fix: rescue gcp ollama through proxy

services/ollama_service.py

@@ -51,6 +51,8 @@ def approved_ollama_env(name: str, default: str = '') -> str:
 OLLAMA_HOST_PRIMARY   = approved_ollama_env('OLLAMA_HOST_PRIMARY',   'http://34.143.170.20:11434')
 OLLAMA_HOST_SECONDARY = approved_ollama_env('OLLAMA_HOST_SECONDARY', 'http://34.21.145.224:11434')
 OLLAMA_HOST_FALLBACK  = approved_ollama_env('OLLAMA_HOST_FALLBACK',  'http://192.168.0.111:11434')
+OLLAMA_HOST_PRIMARY_PROXY = approved_ollama_env('OLLAMA_HOST_PRIMARY_PROXY', 'http://192.168.0.110:11435')
+OLLAMA_HOST_SECONDARY_PROXY = approved_ollama_env('OLLAMA_HOST_SECONDARY_PROXY', 'http://192.168.0.110:11436')
 # 舊 OLLAMA_HOST 只接受核准主機；否則回到 primary，由 resolve_ollama_host() 管控級聯
 OLLAMA_HOST = approved_ollama_env('OLLAMA_HOST', OLLAMA_HOST_PRIMARY)
 DEFAULT_MODEL = os.getenv('OLLAMA_MODEL', 'llama3.1:8b')  # 較快速的模型
@@ -407,6 +409,36 @@ def _canonical_host_chain() -> List[str]:
     return chain
 
 
+def _proxy_rescue_for_primary(primary: str) -> str:
+    """188 直連 GCP-A 不通時，允許先走 110 primary proxy 救援同一順位。"""
+    clean_primary = _normalize_host(primary)
+    if clean_primary != _normalize_host(OLLAMA_HOST_PRIMARY):
+        return ''
+    proxy = _normalize_host(OLLAMA_HOST_PRIMARY_PROXY)
+    if proxy and proxy != clean_primary:
+        return proxy
+    return ''
+
+
+def _proxy_rescue_for_secondary(secondary: str) -> str:
+    """GCP-B direct 不通時，允許走 110 secondary proxy 再進 111。"""
+    clean_secondary = _normalize_host(secondary)
+    if clean_secondary != _normalize_host(OLLAMA_HOST_SECONDARY):
+        return ''
+    proxy = _normalize_host(OLLAMA_HOST_SECONDARY_PROXY)
+    if proxy and proxy != clean_secondary:
+        return proxy
+    return ''
+
+
+def _is_proxy_rescue_host(host: str) -> bool:
+    clean_host = _normalize_host(host)
+    return clean_host in {
+        _normalize_host(OLLAMA_HOST_PRIMARY_PROXY),
+        _normalize_host(OLLAMA_HOST_SECONDARY_PROXY),
+    }
+
+
 def _is_unhealthy(host: str) -> bool:
     """檢查 host 是否在 unhealthy TTL 內"""
     import time
@@ -473,13 +505,28 @@ def resolve_ollama_host(primary: str = OLLAMA_HOST_PRIMARY,
         except Exception:
             return False
 
-    # B4: primary 若被標 unhealthy，嘗試 secondary
+    primary_proxy = _proxy_rescue_for_primary(primary)
+    secondary_proxy = _proxy_rescue_for_secondary(secondary)
+
+    # B4: primary 若被標 unhealthy，先嘗試同順位 110 proxy，再嘗試 secondary
     if not _is_unhealthy(primary) and _is_reachable(primary):
         selected = primary
         logger.info(f"[OllamaHost] Primary 主機可用: {primary}")
+    elif primary_proxy and not _is_unhealthy(primary_proxy) and _is_reachable(primary_proxy):
+        selected = primary_proxy
+        logger.warning(
+            "[OllamaHost] Primary direct 不可用，使用 110 primary proxy: %s",
+            primary_proxy,
+        )
     elif not _is_unhealthy(secondary) and _is_reachable(secondary):
         selected = secondary
         logger.info(f"[OllamaHost] Primary 不可用，使用 Secondary: {secondary}")
+    elif secondary_proxy and not _is_unhealthy(secondary_proxy) and _is_reachable(secondary_proxy):
+        selected = secondary_proxy
+        logger.warning(
+            "[OllamaHost] Secondary direct 不可用，使用 110 secondary proxy: %s",
+            secondary_proxy,
+        )
     else:
         selected = fallback
         logger.warning(f"[OllamaHost] Primary 與 Secondary 皆無法連線，切換 Fallback: {fallback}")
@@ -668,7 +715,10 @@ class OllamaService:
                 # unhealthy TTL(30s) 時第三輪又 resolve 回 primary，導致 111
                 # final fallback 永遠沒被打到。
                 next_host = None
-                if self._explicit_host is None and current_host in allowed_hosts:
+                if (
+                    self._explicit_host is None
+                    and (current_host in allowed_hosts or _is_proxy_rescue_host(current_host))
+                ):
                     next_host = next((host for host in allowed_hosts if host not in attempted_hosts), None)
                 if not next_host:
                     # 非標準 host 或 explicit host 維持原行為：跳出避免無限迴圈。
@@ -1250,7 +1300,7 @@ class OllamaService:
             visited_hosts = attempted_hosts + skipped_hosts
             if target_host in visited_hosts:
                 next_host = None
-                if target_host in allowed_hosts:
+                if not host and (target_host in allowed_hosts or _is_proxy_rescue_host(target_host)):
                     next_host = next((candidate for candidate in allowed_hosts if candidate not in visited_hosts), None)
                 if not next_host:
                     break  # cache 還沒過期或同主機，避免無限迴圈