diff --git a/apps/api/src/services/ai_providers/nemotron.py b/apps/api/src/services/ai_providers/nemotron.py
index b3744b5a..013db91a 100644
--- a/apps/api/src/services/ai_providers/nemotron.py
+++ b/apps/api/src/services/ai_providers/nemotron.py
@@ -171,9 +171,21 @@ class NemotronProvider:
                 timeout = getattr(settings, "NEMOTRON_TIMEOUT_SECONDS", 45)
             nvidia = self._get_nvidia()
 
+            # 2026-04-05 Claude Code: I2 修正 — 補上 system prompt (NIM 標準 message format)
+            # NIM API 要求 system role 先定義模型角色，否則 tool calling 品質下降
             result = await asyncio.wait_for(
                 nvidia.tool_call(
-                    messages=[{"role": "user", "content": tool_prompt}],
+                    messages=[
+                        {
+                            "role": "system",
+                            "content": (
+                                "你是 Kubernetes 操作專家。根據 AI 診斷結果，"
+                                "使用提供的工具生成精確的 kubectl 操作指令。"
+                                "優先選擇最安全、最小影響範圍的操作。"
+                            ),
+                        },
+                        {"role": "user", "content": tool_prompt},
+                    ],
                     tools=_K8S_TOOLS,
                 ),
                 timeout=timeout,
diff --git a/apps/api/src/services/drift_detector.py b/apps/api/src/services/drift_detector.py
index c873ecc1..39c622ef 100644
--- a/apps/api/src/services/drift_detector.py
+++ b/apps/api/src/services/drift_detector.py
@@ -296,22 +296,34 @@ class DriftDetector:
 
         return diffs
 
+    @staticmethod
+    def _pattern_matches(pattern: str, field_path: str) -> bool:
+        """
+        匹配 field_path 是否符合 pattern。
+
+        支援兩種萬用字元：
+          - [*]  → 任意索引 (e.g. containers[*] 匹配 containers[0], containers[1])
+          - *    → 任意字串段
+
+        2026-04-05 Claude Code: I4 修正 — 舊邏輯直接 strip [*] 導致
+        containers[*].image 無法匹配 containers[0].image (首席架構師 Review I4)
+        """
+        import re as _re
+
+        # 將 pattern 轉為正則：[*] → \[\d+\]，* → [^.]+
+        regex = _re.escape(pattern)
+        regex = regex.replace(r"\[\*\]", r"\[\d+\]")
+        regex = regex.replace(r"\*", r"[^.]+")
+        # 允許 pattern 是前綴（field_path 可能更深，. 或 [ 或字串結尾均可）
+        return bool(_re.match(f"^{regex}(\\.|\\[|$)", field_path))
+
     def _is_allowlisted(self, field_path: str) -> bool:
         """判斷欄位是否在白名單（靜默記錄不告警）"""
-        for pattern in self._allowlist:
-            # 簡單前綴匹配（*替換為粗略包含）
-            clean_pattern = pattern.replace("[*]", "")
-            if field_path.startswith(clean_pattern.replace("*", "")):
-                return True
-        return False
+        return any(self._pattern_matches(p, field_path) for p in self._allowlist)
 
     def _is_critical(self, field_path: str) -> bool:
         """判斷欄位是否為關鍵欄位（HIGH 等級）"""
-        for pattern in self._critical_fields:
-            clean_pattern = pattern.replace("[*]", "")
-            if clean_pattern.replace("*", "") in field_path:
-                return True
-        return False
+        return any(self._pattern_matches(p, field_path) for p in self._critical_fields)
 
 
 # =============================================================================