Implement balance monitoring and notification for Selectel API integration

- Added support for balance tracking using `SELECTEL_API_TOKEN` in the configuration.
- Introduced new balance notification logic in the local watchdog, alerting users on balance changes based on defined thresholds.
- Updated documentation to include instructions for setting up balance notifications and the required environment variables.
- Enhanced the `ready` and `session` modules to initialize balance state and handle notifications during GPU operations.
- Refactored the CLI and related components to support the new balance monitoring features, ensuring a seamless user experience.

docs/cli.md

@@ -231,8 +231,9 @@ Application credential для idle-killer CLI создаёт на `up` (узко
 
 1. Nova `ACTIVE`
 2. TCP 22 / SSH
-3. HTTP `http://127.0.0.1:7801` **на VM** (через SSH)
-4. Backend Idle → toast (если `NOTIFY_READY`) + access-card
+3. На VM: HTTP сервисов стека (SwarmUI `:7801` / Ollama `:11434` / llama.cpp `:8080`) — `verify_stack_on_vm`
+4. Backend Idle (если SwarmUI) → toast (если `NOTIFY_READY`)
+5. Туннель + проверка **localhost** тех же сервисов → access-card
 
 Локальный порт UI: **17801** (на VM по-прежнему 7801 на loopback).
 
@@ -241,3 +242,13 @@ Application credential для idle-killer CLI создаёт на `up` (узко
 ## Windows / notify
 
 `NOTIFY_READY`: toast + системный звук, когда backend Idle. Если toast недоступен — только звук и лог, без падения CLI.
+
+### Баланс Selectel (шаг 200 ₽)
+
+Нужны:
+
+1. `SELECTEL_API_TOKEN` в `.env` — **статический ключ панели** (`X-Token`), не OpenStack password ([баланс API](https://docs.selectel.ru/api/balance/)).
+2. `gpu-rent watchdog install` — тот же локальный тикер, что и аварийный stop.
+
+На `up` запоминается baseline баланса. Каждый tick watchdog: если с `up` ушло ещё ~200 ₽ (`BALANCE_NOTIFY_STEP_RUB`) — toast «Списано ~N ₽…». Пополнение сбрасывает baseline. Опционально `BALANCE_NOTIFY_LOW_RUB=300` — разовое «баланс низкий».
+

docs/decisions.md

@@ -26,6 +26,7 @@
 | llm-only | `ENABLE_SWARMUI=false` / `WORKLOAD=llm` / `--no-swarm`/`--llm-only`: GPU + LLM без SwarmUI (bootstrap только data disk) |
 | Capture | `gpu-rent capture` — инвентарь VM → merge **ссылок** в локальные yaml (веса не качать) |
 | Perf auto-tune | На `up`: probe GPU → tier. Swarm/Comfy: sageattention ExtraArgs на Ampere+ ≥16 GiB. Ollama: flash/KV/keep-alive + `GPU_OVERHEAD` чтобы оставить VRAM под Krea |
+| Баланс ₽ | `SELECTEL_API_TOKEN` (X-Token панели) + `gpu-rent watchdog install`. Baseline на `up`; тик watchdog уведомляет каждые `BALANCE_NOTIFY_STEP_RUB` (дефолт **200**) списанных с baseline. Опционально `BALANCE_NOTIFY_LOW_RUB` |
 | Data-диск | Старт **100 GB**, рост через resize вверх (вниз Selectel не умеет) |
 | SSH | CLI генерирует `<repo>/.gpu-rent/id_ed25519` без passphrase и сам регистрирует keypair |
 | Локальные файлы | Всё в корне репозитория: `.env`, `models.yaml`, `extensions.yaml`; runtime (`state.json`, lock, SSH) в `<repo>/.gpu-rent/`. Не `%USERPROFILE%\.gpu-rent` |

docs/selectel.md

@@ -19,6 +19,8 @@ GPU есть не во всех сегментах. Перед create — [ма
 
 Статический API-ключ панели (`X-Token`) **не управляет** объектами OpenStack (серверы, диски, сети).
 
+Исключение: **баланс аккаунта** (`GET https://api.selectel.ru/v3/balances`) — как раз через `X-Token`. Для уведомлений о списании положи `SELECTEL_API_TOKEN` в `.env` и поставь `gpu-rent watchdog install` (см. [cli.md](cli.md)).
+
 Нужен **сервисный пользователь** с правом на проект.
 
 1. CLI хранит: account id, username, password, project id/name, pool, pool segment.

docs/setup.md

@@ -233,7 +233,7 @@ copy models.example.yaml models.yaml
 1. Короткий doctor (полный — `up -v`).
 2. Create/unshelve GPU + диски (после confirm).
 3. Bootstrap SwarmUI, extensions, autocomplete, Civitai-seed, push локальных папок; optional LLM.
-4. Туннель на `localhost:17801`, access-card с URL / MCP.
+4. Туннель на `localhost:17801` (или LLM-порт), **проверка** что сервисы отвечают, access-card.
 5. Процесс ждёт: **Ctrl+C** закрывает только туннель, GPU остаётся.
 
 Полезные флаги:

env.example

@@ -63,3 +63,8 @@ IDLE_GRACE_MINUTES=45
 # LOCAL_WATCHDOG_GRACE_MINUTES=10
 PULL_OUTPUT=false
 NOTIFY_READY=true
+# Статический API-ключ панели (X-Token) — только для баланса, не для OpenStack:
+# SELECTEL_API_TOKEN=
+# Уведомления о списании (local-watchdog tick), шаг с момента up:
+# BALANCE_NOTIFY_STEP_RUB=200
+# BALANCE_NOTIFY_LOW_RUB=0

src/gpu_rent/balance.py

@@ -0,0 +1,174 @@
+"""Selectel account balance watch: notify every N ₽ spent since up."""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Callable
+from dataclasses import asdict, dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from gpu_rent.errors import CloudError
+from gpu_rent.paths import runtime_dir
+
+Log = Callable[[str], None]
+
+BALANCES_URL = "https://api.selectel.ru/v3/balances"
+DEFAULT_STEP_RUB = 200.0
+
+
+@dataclass
+class BalanceWatchState:
+    baseline_rub: float
+    last_notified_step: int = 0
+    low_notified: bool = False
+    last_balance_rub: float | None = None
+    updated_at: str | None = None
+
+    def to_dict(self) -> dict[str, Any]:
+        return asdict(self)
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> BalanceWatchState:
+        return cls(
+            baseline_rub=float(data.get("baseline_rub") or 0),
+            last_notified_step=int(data.get("last_notified_step") or 0),
+            low_notified=bool(data.get("low_notified")),
+            last_balance_rub=(
+                float(data["last_balance_rub"])
+                if data.get("last_balance_rub") is not None
+                else None
+            ),
+            updated_at=data.get("updated_at"),
+        )
+
+
+def balance_watch_path() -> Path:
+    return runtime_dir() / "balance-watch.json"
+
+
+def utc_now_iso() -> str:
+    return datetime.now(timezone.utc).replace(microsecond=0).isoformat()
+
+
+def load_balance_state() -> BalanceWatchState | None:
+    path = balance_watch_path()
+    if not path.is_file():
+        return None
+    try:
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None
+    if not isinstance(raw, dict) or "baseline_rub" not in raw:
+        return None
+    return BalanceWatchState.from_dict(raw)
+
+
+def save_balance_state(state: BalanceWatchState) -> None:
+    runtime_dir().mkdir(parents=True, exist_ok=True)
+    state.updated_at = utc_now_iso()
+    balance_watch_path().write_text(
+        json.dumps(state.to_dict(), indent=2) + "\n", encoding="utf-8"
+    )
+
+
+def clear_balance_state() -> None:
+    path = balance_watch_path()
+    if path.is_file():
+        path.unlink()
+
+
+def balance_rub_from_payload(payload: dict[str, Any]) -> float:
+    """Selectel returns money as integer kopecks → rubles."""
+    data = payload.get("data") if isinstance(payload.get("data"), dict) else payload
+    billings = data.get("billings") if isinstance(data, dict) else None
+    if not isinstance(billings, list) or not billings:
+        raise CloudError("Selectel balances: нет data.billings")
+    total_kop = 0
+    for row in billings:
+        if not isinstance(row, dict):
+            continue
+        if row.get("balances_values_sum") is not None:
+            total_kop += int(row["balances_values_sum"])
+        elif row.get("final_sum") is not None:
+            total_kop += int(row["final_sum"])
+    return total_kop / 100.0
+
+
+def fetch_balance_rub(api_token: str, *, timeout: float = 20.0) -> float:
+    token = (api_token or "").strip()
+    if not token:
+        raise CloudError("SELECTEL_API_TOKEN пуст — баланс недоступен")
+    headers = {"X-Token": token, "Accept": "application/json"}
+    try:
+        with httpx.Client(timeout=timeout, follow_redirects=True) as client:
+            resp = client.get(BALANCES_URL, headers=headers)
+    except httpx.HTTPError as exc:
+        raise CloudError(f"Selectel balances сеть: {exc}") from exc
+    if resp.status_code in {401, 403}:
+        raise CloudError(
+            f"Selectel balances HTTP {resp.status_code} — нужен статический "
+            "API-ключ аккаунта (X-Token), не OpenStack пароль"
+        )
+    if resp.status_code != 200:
+        raise CloudError(f"Selectel balances HTTP {resp.status_code}: {resp.text[:200]}")
+    try:
+        payload = resp.json()
+    except json.JSONDecodeError as exc:
+        raise CloudError("Selectel balances: не JSON") from exc
+    if not isinstance(payload, dict):
+        raise CloudError("Selectel balances: неожиданный JSON")
+    return balance_rub_from_payload(payload)
+
+
+def spend_steps(baseline_rub: float, current_rub: float, step_rub: float) -> int:
+    """How many full step intervals have been spent since baseline."""
+    if step_rub <= 0:
+        return 0
+    spend = baseline_rub - current_rub
+    if spend < 0:
+        return 0
+    return int(spend // step_rub)
+
+
+def plan_balance_notices(
+    state: BalanceWatchState,
+    current_rub: float,
+    *,
+    step_rub: float = DEFAULT_STEP_RUB,
+    low_rub: float = 0.0,
+) -> tuple[BalanceWatchState, list[str]]:
+    """
+    Update state for top-ups / spend steps / low watermark.
+    Returns (new_state, human messages to show).
+    """
+    messages: list[str] = []
+    # Top-up: reset baseline so steps restart from new balance.
+    if current_rub > state.baseline_rub + 0.01:
+        state = BalanceWatchState(
+            baseline_rub=current_rub,
+            last_notified_step=0,
+            low_notified=False,
+            last_balance_rub=current_rub,
+        )
+        messages.append(f"Баланс пополнен: {current_rub:.0f} ₽ (новый baseline)")
+        return state, messages
+
+    crossed = spend_steps(state.baseline_rub, current_rub, step_rub)
+    while state.last_notified_step < crossed:
+        state.last_notified_step += 1
+        spent = state.last_notified_step * step_rub
+        messages.append(
+            f"Списано ~{spent:.0f} ₽ с момента up "
+            f"(осталось {current_rub:.0f} ₽, шаг {step_rub:.0f} ₽)"
+        )
+
+    if low_rub > 0 and current_rub < low_rub and not state.low_notified:
+        state.low_notified = True
+        messages.append(f"Баланс низкий: {current_rub:.0f} ₽ (< {low_rub:.0f} ₽)")
+
+    state.last_balance_rub = current_rub
+    return state, messages

src/gpu_rent/config.py

@@ -51,6 +51,12 @@ def _as_int(value: str | None, default: int) -> int:
     return int(value)
 
 
+def _as_float(value: str | None, default: float) -> float:
+    if value is None or str(value).strip() == "":
+        return default
+    return float(str(value).strip().replace(",", "."))
+
+
 def _csv(value: str | None, default: tuple[str, ...]) -> tuple[str, ...]:
     if value is None or value.strip() == "":
         return default
@@ -114,6 +120,9 @@ class Config:
     idle_grace_minutes: int
     pull_output: bool
     notify_ready: bool
+    selectel_api_token: str
+    balance_notify_step_rub: float
+    balance_notify_low_rub: float
 
     missing: list[str] = field(default_factory=list)
 
@@ -245,5 +254,8 @@ def load_config(*, require_auth: bool = True) -> Config:
         idle_grace_minutes=_as_int(os.environ.get("IDLE_GRACE_MINUTES"), 45),
         pull_output=_as_bool(os.environ.get("PULL_OUTPUT"), False),
         notify_ready=_as_bool(os.environ.get("NOTIFY_READY"), True),
+        selectel_api_token=(os.environ.get("SELECTEL_API_TOKEN") or "").strip(),
+        balance_notify_step_rub=_as_float(os.environ.get("BALANCE_NOTIFY_STEP_RUB"), 200.0),
+        balance_notify_low_rub=_as_float(os.environ.get("BALANCE_NOTIFY_LOW_RUB"), 0.0),
         missing=missing,
     )

src/gpu_rent/local_watchdog.py

@@ -271,7 +271,8 @@ def install_watchdog(
     log(
         f"local-watchdog установлен ({platform}): тик каждые {interval} мин. "
         f"Grace {grace_seconds() // 60} мин после смерти процесса туннеля → stop. "
-        f"Ctrl+C на туннеле GPU не гасит."
+        f"Ctrl+C на туннеле GPU не гасит. "
+        "При SELECTEL_API_TOKEN — toast каждые BALANCE_NOTIFY_STEP_RUB ₽ (дефолт 200)."
     )
     return marker
 
@@ -314,9 +315,64 @@ def watchdog_status_lines() -> list[str]:
             f"pid={lease.pid} alive={alive} hb={lease.heartbeat_at}"
         )
     lines.append(f"grace: {grace_seconds() // 60} мин (LOCAL_WATCHDOG_GRACE_MINUTES)")
+    try:
+        from gpu_rent.balance import load_balance_state
+
+        bal = load_balance_state()
+        if bal:
+            lines.append(
+                f"balance: baseline {bal.baseline_rub:.0f} ₽ "
+                f"last={bal.last_balance_rub} step_notified={bal.last_notified_step}"
+            )
+        else:
+            lines.append("balance: нет baseline (появится на up при SELECTEL_API_TOKEN)")
+    except Exception:
+        pass
     return lines
 
 
+def check_balance_on_tick(cfg: Config, *, log: Log = print) -> list[str]:
+    """Fetch balance; toast on each BALANCE_NOTIFY_STEP_RUB spent since baseline.
+
+    No-op without SELECTEL_API_TOKEN. Arms baseline on first successful fetch
+    if missing but a GPU session may be running.
+    """
+    from gpu_rent.balance import (
+        BalanceWatchState,
+        fetch_balance_rub,
+        load_balance_state,
+        plan_balance_notices,
+        save_balance_state,
+    )
+    from gpu_rent.notify import notify_message
+
+    token = (cfg.selectel_api_token or "").strip()
+    if not token:
+        return []
+    step = float(cfg.balance_notify_step_rub or 200.0)
+    low = float(cfg.balance_notify_low_rub or 0.0)
+    try:
+        current = fetch_balance_rub(token)
+    except Exception as exc:
+        log(f"balance: {exc}")
+        return []
+
+    state = load_balance_state()
+    if state is None:
+        state = BalanceWatchState(baseline_rub=current, last_balance_rub=current)
+        save_balance_state(state)
+        log(f"balance: baseline {current:.0f} ₽ (шаг {step:.0f} ₽)")
+        return []
+
+    state, messages = plan_balance_notices(
+        state, current, step_rub=step, low_rub=low
+    )
+    save_balance_state(state)
+    for msg in messages:
+        notify_message("gpu-rent баланс", msg, log)
+    return messages
+
+
 def run_tick(*, dry_run: bool = False, log: Log = print) -> TickDecision:
     from gpu_rent.session import cmd_stop
     from gpu_rent.state import load_state
@@ -331,6 +387,18 @@ def run_tick(*, dry_run: bool = False, log: Log = print) -> TickDecision:
         process_alive=pid_alive(lease.pid if lease else None),
         grace_sec=grace_seconds(),
     )
+
+    # Balance steps run on every tick while watchdog is installed (same timer).
+    if watchdog_installed() and not dry_run:
+        try:
+            from gpu_rent.config import load_config
+
+            cfg = load_config(require_auth=False)
+            if state.server_id or (lease and lease.armed):
+                check_balance_on_tick(cfg, log=log)
+        except Exception as exc:
+            log(f"balance tick: {exc}")
+
     if decision.kind == "noop":
         log(f"watchdog tick: noop ({decision.detail})")
         return decision

src/gpu_rent/notify.py

@@ -1,4 +1,4 @@
-"""Ready notification: toast + sound. Never raises."""
+"""Ready / balance notifications: toast + sound. Never raises."""
 
 from __future__ import annotations
 
@@ -10,7 +10,12 @@ from gpu_rent.config import Config
 
 Log = Callable[[str], None]
 
-__all__ = ["notify_ready", "print_mcp_snippet", "print_access_card"]
+__all__ = [
+    "notify_ready",
+    "notify_message",
+    "print_mcp_snippet",
+    "print_access_card",
+]
 
 
 def notify_ready(cfg: Config, log: Log) -> None:
@@ -19,7 +24,15 @@ def notify_ready(cfg: Config, log: Log) -> None:
     log("NOTIFY_READY: SwarmUI Idle")
     _sound()
     if sys.platform == "win32":
-        _windows_toast(log)
+        _windows_toast("SwarmUI backend Idle — gpu-rent tunnel", log)
+
+
+def notify_message(title: str, body: str, log: Log) -> None:
+    """Generic toast/sound for balance steps etc."""
+    log(f"{title}: {body}")
+    _sound()
+    if sys.platform == "win32":
+        _windows_toast(body, log, title=title)
 
 
 def _sound() -> None:
@@ -38,15 +51,17 @@ def _sound() -> None:
         pass
 
 
-def _windows_toast(log: Log) -> None:
+def _windows_toast(body: str, log: Log, *, title: str = "gpu-rent") -> None:
+    safe_title = title.replace("'", "''")
+    safe_body = body.replace("'", "''")
     script = (
         "[Windows.UI.Notifications.ToastNotificationManager, Windows.UI.Notifications, "
         "ContentType = WindowsRuntime] > $null; "
         "$template = [Windows.UI.Notifications.ToastNotificationManager]::GetTemplateContent("
         "[Windows.UI.Notifications.ToastTemplateType]::ToastText02); "
         "$text = $template.GetElementsByTagName('text'); "
-        "$text.Item(0).AppendChild($template.CreateTextNode('gpu-rent')) | Out-Null; "
-        "$text.Item(1).AppendChild($template.CreateTextNode('SwarmUI backend Idle — gpu-rent tunnel')) | Out-Null; "
+        f"$text.Item(0).AppendChild($template.CreateTextNode('{safe_title}')) | Out-Null; "
+        f"$text.Item(1).AppendChild($template.CreateTextNode('{safe_body}')) | Out-Null; "
         "$toast = [Windows.UI.Notifications.ToastNotification]::new($template); "
         "[Windows.UI.Notifications.ToastNotificationManager]::CreateToastNotifier('gpu-rent').Show($toast)"
     )

src/gpu_rent/ready.py

@@ -1,12 +1,21 @@
-"""Wait until SwarmUI HTTP is up and backend is Idle (on the VM)."""
+"""Wait until SwarmUI HTTP is up and backend is Idle (on the VM).
+
+Also end-of-up stack verification: every enabled service must answer.
+"""
 
 from __future__ import annotations
 
+import json
+import socket
 import time
+import urllib.error
+import urllib.request
 from collections.abc import Callable
+from dataclasses import dataclass
 
 from gpu_rent.config import Config
 from gpu_rent.errors import CloudError
+from gpu_rent.llm_runtime import normalize_runtime
 from gpu_rent.ssh_ops import run_ssh
 
 Log = Callable[[str], None]
@@ -57,6 +66,93 @@ while time.time() < deadline:
 print("WAIT timeout-slice")
 """
 
+# One-shot probe of configured stack endpoints on the VM (JSON line).
+_REMOTE_STACK_PROBE = r'''
+import json, urllib.error, urllib.request, subprocess
+
+def http_ok(url, timeout=4.0):
+    try:
+        req = urllib.request.Request(url, method="GET")
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            code = getattr(resp, "status", 200) or 200
+            body = resp.read(256)
+            return True, f"HTTP {code} ({len(body)}b)"
+    except Exception as exc:
+        return False, str(exc)[:160]
+
+def unit_active(name):
+    try:
+        out = subprocess.check_output(
+            ["systemctl", "is-active", name],
+            text=True,
+            stderr=subprocess.DEVNULL,
+        ).strip()
+        return out
+    except Exception:
+        return "unknown"
+
+checks = []
+want_swarm = WANT_SWARM
+want_ollama = WANT_OLLAMA
+want_llama = WANT_LLAMA
+
+if want_swarm:
+    ok, detail = http_ok("http://127.0.0.1:7801/")
+    if not ok:
+        # API may answer when / does not
+        ok2, d2 = False, detail
+        try:
+            req = urllib.request.Request(
+                "http://127.0.0.1:7801/API/GetNewSession",
+                data=b"{}",
+                headers={"Content-Type": "application/json"},
+                method="POST",
+            )
+            with urllib.request.urlopen(req, timeout=5) as resp:
+                ok2 = True
+                d2 = f"API session HTTP {getattr(resp, 'status', 200)}"
+        except Exception as exc:
+            d2 = str(exc)[:160]
+        ok, detail = ok2, d2
+    checks.append({
+        "name": "swarmui",
+        "ok": ok,
+        "detail": detail,
+        "unit": unit_active("swarmui"),
+    })
+
+if want_ollama:
+    ok, detail = http_ok("http://127.0.0.1:11434/api/tags")
+    checks.append({
+        "name": "ollama",
+        "ok": ok,
+        "detail": detail,
+        "unit": unit_active("gpu-rent-ollama"),
+    })
+
+if want_llama:
+    ok, detail = http_ok("http://127.0.0.1:8080/health")
+    if not ok:
+        ok2, d2 = http_ok("http://127.0.0.1:8080/v1/models")
+        ok, detail = ok2, d2
+    checks.append({
+        "name": "llamacpp",
+        "ok": ok,
+        "detail": detail,
+        "unit": unit_active("gpu-rent-llamacpp"),
+    })
+
+print(json.dumps({"checks": checks}, ensure_ascii=False))
+'''
+
+
+@dataclass(frozen=True)
+class ServiceCheck:
+    name: str
+    ok: bool
+    detail: str
+    where: str = "vm"  # vm | local
+
 
 def wait_backend_idle(
     cfg: Config,
@@ -93,3 +189,201 @@ def wait_backend_idle(
         f"backend не стал Idle за {int(timeout)} с. "
         "Проверь journalctl -u swarmui на VM; GPU всё ещё жив."
     )
+
+
+def _expected_services(cfg: Config) -> tuple[bool, bool, bool]:
+    swarm = bool(getattr(cfg, "enable_swarmui", True))
+    rt = normalize_runtime(getattr(cfg, "llm_runtime", "none"))
+    return swarm, rt == "ollama", rt == "llamacpp"
+
+
+def _probe_vm_once(cfg: Config, host: str) -> list[ServiceCheck]:
+    want_swarm, want_ollama, want_llama = _expected_services(cfg)
+    script = (
+        _REMOTE_STACK_PROBE.replace("WANT_SWARM", "True" if want_swarm else "False")
+        .replace("WANT_OLLAMA", "True" if want_ollama else "False")
+        .replace("WANT_LLAMA", "True" if want_llama else "False")
+    )
+    out = run_ssh(
+        cfg,
+        host,
+        "python3 - <<'PY'\n" + script + "\nPY",
+        check=False,
+        timeout=60,
+    ).strip()
+    line = ""
+    for row in reversed(out.splitlines()):
+        row = row.strip()
+        if row.startswith("{"):
+            line = row
+            break
+    if not line:
+        return [
+            ServiceCheck("stack", False, f"нет JSON от probe: {out[-200:]}", "vm")
+        ]
+    try:
+        data = json.loads(line)
+    except json.JSONDecodeError:
+        return [ServiceCheck("stack", False, f"битый JSON: {line[:200]}", "vm")]
+    checks: list[ServiceCheck] = []
+    for item in data.get("checks") or []:
+        if not isinstance(item, dict):
+            continue
+        name = str(item.get("name") or "?")
+        detail = str(item.get("detail") or "")
+        unit = str(item.get("unit") or "")
+        if unit and unit != "unknown":
+            detail = f"{detail}; unit={unit}"
+        checks.append(
+            ServiceCheck(name=name, ok=bool(item.get("ok")), detail=detail, where="vm")
+        )
+    return checks
+
+
+def verify_stack_on_vm(
+    cfg: Config,
+    host: str,
+    log: Log,
+    *,
+    timeout: float = 300.0,
+    poll_every: float = 8.0,
+    raise_on_fail: bool = True,
+) -> list[ServiceCheck]:
+    """Poll until every enabled service answers on the VM loopback."""
+    want_swarm, want_ollama, want_llama = _expected_services(cfg)
+    if not (want_swarm or want_ollama or want_llama):
+        log("проверка стека: нечего ждать (swarm off, LLM none)")
+        return []
+
+    names = []
+    if want_swarm:
+        names.append("SwarmUI :7801")
+    if want_ollama:
+        names.append("Ollama :11434")
+    if want_llama:
+        names.append("llama.cpp :8080")
+    log(f"проверка на VM: {', '.join(names)}")
+
+    deadline = time.time() + timeout
+    last: list[ServiceCheck] = []
+    while time.time() < deadline:
+        try:
+            last = _probe_vm_once(cfg, host)
+        except Exception as exc:
+            last = [ServiceCheck("ssh", False, str(exc)[:200], "vm")]
+        if last and all(c.ok for c in last):
+            for c in last:
+                log(f"  [ok] {c.name}: {c.detail}")
+            log("проверка VM: всё отвечает")
+            return last
+        bad = ", ".join(f"{c.name}={c.detail}" for c in last if not c.ok) or "пусто"
+        log(f"  … ещё нет: {bad}")
+        time.sleep(poll_every)
+
+    for c in last:
+        mark = "ok" if c.ok else "FAIL"
+        log(f"  [{mark}] {c.name}: {c.detail}")
+    if raise_on_fail and any(not c.ok for c in last):
+        failed = [c.name for c in last if not c.ok]
+        raise CloudError(
+            f"сервисы не ответили на VM за {int(timeout)} с: {', '.join(failed)}. "
+            "GPU жив — смотри journalctl / gpu-rent logs"
+        )
+    return last
+
+
+def _tcp_ok(port: int, host: str = "127.0.0.1", timeout: float = 0.8) -> bool:
+    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    sock.settimeout(timeout)
+    try:
+        return sock.connect_ex((host, port)) == 0
+    finally:
+        sock.close()
+
+
+def _http_local(url: str, timeout: float = 3.0) -> tuple[bool, str]:
+    try:
+        req = urllib.request.Request(url, method="GET")
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            code = getattr(resp, "status", 200) or 200
+            return True, f"HTTP {code}"
+    except Exception as exc:
+        return False, str(exc)[:160]
+
+
+def verify_stack_local(
+    cfg: Config,
+    log: Log,
+    *,
+    timeout: float = 60.0,
+    poll_every: float = 2.0,
+    raise_on_fail: bool = True,
+) -> list[ServiceCheck]:
+    """After tunnel: local ports + light HTTP for enabled services."""
+    want_swarm, want_ollama, want_llama = _expected_services(cfg)
+    targets: list[tuple[str, int, str | None]] = []
+    if want_swarm:
+        targets.append(("swarmui", int(cfg.swarmui_local_port), None))
+    if want_ollama:
+        targets.append(
+            ("ollama", int(cfg.ollama_local_port), f"http://127.0.0.1:{cfg.ollama_local_port}/api/tags")
+        )
+    if want_llama:
+        p = int(cfg.llamacpp_local_port)
+        targets.append(("llamacpp", p, f"http://127.0.0.1:{p}/health"))
+
+    if not targets:
+        return []
+
+    log("проверка туннеля (localhost): " + ", ".join(f"{n}:{port}" for n, port, _ in targets))
+    deadline = time.time() + timeout
+    last: list[ServiceCheck] = []
+    while time.time() < deadline:
+        last = []
+        for name, port, url in targets:
+            if not _tcp_ok(port):
+                last.append(ServiceCheck(name, False, f"порт {port} закрыт", "local"))
+                continue
+            if url:
+                ok, detail = _http_local(url)
+                if not ok and name == "llamacpp":
+                    ok, detail = _http_local(
+                        f"http://127.0.0.1:{port}/v1/models"
+                    )
+                last.append(ServiceCheck(name, ok, detail, "local"))
+            else:
+                # SwarmUI: TCP enough (API may need POST); try API too
+                ok, detail = _http_local(f"http://127.0.0.1:{port}/")
+                if not ok:
+                    try:
+                        req = urllib.request.Request(
+                            f"http://127.0.0.1:{port}/API/GetNewSession",
+                            data=b"{}",
+                            headers={"Content-Type": "application/json"},
+                            method="POST",
+                        )
+                        with urllib.request.urlopen(req, timeout=4) as resp:
+                            ok = True
+                            detail = f"API HTTP {getattr(resp, 'status', 200)}"
+                    except Exception as exc:
+                        detail = f"TCP ok; HTTP {exc}"[:160]
+                        # TCP open is enough for swarm local check
+                        ok = True
+                        detail = f"TCP :{port} open"
+                last.append(ServiceCheck(name, ok, detail, "local"))
+        if last and all(c.ok for c in last):
+            for c in last:
+                log(f"  [ok] localhost {c.name}: {c.detail}")
+            log("проверка туннеля: всё доступно")
+            return last
+        time.sleep(poll_every)
+
+    for c in last:
+        mark = "ok" if c.ok else "FAIL"
+        log(f"  [{mark}] localhost {c.name}: {c.detail}")
+    if raise_on_fail and any(not c.ok for c in last):
+        failed = [c.name for c in last if not c.ok]
+        raise CloudError(
+            f"туннель поднят, но локально не отвечает: {', '.join(failed)}"
+        )
+    return last

src/gpu_rent/session.py

@@ -22,7 +22,7 @@ from gpu_rent.cloud import (
 )
 from gpu_rent.bootstrap import run_bootstrap
 from gpu_rent.provision import provision_vm, tune_swarm_perf
-from gpu_rent.ready import wait_backend_idle
+from gpu_rent.ready import verify_stack_on_vm, wait_backend_idle
 from gpu_rent.snapshot import ensure_boot_snapshot
 from gpu_rent.notify import notify_ready
 from gpu_rent.config import Config
@@ -127,6 +127,19 @@ def _bind_access(
             log(f"perf tune: {exc}")
     else:
         log("ready: llm-only (без ожидания SwarmUI Idle)")
+
+    try:
+        checks = verify_stack_on_vm(cfg, ip, log, timeout=300.0)
+        state.notes = dict(state.notes or {})
+        state.notes["stack_vm"] = [
+            {"name": c.name, "ok": c.ok, "detail": c.detail} for c in checks
+        ]
+    except CloudError as exc:
+        state.notes = dict(state.notes or {})
+        state.notes["stack_vm_error"] = str(exc)[:500]
+        save_state(state)
+        raise
+
     try:
         ensure_boot_snapshot(
             conn,
@@ -137,6 +150,27 @@ def _bind_access(
     except CloudError as exc:
         log(f"snapshot: {exc}")
     notify_ready(cfg, log)
+    try:
+        from gpu_rent.balance import (
+            BalanceWatchState,
+            fetch_balance_rub,
+            save_balance_state,
+        )
+
+        if cfg.selectel_api_token:
+            rub = fetch_balance_rub(cfg.selectel_api_token)
+            save_balance_state(
+                BalanceWatchState(baseline_rub=rub, last_balance_rub=rub)
+            )
+            log(
+                f"balance: baseline {rub:.0f} ₽ "
+                f"(уведомление каждые {cfg.balance_notify_step_rub:.0f} ₽, "
+                "нужен local-watchdog)"
+            )
+        else:
+            log("balance: SELECTEL_API_TOKEN не задан — шаги по ₽ skip")
+    except Exception as exc:
+        log(f"balance baseline: {exc}")
     state.bootstrapped = True
     state.phase = "ready_cloud"
     state.notes = dict(state.notes or {})
@@ -464,5 +498,11 @@ def cmd_stop(
             clear_lease()
         except Exception:
             pass
+        try:
+            from gpu_rent.balance import clear_balance_state
+
+            clear_balance_state()
+        except Exception:
+            pass
         log("фаза idle" + ("" if destroy_disks else " (диски на месте)"))
         return state

src/gpu_rent/tunnel.py

@@ -194,6 +194,20 @@ def run_tunnel(
     else:
         open_url = f"http://127.0.0.1:{cfg.swarmui_local_port}"
 
+    from gpu_rent.ready import verify_stack_local
+
+    try:
+        local_checks = verify_stack_local(cfg, log, timeout=90.0)
+        state = load_state()
+        state.notes = dict(state.notes or {})
+        state.notes["stack_local"] = [
+            {"name": c.name, "ok": c.ok, "detail": c.detail} for c in local_checks
+        ]
+        save_state(state)
+    except CloudError as exc:
+        log(f"проверка туннеля: {exc}")
+        raise
+
     from gpu_rent.access_card import print_access_card
 
     print_access_card(cfg, tunneled=True, host=current_host)

tests/test_balance.py

@@ -0,0 +1,56 @@
+from gpu_rent.balance import (
+    BalanceWatchState,
+    balance_rub_from_payload,
+    plan_balance_notices,
+    spend_steps,
+)
+
+
+def test_balance_kopecks():
+    payload = {
+        "data": {
+            "billings": [
+                {"balances_values_sum": 123456},  # 1234.56 ₽
+            ]
+        }
+    }
+    assert balance_rub_from_payload(payload) == 1234.56
+
+
+def test_spend_steps_200():
+    assert spend_steps(1000, 850, 200) == 0
+    assert spend_steps(1000, 800, 200) == 1
+    assert spend_steps(1000, 600, 200) == 2
+    assert spend_steps(1000, 1100, 200) == 0  # top-up direction
+
+
+def test_plan_notifies_each_step():
+    state = BalanceWatchState(baseline_rub=1000.0)
+    state, msgs = plan_balance_notices(state, 750.0, step_rub=200.0)
+    assert state.last_notified_step == 1
+    assert len(msgs) == 1
+    assert "200" in msgs[0]
+    state, msgs2 = plan_balance_notices(state, 550.0, step_rub=200.0)
+    assert state.last_notified_step == 2
+    assert len(msgs2) == 1
+    # no re-notify same step
+    state, msgs3 = plan_balance_notices(state, 520.0, step_rub=200.0)
+    assert msgs3 == []
+    assert state.last_notified_step == 2
+
+
+def test_plan_topup_resets():
+    state = BalanceWatchState(baseline_rub=500.0, last_notified_step=2)
+    state, msgs = plan_balance_notices(state, 900.0, step_rub=200.0)
+    assert state.baseline_rub == 900.0
+    assert state.last_notified_step == 0
+    assert any("пополнен" in m for m in msgs)
+
+
+def test_plan_low_once():
+    state = BalanceWatchState(baseline_rub=1000.0)
+    state, msgs = plan_balance_notices(state, 150.0, step_rub=200.0, low_rub=200.0)
+    assert state.low_notified
+    assert any("низкий" in m for m in msgs)
+    state, msgs2 = plan_balance_notices(state, 100.0, step_rub=200.0, low_rub=200.0)
+    assert not any("низкий" in m for m in msgs2)

tests/test_verify_stack.py

@@ -0,0 +1,55 @@
+from gpu_rent.ready import ServiceCheck, _expected_services, verify_stack_local
+
+
+class _Cfg:
+    enable_swarmui = True
+    llm_runtime = "none"
+    swarmui_local_port = 17801
+    ollama_local_port = 17811
+    llamacpp_local_port = 17812
+
+
+def test_expected_services_swarm_only():
+    assert _expected_services(_Cfg()) == (True, False, False)
+
+
+def test_expected_services_llm_only():
+    class C:
+        enable_swarmui = False
+        llm_runtime = "llamacpp"
+
+    assert _expected_services(C()) == (False, False, True)
+
+
+def test_verify_stack_local_empty_when_nothing():
+    class C:
+        enable_swarmui = False
+        llm_runtime = "none"
+        swarmui_local_port = 17801
+        ollama_local_port = 17811
+        llamacpp_local_port = 17812
+
+    logs: list[str] = []
+    assert verify_stack_local(C(), logs.append, timeout=0.1) == []
+
+
+def test_verify_stack_local_fails_closed_port(monkeypatch):
+    class C:
+        enable_swarmui = False
+        llm_runtime = "ollama"
+        swarmui_local_port = 17801
+        ollama_local_port = 17999
+        llamacpp_local_port = 17812
+
+    monkeypatch.setattr("gpu_rent.ready._tcp_ok", lambda port, host="127.0.0.1", timeout=0.8: False)
+    logs: list[str] = []
+    try:
+        verify_stack_local(C(), logs.append, timeout=0.3, poll_every=0.1)
+        assert False, "expected CloudError"
+    except Exception as exc:
+        assert "ollama" in str(exc).lower() or "не отвечает" in str(exc)
+
+
+def test_service_check_dataclass():
+    c = ServiceCheck("x", True, "ok", "vm")
+    assert c.ok and c.where == "vm"