Improve Windows training agent progress

tools/windows_training_agent.py

@@ -7,6 +7,7 @@ import json
 import os
 import platform
 import queue
+import re
 import subprocess
 import sys
 import threading
@@ -124,6 +125,7 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
         text=True,
         encoding="utf-8",
         errors="replace",
+        **hidden_subprocess_kwargs(),
     ) as process:
         reader = threading.Thread(target=read_output, name="training-output-reader", daemon=True)
         reader.start()
@@ -140,7 +142,7 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
                 log(log_path, message)
                 line_count += 1
                 if message:
-                    last_message = message[-220:]
+                    last_message = friendly_training_message(message)
             except queue.Empty:
                 pass
 
@@ -163,6 +165,78 @@ def run_retrain(args: argparse.Namespace, job_id: str, job: dict[str, Any], repo
     report_progress(args, job_id, "running", "guard", 70, "Guard завершён, подготавливаю артефакты")
 
 
+def friendly_training_message(message: str) -> str:
+    cleaned = message.strip()
+    if not cleaned:
+        return "PyTorch обучает модель"
+
+    if "Starting PyTorch recurrent retrain:" in cleaned:
+        return "PyTorch LSTM/GRU запущен: готовлю данные и варианты модели"
+
+    started = re.search(
+        r"training started: symbols=(?P<symbols>\d+) interval=(?P<interval>\d+) "
+        r"limit=(?P<limit>\d+) epochs=(?P<epochs>\d+)",
+        cleaned,
+    )
+    if started:
+        interval = started.group("interval")
+        timeframe = "1h" if interval == "60" else f"{interval}m"
+        return (
+            f"Старт обучения: {started.group('symbols')} пар, таймфрейм {timeframe}, "
+            f"история {started.group('limit')} свечей, до {started.group('epochs')} эпох"
+        )
+
+    pair_started = re.search(r"^(?P<symbol>[A-Z0-9]+): training started \((?P<index>\d+)/(?P<total>\d+)\)", cleaned)
+    if pair_started:
+        return (
+            f"{pair_started.group('symbol')}: обучение пары "
+            f"{pair_started.group('index')}/{pair_started.group('total')}"
+        )
+
+    preparing = re.search(r"^(?P<symbol>[A-Z0-9]+): preparing lookback=(?P<lookback>\d+)", cleaned)
+    if preparing:
+        return f"{preparing.group('symbol')}: готовлю окно {preparing.group('lookback')} свечей"
+
+    fitting = re.search(
+        r"^(?P<symbol>[A-Z0-9]+): fitting (?P<arch>lstm|gru) "
+        r"lookback=(?P<lookback>\d+) hidden=(?P<hidden>\d+) "
+        r"layers=(?P<layers>\d+) dropout=(?P<dropout>[0-9.]+)",
+        cleaned,
+    )
+    if fitting:
+        return (
+            f"{fitting.group('symbol')}: обучаю {fitting.group('arch').upper()}, "
+            f"окно {fitting.group('lookback')}, нейронов {fitting.group('hidden')}, "
+            f"слоёв {fitting.group('layers')}, dropout {fitting.group('dropout')}"
+        )
+
+    model = re.search(
+        r"^(?P<symbol>[A-Z0-9]+): model=torch_(?P<arch>lstm|gru).*?"
+        r"mae=(?P<mae>[0-9.]+)%.*?skill=(?P<skill>-?[0-9.]+).*?dir=(?P<direction>[0-9.]+)",
+        cleaned,
+    )
+    if model:
+        direction = float(model.group("direction")) * 100
+        skill = float(model.group("skill")) * 100
+        return (
+            f"{model.group('symbol')}: выбран {model.group('arch').upper()}, "
+            f"ошибка {model.group('mae')}%, skill {skill:.1f}%, направление {direction:.1f}%"
+        )
+
+    if "Calibrating current artifact" in cleaned:
+        return "Проверяю текущую модель на replay"
+    if "Calibrating candidate artifact" in cleaned:
+        return "Проверяю новую модель на replay"
+    if "Running retrain guard" in cleaned:
+        return "Gate сравнивает новую модель с текущей"
+    if "Candidate rejected by guard" in cleaned:
+        return "Новая модель обучилась, но gate не дал ей ходу"
+    if "Candidate accepted by guard" in cleaned:
+        return "Новая модель прошла gate и стала активной"
+
+    return cleaned[-220:]
+
+
 def training_heartbeat_message(now: float, started_at: float, last_output_at: float, last_message: str) -> str:
     elapsed = format_duration(now - started_at)
     idle_seconds = max(0.0, now - last_output_at)
@@ -313,6 +387,18 @@ def read_json(path: Path) -> dict[str, Any]:
     return data if isinstance(data, dict) else {}
 
 
+def hidden_subprocess_kwargs() -> dict[str, Any]:
+    if os.name != "nt":
+        return {}
+    startupinfo = subprocess.STARTUPINFO()
+    startupinfo.dwFlags |= subprocess.STARTF_USESHOWWINDOW
+    startupinfo.wShowWindow = 0
+    return {
+        "creationflags": getattr(subprocess, "CREATE_NO_WINDOW", 0),
+        "startupinfo": startupinfo,
+    }
+
+
 def quote_for_log(value: str) -> str:
     return f'"{value}"' if " " in value else value