feat: auto-heal checks key availability before restarting gateway

xmpp_logger.py

@@ -305,6 +305,18 @@ def auto_heal():
     h = health()
     actions = []
 
+    # 0. LLM Provider 异常 → 先查是否有更好的 Key
+    if h.get("llm_provider", {}).get("status") in ("timeout", "error"):
+        bk = best_key()
+        if bk:
+            actions.append({
+                "action": "check_keys",
+                "best_key": bk["key_id"],
+                "rolling_pct": bk["rolling"]["usage_percent"],
+                "weekly_pct": bk["weekly"]["usage_percent"],
+                "issues": bk["issues"],
+            })
+
     # 1. LLM Provider 超时 → 重启 Hermes Gateway
     if h.get("llm_provider", {}).get("status") in ("timeout", "error"):
         gw = h.get("gateways", {}).get("zhiwei", {})
@@ -341,3 +353,12 @@ def auto_heal():
                           "success": False, "detail": str(e)[:100]})
 
     return {"actions": actions, "status": h.get("status", "unknown")}
+
+
+def _verify_heal():
+    """自愈后验证：等 Gateway 启动完成，检测 LLM 是否恢复"""
+    _time.sleep(8)  # 等 Gateway 完全启动
+    h = health()
+    llm_ok = h.get("llm_provider", {}).get("status") == "ok"
+    gw_ok = h.get("gateways", {}).get("zhiwei", {}).get("alive", False)
+    return {"llm_recovered": llm_ok, "gateway_alive": gw_ok, "status": h.get("status")}