refactor: 增强跨格式转换系统，支持 thinking/文件/音频/响应格式等完整转换

- 新增 ThinkingConfig/ResponseFormatConfig/FileBlock/AudioBlock 内部表示
- 实现 OpenAI reasoning_effort <-> Claude thinking <-> Gemini thinkingConfig 互转
- 支持 OpenAI web_search_options -> Claude web_search tool 转换
- 新增异步 convert_request_async，在转换阶段解析图片 URL 为 base64
- 将 GlobalModel.output_limit 传播至跨格式转换用于 max_tokens 默认值
- 前端模型表单增加最大输出 Token 和上下文窗口配置
- Claude normalizer 保留 cache_control 透传（system 数组和 content block）
- Gemini normalizer 支持内置工具（googleSearch/codeExecution/urlContext）
- OpenAI normalizer 补全采样参数、response_format、usage details 转换
- 各 normalizer 工具方法提取至基类消除重复代码
- 简化 Kiro model mapping 为直接透传
- 预置模型列表添加 claude-sonnet-4.6

src/services/provider/adapters/kiro/converter.py

@@ -20,41 +20,9 @@ from src.services.provider.adapters.kiro.constants import (
 
 
 def map_model(model: str) -> str | None:
-    """Map an Anthropic model name to a Kiro-compatible model ID.
-
-    Kiro upstream expects specific model IDs (e.g. ``claude-sonnet-4.5``).
-    If *model* already matches a Kiro ID it is returned as-is; otherwise we
-    attempt a best-effort fuzzy mapping.  Returns ``None`` when the model is
-    unrecognised.
-    """
+    """Pass through the model name as-is to Kiro upstream."""
     raw = str(model or "").strip()
-    if not raw:
-        return None
-
-    model_lower = raw.lower()
-
-    # Already a Kiro-native ID?
-    _KIRO_NATIVE_IDS = {
-        "claude-sonnet-4.5",
-        "claude-opus-4.5",
-        "claude-opus-4.6",
-        "claude-haiku-4.5",
-    }
-    if model_lower in _KIRO_NATIVE_IDS:
-        return model_lower
-
-    # Fuzzy mapping from Anthropic-style model names
-    if "sonnet" in model_lower:
-        return "claude-sonnet-4.5"
-    if "opus" in model_lower:
-        if "4-5" in model_lower or "4.5" in model_lower:
-            return "claude-opus-4.5"
-        return "claude-opus-4.6"
-    if "haiku" in model_lower:
-        return "claude-haiku-4.5"
-
-    # Unrecognised — pass through as-is (let the upstream decide)
-    return raw
+    return raw or None
 
 
 def _extract_session_id(user_id: str) -> str | None:

src/services/provider/preset_models.py

@@ -35,6 +35,12 @@ PRESET_MODELS: dict[str, list[dict[str, Any]]] = {
             "owned_by": "anthropic",
             "display_name": "Claude Sonnet 4.5",
         },
+        {
+            "id": "claude-sonnet-4.6",
+            "object": "model",
+            "owned_by": "anthropic",
+            "display_name": "Claude Sonnet 4.6",
+        },
         {
             "id": "claude-opus-4.5",
             "object": "model",

src/services/scheduling/candidate_builder.py

@@ -386,6 +386,17 @@ class CandidateBuilder:
         client_format_str = normalize_endpoint_signature(client_format)
         client_sig = parse_signature_key(client_format_str)
         client_family, client_kind = client_sig.api_family, client_sig.endpoint_kind
+
+        # 提取 GlobalModel 配置的 output_limit（用于跨格式转换时的 max_tokens 默认值）
+        output_limit: int | None = None
+        normalized_name = model_name.strip() if isinstance(model_name, str) else ""
+        if normalized_name:
+            gm = await ModelCacheService.get_global_model_by_name(db, normalized_name)
+            if gm and isinstance(gm.config, dict):
+                raw = gm.config.get("output_limit")
+                if isinstance(raw, int) and raw > 0:
+                    output_limit = raw
+
         # chat/cli 互相可回退（用于同协议族下的端点变体），video/image 等不跨类回退
         if client_kind in {EndpointKind.CHAT, EndpointKind.CLI}:
             allowed_kinds = {EndpointKind.CHAT, EndpointKind.CLI}
@@ -581,6 +592,7 @@ class CandidateBuilder:
                         mapping_matched_model=mapping_matched_model,
                         needs_conversion=needs_conversion,
                         provider_api_format=str(endpoint_format_str or ""),
+                        output_limit=output_limit,
                     )
 
                     if needs_conversion:

src/services/scheduling/schemas.py

@@ -28,6 +28,7 @@ class ProviderCandidate:
     mapping_matched_model: str | None = None  # 通过映射匹配到的模型名（用于实际请求）
     needs_conversion: bool = False  # 是否需要格式转换
     provider_api_format: str = ""  # Provider 端点实际格式（用于健康度/熔断 bucket）
+    output_limit: int | None = None  # GlobalModel 配置的模型输出上限
 
     def _stable_order_key(self) -> tuple[int, int, str, str, str]:
         """