feat(conversion): 支持 Claude output_config.effort 跨格式转换，新增 xhigh 档位

- 新增 Claude output_config.effort 与标准化 reasoning_effort 的双向映射
- 新增 xhigh 档位(budget_tokens=8192)，对应 Claude effort=max
- reasoning_effort 独立于 thinking 存入 extra，支持无 thinking 场景的跨格式传递
- OpenAI/Responses API 输出时 xhigh 自动降级为 high

src/core/api_format/conversion/field_mappings.py

@@ -130,15 +130,34 @@ REASONING_EFFORT_TO_THINKING_BUDGET: dict[str, int] = {
     "low": 1280,
     "medium": 2048,
     "high": 4096,
+    "xhigh": 8192,
 }
 
 # thinking budget_tokens -> OpenAI reasoning_effort（反向映射，取最近区间）
 THINKING_BUDGET_TO_REASONING_EFFORT: list[tuple[int, str]] = [
     (1664, "low"),  # <= 1664 -> low  (midpoint of 1280..2048)
     (3072, "medium"),  # <= 3072 -> medium  (midpoint of 2048..4096)
-    (2**31, "high"),  # > 3072 -> high
+    (6144, "high"),  # <= 6144 -> high  (midpoint of 4096..8192)
+    (2**31, "xhigh"),  # > 6144 -> xhigh
 ]
 
+# Claude output_config.effort -> 标准化 reasoning_effort
+CLAUDE_EFFORT_TO_REASONING_EFFORT: dict[str, str] = {
+    "low": "low",
+    "medium": "medium",
+    "high": "high",
+    "max": "xhigh",
+    # "auto" 不映射，让目标格式使用默认行为
+}
+
+# 标准化 reasoning_effort -> Claude output_config.effort
+REASONING_EFFORT_TO_CLAUDE_EFFORT: dict[str, str] = {
+    "low": "low",
+    "medium": "medium",
+    "high": "high",
+    "xhigh": "max",
+}
+
 
 # OpenAI web_search_options.search_context_size -> Claude web_search max_uses
 WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES: dict[str, int] = {
@@ -190,6 +209,8 @@ __all__ = [
     "RETRYABLE_ERROR_TYPES",
     "REASONING_EFFORT_TO_THINKING_BUDGET",
     "THINKING_BUDGET_TO_REASONING_EFFORT",
+    "CLAUDE_EFFORT_TO_REASONING_EFFORT",
+    "REASONING_EFFORT_TO_CLAUDE_EFFORT",
     "WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES",
     "CLAUDE_DEFAULT_MAX_TOKENS",
     "THINKING_BUDGET_TOKENS_PERCENTAGE",

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

@@ -11,7 +11,9 @@ import json
 from typing import Any
 
 from src.core.api_format.conversion.field_mappings import (
+    CLAUDE_EFFORT_TO_REASONING_EFFORT,
     ERROR_TYPE_MAPPINGS,
+    REASONING_EFFORT_TO_CLAUDE_EFFORT,
     RETRYABLE_ERROR_TYPES,
     STOP_REASON_MAPPINGS,
     THINKING_BUDGET_TOKENS_MIN,
@@ -187,11 +189,19 @@ class ClaudeNormalizer(FormatNormalizer):
                         "tools",
                         "tool_choice",
                         "thinking",
+                        "output_config",
                     },
                 )
             },
         )
 
+        # output_config.effort -> reasoning_effort (独立于 thinking)
+        output_config = request.get("output_config")
+        if isinstance(output_config, dict):
+            effort = output_config.get("effort")
+            if isinstance(effort, str) and effort in CLAUDE_EFFORT_TO_REASONING_EFFORT:
+                internal.extra["reasoning_effort"] = CLAUDE_EFFORT_TO_REASONING_EFFORT[effort]
+
         if dropped:
             internal.extra.setdefault("raw", {})["dropped_blocks"] = dropped
 
@@ -314,6 +324,11 @@ class ClaudeNormalizer(FormatNormalizer):
                     result["max_tokens"] = bt + 1
                 result["thinking"] = thinking_out
 
+        # reasoning_effort -> output_config.effort (独立于 thinking)
+        effort = internal.extra.get("reasoning_effort") if internal.extra else None
+        if isinstance(effort, str) and effort in REASONING_EFFORT_TO_CLAUDE_EFFORT:
+            result["output_config"] = {"effort": REASONING_EFFORT_TO_CLAUDE_EFFORT[effort]}
+
         # web_search_options -> Claude web_search tool
         web_search_opts = internal.extra.get("web_search_options") if internal.extra else None
         if isinstance(web_search_opts, dict):

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

@@ -216,6 +216,7 @@ class OpenAINormalizer(FormatNormalizer):
                 budget_tokens=REASONING_EFFORT_TO_THINKING_BUDGET[reasoning_effort],
                 extra={"reasoning_effort": reasoning_effort},
             )
+            extra["reasoning_effort"] = reasoning_effort
 
         # web_search_options 存入 extra 供目标 normalizer 使用
         web_search_options = request.get("web_search_options")
@@ -344,6 +345,7 @@ class OpenAINormalizer(FormatNormalizer):
             result["tool_choice"] = self._tool_choice_to_openai(internal.tool_choice)
 
         # thinking -> reasoning_effort
+        effort: str | None = None
         if internal.thinking and internal.thinking.enabled:
             effort = internal.thinking.extra.get("reasoning_effort")
             if not effort and internal.thinking.budget_tokens is not None:
@@ -351,8 +353,14 @@ class OpenAINormalizer(FormatNormalizer):
                     if internal.thinking.budget_tokens <= threshold:
                         effort = level
                         break
-            if effort:
+        # 兜底: 从 internal.extra 读取 (支持 output_config.effort 独立于 thinking 的场景)
-                result["reasoning_effort"] = effort
+        if not effort and internal.extra:
+            effort = internal.extra.get("reasoning_effort")
+        if effort:
+            # OpenAI Chat Completions 仅支持 low/medium/high，xhigh 降级为 high
+            if effort == "xhigh":
+                effort = "high"
+            result["reasoning_effort"] = effort
 
         # parallel_tool_calls
         if internal.parallel_tool_calls is not None:

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

@@ -196,6 +196,10 @@ class OpenAICliNormalizer(FormatNormalizer):
             },
         )
 
+        # reasoning_effort 同步存入 extra (支持独立于 thinking 的跨格式转换)
+        if thinking and thinking.extra.get("reasoning_effort"):
+            internal.extra["reasoning_effort"] = thinking.extra["reasoning_effort"]
+
         return internal
 
     def request_from_internal(
@@ -259,11 +263,13 @@ class OpenAICliNormalizer(FormatNormalizer):
             result["tool_choice"] = self._tool_choice_to_openai(internal.tool_choice)
 
         # thinking -> reasoning (Responses API)
+        effort: str | None = None
         if internal.thinking and internal.thinking.enabled:
             # 优先还原原始 reasoning 对象
             original_reasoning = internal.thinking.extra.get("reasoning")
             if isinstance(original_reasoning, dict):
                 result["reasoning"] = original_reasoning
+                effort = None  # 已还原，不需要再构造
             else:
                 effort = internal.thinking.extra.get("reasoning_effort")
                 if not effort and internal.thinking.budget_tokens is not None:
@@ -271,8 +277,14 @@ class OpenAICliNormalizer(FormatNormalizer):
                         if internal.thinking.budget_tokens <= threshold:
                             effort = level
                             break
-                if effort:
+        # 兜底: 从 internal.extra 读取
-                    result["reasoning"] = {"effort": effort}
+        if not effort and internal.extra and "reasoning" not in result:
+            effort = internal.extra.get("reasoning_effort")
+        if effort:
+            # xhigh 降级为 high (Responses API 也仅支持 low/medium/high)
+            if effort == "xhigh":
+                effort = "high"
+            result["reasoning"] = {"effort": effort}
 
         # parallel_tool_calls
         if internal.parallel_tool_calls is not None: