ErrNoConsumer sentinel)	2026-09-21 09:05:46 +02:00
mram	e46e92bee8	Sample env carries a version marker; install appends settings missing since the writing version	2026-09-21 08:54:47 +02:00
mram	cf48796183	Install writes a fully commented sample env file when no config exists	2026-09-21 08:48:51 +02:00
mram	dd3187f951	Install sets a default LOG_FILE in the env file on Windows (no console as a service)	2026-09-21 08:37:25 +02:00
mram	1394a76eea	Make service install converge: stop running service, refresh binary/config only on change, restart only if it was running	2026-09-21 08:30:55 +02:00