From 8f8ab15782497549506f9ceb9b71ec57aa8583d2 Mon Sep 17 00:00:00 2001
From: ShaerWare <shaerware@users.noreply.github.com>
Date: Tue, 31 Mar 2026 21:57:50 +0500
Subject: [PATCH] fix: prevent bridge from dying on failed backend switch + add
 auto-restart

Bridge was killed BEFORE verifying the target backend (vLLM) is available.
If vLLM failed to start, bridge stayed dead with no auto-recovery.
Now bridge is stopped only AFTER successful switch, and a periodic
health check (60s) auto-restarts it if it crashes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 modules/llm/router.py  | 22 ++++++++++++++--------
 modules/llm/startup.py | 26 ++++++++++++++++++++++++++
 orchestrator.py        |  5 ++++-
 3 files changed, 44 insertions(+), 9 deletions(-)

diff --git a/modules/llm/router.py b/modules/llm/router.py
index 697af37..ace4a45 100644
--- a/modules/llm/router.py
+++ b/modules/llm/router.py
@@ -392,14 +392,8 @@ async def admin_set_llm_backend(
     """Переключить LLM backend с горячей перезагрузкой сервиса"""
     container = get_container()
 
-    # Stop bridge if currently running and switching away from it
-    current_service = container.llm_service
-    if current_service and getattr(current_service, "provider_type", None) == "claude_bridge":
-        from bridge_manager import bridge_manager
-
-        if bridge_manager.is_running:
-            logger.info("🛑 Stopping bridge (switching backend)...")
-            await bridge_manager.stop()
+    # NOTE: bridge is stopped AFTER successful switch (not before),
+    # so if the target backend fails, bridge remains running.
 
     # Auto-convert "gemini" to default cloud Gemini provider
     if request.backend == "gemini":
@@ -513,6 +507,18 @@ async def check_vllm_health() -> bool:
             except ImportError:
                 raise HTTPException(status_code=503, detail="VLLMLLMService не доступен")
 
+            # Stop bridge only after vLLM is confirmed working
+            current_service = container.llm_service
+            if (
+                current_service
+                and getattr(current_service, "provider_type", None) == "claude_bridge"
+            ):
+                from bridge_manager import bridge_manager
+
+                if bridge_manager.is_running:
+                    logger.info("🛑 Stopping bridge (switching to vLLM)...")
+                    await bridge_manager.stop()
+
             container.llm_service = new_service
             os.environ["LLM_BACKEND"] = "vllm"
 
diff --git a/modules/llm/startup.py b/modules/llm/startup.py
index ef71033..5675f28 100644
--- a/modules/llm/startup.py
+++ b/modules/llm/startup.py
@@ -255,3 +255,29 @@ async def auto_start_bridge() -> None:
             logger.warning(f"🌉 Bridge auto-start failed: {result.get('error', 'unknown')}")
     except Exception as e:
         logger.error(f"🌉 Error during bridge auto-start: {e}")
+
+
+async def bridge_health_check() -> None:
+    """Periodic health check for bridge — auto-restart if crashed."""
+    from bridge_manager import bridge_manager
+    from db.integration import async_cloud_provider_manager
+
+    try:
+        bridge_providers = await async_cloud_provider_manager.get_by_type(
+            "claude_bridge", enabled_only=True
+        )
+        if not bridge_providers:
+            return
+
+        if bridge_manager.is_running:
+            return
+
+        # Bridge should be running but isn't — restart it
+        logger.warning("🌉 Bridge not running, auto-restarting...")
+        result = await bridge_manager.start()
+        if result.get("status") == "ok":
+            logger.info(f"🌉 Bridge auto-restarted on port {result.get('port', 8787)}")
+        else:
+            logger.warning(f"🌉 Bridge auto-restart failed: {result.get('error', 'unknown')}")
+    except Exception as e:
+        logger.error(f"🌉 Bridge health check error: {e}")
diff --git a/orchestrator.py b/orchestrator.py
index 48e4567..e74ac76 100644
--- a/orchestrator.py
+++ b/orchestrator.py
@@ -247,7 +247,7 @@ async def startup_event():
         # Auto-start bots and bridge
         from modules.channels.telegram.startup import auto_start_bots as auto_start_telegram
         from modules.channels.whatsapp.startup import auto_start_bots as auto_start_whatsapp
-        from modules.llm.startup import auto_start_bridge
+        from modules.llm.startup import auto_start_bridge, bridge_health_check
 
         await auto_start_telegram()
         await auto_start_whatsapp()
@@ -260,6 +260,9 @@ async def startup_event():
         from modules.kanban.tasks import sync_kanban_issues
 
         task_registry.register("session-cleanup", cleanup_expired_sessions, interval=3600)
+        task_registry.register(
+            "bridge-health-check", bridge_health_check, interval=60, initial_delay=30
+        )
         task_registry.register(
             "periodic-vacuum", periodic_vacuum, interval=7 * 24 * 3600, initial_delay=24 * 3600
         )