diff --git a/config.toml.example b/config.toml.example
index b186b773..c6705d68 100644
--- a/config.toml.example
+++ b/config.toml.example
@@ -95,8 +95,8 @@ api_mode = "chat_completions"
 # zh: 是否启用 reasoning.effort。
 # en: Enable reasoning.effort.
 reasoning_enabled = false
-# zh: reasoning.effort 档位：none / minimal / low / medium / high / xhigh。
-# en: reasoning.effort level: none / minimal / low / medium / high / xhigh.
+# zh: reasoning effort 档位。
+# en: reasoning effort level.
 reasoning_effort = "medium"
 # zh: 是否启用 thinking（思维链）。
 # en: Enable thinking (reasoning).
@@ -107,6 +107,9 @@ thinking_budget_tokens = 20000
 # zh: 是否在请求中发送 budget_tokens（关闭后由提供商决定思维预算）。
 # en: Whether to include budget_tokens in the request (if disabled, the provider decides the thinking budget).
 thinking_include_budget = true
+# zh: reasoning effort 传参风格：openai（reasoning.effort）/ anthropic（output_config.effort）。
+# en: Reasoning effort wire format: openai (reasoning.effort) / anthropic (output_config.effort).
+reasoning_effort_style = "openai"
 # zh: 思维链工具调用兼容：启用后在多轮工具调用中回传 reasoning_content，避免部分模型返回 400。
 # en: Thinking tool-call compatibility: pass back reasoning_content in multi-turn tool calls to avoid 400 errors from some models.
 thinking_tool_call_compat = true
@@ -155,8 +158,8 @@ api_mode = "chat_completions"
 # zh: 是否启用 reasoning.effort。
 # en: Enable reasoning.effort.
 reasoning_enabled = false
-# zh: reasoning.effort 档位：none / minimal / low / medium / high / xhigh。
-# en: reasoning.effort level: none / minimal / low / medium / high / xhigh.
+# zh: reasoning effort 档位。
+# en: reasoning effort level.
 reasoning_effort = "medium"
 # zh: 是否启用 thinking（思维链）。
 # en: Enable thinking (reasoning).
@@ -167,6 +170,9 @@ thinking_budget_tokens = 20000
 # zh: 是否在请求中发送 budget_tokens（关闭后由提供商决定思维预算）。
 # en: Whether to include budget_tokens in the request (if disabled, the provider decides the thinking budget).
 thinking_include_budget = true
+# zh: reasoning effort 传参风格：openai（reasoning.effort）/ anthropic（output_config.effort）。
+# en: Reasoning effort wire format: openai (reasoning.effort) / anthropic (output_config.effort).
+reasoning_effort_style = "openai"
 # zh: 思维链工具调用兼容：启用后在多轮工具调用中回传 reasoning_content，避免部分模型返回 400。
 # en: Thinking tool-call compatibility: pass back reasoning_content in multi-turn tool calls to avoid 400 errors from some models.
 thinking_tool_call_compat = true
@@ -208,8 +214,8 @@ api_mode = "chat_completions"
 # zh: 是否启用 reasoning.effort。
 # en: Enable reasoning.effort.
 reasoning_enabled = false
-# zh: reasoning.effort 档位：none / minimal / low / medium / high / xhigh。
-# en: reasoning.effort level: none / minimal / low / medium / high / xhigh.
+# zh: reasoning effort 档位。
+# en: reasoning effort level.
 reasoning_effort = "medium"
 # zh: 是否启用 thinking（思维链）。
 # en: Enable thinking (reasoning).
@@ -220,6 +226,9 @@ thinking_budget_tokens = 0
 # zh: 是否在请求中发送 budget_tokens（关闭后由提供商决定思维预算）。
 # en: Whether to include budget_tokens in the request (if disabled, the provider decides the thinking budget).
 thinking_include_budget = true
+# zh: reasoning effort 传参风格：openai（reasoning.effort）/ anthropic（output_config.effort）。
+# en: Reasoning effort wire format: openai (reasoning.effort) / anthropic (output_config.effort).
+reasoning_effort_style = "openai"
 # zh: 思维链工具调用兼容：启用后在多轮工具调用中回传 reasoning_content，避免部分模型返回 400。
 # en: Thinking tool-call compatibility: pass back reasoning_content in multi-turn tool calls to avoid 400 errors from some models.
 thinking_tool_call_compat = true
@@ -258,8 +267,8 @@ api_mode = "chat_completions"
 # zh: 是否启用 reasoning.effort。
 # en: Enable reasoning.effort.
 reasoning_enabled = false
-# zh: reasoning.effort 档位：none / minimal / low / medium / high / xhigh。
-# en: reasoning.effort level: none / minimal / low / medium / high / xhigh.
+# zh: reasoning effort 档位。
+# en: reasoning effort level.
 reasoning_effort = "medium"
 # zh: 是否启用 thinking（思维链）。
 # en: Enable thinking (reasoning).
@@ -270,6 +279,9 @@ thinking_budget_tokens = 0
 # zh: 是否在请求中发送 budget_tokens（关闭后由提供商决定思维预算）。
 # en: Whether to include budget_tokens in the request (if disabled, the provider decides the thinking budget).
 thinking_include_budget = true
+# zh: reasoning effort 传参风格：openai（reasoning.effort）/ anthropic（output_config.effort）。
+# en: Reasoning effort wire format: openai (reasoning.effort) / anthropic (output_config.effort).
+reasoning_effort_style = "openai"
 # zh: 思维链工具调用兼容：启用后在多轮工具调用中回传 reasoning_content，避免部分模型返回 400。
 # en: Thinking tool-call compatibility: pass back reasoning_content in multi-turn tool calls to avoid 400 errors from some models.
 thinking_tool_call_compat = true
@@ -321,8 +333,8 @@ api_mode = "chat_completions"
 # zh: 是否启用 reasoning.effort。
 # en: Enable reasoning.effort.
 reasoning_enabled = false
-# zh: reasoning.effort 档位：none / minimal / low / medium / high / xhigh。
-# en: reasoning.effort level: none / minimal / low / medium / high / xhigh.
+# zh: reasoning effort 档位。
+# en: reasoning effort level.
 reasoning_effort = "medium"
 # zh: 是否启用 thinking（思维链）。
 # en: Enable thinking (reasoning).
@@ -333,6 +345,9 @@ thinking_budget_tokens = 0
 # zh: 是否在请求中发送 budget_tokens（关闭后由提供商决定思维预算）。
 # en: Whether to include budget_tokens in the request (if disabled, the provider decides the thinking budget).
 thinking_include_budget = true
+# zh: reasoning effort 传参风格：openai（reasoning.effort）/ anthropic（output_config.effort）。
+# en: Reasoning effort wire format: openai (reasoning.effort) / anthropic (output_config.effort).
+reasoning_effort_style = "openai"
 # zh: 思维链工具调用兼容：启用后在多轮工具调用中回传 reasoning_content，避免部分模型返回 400。
 # en: Thinking tool-call compatibility: pass back reasoning_content in multi-turn tool calls to avoid 400 errors from some models.
 thinking_tool_call_compat = true
diff --git a/scripts/sync_config_template.py b/scripts/sync_config_template.py
index 93cff1bb..e4320012 100755
--- a/scripts/sync_config_template.py
+++ b/scripts/sync_config_template.py
@@ -39,16 +39,49 @@ def build_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="将同步后的完整 TOML 输出到标准输出。",
     )
+    parser.add_argument(
+        "--prune",
+        action="store_true",
+        help="删除存在于 config.toml 但不存在于 config.toml.example 中的配置项（危险操作，需二次确认）。",
+    )
     return parser
 
 
+def _confirm_prune(removed_paths: list[str]) -> bool:
+    """显示即将删除的路径并请求用户二次确认。"""
+    print(
+        "\n\033[1;31m[sync-config] ⚠ 危险操作：以下配置项不存在于模板中，将被永久删除：\033[0m"
+    )
+    for path in removed_paths:
+        print(f"  \033[31m- {path}\033[0m")
+    print()
+    try:
+        answer = input(
+            "\033[1;33m确认删除以上配置项？此操作不可撤销。输入 yes 确认: \033[0m"
+        )
+    except (EOFError, KeyboardInterrupt):
+        print()
+        return False
+    return answer.strip().lower() == "yes"
+
+
+def _initial_action_label(*, dry_run: bool, prune: bool) -> str:
+    if dry_run:
+        return "预览完成"
+    if prune:
+        return "分析完成"
+    return "同步完成"
+
+
 def main() -> int:
     args = build_parser().parse_args()
+
+    # 第一轮：不带 prune 的常规同步（或 dry-run 预览）
     try:
         result = sync_config_file(
             config_path=args.config,
             example_path=args.example,
-            write=not args.dry_run,
+            write=not args.dry_run and not args.prune,
         )
     except FileNotFoundError as exc:
         print(f"[sync-config] 未找到示例配置：{exc}", file=sys.stderr)
@@ -57,12 +90,53 @@ def main() -> int:
         print(f"[sync-config] 配置解析失败：{exc}", file=sys.stderr)
         return 1
 
-    action = "预览完成" if args.dry_run else "同步完成"
+    action = _initial_action_label(dry_run=args.dry_run, prune=args.prune)
     print(f"[sync-config] {action}: {args.config}")
     print(f"[sync-config] 新增路径数量: {len(result.added_paths)}")
     for path in result.added_paths:
         print(f"  + {path}")
 
+    if result.removed_paths:
+        print(f"[sync-config] 多余路径数量: {len(result.removed_paths)}")
+        for path in result.removed_paths:
+            print(f"  - {path}")
+
+    # --prune 流程：确认后带 prune 重新同步
+    if args.prune and result.removed_paths:
+        if args.dry_run:
+            print("\n[sync-config] --dry-run 模式，跳过删除。")
+        elif _confirm_prune(result.removed_paths):
+            result = sync_config_file(
+                config_path=args.config,
+                example_path=args.example,
+                write=True,
+                prune=True,
+            )
+            print(
+                f"\033[1;32m[sync-config] 已删除 {len(result.removed_paths)} 个多余配置项并写回文件。\033[0m"
+            )
+        else:
+            # 用户取消 prune，仍执行不带 prune 的常规同步
+            sync_config_file(
+                config_path=args.config,
+                example_path=args.example,
+                write=True,
+                prune=False,
+            )
+            print("[sync-config] 已取消删除，仅执行常规同步。")
+    elif args.prune and not result.removed_paths:
+        if not args.dry_run:
+            # 无多余项但仍需写入常规同步结果
+            sync_config_file(
+                config_path=args.config,
+                example_path=args.example,
+                write=True,
+                prune=False,
+            )
+            print("[sync-config] 无多余配置项需要删除，已完成常规同步。")
+        else:
+            print("[sync-config] 无多余配置项需要删除。")
+
     if args.stdout:
         print("\n--- merged config.toml ---\n")
         print(result.content, end="")
diff --git a/src/Undefined/ai/llm.py b/src/Undefined/ai/llm.py
index 41d93b61..931f13bc 100644
--- a/src/Undefined/ai/llm.py
+++ b/src/Undefined/ai/llm.py
@@ -25,7 +25,9 @@
     API_MODE_RESPONSES,
     build_responses_request_body,
     get_api_mode,
-    get_reasoning_payload,
+    get_effort_payload,
+    get_effort_style,
+    get_thinking_payload,
     normalize_responses_result,
 )
 from Undefined.ai.retrieval import RetrievalRequester
@@ -117,8 +119,10 @@
             "tool_choice",
             "stream",
             "stream_options",
+            "thinking",
             "reasoning",
             "reasoning_effort",
+            "output_config",
         }
     )
     | _SDK_REQUEST_OPTION_FIELDS
@@ -139,6 +143,7 @@
             "thinking",
             "reasoning",
             "reasoning_effort",
+            "output_config",
         }
     )
     | _SDK_REQUEST_OPTION_FIELDS
@@ -870,6 +875,8 @@ def _build_effective_request_kwargs(
         getattr(model_config, "request_params", {}),
         overrides,
     )
+    thinking_override = overrides["thinking"] if "thinking" in overrides else None
+    has_thinking_override = "thinking" in overrides
     reserved_fields = (
         _RESPONSES_RESERVED_FIELDS
         if get_api_mode(model_config) == API_MODE_RESPONSES
@@ -879,11 +886,15 @@ def _build_effective_request_kwargs(
         merged,
         reserved_fields,
     )
+    if has_thinking_override:
+        ignored.pop("thinking", None)
     _warn_ignored_request_params(
         call_type=call_type,
         model_name=model_config.model_name,
         ignored=ignored,
     )
+    if has_thinking_override:
+        allowed["thinking"] = thinking_override
     return allowed
 
 
@@ -1447,12 +1458,20 @@ def build_request_body(
     """构建 API 请求体。"""
     api_mode = get_api_mode(model_config)
     extra_kwargs: dict[str, Any] = dict(kwargs)
-    reasoning_payload = get_reasoning_payload(model_config)
+
+    if "thinking" in extra_kwargs:
+        normalized = _normalize_thinking_override(
+            extra_kwargs.get("thinking"), model_config
+        )
+        if normalized is None:
+            extra_kwargs.pop("thinking", None)
+        else:
+            extra_kwargs["thinking"] = normalized
 
     if api_mode == API_MODE_RESPONSES:
-        extra_kwargs.pop("thinking", None)
         extra_kwargs.pop("reasoning", None)
         extra_kwargs.pop("reasoning_effort", None)
+        extra_kwargs.pop("output_config", None)
         return build_responses_request_body(
             model_config,
             messages,
@@ -1470,28 +1489,21 @@ def build_request_body(
         "max_tokens": max_tokens,
     }
 
-    if "thinking" in extra_kwargs:
-        normalized = _normalize_thinking_override(
-            extra_kwargs.get("thinking"), model_config
-        )
-        if normalized is None:
-            extra_kwargs.pop("thinking", None)
-        else:
-            extra_kwargs["thinking"] = normalized
-
     extra_kwargs.pop("reasoning", None)
     extra_kwargs.pop("reasoning_effort", None)
+    extra_kwargs.pop("output_config", None)
 
-    if getattr(model_config, "thinking_enabled", False):
-        thinking_param: dict[str, Any] = {"type": "enabled"}
-        if getattr(model_config, "thinking_include_budget", True):
-            thinking_param["budget_tokens"] = getattr(
-                model_config, "thinking_budget_tokens", 0
-            )
-        body["thinking"] = thinking_param
+    thinking = get_thinking_payload(model_config)
+    if thinking is not None:
+        body["thinking"] = thinking
 
-    if reasoning_payload is not None:
-        body["reasoning"] = reasoning_payload
+    effort_payload = get_effort_payload(model_config)
+    if effort_payload is not None:
+        style = get_effort_style(model_config)
+        if style == "anthropic":
+            body["output_config"] = effort_payload
+        else:
+            body["reasoning"] = effort_payload
 
     if tools:
         body["tools"] = tools
diff --git a/src/Undefined/ai/transports/__init__.py b/src/Undefined/ai/transports/__init__.py
index 9f33ccf8..74f527fc 100644
--- a/src/Undefined/ai/transports/__init__.py
+++ b/src/Undefined/ai/transports/__init__.py
@@ -5,7 +5,10 @@
     API_MODE_RESPONSES,
     build_responses_request_body,
     get_api_mode,
+    get_effort_payload,
+    get_effort_style,
     get_reasoning_payload,
+    get_thinking_payload,
     normalize_api_mode,
     normalize_reasoning_effort,
     normalize_responses_result,
@@ -16,7 +19,10 @@
     "API_MODE_RESPONSES",
     "build_responses_request_body",
     "get_api_mode",
+    "get_effort_payload",
+    "get_effort_style",
     "get_reasoning_payload",
+    "get_thinking_payload",
     "normalize_api_mode",
     "normalize_reasoning_effort",
     "normalize_responses_result",
diff --git a/src/Undefined/ai/transports/openai_transport.py b/src/Undefined/ai/transports/openai_transport.py
index 5c6d7e7e..c5ab32a2 100644
--- a/src/Undefined/ai/transports/openai_transport.py
+++ b/src/Undefined/ai/transports/openai_transport.py
@@ -6,7 +6,6 @@
 API_MODE_CHAT_COMPLETIONS = "chat_completions"
 API_MODE_RESPONSES = "responses"
 _VALID_API_MODES = {API_MODE_CHAT_COMPLETIONS, API_MODE_RESPONSES}
-_VALID_REASONING_EFFORTS = {"none", "minimal", "low", "medium", "high", "xhigh"}
 
 
 def normalize_api_mode(value: Any, default: str = API_MODE_CHAT_COMPLETIONS) -> str:
@@ -23,13 +22,28 @@ def get_api_mode(model_config: Any) -> str:
 
 
 def normalize_reasoning_effort(value: Any, default: str = "medium") -> str:
-    text = str(value or default).strip().lower()
-    if text not in _VALID_REASONING_EFFORTS:
-        return default
-    return text
+    return str(value or default).strip().lower()
 
 
 def get_reasoning_payload(model_config: Any) -> dict[str, Any] | None:
+    return get_effort_payload(model_config)
+
+
+_VALID_REASONING_EFFORT_STYLES = {"openai", "anthropic"}
+
+
+def get_thinking_payload(model_config: Any) -> dict[str, Any] | None:
+    """构建 thinking 请求参数，仅由 thinking_* 配置控制。"""
+    if not bool(getattr(model_config, "thinking_enabled", False)):
+        return None
+    param: dict[str, Any] = {"type": "enabled"}
+    if bool(getattr(model_config, "thinking_include_budget", True)):
+        param["budget_tokens"] = int(getattr(model_config, "thinking_budget_tokens", 0))
+    return param
+
+
+def get_effort_payload(model_config: Any) -> dict[str, Any] | None:
+    """构建 effort 请求参数（仅在 reasoning_enabled 启用时生效）。"""
     if not bool(getattr(model_config, "reasoning_enabled", False)):
         return None
     return {
@@ -39,6 +53,15 @@ def get_reasoning_payload(model_config: Any) -> dict[str, Any] | None:
     }
 
 
+def get_effort_style(model_config: Any) -> str:
+    style = (
+        str(getattr(model_config, "reasoning_effort_style", "openai") or "openai")
+        .strip()
+        .lower()
+    )
+    return style if style in _VALID_REASONING_EFFORT_STYLES else "openai"
+
+
 def _stringify_content(value: Any) -> str:
     if value is None:
         return ""
@@ -334,9 +357,16 @@ def build_responses_request_body(
         "model": getattr(model_config, "model_name"),
         "max_output_tokens": max_tokens,
     }
-    reasoning = get_reasoning_payload(model_config)
-    if reasoning is not None:
-        body["reasoning"] = reasoning
+    thinking = get_thinking_payload(model_config)
+    effort_payload = get_effort_payload(model_config)
+    if effort_payload is not None:
+        style = get_effort_style(model_config)
+        if style == "anthropic":
+            body["output_config"] = effort_payload
+        else:
+            body["reasoning"] = effort_payload
+    if thinking is not None:
+        body["thinking"] = thinking
     if tools:
         normalized_tools = _normalize_responses_tools(tools, internal_to_api)
         normalized_tool_choice, selected_tool_name = _normalize_responses_tool_choice(
diff --git a/src/Undefined/config/loader.py b/src/Undefined/config/loader.py
index 2d36a049..7ff8f5f2 100644
--- a/src/Undefined/config/loader.py
+++ b/src/Undefined/config/loader.py
@@ -257,7 +257,14 @@ def _get_value(
 
 
 _VALID_API_MODES = {"chat_completions", "responses"}
-_VALID_REASONING_EFFORTS = {"none", "minimal", "low", "medium", "high", "xhigh"}
+_VALID_REASONING_EFFORT_STYLES = {"openai", "anthropic"}
+
+
+def _resolve_reasoning_effort_style(value: Any, default: str = "openai") -> str:
+    style = _coerce_str(value, default).strip().lower()
+    if style not in _VALID_REASONING_EFFORT_STYLES:
+        return default
+    return style
 
 
 def _resolve_thinking_compat_flags(
@@ -311,10 +318,7 @@ def _resolve_api_mode(
 
 
 def _resolve_reasoning_effort(value: Any, default: str = "medium") -> str:
-    effort = _coerce_str(value, default).strip().lower()
-    if effort not in _VALID_REASONING_EFFORTS:
-        return default
-    return effort
+    return _coerce_str(value, default).strip().lower()
 
 
 def _resolve_responses_tool_choice_compat(
@@ -1599,6 +1603,10 @@ def _parse_model_pool(
                         item.get("thinking_include_budget"),
                         primary_config.thinking_include_budget,
                     ),
+                    reasoning_effort_style=_resolve_reasoning_effort_style(
+                        item.get("reasoning_effort_style"),
+                        primary_config.reasoning_effort_style,
+                    ),
                     thinking_tool_call_compat=_coerce_bool(
                         item.get("thinking_tool_call_compat"),
                         primary_config.thinking_tool_call_compat,
@@ -1785,6 +1793,13 @@ def _parse_chat_model_config(data: dict[str, Any]) -> ChatModelConfig:
                 20000,
             ),
             thinking_include_budget=thinking_include_budget,
+            reasoning_effort_style=_resolve_reasoning_effort_style(
+                _get_value(
+                    data,
+                    ("models", "chat", "reasoning_effort_style"),
+                    "CHAT_MODEL_REASONING_EFFORT_STYLE",
+                ),
+            ),
             thinking_tool_call_compat=thinking_tool_call_compat,
             responses_tool_choice_compat=responses_tool_choice_compat,
             responses_force_stateless_replay=responses_force_stateless_replay,
@@ -1877,6 +1892,13 @@ def _parse_vision_model_config(data: dict[str, Any]) -> VisionModelConfig:
                 20000,
             ),
             thinking_include_budget=thinking_include_budget,
+            reasoning_effort_style=_resolve_reasoning_effort_style(
+                _get_value(
+                    data,
+                    ("models", "vision", "reasoning_effort_style"),
+                    "VISION_MODEL_REASONING_EFFORT_STYLE",
+                ),
+            ),
             thinking_tool_call_compat=thinking_tool_call_compat,
             responses_tool_choice_compat=responses_tool_choice_compat,
             responses_force_stateless_replay=responses_force_stateless_replay,
@@ -1983,6 +2005,13 @@ def _parse_security_model_config(
                     0,
                 ),
                 thinking_include_budget=thinking_include_budget,
+                reasoning_effort_style=_resolve_reasoning_effort_style(
+                    _get_value(
+                        data,
+                        ("models", "security", "reasoning_effort_style"),
+                        "SECURITY_MODEL_REASONING_EFFORT_STYLE",
+                    ),
+                ),
                 thinking_tool_call_compat=thinking_tool_call_compat,
                 responses_tool_choice_compat=responses_tool_choice_compat,
                 responses_force_stateless_replay=responses_force_stateless_replay,
@@ -2002,6 +2031,7 @@ def _parse_security_model_config(
             thinking_enabled=False,
             thinking_budget_tokens=0,
             thinking_include_budget=True,
+            reasoning_effort_style="openai",
             thinking_tool_call_compat=chat_model.thinking_tool_call_compat,
             responses_tool_choice_compat=chat_model.responses_tool_choice_compat,
             responses_force_stateless_replay=chat_model.responses_force_stateless_replay,
@@ -2092,6 +2122,13 @@ def _parse_agent_model_config(data: dict[str, Any]) -> AgentModelConfig:
                 0,
             ),
             thinking_include_budget=thinking_include_budget,
+            reasoning_effort_style=_resolve_reasoning_effort_style(
+                _get_value(
+                    data,
+                    ("models", "agent", "reasoning_effort_style"),
+                    "AGENT_MODEL_REASONING_EFFORT_STYLE",
+                ),
+            ),
             thinking_tool_call_compat=thinking_tool_call_compat,
             responses_tool_choice_compat=responses_tool_choice_compat,
             responses_force_stateless_replay=responses_force_stateless_replay,
@@ -2268,6 +2305,14 @@ def _parse_historian_model_config(
                 h.get("thinking_budget_tokens"), fallback.thinking_budget_tokens
             ),
             thinking_include_budget=thinking_include_budget,
+            reasoning_effort_style=_resolve_reasoning_effort_style(
+                _get_value(
+                    {"models": {"historian": h}},
+                    ("models", "historian", "reasoning_effort_style"),
+                    "HISTORIAN_MODEL_REASONING_EFFORT_STYLE",
+                ),
+                fallback.reasoning_effort_style,
+            ),
             thinking_tool_call_compat=thinking_tool_call_compat,
             responses_tool_choice_compat=responses_tool_choice_compat,
             responses_force_stateless_replay=responses_force_stateless_replay,
diff --git a/src/Undefined/config/models.py b/src/Undefined/config/models.py
index 9454a9b9..3b073d11 100644
--- a/src/Undefined/config/models.py
+++ b/src/Undefined/config/models.py
@@ -38,6 +38,7 @@ class ModelPoolEntry:
     thinking_enabled: bool = False
     thinking_budget_tokens: int = 0
     thinking_include_budget: bool = True
+    reasoning_effort_style: str = "openai"  # effort 传参风格：openai / anthropic
     thinking_tool_call_compat: bool = True
     responses_tool_choice_compat: bool = False
     responses_force_stateless_replay: bool = False
@@ -68,6 +69,7 @@ class ChatModelConfig:
     thinking_enabled: bool = False  # 是否启用 thinking
     thinking_budget_tokens: int = 20000  # 思维预算 token 数量
     thinking_include_budget: bool = True  # 是否在请求中发送 budget_tokens
+    reasoning_effort_style: str = "openai"  # effort 传参风格：openai / anthropic
     thinking_tool_call_compat: bool = (
         True  # 思维链 + 工具调用兼容（回传 reasoning_content）
     )
@@ -93,6 +95,7 @@ class VisionModelConfig:
     thinking_enabled: bool = False  # 是否启用 thinking
     thinking_budget_tokens: int = 20000  # 思维预算 token 数量
     thinking_include_budget: bool = True  # 是否在请求中发送 budget_tokens
+    reasoning_effort_style: str = "openai"  # effort 传参风格：openai / anthropic
     thinking_tool_call_compat: bool = (
         True  # 思维链 + 工具调用兼容（回传 reasoning_content）
     )
@@ -118,6 +121,7 @@ class SecurityModelConfig:
     thinking_enabled: bool = False  # 是否启用 thinking
     thinking_budget_tokens: int = 0  # 思维预算 token 数量
     thinking_include_budget: bool = True  # 是否在请求中发送 budget_tokens
+    reasoning_effort_style: str = "openai"  # effort 传参风格：openai / anthropic
     thinking_tool_call_compat: bool = (
         True  # 思维链 + 工具调用兼容（回传 reasoning_content）
     )
@@ -169,6 +173,7 @@ class AgentModelConfig:
     thinking_enabled: bool = False  # 是否启用 thinking
     thinking_budget_tokens: int = 0  # 思维预算 token 数量
     thinking_include_budget: bool = True  # 是否在请求中发送 budget_tokens
+    reasoning_effort_style: str = "openai"  # effort 传参风格：openai / anthropic
     thinking_tool_call_compat: bool = (
         True  # 思维链 + 工具调用兼容（回传 reasoning_content）
     )
diff --git a/src/Undefined/skills/agents/README.md b/src/Undefined/skills/agents/README.md
index 8f46a53a..81f7d794 100644
--- a/src/Undefined/skills/agents/README.md
+++ b/src/Undefined/skills/agents/README.md
@@ -44,8 +44,8 @@ responses_force_stateless_replay = false
 ```
 
 说明：
-- `api_mode = "chat_completions"` 时，旧 `thinking_*` 仍按原逻辑生效；若开启 `reasoning_enabled`，也会额外发送 `reasoning.effort`。
-- `api_mode = "responses"` 时，Agent 的多轮工具调用默认使用 `previous_response_id + function_call_output` 续轮；若开启 `responses_force_stateless_replay`，则会始终改为完整消息重放；旧 `thinking_*` 不会发到 `responses`。
+- `api_mode = "chat_completions"` 时，`thinking_*` 仍按原逻辑生效；若开启 `reasoning_enabled`，会额外发送 `reasoning.effort`。
+- `api_mode = "responses"` 时，`thinking_*` 与 `reasoning_*` 分别独立控制 `thinking` 和 `reasoning.effort` / `output_config.effort`；Agent 的多轮工具调用默认使用 `previous_response_id + function_call_output` 续轮；若开启 `responses_force_stateless_replay`，则会始终改为完整消息重放。
 - `thinking_tool_call_compat` 默认 `true`，会把 `reasoning_content` 回填到本地消息历史，便于日志、回放和兼容读取。
 
 兼容的环境变量（会覆盖 `config.toml`）：
diff --git a/src/Undefined/skills/commands/naga/README.md b/src/Undefined/skills/commands/naga/README.md
new file mode 100644
index 00000000..a7c285a5
--- /dev/null
+++ b/src/Undefined/skills/commands/naga/README.md
@@ -0,0 +1,66 @@
+# /naga 命令说明
+
+## 这是什么？
+
+NagaAgent 是一个可以接入 Undefined 的外部 AI 助手。
+通过 `/naga` 命令，你可以把自己的 NagaAgent 绑定到 QQ 群，绑定之后，NagaAgent 里的特定功能就可以向你发送消息。
+
+## 普通用户
+
+普通用户只需要用到一个子命令：`bind`（绑定）。
+
+### 如何绑定？
+
+1. 在**群聊**中发送：`/naga bind <你的naga_id>`
+2. 系统会提示"申请已提交，等待超管审核"
+3. 超级管理员审核通过后，你会收到私聊通知
+4. 绑定完成！你的 NagaAgent 即可开始使用
+
+### 注意事项
+
+- `naga_id` 是你在 NagaAgent 中设置的标识，不是 QQ 号
+- 每个 `naga_id` 只能绑定一次，不能重复申请
+- 如果已在审核队列中，无需重复提交
+
+## 管理员命令（仅超级管理员）
+
+以下命令仅超级管理员可使用，用于管理所有绑定：
+
+| 子命令 | 用法 | 说明 |
+|--------|------|------|
+| approve | `/naga approve <naga_id>` | 通过绑定申请，系统会自动生成 Token 并通知申请人 |
+| reject | `/naga reject <naga_id>` | 拒绝绑定申请，申请人会收到私聊通知 |
+| revoke | `/naga revoke <naga_id>` | 吊销已有绑定，该 NagaAgent 将无法继续使用 |
+| list | `/naga list` | 查看所有活跃的绑定（含使用次数） |
+| pending | `/naga pending` | 查看等待审核的申请列表 |
+| info | `/naga info <naga_id>` | 查看指定绑定的详细信息（Token、使用次数、创建时间等） |
+
+## 完整示例
+
+```
+# 普通用户：在群聊中提交绑定申请
+/naga bind my-naga-001
+
+# 超级管理员：查看待审核列表
+/naga pending
+
+# 超级管理员：通过申请
+/naga approve my-naga-001
+
+# 超级管理员：查看绑定详情
+/naga info my-naga-001
+
+# 超级管理员：吊销绑定
+/naga revoke my-naga-001
+```
+
+## 常见问题
+
+**Q: 提示"Naga 集成未启用"？**
+A: 请联系管理员开启相关配置开关。
+
+**Q: 在群里发了 /naga bind 没有任何反应？**
+A: 该群可能不在白名单中，请联系管理员添加。
+
+**Q: 绑定通过后 NagaAgent 怎么用？**
+A: 请参考 NagaAgent 相关文档，填入QQ号以及其他几个参数即可完成对接。
\ No newline at end of file
diff --git a/src/Undefined/skills/commands/naga/config.json b/src/Undefined/skills/commands/naga/config.json
index 5086e70f..2f840a79 100644
--- a/src/Undefined/skills/commands/naga/config.json
+++ b/src/Undefined/skills/commands/naga/config.json
@@ -1,7 +1,8 @@
 {
   "name": "naga",
-  "description": "NagaAgent 集成管理",
-  "usage": "/naga <bind|approve|reject|revoke|list|pending|info> [参数]",
+  "description": "NagaAgent 联动命令",
+  "usage": "/naga <子命令> [参数]",
+  "example": "/naga bind my-naga",
   "permission": "public",
   "allow_in_private": true,
   "show_in_help": true,
diff --git a/src/Undefined/webui/routes/_config.py b/src/Undefined/webui/routes/_config.py
index a284d807..0afeaa01 100644
--- a/src/Undefined/webui/routes/_config.py
+++ b/src/Undefined/webui/routes/_config.py
@@ -1,3 +1,4 @@
+import asyncio
 import tomllib
 from pathlib import Path
 from tempfile import NamedTemporaryFile
@@ -151,16 +152,26 @@ async def config_patch_handler(request: web.Request) -> Response:
 async def sync_config_template_handler(request: web.Request) -> Response:
     if not check_auth(request):
         return web.json_response({"error": "Unauthorized"}, status=401)
+    prune = request.query.get("prune") == "true"
+    write = request.query.get("write", "true").lower() != "false"
     try:
-        result = sync_config_file()
-        get_config_manager().reload()
-        validation_ok, validation_msg = validate_required_config()
+        result = await asyncio.to_thread(sync_config_file, prune=prune, write=write)
+        validation_ok = True
+        validation_msg: str | None = None
+        if write:
+            await asyncio.to_thread(get_config_manager().reload)
+            validation_ok, validation_msg = await asyncio.to_thread(
+                validate_required_config
+            )
         return web.json_response(
             {
                 "success": True,
                 "message": "Synced",
+                "preview": not write,
                 "added_paths": result.added_paths,
                 "added_count": len(result.added_paths),
+                "removed_paths": result.removed_paths,
+                "removed_count": len(result.removed_paths),
                 "warning": None if validation_ok else validation_msg,
             }
         )
diff --git a/src/Undefined/webui/static/js/config-form.js b/src/Undefined/webui/static/js/config-form.js
index 328b6cf6..aafb0048 100644
--- a/src/Undefined/webui/static/js/config-form.js
+++ b/src/Undefined/webui/static/js/config-form.js
@@ -251,7 +251,7 @@ function isLongText(value) {
 
 const FIELD_SELECT_OPTIONS = {
     api_mode: ["chat_completions", "responses"],
-    reasoning_effort: ["none", "minimal", "low", "medium", "high", "xhigh"],
+    reasoning_effort_style: ["openai", "anthropic"],
 };
 
 function getFieldSelectOptions(path) {
@@ -905,6 +905,14 @@ function buildAotTemplate(path, arr) {
             ) {
                 template.reasoning_effort = "medium";
             }
+            if (
+                !Object.prototype.hasOwnProperty.call(
+                    template,
+                    "reasoning_effort_style",
+                )
+            ) {
+                template.reasoning_effort_style = "openai";
+            }
         }
         return template;
     }
@@ -914,6 +922,7 @@ function buildAotTemplate(path, arr) {
         api_key: "",
         api_mode: "chat_completions",
         thinking_tool_call_compat: true,
+        reasoning_effort_style: "openai",
         responses_tool_choice_compat: false,
         responses_force_stateless_replay: false,
         reasoning_enabled: false,
@@ -1058,30 +1067,74 @@ async function syncConfigTemplate(button) {
     setButtonLoading(button, true);
     showSaveStatus("saving", t("config.syncing"));
     try {
-        const res = await api("/api/config/sync-template", { method: "POST" });
-        const data = await res.json();
-        if (!data.success) {
+        const previewRes = await api("/api/config/sync-template?write=false", {
+            method: "POST",
+        });
+        const previewData = await previewRes.json();
+        if (!previewData.success) {
             showSaveStatus("error", t("config.save_error"));
             showToast(
-                `${t("common.error")}: ${data.error || t("config.sync_error")}`,
+                `${t("common.error")}: ${previewData.error || t("config.sync_error")}`,
                 "error",
                 5000,
             );
             return;
         }
+
+        let shouldPrune = false;
+        if (
+            Array.isArray(previewData.removed_paths) &&
+            previewData.removed_paths.length > 0
+        ) {
+            const listing = previewData.removed_paths
+                .map((p) => `  - ${p}`)
+                .join("\n");
+            shouldPrune = confirm(
+                `${t("config.prune_confirm")}\n\n${listing}\n\n${t("config.prune_confirm_action")}`,
+            );
+        }
+
+        const finalUrl = shouldPrune
+            ? "/api/config/sync-template?prune=true"
+            : "/api/config/sync-template";
+        const finalRes = await api(finalUrl, { method: "POST" });
+        const finalData = await finalRes.json();
+        if (!finalData.success) {
+            showSaveStatus("error", t("config.save_error"));
+            showToast(
+                `${t("common.error")}: ${finalData.error || t("config.sync_error")}`,
+                "error",
+                5000,
+            );
+            return;
+        }
+
         await loadConfig();
         showSaveStatus("saved", t("config.saved"));
-        if (data.warning) {
+        if (finalData.warning) {
             showToast(
-                `${t("common.warning")}: ${data.warning}`,
+                `${t("common.warning")}: ${finalData.warning}`,
                 "warning",
                 5000,
             );
         }
-        const suffix = Number.isFinite(data.added_count)
-            ? ` (+${data.added_count})`
-            : "";
-        showToast(`${t("config.sync_success")}${suffix}`, "info", 4000);
+        if (shouldPrune) {
+            const removedCount = Number.isFinite(finalData.removed_count)
+                ? finalData.removed_count
+                : Array.isArray(finalData.removed_paths)
+                  ? finalData.removed_paths.length
+                  : 0;
+            showToast(
+                `${t("config.prune_success")} (-${removedCount})`,
+                "info",
+                4000,
+            );
+        } else {
+            const suffix = Number.isFinite(finalData.added_count)
+                ? ` (+${finalData.added_count})`
+                : "";
+            showToast(`${t("config.sync_success")}${suffix}`, "info", 4000);
+        }
     } catch (e) {
         showSaveStatus("error", t("config.sync_error"));
         showToast(`${t("common.error")}: ${e.message}`, "error", 5000);
diff --git a/src/Undefined/webui/static/js/i18n.js b/src/Undefined/webui/static/js/i18n.js
index 2959eec8..b6027322 100644
--- a/src/Undefined/webui/static/js/i18n.js
+++ b/src/Undefined/webui/static/js/i18n.js
@@ -79,6 +79,9 @@ const I18N = {
         "config.syncing": "同步模板中...",
         "config.sync_success": "模板同步完成",
         "config.sync_error": "模板同步失败",
+        "config.prune_confirm": "以下配置项不存在于模板中，是否删除？",
+        "config.prune_confirm_action": "点击「确定」将永久删除以上配置项。",
+        "config.prune_success": "已清理多余配置项",
         "config.search_placeholder": "搜索配置...",
         "config.clear_search": "清除搜索",
         "config.expand_all": "全部展开",
@@ -307,6 +310,11 @@ const I18N = {
         "config.syncing": "Syncing template...",
         "config.sync_success": "Template sync completed",
         "config.sync_error": "Template sync failed",
+        "config.prune_confirm":
+            "The following config keys do not exist in the template. Delete them?",
+        "config.prune_confirm_action":
+            "Click OK to permanently remove the listed keys.",
+        "config.prune_success": "Removed obsolete config keys",
         "config.search_placeholder": "Search config...",
         "config.clear_search": "Clear search",
         "config.expand_all": "Expand all",
diff --git a/src/Undefined/webui/utils/config_sync.py b/src/Undefined/webui/utils/config_sync.py
index 61a88626..46c16eef 100644
--- a/src/Undefined/webui/utils/config_sync.py
+++ b/src/Undefined/webui/utils/config_sync.py
@@ -12,11 +12,14 @@
 from .config_io import CONFIG_EXAMPLE_PATH, _resolve_config_example_path
 from .toml_render import TomlData, merge_defaults, render_toml
 
+_PASSTHROUGH_KEYS = {"request_params"}
+
 
 @dataclass(frozen=True)
 class ConfigTemplateSyncResult:
     content: str
     added_paths: list[str]
+    removed_paths: list[str]
     comments: CommentMap
 
 
@@ -64,6 +67,73 @@ def _collect_added_paths(
     return added
 
 
+def _collect_removed_paths(
+    defaults: TomlData, current: TomlData, prefix: str = ""
+) -> list[str]:
+    """收集存在于 current 但不在 defaults 中的路径（与 _collect_added_paths 互逆）。"""
+    removed: list[str] = []
+    for key, current_value in current.items():
+        path = f"{prefix}.{key}" if prefix else key
+        if key not in defaults:
+            removed.append(path)
+            continue
+        default_value = defaults[key]
+        if isinstance(current_value, dict) and isinstance(default_value, dict):
+            if _should_skip_passthrough_recursion(key, default_value):
+                continue
+            removed.extend(_collect_removed_paths(default_value, current_value, path))
+        elif _is_array_of_tables(current_value) and _is_array_of_tables(default_value):
+            template_item = default_value[0]
+            for index, current_item in enumerate(current_value):
+                default_item = (
+                    default_value[index]
+                    if index < len(default_value)
+                    else template_item
+                )
+                removed.extend(
+                    _collect_removed_paths(
+                        default_item,
+                        current_item,
+                        f"{path}[{index}]",
+                    )
+                )
+    return removed
+
+
+def _prune_to_template(
+    data: TomlData, template: TomlData, prefix: str = ""
+) -> TomlData:
+    """递归移除 data 中不存在于 template 的键。"""
+    pruned: TomlData = {}
+    for key, value in data.items():
+        path = f"{prefix}.{key}" if prefix else key
+        if key not in template:
+            continue
+        template_value = template[key]
+        if isinstance(value, dict) and isinstance(template_value, dict):
+            if _should_skip_passthrough_recursion(key, template_value):
+                pruned[key] = value
+            else:
+                pruned[key] = _prune_to_template(value, template_value, path)
+        elif _is_array_of_tables(value) and _is_array_of_tables(template_value):
+            tpl_item = template_value[0]
+            pruned[key] = [
+                _prune_to_template(
+                    item,
+                    template_value[idx] if idx < len(template_value) else tpl_item,
+                    f"{path}[{idx}]",
+                )
+                for idx, item in enumerate(value)
+            ]
+        else:
+            pruned[key] = value
+    return pruned
+
+
+def _should_skip_passthrough_recursion(key: str, template_value: TomlData) -> bool:
+    return not template_value and key in _PASSTHROUGH_KEYS
+
+
 def _is_array_of_tables(value: Any) -> bool:
     return (
         isinstance(value, list)
@@ -163,12 +233,20 @@ def _merge_comment_maps(current: CommentMap, example: CommentMap) -> CommentMap:
     return merged
 
 
-def sync_config_text(current_text: str, example_text: str) -> ConfigTemplateSyncResult:
+def sync_config_text(
+    current_text: str,
+    example_text: str,
+    *,
+    prune: bool = False,
+) -> ConfigTemplateSyncResult:
     current_data = _parse_toml_text(current_text, label="current config")
     example_data = _parse_toml_text(example_text, label="config example")
     prepared_example_data = _prepare_pool_model_templates(example_data, current_data)
     added_paths = _collect_added_paths(prepared_example_data, current_data)
+    removed_paths = _collect_removed_paths(prepared_example_data, current_data)
     merged = merge_defaults(prepared_example_data, current_data)
+    if prune and removed_paths:
+        merged = _prune_to_template(merged, prepared_example_data)
     example_comments = parse_comment_map_text(example_text)
     comments = _merge_comment_maps(
         parse_comment_map_text(current_text),
@@ -178,6 +256,7 @@ def sync_config_text(current_text: str, example_text: str) -> ConfigTemplateSync
     return ConfigTemplateSyncResult(
         content=content,
         added_paths=added_paths,
+        removed_paths=removed_paths,
         comments=comments,
     )
 
@@ -187,6 +266,7 @@ def sync_config_file(
     example_path: Path = CONFIG_EXAMPLE_PATH,
     *,
     write: bool = True,
+    prune: bool = False,
 ) -> ConfigTemplateSyncResult:
     resolved_example = _resolve_config_example_path(example_path)
     if resolved_example is None or not resolved_example.exists():
@@ -196,7 +276,7 @@ def sync_config_file(
         config_path.read_text(encoding="utf-8") if config_path.exists() else ""
     )
     example_text = resolved_example.read_text(encoding="utf-8")
-    result = sync_config_text(current_text, example_text)
+    result = sync_config_text(current_text, example_text, prune=prune)
     if write:
         config_path.write_text(result.content, encoding="utf-8")
     return result
diff --git a/tests/test_config_template_sync.py b/tests/test_config_template_sync.py
index e57c2189..c26e107b 100644
--- a/tests/test_config_template_sync.py
+++ b/tests/test_config_template_sync.py
@@ -115,3 +115,74 @@ def test_sync_config_text_merges_new_fields_into_existing_pool_model_entries() -
     assert model["request_params"]["temperature"] == 0.2
     assert "models.chat.pool.models[0].api_mode" in result.added_paths
     assert "models.chat.pool.models[0].request_params" in result.added_paths
+
+
+def test_sync_config_text_prune_preserves_passthrough_request_params() -> None:
+    current = """
+[models.chat]
+api_url = "https://primary.example/v1"
+api_key = "primary-key"
+model_name = "primary-model"
+
+[models.chat.request_params]
+temperature = 0.2
+
+[models.chat.request_params.metadata]
+tier = "gold"
+
+[[models.chat.request_params.tags]]
+name = "alpha"
+
+[models.chat.extra]
+flag = true
+"""
+    example = """
+[models.chat]
+api_url = ""
+api_key = ""
+model_name = ""
+
+[models.chat.request_params]
+"""
+
+    result = sync_config_text(current, example, prune=True)
+    parsed = tomllib.loads(result.content)
+    request_params = parsed["models"]["chat"]["request_params"]
+
+    assert result.removed_paths == ["models.chat.extra"]
+    assert request_params["temperature"] == 0.2
+    assert request_params["metadata"]["tier"] == "gold"
+    assert request_params["tags"][0]["name"] == "alpha"
+    assert "extra" not in parsed["models"]["chat"]
+
+
+def test_sync_config_text_prune_only_preserves_exact_passthrough_keys() -> None:
+    current = """
+[models.chat]
+api_url = "https://primary.example/v1"
+api_key = "primary-key"
+model_name = "primary-model"
+
+[models.chat.request_params]
+temperature = 0.2
+
+[models.chat.custom_request_params]
+temperature = 0.8
+"""
+    example = """
+[models.chat]
+api_url = ""
+api_key = ""
+model_name = ""
+
+[models.chat.request_params]
+
+[models.chat.custom_request_params]
+"""
+
+    result = sync_config_text(current, example, prune=True)
+    parsed = tomllib.loads(result.content)
+
+    assert result.removed_paths == ["models.chat.custom_request_params.temperature"]
+    assert parsed["models"]["chat"]["request_params"]["temperature"] == 0.2
+    assert parsed["models"]["chat"]["custom_request_params"] == {}
diff --git a/tests/test_llm_request_params.py b/tests/test_llm_request_params.py
index 3c2cb7b5..2d8e68ec 100644
--- a/tests/test_llm_request_params.py
+++ b/tests/test_llm_request_params.py
@@ -193,7 +193,6 @@ async def test_responses_request_normalizes_tool_calls_and_usage() -> None:
             }
         ],
         tool_choice=cast(Any, {"type": "function", "function": {"name": "lookup"}}),
-        thinking={"enabled": False, "budget_tokens": 0},
     )
 
     assert fake_client.responses.last_kwargs is not None
@@ -235,6 +234,61 @@ async def test_responses_request_normalizes_tool_calls_and_usage() -> None:
         "tool_result_start_index": 2,
     }
 
+    await requester._http_client.aclose()
+
+
+@pytest.mark.asyncio
+async def test_responses_request_respects_explicit_thinking_override() -> None:
+    requester = ModelRequester(
+        http_client=httpx.AsyncClient(),
+        token_usage_storage=cast(TokenUsageStorage, _FakeUsageStorage()),
+    )
+    fake_client = _FakeClient(
+        responses=[
+            {
+                "id": "resp_1",
+                "output": [
+                    {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": [{"type": "output_text", "text": "hi"}],
+                    },
+                ],
+                "usage": {"input_tokens": 5, "output_tokens": 3, "total_tokens": 8},
+            }
+        ]
+    )
+    setattr(
+        requester,
+        "_get_openai_client_for_model",
+        lambda _cfg: cast(AsyncOpenAI, fake_client),
+    )
+    cfg = ChatModelConfig(
+        api_url="https://api.openai.com/v1",
+        api_key="sk-test",
+        model_name="gpt-test",
+        max_tokens=512,
+        api_mode="responses",
+        reasoning_enabled=True,
+        reasoning_effort="low",
+    )
+
+    await requester.request(
+        model_config=cfg,
+        messages=[{"role": "user", "content": "hello"}],
+        max_tokens=128,
+        call_type="chat",
+        thinking={"enabled": False, "budget_tokens": 0},
+    )
+
+    assert fake_client.responses.last_kwargs is not None
+    assert fake_client.responses.last_kwargs["reasoning"] == {"effort": "low"}
+    assert fake_client.responses.last_kwargs["extra_body"] == {
+        "thinking": {"budget_tokens": 0, "type": "disabled"},
+    }
+
+    await requester._http_client.aclose()
+
 
 def test_normalize_responses_result_falls_back_to_output_text_and_scalar_content() -> (
     None
@@ -928,3 +982,136 @@ async def test_responses_tools_and_tool_choice_use_sanitized_api_names() -> None
     )
 
     await requester._http_client.aclose()
+
+
+@pytest.mark.asyncio
+async def test_thinking_effort_anthropic_style_chat_completions() -> None:
+    """thinking_enabled + anthropic style → legacy thinking + output_config.effort."""
+    requester = ModelRequester(
+        http_client=httpx.AsyncClient(),
+        token_usage_storage=cast(TokenUsageStorage, _FakeUsageStorage()),
+    )
+    fake_client = _FakeClient()
+    setattr(
+        requester,
+        "_get_openai_client_for_model",
+        lambda _cfg: cast(AsyncOpenAI, fake_client),
+    )
+    cfg = ChatModelConfig(
+        api_url="https://api.anthropic.com/v1",
+        api_key="sk-test",
+        model_name="claude-test",
+        max_tokens=4096,
+        thinking_enabled=True,
+        thinking_budget_tokens=8000,
+        reasoning_enabled=True,
+        reasoning_effort="max",
+        reasoning_effort_style="anthropic",
+    )
+
+    await requester.request(
+        model_config=cfg,
+        messages=[{"role": "user", "content": "hello"}],
+        max_tokens=1024,
+        call_type="chat",
+    )
+
+    kw = fake_client.chat.completions.last_kwargs
+    assert kw is not None
+    assert kw["extra_body"]["thinking"] == {"type": "enabled", "budget_tokens": 8000}
+    assert kw["extra_body"]["output_config"] == {"effort": "max"}
+    assert "reasoning" not in kw.get("extra_body", {})
+
+    await requester._http_client.aclose()
+
+
+@pytest.mark.asyncio
+async def test_thinking_effort_openai_style_responses() -> None:
+    """thinking_enabled + openai style → legacy thinking + reasoning.effort."""
+    requester = ModelRequester(
+        http_client=httpx.AsyncClient(),
+        token_usage_storage=cast(TokenUsageStorage, _FakeUsageStorage()),
+    )
+    fake_client = _FakeClient(
+        responses=[
+            {
+                "id": "resp_1",
+                "output": [
+                    {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": [{"type": "output_text", "text": "hi"}],
+                    },
+                ],
+                "usage": {"input_tokens": 5, "output_tokens": 3, "total_tokens": 8},
+            }
+        ]
+    )
+    setattr(
+        requester,
+        "_get_openai_client_for_model",
+        lambda _cfg: cast(AsyncOpenAI, fake_client),
+    )
+    cfg = ChatModelConfig(
+        api_url="https://api.openai.com/v1",
+        api_key="sk-test",
+        model_name="gpt-test",
+        max_tokens=4096,
+        api_mode="responses",
+        thinking_enabled=True,
+        thinking_budget_tokens=8000,
+        reasoning_enabled=True,
+        reasoning_effort="high",
+        reasoning_effort_style="openai",
+    )
+
+    await requester.request(
+        model_config=cfg,
+        messages=[{"role": "user", "content": "hello"}],
+        max_tokens=1024,
+        call_type="chat",
+    )
+
+    kw = fake_client.responses.last_kwargs
+    assert kw is not None
+    assert kw["extra_body"]["thinking"] == {"type": "enabled", "budget_tokens": 8000}
+    assert kw["reasoning"] == {"effort": "high"}
+    assert "output_config" not in kw and "output_config" not in kw.get("extra_body", {})
+
+    await requester._http_client.aclose()
+
+
+@pytest.mark.asyncio
+async def test_thinking_enabled_legacy_budget_tokens() -> None:
+    """thinking_enabled=True + no effort → legacy budget_tokens mode."""
+    requester = ModelRequester(
+        http_client=httpx.AsyncClient(),
+        token_usage_storage=cast(TokenUsageStorage, _FakeUsageStorage()),
+    )
+    fake_client = _FakeClient()
+    setattr(
+        requester,
+        "_get_openai_client_for_model",
+        lambda _cfg: cast(AsyncOpenAI, fake_client),
+    )
+    cfg = ChatModelConfig(
+        api_url="https://api.openai.com/v1",
+        api_key="sk-test",
+        model_name="gpt-test",
+        max_tokens=4096,
+        thinking_enabled=True,
+        thinking_budget_tokens=8000,
+    )
+
+    await requester.request(
+        model_config=cfg,
+        messages=[{"role": "user", "content": "hello"}],
+        max_tokens=1024,
+        call_type="chat",
+    )
+
+    kw = fake_client.chat.completions.last_kwargs
+    assert kw is not None
+    assert kw["extra_body"]["thinking"] == {"type": "enabled", "budget_tokens": 8000}
+
+    await requester._http_client.aclose()
diff --git a/tests/test_sync_config_template_script.py b/tests/test_sync_config_template_script.py
new file mode 100644
index 00000000..3f14aec4
--- /dev/null
+++ b/tests/test_sync_config_template_script.py
@@ -0,0 +1,56 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType, SimpleNamespace
+
+import pytest
+
+
+def _load_script_module() -> ModuleType:
+    script_path = (
+        Path(__file__).resolve().parent.parent / "scripts" / "sync_config_template.py"
+    )
+    spec = importlib.util.spec_from_file_location(
+        "sync_config_template_script", script_path
+    )
+    assert spec is not None
+    assert spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_prune_mode_reports_analysis_before_write(
+    monkeypatch: pytest.MonkeyPatch,
+    capsys: pytest.CaptureFixture[str],
+) -> None:
+    module = _load_script_module()
+    calls: list[tuple[bool, bool]] = []
+
+    def fake_sync_config_file(
+        *,
+        config_path: Path,
+        example_path: Path,
+        write: bool = True,
+        prune: bool = False,
+    ) -> SimpleNamespace:
+        del config_path, example_path
+        calls.append((write, prune))
+        return SimpleNamespace(
+            content="",
+            added_paths=[],
+            removed_paths=["models.chat.extra"],
+            comments={},
+        )
+
+    monkeypatch.setattr(module, "sync_config_file", fake_sync_config_file)
+    monkeypatch.setattr(module, "_confirm_prune", lambda _paths: False)
+    monkeypatch.setattr(sys, "argv", ["sync_config_template.py", "--prune"])
+
+    assert module.main() == 0
+
+    output = capsys.readouterr().out
+    assert "[sync-config] 分析完成:" in output
+    assert calls == [(False, False), (True, False)]
diff --git a/tests/test_webui_management_api.py b/tests/test_webui_management_api.py
index d4caedfa..be44bffe 100644
--- a/tests/test_webui_management_api.py
+++ b/tests/test_webui_management_api.py
@@ -10,7 +10,7 @@
 from Undefined.webui import app as webui_app
 from Undefined.webui.app import create_app
 from Undefined.webui.core import SessionStore
-from Undefined.webui.routes import _auth, _index, _shared, _system
+from Undefined.webui.routes import _auth, _config, _index, _shared, _system
 from Undefined.webui.routes._shared import (
     REDIRECT_TO_CONFIG_ONCE_APP_KEY,
     SESSION_COOKIE,
@@ -156,6 +156,106 @@ async def _fake_runtime() -> tuple[bool, bool, str]:
     assert payload["advice"]
 
 
+async def test_sync_config_template_handler_preview_skips_reload(
+    monkeypatch: Any,
+) -> None:
+    request = _request(query={"write": "false"})
+    calls: list[tuple[str, object, object]] = []
+
+    async def _fake_to_thread(func: Any, *args: Any, **kwargs: Any) -> Any:
+        return func(*args, **kwargs)
+
+    def _fake_validate_required_config() -> tuple[bool, str]:
+        calls.append(("validate", None, None))
+        return True, "OK"
+
+    def _fake_sync_config_file(*, write: bool = True, prune: bool = False) -> Any:
+        calls.append(("sync", write, prune))
+        return SimpleNamespace(
+            added_paths=["models.chat.api_mode"],
+            removed_paths=["models.chat.extra"],
+        )
+
+    monkeypatch.setattr(_config, "check_auth", lambda _request: True)
+    monkeypatch.setattr(
+        cast(Any, getattr(_config, "asyncio")), "to_thread", _fake_to_thread
+    )
+    monkeypatch.setattr(_config, "sync_config_file", _fake_sync_config_file)
+    monkeypatch.setattr(
+        _config,
+        "get_config_manager",
+        lambda: SimpleNamespace(
+            reload=lambda: calls.append(("reload", None, None)),
+        ),
+    )
+    monkeypatch.setattr(
+        _config, "validate_required_config", _fake_validate_required_config
+    )
+
+    response = await _config.sync_config_template_handler(
+        cast(web.Request, cast(Any, request))
+    )
+    payload = _json_payload(response)
+
+    assert payload["success"] is True
+    assert payload["preview"] is True
+    assert payload["warning"] is None
+    assert payload["added_count"] == 1
+    assert payload["removed_count"] == 1
+    assert calls == [("sync", False, False)]
+
+
+async def test_sync_config_template_handler_write_reloads_and_validates(
+    monkeypatch: Any,
+) -> None:
+    request = _request(query={"prune": "true"})
+    calls: list[tuple[str, object, object]] = []
+
+    async def _fake_to_thread(func: Any, *args: Any, **kwargs: Any) -> Any:
+        return func(*args, **kwargs)
+
+    def _fake_validate_required_config() -> tuple[bool, str]:
+        calls.append(("validate", None, None))
+        return False, "missing required field"
+
+    def _fake_sync_config_file(*, write: bool = True, prune: bool = False) -> Any:
+        calls.append(("sync", write, prune))
+        return SimpleNamespace(
+            added_paths=[],
+            removed_paths=["models.chat.extra"],
+        )
+
+    monkeypatch.setattr(_config, "check_auth", lambda _request: True)
+    monkeypatch.setattr(
+        cast(Any, getattr(_config, "asyncio")), "to_thread", _fake_to_thread
+    )
+    monkeypatch.setattr(_config, "sync_config_file", _fake_sync_config_file)
+    monkeypatch.setattr(
+        _config,
+        "get_config_manager",
+        lambda: SimpleNamespace(
+            reload=lambda: calls.append(("reload", None, None)),
+        ),
+    )
+    monkeypatch.setattr(
+        _config, "validate_required_config", _fake_validate_required_config
+    )
+
+    response = await _config.sync_config_template_handler(
+        cast(web.Request, cast(Any, request))
+    )
+    payload = _json_payload(response)
+
+    assert payload["success"] is True
+    assert payload["preview"] is False
+    assert payload["warning"] == "missing required field"
+    assert calls == [
+        ("sync", True, True),
+        ("reload", None, None),
+        ("validate", None, None),
+    ]
+
+
 def test_create_app_registers_management_routes() -> None:
     app = create_app()
     routes = {