feat(normalizer): 稳定请求体字段顺序以提升 prompt cache 命中率

在 FormatNormalizer 基类新增 _reorder_request_keys 方法，OpenAI/OpenAI CLI
normalizer 各自定义前缀字段顺序（model, tools, messages / model, instructions,
tools, input），确保静态字段前置、动态内容后置。同时调整 OpenAI CLI 中
instructions 字段的构建位置使其在 input 之前。

src/core/api_format/conversion/normalizer.py

@@ -160,6 +160,21 @@ class FormatNormalizer(ABC):
     def _extract_extra(self, payload: dict[str, Any], known_keys: set[str]) -> dict[str, Any]:
         return {k: v for k, v in payload.items() if k not in known_keys}
 
+    @staticmethod
+    def _reorder_request_keys(
+        payload: dict[str, Any],
+        prefix_keys: tuple[str, ...],
+    ) -> dict[str, Any]:
+        """将指定字段前置，其余保持原序，用于稳定请求体前缀以提升 prompt cache 命中率。"""
+        ordered: dict[str, Any] = {}
+        for key in prefix_keys:
+            if key in payload:
+                ordered[key] = payload[key]
+        for key, value in payload.items():
+            if key not in ordered:
+                ordered[key] = value
+        return ordered
+
     def _merge_dropped(self, target: dict[str, int], source: dict[str, int]) -> None:
         for k, v in source.items():
             target[k] = target.get(k, 0) + int(v)

src/core/api_format/conversion/normalizers/openai.py

@@ -453,7 +453,7 @@ class OpenAINormalizer(FormatNormalizer):
                 if key in _CHAT_PASSTHROUGH_KEYS and key not in result and key not in _HANDLED_KEYS:
                     result[key] = value
 
-        return result
+        return self._reorder_request_prefix_keys(result)
 
     # =========================
     # Responses
@@ -1401,6 +1401,11 @@ class OpenAINormalizer(FormatNormalizer):
         joined = "\n\n".join(parts)
         return joined or None
 
+    _REQUEST_PREFIX_KEYS = ("model", "tools", "messages")
+
+    def _reorder_request_prefix_keys(self, payload: dict[str, Any]) -> dict[str, Any]:
+        return self._reorder_request_keys(payload, self._REQUEST_PREFIX_KEYS)
+
     def _openai_tools_to_internal(self, tools: Any) -> list[ToolDefinition] | None:
         if not tools or not isinstance(tools, list):
             return None

src/core/api_format/conversion/normalizers/openai_cli.py

@@ -262,11 +262,6 @@ class OpenAICliNormalizer(FormatNormalizer):
             isinstance(request_flags, dict) and request_flags.get("has_stream")
         )
 
-        result: dict[str, Any] = {
-            "model": internal.model,
-            "input": self._internal_messages_to_input(internal.messages, system_to_developer=False),
-        }
-
         # 合并 instructions，如果没有则使用 system。
         # 仅在 internal 中确有内容或原请求显式提供过时输出，
         # 让 Codex 默认 body_rules 仍可在字段缺失时注入默认 instructions。
@@ -275,8 +270,14 @@ class OpenAICliNormalizer(FormatNormalizer):
             if internal.instructions
             else internal.system
         )
+        result: dict[str, Any] = {"model": internal.model}
         if instructions_text or has_explicit_instructions:
             result["instructions"] = instructions_text or ""
+        # Keep the stable system prefix ahead of the typically dynamic input payload.
+        result["input"] = self._internal_messages_to_input(
+            internal.messages,
+            system_to_developer=False,
+        )
 
         if internal.max_tokens is not None:
             # Responses API 使用 max_output_tokens
@@ -402,7 +403,7 @@ class OpenAICliNormalizer(FormatNormalizer):
         if is_codex_variant and "store" not in result:
             result["store"] = False
 
-        return result
+        return self._reorder_request_prefix_keys(result)
 
     # =========================
     # Responses
@@ -2128,6 +2129,11 @@ class OpenAICliNormalizer(FormatNormalizer):
         joined = "\n\n".join(parts)
         return joined or None
 
+    _REQUEST_PREFIX_KEYS = ("model", "instructions", "tools", "input")
+
+    def _reorder_request_prefix_keys(self, payload: dict[str, Any]) -> dict[str, Any]:
+        return self._reorder_request_keys(payload, self._REQUEST_PREFIX_KEYS)
+
     def _error_type_from_value(self, value: str) -> ErrorType:
         for t in ErrorType:
             if t.value == value: