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/api/handlers/base/chat_handler_base.py

@@ -706,11 +706,12 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         # 跨格式：先做请求体转换（失败触发 failover）
         registry = get_format_converter_registry()
         if needs_conversion:
-            request_body = registry.convert_request(
+            request_body = await registry.convert_request_async(
                 request_body,
                 str(client_api_format),
                 str(provider_api_format),
                 target_variant=cross_format_variant,
+                output_limit=candidate.output_limit if candidate else None,
             )
             # 格式转换后，为需要 model 字段的格式设置模型名
             self._set_model_after_conversion(

src/api/handlers/base/cli_protocol.py

@@ -155,7 +155,7 @@ class CliHandlerProtocol(Protocol):
         mapped_model: str | None,
     ) -> str | None: ...
 
-    def _convert_request_for_cross_format(
+    async def _convert_request_for_cross_format(
         self,
         request_body: dict[str, Any],
         client_api_format: str,
@@ -165,6 +165,7 @@ class CliHandlerProtocol(Protocol):
         is_stream: bool,
         *,
         target_variant: str | None = ...,
+        output_limit: int | None = ...,
     ) -> tuple[dict[str, Any], str]: ...
 
     def _extract_response_metadata(

src/api/handlers/base/cli_request_mixin.py

@@ -228,7 +228,7 @@ class CliRequestMixin:
             stream_options["include_usage"] = True
             request_body["stream_options"] = stream_options
 
-    def _convert_request_for_cross_format(
+    async def _convert_request_for_cross_format(
         self,
         request_body: dict[str, Any],
         client_api_format: str,
@@ -238,6 +238,7 @@ class CliRequestMixin:
         is_stream: bool,
         *,
         target_variant: str | None = None,
+        output_limit: int | None = None,
     ) -> tuple[dict[str, Any], str]:
         """
         跨格式请求转换的公共逻辑
@@ -252,16 +253,18 @@ class CliRequestMixin:
             fallback_model: 备用模型名（通常是原始请求的 model）
             is_stream: 是否流式请求
             target_variant: 目标变体（如 "codex"），用于同格式但有细微差异的上游
+            output_limit: GlobalModel 配置的模型输出上限
 
         Returns:
             (转换后的请求体, 用于 URL 的模型名)
         """
         registry = get_format_converter_registry()
-        converted_body = registry.convert_request(
+        converted_body = await registry.convert_request_async(
             request_body,
             str(client_api_format),
             str(provider_api_format),
             target_variant=target_variant,
+            output_limit=output_limit,
         )
 
         # 先计算 URL 模型（在清理 body 中的 model 字段之前）

src/api/handlers/base/cli_stream_mixin.py

@@ -317,7 +317,7 @@ class CliStreamMixin:
 
         # 跨格式：先做请求体转换（失败触发 failover）
         if needs_conversion and provider_api_format:
-            request_body, url_model = self._convert_request_for_cross_format(
+            request_body, url_model = await self._convert_request_for_cross_format(
                 request_body,
                 client_api_format,
                 provider_api_format,
@@ -325,6 +325,7 @@ class CliStreamMixin:
                 ctx.model,
                 is_stream=upstream_is_stream,
                 target_variant=conversion_variant,
+                output_limit=candidate.output_limit if candidate else None,
             )
         else:
             # 同格式：按原逻辑做轻量清理（子类可覆盖）

src/api/handlers/base/cli_sync_mixin.py

@@ -150,7 +150,7 @@ class CliSyncMixin:
 
             # 跨格式：先做请求体转换（失败触发 failover）
             if needs_conversion and provider_api_format:
-                request_body, url_model = self._convert_request_for_cross_format(
+                request_body, url_model = await self._convert_request_for_cross_format(
                     request_body,
                     client_api_format,
                     provider_api_format,
@@ -158,6 +158,7 @@ class CliSyncMixin:
                     model,
                     is_stream=upstream_is_stream,
                     target_variant=conversion_variant,
+                    output_limit=candidate.output_limit if candidate else None,
                 )
             else:
                 # 同格式：按原逻辑做轻量清理（子类可覆盖）

src/core/api_format/conversion/field_mappings.py

@@ -30,6 +30,8 @@ STOP_REASON_MAPPINGS: dict[str, dict[str, str]] = {
         "max_tokens": "max_tokens",
         "stop_sequence": "stop_sequence",
         "tool_use": "tool_use",
+        "pause_turn": "end_turn",
+        "refusal": "end_turn",
         # Claude 通常以错误/阻断体现，这里仅兜底
         "content_filtered": "end_turn",
         "unknown": "end_turn",
@@ -40,6 +42,8 @@ STOP_REASON_MAPPINGS: dict[str, dict[str, str]] = {
         "stop_sequence": "stop",
         "tool_use": "tool_calls",
         "content_filtered": "content_filter",
+        "refusal": "content_filter",
+        "pause_turn": "stop",
         "unknown": "stop",
     },
     "GEMINI": {
@@ -49,6 +53,8 @@ STOP_REASON_MAPPINGS: dict[str, dict[str, str]] = {
         # Gemini finishReason 对工具调用并没有稳定等价枚举，这里保守兜底为 STOP
         "tool_use": "STOP",
         "content_filtered": "SAFETY",
+        "refusal": "SAFETY",
+        "pause_turn": "OTHER",
         "unknown": "OTHER",
     },
 }
@@ -118,10 +124,77 @@ RETRYABLE_ERROR_TYPES: set[str] = {
 }
 
 
+# OpenAI reasoning_effort -> thinking budget_tokens
+# 参考 new-api relay-claude.go:178-196
+REASONING_EFFORT_TO_THINKING_BUDGET: dict[str, int] = {
+    "low": 1280,
+    "medium": 2048,
+    "high": 4096,
+}
+
+# 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
+]
+
+
+# OpenAI web_search_options.search_context_size -> Claude web_search max_uses
+WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES: dict[str, int] = {
+    "low": 1,
+    "medium": 5,
+    "high": 10,
+}
+
+
+# Claude max_tokens 兜底默认值（仅在 GlobalModel.output_limit 和请求 max_tokens 均为空时使用）
+# 参考 new-api setting/model_setting/claude.go 的 DefaultMaxTokens["default"]
+CLAUDE_DEFAULT_MAX_TOKENS: int = 8192
+
+
+def get_claude_default_max_tokens(_model: str) -> int:
+    """获取 Claude 的 max_tokens 兜底默认值。
+
+    正常情况下应优先使用 GlobalModel.config.output_limit（通过 InternalRequest.output_limit 传入），
+    此函数仅在 output_limit 不可用时作为最终兜底。
+    """
+    return CLAUDE_DEFAULT_MAX_TOKENS
+
+
+# thinking budget_tokens 占 max_tokens 的比例（参考 new-api: 0.8）
+THINKING_BUDGET_TOKENS_PERCENTAGE: float = 0.8
+
+# thinking budget_tokens 最小值（Claude API 要求 >= 1024）
+THINKING_BUDGET_TOKENS_MIN: int = 1280
+
+
+def get_claude_default_thinking_budget(model: str) -> int:
+    """根据模型名称计算 thinking budget_tokens 默认值。
+
+    budget = max(max_tokens * THINKING_BUDGET_TOKENS_PERCENTAGE, THINKING_BUDGET_TOKENS_MIN)
+    """
+    max_tokens = get_claude_default_max_tokens(model)
+    return max(int(max_tokens * THINKING_BUDGET_TOKENS_PERCENTAGE), THINKING_BUDGET_TOKENS_MIN)
+
+
+# 跨格式 thinking 转换时，非 Claude 模型的默认 budget_tokens 安全值
+CROSS_FORMAT_THINKING_BUDGET_DEFAULT: int = 8192
+
+
 __all__ = [
     "ROLE_MAPPINGS",
     "STOP_REASON_MAPPINGS",
     "USAGE_FIELD_MAPPINGS",
     "ERROR_TYPE_MAPPINGS",
     "RETRYABLE_ERROR_TYPES",
+    "REASONING_EFFORT_TO_THINKING_BUDGET",
+    "THINKING_BUDGET_TO_REASONING_EFFORT",
+    "WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES",
+    "CLAUDE_DEFAULT_MAX_TOKENS",
+    "THINKING_BUDGET_TOKENS_PERCENTAGE",
+    "THINKING_BUDGET_TOKENS_MIN",
+    "CROSS_FORMAT_THINKING_BUDGET_DEFAULT",
+    "get_claude_default_max_tokens",
+    "get_claude_default_thinking_budget",
 ]

src/core/api_format/conversion/image_resolver.py

@@ -0,0 +1,273 @@
+"""
+图片 URL 解析器
+
+当跨格式转换时，目标格式需要 base64 图片数据（如 Claude 不原生支持 URL 图片引用），
+该模块负责自动下载图片 URL 并转换为 base64 内嵌数据。
+
+使用方式：在跨格式转换前/后调用 resolve_image_urls()。
+"""
+
+import asyncio
+import base64
+import ipaddress
+import mimetypes
+import socket
+from urllib.parse import urljoin, urlparse
+
+import httpx
+
+from src.core.api_format.conversion.internal import (
+    FileBlock,
+    ImageBlock,
+    InternalRequest,
+)
+from src.core.logger import logger
+
+# 需要 base64 图片数据的目标格式前缀
+_FORMATS_REQUIRING_BASE64 = frozenset({"CLAUDE"})
+
+# 图片下载超时（秒）
+_DOWNLOAD_TIMEOUT = 15.0
+
+# 单张图片最大大小（字节，20MB）
+_MAX_IMAGE_SIZE = 20 * 1024 * 1024
+
+# 并发下载数量限制
+_MAX_CONCURRENT_DOWNLOADS = 8
+
+# 最大重定向次数
+_MAX_REDIRECTS = 5
+
+
+def _is_private_ip(addr: str) -> bool:
+    """检查 IP 地址是否为私有/内网地址。"""
+    try:
+        ip = ipaddress.ip_address(addr)
+        return bool(ip.is_private or ip.is_loopback or ip.is_link_local or ip.is_reserved)
+    except ValueError:
+        return True
+
+
+async def _resolve_and_validate_host(hostname: str) -> list[str] | None:
+    """DNS 解析并校验所有 IP 均为公网地址（SSRF 防护）。
+
+    返回已校验的公网 IP 列表；如果任一 IP 为私有地址或解析失败，返回 None。
+    """
+    try:
+        loop = asyncio.get_running_loop()
+        infos = await loop.getaddrinfo(hostname, None, socket.AF_UNSPEC, socket.SOCK_STREAM)
+        ips: list[str] = []
+        for _family, _type, _proto, _canonname, sockaddr in infos:
+            addr = sockaddr[0]
+            if _is_private_ip(addr):
+                return None
+            ips.append(addr)
+        return ips or None
+    except (socket.gaierror, ValueError, OSError):
+        # DNS 解析失败：安全默认拒绝
+        return None
+
+
+async def _validate_url(url: str) -> bool:
+    """校验 URL 的 scheme 和主机地址，通过返回 True，否则返回 False。"""
+    parsed = urlparse(url)
+    if parsed.scheme not in ("http", "https"):
+        logger.warning("[ImageResolver] 不支持的 URL scheme: {}", url[:100])
+        return False
+    hostname = parsed.hostname or ""
+    resolved = await _resolve_and_validate_host(hostname)
+    if resolved is None:
+        logger.warning("[ImageResolver] 拒绝下载私有网络地址: {}", url[:100])
+        return False
+    return True
+
+
+def _validate_peer_ip(resp: httpx.Response) -> bool:
+    """校验 HTTP 响应的实际对端 IP 是否为公网地址（防 DNS rebinding TOCTOU 绕过）。
+
+    httpx 通过 extensions["network_stream"] 暴露底层连接，
+    从中可获取对端地址进行二次校验。
+    """
+    try:
+        network_stream = resp.extensions.get("network_stream")
+        if network_stream is None:
+            logger.debug("[ImageResolver] network_stream 不可用, DNS rebinding 检测跳过")
+            return True
+        # asyncio transport 标准 extra info key 是 peername
+        peername = network_stream.get_extra_info("peername")
+        if peername is not None:
+            peer_ip = peername[0] if isinstance(peername, tuple) else str(peername)
+            if _is_private_ip(peer_ip):
+                logger.warning("[ImageResolver] DNS rebinding 检测: 实际连接到私有 IP {}", peer_ip)
+                return False
+    except Exception as e:
+        # 无法获取对端信息时放行（不阻塞正常功能），依赖前置 DNS 校验
+        logger.debug("[ImageResolver] 无法获取对端 IP 信息（DNS rebinding 检测跳过）: {}", e)
+    return True
+
+
+async def _download_file(
+    client: httpx.AsyncClient, url: str, semaphore: asyncio.Semaphore
+) -> tuple[str, str] | None:
+    """下载 URL 并返回 (base64_data, media_type)，失败返回 None。
+
+    手动处理重定向，每一跳都检查目标地址（防止重定向到内网的 SSRF 绕过）。
+    连接建立后二次校验对端 IP（防 DNS rebinding）。
+    """
+    async with semaphore:
+        try:
+            if not await _validate_url(url):
+                return None
+
+            current_url = url
+            for _ in range(_MAX_REDIRECTS + 1):
+                async with client.stream("GET", current_url) as resp:
+                    # DNS rebinding 防护：校验实际连接的对端 IP
+                    if not _validate_peer_ip(resp):
+                        return None
+
+                    if resp.is_redirect:
+                        location = resp.headers.get("location", "")
+                        if not location:
+                            logger.warning("[ImageResolver] 重定向缺少 Location: {}", url[:100])
+                            return None
+                        redirect_url = urljoin(str(current_url), location)
+                        if not await _validate_url(redirect_url):
+                            return None
+                        current_url = redirect_url
+                        continue
+
+                    resp.raise_for_status()
+
+                    content_type = resp.headers.get("content-type", "")
+                    media_type = content_type.split(";")[0].strip().lower()
+                    if not media_type:
+                        media_type = _guess_media_type(current_url)
+
+                    # 检查 MIME 类型是否为目标 API 可接受的类型
+                    if not any(media_type.startswith(p) for p in _ACCEPTED_MIME_PREFIXES):
+                        logger.warning(
+                            "[ImageResolver] 不支持的 MIME 类型 {}, 跳过: {}",
+                            media_type,
+                            url[:100],
+                        )
+                        return None
+
+                    # 预检 Content-Length（如果有）
+                    content_length = resp.headers.get("content-length")
+                    if content_length:
+                        try:
+                            if int(content_length) > _MAX_IMAGE_SIZE:
+                                logger.warning(
+                                    "[ImageResolver] Content-Length 超过大小限制: {}",
+                                    url[:100],
+                                )
+                                return None
+                        except (TypeError, ValueError):
+                            pass
+
+                    # 流式累计读取并检查大小
+                    chunks: list[bytes] = []
+                    total = 0
+                    async for chunk in resp.aiter_bytes():
+                        total += len(chunk)
+                        if total > _MAX_IMAGE_SIZE:
+                            logger.warning(
+                                "[ImageResolver] 文件超过大小限制 ({} bytes > {}): {}",
+                                total,
+                                _MAX_IMAGE_SIZE,
+                                url[:100],
+                            )
+                            return None
+                        chunks.append(chunk)
+
+                    data = b"".join(chunks)
+                    b64 = base64.b64encode(data).decode("ascii")
+                    return b64, media_type
+
+            logger.warning("[ImageResolver] 超过最大重定向次数: {}", url[:100])
+            return None
+        except Exception as e:
+            logger.warning("[ImageResolver] 下载文件失败: {} - {}", url[:100], e)
+            return None
+
+
+def _guess_media_type(url: str) -> str:
+    """从 URL 路径猜测 MIME 类型。"""
+    path = url.split("?")[0]
+    mt, _ = mimetypes.guess_type(path)
+    return mt or "application/octet-stream"
+
+
+# Claude API 支持的 MIME 类型前缀（图片/文档/音频等可直接作为 base64 内嵌的类型）
+_ACCEPTED_MIME_PREFIXES: tuple[str, ...] = (
+    "image/",
+    "application/pdf",
+    "text/",
+    "audio/",
+    "video/",
+)
+
+
+def _is_data_url(url: str) -> bool:
+    """判断是否是 data: URL（已内嵌 base64）。"""
+    return url.startswith("data:")
+
+
+async def resolve_image_urls(
+    internal: InternalRequest,
+    target_format: str,
+) -> None:
+    """遍历 InternalRequest 中所有 ImageBlock/FileBlock，对有 url 无 data 的进行下载转 base64。
+
+    仅当 target_format 需要 base64 时执行下载（如 CLAUDE）。
+    直接修改 internal 对象，无返回值。
+    """
+    target_upper = str(target_format).upper()
+
+    # 检查目标格式是否需要 base64
+    needs_base64 = any(target_upper.startswith(prefix) for prefix in _FORMATS_REQUIRING_BASE64)
+    if not needs_base64:
+        return
+
+    # 收集所有需要下载的 block 及其 URL
+    download_items: list[tuple[ImageBlock | FileBlock, str]] = []
+    for msg in internal.messages:
+        for block in msg.content:
+            if isinstance(block, ImageBlock):
+                if block.url and not block.data and not _is_data_url(block.url):
+                    download_items.append((block, block.url))
+            elif isinstance(block, FileBlock):
+                if block.file_url and not block.data and not _is_data_url(block.file_url):
+                    download_items.append((block, block.file_url))
+
+    if not download_items:
+        return
+
+    semaphore = asyncio.Semaphore(_MAX_CONCURRENT_DOWNLOADS)
+
+    # 复用同一个 client 并发下载所有文件（禁用自动重定向，在 _download_file 中手动处理）
+    async with httpx.AsyncClient(
+        follow_redirects=False,
+        timeout=httpx.Timeout(_DOWNLOAD_TIMEOUT),
+        limits=httpx.Limits(max_connections=_MAX_CONCURRENT_DOWNLOADS),
+    ) as client:
+        tasks = [_download_file(client, url, semaphore) for _, url in download_items]
+        results = await asyncio.gather(*tasks)
+
+    # 回写结果
+    for (block, url), result in zip(download_items, results):
+        if result is not None:
+            b64_data, media_type = result
+            block.data = b64_data
+            if not block.media_type:
+                block.media_type = media_type
+            # 保留原始 URL 以便调试，清除源 URL 字段避免语义模糊
+            block.extra["original_url"] = url
+            if isinstance(block, ImageBlock):
+                block.url = None
+            elif isinstance(block, FileBlock):
+                block.file_url = None
+
+
+__all__ = ["resolve_image_urls"]

src/core/api_format/conversion/internal.py

@@ -28,6 +28,8 @@ class ContentType(str, Enum):
     TEXT = "text"
     THINKING = "thinking"
     IMAGE = "image"
+    FILE = "file"
+    AUDIO = "audio"
     TOOL_USE = "tool_use"
     TOOL_RESULT = "tool_result"
     UNKNOWN = "unknown"
@@ -115,6 +117,30 @@ class ToolResultBlock:
     extra: dict[str, Any] = field(default_factory=dict)
 
 
+@dataclass
+class FileBlock:
+    """文件内容块（PDF、文档等）"""
+
+    type: ContentType = field(default=ContentType.FILE, init=False)
+    data: str | None = None  # base64 编码
+    media_type: str | None = None
+    file_id: str | None = None  # OpenAI file reference
+    file_url: str | None = None  # Gemini fileData URI
+    filename: str | None = None
+    extra: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class AudioBlock:
+    """音频内容块"""
+
+    type: ContentType = field(default=ContentType.AUDIO, init=False)
+    data: str | None = None  # base64 编码
+    media_type: str | None = None  # 完整 MIME（如 audio/mp3）
+    format: str | None = None  # 简短格式名（如 mp3, wav）
+    extra: dict[str, Any] = field(default_factory=dict)
+
+
 @dataclass
 class UnknownBlock:
     """未知内容块（用于前向兼容）"""
@@ -126,7 +152,14 @@ class UnknownBlock:
 
 
 ContentBlock = (
-    TextBlock | ThinkingBlock | ImageBlock | ToolUseBlock | ToolResultBlock | UnknownBlock
+    TextBlock
+    | ThinkingBlock
+    | ImageBlock
+    | FileBlock
+    | AudioBlock
+    | ToolUseBlock
+    | ToolResultBlock
+    | UnknownBlock
 )
 
 
@@ -174,6 +207,24 @@ class InstructionSegment:
     extra: dict[str, Any] = field(default_factory=dict)
 
 
+@dataclass
+class ThinkingConfig:
+    """统一的思考/推理配置（对齐 Claude thinking / Gemini thinkingConfig / OpenAI reasoning_effort）"""
+
+    enabled: bool = False
+    budget_tokens: int | None = None  # None = provider 默认
+    extra: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ResponseFormatConfig:
+    """统一的响应格式配置（JSON mode / structured output）"""
+
+    type: str = "text"  # "text" | "json_object" | "json_schema"
+    json_schema: dict[str, Any] | None = None
+    extra: dict[str, Any] = field(default_factory=dict)
+
+
 @dataclass
 class InternalRequest:
     """统一的请求表示"""
@@ -195,6 +246,27 @@ class InternalRequest:
     stream: bool = False
     tools: list[ToolDefinition] | None = None
     tool_choice: ToolChoice | None = None  # auto/none/required 或指定 tool_name
+
+    # 思考/推理配置
+    thinking: ThinkingConfig | None = None
+
+    # 并行工具调用控制
+    parallel_tool_calls: bool | None = None
+
+    # 采样参数
+    n: int | None = None
+    presence_penalty: float | None = None
+    frequency_penalty: float | None = None
+    seed: int | None = None
+    logprobs: bool | None = None
+    top_logprobs: int | None = None
+
+    # 响应格式
+    response_format: ResponseFormatConfig | None = None
+
+    # 模型输出上限（来自 GlobalModel.config.output_limit，用于跨格式转换时的 max_tokens 默认值）
+    output_limit: int | None = None
+
     extra: dict[str, Any] = field(default_factory=dict)  # 未识别字段透传
 
     def to_debug_dict(self) -> dict[str, Any]:
@@ -293,6 +365,8 @@ __all__ = [
     "TextBlock",
     "ThinkingBlock",
     "ImageBlock",
+    "FileBlock",
+    "AudioBlock",
     "ToolUseBlock",
     "ToolResultBlock",
     "UnknownBlock",
@@ -301,6 +375,8 @@ __all__ = [
     "InstructionSegment",
     "ToolDefinition",
     "ToolChoice",
+    "ThinkingConfig",
+    "ResponseFormatConfig",
     "InternalRequest",
     "UsageInfo",
     "InternalResponse",

src/core/api_format/conversion/normalizer.py

@@ -115,6 +115,40 @@ class FormatNormalizer(ABC):
             return ImageBlock(data=data, media_type=media_type)
         return ImageBlock(url=url)
 
+    # ============ 通用解析工具 ============
+
+    def _optional_int(self, value: Any) -> int | None:
+        if value is None:
+            return None
+        try:
+            return int(value)
+        except (TypeError, ValueError):
+            return None
+
+    def _optional_float(self, value: Any) -> float | None:
+        if value is None:
+            return None
+        try:
+            return float(value)
+        except (TypeError, ValueError):
+            return None
+
+    def _coerce_str_list(self, value: Any) -> list[str] | None:
+        if value is None:
+            return None
+        if isinstance(value, str):
+            return [value]
+        if isinstance(value, list):
+            return [str(x) for x in value if x is not None]
+        return None
+
+    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}
+
+    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)
+
     # ============ 视频转换（可选） ============
 
     def video_request_to_internal(self, request: dict[str, Any]) -> InternalVideoRequest:

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

@@ -14,12 +14,18 @@ from src.core.api_format.conversion.field_mappings import (
     ERROR_TYPE_MAPPINGS,
     RETRYABLE_ERROR_TYPES,
     STOP_REASON_MAPPINGS,
+    THINKING_BUDGET_TOKENS_MIN,
+    THINKING_BUDGET_TOKENS_PERCENTAGE,
     USAGE_FIELD_MAPPINGS,
+    WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES,
+    get_claude_default_max_tokens,
 )
 from src.core.api_format.conversion.internal import (
+    AudioBlock,
     ContentBlock,
     ContentType,
     ErrorType,
+    FileBlock,
     FormatCapabilities,
     ImageBlock,
     InstructionSegment,
@@ -31,6 +37,7 @@ from src.core.api_format.conversion.internal import (
     StopReason,
     TextBlock,
     ThinkingBlock,
+    ThinkingConfig,
     ToolChoice,
     ToolChoiceType,
     ToolDefinition,
@@ -97,9 +104,12 @@ class ClaudeNormalizer(FormatNormalizer):
 
         # 顶层 system 先进入 instructions（保持确定性优先级）
         sys_value = request.get("system")
-        sys_text, sys_dropped = self._collapse_claude_system(sys_value)
+        sys_text, sys_dropped, sys_segments = self._collapse_claude_system(sys_value)
         self._merge_dropped(dropped, sys_dropped)
-        if sys_text:
+        if sys_segments:
+            # 数组格式含 cache_control：保留逐段 InstructionSegment
+            instructions.extend(sys_segments)
+        elif sys_text:
             instructions.append(InstructionSegment(role=Role.SYSTEM, text=sys_text))
 
         messages: list[InternalMessage] = []
@@ -112,7 +122,7 @@ class ClaudeNormalizer(FormatNormalizer):
 
             # 兼容：少数客户端可能把 system/developer 混进 messages[]
             if role in ("system", "developer"):
-                text, md = self._collapse_claude_system(msg.get("content"))
+                text, md, _ = self._collapse_claude_system(msg.get("content"))
                 self._merge_dropped(dropped, md)
                 if text:
                     instructions.append(
@@ -134,6 +144,19 @@ class ClaudeNormalizer(FormatNormalizer):
         tools = self._claude_tools_to_internal(request.get("tools"))
         tool_choice = self._claude_tool_choice_to_internal(request.get("tool_choice"))
 
+        # 解析 Claude 原生 thinking 配置
+        thinking: ThinkingConfig | None = None
+        thinking_raw = request.get("thinking")
+        if isinstance(thinking_raw, dict):
+            thinking_type = str(thinking_raw.get("type") or "")
+            if thinking_type in ("enabled", "adaptive"):
+                budget = self._optional_int(thinking_raw.get("budget_tokens"))
+                thinking = ThinkingConfig(
+                    enabled=True,
+                    budget_tokens=budget,
+                    extra={"claude_thinking": thinking_raw},
+                )
+
         internal = InternalRequest(
             model=model,
             messages=messages,
@@ -147,6 +170,7 @@ class ClaudeNormalizer(FormatNormalizer):
             stream=bool(request.get("stream") or False),
             tools=tools,
             tool_choice=tool_choice,
+            thinking=thinking,
             extra={"claude": self._extract_extra(request, {"messages"})},
         )
 
@@ -161,7 +185,26 @@ class ClaudeNormalizer(FormatNormalizer):
         *,
         target_variant: str | None = None,
     ) -> dict[str, Any]:
-        system_text = internal.system or self._join_instructions(internal.instructions)
+        # system: 如果任一 InstructionSegment 有 cache_control，输出数组格式
+        has_cache_control = any(seg.extra.get("cache_control") for seg in internal.instructions)
+        if has_cache_control and internal.instructions:
+            system_value: str | list[dict[str, Any]] | None = [
+                {
+                    "type": "text",
+                    "text": seg.text,
+                    **(
+                        {"cache_control": seg.extra["cache_control"]}
+                        if seg.extra.get("cache_control")
+                        else {}
+                    ),
+                }
+                for seg in internal.instructions
+                if seg.text
+            ]
+            if not system_value:
+                system_value = None
+        else:
+            system_value = internal.system or self._join_instructions(internal.instructions)
 
         # Claude Messages API: messages[] 仅允许 user/assistant，且需要交替；这里做最小修复
         fixed_messages = self._coerce_claude_message_sequence(internal.messages)
@@ -170,14 +213,23 @@ class ClaudeNormalizer(FormatNormalizer):
             self._internal_message_to_claude(m) for m in fixed_messages
         ]
 
+        # max_tokens: 优先使用请求中的值, 其次 GlobalModel.output_limit, 最后硬编码默认值
+        effective_max_tokens: int
+        if internal.max_tokens is not None:
+            effective_max_tokens = internal.max_tokens
+        elif internal.output_limit is not None:
+            effective_max_tokens = internal.output_limit
+        else:
+            effective_max_tokens = get_claude_default_max_tokens(internal.model)
+
         result: dict[str, Any] = {
             "model": internal.model,
             "messages": out_messages,
-            "max_tokens": internal.max_tokens if internal.max_tokens is not None else 4096,
+            "max_tokens": effective_max_tokens,
         }
 
-        if system_text:
-            result["system"] = system_text
+        if system_value:
+            result["system"] = system_value
 
         if internal.temperature is not None:
             result["temperature"] = internal.temperature
@@ -191,18 +243,75 @@ class ClaudeNormalizer(FormatNormalizer):
             result["stream"] = True
 
         if internal.tools:
-            result["tools"] = [
-                {
+            claude_tools: list[dict[str, Any]] = []
+            for t in internal.tools:
+                # 跳过 Gemini 内置工具（在 Claude 中无对应物）
+                if t.extra.get("gemini_builtin_tool"):
+                    continue
+                tool_def: dict[str, Any] = {
                     "name": t.name,
-                    "description": t.description,
                     "input_schema": t.parameters or {},
                     **(t.extra.get("claude") or {}),
                 }
-                for t in internal.tools
-            ]
+                if t.description is not None:
+                    tool_def["description"] = t.description
+                claude_tools.append(tool_def)
+            if claude_tools:
+                result["tools"] = claude_tools
 
         if internal.tool_choice:
-            result["tool_choice"] = self._tool_choice_to_claude(internal.tool_choice)
+            tc = self._tool_choice_to_claude(internal.tool_choice)
+            # parallel_tool_calls=False -> disable_parallel_tool_use=True
+            if internal.parallel_tool_calls is False and tc.get("type") != "none":
+                tc["disable_parallel_tool_use"] = True
+            result["tool_choice"] = tc
+
+        # thinking 配置
+        if internal.thinking and internal.thinking.enabled:
+            # 优先使用原始 Claude thinking 配置（round-trip 透传）
+            claude_thinking = internal.thinking.extra.get("claude_thinking")
+            if isinstance(claude_thinking, dict):
+                result["thinking"] = claude_thinking
+            else:
+                thinking_out: dict[str, Any] = {"type": "enabled"}
+                # Claude API 要求 budget_tokens 必须提供
+                # 跨格式时 internal.model 可能是非 Claude 模型名，
+                # 仅当模型名以 claude- 开头时用模型感知默认值，否则用固定安全值
+                if internal.thinking.budget_tokens is not None:
+                    thinking_out["budget_tokens"] = internal.thinking.budget_tokens
+                else:
+                    # 基于 effective_max_tokens 计算兜底 budget，避免覆盖用户显式指定的 max_tokens
+                    thinking_out["budget_tokens"] = max(
+                        int(effective_max_tokens * THINKING_BUDGET_TOKENS_PERCENTAGE),
+                        THINKING_BUDGET_TOKENS_MIN,
+                    )
+                # 确保 budget_tokens >= 最小值（参考 new-api: 1280）
+                bt = thinking_out["budget_tokens"]
+                if bt < THINKING_BUDGET_TOKENS_MIN:
+                    thinking_out["budget_tokens"] = THINKING_BUDGET_TOKENS_MIN
+                    bt = THINKING_BUDGET_TOKENS_MIN
+                # Claude API 要求 budget_tokens < max_tokens，确保不违反约束
+                max_t = result.get("max_tokens")
+                if max_t is not None and bt >= max_t:
+                    result["max_tokens"] = bt + 1
+                result["thinking"] = thinking_out
+
+        # 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):
+            context_size = str(web_search_opts.get("search_context_size") or "medium")
+            max_uses = WEB_SEARCH_CONTEXT_SIZE_TO_MAX_USES.get(context_size, 5)
+            ws_tool: dict[str, Any] = {
+                "type": "web_search_20250305",
+                "name": "web_search",
+                "max_uses": max_uses,
+            }
+            user_location = web_search_opts.get("user_location")
+            if isinstance(user_location, dict):
+                ws_tool["user_location"] = user_location
+            if "tools" not in result:
+                result["tools"] = []
+            result["tools"].append(ws_tool)
 
         # 恢复 Claude 特有字段（如 metadata）
         claude_extra = internal.extra.get("claude") if isinstance(internal.extra, dict) else None
@@ -296,7 +405,45 @@ class ClaudeNormalizer(FormatNormalizer):
                         }
                     )
                 elif b.url:
-                    content.append({"type": "text", "text": f"[Image: {b.url}]"})
+                    content.append(
+                        {
+                            "type": "image",
+                            "source": {"type": "url", "url": b.url},
+                        }
+                    )
+                continue
+            if isinstance(b, FileBlock):
+                if b.data and b.media_type:
+                    content.append(
+                        {
+                            "type": "document",
+                            "source": {
+                                "type": "base64",
+                                "media_type": b.media_type,
+                                "data": b.data,
+                            },
+                        }
+                    )
+                elif b.file_url:
+                    content.append(
+                        {
+                            "type": "document",
+                            "source": {"type": "url", "url": b.file_url},
+                        }
+                    )
+                continue
+            if isinstance(b, AudioBlock):
+                if b.data and b.media_type:
+                    content.append(
+                        {
+                            "type": "document",
+                            "source": {
+                                "type": "base64",
+                                "media_type": b.media_type,
+                                "data": b.data,
+                            },
+                        }
+                    )
                 continue
             # Unknown/ToolResult 默认丢弃
 
@@ -740,9 +887,12 @@ class ClaudeNormalizer(FormatNormalizer):
             if btype == "text":
                 text = str(block.get("text") or "")
                 if text:
-                    blocks.append(
-                        TextBlock(text=text, extra=self._extract_extra(block, {"type", "text"}))
-                    )
+                    block_extra = self._extract_extra(block, {"type", "text"})
+                    # cache_control 透传
+                    cc = block.get("cache_control")
+                    if isinstance(cc, dict):
+                        block_extra["cache_control"] = cc
+                    blocks.append(TextBlock(text=text, extra=block_extra))
                 continue
 
             if btype == "image":
@@ -760,10 +910,42 @@ class ClaudeNormalizer(FormatNormalizer):
                     ):
                         blocks.append(ImageBlock(data=data, media_type=media_type))
                         continue
+                elif stype == "url":
+                    url = src.get("url")
+                    if isinstance(url, str) and url:
+                        blocks.append(ImageBlock(url=url))
+                        continue
                 dropped["claude_image_unsupported"] = dropped.get("claude_image_unsupported", 0) + 1
                 blocks.append(UnknownBlock(raw_type="image", payload=block))
                 continue
 
+            if btype == "document":
+                doc_src = block.get("source") or {}
+                if isinstance(doc_src, dict):
+                    doc_stype = doc_src.get("type")
+                    if doc_stype == "base64":
+                        blocks.append(
+                            FileBlock(
+                                data=doc_src.get("data"),
+                                media_type=doc_src.get("media_type"),
+                                extra=self._extract_extra(block, {"type", "source"}),
+                            )
+                        )
+                        continue
+                    elif doc_stype == "url":
+                        blocks.append(
+                            FileBlock(
+                                file_url=doc_src.get("url"),
+                                media_type=doc_src.get("media_type"),
+                                extra=self._extract_extra(block, {"type", "source"}),
+                            )
+                        )
+                        continue
+                dropped_key = f"claude_block:{btype}"
+                dropped[dropped_key] = dropped.get(dropped_key, 0) + 1
+                blocks.append(UnknownBlock(raw_type=btype, payload=block))
+                continue
+
             if btype == "tool_use":
                 tool_id = str(block.get("id") or "")
                 tool_name = str(block.get("name") or "")
@@ -862,15 +1044,21 @@ class ClaudeNormalizer(FormatNormalizer):
             extra={"claude": raw_block},
         )
 
-    def _collapse_claude_system(self, system_value: Any) -> tuple[str | None, dict[str, int]]:
+    def _collapse_claude_system(
+        self, system_value: Any
+    ) -> tuple[str | None, dict[str, int], list[InstructionSegment] | None]:
+        """解析 Claude system 字段。返回 (text, dropped, segments)。
+        segments 仅在 system 为数组且含 cache_control 时非空。"""
         dropped: dict[str, int] = {}
         if system_value is None:
-            return None, dropped
+            return None, dropped, None
         if isinstance(system_value, str):
-            return (system_value or None), dropped
+            return (system_value or None), dropped, None
 
         if isinstance(system_value, list):
             texts: list[str] = []
+            segments: list[InstructionSegment] = []
+            has_cache_control = False
             for item in system_value:
                 if not isinstance(item, dict):
                     dropped["claude_system_item_non_dict"] = (
@@ -881,14 +1069,26 @@ class ClaudeNormalizer(FormatNormalizer):
                     text = item.get("text")
                     if text:
                         texts.append(str(text))
+                        seg_extra: dict[str, Any] = {}
+                        cc = item.get("cache_control")
+                        if isinstance(cc, dict):
+                            seg_extra["cache_control"] = cc
+                            has_cache_control = True
+                        segments.append(
+                            InstructionSegment(role=Role.SYSTEM, text=str(text), extra=seg_extra)
+                        )
                 else:
                     dropped_key = f"claude_system_item:{item.get('type')}"
                     dropped[dropped_key] = dropped.get(dropped_key, 0) + 1
             joined = "\n\n".join(texts)
-            return (joined or None), dropped
+            return (
+                (joined or None),
+                dropped,
+                segments if has_cache_control else None,
+            )
 
         dropped["claude_system_unsupported"] = dropped.get("claude_system_unsupported", 0) + 1
-        return None, dropped
+        return None, dropped, None
 
     def _join_instructions(self, instructions: list[InstructionSegment]) -> str | None:
         parts = [seg.text for seg in instructions if seg.text]
@@ -957,6 +1157,11 @@ class ClaudeNormalizer(FormatNormalizer):
     def _internal_message_to_claude(self, msg: InternalMessage) -> dict[str, Any]:
         role = "user" if msg.role == Role.USER else "assistant"
 
+        # 预扫描：如果任一 TextBlock 带 cache_control，所有 TextBlock 都走结构化路径以保持顺序
+        force_structured_text = any(
+            isinstance(b, TextBlock) and b.extra.get("cache_control") for b in msg.content
+        )
+
         blocks: list[dict[str, Any]] = []
         text_parts: list[str] = []
 
@@ -966,7 +1171,14 @@ class ClaudeNormalizer(FormatNormalizer):
 
             if isinstance(b, TextBlock):
                 if b.text:
-                    text_parts.append(b.text)
+                    if force_structured_text:
+                        text_block: dict[str, Any] = {"type": "text", "text": b.text}
+                        cc = b.extra.get("cache_control") if b.extra else None
+                        if isinstance(cc, dict):
+                            text_block["cache_control"] = cc
+                        blocks.append(text_block)
+                    else:
+                        text_parts.append(b.text)
                 continue
 
             if isinstance(b, ImageBlock):
@@ -989,7 +1201,52 @@ class ClaudeNormalizer(FormatNormalizer):
                         }
                     )
                 elif b.url:
-                    text_parts.append(f"[Image: {b.url}]")
+                    blocks.append(
+                        {
+                            "type": "image",
+                            "source": {"type": "url", "url": b.url},
+                        }
+                    )
+                continue
+
+            if isinstance(b, FileBlock):
+                if b.data and b.media_type:
+                    blocks.append(
+                        {
+                            "type": "document",
+                            "source": {
+                                "type": "base64",
+                                "media_type": b.media_type,
+                                "data": b.data,
+                            },
+                        }
+                    )
+                elif b.file_url:
+                    blocks.append(
+                        {
+                            "type": "document",
+                            "source": {
+                                "type": "url",
+                                "url": b.file_url,
+                            },
+                        }
+                    )
+                elif b.file_id:
+                    text_parts.append(f"[File: {b.file_id}]")
+                continue
+
+            if isinstance(b, AudioBlock):
+                if b.data and b.media_type:
+                    blocks.append(
+                        {
+                            "type": "document",
+                            "source": {
+                                "type": "base64",
+                                "media_type": b.media_type,
+                                "data": b.data,
+                            },
+                        }
+                    )
                 continue
 
             if isinstance(b, ToolUseBlock) and role == "assistant":
@@ -1062,7 +1319,11 @@ class ClaudeNormalizer(FormatNormalizer):
 
         mapping = USAGE_FIELD_MAPPINGS.get("CLAUDE", {})
         fields: dict[str, int] = {}
-        extra = self._extract_extra(usage, set(mapping.keys()))
+        extra = self._extract_extra(
+            usage,
+            set(mapping.keys())
+            | {"cache_creation_input_tokens_5m", "cache_creation_input_tokens_1h"},
+        )
 
         for provider_key, internal_key in mapping.items():
             if provider_key in usage and usage.get(provider_key) is not None:
@@ -1076,6 +1337,21 @@ class ClaudeNormalizer(FormatNormalizer):
                 fields.get("input_tokens", 0) + fields.get("output_tokens", 0)
             )
 
+        # Claude cache_creation 5m/1h 细分（存入 extra）
+        cache_details: dict[str, int] = {}
+        for detail_key in (
+            "cache_creation_input_tokens_5m",
+            "cache_creation_input_tokens_1h",
+        ):
+            val = usage.get(detail_key)
+            if val is not None:
+                try:
+                    cache_details[detail_key] = int(val)
+                except (TypeError, ValueError):
+                    pass
+        if cache_details:
+            extra["cache_creation_details"] = cache_details
+
         return UsageInfo(
             input_tokens=int(fields.get("input_tokens", 0)),
             output_tokens=int(fields.get("output_tokens", 0)),
@@ -1094,6 +1370,16 @@ class ClaudeNormalizer(FormatNormalizer):
             result["cache_read_input_tokens"] = int(usage.cache_read_tokens)
         if usage.cache_write_tokens:
             result["cache_creation_input_tokens"] = int(usage.cache_write_tokens)
+        # 回写 5m/1h 细分
+        claude_extra = usage.extra.get("claude", {})
+        if isinstance(claude_extra, dict):
+            cache_details = claude_extra.get("cache_creation_details")
+            if isinstance(cache_details, dict):
+                for k, v in cache_details.items():
+                    try:
+                        result[k] = int(v)
+                    except (TypeError, ValueError):
+                        pass
         return result
 
     def _error_type_from_value(self, value: str) -> ErrorType:
@@ -1102,37 +1388,5 @@ class ClaudeNormalizer(FormatNormalizer):
         except ValueError:
             return ErrorType.UNKNOWN
 
-    def _optional_int(self, value: Any) -> int | None:
-        if value is None:
-            return None
-        try:
-            return int(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _optional_float(self, value: Any) -> float | None:
-        if value is None:
-            return None
-        try:
-            return float(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _coerce_str_list(self, value: Any) -> list[str] | None:
-        if value is None:
-            return None
-        if isinstance(value, str):
-            return [value]
-        if isinstance(value, list):
-            return [str(x) for x in value if x is not None]
-        return None
-
-    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}
-
-    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)
-
 
 __all__ = ["ClaudeNormalizer"]

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

@@ -12,6 +12,7 @@ Gemini (GenerateContent / streamGenerateContent) Normalizer
 """
 
 import json
+import re
 from typing import Any
 
 from src.core.api_format.conversion.field_mappings import (
@@ -21,9 +22,11 @@ from src.core.api_format.conversion.field_mappings import (
     USAGE_FIELD_MAPPINGS,
 )
 from src.core.api_format.conversion.internal import (
+    AudioBlock,
     ContentBlock,
     ContentType,
     ErrorType,
+    FileBlock,
     FormatCapabilities,
     ImageBlock,
     InstructionSegment,
@@ -31,10 +34,12 @@ from src.core.api_format.conversion.internal import (
     InternalMessage,
     InternalRequest,
     InternalResponse,
+    ResponseFormatConfig,
     Role,
     StopReason,
     TextBlock,
     ThinkingBlock,
+    ThinkingConfig,
     ToolChoice,
     ToolChoiceType,
     ToolDefinition,
@@ -131,6 +136,60 @@ def compact_gemini_contents(contents: list[dict[str, Any]]) -> list[dict[str, An
     return merged
 
 
+# Gemini 内置工具名称映射（输入侧：camelCase/snake_case key -> 标准化 ToolDefinition name）
+_BUILTIN_TOOL_KEYS: dict[str, str] = {
+    "googleSearch": "googleSearch",
+    "google_search": "googleSearch",
+    "codeExecution": "codeExecution",
+    "code_execution": "codeExecution",
+    "urlContext": "urlContext",
+    "url_context": "urlContext",
+}
+
+# Gemini 内置工具名称集合（输出侧：ToolDefinition.name -> 特殊处理）
+_GEMINI_BUILTIN_TOOLS: frozenset[str] = frozenset(
+    {
+        "google_search",
+        "googleSearch",
+        "code_execution",
+        "codeExecution",
+        "url_context",
+        "urlContext",
+    }
+)
+
+
+# Markdown 内嵌 base64 图片正则：![alt](data:image/xxx;base64,...)
+# 使用 [^)]+ 贪婪匹配至右括号，避免字符类量词的回溯风险
+_MD_IMAGE_RE = re.compile(r"!\[[^\]]*\]\(data:(image/[a-zA-Z0-9.+-]+);base64,([^)]+)\)")
+
+
+def _extract_markdown_images(
+    text: str,
+) -> tuple[str, list[dict[str, Any]]]:
+    """从文本中提取 Markdown 内嵌 base64 图片，返回 (剩余文本, inlineData parts)。
+
+    如果没有匹配到任何内嵌图片，返回 (原文本, [])。
+    """
+    # 快速预检：避免对纯文本执行正则
+    if "data:image/" not in text:
+        return text, []
+
+    image_parts: list[dict[str, Any]] = []
+
+    def _collect(m: re.Match[str]) -> str:
+        mime_type = m.group(1)
+        b64_data = m.group(2).replace("\n", "").replace(" ", "")
+        image_parts.append({"inline_data": {"mime_type": mime_type, "data": b64_data}})
+        return ""
+
+    remaining = _MD_IMAGE_RE.sub(_collect, text)
+    if not image_parts:
+        return text, []
+
+    return remaining, image_parts
+
+
 class GeminiNormalizer(FormatNormalizer):
     FORMAT_ID = "gemini:chat"
     capabilities = FormatCapabilities(
@@ -226,9 +285,28 @@ class GeminiNormalizer(FormatNormalizer):
 
         # 保留 generationConfig 中的特殊字段（responseModalities, thinkingConfig 等）
         # 这些字段在 _get_generation_config 中已提取，需要单独存储以便转换时使用
+        thinking: ThinkingConfig | None = None
+        n: int | None = None
+        response_format: ResponseFormatConfig | None = None
+
         if isinstance(generation_config, dict):
             response_modalities = generation_config.get("response_modalities")
             thinking_config = generation_config.get("thinking_config")
+
+            # 解析 thinkingConfig -> ThinkingConfig
+            if isinstance(thinking_config, dict):
+                include_thoughts = thinking_config.get(
+                    "include_thoughts", thinking_config.get("includeThoughts")
+                )
+                thinking_budget = thinking_config.get(
+                    "thinking_budget", thinking_config.get("thinkingBudget")
+                )
+                thinking = ThinkingConfig(
+                    enabled=bool(include_thoughts) if include_thoughts is not None else True,
+                    budget_tokens=self._optional_int(thinking_budget),
+                    extra={"gemini_thinking_config": thinking_config},
+                )
+
             if response_modalities or thinking_config:
                 google_extra: dict[str, Any] = {}
                 if response_modalities:
@@ -237,6 +315,32 @@ class GeminiNormalizer(FormatNormalizer):
                     google_extra["thinking_config"] = thinking_config
                 extra["google"] = google_extra
 
+            # 解析 candidateCount -> n
+            n = self._optional_int(
+                generation_config.get("candidate_count", generation_config.get("candidateCount"))
+            )
+
+            # 解析 response_format (responseMimeType / responseSchema)
+            resp_mime = generation_config.get(
+                "response_mime_type", generation_config.get("responseMimeType")
+            )
+            resp_schema = generation_config.get(
+                "response_schema", generation_config.get("responseSchema")
+            )
+            if resp_mime or resp_schema:
+                if resp_schema and isinstance(resp_schema, dict):
+                    response_format = ResponseFormatConfig(
+                        type="json_schema",
+                        json_schema=resp_schema,
+                        extra={"response_mime_type": resp_mime} if resp_mime else {},
+                    )
+                elif resp_mime and "json" in str(resp_mime).lower():
+                    response_format = ResponseFormatConfig(type="json_object")
+                elif resp_mime:
+                    response_format = ResponseFormatConfig(
+                        type="text", extra={"response_mime_type": resp_mime}
+                    )
+
         internal = InternalRequest(
             model=model,
             messages=messages,
@@ -250,6 +354,9 @@ class GeminiNormalizer(FormatNormalizer):
             stream=bool(request.get("stream") or False),
             tools=tools,
             tool_choice=tool_choice,
+            thinking=thinking,
+            n=n,
+            response_format=response_format,
             extra=extra,
         )
 
@@ -276,21 +383,36 @@ class GeminiNormalizer(FormatNormalizer):
         )
 
         # tools/tool_choice — clean unsupported JSON Schema fields from parameters
+        # Gemini 特殊内置工具名称需要单独处理
         tools = None
         if internal.tools:
             func_decls: list[dict[str, Any]] = []
+            builtin_tools: list[dict[str, Any]] = []
             for t in internal.tools:
+                # 检测 Gemini 内置工具
+                if t.name in _GEMINI_BUILTIN_TOOLS:
+                    canonical = _BUILTIN_TOOL_KEYS.get(t.name, t.name)
+                    builtin_tools.append({canonical: {}})
+                    continue
+
                 params = dict(t.parameters) if t.parameters else {}
                 if params:
                     _clean_gemini_schema(params)
                 decl: dict[str, Any] = {
                     "name": t.name,
-                    "description": t.description,
                     "parameters": params,
                     **(t.extra.get("gemini_function_declaration") or {}),
                 }
+                if t.description is not None:
+                    decl["description"] = t.description
                 func_decls.append(decl)
-            tools = [{"function_declarations": func_decls}]
+            if func_decls:
+                tools = [{"function_declarations": func_decls}]
+            else:
+                tools = []
+            tools.extend(builtin_tools)
+            if not tools:
+                tools = None
 
         tool_config = None
         if internal.tool_choice:
@@ -308,12 +430,40 @@ class GeminiNormalizer(FormatNormalizer):
         if internal.stop_sequences:
             generation_config["stop_sequences"] = list(internal.stop_sequences)
 
+        # 从 internal.thinking 输出 thinkingConfig（跨格式转换的标准路径）
+        if (
+            internal.thinking
+            and internal.thinking.enabled
+            and "thinkingConfig" not in generation_config
+        ):
+            gemini_tc: dict[str, Any] = {"includeThoughts": True}
+            if internal.thinking.budget_tokens is not None:
+                gemini_tc["thinkingBudget"] = internal.thinking.budget_tokens
+            generation_config["thinkingConfig"] = gemini_tc
+
+        # 从 internal.response_format 输出 responseMimeType / responseSchema
+        if internal.response_format and "responseMimeType" not in generation_config:
+            if (
+                internal.response_format.type == "json_schema"
+                and internal.response_format.json_schema
+            ):
+                generation_config["responseMimeType"] = "application/json"
+                schema = dict(internal.response_format.json_schema)
+                _clean_gemini_schema(schema)
+                generation_config["responseSchema"] = schema
+            elif internal.response_format.type == "json_object":
+                generation_config["responseMimeType"] = "application/json"
+
+        # 从 internal.n 输出 candidateCount
+        if internal.n is not None and internal.n > 1 and "candidateCount" not in generation_config:
+            generation_config["candidateCount"] = internal.n
+
         # 从 internal.extra["google"] 读取 OpenAI extra_body.google 透传的配置
         google_extra = internal.extra.get("google", {})
         if isinstance(google_extra, dict):
-            # 处理 thinking_config -> thinkingConfig
+            # 处理 thinking_config -> thinkingConfig（仅当上面标准路径未设置时）
             thinking_config = google_extra.get("thinking_config")
-            if isinstance(thinking_config, dict):
+            if isinstance(thinking_config, dict) and "thinkingConfig" not in generation_config:
                 # snake_case -> camelCase 转换
                 gemini_thinking: dict[str, Any] = {}
                 if "thinking_budget" in thinking_config:
@@ -472,6 +622,13 @@ class GeminiNormalizer(FormatNormalizer):
     ) -> dict[str, Any]:
         parts: list[dict[str, Any]] = []
         for b in internal.content:
+            if isinstance(b, ThinkingBlock):
+                if b.thinking:
+                    thought_part: dict[str, Any] = {"text": b.thinking, "thought": True}
+                    if b.signature:
+                        thought_part["thoughtSignature"] = b.signature
+                    parts.append(thought_part)
+                continue
             if isinstance(b, TextBlock):
                 if b.text:
                     parts.append({"text": b.text})
@@ -497,7 +654,28 @@ class GeminiNormalizer(FormatNormalizer):
                         }
                     )
                 elif b.url:
-                    parts.append({"text": f"[Image: {b.url}]"})
+                    parts.append(
+                        {"fileData": {"fileUri": b.url, "mimeType": b.media_type or "image/jpeg"}}
+                    )
+                continue
+            if isinstance(b, FileBlock):
+                if b.data and b.media_type:
+                    parts.append({"inlineData": {"mimeType": b.media_type, "data": b.data}})
+                elif b.file_url:
+                    parts.append(
+                        {
+                            "fileData": {
+                                "fileUri": b.file_url,
+                                "mimeType": b.media_type or "application/octet-stream",
+                            }
+                        }
+                    )
+                continue
+            if isinstance(b, AudioBlock):
+                if b.data and b.media_type:
+                    parts.append({"inlineData": {"mimeType": b.media_type, "data": b.data}})
+                elif b.data and b.format:
+                    parts.append({"inlineData": {"mimeType": f"audio/{b.format}", "data": b.data}})
                 continue
             # Unknown/ToolResult 默认丢弃
 
@@ -1319,7 +1497,15 @@ class GeminiNormalizer(FormatNormalizer):
                 )
                 data = inline.get("data")
                 if isinstance(mime_type, str) and mime_type and isinstance(data, str) and data:
-                    blocks.append(ImageBlock(data=data, media_type=mime_type))
+                    if mime_type.startswith("audio/"):
+                        # 音频 -> AudioBlock
+                        fmt = mime_type.split("/", 1)[1] if "/" in mime_type else None
+                        blocks.append(AudioBlock(data=data, media_type=mime_type, format=fmt))
+                    elif mime_type.startswith("image/"):
+                        blocks.append(ImageBlock(data=data, media_type=mime_type))
+                    else:
+                        # 其他类型 (PDF 等) -> FileBlock
+                        blocks.append(FileBlock(data=data, media_type=mime_type))
                 else:
                     dropped["gemini_inline_data_invalid"] = (
                         dropped.get("gemini_inline_data_invalid", 0) + 1
@@ -1327,6 +1513,27 @@ class GeminiNormalizer(FormatNormalizer):
                     blocks.append(UnknownBlock(raw_type="inline_data", payload=part))
                 continue
 
+            # fileData / file_data
+            file_data = part.get("fileData")
+            if file_data is None:
+                file_data = part.get("file_data")
+            if isinstance(file_data, dict):
+                file_uri = file_data.get("fileUri") or file_data.get("file_uri") or ""
+                file_mime = file_data.get("mimeType") or file_data.get("mime_type") or ""
+                if isinstance(file_mime, str) and file_mime.startswith("image/"):
+                    # 图片 MIME -> ImageBlock
+                    blocks.append(ImageBlock(url=str(file_uri), media_type=file_mime))
+                else:
+                    # 非图片 -> FileBlock
+                    blocks.append(
+                        FileBlock(
+                            file_url=str(file_uri),
+                            media_type=str(file_mime) if file_mime else None,
+                            extra={"gemini": part},
+                        )
+                    )
+                continue
+
             func_call = part.get("function_call")
             if func_call is None:
                 func_call = part.get("functionCall")
@@ -1457,14 +1664,47 @@ class GeminiNormalizer(FormatNormalizer):
 
             if isinstance(b, TextBlock):
                 if b.text:
-                    parts.append({"text": b.text})
+                    # 检测 Markdown 内嵌 base64 图片并提取为独立 inlineData part
+                    remaining_text, image_parts = _extract_markdown_images(b.text)
+                    if image_parts:
+                        if remaining_text.strip():
+                            parts.append({"text": remaining_text})
+                        parts.extend(image_parts)
+                    else:
+                        parts.append({"text": b.text})
                 continue
 
             if isinstance(b, ImageBlock):
                 if b.data and b.media_type:
                     parts.append({"inline_data": {"mime_type": b.media_type, "data": b.data}})
                 elif b.url:
-                    parts.append({"text": f"[Image: {b.url}]"})
+                    # Gemini 支持 fileData URI 引用
+                    parts.append(
+                        {"fileData": {"fileUri": b.url, "mimeType": b.media_type or "image/jpeg"}}
+                    )
+                continue
+
+            if isinstance(b, FileBlock):
+                if b.data and b.media_type:
+                    parts.append({"inline_data": {"mime_type": b.media_type, "data": b.data}})
+                elif b.file_url:
+                    parts.append(
+                        {
+                            "fileData": {
+                                "fileUri": b.file_url,
+                                "mimeType": b.media_type or "application/octet-stream",
+                            }
+                        }
+                    )
+                continue
+
+            if isinstance(b, AudioBlock):
+                if b.data and b.media_type:
+                    parts.append({"inline_data": {"mime_type": b.media_type, "data": b.data}})
+                elif b.data and b.format:
+                    parts.append(
+                        {"inline_data": {"mime_type": f"audio/{b.format}", "data": b.data}}
+                    )
                 continue
 
             if isinstance(b, ToolUseBlock) and role == "model":
@@ -1646,6 +1886,19 @@ class GeminiNormalizer(FormatNormalizer):
         if thinking_config:
             normalized["thinking_config"] = thinking_config
 
+        # 保留 candidateCount
+        candidate_count = pick("candidate_count", "candidateCount")
+        if candidate_count is not None:
+            normalized["candidate_count"] = candidate_count
+
+        # 保留 responseMimeType / responseSchema
+        resp_mime = pick("response_mime_type", "responseMimeType")
+        if resp_mime is not None:
+            normalized["response_mime_type"] = resp_mime
+        resp_schema = pick("response_schema", "responseSchema")
+        if resp_schema is not None:
+            normalized["response_schema"] = resp_schema
+
         return {k: v for k, v in normalized.items() if v is not None}
 
     def _gemini_tools_to_internal(self, tools: Any) -> list[ToolDefinition] | None:
@@ -1657,35 +1910,49 @@ class GeminiNormalizer(FormatNormalizer):
             if not isinstance(tool, dict):
                 continue
 
-            decls = tool.get("function_declarations")
-            if decls is None:
-                decls = tool.get("functionDeclarations")
-
-            if not isinstance(decls, list):
-                continue
-
-            for decl in decls:
-                if not isinstance(decl, dict):
-                    continue
-                name = str(decl.get("name") or "")
-                if not name:
-                    continue
-                out.append(
-                    ToolDefinition(
-                        name=name,
-                        description=decl.get("description"),
-                        parameters=(
-                            decl.get("parameters")
-                            if isinstance(decl.get("parameters"), dict)
-                            else None
-                        ),
-                        extra={
-                            "gemini_function_declaration": self._extract_extra(
-                                decl, {"name", "description", "parameters"}
-                            )
-                        },
+            # 检测 Gemini 内置工具（如 {"googleSearch": {}}, {"codeExecution": {}}）
+            for key, canonical_name in _BUILTIN_TOOL_KEYS.items():
+                if key in tool:
+                    out.append(
+                        ToolDefinition(
+                            name=canonical_name,
+                            description=None,
+                            parameters=None,
+                            extra={"gemini_builtin_tool": True},
+                        )
+                    )
+                    break
+            else:
+                # 非内置工具：解析 functionDeclarations
+                decls = tool.get("function_declarations")
+                if decls is None:
+                    decls = tool.get("functionDeclarations")
+
+                if not isinstance(decls, list):
+                    continue
+
+                for decl in decls:
+                    if not isinstance(decl, dict):
+                        continue
+                    name = str(decl.get("name") or "")
+                    if not name:
+                        continue
+                    out.append(
+                        ToolDefinition(
+                            name=name,
+                            description=decl.get("description"),
+                            parameters=(
+                                decl.get("parameters")
+                                if isinstance(decl.get("parameters"), dict)
+                                else None
+                            ),
+                            extra={
+                                "gemini_function_declaration": self._extract_extra(
+                                    decl, {"name", "description", "parameters"}
+                                )
+                            },
+                        )
                     )
-                )
 
         return out or None
 
@@ -1784,37 +2051,5 @@ class GeminiNormalizer(FormatNormalizer):
         except ValueError:
             return ErrorType.UNKNOWN
 
-    def _optional_int(self, value: Any) -> int | None:
-        if value is None:
-            return None
-        try:
-            return int(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _optional_float(self, value: Any) -> float | None:
-        if value is None:
-            return None
-        try:
-            return float(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _coerce_str_list(self, value: Any) -> list[str] | None:
-        if value is None:
-            return None
-        if isinstance(value, str):
-            return [value]
-        if isinstance(value, list):
-            return [str(x) for x in value if x is not None]
-        return None
-
-    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}
-
-    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)
-
 
 __all__ = ["GeminiNormalizer"]

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

@@ -14,12 +14,16 @@ from typing import Any
 
 from src.core.api_format.conversion.field_mappings import (
     ERROR_TYPE_MAPPINGS,
+    REASONING_EFFORT_TO_THINKING_BUDGET,
     RETRYABLE_ERROR_TYPES,
+    THINKING_BUDGET_TO_REASONING_EFFORT,
 )
 from src.core.api_format.conversion.internal import (
+    AudioBlock,
     ContentBlock,
     ContentType,
     ErrorType,
+    FileBlock,
     FormatCapabilities,
     ImageBlock,
     InstructionSegment,
@@ -27,10 +31,12 @@ from src.core.api_format.conversion.internal import (
     InternalMessage,
     InternalRequest,
     InternalResponse,
+    ResponseFormatConfig,
     Role,
     StopReason,
     TextBlock,
     ThinkingBlock,
+    ThinkingConfig,
     ToolChoice,
     ToolChoiceType,
     ToolDefinition,
@@ -86,6 +92,8 @@ class OpenAINormalizer(FormatNormalizer):
         StopReason.STOP_SEQUENCE: "stop",
         StopReason.TOOL_USE: "tool_calls",
         StopReason.CONTENT_FILTERED: "content_filter",
+        StopReason.REFUSAL: "content_filter",
+        StopReason.PAUSE_TURN: "stop",
         StopReason.UNKNOWN: "stop",
     }
 
@@ -181,6 +189,52 @@ class OpenAINormalizer(FormatNormalizer):
             if isinstance(google_extra, dict) and google_extra:
                 extra["google"] = google_extra
 
+        # reasoning_effort -> ThinkingConfig
+        thinking: ThinkingConfig | None = None
+        reasoning_effort = request.get("reasoning_effort")
+        if (
+            isinstance(reasoning_effort, str)
+            and reasoning_effort in REASONING_EFFORT_TO_THINKING_BUDGET
+        ):
+            thinking = ThinkingConfig(
+                enabled=True,
+                budget_tokens=REASONING_EFFORT_TO_THINKING_BUDGET[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")
+        if isinstance(web_search_options, dict):
+            extra["web_search_options"] = web_search_options
+
+        # parallel_tool_calls
+        parallel_tool_calls: bool | None = None
+        ptc = request.get("parallel_tool_calls")
+        if ptc is not None:
+            parallel_tool_calls = bool(ptc)
+
+        # 采样参数
+        n_value = self._optional_int(request.get("n"))
+        presence_penalty = self._optional_float(request.get("presence_penalty"))
+        frequency_penalty = self._optional_float(request.get("frequency_penalty"))
+        seed = self._optional_int(request.get("seed"))
+        logprobs = request.get("logprobs")
+        logprobs_val: bool | None = bool(logprobs) if logprobs is not None else None
+        top_logprobs = self._optional_int(request.get("top_logprobs"))
+
+        # response_format
+        response_format: ResponseFormatConfig | None = None
+        rf = request.get("response_format")
+        if isinstance(rf, dict):
+            rf_type = str(rf.get("type") or "text")
+            rf_schema = rf.get("json_schema") if isinstance(rf.get("json_schema"), dict) else None
+            if rf_type != "text":
+                response_format = ResponseFormatConfig(
+                    type=rf_type,
+                    json_schema=rf_schema,
+                    extra=self._extract_extra(rf, {"type", "json_schema"}),
+                )
+
         internal = InternalRequest(
             model=model,
             messages=messages,
@@ -193,6 +247,15 @@ class OpenAINormalizer(FormatNormalizer):
             stream=bool(request.get("stream") or False),
             tools=tools,
             tool_choice=tool_choice,
+            thinking=thinking,
+            parallel_tool_calls=parallel_tool_calls,
+            n=n_value,
+            presence_penalty=presence_penalty,
+            frequency_penalty=frequency_penalty,
+            seed=seed,
+            logprobs=logprobs_val,
+            top_logprobs=top_logprobs,
+            response_format=response_format,
             extra=extra,
         )
 
@@ -240,24 +303,67 @@ class OpenAINormalizer(FormatNormalizer):
             result["stream_options"] = {"include_usage": True}
 
         if internal.tools:
-            result["tools"] = [
-                {
-                    "type": "function",
-                    "function": {
-                        "name": t.name,
-                        "description": t.description,
-                        "parameters": t.parameters or {},
-                        **(t.extra.get("openai_function") or {}),
-                    },
-                    **(t.extra.get("openai_tool") or {}),
+            openai_tools: list[dict[str, Any]] = []
+            for t in internal.tools:
+                # 跳过 Gemini 内置工具（在 OpenAI 中无对应物）
+                if t.extra.get("gemini_builtin_tool"):
+                    continue
+                func: dict[str, Any] = {
+                    "name": t.name,
+                    "parameters": t.parameters or {},
+                    **(t.extra.get("openai_function") or {}),
                 }
-                for t in internal.tools
-            ]
+                if t.description is not None:
+                    func["description"] = t.description
+                openai_tools.append(
+                    {
+                        "type": "function",
+                        "function": func,
+                        **(t.extra.get("openai_tool") or {}),
+                    }
+                )
+            if openai_tools:
+                result["tools"] = openai_tools
 
         if internal.tool_choice:
             result["tool_choice"] = self._tool_choice_to_openai(internal.tool_choice)
 
-        # 其余字段：保留在 internal.extra 中，但不默认回写到 OpenAI body（兼容优先）
+        # thinking -> reasoning_effort
+        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:
+                for threshold, level in THINKING_BUDGET_TO_REASONING_EFFORT:
+                    if internal.thinking.budget_tokens <= threshold:
+                        effort = level
+                        break
+            if effort:
+                result["reasoning_effort"] = effort
+
+        # parallel_tool_calls
+        if internal.parallel_tool_calls is not None:
+            result["parallel_tool_calls"] = internal.parallel_tool_calls
+
+        # 采样参数
+        if internal.n is not None and internal.n > 1:
+            result["n"] = internal.n
+        for attr, key in (
+            ("presence_penalty", "presence_penalty"),
+            ("frequency_penalty", "frequency_penalty"),
+            ("seed", "seed"),
+            ("logprobs", "logprobs"),
+            ("top_logprobs", "top_logprobs"),
+        ):
+            val = getattr(internal, attr, None)
+            if val is not None:
+                result[key] = val
+
+        # response_format
+        if internal.response_format and internal.response_format.type != "text":
+            rf: dict[str, Any] = {"type": internal.response_format.type}
+            if internal.response_format.json_schema:
+                rf["json_schema"] = internal.response_format.json_schema
+            result["response_format"] = rf
+
         return result
 
     # =========================
@@ -344,6 +450,8 @@ class OpenAINormalizer(FormatNormalizer):
         content_value = self._blocks_to_openai_content(content_blocks)
         if content_value is not None:
             message["content"] = content_value
+        else:
+            message["content"] = None
 
         if tool_blocks:
             message["tool_calls"] = [
@@ -366,11 +474,23 @@ class OpenAINormalizer(FormatNormalizer):
             total = internal.usage.total_tokens or (
                 internal.usage.input_tokens + internal.usage.output_tokens
             )
-            out["usage"] = {
+            usage_out: dict[str, Any] = {
                 "prompt_tokens": int(internal.usage.input_tokens),
                 "completion_tokens": int(internal.usage.output_tokens),
                 "total_tokens": int(total),
             }
+            # 回写 prompt_tokens_details（cached_tokens）
+            if internal.usage.cache_read_tokens:
+                usage_out["prompt_tokens_details"] = {
+                    "cached_tokens": int(internal.usage.cache_read_tokens),
+                }
+            # 回写 completion_tokens_details
+            openai_extra = internal.usage.extra.get("openai", {})
+            if isinstance(openai_extra, dict):
+                ctd = openai_extra.get("completion_tokens_details")
+                if isinstance(ctd, dict):
+                    usage_out["completion_tokens_details"] = ctd
+            out["usage"] = usage_out
 
         return out
 
@@ -1061,6 +1181,45 @@ class OpenAINormalizer(FormatNormalizer):
                     blocks.append(UnknownBlock(raw_type="image_url", payload=part))
                 continue
 
+            if ptype == "file":
+                file_data = part.get("file_data")
+                file_id = part.get("file_id")
+                if isinstance(file_data, dict):
+                    blocks.append(
+                        FileBlock(
+                            data=file_data.get("data"),
+                            media_type=file_data.get("mime_type"),
+                            filename=file_data.get("filename"),
+                            extra=self._extract_extra(part, {"type", "file_data"}),
+                        )
+                    )
+                elif isinstance(file_id, str) and file_id:
+                    blocks.append(
+                        FileBlock(
+                            file_id=file_id,
+                            extra=self._extract_extra(part, {"type", "file_id"}),
+                        )
+                    )
+                else:
+                    blocks.append(UnknownBlock(raw_type="file", payload=part))
+                continue
+
+            if ptype == "input_audio":
+                audio_data = part.get("input_audio") or {}
+                if isinstance(audio_data, dict):
+                    audio_fmt = str(audio_data.get("format") or "")
+                    blocks.append(
+                        AudioBlock(
+                            data=audio_data.get("data"),
+                            media_type=f"audio/{audio_fmt}" if audio_fmt else None,
+                            format=audio_fmt or None,
+                            extra=self._extract_extra(part, {"type", "input_audio"}),
+                        )
+                    )
+                else:
+                    blocks.append(UnknownBlock(raw_type="input_audio", payload=part))
+                continue
+
             # 其他类型：UnknownBlock
             dropped_key = f"openai_part:{ptype}"
             dropped[dropped_key] = dropped.get(dropped_key, 0) + 1
@@ -1288,14 +1447,35 @@ class OpenAINormalizer(FormatNormalizer):
         if input_tokens == 0 and output_tokens == 0 and total_tokens_int == 0:
             return None
 
+        # prompt_tokens_details -> cache_read_tokens
+        cache_read = 0
+        ptd = usage.get("prompt_tokens_details")
+        if isinstance(ptd, dict):
+            cache_read = int(ptd.get("cached_tokens") or 0)
+
+        # completion_tokens_details（reasoning_tokens 等存入 extra）
+        ctd = usage.get("completion_tokens_details")
+
         extra = self._extract_extra(
             usage,
-            {"prompt_tokens", "completion_tokens", "total_tokens"},
+            {
+                "prompt_tokens",
+                "completion_tokens",
+                "total_tokens",
+                "prompt_tokens_details",
+                "completion_tokens_details",
+            },
         )
+
+        # 保留 completion_tokens_details 细分到 extra
+        if isinstance(ctd, dict):
+            extra["completion_tokens_details"] = ctd
+
         return UsageInfo(
             input_tokens=input_tokens,
             output_tokens=output_tokens,
             total_tokens=total_tokens_int,
+            cache_read_tokens=cache_read,
             extra={"openai": extra} if extra else {},
         )
 
@@ -1326,9 +1506,47 @@ class OpenAINormalizer(FormatNormalizer):
                     parts.append({"type": "image_url", "image_url": {"url": b.url}})
                 continue
 
+            if isinstance(b, FileBlock):
+                if b.data and b.media_type:
+                    # OpenAI file content part
+                    file_part: dict[str, Any] = {
+                        "type": "file",
+                        "file_data": {
+                            "mime_type": b.media_type,
+                            "data": b.data,
+                        },
+                    }
+                    if b.filename:
+                        file_part["file_data"]["filename"] = b.filename
+                    parts.append(file_part)
+                elif b.file_id:
+                    parts.append({"type": "file", "file_id": b.file_id})
+                elif b.file_url:
+                    # 回退为文本描述
+                    text_parts.append(f"[File: {b.file_url}]")
+                continue
+
+            if isinstance(b, AudioBlock):
+                if b.data:
+                    audio_fmt = b.format or (
+                        b.media_type.split("/", 1)[1]
+                        if b.media_type and "/" in b.media_type
+                        else "mp3"
+                    )
+                    parts.append(
+                        {
+                            "type": "input_audio",
+                            "input_audio": {
+                                "data": b.data,
+                                "format": audio_fmt,
+                            },
+                        }
+                    )
+                continue
+
             # Unknown / Tool blocks 不进入 OpenAI content
 
-        # 如果有 OpenAI 原生格式的图片（URL 引用），使用 multipart content
+        # 如果有 OpenAI 原生格式的图片/文件/音频（multipart content parts），使用 multipart content
         if parts:
             if text_parts:
                 parts = [{"type": "text", "text": "\n".join(text_parts)}] + parts
@@ -1355,10 +1573,10 @@ class OpenAINormalizer(FormatNormalizer):
                 tool_blocks.append(b)
                 continue
             if isinstance(b, ToolResultBlock):
-                # InternalResponse.content 不应该包含 tool_result；忽略
                 continue
             if isinstance(b, UnknownBlock):
                 continue
+            # TextBlock, ImageBlock, FileBlock, AudioBlock -> content
             content_blocks.append(b)
         return thinking_blocks, content_blocks, tool_blocks
 
@@ -1495,38 +1713,6 @@ class OpenAINormalizer(FormatNormalizer):
         except ValueError:
             return ErrorType.UNKNOWN
 
-    def _optional_int(self, value: Any) -> int | None:
-        if value is None:
-            return None
-        try:
-            return int(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _optional_float(self, value: Any) -> float | None:
-        if value is None:
-            return None
-        try:
-            return float(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _coerce_str_list(self, value: Any) -> list[str] | None:
-        if value is None:
-            return None
-        if isinstance(value, str):
-            return [value]
-        if isinstance(value, list):
-            return [str(x) for x in value if x is not None]
-        return None
-
-    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}
-
-    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)
-
     def _ensure_tool_block_index(self, ss: dict[str, Any], tool_key: str) -> int:
         mapping = ss.get("tool_id_to_block_index")
         if not isinstance(mapping, dict):

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

@@ -1343,22 +1343,6 @@ class OpenAICliNormalizer(FormatNormalizer):
             return "tool"
         return "user"
 
-    def _optional_int(self, value: Any) -> int | None:
-        if value is None:
-            return None
-        try:
-            return int(value)
-        except (TypeError, ValueError):
-            return None
-
-    def _optional_float(self, value: Any) -> float | None:
-        if value is None:
-            return None
-        try:
-            return float(value)
-        except (TypeError, ValueError):
-            return None
-
     def _coerce_str_list(self, value: Any) -> list[str] | None:
         if value is None:
             return None

src/core/api_format/conversion/registry.py

@@ -16,6 +16,7 @@ from contextlib import contextmanager
 from typing import Any
 
 from src.core.api_format.conversion.exceptions import FormatConversionError
+from src.core.api_format.conversion.image_resolver import resolve_image_urls
 from src.core.api_format.conversion.internal import InternalRequest, ToolResultBlock, ToolUseBlock
 from src.core.api_format.conversion.normalizer import FormatNormalizer
 from src.core.api_format.conversion.stream_state import StreamState
@@ -104,6 +105,7 @@ class FormatConversionRegistry:
         target_format: str,
         *,
         target_variant: str | None = None,
+        output_limit: int | None = None,
     ) -> dict[str, Any]:
         if str(source_format).upper() == str(target_format).upper() and not target_variant:
             return request
@@ -116,11 +118,43 @@ class FormatConversionRegistry:
         ):
             try:
                 internal = src.request_to_internal(request)
+                internal.output_limit = output_limit
                 self._repair_internal_tool_call_ids(internal)
                 return tgt.request_from_internal(internal, target_variant=target_variant)
             except Exception as e:
                 raise FormatConversionError(source_format, target_format, str(e)) from e
 
+    async def convert_request_async(
+        self,
+        request: dict[str, Any],
+        source_format: str,
+        target_format: str,
+        *,
+        target_variant: str | None = None,
+        output_limit: int | None = None,
+    ) -> dict[str, Any]:
+        """异步版本的 convert_request，在 internal 阶段执行图片 URL 下载等异步操作。"""
+        if str(source_format).upper() == str(target_format).upper() and not target_variant:
+            return request
+
+        src = self._require_normalizer(source_format)
+        tgt = self._require_normalizer(target_format)
+
+        with _track_conversion_metrics(
+            "request", str(source_format).upper(), str(target_format).upper()
+        ):
+            try:
+                internal = src.request_to_internal(request)
+                internal.output_limit = output_limit
+                self._repair_internal_tool_call_ids(internal)
+
+                # 异步阶段：解析图片 URL -> base64（仅在目标格式需要时）
+                await resolve_image_urls(internal, str(target_format).upper())
+
+                return tgt.request_from_internal(internal, target_variant=target_variant)
+            except Exception as e:
+                raise FormatConversionError(source_format, target_format, str(e)) from e
+
     def convert_response(
         self,
         response: dict[str, Any],

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]:
         """