feat: 多项功能优化与问题修复

- 添加高频轮询端点日志抑制，减少 debug 日志噪音
- Gemini 格式转换支持 responseModalities、thinkingConfig 透传和图片生成输出
- OpenAI 格式转换支持流式图片内容块，区分 URL 引用和 base64 内嵌图片
- Provider 余额缓存认证失败时使用短 TTL，避免前端无限加载中
- Usage 服务支持更新 pending/streaming 记录，处理重复 request_id 冲突
- Usage 超时检测增强，避免错误标记已完成请求为超时

src/api/base/context.py

@@ -45,6 +45,9 @@ class ApiRequestContext:
     extra: Dict[str, Any] = field(default_factory=dict)
     audit_metadata: Dict[str, Any] = field(default_factory=dict)
 
+    # 高频轮询端点日志抑制标志
+    quiet_logging: bool = False
+
     def ensure_json_body(self) -> Dict[str, Any]:
         """确保请求体已解析为JSON并返回。"""
         if self.json_body is not None:

src/api/base/pipeline.py

@@ -22,6 +22,14 @@ if TYPE_CHECKING:
 from .adapter import ApiAdapter, ApiMode
 from .context import ApiRequestContext
 
+# 高频轮询端点，抑制其 debug 日志以减少噪音
+QUIET_POLLING_PATHS: set[str] = {
+    "/api/admin/usage/active",
+    "/api/admin/usage/records",
+    "/api/admin/usage/stats",
+    "/api/admin/usage/aggregation/stats",
+    "/api/admin/health/status",
+}
 
 
 class ApiRequestPipeline:
@@ -47,9 +55,12 @@ class ApiRequestPipeline:
         api_format_hint: Optional[str] = None,
         path_params: Optional[dict[str, Any]] = None,
     ):
-        logger.debug(f"[Pipeline] START | path={http_request.url.path}")
-        logger.debug(f"[Pipeline] Running with mode={mode}, adapter={adapter.__class__.__name__}, "
-            f"adapter.mode={adapter.mode}, path={http_request.url.path}")
+        # 高频轮询端点抑制 debug 日志
+        is_quiet = http_request.url.path in QUIET_POLLING_PATHS
+        if not is_quiet:
+            logger.debug(f"[Pipeline] START | path={http_request.url.path}")
+            logger.debug(f"[Pipeline] Running with mode={mode}, adapter={adapter.__class__.__name__}, "
+                f"adapter.mode={adapter.mode}, path={http_request.url.path}")
         if mode == ApiMode.ADMIN:
             user, management_token = await self._authenticate_admin(http_request, db)
             api_key = None
@@ -64,10 +75,12 @@ class ApiRequestPipeline:
             user, management_token = await self._authenticate_management(http_request, db)
             api_key = None
         else:
-            logger.debug("[Pipeline] 调用 _authenticate_client")
-            user, api_key = self._authenticate_client(http_request, db, adapter)
+            if not is_quiet:
+                logger.debug("[Pipeline] 调用 _authenticate_client")
+            user, api_key = self._authenticate_client(http_request, db, adapter, quiet=is_quiet)
             management_token = None
-            logger.debug(f"[Pipeline] 认证完成 | user={user.username if user else None}")
+            if not is_quiet:
+                logger.debug(f"[Pipeline] 认证完成 | user={user.username if user else None}")
 
         raw_body = None
         if http_request.method in {"POST", "PUT", "PATCH"}:
@@ -78,7 +91,8 @@ class ApiRequestPipeline:
                 raw_body = await asyncio.wait_for(
                     http_request.body(), timeout=config.request_body_timeout
                 )
-                logger.debug(f"[Pipeline] Raw body读取完成 | size={len(raw_body) if raw_body is not None else 0} bytes")
+                if not is_quiet:
+                    logger.debug(f"[Pipeline] Raw body读取完成 | size={len(raw_body) if raw_body is not None else 0} bytes")
             except asyncio.TimeoutError:
                 timeout_sec = int(config.request_body_timeout)
                 logger.error(f"读取请求体超时({timeout_sec}s),可能客户端未发送完整请求体")
@@ -87,7 +101,8 @@ class ApiRequestPipeline:
                     detail=f"Request timeout: body not received within {timeout_sec} seconds",
                 )
         else:
-            logger.debug(f"[Pipeline] 非写请求跳过读取Body | method={http_request.method}")
+            if not is_quiet:
+                logger.debug(f"[Pipeline] 非写请求跳过读取Body | method={http_request.method}")
 
         context = ApiRequestContext.build(
             request=http_request,
@@ -102,14 +117,17 @@ class ApiRequestPipeline:
         # 存储 management_token 到 context（用于权限检查）
         if management_token:
             context.management_token = management_token
-        logger.debug(f"[Pipeline] Context构建完成 | adapter={adapter.name} | request_id={context.request_id}")
+        # 存储 quiet 标志到 context，用于审计日志判断
+        context.quiet_logging = is_quiet
+        if not is_quiet:
+            logger.debug(f"[Pipeline] Context构建完成 | adapter={adapter.name} | request_id={context.request_id}")
 
         if mode != ApiMode.ADMIN and user:
             context.quota_remaining = self._calculate_quota_remaining(user)
 
-        logger.debug(f"[Pipeline] Adapter={adapter.name} | RequestID={context.request_id}")
-
-        logger.debug(f"[Pipeline] Calling authorize on {adapter.__class__.__name__}, user={context.user}")
+        if not is_quiet:
+            logger.debug(f"[Pipeline] Adapter={adapter.name} | RequestID={context.request_id}")
+            logger.debug(f"[Pipeline] Calling authorize on {adapter.__class__.__name__}, user={context.user}")
         # authorize 可能是异步的，需要检查并 await
         authorize_result = adapter.authorize(context)
         if hasattr(authorize_result, "__await__"):
@@ -145,18 +163,22 @@ class ApiRequestPipeline:
     # --------------------------------------------------------------------- #
 
     def _authenticate_client(
-        self, request: Request, db: Session, adapter: ApiAdapter
+        self, request: Request, db: Session, adapter: ApiAdapter, *, quiet: bool = False
     ) -> Tuple[User, ApiKey]:
-        logger.debug("[Pipeline._authenticate_client] 开始")
+        if not quiet:
+            logger.debug("[Pipeline._authenticate_client] 开始")
         # 使用 adapter 的 extract_api_key 方法，支持不同 API 格式的认证头
         client_api_key = adapter.extract_api_key(request)
-        logger.debug(f"[Pipeline._authenticate_client] 提取API密钥完成 | key_prefix={client_api_key[:8] if client_api_key else None}...")
+        if not quiet:
+            logger.debug(f"[Pipeline._authenticate_client] 提取API密钥完成 | key_prefix={client_api_key[:8] if client_api_key else None}...")
         if not client_api_key:
             raise HTTPException(status_code=401, detail="请提供API密钥")
 
-        logger.debug("[Pipeline._authenticate_client] 调用 auth_service.authenticate_api_key")
+        if not quiet:
+            logger.debug("[Pipeline._authenticate_client] 调用 auth_service.authenticate_api_key")
         auth_result = self.auth_service.authenticate_api_key(db, client_api_key)
-        logger.debug(f"[Pipeline._authenticate_client] 认证结果 | result={bool(auth_result)}")
+        if not quiet:
+            logger.debug(f"[Pipeline._authenticate_client] 认证结果 | result={bool(auth_result)}")
         if not auth_result:
             raise HTTPException(status_code=401, detail="无效的API密钥")
 
@@ -435,6 +457,8 @@ class ApiRequestPipeline:
             "request_content_type": request.headers.get("content-type"),
             "quota_remaining": context.quota_remaining,
             "success": success,
+            # 传递 quiet_logging 标志给审计服务，用于抑制高频轮询日志
+            "quiet_logging": getattr(context, "quiet_logging", False),
         }
         if status_code is not None:
             metadata["status_code"] = status_code

src/api/handlers/base/cli_handler_base.py

@@ -1851,6 +1851,9 @@ class CliMessageHandlerBase(BaseMessageHandler):
                 api_key = bg_db.query(ApiKeyModel).filter(ApiKeyModel.id == ctx.api_key_id).first()
 
                 if not user or not api_key:
+                    logger.warning(
+                        f"[{ctx.request_id}] 无法记录统计: user={user is not None}, api_key={api_key is not None}"
+                    )
                     return
 
                 bg_telemetry = MessageTelemetry(
@@ -1925,6 +1928,11 @@ class CliMessageHandlerBase(BaseMessageHandler):
                         }
                     )
 
+                    logger.debug(
+                        f"[{ctx.request_id}] 开始记录 Usage: "
+                        f"provider={ctx.provider_name}, model={ctx.model}, "
+                        f"in={actual_input_tokens}, out={ctx.output_tokens}"
+                    )
                     total_cost = await bg_telemetry.record_success(
                         provider=ctx.provider_name,
                         model=ctx.model,
@@ -1955,7 +1963,9 @@ class CliMessageHandlerBase(BaseMessageHandler):
                         # Provider 响应元数据（如 Gemini 的 modelVersion）
                         response_metadata=ctx.response_metadata if ctx.response_metadata else None,
                     )
-                    logger.debug(f"{self.FORMAT_ID} 流式响应完成")
+                    logger.debug(
+                        f"[{ctx.request_id}] Usage 记录完成: cost=${total_cost:.6f}"
+                    )
                     # 简洁的请求完成摘要（两行格式）
                     line1 = f"[OK] {self.request_id[:8]} | {ctx.model} | {ctx.provider_name}"
                     if ctx.first_byte_time_ms:

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

@@ -151,6 +151,22 @@ class GeminiNormalizer(FormatNormalizer):
             else request.get("toolConfig")
         )
 
+        # 构建 extra，保留原始 gemini 字段
+        extra: Dict[str, Any] = {"gemini": self._extract_extra(request, {"contents"})}
+
+        # 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
+        if isinstance(generation_config, dict):
+            response_modalities = generation_config.get("response_modalities")
+            thinking_config = generation_config.get("thinking_config")
+            if response_modalities or thinking_config:
+                google_extra: Dict[str, Any] = {}
+                if response_modalities:
+                    google_extra["response_modalities"] = response_modalities
+                if thinking_config:
+                    google_extra["thinking_config"] = thinking_config
+                extra["google"] = google_extra
+
         internal = InternalRequest(
             model=model,
             messages=messages,
@@ -164,7 +180,7 @@ class GeminiNormalizer(FormatNormalizer):
             stream=bool(request.get("stream") or False),
             tools=tools,
             tool_choice=tool_choice,
-            extra={"gemini": self._extract_extra(request, {"contents"})},
+            extra=extra,
         )
 
         if dropped:
@@ -208,6 +224,47 @@ class GeminiNormalizer(FormatNormalizer):
         if internal.stop_sequences:
             generation_config["stop_sequences"] = list(internal.stop_sequences)
 
+        # 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
+            thinking_config = google_extra.get("thinking_config")
+            if isinstance(thinking_config, dict):
+                # snake_case -> camelCase 转换
+                gemini_thinking: Dict[str, Any] = {}
+                if "thinking_budget" in thinking_config:
+                    gemini_thinking["thinkingBudget"] = thinking_config["thinking_budget"]
+                if "include_thoughts" in thinking_config:
+                    gemini_thinking["includeThoughts"] = thinking_config["include_thoughts"]
+                # 保留其他可能的字段
+                for k, v in thinking_config.items():
+                    if k not in ("thinking_budget", "include_thoughts"):
+                        gemini_thinking[k] = v
+                if gemini_thinking:
+                    generation_config["thinkingConfig"] = gemini_thinking
+
+            # 处理 response_modalities -> responseModalities
+            response_modalities = google_extra.get("response_modalities")
+            if response_modalities:
+                generation_config["responseModalities"] = response_modalities
+
+        # 从 internal.extra["gemini"] 读取原生 Gemini 配置（Gemini -> Gemini 场景）
+        gemini_extra = internal.extra.get("gemini", {})
+        if isinstance(gemini_extra, dict):
+            # 保留原生 Gemini generationConfig 中的额外字段
+            orig_gc = gemini_extra.get("generation_config") or gemini_extra.get("generationConfig")
+            if isinstance(orig_gc, dict):
+                # responseModalities
+                if "responseModalities" in orig_gc and "responseModalities" not in generation_config:
+                    generation_config["responseModalities"] = orig_gc["responseModalities"]
+                if "response_modalities" in orig_gc and "responseModalities" not in generation_config:
+                    generation_config["responseModalities"] = orig_gc["response_modalities"]
+                # thinkingConfig
+                if "thinkingConfig" in orig_gc and "thinkingConfig" not in generation_config:
+                    generation_config["thinkingConfig"] = orig_gc["thinkingConfig"]
+                if "thinking_config" in orig_gc and "thinkingConfig" not in generation_config:
+                    generation_config["thinkingConfig"] = orig_gc["thinking_config"]
+
         contents: List[Dict[str, Any]] = []
         for msg in internal.messages:
             contents.append(self._internal_message_to_content(msg))
@@ -431,6 +488,35 @@ class GeminiNormalizer(FormatNormalizer):
                     events.append(ContentBlockStopEvent(block_index=block_index))
                     continue
 
+                # inlineData（图片生成等多模态输出）
+                inline_data = part.get("inlineData")
+                if inline_data is None:
+                    inline_data = part.get("inline_data")
+
+                if isinstance(inline_data, dict):
+                    mime_type = str(inline_data.get("mimeType") or inline_data.get("mime_type") or "").strip()
+                    data = str(inline_data.get("data") or "").strip()
+
+                    # 确保 mime_type 和 data 都非空
+                    if mime_type and data and len(data) > 10:  # base64 图片数据至少几十个字符
+                        block_index = int(ss.get("next_block_index") or 1)
+                        ss["next_block_index"] = block_index + 1
+
+                        # 使用 ContentBlockStartEvent 传递图片数据
+                        # 图片数据存储在 extra 中，供 target normalizer 处理
+                        events.append(
+                            ContentBlockStartEvent(
+                                block_index=block_index,
+                                block_type=ContentType.IMAGE,
+                                extra={
+                                    "image_data": data,
+                                    "image_media_type": mime_type,
+                                },
+                            )
+                        )
+                        events.append(ContentBlockStopEvent(block_index=block_index))
+                    continue
+
         finish_reason = candidate0.get("finishReason")
         if finish_reason is not None:
             stop_reason = self._FINISH_REASON_TO_STOP.get(str(finish_reason), StopReason.UNKNOWN)
@@ -488,6 +574,25 @@ class GeminiNormalizer(FormatNormalizer):
             }
             return out
 
+        # 图片内容块（来自其他格式的图像生成输出）
+        if isinstance(event, ContentBlockStartEvent) and event.block_type == ContentType.IMAGE:
+            image_data = event.extra.get("image_data")
+            image_media_type = event.extra.get("image_media_type")
+            if image_data and image_media_type:
+                out.append(
+                    base_chunk(
+                        [
+                            {
+                                "inlineData": {
+                                    "mimeType": image_media_type,
+                                    "data": image_data,
+                                }
+                            }
+                        ]
+                    )
+                )
+            return out
+
         if isinstance(event, ToolCallDeltaEvent):
             tool_blocks = ss.get("tool_blocks")
             if isinstance(tool_blocks, dict):
@@ -784,6 +889,17 @@ class GeminiNormalizer(FormatNormalizer):
         normalized["top_p"] = pick("top_p", "topP")
         normalized["top_k"] = pick("top_k", "topK")
         normalized["stop_sequences"] = pick("stop_sequences", "stopSequences")
+
+        # 保留 responseModalities（图像生成等多模态输出必需）
+        response_modalities = pick("response_modalities", "responseModalities")
+        if response_modalities:
+            normalized["response_modalities"] = response_modalities
+
+        # 保留 thinkingConfig（思考模式配置）
+        thinking_config = pick("thinking_config", "thinkingConfig")
+        if thinking_config:
+            normalized["thinking_config"] = thinking_config
+
         return {k: v for k, v in normalized.items() if v is not None}
 
     def _gemini_tools_to_internal(self, tools: Any) -> Optional[List[ToolDefinition]]:

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

@@ -13,6 +13,7 @@ import json
 import time
 from typing import Any, Dict, List, Optional, Tuple, Union
 
+from src.core.logger import logger
 from src.core.api_format.conversion.field_mappings import (
     ERROR_TYPE_MAPPINGS,
     RETRYABLE_ERROR_TYPES,
@@ -144,6 +145,16 @@ class OpenAINormalizer(FormatNormalizer):
             mct if mct is not None else request.get("max_tokens")
         )
 
+        # 构建 extra，保留未识别字段
+        extra: Dict[str, Any] = {"openai": self._extract_extra(request, {"messages"})}
+
+        # 处理 extra_body.google (用于 Gemini 特定功能透传，如 thinkingConfig, responseModalities)
+        extra_body = request.get("extra_body")
+        if isinstance(extra_body, dict):
+            google_extra = extra_body.get("google")
+            if isinstance(google_extra, dict) and google_extra:
+                extra["google"] = google_extra
+
         internal = InternalRequest(
             model=model,
             messages=messages,
@@ -156,7 +167,7 @@ class OpenAINormalizer(FormatNormalizer):
             stream=bool(request.get("stream") or False),
             tools=tools,
             tool_choice=tool_choice,
-            extra={"openai": self._extract_extra(request, {"messages"})},
+            extra=extra,
         )
 
         if dropped:
@@ -470,6 +481,50 @@ class OpenAINormalizer(FormatNormalizer):
             )
             return out
 
+        # 图片内容块（来自 Gemini 图像生成等多模态输出）
+        if isinstance(event, ContentBlockStartEvent) and event.block_type == ContentType.IMAGE:
+            image_data = event.extra.get("image_data")
+            image_media_type = event.extra.get("image_media_type")
+            # 确保图片数据有效（base64 数据至少有一定长度）
+            if image_data and image_media_type and isinstance(image_data, str) and len(image_data) > 10:
+                # 构造 data URL 格式的图片
+                data_url = f"data:{image_media_type};base64,{image_data}"
+                # 存储图片数据，在 ContentBlockStopEvent 时输出
+                image_blocks = ss.get("image_blocks")
+                if not isinstance(image_blocks, dict):
+                    image_blocks = {}
+                    ss["image_blocks"] = image_blocks
+                image_blocks[int(event.block_index)] = {
+                    "url": data_url,
+                    "media_type": image_media_type,
+                }
+            return out
+
+        # 图片内容块结束时输出
+        if isinstance(event, ContentBlockStopEvent):
+            image_blocks = ss.get("image_blocks")
+            if isinstance(image_blocks, dict):
+                entry = image_blocks.get(int(event.block_index))
+                if isinstance(entry, dict):
+                    url = entry.get("url")
+                    # 确保 URL 有效（data URL 至少包含 "data:" 前缀 + 一些数据）
+                    if url and isinstance(url, str) and len(url) > 20:
+                        # OpenAI 流式响应中 delta.content 必须是字符串
+                        # 使用 markdown 图片格式，兼容各种客户端渲染
+                        # 注意：base64 data URL 可能很长（几百KB~几MB），客户端需支持长内容
+                        # 典型图片大小：100KB 原图 ≈ 130KB base64，1MB 原图 ≈ 1.3MB base64
+                        if len(url) > 1_000_000:  # > 1MB
+                            logger.warning(
+                                f"Large image in stream response: {len(url)} bytes, "
+                                "client may have rendering issues"
+                            )
+                        out.append(base_chunk({"content": f"![image]({url})"}))
+                    # 清理已处理的图片
+                    del image_blocks[int(event.block_index)]
+                    return out
+            # 其他 ContentBlockStopEvent 不处理
+            return out
+
         if isinstance(event, ToolCallDeltaEvent):
             tool_index = self._ensure_tool_call_index(ss, event.tool_id)
             out.append(
@@ -854,20 +909,30 @@ class OpenAINormalizer(FormatNormalizer):
                     text_parts.append(b.text)
                 continue
             if isinstance(b, ImageBlock):
-                url = b.url
-                if not url and b.data and b.media_type:
-                    url = f"data:{b.media_type};base64,{b.data}"
-                if url:
-                    parts.append({"type": "image_url", "image_url": {"url": url}})
+                # 区分两种图片来源：
+                # 1. URL 引用（OpenAI 原生格式）-> multipart content
+                # 2. base64 内嵌数据（格式转换来的）-> markdown 格式
+                if b.url and not b.data:
+                    # OpenAI 原生格式：URL 引用的图片，使用 multipart content
+                    parts.append({"type": "image_url", "image_url": {"url": b.url}})
+                elif b.data and b.media_type:
+                    # 格式转换来的图片（base64 内嵌），使用 markdown 格式
+                    data_url = f"data:{b.media_type};base64,{b.data}"
+                    text_parts.append(f"![image]({data_url})")
+                elif b.url:
+                    # 有 URL 也有 data，优先使用 URL
+                    parts.append({"type": "image_url", "image_url": {"url": b.url}})
                 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
         if parts:
             if text_parts:
                 parts = [{"type": "text", "text": "\n".join(text_parts)}] + parts
             return parts
 
+        # 纯文本或包含 markdown 图片
         if text_parts:
             return "\n".join(text_parts)
 

src/services/provider_ops/service.py

@@ -33,6 +33,8 @@ from src.services.provider_ops.types import (
 
 # 余额缓存 TTL（24 小时）
 BALANCE_CACHE_TTL = 86400
+# 认证失败缓存 TTL（60 秒，避免频繁重试但允许用户修正后快速重试）
+AUTH_FAILED_CACHE_TTL = 60
 
 
 def _get_batch_balance_concurrency() -> int:
@@ -388,9 +390,9 @@ class ProviderOpsService:
         # 成功或 auth_expired 时缓存（auth_expired 带有 cookie_expired 信息供前端显示警告）
         if result.status in (ActionStatus.SUCCESS, ActionStatus.AUTH_EXPIRED) and result.data:
             await self._cache_balance(provider_id, result)
-        # auth_failed 时清除缓存（配置错误，用户修正后应立即重试）
+        # auth_failed 时也缓存（使用较短 TTL），避免前端无限显示"加载中..."
         elif result.status == ActionStatus.AUTH_FAILED:
-            await self._clear_balance_cache(provider_id)
+            await self._cache_auth_failed(provider_id, result)
 
         return result
 
@@ -460,11 +462,29 @@ class ProviderOpsService:
                     pass
 
     async def _clear_balance_cache(self, provider_id: str) -> None:
-        """清除余额缓存（认证失败时调用）"""
+        """清除余额缓存"""
         cache_key = f"provider_ops:balance:{provider_id}"
         await CacheService.delete(cache_key)
         logger.info(f"余额缓存已清除: provider_id={provider_id}")
 
+    async def _cache_auth_failed(self, provider_id: str, result: ActionResult) -> None:
+        """
+        缓存认证失败结果（使用较短 TTL）
+
+        这样前端可以立即显示错误信息，而不是无限显示"加载中..."。
+        用户修正配置后，等待 60 秒或手动刷新即可重试。
+        """
+        cache_key = f"provider_ops:balance:{provider_id}"
+        cache_data = {
+            "status": result.status.value,
+            "data": None,
+            "message": result.message,
+            "executed_at": result.executed_at.isoformat() if result.executed_at else None,
+            "response_time_ms": result.response_time_ms,
+        }
+        await CacheService.set(cache_key, cache_data, AUTH_FAILED_CACHE_TTL)
+        logger.info(f"余额缓存已写入（认证失败）: provider_id={provider_id}, message={result.message}")
+
     async def _cache_balance(self, provider_id: str, result: ActionResult) -> None:
         """缓存余额结果"""
         cache_key = f"provider_ops:balance:{provider_id}"
@@ -546,13 +566,17 @@ class ProviderOpsService:
                 else datetime.now(timezone.utc)
             )
 
+            status = ActionStatus(cached.get("status", "success"))
+            # 认证失败使用较短的缓存 TTL
+            ttl = AUTH_FAILED_CACHE_TTL if status == ActionStatus.AUTH_FAILED else BALANCE_CACHE_TTL
             return ActionResult(
-                status=ActionStatus(cached.get("status", "success")),
+                status=status,
                 action_type=ProviderActionType.QUERY_BALANCE,
                 data=data,
+                message=cached.get("message"),
                 executed_at=executed_at,
                 response_time_ms=cached.get("response_time_ms"),
-                cache_ttl_seconds=BALANCE_CACHE_TTL,
+                cache_ttl_seconds=ttl,
             )
         except Exception as e:
             logger.warning(f"解析缓存余额失败: provider_id={provider_id}, error={e}")

src/services/system/audit.py

@@ -78,20 +78,24 @@ class AuditService:
         db.flush()
 
         # 同时记录到系统日志
-        log_message = (
-            f"AUDIT [{event_type.value}] - {description} | "
-            f"user_id={user_id}, ip={ip_address}"
-        )
+        # 检查 metadata 中是否有 quiet_logging 标志（由高频轮询端点设置）
+        quiet_logging = metadata.get("quiet_logging", False) if metadata else False
 
-        if event_type in [
-            AuditEventType.UNAUTHORIZED_ACCESS,
-            AuditEventType.SUSPICIOUS_ACTIVITY,
-        ]:
-            logger.warning(log_message)
-        elif event_type in [AuditEventType.LOGIN_FAILED, AuditEventType.REQUEST_FAILED]:
-            logger.info(log_message)
-        else:
-            logger.debug(log_message)
+        if not quiet_logging:
+            log_message = (
+                f"AUDIT [{event_type.value}] - {description} | "
+                f"user_id={user_id}, ip={ip_address}"
+            )
+
+            if event_type in [
+                AuditEventType.UNAUTHORIZED_ACCESS,
+                AuditEventType.SUSPICIOUS_ACTIVITY,
+            ]:
+                logger.warning(log_message)
+            elif event_type in [AuditEventType.LOGIN_FAILED, AuditEventType.REQUEST_FAILED]:
+                logger.info(log_message)
+            else:
+                logger.debug(log_message)
 
         return audit_log
 

src/services/usage/consumer_streams.py

@@ -14,6 +14,7 @@ import time
 from typing import Any, Dict, List, Optional, Tuple
 
 from redis.exceptions import ResponseError
+from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
 from src.clients.redis_client import get_redis_client
@@ -136,6 +137,12 @@ class UsageQueueConsumer:
         self._dlq_maxlen = config.usage_queue_dlq_maxlen
         self._metrics_interval = config.usage_queue_metrics_interval_seconds
 
+    @staticmethod
+    def _is_duplicate_key_error(exc: IntegrityError) -> bool:
+        """判断是否为重复键错误（唯一约束冲突）"""
+        err_str = str(exc).lower()
+        return "unique" in err_str or "duplicate" in err_str
+
     async def start(self) -> None:
         if self._running:
             return
@@ -303,6 +310,19 @@ class UsageQueueConsumer:
                 try:
                     await self._apply_record_event(event, db=db)
                     success_ids.append(message_id)
+                except IntegrityError as ie:
+                    # 重复 request_id 导致的唯一约束冲突，视为成功（记录已存在）
+                    try:
+                        db.rollback()
+                    except Exception:
+                        pass
+                    if self._is_duplicate_key_error(ie):
+                        logger.debug(
+                            f"[usage-queue] Duplicate request_id, skipping: {event.request_id}"
+                        )
+                        success_ids.append(message_id)
+                    else:
+                        await self._handle_processing_error(redis_client, message_id, fields, ie)
                 except Exception as individual_exc:
                     await self._handle_processing_error(redis_client, message_id, fields, individual_exc)
             # 批量 ACK 成功处理的消息

src/services/usage/service.py

@@ -12,7 +12,15 @@ from sqlalchemy.orm import Session
 
 from src.core.api_format.metadata import can_passthrough
 from src.core.logger import logger
-from src.models.database import ApiKey, Provider, ProviderAPIKey, Usage, User, UserRole
+from src.models.database import (
+    ApiKey,
+    Provider,
+    ProviderAPIKey,
+    RequestCandidate,
+    Usage,
+    User,
+    UserRole,
+)
 from src.services.model.cost import ModelCostService
 from src.services.system.config import SystemConfigService
 
@@ -1025,6 +1033,7 @@ class UsageService:
         - 批量插入 Usage 记录，减少 commit 次数
         - 聚合更新用户/API Key 统计（按 user_id/api_key_id 分组）
         - 聚合更新 GlobalModel 和 Provider 统计
+        - 支持更新已存在的 pending/streaming 状态记录
 
         Args:
             db: 数据库会话
@@ -1040,6 +1049,43 @@ class UsageService:
         from sqlalchemy import update
         from src.models.database import ApiKey as ApiKeyModel, User as UserModel, GlobalModel
 
+        # 分离需要更新和需要新建的记录
+        request_ids = [r.get("request_id") for r in records if r.get("request_id")]
+        existing_usages: Dict[str, Usage] = {}
+        records_to_update: List[Dict[str, Any]] = []
+        records_to_insert: List[Dict[str, Any]] = []
+
+        if request_ids:
+            # 查询已存在的 Usage 记录（包括 pending/streaming 状态）
+            existing_records = (
+                db.query(Usage)
+                .filter(Usage.request_id.in_(request_ids))
+                .all()
+            )
+            existing_usages = {u.request_id: u for u in existing_records}
+
+            for record in records:
+                req_id = record.get("request_id")
+                if req_id and req_id in existing_usages:
+                    existing_usage = existing_usages[req_id]
+                    # 只更新 pending/streaming 状态的记录
+                    # 已经是 completed/failed/cancelled 的记录跳过
+                    if existing_usage.status in ("pending", "streaming"):
+                        records_to_update.append(record)
+                    else:
+                        logger.debug(
+                            f"批量记录预过滤: 跳过已完成的 request_id={req_id} (status={existing_usage.status})"
+                        )
+                else:
+                    records_to_insert.append(record)
+        else:
+            records_to_insert = list(records)
+
+        if records_to_update:
+            logger.debug(
+                f"批量记录: 需要更新 {len(records_to_update)} 条已存在的 pending/streaming 记录"
+            )
+
         usages: List[Usage] = []
         user_costs: Dict[str, float] = defaultdict(float)  # user_id -> total_cost
         apikey_stats: Dict[str, Dict[str, Any]] = defaultdict(
@@ -1048,9 +1094,12 @@ class UsageService:
         model_counts: Dict[str, int] = defaultdict(int)  # model -> count
         provider_costs: Dict[str, float] = defaultdict(float)  # provider_id -> cost
 
+        # 合并所有需要处理的记录（用于预取 user/api_key）
+        all_records = records_to_insert + records_to_update
+
         # 批量预取 User 和 ApiKey，避免 N+1 查询
-        user_ids = {r.get("user_id") for r in records if r.get("user_id")}
-        api_key_ids = {r.get("api_key_id") for r in records if r.get("api_key_id")}
+        user_ids = {r.get("user_id") for r in all_records if r.get("user_id")}
+        api_key_ids = {r.get("api_key_id") for r in all_records if r.get("api_key_id")}
 
         users_map: Dict[str, User] = {}
         if user_ids:
@@ -1063,9 +1112,95 @@ class UsageService:
             api_keys_map = {str(k.id): k for k in api_keys}
 
         skipped_count = 0
-        total_count = len(records)
+        updated_count = 0
+        total_count = len(all_records)
 
-        for record in records:
+        # 1. 处理需要更新的记录（pending/streaming -> completed/failed/cancelled）
+        for record in records_to_update:
+            try:
+                request_id = record.get("request_id")
+                existing_usage = existing_usages.get(request_id)
+                if not existing_usage:
+                    skipped_count += 1
+                    continue
+
+                # 从预取的 map 中获取 user 和 api_key 对象
+                user_id = record.get("user_id")
+                api_key_id = record.get("api_key_id")
+                user = users_map.get(str(user_id)) if user_id else None
+                api_key = api_keys_map.get(str(api_key_id)) if api_key_id else None
+
+                # 准备记录参数
+                params = UsageRecordParams(
+                    db=db,
+                    user=user,
+                    api_key=api_key,
+                    provider=record.get("provider") or "unknown",
+                    model=record.get("model") or "unknown",
+                    input_tokens=int(record.get("input_tokens") or 0),
+                    output_tokens=int(record.get("output_tokens") or 0),
+                    cache_creation_input_tokens=int(record.get("cache_creation_input_tokens") or 0),
+                    cache_read_input_tokens=int(record.get("cache_read_input_tokens") or 0),
+                    request_type=record.get("request_type") or "chat",
+                    api_format=record.get("api_format"),
+                    endpoint_api_format=record.get("endpoint_api_format"),
+                    has_format_conversion=bool(record.get("has_format_conversion")),
+                    is_stream=bool(record.get("is_stream", True)),
+                    response_time_ms=record.get("response_time_ms"),
+                    first_byte_time_ms=record.get("first_byte_time_ms"),
+                    status_code=int(record.get("status_code") or 200),
+                    error_message=record.get("error_message"),
+                    metadata=record.get("metadata"),
+                    request_headers=record.get("request_headers"),
+                    request_body=record.get("request_body"),
+                    provider_request_headers=record.get("provider_request_headers"),
+                    response_headers=record.get("response_headers"),
+                    client_response_headers=record.get("client_response_headers"),
+                    response_body=record.get("response_body"),
+                    request_id=request_id,
+                    provider_id=record.get("provider_id"),
+                    provider_endpoint_id=record.get("provider_endpoint_id"),
+                    provider_api_key_id=record.get("provider_api_key_id"),
+                    status=record.get("status") or "completed",
+                    cache_ttl_minutes=record.get("cache_ttl_minutes"),
+                    use_tiered_pricing=record.get("use_tiered_pricing", True),
+                    target_model=record.get("target_model"),
+                )
+
+                usage_params, total_cost = await cls._prepare_usage_record(params)
+
+                # 更新已存在的 Usage 记录
+                cls._update_existing_usage(existing_usage, usage_params, record.get("target_model"))
+                usages.append(existing_usage)
+                updated_count += 1
+
+                # 聚合统计（更新记录也需要更新统计，因为 pending 状态没有计费）
+                model_name = record.get("model") or "unknown"
+                model_counts[model_name] += 1
+
+                provider_id = record.get("provider_id")
+                if provider_id:
+                    actual_cost = usage_params.get("actual_total_cost_usd", 0)
+                    provider_costs[provider_id] += actual_cost
+
+                # 用户统计（独立 Key 不计入创建者）
+                if user and not (api_key and api_key.is_standalone):
+                    user_costs[str(user.id)] += total_cost
+
+                # API Key 统计
+                if api_key:
+                    key_id = str(api_key.id)
+                    apikey_stats[key_id]["requests"] += 1
+                    apikey_stats[key_id]["cost"] += total_cost
+                    apikey_stats[key_id]["is_standalone"] = api_key.is_standalone
+
+            except Exception as e:
+                skipped_count += 1
+                logger.warning(f"批量记录中更新失败: {e}, request_id={record.get('request_id')}")
+                continue
+
+        # 2. 处理需要新建的记录
+        for record in records_to_insert:
             try:
                 # 从预取的 map 中获取 user 和 api_key 对象
                 user_id = record.get("user_id")
@@ -1215,7 +1350,13 @@ class UsageService:
         # 单次提交所有更改
         try:
             db.commit()
-            logger.debug(f"批量记录 {len(usages)} 条使用记录成功")
+            inserted_count = len(usages) - updated_count
+            if updated_count > 0:
+                logger.debug(
+                    f"批量记录成功: 更新 {updated_count} 条, 新建 {inserted_count} 条"
+                )
+            else:
+                logger.debug(f"批量记录 {len(usages)} 条使用记录成功")
         except Exception as e:
             logger.error(f"批量提交使用记录时出错: {e}")
             db.rollback()
@@ -1989,7 +2130,8 @@ class UsageService:
         records = query.all()
 
         # 检查超时的 pending/streaming 请求
-        timeout_ids = []
+        # 收集可能超时的 usage_id 列表
+        timeout_candidates: List[str] = []
         for r in records:
             if r.status in ("pending", "streaming") and r.created_at:
                 # 使用全局配置的超时时间
@@ -2001,15 +2143,90 @@ class UsageService:
                     created_at = created_at.replace(tzinfo=timezone.utc)
                 elapsed = (now - created_at).total_seconds()
                 if elapsed > timeout_seconds:
-                    timeout_ids.append(r.id)
+                    # 需要获取 request_id 以便检查 RequestCandidate 表
+                    # r.id 是 usage_id，需要查询 request_id
+                    timeout_candidates.append(r.id)
 
-        # 批量更新超时的请求
-        if timeout_ids:
-            db.query(Usage).filter(Usage.id.in_(timeout_ids)).update(
-                {"status": "failed", "error_message": "请求超时（服务器可能已重启）"},
-                synchronize_session=False,
+        # 批量更新超时的请求（排除已有成功完成记录的请求）
+        timeout_ids = []
+        if timeout_candidates:
+            # 检查 RequestCandidate 表是否有成功完成的记录
+            # 如果流已经成功完成（stream_completed: true），不应该标记为超时
+            # 先获取这些 Usage 的 request_id
+            usage_request_ids = (
+                db.query(Usage.id, Usage.request_id)
+                .filter(Usage.id.in_(timeout_candidates))
+                .all()
             )
-            db.commit()
+            usage_id_to_request_id = {u.id: u.request_id for u in usage_request_ids}
+            request_id_to_usage_id = {u.request_id: u.id for u in usage_request_ids}
+            request_ids = list(request_id_to_usage_id.keys())
+
+            # 查询这些请求中已有成功完成记录的 request_id
+            # 包括两种情况：
+            # 1. status='success' 且 stream_completed=True（正常完成）
+            # 2. status='streaming' 且 status_code=200（流传输中但 Provider 已返回 200，可能是服务重启导致回调丢失）
+            completed_usage_ids = set()
+            if request_ids:
+                from sqlalchemy import or_
+
+                candidates = (
+                    db.query(
+                        RequestCandidate.request_id,
+                        RequestCandidate.status,
+                        RequestCandidate.status_code,
+                        RequestCandidate.extra_data,
+                    )
+                    .filter(
+                        RequestCandidate.request_id.in_(request_ids),
+                        or_(
+                            RequestCandidate.status == "success",
+                            # streaming 状态且 status_code=200，说明 Provider 响应成功
+                            # 但流传输可能因服务重启而中断
+                            (RequestCandidate.status == "streaming")
+                            & (RequestCandidate.status_code == 200),
+                        ),
+                    )
+                    .all()
+                )
+                for candidate in candidates:
+                    extra_data = candidate.extra_data or {}
+                    # 情况1：status='success' 且 stream_completed=True
+                    if candidate.status == "success" and extra_data.get(
+                        "stream_completed", False
+                    ):
+                        usage_id = request_id_to_usage_id.get(candidate.request_id)
+                        if usage_id:
+                            completed_usage_ids.add(usage_id)
+                    # 情况2：status='streaming' 且 status_code=200
+                    # 这表示 Provider 返回了 200，但流传输可能因服务重启而未正常结束
+                    # 此时应该恢复为 completed 而不是标记为 failed
+                    elif candidate.status == "streaming" and candidate.status_code == 200:
+                        usage_id = request_id_to_usage_id.get(candidate.request_id)
+                        if usage_id:
+                            completed_usage_ids.add(usage_id)
+
+            # 只对没有成功完成记录的请求标记超时
+            timeout_ids = [uid for uid in timeout_candidates if uid not in completed_usage_ids]
+
+            if timeout_ids:
+                db.query(Usage).filter(Usage.id.in_(timeout_ids)).update(
+                    {"status": "failed", "error_message": "请求超时（服务器可能已重启）"},
+                    synchronize_session=False,
+                )
+                db.commit()
+
+            # 对于已完成但状态未更新的请求，主动恢复状态为 completed
+            # 这处理了遥测回调丢失的情况（例如服务重启、后台任务未执行等）
+            if completed_usage_ids:
+                db.query(Usage).filter(Usage.id.in_(list(completed_usage_ids))).update(
+                    {"status": "completed"},
+                    synchronize_session=False,
+                )
+                db.commit()
+                logger.info(
+                    f"[Usage] 恢复 {len(completed_usage_ids)} 个已完成请求的状态（遥测回调丢失）"
+                )
 
         result: List[Dict[str, Any]] = []
         for r in records: