feat: production paper trading platform

2026-07-14 22:52:47 +03:00
parent 7186acb9a1
commit 5c4aecfe5f
29 changed files with 396 additions and 564 deletions
@@ -3,13 +3,18 @@ HOST=127.0.0.1
 PORT=8787

 BYBIT_TESTNET=false
+# Official regional public endpoints for this deployment. Before future live
+# trading they must match the Bybit site where the API key was created.
+BYBIT_REST_BASE_URL=https://api.bybit.kz
+BYBIT_WEBSOCKET_URL=wss://stream.bybit.kz/v5/public/spot
 BYBIT_API_KEY=
 BYBIT_API_SECRET=

 STARTING_BALANCE_USDT=100
-AUTO_SELECT_SYMBOLS=false
+AUTO_SELECT_SYMBOLS=true
 TOP_SYMBOLS_COUNT=12
-SYMBOLS=BTCUSDT,ETHUSDT,HYPEUSDT,SOLUSDT,XRPUSDT,XPLUSDT,WLDUSDT,MNTUSDT,HUSDT,XAUTUSDT,IPUSDT,AAVEUSDT
+# Leave empty to discover the most liquid eligible USDT Spot pairs from Bybit.
+SYMBOLS=

 STRATEGY_MODE=torch_forecast
 BASE_INTERVAL=60
@@ -5,12 +5,16 @@ ENV PYTHONDONTWRITEBYTECODE=1 \

 WORKDIR /app
 COPY requirements.txt /app/requirements.txt
-RUN pip install --no-cache-dir --upgrade pip \
-    && pip install --no-cache-dir -r /app/requirements.txt
+RUN pip install --no-cache-dir --disable-pip-version-check --upgrade pip \
+    && pip install --no-cache-dir --disable-pip-version-check -r /app/requirements.txt \
+    && groupadd --gid 1000 tradebot \
+    && useradd --uid 1000 --gid tradebot --home-dir /app --shell /usr/sbin/nologin tradebot

-COPY crypto_spot_bot /app/crypto_spot_bot
-COPY README.md /app/README.md
-RUN mkdir -p /app/runtime
+COPY --chown=1000:1000 crypto_spot_bot /app/crypto_spot_bot
+COPY --chown=1000:1000 README.md /app/README.md
+RUN mkdir -p /app/runtime && chown -R 1000:1000 /app/runtime

 EXPOSE 8787
+USER 1000:1000
+STOPSIGNAL SIGTERM
 CMD ["python", "-m", "crypto_spot_bot.main"]
@@ -5,7 +5,7 @@ Spot-бот для демо-торговли криптовалютой на р
 ## Что реализовано

 - Реальные market data Bybit Spot: REST bootstrap и WebSocket-обновления.
- Фиксированный набор 12 USDT spot-пар для основной стратегии: `BTCUSDT`, `ETHUSDT`, `HYPEUSDT`, `SOLUSDT`, `XRPUSDT`, `XPLUSDT`, `WLDUSDT`, `MNTUSDT`, `HUSDT`, `XAUTUSDT`, `IPUSDT`, `AAVEUSDT`.
+- Торговый universe автоматически строится из актуальных Bybit Spot-инструментов: выбираются до 12 ликвидных USDT-пар по `turnover24h`, исключаются stablecoin-to-stablecoin и leveraged-token пары; фиксированный список можно задать только явным `SYMBOLS`.
 - Paper trading с учетом cash, комиссий, проскальзывания, stop-loss, take-profit и trailing stop.
 - Spot-only логика: покупка базовой монеты за USDT и продажа обратно, без short и без плеча.
 - Live spot-ордеры явно отправляются без плеча: `category=spot`, `isLeverage=0`.
@@ -18,13 +18,14 @@ Spot-бот для демо-торговли криптовалютой на р
 - DCA/мартингейл отключены: в режиме `trend_macd` брокер не разрешает вторую позицию по той же паре.
 - Grid, rebound, adaptive learning, Kelly sizing и time-series forecast выключены по умолчанию и не участвуют в принятии решений `trend_macd`.
 - Быстрый режим торговли: отдельный короткий интервал цикла, короткий cooldown после выхода и лимит новых входов в минуту; выходы по риску этим лимитом не блокируются.
- Веб-dashboard на русском: equity, cash, PnL, позиции, сделки, сигналы, события, свечные графики, переключатель быстрой торговли и индикаторы работы обучения.
- Android-монитор в `android/TradeBotMonitor`: русский мобильный интерфейс для просмотра 12 пар, свечей, Torch/Kelly параметров, расписания удалённого retrain и live-чеклиста.
+- Защищённый JSON API: equity, cash, PnL, позиции, сделки, сигналы, события, свечи, управление paper-циклом и состоянием обучения.
+- Android-монитор в `android/TradeBotMonitor`: русский мобильный интерфейс для динамического списка Bybit-пар, свечей, Torch/Kelly параметров, WorkManager-расписания удалённого retrain и live-чеклиста.
 - SQLite runtime-хранилище в `runtime/tradebot.sqlite3`.
 - Liveness `/api/health`, readiness `/api/ready`, объединенный mobile snapshot `/api/mobile/snapshot` и Prometheus-compatible `/metrics`.
 - Все приватные API endpoints требуют токен или подтвержденный reverse-proxy user header; health и metrics остаются доступными для локального мониторинга.
- Docker Compose для установки на Raspberry Pi 5 или другой Linux-хост.
+- Hardened Docker Compose для установки на Dell/Linux: non-root user, read-only root filesystem, dropped capabilities, healthcheck и ротация container logs.
 - Live trading guard: live не стартует без `ENABLE_LIVE_TRADING=true`, `LIVE_TRADING_CONFIRM=I_ACCEPT_REAL_RISK` и Bybit API-ключей.
+- Внешний production endpoint сохраняется на `https://tb.kusoft.xyz`; Caddy завершает TLS и проксирует API к контейнеру на loopback.

 ## Источники и принятые параметры

@@ -34,6 +35,8 @@ Spot-бот для демо-торговли криптовалютой на р

 Популярность пар определяется через `/v5/market/tickers`, потому что Bybit Spot ticker возвращает `turnover24h`, `volume24h`, `bid1Price`, `ask1Price` и `lastPrice`: <https://bybit-exchange.github.io/docs/v5/market/tickers>.

+Для текущего paper/training-развёртывания используется официальный региональный endpoint `api.bybit.kz`; Bybit перечисляет его в Integration Guidance. `BYBIT_REST_BASE_URL` и `BYBIT_WEBSOCKET_URL` остаются явными настройками, потому что перед будущим live-режимом домен обязан соответствовать площадке выпуска API-ключа: <https://bybit-exchange.github.io/docs/v5/guide>.
+
 Лучшие bid/ask берутся из `/v5/market/orderbook`; документация Bybit описывает `GET /v5/market/orderbook` с `category=spot`: <https://bybit-exchange.github.io/docs/v5/market/orderbook>.

 WebSocket-стакан использует topic `orderbook.{depth}.{symbol}`; Bybit документирует snapshot/delta-поведение и частоты push для Spot depth 1/50/200/1000: <https://bybit-exchange.github.io/docs/v5/websocket/public/orderbook>.
@@ -45,7 +48,7 @@ Live market orders используют `/v5/order/create`; Bybit докумен
 - Investopedia перечисляет важные свойства algo trading software: real-time market data, low latency, configurability, backtesting, broker/exchange integration, fees/costs и APIs: <https://www.investopedia.com/articles/active-trading/090815/picking-right-algorithmic-trading-software.asp>.
 - Investopedia отдельно указывает, что automated trading systems задают правила entry/exit/money management, но требуют мониторинга и несут риск mechanical failures и over-optimization: <https://www.investopedia.com/articles/trading/11/automated-trading-systems.asp>.
 - QuantInsti описывает типовой путь разработки: стратегия, backtesting, paper trading, затем live trading, плюс GUI, order management и risk management: <https://www.quantinsti.com/articles/automated-trading-system/>.
- Hochreiter и Schmidhuber описали LSTM как recurrent neural network architecture для последовательностей; обучение LSTM/GRU в проекте выполняется локально через PyTorch, а Raspberry Pi исполняет только экспортированные JSON-веса без PyTorch runtime: <https://direct.mit.edu/neco/article/9/8/1735/6109/Long-Short-Term-Memory>.
+- Hochreiter и Schmidhuber описали LSTM как recurrent neural network architecture для последовательностей; обучение LSTM/GRU в проекте выполняется локально через PyTorch, а Dell исполняет только прошедшие quality gate экспортированные JSON-веса без PyTorch runtime: <https://direct.mit.edu/neco/article/9/8/1735/6109/Long-Short-Term-Memory>.

 Я не могу подтвердить, что эта стратегия будет прибыльной. Источники выше описывают технические свойства и риски автоматической торговли, но не гарантируют прибыль.

@@ -59,17 +62,17 @@ Copy-Item .env.example .env
 python -m crypto_spot_bot.main
 ```

-Dashboard: <http://127.0.0.1:8787/>
+Liveness: <http://127.0.0.1:8787/api/health>

-## Локальное обучение PyTorch LSTM/GRU
+## Локальное обучение PyTorch LSTM

-Обучение запускается на основной Windows-машине, а Raspberry Pi остается только для исполнения торгового цикла. PyTorch нужен только на машине обучения; в JSON экспортируются веса, а runtime на Raspberry Pi считает inference обычным Python-кодом:
+Обучение запускается на основной Windows-машине, а Dell остается для исполнения торгового цикла. PyTorch нужен только на машине обучения; в JSON экспортируются веса, а runtime на Dell считает inference обычным Python-кодом:

 ```powershell
 .\.venv\Scripts\python.exe -m pip install torch --index-url https://download.pytorch.org/whl/cpu
 .\.venv\Scripts\python.exe tools\train_torch_recurrent_forecaster.py `
  --limit 3000 `
-  --architectures lstm,gru `
+  --architectures lstm `
  --lookbacks 64 `
  --hidden-sizes 64,96 `
  --layers 2 `
@@ -86,11 +89,10 @@ Dashboard: <http://127.0.0.1:8787/>

 Файл из `TIME_SERIES_LSTM_MODEL_PATH` читается ботом автоматически, если `TIME_SERIES_FORECAST_ENABLED=true`. В стратегии `torch_forecast` экспортированная PyTorch LSTM/GRU модель является единственным направляющим сигналом для входа и forecast-выхода. Экспортированные модели появляются в dashboard как `PyTorch LSTM` или `PyTorch GRU`; старый легкий reservoir LSTM-кандидат и все встроенные не-torch прогнозы удалены.

-Автопереобучение на Windows запускает PyTorch trainer, пишет лог в `runtime/torch_retrain.log` и защищается от параллельных запусков:
+Локальный retrain на Windows запускает PyTorch trainer, пишет лог в `runtime/torch_retrain.log` и защищается от параллельных запусков:

 ```powershell
 powershell -ExecutionPolicy Bypass -File tools\run_torch_retrain.ps1
-powershell -ExecutionPolicy Bypass -File tools\install_windows_torch_retrainer.ps1
 ```

 Для удалённого запуска с телефона или с бота используется Windows training agent. Бот на `tb.kusoft.xyz` хранит очередь заданий, а Windows-машина сама подключается к интернету, забирает задания, обучает модель и загружает артефакты обратно:
@@ -99,21 +101,15 @@ powershell -ExecutionPolicy Bypass -File tools\install_windows_torch_retrainer.p
 powershell -ExecutionPolicy Bypass -File tools\install_windows_training_agent.ps1 -ApiAuth "<TRADEBOT_TRAINING_TOKEN>" -StartNow
 ```

-Установщик сохраняет worker-токен через Windows DPAPI, удаляет его старую plaintext-копию из пользовательского окружения и включает постоянный запуск агента. С правами администратора используется Scheduled Task с watchdog; без повышения прав — штатный ярлык в пользовательской папке Startup. Старые локальные retrain-задачи удаляются, чтобы обучение запускалось через очередь, а не двумя независимыми механизмами.
+Установщик сохраняет worker-токен через Windows DPAPI, удаляет его старую plaintext-копию из пользовательского окружения и включает постоянный запуск агента. С правами администратора используется Scheduled Task с watchdog; без повышения прав — штатный ярлык в пользовательской папке Startup. Сервер выдаёт каждой попытке 10-минутную возобновляемую lease; зависшая попытка автоматически возвращается в очередь, а устаревший процесс не может загрузить артефакты по старой lease.

-По умолчанию Windows-agent обучает отдельную PyTorch `LSTM/GRU` для каждой пары на `6000` часовых свечах. Это не заставляет разнородные активы делить одну архитектуру и один набор recurrent-весов. Прогноз усредняется по seed `7/19`, модели сравниваются на validation-folds, а пороги калибруются отдельно для каждой пары. Ensemble guard выполняется пакетно на GPU, а экспорт не дублирует первый набор весов. Search space использует lookback `32/64/128`, hidden `64/96`, dropout `0.20`, AdamW learning rate `0.0007` и weight decay `0.0005`; untouched holdout и quality gate не ослабляются. Для диагностического pooled-запуска используется ключ `-Pooled`. Параметры можно переопределить через env: `TORCH_RETRAIN_SYMBOLS`, `TORCH_RETRAIN_LIMIT`, `TORCH_RETRAIN_LOOKBACKS`, `TORCH_RETRAIN_ARCHITECTURES`, `TORCH_RETRAIN_HIDDEN_SIZES`, `TORCH_RETRAIN_LAYERS`, `TORCH_RETRAIN_DROPOUTS`, `TORCH_RETRAIN_HORIZON`, `TORCH_RETRAIN_HORIZONS`, `TORCH_RETRAIN_CONTEXT_SYMBOLS`, `TORCH_RETRAIN_FEATURES`, `TORCH_RETRAIN_SEED`, `TORCH_RETRAIN_ENSEMBLE_SEEDS`, `TORCH_RETRAIN_SELECTION_FOLDS`, `TORCH_RETRAIN_LEARNING_RATE`, `TORCH_RETRAIN_WEIGHT_DECAY`, `TORCH_RETRAIN_EPOCHS`, `TORCH_RETRAIN_PATIENCE`, `TORCH_RETRAIN_INTERVAL`, `TORCH_RETRAIN_ENV`.
+По умолчанию Windows-agent обучает одну pooled PyTorch LSTM на динамическом наборе пар и `4000` часовых свечах на пару. Базовый профиль использует lookback `64`, hidden size `64`, два recurrent-слоя, dropout `0.20`, до `50` эпох, seed-ensemble `7/19`, три validation-fold и AdamW с learning rate `0.0007`/weight decay `0.0005`. Untouched holdout и quality gate не ослабляются. Параметр задания `pooled=false` включает независимые модели по парам; `architectures=gru` оставлен только как явная экспериментальная опция. Параметры можно переопределить через env: `TORCH_RETRAIN_SYMBOLS`, `TORCH_RETRAIN_LIMIT`, `TORCH_RETRAIN_LOOKBACKS`, `TORCH_RETRAIN_ARCHITECTURES`, `TORCH_RETRAIN_HIDDEN_SIZES`, `TORCH_RETRAIN_LAYERS`, `TORCH_RETRAIN_DROPOUTS`, `TORCH_RETRAIN_HORIZON`, `TORCH_RETRAIN_HORIZONS`, `TORCH_RETRAIN_CONTEXT_SYMBOLS`, `TORCH_RETRAIN_FEATURES`, `TORCH_RETRAIN_SEED`, `TORCH_RETRAIN_ENSEMBLE_SEEDS`, `TORCH_RETRAIN_SELECTION_FOLDS`, `TORCH_RETRAIN_LEARNING_RATE`, `TORCH_RETRAIN_WEIGHT_DECAY`, `TORCH_RETRAIN_EPOCHS`, `TORCH_RETRAIN_PATIENCE`, `TORCH_RETRAIN_INTERVAL`, `TORCH_RETRAIN_ENV`.

 Loss и выбор гиперпараметров учитывают after-cost trading utility, ошибку ожидаемого чистого PnL, quantile-loss и focal BCE для события `TP before SL`, а не только MAE направления цены. В каждом walk-forward fold вероятность успеха калибруется Platt-моделью исключительно на train-части; затем на этой же train-части выбираются глобальные и per-symbol пороги, которые применяются к test-части. Для выбора порога требуется минимум 24 непересекающиеся сделки, а финальный quality gate по-прежнему требует не менее 30 OOS-сделок. Калибратор не имеет fallback на единичные сделки: если минимальная статистика не набрана, кандидат получает `calibration_insufficient` и не может пройти gate.

 Основной decision horizon — `12h`, дополнительные горизонты — `3/6/12/24`. Размеры обучающих барьеров берутся из `STOP_LOSS_PERCENT` и `TAKE_PROFIT_PERCENT`, а round-trip cost — из fee/slippage настроек. Threshold search оценивается тем же execution replay со stop-loss, take-profit, ATR trailing и forecast-exit, который используется в walk-forward. `holdout_skill` остаётся только в финальном отчёте и никогда не участвует в фильтрации входов или подборе порогов.

-Если retrain запускается с `-DeployToPi`, после успешного guard он синхронизирует `runtime/lstm_forecaster.json`, `runtime/torch_retrain_guard.json` и `runtime/torch_threshold_calibration.json` на Raspberry Pi через SSH-ключ и перезапускает сервис `tradebot`. Отдельный запуск sync:
-
-```powershell
-powershell -ExecutionPolicy Bypass -File tools\sync_torch_artifacts_to_pi.ps1 -RemoteHost 192.168.0.185 -RemoteUser sevenhill -RemoteRoot /mnt/data/tradebot
-```
-
-Внутри recurrent модели используются exportable attention pooling и LayerNorm. После recurrent-контекста добавлена нелинейная GELU-проекция и две отдельные экспортируемые головы: одна для ожидаемого PnL/quantiles, вторая для `P(TP before SL)`. Raspberry Pi по-прежнему исполняет модель из JSON без PyTorch runtime.
+Внутри recurrent модели используются exportable attention pooling и LayerNorm. После recurrent-контекста добавлена нелинейная GELU-проекция и две отдельные экспортируемые головы: одна для ожидаемого PnL/quantiles, вторая для `P(TP before SL)`. Принятый bundle загружается агентом через защищённый API `tb.kusoft.xyz`, проходит серверную проверку SHA-256/guard/calibration и атомарно становится активным на Dell.

 ## Docker

@@ -123,18 +119,20 @@ docker compose up -d --build
 docker compose logs -f tradebot
 ```

-Dashboard: `http://<host>:8787/`
+Локальная проверка: `http://127.0.0.1:8787/api/health`; внешний адрес: `https://tb.kusoft.xyz`.

-Для Raspberry Pi 5 проект использует `python:3.12-slim`, без Node.js build step. Runtime-данные лежат в volume `./runtime:/app/runtime`; на внешнем диске можно разместить папку проекта или заменить volume на абсолютный путь внешнего диска.
+На Dell проект использует `python:3.12-slim`, без Node.js build step. Runtime-данные лежат в bind mount `./runtime:/app/runtime`; корневая файловая система контейнера read-only, процесс работает как UID/GID 1000, а container logs ротируются по `10 MiB × 3`.

 ## Основные env-параметры

 ```env
 TRADING_MODE=paper
 STARTING_BALANCE_USDT=100
-AUTO_SELECT_SYMBOLS=false
+BYBIT_REST_BASE_URL=https://api.bybit.kz
+BYBIT_WEBSOCKET_URL=wss://stream.bybit.kz/v5/public/spot
+AUTO_SELECT_SYMBOLS=true
 TOP_SYMBOLS_COUNT=12
-SYMBOLS=BTCUSDT,ETHUSDT,HYPEUSDT,SOLUSDT,XRPUSDT,XPLUSDT,WLDUSDT,MNTUSDT,HUSDT,XAUTUSDT,IPUSDT,AAVEUSDT
+SYMBOLS=
 STRATEGY_MODE=torch_forecast
 BASE_INTERVAL=60
 TREND_INTERVAL=D
@@ -11,7 +11,7 @@
 - Kelly/размер позиции: текущий размер, Kelly-цель, занятая экспозиция, остаток, множители edge/P(up)/skill.
 - Обзор equity/cash/exposure/PnL и последних решений.
 - Удалённый запуск retrain через очередь заданий на боте и закреплённый Windows-компьютер обучения.
- Расписание retrain на телефоне: Android отправляет команду по расписанию, но обучение идёт на Windows-машине.
+- Расписание retrain через Android WorkManager: команда отправляется только при наличии сети, а обучение идёт на Windows-машине.
 - Настройки API, токена команд, тёмной/светлой темы.
 - Live-чеклист: приложение показывает, готов ли сервер к реальной торговле, и не включает live одной опасной кнопкой.

@@ -42,7 +42,7 @@ https://tb.kusoft.xyz

 ## Переобучение

-Телефон не обучает модель локально. Вкладка `Обучение` ставит задание в очередь на `tb.kusoft.xyz`, а Windows-agent на закреплённой машине `SEVENHILL` (`G:\Repos\TradeBot`) сам выходит в интернет, забирает задание, обучает модель и отправляет артефакты обратно боту. Так телефон становится пультом запуска/расписания, а тяжёлый PyTorch retrain остаётся на нормальном компьютере даже если он находится в другой сети.
+Телефон не обучает модель локально. Вкладка `Обучение` ставит задание в очередь на `tb.kusoft.xyz`, а Windows-agent на этой машине сам выходит в интернет, забирает задание, обучает модель и отправляет проверенный bundle обратно боту. Имя и путь активного worker приложение получает от сервера, без прошитого имени компьютера.

 ## Live-торговля

@@ -4,13 +4,17 @@ plugins {

 android {
    namespace = "xyz.kusoft.tradebotmonitor"
-    compileSdk = 36
+    compileSdk = 37

    defaultConfig {
        applicationId = "xyz.kusoft.tradebotmonitor"
        minSdk = 26
-        targetSdk = 36
-        versionCode = 21
-        versionName = "0.4.2"
+        targetSdk = 37
+        versionCode = 22
+        versionName = "0.5.0"
    }
 }
+
+dependencies {
+    implementation("androidx.work:work-runtime:2.11.2")
+}
@@ -2,12 +2,13 @@
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
    <uses-permission android:name="android.permission.INTERNET" />
    <uses-permission android:name="android.permission.ACCESS_NETWORK_STATE" />
-    <uses-permission android:name="android.permission.RECEIVE_BOOT_COMPLETED" />
-
    <application
        android:allowBackup="false"
+        android:dataExtractionRules="@xml/data_extraction_rules"
+        android:fullBackupContent="@xml/backup_rules"
        android:icon="@drawable/ic_launcher"
        android:label="TradeBot AI"
+        android:networkSecurityConfig="@xml/network_security_config"
        android:roundIcon="@drawable/ic_launcher"
        android:supportsRtl="true"
        android:theme="@style/AppTheme"
@@ -22,16 +23,5 @@
            </intent-filter>
        </activity>

-        <receiver
-            android:name=".RetrainAlarmReceiver"
-            android:exported="false" />
-
-        <receiver
-            android:name=".BootReceiver"
-            android:exported="false">
-            <intent-filter>
-                <action android:name="android.intent.action.BOOT_COMPLETED" />
-            </intent-filter>
-        </receiver>
    </application>
 </manifest>
@@ -18,12 +18,9 @@ class AppPrefs(context: Context) {
        }
        val trainingComputerName = prefs.getString("training_computer_name", null)?.trim()
        val trainingComputerPath = prefs.getString("training_computer_path", null)?.trim()
-        if (
-            trainingComputerName.isNullOrBlank() ||
-            trainingComputerName == LEGACY_TRAINING_COMPUTER_NAME ||
-            trainingComputerPath.isNullOrBlank() ||
-            trainingComputerPath == LEGACY_TRAINING_COMPUTER_PATH
-        ) {
+        val staleFallback = trainingComputerName in setOf("SEVENHILL", "DESKTOP-TMFDL0H") ||
+            trainingComputerPath in setOf("G:\\Repos\\TradeBot", "C:\\Repos\\TradeBot")
+        if (trainingComputerName.isNullOrBlank() || trainingComputerPath.isNullOrBlank() || staleFallback) {
            pinDefaultTrainingComputer()
        }
    }
@@ -141,10 +138,8 @@ class AppPrefs(context: Context) {
    private companion object {
        const val DEFAULT_API_BASE_URL = "https://tb.kusoft.xyz"
        const val LEGACY_PI_API_BASE_URL = "http://192.168.0.185:8787"
-        const val DEFAULT_TRAINING_COMPUTER_NAME = "SEVENHILL"
-        const val DEFAULT_TRAINING_COMPUTER_PATH = "G:\\Repos\\TradeBot"
-        const val LEGACY_TRAINING_COMPUTER_NAME = "DESKTOP-TMFDL0H"
-        const val LEGACY_TRAINING_COMPUTER_PATH = "C:\\Repos\\TradeBot"
+        const val DEFAULT_TRAINING_COMPUTER_NAME = "Ожидание Windows-agent"
+        const val DEFAULT_TRAINING_COMPUTER_PATH = "Имя и путь поступят от сервера"
        const val TOKEN_KEY_ALIAS = "tradebot_api_auth_v1"
    }
 }
@@ -1512,12 +1512,15 @@ class MainActivity : Activity() {
    private fun rankedMarkets(data: BotSnapshot): List<MarketItem> =
        orderedMarkets(data.markets).sortedWith(
            compareByDescending<MarketItem> { marketRankScore(it, data.signalsBySymbol[it.symbol]) }
-                .thenBy { fixedSymbolIndex(it.symbol) }
+                .thenByDescending { it.ticker?.turnover24h ?: 0.0 }
                .thenBy { it.symbol },
        )

    private fun orderedMarkets(markets: List<MarketItem>): List<MarketItem> =
-        markets.sortedWith(compareBy({ fixedSymbolIndex(it.symbol) }, { it.symbol }))
+        markets.sortedWith(
+            compareByDescending<MarketItem> { it.ticker?.turnover24h ?: 0.0 }
+                .thenBy { it.symbol },
+        )

    private fun marketRankScore(market: MarketItem, signal: SignalData?): Double {
        val actionScore = when (normalizedAction(signal?.action)) {
@@ -1615,11 +1618,6 @@ class MainActivity : Activity() {
        }
    }

-    private fun fixedSymbolIndex(symbol: String): Int {
-        val index = FIXED_SYMBOLS.indexOf(symbol.uppercase(Locale.US))
-        return if (index >= 0) index else FIXED_SYMBOLS.size + 1
-    }
-
    private fun trainingStatusSignature(retrain: JSONObject): String =
        (retrain.optJSONObject("coordination") ?: JSONObject()).toString()

@@ -1895,20 +1893,4 @@ class MainActivity : Activity() {
        Toast.makeText(this, message, Toast.LENGTH_LONG).show()
    }

-    private companion object {
-        val FIXED_SYMBOLS = listOf(
-            "BTCUSDT",
-            "ETHUSDT",
-            "HYPEUSDT",
-            "SOLUSDT",
-            "XRPUSDT",
-            "XPLUSDT",
-            "WLDUSDT",
-            "MNTUSDT",
-            "HUSDT",
-            "XAUTUSDT",
-            "IPUSDT",
-            "AAVEUSDT",
-        )
-    }
 }
@@ -1,65 +1,55 @@
 package xyz.kusoft.tradebotmonitor

-import android.app.AlarmManager
-import android.app.PendingIntent
-import android.content.BroadcastReceiver
 import android.content.Context
-import android.content.Intent
-import java.util.concurrent.Executors
+import androidx.work.Constraints
+import androidx.work.ExistingPeriodicWorkPolicy
+import androidx.work.NetworkType
+import androidx.work.PeriodicWorkRequest
+import androidx.work.WorkManager
+import androidx.work.Worker
+import androidx.work.WorkerParameters
+import java.util.concurrent.TimeUnit

 object RetrainScheduler {
-    private const val ACTION_RETRAIN = "xyz.kusoft.tradebotmonitor.RETRAIN"
-    private const val REQUEST_CODE = 6406
+    private const val UNIQUE_WORK_NAME = "tradebot-periodic-retrain"

    fun schedule(context: Context, hours: Int) {
-        val interval = hours.coerceAtLeast(1) * 60L * 60L * 1000L
-        val manager = context.getSystemService(Context.ALARM_SERVICE) as AlarmManager
-        manager.setInexactRepeating(
-            AlarmManager.RTC_WAKEUP,
-            System.currentTimeMillis() + interval,
-            interval,
-            pendingIntent(context),
+        val constraints = Constraints.Builder()
+            .setRequiredNetworkType(NetworkType.CONNECTED)
+            .build()
+        val request = PeriodicWorkRequest.Builder(
+            RetrainWorker::class.java,
+            hours.coerceAtLeast(1).toLong(),
+            TimeUnit.HOURS,
+        )
+            .setConstraints(constraints)
+            .build()
+        WorkManager.getInstance(context).enqueueUniquePeriodicWork(
+            UNIQUE_WORK_NAME,
+            ExistingPeriodicWorkPolicy.UPDATE,
+            request,
        )
    }

    fun cancel(context: Context) {
-        val manager = context.getSystemService(Context.ALARM_SERVICE) as AlarmManager
-        manager.cancel(pendingIntent(context))
+        WorkManager.getInstance(context).cancelUniqueWork(UNIQUE_WORK_NAME)
    }
-
-    private fun pendingIntent(context: Context): PendingIntent =
-        PendingIntent.getBroadcast(
-            context,
-            REQUEST_CODE,
-            Intent(context, RetrainAlarmReceiver::class.java).setAction(ACTION_RETRAIN),
-            PendingIntent.FLAG_UPDATE_CURRENT or PendingIntent.FLAG_IMMUTABLE,
-        )
 }

-class RetrainAlarmReceiver : BroadcastReceiver() {
-    override fun onReceive(context: Context, intent: Intent) {
-        val pending = goAsync()
-        val executor = Executors.newSingleThreadExecutor()
-        executor.execute {
-            try {
-                val prefs = AppPrefs(context)
-                if (prefs.retrainScheduleEnabled) {
-                    TradeBotApi(prefs.apiBaseUrl, prefs.commandToken).requestRetrain()
-                }
-            } finally {
-                pending.finish()
-                executor.shutdown()
-            }
+class RetrainWorker(
+    context: Context,
+    parameters: WorkerParameters,
+) : Worker(context, parameters) {
+    override fun doWork(): Result {
+        val prefs = AppPrefs(applicationContext)
+        if (!prefs.retrainScheduleEnabled || prefs.commandToken.isBlank()) {
+            return Result.success()
        }
-    }
-}
-
-class BootReceiver : BroadcastReceiver() {
-    override fun onReceive(context: Context, intent: Intent) {
-        if (intent.action != Intent.ACTION_BOOT_COMPLETED) return
-        val prefs = AppPrefs(context)
-        if (prefs.retrainScheduleEnabled) {
-            RetrainScheduler.schedule(context, prefs.retrainIntervalHours)
+        return try {
+            TradeBotApi(prefs.apiBaseUrl, prefs.commandToken).requestRetrain()
+            Result.success()
+        } catch (_: Exception) {
+            Result.retry()
        }
    }
 }
@@ -183,7 +183,7 @@ class TradeBotApi(
                qualityScore = quality.optDouble("score", 0.0),
            )
        }
-        return output.sortedBy { it.symbol }
+        return output
    }

    private fun parseTicker(row: JSONObject): TickerData =
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="utf-8"?>
+<full-backup-content>
+    <exclude domain="root" path="." />
+    <exclude domain="file" path="." />
+    <exclude domain="database" path="." />
+    <exclude domain="sharedpref" path="." />
+    <exclude domain="external" path="." />
+</full-backup-content>
@@ -0,0 +1,17 @@
+<?xml version="1.0" encoding="utf-8"?>
+<data-extraction-rules>
+    <cloud-backup disableIfNoEncryptionCapabilities="true">
+        <exclude domain="root" path="." />
+        <exclude domain="file" path="." />
+        <exclude domain="database" path="." />
+        <exclude domain="sharedpref" path="." />
+        <exclude domain="external" path="." />
+    </cloud-backup>
+    <device-transfer>
+        <exclude domain="root" path="." />
+        <exclude domain="file" path="." />
+        <exclude domain="database" path="." />
+        <exclude domain="sharedpref" path="." />
+        <exclude domain="external" path="." />
+    </device-transfer>
+</data-extraction-rules>
@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="utf-8"?>
+<network-security-config>
+    <base-config cleartextTrafficPermitted="false">
+        <trust-anchors>
+            <certificates src="system" />
+        </trust-anchors>
+    </base-config>
+</network-security-config>
@@ -142,7 +142,7 @@
    <rect x="168" y="29" width="84" height="8" rx="4" fill="#151922"/>
    <text x="338" y="48" class="text small">91%</text>
    <text x="34" y="88" class="text h2">Рынки</text>
-    <text x="34" y="114" class="muted small">12 фиксированных spot-пар</text>
+    <text x="34" y="114" class="muted small">Динамические Bybit spot-пары</text>
    <rect x="34" y="136" width="340" height="42" rx="7" fill="#11141a" stroke="#242a36"/>
    <text x="52" y="162" class="dim small">Поиск пары или сигнала</text>

@@ -1,7 +1,7 @@
 distributionBase=GRADLE_USER_HOME
 distributionPath=wrapper/dists
 distributionUrl=https\://services.gradle.org/distributions/gradle-9.4.1-bin.zip
-networkTimeout=10000
+networkTimeout=60000
 validateDistributionUrl=true
 zipStoreBase=GRADLE_USER_HOME
 zipStorePath=wrapper/dists
@@ -1,3 +1,3 @@
 """Crypto spot trading bot package."""

-__version__ = "0.1.0"
+__version__ = "1.0.0"
@@ -169,17 +169,23 @@ class Settings:
    hold_signal_sample_seconds: int = 60
    storage_retention_days: int = 30
    storage_prune_interval_seconds: int = 3600
+    bybit_rest_base_url_override: str = ""
+    bybit_websocket_url_override: str = ""

    @property
    def rest_base_url(self) -> str:
-        return "https://api-testnet.bybit.com" if self.bybit_testnet else "https://api.bybit.com"
+        if self.bybit_rest_base_url_override:
+            return self.bybit_rest_base_url_override.rstrip("/")
+        return "https://api-testnet.bybit.com" if self.bybit_testnet else "https://api.bybit.kz"

    @property
    def websocket_url(self) -> str:
+        if self.bybit_websocket_url_override:
+            return self.bybit_websocket_url_override
        return (
            "wss://stream-testnet.bybit.com/v5/public/spot"
            if self.bybit_testnet
-            else "wss://stream.bybit.com/v5/public/spot"
+            else "wss://stream.bybit.kz/v5/public/spot"
        )

    @property
@@ -220,7 +226,7 @@ def load_settings(env_file: str | Path | None = None) -> Settings:
    strategy_mode = os.getenv("STRATEGY_MODE", "torch_forecast").strip().lower()
    if strategy_mode not in STRATEGY_MODES:
        raise ValueError("STRATEGY_MODE must be legacy, trend_macd or torch_forecast")
-    auto_select_symbols = _bool_env("AUTO_SELECT_SYMBOLS", False)
+    auto_select_symbols = _bool_env("AUTO_SELECT_SYMBOLS", True)
    top_symbols_count = _int_env("TOP_SYMBOLS_COUNT", len(FIXED_SPOT_SYMBOLS))
    requested_symbols = _symbols_env("SYMBOLS")
    symbols = requested_symbols if requested_symbols else (() if auto_select_symbols else FIXED_SPOT_SYMBOLS)
@@ -343,6 +349,11 @@ def load_settings(env_file: str | Path | None = None) -> Settings:
        hold_signal_sample_seconds=_int_env("HOLD_SIGNAL_SAMPLE_SECONDS", 60),
        storage_retention_days=_int_env("STORAGE_RETENTION_DAYS", 30),
        storage_prune_interval_seconds=_int_env("STORAGE_PRUNE_INTERVAL_SECONDS", 3600),
+        bybit_rest_base_url_override=os.getenv(
+            "BYBIT_REST_BASE_URL",
+            "" if _bool_env("BYBIT_TESTNET", False) else "https://api.bybit.kz",
+        ).strip(),
+        bybit_websocket_url_override=os.getenv("BYBIT_WEBSOCKET_URL", "").strip(),
    )
    _validate_settings(settings)
    if settings.trading_mode == "live" and not settings.live_ready:
@@ -12,6 +12,7 @@ from fastapi.responses import JSONResponse, PlainTextResponse
 from crypto_spot_bot.analytics import analytics_snapshot
 from crypto_spot_bot.auth import ApiAuthorizer
 from crypto_spot_bot.bot import CryptoSpotBot
+from crypto_spot_bot import __version__
 from crypto_spot_bot.bybit import BybitClient
 from crypto_spot_bot.config import Settings, load_settings, update_env_value
 from crypto_spot_bot.execution import LiveBroker, PaperBroker
@@ -58,7 +59,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
        finally:
            await bot.stop()

-    app = FastAPI(title="Крипто спот-бот", lifespan=lifespan)
+    app = FastAPI(title="Крипто спот-бот", version=__version__, lifespan=lifespan)
    app.state.settings = settings
    app.state.storage = storage
    app.state.bot = bot
@@ -76,6 +77,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
            "running": bot.running,
            "mode": settings.trading_mode,
            "auth_configured": authorizer.configured(),
+            "version": __version__,
        }

    @app.get("/api/ready")
@@ -2,9 +2,11 @@ from __future__ import annotations

 import base64
 import hashlib
+import hmac
 import json
 import os
 import re
+import secrets
 import shutil
 import uuid
 from datetime import UTC
@@ -20,11 +22,11 @@ ALLOWED_TRAINING_ARTIFACTS = {
    "torch_retrain_guard.json",
    "torch_threshold_calibration.json",
 }
-RUNNING_TIMEOUT = timedelta(hours=12)
+RUNNING_LEASE_TIMEOUT = timedelta(minutes=10)
 ONLINE_WINDOW = timedelta(minutes=3)
+MAX_JOB_ATTEMPTS = 3
 MAX_ARTIFACT_CHUNK_BYTES = 1024 * 1024
-# Independent per-symbol ensembles are intentionally larger than pooled models.
-# Keep a bounded limit, but leave enough room for the supported 12-symbol bundle.
+# Keep uploads bounded while leaving room for explicitly requested per-symbol bundles.
 MAX_ARTIFACT_BYTES = 256 * 1024 * 1024
 MAX_ARTIFACT_CHUNKS = 1024
 REQUIRED_MODEL_BUNDLE = set(ALLOWED_TRAINING_ARTIFACTS)
@@ -52,7 +54,12 @@ class TrainingCoordinator:
            existing = self._active_job(state)
            if existing is not None:
                self._save_state(state)
-                return {"queued": False, "reason": "active_job_exists", "job": existing, "status": self._public_status(state)}
+                return {
+                    "queued": False,
+                    "reason": "active_job_exists",
+                    "job": self._public_job(existing),
+                    "status": self._public_status(state),
+                }

            now = _now()
            job = {
@@ -63,11 +70,16 @@ class TrainingCoordinator:
                "parameters": _safe_parameters(payload.get("parameters")),
                "message": "",
                "artifacts": [],
+                "attempts": 0,
            }
            state.setdefault("jobs", []).append(job)
            self._trim_jobs(state)
            self._save_state(state)
-            return {"queued": True, "job": job, "status": self._public_status(state)}
+            return {
+                "queued": True,
+                "job": self._public_job(job),
+                "status": self._public_status(state),
+            }

    def heartbeat(self, payload: dict[str, Any] | None = None) -> dict[str, Any]:
        payload = payload or {}
@@ -91,12 +103,21 @@ class TrainingCoordinator:
                return {"claimed": False, "job": None, "status": self._public_status(state)}

            now = _now()
+            lease_token = secrets.token_urlsafe(32)
            job["status"] = "running"
            job["claimed_at"] = now
+            job["updated_at"] = now
            job["claimed_by"] = worker["id"]
            job["worker"] = worker
+            job["lease_token"] = lease_token
+            job["attempts"] = int(job.get("attempts", 0)) + 1
            self._save_state(state)
-            return {"claimed": True, "job": job, "status": self._public_status(state)}
+            return {
+                "claimed": True,
+                "job": self._public_job(job),
+                "lease_token": lease_token,
+                "status": self._public_status(state),
+            }

    def save_artifact_chunk(self, job_id: str, payload: dict[str, Any]) -> dict[str, Any]:
        job_id = _valid_job_id(job_id)
@@ -126,6 +147,7 @@ class TrainingCoordinator:
                raise ValueError(f"training job not found: {job_id}")
            if job.get("status") != "running" or not job.get("claimed_by"):
                raise ValueError("training job is not claimed and running")
+            self._require_lease(job, payload)
            uploads = job.setdefault("uploads", {})
            upload = uploads.setdefault(name, {"sha256": sha256, "total": total})
            if upload.get("sha256") != sha256 or int(upload.get("total", 0)) != total:
@@ -138,6 +160,7 @@ class TrainingCoordinator:
            received = sum(1 for part in range(total) if (chunk_dir / f"{part:06d}.part").is_file())
            if received < total:
                upload["received"] = received
+                job["updated_at"] = _now()
                self._save_state(state)
                return {"complete": False, "received": received, "total": total}

@@ -169,6 +192,7 @@ class TrainingCoordinator:
                {"name": name, "sha256": sha256, "size": size, "staged_at": _now()}
            )
            job["artifacts"] = artifacts
+            job["updated_at"] = _now()
            upload["received"] = total
            upload["complete"] = True
            self._save_state(state)
@@ -184,6 +208,7 @@ class TrainingCoordinator:
                raise ValueError(f"training job not found: {job_id}")
            if job.get("status") != "running" or not job.get("claimed_by"):
                raise ValueError("training job is not claimed and running")
+            self._require_lease(job, payload)
            if isinstance(payload.get("worker"), dict):
                state["worker"] = self._worker_from_payload(payload["worker"])
            job["status"] = "running"
@@ -194,7 +219,11 @@ class TrainingCoordinator:
            if isinstance(payload.get("details"), dict):
                job["details"] = payload["details"]
            self._save_state(state)
-            return {"ok": True, "job": job, "status": self._public_status(state)}
+            return {
+                "ok": True,
+                "job": self._public_job(job),
+                "status": self._public_status(state),
+            }

    def complete(self, job_id: str, payload: dict[str, Any] | None = None) -> dict[str, Any]:
        payload = payload or {}
@@ -206,6 +235,7 @@ class TrainingCoordinator:
                raise ValueError(f"training job not found: {job_id}")
            if job.get("status") != "running" or not job.get("claimed_by"):
                raise ValueError("training job is not claimed and running")
+            self._require_lease(job, payload)
            success = bool(payload.get("success", payload.get("status") == "completed"))
            if success and job.get("artifacts"):
                promoted = self._validate_and_promote(job_id, job)
@@ -221,8 +251,13 @@ class TrainingCoordinator:
                    job["model_decision"] = (
                        "accepted" if payload["summary"]["accepted"] else "rejected"
                    )
+            job.pop("lease_token", None)
            self._save_state(state)
-            return {"ok": True, "job": job, "status": self._public_status(state)}
+            return {
+                "ok": True,
+                "job": self._public_job(job),
+                "status": self._public_status(state),
+            }

    def _validate_and_promote(self, job_id: str, job: dict[str, Any]) -> list[dict[str, Any]]:
        ready_dir = self.upload_root / job_id / "ready"
@@ -327,11 +362,27 @@ class TrainingCoordinator:
            "agent_recently_seen": recently_seen,
            "agent_busy": agent_busy,
            "worker": worker,
-            "active_job": active,
-            "latest_job": latest,
+            "active_job": self._public_job(active),
+            "latest_job": self._public_job(latest),
            "pending_jobs": sum(1 for job in state.get("jobs", []) if job.get("status") == "pending"),
        }

+    @staticmethod
+    def _public_job(job: dict[str, Any] | None) -> dict[str, Any] | None:
+        if job is None:
+            return None
+        public = dict(job)
+        public.pop("lease_token", None)
+        public.pop("uploads", None)
+        return public
+
+    @staticmethod
+    def _require_lease(job: dict[str, Any], payload: dict[str, Any]) -> None:
+        expected = str(job.get("lease_token") or "")
+        supplied = str(payload.get("lease_token") or "")
+        if not expected or not supplied or not hmac.compare_digest(expected, supplied):
+            raise ValueError("training job lease is invalid or expired")
+
    def _active_job(self, state: dict[str, Any]) -> dict[str, Any] | None:
        for job in reversed(state.get("jobs", [])):
            if job.get("status") in {"pending", "running"}:
@@ -355,11 +406,30 @@ class TrainingCoordinator:
        for job in state.get("jobs", []):
            if job.get("status") != "running":
                continue
-            claimed_at = _parse_time(str(job.get("claimed_at") or ""))
-            if claimed_at and now - claimed_at > RUNNING_TIMEOUT:
+            lease_updated_at = _parse_time(
+                str(job.get("updated_at") or job.get("claimed_at") or "")
+            )
+            if not lease_updated_at or now - lease_updated_at <= RUNNING_LEASE_TIMEOUT:
+                continue
+            job_id = str(job.get("id") or "")
+            if job_id:
+                _remove_tree(self.upload_root / job_id)
+            job.pop("lease_token", None)
+            job.pop("uploads", None)
+            attempts = int(job.get("attempts", 0))
+            if attempts < MAX_JOB_ATTEMPTS:
+                job["status"] = "pending"
+                job["phase"] = "queued"
+                job["progress_percent"] = 0
+                job["message"] = "training worker lease expired; queued for retry"
+                job["retry_queued_at"] = _now()
+                for key in ("claimed_at", "claimed_by", "worker", "updated_at"):
+                    job.pop(key, None)
+            else:
                job["status"] = "failed"
+                job["phase"] = "failed"
                job["completed_at"] = _now()
-                job["message"] = "training worker timeout"
+                job["message"] = "training worker lease expired after maximum retries"

    def _trim_jobs(self, state: dict[str, Any]) -> None:
        jobs = state.get("jobs", [])
@@ -7,9 +7,9 @@ services:
    environment:
      HOST: 0.0.0.0
      PYTHONDONTWRITEBYTECODE: "1"
-    user: "1000:1000"
    init: true
    read_only: true
+    pids_limit: 128
    cap_drop:
      - ALL
    security_opt:
@@ -27,4 +27,10 @@ services:
      timeout: 10s
      retries: 3
      start_period: 30s
+    stop_grace_period: 30s
+    logging:
+      driver: json-file
+      options:
+        max-size: "10m"
+        max-file: "3"
    restart: unless-stopped
@@ -1,2 +1,2 @@
 -r requirements.txt
-pytest==8.4.2
+pytest==9.1.1
@@ -1,4 +1,4 @@
-fastapi==0.115.6
-uvicorn[standard]==0.34.0
-requests==2.32.3
-websockets==14.1
+fastapi==0.139.0
+uvicorn[standard]==0.51.0
+requests==2.34.2
+websockets==16.1
@@ -78,7 +78,7 @@ def test_llm_advisor_is_disabled_by_default(tmp_path, monkeypatch) -> None:
    assert settings.llm_advisor_enabled is False


-def test_default_symbols_are_fixed_trend_pairs(tmp_path, monkeypatch) -> None:
+def test_default_symbols_are_discovered_from_bybit(tmp_path, monkeypatch) -> None:
    monkeypatch.delenv("AUTO_SELECT_SYMBOLS", raising=False)
    monkeypatch.delenv("TOP_SYMBOLS_COUNT", raising=False)
    monkeypatch.delenv("SYMBOLS", raising=False)
@@ -89,9 +89,9 @@ def test_default_symbols_are_fixed_trend_pairs(tmp_path, monkeypatch) -> None:

    settings = load_settings(env_file)

-    assert settings.auto_select_symbols is False
+    assert settings.auto_select_symbols is True
    assert settings.top_symbols_count == len(FIXED_SPOT_SYMBOLS)
-    assert settings.symbols == FIXED_SPOT_SYMBOLS
+    assert settings.symbols == ()
    assert settings.strategy_mode == "torch_forecast"
    assert settings.base_interval == "60"
    assert settings.trend_interval == "D"
@@ -3,6 +3,7 @@ from __future__ import annotations
 import base64
 import hashlib
 import json
+from datetime import UTC, datetime, timedelta

 import pytest

@@ -16,6 +17,7 @@ def test_training_coordinator_claims_and_completes_job(tmp_path) -> None:
    job_id = requested["job"]["id"]
    heartbeat = coordinator.heartbeat({"worker_id": "win-1", "name": "DESKTOP-TMFDL0H"})
    claimed = coordinator.claim({"worker_id": "win-1", "name": "DESKTOP-TMFDL0H"})
+    lease_token = claimed["lease_token"]

    assert requested["queued"] is True
    assert heartbeat["status"]["agent_online"] is True
@@ -25,14 +27,23 @@ def test_training_coordinator_claims_and_completes_job(tmp_path) -> None:

    progress = coordinator.progress(
        job_id,
-        {"status": "running", "phase": "training", "progress_percent": 42, "message": "epoch 1"},
+        {
+            "status": "running",
+            "phase": "training",
+            "progress_percent": 42,
+            "message": "epoch 1",
+            "lease_token": lease_token,
+        },
    )

    assert progress["job"]["phase"] == "training"
    assert progress["job"]["progress_percent"] == 42
    assert coordinator.status()["active_job"]["message"] == "epoch 1"

-    completed = coordinator.complete(job_id, {"success": True, "message": "ok"})
+    completed = coordinator.complete(
+        job_id,
+        {"success": True, "message": "ok", "lease_token": lease_token},
+    )

    assert completed["job"]["status"] == "completed"
    assert coordinator.status()["active_job"] is None
@@ -104,7 +115,7 @@ def test_training_coordinator_reports_worker_identity_from_heartbeat(tmp_path) -
 def test_training_coordinator_records_rejected_candidate_as_completed_training(tmp_path) -> None:
    coordinator = TrainingCoordinator(tmp_path)
    job = coordinator.request_retrain({"source": "android"})["job"]
-    coordinator.claim({"worker_id": "worker-1"})
+    lease_token = coordinator.claim({"worker_id": "worker-1"})["lease_token"]

    completed = coordinator.complete(
        job["id"],
@@ -112,6 +123,7 @@ def test_training_coordinator_records_rejected_candidate_as_completed_training(t
            "success": True,
            "message": "training completed; candidate rejected by quality gate",
            "summary": {"accepted": False, "reason": "candidate_failed_honest_validation"},
+            "lease_token": lease_token,
        },
    )

@@ -124,7 +136,7 @@ def test_training_coordinator_records_rejected_candidate_as_completed_training(t
 def test_training_coordinator_accepts_chunked_artifact_upload(tmp_path) -> None:
    coordinator = TrainingCoordinator(tmp_path)
    job = coordinator.request_retrain({"source": "test"})["job"]
-    coordinator.claim({"worker_id": "test-worker"})
+    lease_token = coordinator.claim({"worker_id": "test-worker"})["lease_token"]
    payload = b'{"type":"pytorch_recurrent_forecaster","symbols":{}}\n'
    sha256 = hashlib.sha256(payload).hexdigest()
    first = payload[:20]
@@ -138,6 +150,7 @@ def test_training_coordinator_accepts_chunked_artifact_upload(tmp_path) -> None:
            "total": 2,
            "sha256": sha256,
            "data_base64": base64.b64encode(first).decode("ascii"),
+            "lease_token": lease_token,
        },
    )
    part_2 = coordinator.save_artifact_chunk(
@@ -148,6 +161,7 @@ def test_training_coordinator_accepts_chunked_artifact_upload(tmp_path) -> None:
            "total": 2,
            "sha256": sha256,
            "data_base64": base64.b64encode(second).decode("ascii"),
+            "lease_token": lease_token,
        },
    )

@@ -199,6 +213,35 @@ def test_running_claimed_job_keeps_agent_online_when_heartbeat_is_stale(tmp_path
    assert status["agent_online"] is True


+def test_stale_training_lease_is_requeued_and_old_lease_is_rejected(tmp_path) -> None:
+    coordinator = TrainingCoordinator(tmp_path)
+    job = coordinator.request_retrain({"source": "android"})["job"]
+    first_claim = coordinator.claim({"worker_id": "worker-1"})
+
+    state_path = tmp_path / "training_coordination.json"
+    state = json.loads(state_path.read_text(encoding="utf-8"))
+    state["jobs"][0]["updated_at"] = (
+        datetime.now(UTC) - timedelta(minutes=11)
+    ).isoformat()
+    state_path.write_text(json.dumps(state), encoding="utf-8")
+
+    second_claim = coordinator.claim({"worker_id": "worker-2"})
+
+    assert second_claim["claimed"] is True
+    assert second_claim["job"]["id"] == job["id"]
+    assert second_claim["job"]["attempts"] == 2
+    assert second_claim["lease_token"] != first_claim["lease_token"]
+    with pytest.raises(ValueError, match="lease"):
+        coordinator.progress(
+            job["id"],
+            {
+                "phase": "training",
+                "progress_percent": 10,
+                "lease_token": first_claim["lease_token"],
+            },
+        )
+
+
 def test_training_upload_rejects_unknown_job(tmp_path) -> None:
    coordinator = TrainingCoordinator(tmp_path)
    payload = b"{}"
@@ -219,7 +262,7 @@ def test_training_upload_rejects_unknown_job(tmp_path) -> None:
 def test_training_bundle_promotes_only_after_successful_guard(tmp_path) -> None:
    coordinator = TrainingCoordinator(tmp_path)
    job = coordinator.request_retrain({"source": "test"})["job"]
-    coordinator.claim({"worker_id": "worker-1"})
+    lease_token = coordinator.claim({"worker_id": "worker-1"})["lease_token"]
    model = {
        "type": "pytorch_recurrent_forecaster",
        "symbols": {
@@ -260,10 +303,14 @@ def test_training_bundle_promotes_only_after_successful_guard(tmp_path) -> None:
                "total": 1,
                "sha256": hashlib.sha256(payload).hexdigest(),
                "data_base64": base64.b64encode(payload).decode("ascii"),
+                "lease_token": lease_token,
            },
        )

-    completed = coordinator.complete(job["id"], {"success": True})
+    completed = coordinator.complete(
+        job["id"],
+        {"success": True, "lease_token": lease_token},
+    )

    assert completed["job"]["status"] == "completed"
    assert json.loads((tmp_path / "lstm_forecaster.json").read_text())["symbols"]["BTCUSDT"]
@@ -1,94 +0,0 @@
-[CmdletBinding()]
-param(
-    [string]$TaskName = "TradeBot PyTorch Forecaster Retrainer",
-    [int]$EveryHours = 6,
-    [string]$Symbols = "",
-    [int]$Limit = 3000,
-    [int]$Horizon = 0,
-    [string]$Horizons = "",
-    [string]$Features = "",
-    [string]$ContextSymbols = "",
-    [int]$FirstRunMinutes = 0,
-    [switch]$DeployToPi,
-    [string]$PiHost = "192.168.0.185",
-    [string]$PiUser = "sevenhill",
-    [string]$PiRoot = "/mnt/data/tradebot",
-    [string]$PiSshKeyPath = ""
-)
-
-$ErrorActionPreference = "Stop"
-
-$RepoRoot = (Resolve-Path (Join-Path $PSScriptRoot "..")).Path
-$Runner = Join-Path $RepoRoot "tools\run_torch_retrain.ps1"
-if (-not (Test-Path $Runner)) {
-    throw "Runner not found: $Runner"
-}
-
-$LegacyTaskName = "TradeBot LSTM Retrainer"
-if ($TaskName -ne $LegacyTaskName) {
-    $legacyTask = Get-ScheduledTask -TaskName $LegacyTaskName -ErrorAction SilentlyContinue
-    if ($legacyTask) {
-        Unregister-ScheduledTask -TaskName $LegacyTaskName -Confirm:$false
-    }
-}
-
-$actionArgs = "-NoProfile -ExecutionPolicy Bypass -File `"$Runner`""
-if ($Symbols) {
-    $actionArgs += " -Symbols `"$Symbols`""
-}
-if ($Limit -gt 0) {
-    $actionArgs += " -Limit $Limit"
-}
-if ($Horizon -gt 0) {
-    $actionArgs += " -Horizon $Horizon"
-}
-if ($Horizons) {
-    $actionArgs += " -Horizons `"$Horizons`""
-}
-if ($Features) {
-    $actionArgs += " -Features `"$Features`""
-}
-if ($ContextSymbols) {
-    $actionArgs += " -ContextSymbols `"$ContextSymbols`""
-}
-if ($DeployToPi) {
-    $actionArgs += " -DeployToPi"
-}
-if ($PiHost) {
-    $actionArgs += " -PiHost `"$PiHost`""
-}
-if ($PiUser) {
-    $actionArgs += " -PiUser `"$PiUser`""
-}
-if ($PiRoot) {
-    $actionArgs += " -PiRoot `"$PiRoot`""
-}
-if ($PiSshKeyPath) {
-    $actionArgs += " -PiSshKeyPath `"$PiSshKeyPath`""
-}
-$action = New-ScheduledTaskAction -Execute "powershell.exe" -Argument $actionArgs -WorkingDirectory $RepoRoot
-$trigger = New-ScheduledTaskTrigger `
-    -Once `
-    -At (Get-Date).AddMinutes($(if ($FirstRunMinutes -gt 0) { $FirstRunMinutes } else { $EveryHours * 60 })) `
-    -RepetitionInterval (New-TimeSpan -Hours $EveryHours) `
-    -RepetitionDuration (New-TimeSpan -Days 3650)
-$principal = New-ScheduledTaskPrincipal `
-    -UserId ([System.Security.Principal.WindowsIdentity]::GetCurrent().Name) `
-    -LogonType Interactive `
-    -RunLevel Limited
-$settings = New-ScheduledTaskSettingsSet `
-    -StartWhenAvailable `
-    -MultipleInstances IgnoreNew `
-    -AllowStartIfOnBatteries `
-    -DontStopIfGoingOnBatteries
-
-Register-ScheduledTask `
-    -TaskName $TaskName `
-    -Action $action `
-    -Trigger $trigger `
-    -Principal $principal `
-    -Settings $settings `
-    -Description "Retrains TradeBot PyTorch recurrent forecast parameters every $EveryHours hours." `
-    -Force | Out-Null
-
-Write-Host "Registered scheduled task '$TaskName' every $EveryHours hours."
@@ -1,152 +0,0 @@
-[CmdletBinding()]
-param(
-    [int]$MinReplayTrades = 8,
-    [int]$MaxAttempts = 0,
-    [string]$Symbols = "",
-    [int]$Limit = 3000,
-    [switch]$DeployToPi,
-    [string]$PiHost = "192.168.0.185",
-    [string]$PiUser = "sevenhill",
-    [string]$PiRoot = "/mnt/data/tradebot",
-    [string]$PiSshKeyPath = "",
-    [int]$SeedStart = 0
-)
-
-$ErrorActionPreference = "Stop"
-
-$RepoRoot = (Resolve-Path (Join-Path $PSScriptRoot "..")).Path
-$RuntimeDir = Join-Path $RepoRoot "runtime"
-$LoopLog = Join-Path $RuntimeDir "torch_retrain_until_replay8.log"
-$GuardReport = Join-Path $RuntimeDir "torch_retrain_guard.json"
-$ActiveCalibration = Join-Path $RuntimeDir "torch_threshold_calibration.json"
-$Runner = Join-Path $RepoRoot "tools\run_torch_retrain.ps1"
-New-Item -ItemType Directory -Force -Path $RuntimeDir | Out-Null
-
-function Write-LoopLog {
-    param([string]$Message)
-    $timestamp = Get-Date -Format "yyyy-MM-dd HH:mm:ssK"
-    "[$timestamp] $Message" | Tee-Object -FilePath $LoopLog -Append
-}
-
-function ConvertTo-IntOrZero {
-    param($Value)
-    try {
-        if ($null -eq $Value) {
-            return 0
-        }
-        return [int]$Value
-    }
-    catch {
-        return 0
-    }
-}
-
-function Read-GuardSummary {
-    if (-not (Test-Path $GuardReport)) {
-        return [pscustomobject]@{
-            Accepted = $false
-            Reason = "guard_report_missing"
-            CandidateReplayTrades = 0
-            CurrentReplayTrades = 0
-            WalkForwardTrades = 0
-        }
-    }
-    try {
-        $payload = Get-Content -Raw -LiteralPath $GuardReport | ConvertFrom-Json
-        return [pscustomobject]@{
-            Accepted = [bool]$payload.accepted
-            Reason = [string]$payload.reason
-            CandidateReplayTrades = ConvertTo-IntOrZero $payload.candidate.full_replay.trades
-            CurrentReplayTrades = ConvertTo-IntOrZero $payload.current.full_replay.trades
-            WalkForwardTrades = ConvertTo-IntOrZero $payload.candidate.walk_forward_summary.trades
-        }
-    }
-    catch {
-        return [pscustomobject]@{
-            Accepted = $false
-            Reason = "guard_report_unreadable"
-            CandidateReplayTrades = 0
-            CurrentReplayTrades = 0
-            WalkForwardTrades = 0
-        }
-    }
-}
-
-function Read-ActiveReplayTrades {
-    if (-not (Test-Path $ActiveCalibration)) {
-        return 0
-    }
-    try {
-        $payload = Get-Content -Raw -LiteralPath $ActiveCalibration | ConvertFrom-Json
-        return ConvertTo-IntOrZero $payload.full_replay.trades
-    }
-    catch {
-        return 0
-    }
-}
-
-function Read-ActiveValidationPassed {
-    if (-not (Test-Path $ActiveCalibration)) {
-        return $false
-    }
-    try {
-        $payload = Get-Content -Raw -LiteralPath $ActiveCalibration | ConvertFrom-Json
-        return [bool]$payload.validation.passed
-    }
-    catch {
-        return $false
-    }
-}
-
-$attempt = 0
-while ($true) {
-    $activeReplayTrades = Read-ActiveReplayTrades
-    if (Read-ActiveValidationPassed) {
-        Write-LoopLog "Stop condition reached: active calibration passed honest validation with full_replay.trades=$activeReplayTrades."
-        exit 0
-    }
-
-    $attempt += 1
-    if ($SeedStart -gt 0) {
-        $attemptSeed = $SeedStart + $attempt - 1
-    }
-    else {
-        $attemptSeed = Get-Random -Minimum 1 -Maximum 2147483647
-    }
-    Write-LoopLog "Attempt $attempt started; seed=$attemptSeed; target full_replay.trades >= $MinReplayTrades."
-
-    $runnerArgs = @(
-        "-NoProfile",
-        "-ExecutionPolicy", "Bypass",
-        "-File", $Runner,
-        "-Limit", $Limit.ToString(),
-        "-Seed", $attemptSeed.ToString()
-    )
-    if ($Symbols) {
-        $runnerArgs += @("-Symbols", $Symbols)
-    }
-    if ($DeployToPi) {
-        $runnerArgs += "-DeployToPi"
-        if ($PiHost) { $runnerArgs += @("-PiHost", $PiHost) }
-        if ($PiUser) { $runnerArgs += @("-PiUser", $PiUser) }
-        if ($PiRoot) { $runnerArgs += @("-PiRoot", $PiRoot) }
-        if ($PiSshKeyPath) { $runnerArgs += @("-PiSshKeyPath", $PiSshKeyPath) }
-    }
-
-    & powershell.exe @runnerArgs 2>&1 | Tee-Object -FilePath $LoopLog -Append
-    $runnerExit = $LASTEXITCODE
-    $summary = Read-GuardSummary
-    Write-LoopLog "Attempt $attempt finished; runner_exit=$runnerExit accepted=$($summary.Accepted) reason=$($summary.Reason) candidate_full_replay.trades=$($summary.CandidateReplayTrades) current_full_replay.trades=$($summary.CurrentReplayTrades) walk_forward.trades=$($summary.WalkForwardTrades)."
-
-    if ($summary.Accepted -and (Read-ActiveValidationPassed)) {
-        Write-LoopLog "Stop condition reached: accepted candidate passed honest validation with full_replay.trades=$($summary.CandidateReplayTrades)."
-        exit 0
-    }
-
-    if ($MaxAttempts -gt 0 -and $attempt -ge $MaxAttempts) {
-        Write-LoopLog "MaxAttempts=$MaxAttempts reached before replay target."
-        exit 2
-    }
-
-    Start-Sleep -Seconds 10
-}
@@ -22,12 +22,6 @@ param(
    [int]$HoldoutWindow = 0,
    [string]$Interval = "",
    [string]$EnvFile = "",
-    [switch]$DeployToPi,
-    [string]$PiHost = "",
-    [string]$PiUser = "",
-    [string]$PiRoot = "",
-    [string]$PiSshKeyPath = "",
-    [switch]$NoPiRestart,
    [switch]$Pooled,
    [switch]$SkipGuard,
    [switch]$ResumeCandidate
@@ -105,44 +99,13 @@ function Test-TorchArtifactFile {
    }
 }

-function Sync-AcceptedArtifactsToPi {
-    if (-not ($DeployToPi -or $env:TORCH_RETRAIN_DEPLOY_TO_PI)) {
-        Write-RetrainLog "Pi artifact sync disabled."
-        return
-    }
-
-    $syncScript = Join-Path $RepoRoot "tools\sync_torch_artifacts_to_pi.ps1"
-    if (-not (Test-Path $syncScript)) {
-        throw "Pi sync script not found: $syncScript"
-    }
-
-    $syncArgs = @(
-        "-NoProfile",
-        "-ExecutionPolicy", "Bypass",
-        "-File", $syncScript,
-        "-RepoRoot", $RepoRoot
-    )
-    if ($PiHost) { $syncArgs += @("-RemoteHost", $PiHost) }
-    if ($PiUser) { $syncArgs += @("-RemoteUser", $PiUser) }
-    if ($PiRoot) { $syncArgs += @("-RemoteRoot", $PiRoot) }
-    if ($PiSshKeyPath) { $syncArgs += @("-SshKeyPath", $PiSshKeyPath) }
-    if ($NoPiRestart) { $syncArgs += "-NoRestart" }
-
-    Write-RetrainLog "Syncing accepted Torch artifacts to Raspberry Pi."
-    & powershell.exe @syncArgs 2>&1 | Tee-Object -FilePath $LogFile -Append
-    if ($LASTEXITCODE -ne 0) {
-        throw "Pi artifact sync failed with exit code $LASTEXITCODE."
-    }
-    Write-RetrainLog "Pi artifact sync completed."
-}
-
 if (-not $Symbols -and $env:TORCH_RETRAIN_SYMBOLS) { $Symbols = $env:TORCH_RETRAIN_SYMBOLS }
 if ($Limit -le 0) {
-    $Limit = if ($env:TORCH_RETRAIN_LIMIT) { [int]$env:TORCH_RETRAIN_LIMIT } else { 6000 }
+    $Limit = if ($env:TORCH_RETRAIN_LIMIT) { [int]$env:TORCH_RETRAIN_LIMIT } else { 4000 }
 }
-if (-not $Lookbacks) { $Lookbacks = if ($env:TORCH_RETRAIN_LOOKBACKS) { $env:TORCH_RETRAIN_LOOKBACKS } else { "32,64,128" } }
-if (-not $Architectures) { $Architectures = if ($env:TORCH_RETRAIN_ARCHITECTURES) { $env:TORCH_RETRAIN_ARCHITECTURES } else { "lstm,gru" } }
-if (-not $HiddenSizes) { $HiddenSizes = if ($env:TORCH_RETRAIN_HIDDEN_SIZES) { $env:TORCH_RETRAIN_HIDDEN_SIZES } else { "64,96" } }
+if (-not $Lookbacks) { $Lookbacks = if ($env:TORCH_RETRAIN_LOOKBACKS) { $env:TORCH_RETRAIN_LOOKBACKS } else { "64" } }
+if (-not $Architectures) { $Architectures = if ($env:TORCH_RETRAIN_ARCHITECTURES) { $env:TORCH_RETRAIN_ARCHITECTURES } else { "lstm" } }
+if (-not $HiddenSizes) { $HiddenSizes = if ($env:TORCH_RETRAIN_HIDDEN_SIZES) { $env:TORCH_RETRAIN_HIDDEN_SIZES } else { "64" } }
 if (-not $Layers) { $Layers = if ($env:TORCH_RETRAIN_LAYERS) { $env:TORCH_RETRAIN_LAYERS } else { "2" } }
 if (-not $Dropouts) { $Dropouts = if ($env:TORCH_RETRAIN_DROPOUTS) { $env:TORCH_RETRAIN_DROPOUTS } else { "0.20" } }
 if ($Horizon -le 0) { $Horizon = if ($env:TORCH_RETRAIN_HORIZON) { [int]$env:TORCH_RETRAIN_HORIZON } else { 12 } }
@@ -154,7 +117,7 @@ if (-not $EnsembleSeeds) { $EnsembleSeeds = if ($env:TORCH_RETRAIN_ENSEMBLE_SEED
 if ($SelectionFolds -le 0) { $SelectionFolds = if ($env:TORCH_RETRAIN_SELECTION_FOLDS) { [int]$env:TORCH_RETRAIN_SELECTION_FOLDS } else { 3 } }
 if ($LearningRate -le 0) { $LearningRate = if ($env:TORCH_RETRAIN_LEARNING_RATE) { [double]$env:TORCH_RETRAIN_LEARNING_RATE } else { 0.0007 } }
 if ($WeightDecay -le 0) { $WeightDecay = if ($env:TORCH_RETRAIN_WEIGHT_DECAY) { [double]$env:TORCH_RETRAIN_WEIGHT_DECAY } else { 0.0005 } }
-if ($Epochs -le 0) { $Epochs = if ($env:TORCH_RETRAIN_EPOCHS) { [int]$env:TORCH_RETRAIN_EPOCHS } else { 70 } }
+if ($Epochs -le 0) { $Epochs = if ($env:TORCH_RETRAIN_EPOCHS) { [int]$env:TORCH_RETRAIN_EPOCHS } else { 50 } }
 if ($Patience -le 0) { $Patience = if ($env:TORCH_RETRAIN_PATIENCE) { [int]$env:TORCH_RETRAIN_PATIENCE } else { 8 } }
 if ($ValidationWindow -le 0) { $ValidationWindow = if ($env:TORCH_RETRAIN_VALIDATION_WINDOW) { [int]$env:TORCH_RETRAIN_VALIDATION_WINDOW } else { 720 } }
 if ($HoldoutWindow -le 0) { $HoldoutWindow = if ($env:TORCH_RETRAIN_HOLDOUT_WINDOW) { [int]$env:TORCH_RETRAIN_HOLDOUT_WINDOW } else { 1000 } }
@@ -302,7 +265,6 @@ try {
        Write-RetrainLog "Updated active threshold calibration: $(Join-Path $RuntimeDir "torch_threshold_calibration.json")"
    }
    Write-RetrainLog "Candidate accepted by guard. Active artifact: $ModelFile"
-    Sync-AcceptedArtifactsToPi
 }
 catch {
    Write-RetrainLog "ERROR: $($_.Exception.Message)"
@@ -1,95 +0,0 @@
-[CmdletBinding()]
-param(
-    [string]$RepoRoot = "",
-    [string]$RemoteHost = "",
-    [string]$RemoteUser = "",
-    [string]$RemoteRoot = "",
-    [string]$SshKeyPath = "",
-    [string]$ServiceName = "tradebot",
-    [switch]$NoRestart,
-    [switch]$DryRun
-)
-
-$ErrorActionPreference = "Stop"
-
-if (-not $RepoRoot) { $RepoRoot = (Resolve-Path (Join-Path $PSScriptRoot "..")).Path }
-if (-not $RemoteHost -and $env:TORCH_DEPLOY_PI_HOST) { $RemoteHost = $env:TORCH_DEPLOY_PI_HOST }
-if (-not $RemoteUser -and $env:TORCH_DEPLOY_PI_USER) { $RemoteUser = $env:TORCH_DEPLOY_PI_USER }
-if (-not $RemoteRoot -and $env:TORCH_DEPLOY_PI_ROOT) { $RemoteRoot = $env:TORCH_DEPLOY_PI_ROOT }
-if (-not $SshKeyPath -and $env:TORCH_DEPLOY_PI_SSH_KEY) { $SshKeyPath = $env:TORCH_DEPLOY_PI_SSH_KEY }
-if (-not $RemoteHost) { $RemoteHost = "192.168.0.185" }
-if (-not $RemoteUser) { $RemoteUser = "sevenhill" }
-if (-not $RemoteRoot) { $RemoteRoot = "/mnt/data/tradebot" }
-
-$RuntimeDir = Join-Path $RepoRoot "runtime"
-$artifactNames = @(
-    "lstm_forecaster.json",
-    "torch_retrain_guard.json",
-    "torch_threshold_calibration.json"
-)
-$localFiles = @()
-foreach ($name in $artifactNames) {
-    $path = Join-Path $RuntimeDir $name
-    if (Test-Path $path) {
-        $localFiles += (Resolve-Path $path).Path
-    }
-}
-if ($localFiles.Count -eq 0) {
-    throw "No Torch artifacts found in $RuntimeDir."
-}
-
-function ConvertTo-RemoteSingleQuoted {
-    param([string]$Value)
-    return "'" + ($Value -replace "'", "'\''") + "'"
-}
-
-function Invoke-LoggedCommand {
-    param(
-        [string]$Exe,
-        [string[]]$Arguments
-    )
-    $rendered = @($Exe) + $Arguments
-    Write-Host ($rendered -join " ")
-    if ($DryRun) {
-        return
-    }
-    & $Exe @Arguments
-    if ($LASTEXITCODE -ne 0) {
-        throw "$Exe failed with exit code $LASTEXITCODE."
-    }
-}
-
-$ssh = (Get-Command "ssh.exe" -ErrorAction SilentlyContinue)
-if (-not $ssh) { $ssh = Get-Command "ssh" -ErrorAction Stop }
-$scp = (Get-Command "scp.exe" -ErrorAction SilentlyContinue)
-if (-not $scp) { $scp = Get-Command "scp" -ErrorAction Stop }
-
-$commonSshArgs = @("-o", "BatchMode=yes", "-o", "StrictHostKeyChecking=accept-new", "-o", "ConnectTimeout=15")
-if ($SshKeyPath) {
-    $expandedKey = $ExecutionContext.SessionState.Path.GetUnresolvedProviderPathFromPSPath($SshKeyPath)
-    $commonSshArgs += @("-i", $expandedKey)
-}
-
-$remote = "${RemoteUser}@${RemoteHost}"
-$remoteRuntime = "$RemoteRoot/runtime"
-$remoteIncoming = "$remoteRuntime/.incoming-torch"
-$mkdirCommand = "mkdir -p $(ConvertTo-RemoteSingleQuoted $remoteIncoming) $(ConvertTo-RemoteSingleQuoted $remoteRuntime)"
-Invoke-LoggedCommand $ssh.Source (@($commonSshArgs + @($remote, $mkdirCommand)))
-
-$destination = "${remote}:$remoteIncoming/"
-Invoke-LoggedCommand $scp.Source (@($commonSshArgs + $localFiles + @($destination)))
-
-$moveParts = @()
-foreach ($path in $localFiles) {
-    $name = Split-Path $path -Leaf
-    $moveParts += "mv -f $(ConvertTo-RemoteSingleQuoted "$remoteIncoming/$name") $(ConvertTo-RemoteSingleQuoted "$remoteRuntime/$name")"
-}
-$moveCommand = $moveParts -join " && "
-Invoke-LoggedCommand $ssh.Source (@($commonSshArgs + @($remote, $moveCommand)))
-
-if (-not $NoRestart) {
-    $restartCommand = "cd $(ConvertTo-RemoteSingleQuoted $RemoteRoot) && docker compose restart $(ConvertTo-RemoteSingleQuoted $ServiceName)"
-    Invoke-LoggedCommand $ssh.Source (@($commonSshArgs + @($remote, $restartCommand)))
-}
-
-Write-Host "Synced Torch artifacts to ${remote}:$remoteRuntime"
@@ -54,23 +54,34 @@ def poll_once(args: argparse.Namespace, repo_root: Path, runtime_dir: Path, log_
        return
    job = claim.get("job") if isinstance(claim.get("job"), dict) else {}
    job_id = str(job.get("id") or "")
+    lease_token = str(claim.get("lease_token") or "")
    if not job_id:
        return
+    if not lease_token:
+        raise RuntimeError("training server did not issue a job lease")
    log(log_path, f"Claimed retrain job {job_id}")
-    report_progress(args, job_id, "running", "claimed", 2, "Задание получено Windows-agent")
+    report_progress(args, job_id, lease_token, "running", "claimed", 2, "Задание получено Windows-agent")
    success = False
    message = ""
    summary: dict[str, Any] = {}
    try:
-        run_retrain(args, job_id, job, repo_root, log_path)
+        run_retrain(args, job_id, lease_token, job, repo_root, log_path)
        summary = read_json(runtime_dir / "torch_retrain_guard.json")
        accepted = summary.get("accepted") is True
        if accepted:
-            report_progress(args, job_id, "running", "uploading", 72, "Обучение завершено, загружаю артефакты")
+            report_progress(
+                args,
+                job_id,
+                lease_token,
+                "running",
+                "uploading",
+                72,
+                "Обучение завершено, загружаю артефакты",
+            )
            for name in ARTIFACT_NAMES:
                path = runtime_dir / name
                if path.is_file():
-                    upload_artifact(args, job_id, path, log_path)
+                    upload_artifact(args, job_id, lease_token, path, log_path)
            message = "training completed; candidate accepted"
            log(log_path, f"Completed retrain job {job_id}; candidate accepted")
        else:
@@ -82,11 +93,23 @@ def poll_once(args: argparse.Namespace, repo_root: Path, runtime_dir: Path, log_
        message = str(exc)
        log(log_path, f"Job {job_id} failed: {message}")
    finally:
-        payload = {"success": success, "message": message, "summary": summary}
+        payload = {
+            "success": success,
+            "message": message,
+            "summary": summary,
+            "lease_token": lease_token,
+        }
        api_json(args, f"/api/training/jobs/{job_id}/complete", payload)


-def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo_root: Path, log_path: Path) -> None:
+def run_retrain(
+    args: argparse.Namespace,
+    job_id: str,
+    lease_token: str,
+    job: dict[str, Any],
+    repo_root: Path,
+    log_path: Path,
+) -> None:
    script = repo_root / "tools" / "run_torch_retrain.ps1"
    if not script.is_file():
        raise RuntimeError(f"retrain script not found: {script}")
@@ -126,12 +149,20 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
        value = parameters.get(key)
        if value not in (None, ""):
            cmd.extend([ps_arg, str(value)])
-    if parameters.get("pooled") is True:
+    if parameters.get("pooled", True) is True:
        cmd.append("-Pooled")
    if parameters.get("resume_candidate") is True:
        cmd.append("-ResumeCandidate")
    log(log_path, "Running retrain: " + " ".join(quote_for_log(part) for part in cmd))
-    report_progress(args, job_id, "running", "training", 8, "PyTorch retrain запущен")
+    report_progress(
+        args,
+        job_id,
+        lease_token,
+        "running",
+        "training",
+        8,
+        "PyTorch retrain запущен",
+    )
    line_count = 0
    output_queue: queue.Queue[str] = queue.Queue()

@@ -175,7 +206,16 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
                report_message = last_message
                if not got_line:
                    report_message = training_heartbeat_message(now, started_at, last_output_at, last_message)
-                safe_report_progress(args, job_id, "running", "training", progress, report_message, log_path)
+                safe_report_progress(
+                    args,
+                    job_id,
+                    lease_token,
+                    "running",
+                    "training",
+                    progress,
+                    report_message,
+                    log_path,
+                )
                last_report_at = now

            if process.poll() is not None and output_queue.empty():
@@ -185,7 +225,15 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
        code = process.wait()
    if code != 0:
        raise RuntimeError(f"retrain failed with exit code {code}")
-    report_progress(args, job_id, "running", "guard", 70, "Guard завершён, подготавливаю артефакты")
+    report_progress(
+        args,
+        job_id,
+        lease_token,
+        "running",
+        "guard",
+        70,
+        "Guard завершён, подготавливаю артефакты",
+    )


 def friendly_training_message(message: str) -> str:
@@ -282,7 +330,13 @@ def format_duration(seconds: float) -> str:
    return f"{seconds_part}с"


-def upload_artifact(args: argparse.Namespace, job_id: str, path: Path, log_path: Path) -> None:
+def upload_artifact(
+    args: argparse.Namespace,
+    job_id: str,
+    lease_token: str,
+    path: Path,
+    log_path: Path,
+) -> None:
    digest = hashlib.sha256(path.read_bytes()).hexdigest()
    size = path.stat().st_size
    chunk_size = max(64 * 1024, args.chunk_size)
@@ -297,16 +351,26 @@ def upload_artifact(args: argparse.Namespace, job_id: str, path: Path, log_path:
                "total": total,
                "sha256": digest,
                "data_base64": base64.b64encode(data).decode("ascii"),
+                "lease_token": lease_token,
            }
            api_json(args, f"/api/training/jobs/{job_id}/artifacts/chunk", payload, timeout=120)
            if index == 0 or index == total - 1 or index % 10 == 0:
                progress = 72 + int(((index + 1) / total) * 23)
-                report_progress(args, job_id, "running", "uploading", progress, f"Загружаю {path.name}: {index + 1}/{total}")
+                report_progress(
+                    args,
+                    job_id,
+                    lease_token,
+                    "running",
+                    "uploading",
+                    progress,
+                    f"Загружаю {path.name}: {index + 1}/{total}",
+                )


 def report_progress(
    args: argparse.Namespace,
    job_id: str,
+    lease_token: str,
    status: str,
    phase: str,
    progress_percent: int,
@@ -321,6 +385,7 @@ def report_progress(
            "progress_percent": progress_percent,
            "message": message,
            "worker": worker_payload(args, Path(args.repo_root).resolve()),
+            "lease_token": lease_token,
        },
    )

@@ -328,6 +393,7 @@ def report_progress(
 def safe_report_progress(
    args: argparse.Namespace,
    job_id: str,
+    lease_token: str,
    status: str,
    phase: str,
    progress_percent: int,
@@ -337,7 +403,15 @@ def safe_report_progress(
    last_error: Exception | None = None
    for attempt in range(1, 4):
        try:
-            report_progress(args, job_id, status, phase, progress_percent, message)
+            report_progress(
+                args,
+                job_id,
+                lease_token,
+                status,
+                phase,
+                progress_percent,
+                message,
+            )
            return
        except Exception as exc:  # noqa: BLE001 - keep the local training process alive.
            last_error = exc
@@ -385,7 +459,7 @@ def worker_payload(args: argparse.Namespace, repo_root: Path) -> dict[str, Any]:
        "worker_id": args.worker_id or f"{name}:{repo_root}",
        "name": name,
        "path": str(repo_root),
-        "version": "1",
+        "version": "2",
    }