feat: 配置同步增强 + AI 推理参数重构

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

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="")

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

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",