feat: OAuth 导入导出、提供商筛选、Gemini 图像生成支持与流式处理增强

- OAuth: 支持通过 Refresh Token 导入账号（文件拖拽/粘贴），OAuth Key 可导出为 JSON
- OAuth: 所有 OAuth 端点添加 require_admin 鉴权
- 提供商管理: 新增状态/API格式/模型三级筛选，后端返回 global_model_ids
- Gemini: 新增图像生成模型适配（finalize_provider_request 钩子 + envelope 跳过不兼容字段）
- 流式处理: buffer 残留数据 flush 与 token 兜底估算
- 上游元数据: 提取 merge_upstream_metadata，配额耗尽模型保留与深度合并
- Antigravity 配额: 无 quotaInfo 时视为耗尽，移除 Other 兜底分组
- README: 新增升级备份与回滚指南

src/api/admin/endpoints/keys.py

@@ -30,6 +30,7 @@ from src.models.endpoint_models import (
     EndpointAPIKeyUpdate,
 )
 from src.services.cache.provider_cache import ProviderCacheService
+from src.services.model.upstream_fetcher import merge_upstream_metadata
 from src.utils.auth_utils import require_admin
 
 router = APIRouter(tags=["Provider Keys"])
@@ -467,6 +468,29 @@ class AdminRevealEndpointKeyAdapter(AdminApiAdapter):
                     "无法解密认证配置，可能是加密密钥已更改。请重新添加该密钥。"
                 )
 
+        # OAuth 类型：返回 access_token + refresh_token
+        if auth_type == "oauth":
+            try:
+                decrypted_key = crypto_service.decrypt(key.api_key)
+            except Exception as e:
+                logger.error(f"解密 Key 失败: ID={self.key_id}, Error={e}")
+                raise InvalidRequestException(
+                    "无法解密 API Key，可能是加密密钥已更改。请重新添加该密钥。"
+                )
+            result: dict[str, Any] = {"auth_type": "oauth", "api_key": decrypted_key}
+            encrypted_auth_config = getattr(key, "auth_config", None)
+            if encrypted_auth_config:
+                try:
+                    decrypted_config = crypto_service.decrypt(encrypted_auth_config)
+                    auth_config = json.loads(decrypted_config)
+                    refresh_token = auth_config.get("refresh_token")
+                    if refresh_token:
+                        result["refresh_token"] = refresh_token
+                except Exception as e:
+                    logger.error(f"解密 auth_config 失败: ID={self.key_id}, Error={e}")
+            logger.info(f"[REVEAL] 查看 OAuth Key: ID={self.key_id}, Name={key.name}")
+            return result
+
         # API Key 类型返回 api_key
         try:
             decrypted_key = crypto_service.decrypt(key.api_key)
@@ -1178,13 +1202,9 @@ class AdminRefreshProviderQuotaAdapter(AdminApiAdapter):
                 if key.id in metadata_updates:
                     updates = metadata_updates[key.id]
                     if isinstance(updates, dict):
-                        # NOTE: upstream_metadata is a plain JSON column (not MutableDict),
-                        # so in-place mutation won't be persisted reliably. Always assign
-                        # a new dict object to mark the column as dirty.
-                        current = key.upstream_metadata
-                        merged: dict = dict(current) if isinstance(current, dict) else {}
-                        merged.update(updates)
-                        key.upstream_metadata = merged
+                        key.upstream_metadata = merge_upstream_metadata(
+                            key.upstream_metadata, updates
+                        )
                     db.add(key)
 
         # 提交数据库更改

src/api/admin/provider_oauth.py

@@ -35,7 +35,8 @@ from src.core.provider_oauth_utils import enrich_auth_config, post_oauth_token
 from src.core.provider_templates.fixed_providers import FIXED_PROVIDERS
 from src.core.provider_templates.types import ProviderType
 from src.database.database import get_db
-from src.models.database import Provider, ProviderAPIKey
+from src.models.database import Provider, ProviderAPIKey, User
+from src.utils.auth_utils import require_admin
 
 router = APIRouter(prefix="/api/admin/provider-oauth", tags=["Provider OAuth"])
 
@@ -189,7 +190,7 @@ def _parse_callback_params(callback_url: str) -> dict[str, str]:
 
 
 @router.get("/supported-types")
-async def supported_types() -> list[dict[str, Any]]:
+async def supported_types(_: User = Depends(require_admin)) -> list[dict[str, Any]]:
     # 不返回 client_secret
     result: list[dict[str, Any]] = []
     for provider_type, template in FIXED_PROVIDERS.items():
@@ -216,6 +217,7 @@ async def start_oauth(
     key_id: str,
     request: Request,
     db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
 ) -> StartOAuthResponse:
     key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
     if not key:
@@ -295,6 +297,7 @@ async def complete_oauth(
     payload: CompleteOAuthRequest,
     request: Request,
     db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
 ) -> CompleteOAuthResponse:
     redis = await get_redis_client(require_redis=True)
     assert redis is not None
@@ -428,6 +431,7 @@ async def refresh_oauth(
     key_id: str,
     request: Request,
     db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
 ) -> CompleteOAuthResponse:
     key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
     if not key:
@@ -586,6 +590,7 @@ async def start_provider_oauth(
     provider_id: str,
     request: Request,
     db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
 ) -> StartOAuthResponse:
     """基于 Provider 启动 OAuth（不需要预先创建 key）。"""
     provider = db.query(Provider).filter(Provider.id == provider_id).first()
@@ -659,6 +664,7 @@ async def complete_provider_oauth(
     payload: ProviderCompleteOAuthRequest,
     request: Request,
     db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
 ) -> ProviderCompleteOAuthResponse:
     """完成 Provider OAuth 并创建 key。"""
     redis = await get_redis_client(require_redis=True)
@@ -799,3 +805,155 @@ async def complete_provider_oauth(
         has_refresh_token=bool(refresh_token),
         email=auth_config.get("email"),
     )
+
+
+# ==============================================================================
+# Import Refresh Token (从导出文件导入)
+# ==============================================================================
+
+
+class ImportRefreshTokenRequest(BaseModel):
+    refresh_token: str = Field(..., min_length=1, description="Refresh Token")
+    name: str | None = Field(None, max_length=100, description="账号名称（可选）")
+
+
+@router.post(
+    "/providers/{provider_id}/import-refresh-token",
+    response_model=ProviderCompleteOAuthResponse,
+)
+async def import_refresh_token(
+    provider_id: str,
+    payload: ImportRefreshTokenRequest,
+    db: Session = Depends(get_db),
+    _: User = Depends(require_admin),
+) -> ProviderCompleteOAuthResponse:
+    """通过 Refresh Token 导入 OAuth 账号。
+
+    使用导出的 Refresh Token 换取 Access Token 并创建新的 OAuth Key。
+    """
+    provider = db.query(Provider).filter(Provider.id == provider_id).first()
+    if not provider:
+        raise NotFoundException("Provider 不存在", "provider")
+    provider_type = _require_fixed_provider(provider)
+
+    try:
+        template = FIXED_PROVIDERS.get(ProviderType(provider_type))
+    except Exception:
+        template = None
+    if not template:
+        raise InvalidRequestException("不支持的 provider_type")
+
+    # 用 refresh_token 换取 access_token
+    refresh_token = payload.refresh_token.strip()
+    token_url = template.oauth.token_url
+    is_json = "anthropic.com" in token_url
+
+    if is_json:
+        body: dict[str, Any] = {
+            "grant_type": "refresh_token",
+            "client_id": template.oauth.client_id,
+            "refresh_token": refresh_token,
+        }
+        headers = {"Content-Type": "application/json", "Accept": "application/json"}
+        data = None
+        json_body = body
+    else:
+        form: dict[str, str] = {
+            "grant_type": "refresh_token",
+            "client_id": template.oauth.client_id,
+            "refresh_token": refresh_token,
+        }
+        if template.oauth.client_secret:
+            form["client_secret"] = template.oauth.client_secret
+        headers = {
+            "Content-Type": "application/x-www-form-urlencoded",
+            "Accept": "application/json",
+        }
+        data = form
+        json_body = None
+
+    proxy_config = getattr(provider, "proxy", None)
+
+    resp = await post_oauth_token(
+        provider_type=provider_type,
+        token_url=token_url,
+        headers=headers,
+        data=data,
+        json_body=json_body,
+        proxy_config=proxy_config,
+        timeout_seconds=30.0,
+    )
+
+    if resp.status_code < 200 or resp.status_code >= 300:
+        error_reason = f"HTTP {resp.status_code}"
+        try:
+            error_body = resp.json()
+            if "error" in error_body:
+                error_reason = str(error_body.get("error_description") or error_body.get("error"))
+        except Exception:
+            error_reason = resp.text[:100] if resp.text else f"HTTP {resp.status_code}"
+        raise InvalidRequestException(f"Refresh Token 验证失败: {error_reason}")
+
+    token = resp.json()
+    access_token = str(token.get("access_token") or "")
+    new_refresh_token = str(token.get("refresh_token") or "") or refresh_token
+    expires_in = token.get("expires_in")
+    expires_at: int | None = None
+    try:
+        if expires_in is not None:
+            expires_at = int(time.time()) + int(expires_in)
+    except Exception:
+        expires_at = None
+
+    if not access_token:
+        raise InvalidRequestException("token refresh 返回缺少 access_token")
+
+    # 构建 auth_config
+    auth_config: dict[str, Any] = {
+        "provider_type": provider_type,
+        "token_type": token.get("token_type"),
+        "refresh_token": new_refresh_token or None,
+        "expires_at": expires_at,
+        "scope": token.get("scope"),
+        "updated_at": int(time.time()),
+    }
+
+    auth_config = await enrich_auth_config(
+        provider_type=provider_type,
+        auth_config=auth_config,
+        token_response=token,
+        access_token=access_token,
+        proxy_config=proxy_config,
+    )
+
+    # 确定账号名称
+    name = (payload.name or "").strip()
+    if not name:
+        name = auth_config.get("email") or f"账号_{int(time.time())}"
+
+    # 从 Provider 的 endpoints 中提取所有 api_format 作为 Key 的支持格式
+    api_formats = [ep.api_format for ep in provider.endpoints if ep.api_format and ep.is_active]
+
+    # 创建 key
+    from src.models.database import ProviderAPIKey as ProviderAPIKeyModel
+
+    new_key = ProviderAPIKeyModel(
+        provider_id=provider_id,
+        name=name,
+        api_key=crypto_service.encrypt(access_token),
+        auth_type="oauth",
+        auth_config=crypto_service.encrypt(json.dumps(auth_config)),
+        api_formats=api_formats,
+        is_active=True,
+    )
+    db.add(new_key)
+    db.commit()
+    db.refresh(new_key)
+
+    return ProviderCompleteOAuthResponse(
+        key_id=str(new_key.id),
+        provider_type=provider_type,
+        expires_at=expires_at,
+        has_refresh_token=bool(new_refresh_token),
+        email=auth_config.get("email"),
+    )

src/api/admin/providers/summary.py

@@ -240,6 +240,19 @@ def _build_provider_summary(db: Session, provider: Provider) -> ProviderWithEndp
     total_models = model_stats.total or 0
     active_models = int(model_stats.active or 0)
 
+    # 活跃模型关联的全局模型 ID 列表
+    global_model_ids = [
+        row[0]
+        for row in db.query(Model.global_model_id)
+        .filter(
+            Model.provider_id == provider.id,
+            Model.is_active == True,
+            Model.global_model_id.isnot(None),
+        )
+        .distinct()
+        .all()
+    ]
+
     api_formats = [e.api_format for e in endpoints]
 
     # 优化: 一次性加载 Provider 的 keys，避免 N+1 查询
@@ -328,6 +341,7 @@ def _build_provider_summary(db: Session, provider: Provider) -> ProviderWithEndp
         active_keys=active_keys,
         total_models=total_models,
         active_models=active_models,
+        global_model_ids=global_model_ids,
         avg_health_score=avg_health_score,
         unhealthy_endpoints=unhealthy_endpoints,
         api_formats=api_formats,

src/api/handlers/base/chat_handler_base.py

@@ -409,6 +409,32 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
         """
         return request_body
 
+    def finalize_provider_request(
+        self,
+        request_body: dict[str, Any],
+        *,
+        mapped_model: str | None,
+        provider_api_format: str | None,
+    ) -> dict[str, Any]:
+        """
+        格式转换完成后、envelope 之前的模型感知后处理钩子 - 子类可覆盖
+
+        用于根据目标模型的特性对请求体做最终调整，例如：
+        - 图像生成模型需要移除不兼容的 tools/system_instruction 并注入 imageConfig
+        - 特定模型需要注入/移除某些字段
+
+        此方法在流式和非流式路径中均会被调用，且 mapped_model 已确定。
+
+        Args:
+            request_body: 已完成格式转换的请求体
+            mapped_model: 映射后的目标模型名
+            provider_api_format: Provider 侧 API 格式标识
+
+        Returns:
+            调整后的请求体
+        """
+        return request_body
+
     def _set_model_after_conversion(
         self,
         request_body: dict[str, Any],
@@ -827,6 +853,13 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
                     target_variant=same_format_variant,
                 )
 
+        # 模型感知的请求后处理（如图像生成模型移除不兼容字段）
+        request_body = self.finalize_provider_request(
+            request_body,
+            mapped_model=mapped_model,
+            provider_api_format=str(provider_api_format) if provider_api_format else None,
+        )
+
         # Force upstream stream/sync mode in request body (best-effort).
         if provider_api_format:
             enforce_stream_mode_for_upstream(
@@ -1440,6 +1473,13 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
                         target_variant=same_format_variant,
                     )
 
+            # 模型感知的请求后处理（如图像生成模型移除不兼容字段）
+            request_body = self.finalize_provider_request(
+                request_body,
+                mapped_model=mapped_model,
+                provider_api_format=str(provider_api_format) if provider_api_format else None,
+            )
+
             # Force upstream stream/sync mode in request body (best-effort).
             if provider_api_format:
                 enforce_stream_mode_for_upstream(

src/api/handlers/base/cli_handler_base.py

@@ -368,6 +368,32 @@ class CliMessageHandlerBase(BaseMessageHandler):
         """
         return request_body
 
+    def finalize_provider_request(
+        self,
+        request_body: dict[str, Any],
+        *,
+        mapped_model: str | None,
+        provider_api_format: str | None,
+    ) -> dict[str, Any]:
+        """
+        格式转换完成后、envelope 之前的模型感知后处理钩子 - 子类可覆盖
+
+        用于根据目标模型的特性对请求体做最终调整，例如：
+        - 图像生成模型需要移除不兼容的 tools/system_instruction 并注入 imageConfig
+        - 特定模型需要注入/移除某些字段
+
+        此方法在流式和非流式路径中均会被调用，且 mapped_model 已确定。
+
+        Args:
+            request_body: 已完成格式转换的请求体
+            mapped_model: 映射后的目标模型名
+            provider_api_format: Provider 侧 API 格式标识
+
+        Returns:
+            调整后的请求体
+        """
+        return request_body
+
     @staticmethod
     def _get_format_metadata(format_id: str) -> "EndpointDefinition | None":
         """获取 endpoint 元数据（解析失败返回 None）"""
@@ -801,6 +827,13 @@ class CliMessageHandlerBase(BaseMessageHandler):
                     target_variant=target_variant,
                 )
 
+        # 模型感知的请求后处理（如图像生成模型移除不兼容字段）
+        request_body = self.finalize_provider_request(
+            request_body,
+            mapped_model=mapped_model,
+            provider_api_format=provider_api_format,
+        )
+
         # Force upstream stream/sync mode in request body (best-effort).
         if provider_api_format:
             enforce_stream_mode_for_upstream(
@@ -2812,6 +2845,13 @@ class CliMessageHandlerBase(BaseMessageHandler):
                         target_variant=target_variant,
                     )
 
+            # 模型感知的请求后处理（如图像生成模型移除不兼容字段）
+            request_body = self.finalize_provider_request(
+                request_body,
+                mapped_model=mapped_model,
+                provider_api_format=provider_api_format,
+            )
+
             # Force upstream stream/sync mode in request body (best-effort).
             if provider_api_format:
                 enforce_stream_mode_for_upstream(

src/api/handlers/base/stream_processor.py

@@ -708,6 +708,7 @@ class StreamProcessor:
                             f"[{self.request_id}] 处理剩余缓冲区失败: {e}, bytes={buffer[:50]!r}"
                         )
                         line = ""
+                    buffer = b""  # 标记已消费，避免 finally 中重复处理
                     if line:
                         # 需要格式转换时，跳过记录原始数据
                         _process_line_with_perf(line, skip_record=True)
@@ -782,8 +783,26 @@ class StreamProcessor:
                     logger.warning(
                         f"[{self.request_id}] 处理剩余缓冲区失败: {e}, bytes={buffer[:50]!r}"
                     )
+                buffer = b""  # 标记已消费，避免下方重复处理
 
-            # 处理剩余事件
+            # flush 残留的字节 buffer（异常中断时 buffer 可能仍有未解析的数据，
+            # 如包含 usage 的 message_delta/response.completed 事件）
+            # 正常结束时 buffer 已在上方被消费为空，此处为 no-op
+            if buffer:
+                try:
+                    remaining = decoder.decode(buffer, True)
+                    for line in remaining.split("\n"):
+                        stripped = line.rstrip("\r\n")
+                        if stripped:
+                            events = sse_parser.feed_line(stripped)
+                            for event in events:
+                                self.handle_sse_event(
+                                    ctx, event.get("event"), event.get("data") or ""
+                                )
+                except Exception:
+                    pass  # best-effort: 不应因 flush 失败影响后续流程
+
+            # flush SSE parser 内部累积的未完成事件
             for event in sse_parser.flush():
                 self.handle_sse_event(ctx, event.get("event"), event.get("data") or "")
 

src/api/handlers/base/stream_telemetry.py

@@ -8,6 +8,7 @@
 """
 
 import asyncio
+import json
 import time
 from typing import Any
 
@@ -104,6 +105,19 @@ class StreamTelemetryRecorder:
                 if writer is None:
                     return
                 actual_request_body = ctx.provider_request_body or original_request_body
+
+                # 兜底估算：流未正常完成且 token 均为 0 时，从请求体粗略估算
+                # 覆盖 Chat Handler 路径（CLI Handler 在更早的位置已做估算，
+                # 若已估算过则 token > 0，此处条件不会触发）
+                if (
+                    ctx.is_success()
+                    and not ctx.has_completion
+                    and ctx.data_count > 0
+                    and ctx.input_tokens == 0
+                    and ctx.output_tokens == 0
+                ):
+                    self._estimate_tokens_for_incomplete_stream(ctx, actual_request_body)
+
                 should_log_body = SystemConfigService.should_log_body(bg_db)
                 include_bodies = (
                     writer.include_bodies
@@ -536,6 +550,59 @@ class StreamTelemetryRecorder:
             return "cancelled"
         return "failed"
 
+    @staticmethod
+    def _estimate_tokens_for_incomplete_stream(
+        ctx: StreamContext,
+        request_body: dict[str, Any],
+    ) -> None:
+        """
+        流未正常完成（无 response.completed）且 token 均为 0 时的兜底估算。
+
+        从已收集的输出文本和请求体粗略估算 token 数，确保 usage 记录不为 0。
+        估算采用 ~4 字符/token 的保守比例。
+        """
+        # 输出 tokens：从已收集的文本估算
+        collected = ctx.collected_text
+        if collected:
+            ctx.output_tokens = max(1, len(collected) // 4)
+
+        # 输入 tokens：从请求体文本内容估算
+        try:
+            total_input_len = 0
+            instructions = request_body.get("instructions")
+            if isinstance(instructions, str):
+                total_input_len += len(instructions)
+            # OpenAI Responses API 使用 input 字段；Claude 使用 messages
+            input_items = request_body.get("input") or request_body.get("messages") or []
+            if isinstance(input_items, list):
+                for item in input_items:
+                    if isinstance(item, str):
+                        total_input_len += len(item)
+                    elif isinstance(item, dict):
+                        content = item.get("content", "")
+                        if isinstance(content, str):
+                            total_input_len += len(content)
+                        elif isinstance(content, list):
+                            for block in content:
+                                if isinstance(block, dict):
+                                    text = block.get("text", "")
+                                    if isinstance(text, str):
+                                        total_input_len += len(text)
+            if total_input_len > 0:
+                ctx.input_tokens = max(1, total_input_len // 4)
+            else:
+                # fallback: 整个请求体 JSON 大小
+                body_str = json.dumps(request_body, ensure_ascii=False)
+                ctx.input_tokens = max(1, len(body_str) // 4)
+        except Exception:
+            pass
+
+        if ctx.input_tokens > 0 or ctx.output_tokens > 0:
+            logger.warning(
+                f"[{ctx.request_id}] 流未正常完成 (has_completion=False, data_count={ctx.data_count}), "
+                f"使用估算 tokens: in={ctx.input_tokens}, out={ctx.output_tokens}"
+            )
+
     def _build_db_writer(self, bg_db: Session) -> DbTelemetryWriter | None:
         user = bg_db.query(User).filter(User.id == self.user_id).first()
         api_key_obj = bg_db.query(ApiKey).filter(ApiKey.id == self.api_key_id).first()

src/api/handlers/gemini/handler.py

@@ -157,6 +157,22 @@ class GeminiChatHandler(ChatHandlerBase):
             "cache_read_input_tokens": usage.get("cached_tokens", 0),
         }
 
+    def finalize_provider_request(
+        self,
+        request_body: dict[str, Any],
+        *,
+        mapped_model: str | None,
+        provider_api_format: str | None,  # noqa: ARG002
+    ) -> dict[str, Any]:
+        from src.api.handlers.gemini.image_gen import (
+            adapt_request_for_image_gen,
+            is_image_gen_model,
+        )
+
+        if not is_image_gen_model(mapped_model):
+            return request_body
+        return adapt_request_for_image_gen(request_body)
+
     def _normalize_response(self, response: dict) -> dict:
         """
         规范化 Gemini 响应

src/api/handlers/gemini/image_gen.py

@@ -0,0 +1,45 @@
+"""
+Gemini 图像生成模型请求适配
+
+- 图像生成模型不支持 tools / system_instruction，需要移除
+- responseModalities / responseMimeType 与 imageConfig 冲突，需要移除
+"""
+
+from typing import Any
+
+
+def is_image_gen_model(model: str | None) -> bool:
+    """判断是否为图像生成模型（模式匹配，覆盖 gemini-*-image / imagen-* 系列）"""
+    if not model:
+        return False
+    m = model.lower()
+    return "image" in m and ("gemini" in m or "imagen" in m)
+
+
+def adapt_request_for_image_gen(body: dict[str, Any]) -> dict[str, Any]:
+    """为图像生成模型清理不兼容字段"""
+    # 移除图像生成不支持的顶层字段
+    for key in ("tools", "tool_config", "toolConfig", "system_instruction", "systemInstruction"):
+        if key in body:
+            body.pop(key)
+
+    # 处理 generationConfig
+    gc_key = "generationConfig" if "generationConfig" in body else "generation_config"
+    gc = body.get(gc_key)
+    if not isinstance(gc, dict):
+        gc = {}
+        body[gc_key] = gc
+
+    # 移除与图像生成冲突的字段
+    for key in (
+        "responseMimeType",
+        "response_mime_type",
+        "responseModalities",
+        "response_modalities",
+    ):
+        gc.pop(key, None)
+
+    # 设置输出模态
+    gc["responseModalities"] = ["TEXT", "IMAGE"]
+
+    return body

src/api/handlers/gemini_cli/handler.py

@@ -78,6 +78,22 @@ class GeminiCliMessageHandler(CliMessageHandlerBase):
         result.pop("model", None)
         return result
 
+    def finalize_provider_request(
+        self,
+        request_body: dict[str, Any],
+        *,
+        mapped_model: str | None,
+        provider_api_format: str | None,  # noqa: ARG002
+    ) -> dict[str, Any]:
+        from src.api.handlers.gemini.image_gen import (
+            adapt_request_for_image_gen,
+            is_image_gen_model,
+        )
+
+        if not is_image_gen_model(mapped_model):
+            return request_body
+        return adapt_request_for_image_gen(request_body)
+
     def get_model_for_url(
         self,
         request_body: dict[str, Any],

src/models/endpoint_models.py

@@ -759,6 +759,7 @@ class ProviderWithEndpointsSummary(BaseModel):
     # Model 统计
     total_models: int = Field(default=0, description="总模型数量")
     active_models: int = Field(default=0, description="活跃模型数量")
+    global_model_ids: list[str] = Field(default=[], description="活跃模型关联的全局模型 ID 列表")
 
     # API 格式列表
     api_formats: list[str] = Field(default=[], description="支持的 API 格式列表")

src/services/model/fetch_scheduler.py

@@ -30,6 +30,7 @@ from src.models.database import Provider, ProviderAPIKey
 from src.services.model.upstream_fetcher import (
     UpstreamModelsFetchContext,
     fetch_models_for_key,
+    merge_upstream_metadata,
 )
 from src.services.provider.oauth_token import resolve_oauth_access_token
 from src.services.system.scheduler import get_scheduler
@@ -491,13 +492,9 @@ class ModelFetchScheduler:
 
             # 最佳努力：保存上游元数据（如 Antigravity 配额信息）
             if upstream_metadata and isinstance(upstream_metadata, dict):
-                # NOTE: upstream_metadata is a plain JSON column (not MutableDict),
-                # so in-place mutation won't be persisted reliably. Always assign
-                # a new dict object to mark the column as dirty.
-                current = key.upstream_metadata
-                merged: dict[str, Any] = dict(current) if isinstance(current, dict) else {}
-                merged.update(upstream_metadata)
-                key.upstream_metadata = merged
+                key.upstream_metadata = merge_upstream_metadata(
+                    key.upstream_metadata, upstream_metadata
+                )
 
             # 去重获取模型 ID 列表
             fetched_model_ids: set[str] = set()

src/services/model/upstream_fetcher.py

@@ -82,6 +82,50 @@ async def fetch_models_for_key(
     return await fetcher(ctx, timeout_seconds)
 
 
+def merge_upstream_metadata(
+    current: dict[str, Any] | None,
+    incoming: dict[str, Any],
+) -> dict[str, Any]:
+    """合并上游元数据，对 quota_by_model 做模型级深度合并。
+
+    上游 API 在配额耗尽后可能不再返回该模型的 quotaInfo，因此需要：
+    1. 保留旧数据中已有的 reset_time（当新数据缺少时）
+    2. 保留旧数据中存在但新数据中缺失的模型条目（标记为 100% 已用）
+    """
+    merged: dict[str, Any] = dict(current) if isinstance(current, dict) else {}
+    for ns_key, ns_val in incoming.items():
+        old_ns = merged.get(ns_key)
+        if (
+            isinstance(ns_val, dict)
+            and isinstance(old_ns, dict)
+            and "quota_by_model" in ns_val
+            and "quota_by_model" in old_ns
+        ):
+            old_qbm = old_ns["quota_by_model"]
+            new_qbm = ns_val["quota_by_model"]
+            if isinstance(old_qbm, dict) and isinstance(new_qbm, dict):
+                # 保留新数据中已有模型的旧 reset_time
+                for model_id, new_info in new_qbm.items():
+                    if not isinstance(new_info, dict):
+                        continue
+                    old_info = old_qbm.get(model_id)
+                    if (
+                        isinstance(old_info, dict)
+                        and "reset_time" in old_info
+                        and "reset_time" not in new_info
+                    ):
+                        new_info["reset_time"] = old_info["reset_time"]
+                # 保留新数据中缺失但旧数据中存在的模型（配额耗尽后上游可能不返回）
+                for model_id, old_info in old_qbm.items():
+                    if model_id not in new_qbm and isinstance(old_info, dict):
+                        exhausted = dict(old_info)
+                        exhausted["remaining_fraction"] = 0.0
+                        exhausted["used_percent"] = 100.0
+                        new_qbm[model_id] = exhausted
+        merged[ns_key] = ns_val
+    return merged
+
+
 # Provider-specific fetchers are registered by plugin.register_all()
 # (called from envelope.py bootstrap)
 

src/services/provider/adapters/antigravity/envelope.py

@@ -345,31 +345,45 @@ def wrap_v1internal_request(
     1. 移除 model（移到顶层）
     2. 移除 safetySettings（v1internal 不支持）
     3. 深度清理 [undefined] 字符串
-    4. Claude model tool ID 注入
-    5. Thinking budget 处理（自动注入 + Auto Cap）
-    6. 工具声明清洗（schema 清理 + 字段重命名）
-    7. System Instruction 注入
+    4. Claude model tool ID 注入（图像生成模型跳过）
+    5. Thinking budget 处理
+    6. 工具声明清洗（图像生成模型跳过）
+    7. System Instruction 注入（图像生成模型跳过）
     8. 注入 sessionId（对齐 CLIProxyAPI）
     9. 构建 v1internal 信封
     """
+    from src.api.handlers.gemini.image_gen import is_image_gen_model
+
     inner_request = dict(gemini_request)
     inner_request.pop("model", None)
     inner_request.pop("safetySettings", None)
 
+    is_image_gen = is_image_gen_model(model)
+
     # 1. 深度清理 [undefined]
     _deep_clean_undefined(inner_request)
 
-    # 2. Claude tool ID 注入
-    _inject_claude_tool_ids_request(inner_request, model)
+    if not is_image_gen:
+        # 2. Claude tool ID 注入
+        _inject_claude_tool_ids_request(inner_request, model)
 
     # 3. Thinking budget 处理
     _process_thinking_budget(inner_request, model)
 
-    # 4. 工具声明清洗
-    _clean_tool_declarations(inner_request)
+    if not is_image_gen:
+        # 4. 工具声明清洗
+        _clean_tool_declarations(inner_request)
 
-    # 5. System Instruction 注入
-    _inject_system_instruction(inner_request)
+        # 5. System Instruction 注入
+        _inject_system_instruction(inner_request)
+    else:
+        # 图像生成模型：对齐 AM wrapper.rs，移除不兼容字段
+        inner_request.pop("tools", None)
+        inner_request.pop("toolConfig", None)
+        inner_request.pop("tool_config", None)
+        inner_request.pop("systemInstruction", None)
+        inner_request.pop("system_instruction", None)
+        request_type = "image_gen"
 
     # 6. 注入 sessionId（对齐 CLIProxyAPI/sub2api）
     if "sessionId" not in inner_request:

src/services/provider/adapters/antigravity/plugin.py

@@ -255,6 +255,11 @@ async def fetch_models_antigravity(
 
         quota_info = model_data.get("quotaInfo")
         if not isinstance(quota_info, dict):
+            # 没有 quotaInfo 视为配额耗尽
+            quota_by_model[model_id] = {
+                "remaining_fraction": 0.0,
+                "used_percent": 100.0,
+            }
             continue
 
         remaining = quota_info.get("remainingFraction")
@@ -268,6 +273,14 @@ async def fetch_models_antigravity(
             remaining_fraction = None
 
         if remaining_fraction is None:
+            # remainingFraction 缺失视为配额耗尽
+            payload: dict[str, Any] = {
+                "remaining_fraction": 0.0,
+                "used_percent": 100.0,
+            }
+            if isinstance(reset_time, str) and reset_time.strip():
+                payload["reset_time"] = reset_time.strip()
+            quota_by_model[model_id] = payload
             continue
 
         used_percent = (1.0 - remaining_fraction) * 100.0