feat: add provider compatibility error detection for intelligent failover

- Introduce ProviderCompatibilityException for unsupported parameter/feature errors
- Add COMPATIBILITY_ERROR_PATTERNS to detect provider-specific limitations
- Implement _is_compatibility_error() method in ErrorClassifier
- Prioritize compatibility error checking before client error validation
- Remove 'max_tokens' from CLIENT_ERROR_PATTERNS as it can indicate compatibility issues
- Enable automatic failover when provider doesn't support requested features
- Improve error classification accuracy with pattern matching for common compatibility issues

.github/workflows/docker-publish.yml

@@ -15,6 +15,8 @@ env:
   REGISTRY: ghcr.io
   BASE_IMAGE_NAME: fawney19/aether-base
   APP_IMAGE_NAME: fawney19/aether
+  # Files that affect base image - used for hash calculation
+  BASE_FILES: "Dockerfile.base pyproject.toml frontend/package.json frontend/package-lock.json"
 
 jobs:
   check-base-changes:
@@ -23,8 +25,13 @@ jobs:
       base_changed: ${{ steps.check.outputs.base_changed }}
     steps:
       - uses: actions/checkout@v4
+
+      - name: Log in to Container Registry
+        uses: docker/login-action@v3
         with:
-          fetch-depth: 2
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Check if base image needs rebuild
         id: check
@@ -34,10 +41,26 @@ jobs:
             exit 0
           fi
 
-          # Check if base-related files changed
+          # Calculate current hash of base-related files
-          if git diff --name-only HEAD~1 HEAD | grep -qE '^(Dockerfile\.base|pyproject\.toml|frontend/package.*\.json)$'; then
+          CURRENT_HASH=$(cat ${{ env.BASE_FILES }} 2>/dev/null | sha256sum | cut -d' ' -f1)
+          echo "Current base files hash: $CURRENT_HASH"
+
+          # Try to get hash label from remote image config
+          # Pull the image config and extract labels
+          REMOTE_HASH=""
+          if docker pull ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest 2>/dev/null; then
+            REMOTE_HASH=$(docker inspect ${{ env.REGISTRY }}/${{ env.BASE_IMAGE_NAME }}:latest --format '{{ index .Config.Labels "org.opencontainers.image.base.hash" }}' 2>/dev/null) || true
+          fi
+
+          if [ -z "$REMOTE_HASH" ] || [ "$REMOTE_HASH" == "<no value>" ]; then
+            # No remote image or no hash label, need to rebuild
+            echo "No remote base image or hash label found, need rebuild"
+            echo "base_changed=true" >> $GITHUB_OUTPUT
+          elif [ "$CURRENT_HASH" != "$REMOTE_HASH" ]; then
+            echo "Hash mismatch: remote=$REMOTE_HASH, current=$CURRENT_HASH"
             echo "base_changed=true" >> $GITHUB_OUTPUT
           else
+            echo "Hash matches, no rebuild needed"
             echo "base_changed=false" >> $GITHUB_OUTPUT
           fi
 
@@ -61,6 +84,12 @@ jobs:
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Calculate base files hash
+        id: hash
+        run: |
+          HASH=$(cat ${{ env.BASE_FILES }} 2>/dev/null | sha256sum | cut -d' ' -f1)
+          echo "hash=$HASH" >> $GITHUB_OUTPUT
+
       - name: Extract metadata for base image
         id: meta
         uses: docker/metadata-action@v5
@@ -69,6 +98,8 @@ jobs:
           tags: |
             type=raw,value=latest
             type=sha,prefix=
+          labels: |
+            org.opencontainers.image.base.hash=${{ steps.hash.outputs.hash }}
 
       - name: Build and push base image
         uses: docker/build-push-action@v5

Dockerfile.base

@@ -19,7 +19,7 @@ RUN apt-get update && apt-get install -y \
 # Python 依赖（安装到系统，不用 -e 模式）
 COPY pyproject.toml README.md ./
 RUN mkdir -p src && touch src/__init__.py && \
-    pip install --no-cache-dir .
+    SETUPTOOLS_SCM_PRETEND_VERSION=0.1.0 pip install --no-cache-dir .
 
 # 前端依赖
 COPY frontend/package*.json /tmp/frontend/
@@ -70,6 +70,8 @@ RUN printf '%s\n' \
 '        proxy_cache off;' \
 '        proxy_request_buffering off;' \
 '        chunked_transfer_encoding on;' \
+'        gzip off;' \
+'        add_header X-Accel-Buffering no;' \
 '        proxy_connect_timeout 600s;' \
 '        proxy_send_timeout 600s;' \
 '        proxy_read_timeout 600s;' \

Dockerfile.base.local

@@ -74,6 +74,8 @@ RUN printf '%s\n' \
 '        proxy_cache off;' \
 '        proxy_request_buffering off;' \
 '        chunked_transfer_encoding on;' \
+'        gzip off;' \
+'        add_header X-Accel-Buffering no;' \
 '        proxy_connect_timeout 600s;' \
 '        proxy_send_timeout 600s;' \
 '        proxy_read_timeout 600s;' \

deploy.sh

@@ -21,9 +21,9 @@ HASH_FILE=".deps-hash"
 CODE_HASH_FILE=".code-hash"
 MIGRATION_HASH_FILE=".migration-hash"
 
-# 计算依赖文件的哈希值
+# 计算依赖文件的哈希值（包含 Dockerfile.base.local）
 calc_deps_hash() {
-    cat pyproject.toml frontend/package.json frontend/package-lock.json 2>/dev/null | md5sum | cut -d' ' -f1
+    cat pyproject.toml frontend/package.json frontend/package-lock.json Dockerfile.base.local 2>/dev/null | md5sum | cut -d' ' -f1
 }
 
 # 计算代码文件的哈希值
@@ -162,25 +162,32 @@ git pull
 
 # 标记是否需要重启
 NEED_RESTART=false
+BASE_REBUILT=false
 
 # 检查基础镜像是否存在，或依赖是否变化
 if ! docker image inspect aether-base:latest >/dev/null 2>&1; then
     echo ">>> Base image not found, building..."
     build_base
+    BASE_REBUILT=true
     NEED_RESTART=true
 elif check_deps_changed; then
     echo ">>> Dependencies changed, rebuilding base image..."
     build_base
+    BASE_REBUILT=true
     NEED_RESTART=true
 else
     echo ">>> Dependencies unchanged."
 fi
 
-# 检查代码是否变化
+# 检查代码是否变化，或者 base 重建了（app 依赖 base）
 if ! docker image inspect aether-app:latest >/dev/null 2>&1; then
     echo ">>> App image not found, building..."
     build_app
     NEED_RESTART=true
+elif [ "$BASE_REBUILT" = true ]; then
+    echo ">>> Base image rebuilt, rebuilding app image..."
+    build_app
+    NEED_RESTART=true
 elif check_code_changed; then
     echo ">>> Code changed, rebuilding app image..."
     build_app

frontend/src/api/cache.ts

@@ -66,6 +66,7 @@ export interface UserAffinity {
   key_name: string | null
   key_prefix: string | null  // Provider Key 脱敏显示（前4...后4）
   rate_multiplier: number
+  global_model_id: string | null  // 原始的 global_model_id（用于删除）
   model_name: string | null  // 模型名称（如 claude-haiku-4-5-20250514）
   model_display_name: string | null  // 模型显示名称（如 Claude Haiku 4.5）
   api_format: string | null  // API 格式 (claude/openai)
@@ -119,6 +120,18 @@ export const cacheApi = {
     await api.delete(`/api/admin/monitoring/cache/users/${userIdentifier}`)
   },
 
+  /**
+   * 清除单条缓存亲和性
+   *
+   * @param affinityKey API Key ID
+   * @param endpointId Endpoint ID
+   * @param modelId GlobalModel ID
+   * @param apiFormat API 格式 (claude/openai)
+   */
+  async clearSingleAffinity(affinityKey: string, endpointId: string, modelId: string, apiFormat: string): Promise<void> {
+    await api.delete(`/api/admin/monitoring/cache/affinity/${affinityKey}/${endpointId}/${modelId}/${apiFormat}`)
+  },
+
   /**
    * 清除所有缓存
    */

frontend/src/views/admin/CacheMonitoring.vue

@@ -142,32 +142,37 @@ async function resetAffinitySearch() {
   await fetchAffinityList()
 }
 
-async function clearUserCache(identifier: string, displayName?: string) {
+async function clearSingleAffinity(item: UserAffinity) {
-  const target = identifier?.trim()
+  const affinityKey = item.affinity_key?.trim()
-  if (!target) {
+  const endpointId = item.endpoint_id?.trim()
-    showError('无法识别标识符')
+  const modelId = item.global_model_id?.trim()
+  const apiFormat = item.api_format?.trim()
+
+  if (!affinityKey || !endpointId || !modelId || !apiFormat) {
+    showError('缓存记录信息不完整，无法删除')
     return
   }
 
-  const label = displayName || target
+  const label = item.user_api_key_name || affinityKey
+  const modelLabel = item.model_display_name || item.model_name || modelId
   const confirmed = await showConfirm({
     title: '确认清除',
-    message: `确定要清除 ${label} 的缓存吗？`,
+    message: `确定要清除 ${label} 在模型 ${modelLabel} 上的缓存亲和性吗？`,
     confirmText: '确认清除',
     variant: 'destructive'
   })
 
   if (!confirmed) return
 
-  clearingRowAffinityKey.value = target
+  clearingRowAffinityKey.value = affinityKey
   try {
-    await cacheApi.clearUserCache(target)
+    await cacheApi.clearSingleAffinity(affinityKey, endpointId, modelId, apiFormat)
     showSuccess('清除成功')
     await fetchCacheStats()
     await fetchAffinityList(tableKeyword.value.trim() || undefined)
   } catch (error) {
     showError('清除失败')
-    log.error('清除用户缓存失败', error)
+    log.error('清除单条缓存失败', error)
   } finally {
     clearingRowAffinityKey.value = null
   }
@@ -618,7 +623,7 @@ onBeforeUnmount(() => {
                 class="h-7 w-7 text-muted-foreground/70 hover:text-destructive"
                 :disabled="clearingRowAffinityKey === item.affinity_key"
                 title="清除缓存"
-                @click="clearUserCache(item.affinity_key, item.user_api_key_name || item.affinity_key)"
+                @click="clearSingleAffinity(item)"
               >
                 <Trash2 class="h-3.5 w-3.5" />
               </Button>
@@ -668,7 +673,7 @@ onBeforeUnmount(() => {
               variant="ghost"
               class="h-7 w-7 text-muted-foreground/70 hover:text-destructive shrink-0"
               :disabled="clearingRowAffinityKey === item.affinity_key"
-              @click="clearUserCache(item.affinity_key, item.user_api_key_name || item.affinity_key)"
+              @click="clearSingleAffinity(item)"
             >
               <Trash2 class="h-3.5 w-3.5" />
             </Button>

frontend/src/views/admin/SystemSettings.vue

@@ -464,6 +464,7 @@
           </div>
         </div>
       </CardSection>
+
     </div>
 
     <!-- 导入配置对话框 -->

src/api/admin/monitoring/cache.py

@@ -186,6 +186,30 @@ async def clear_user_cache(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.delete("/affinity/{affinity_key}/{endpoint_id}/{model_id}/{api_format}")
+async def clear_single_affinity(
+    affinity_key: str,
+    endpoint_id: str,
+    model_id: str,
+    api_format: str,
+    request: Request,
+    db: Session = Depends(get_db),
+) -> Any:
+    """
+    Clear a single cache affinity entry
+
+    Parameters:
+    - affinity_key: API Key ID
+    - endpoint_id: Endpoint ID
+    - model_id: Model ID (GlobalModel ID)
+    - api_format: API format (claude/openai)
+    """
+    adapter = AdminClearSingleAffinityAdapter(
+        affinity_key=affinity_key, endpoint_id=endpoint_id, model_id=model_id, api_format=api_format
+    )
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 @router.delete("")
 async def clear_all_cache(
     request: Request,
@@ -655,6 +679,7 @@ class AdminListAffinitiesAdapter(AdminApiAdapter):
                 "key_name": key.name if key else None,
                 "key_prefix": provider_key_masked,
                 "rate_multiplier": key.rate_multiplier if key else 1.0,
+                "global_model_id": affinity.get("model_name"),  # 原始的 global_model_id
                 "model_name": (
                     global_model_map.get(affinity.get("model_name")).name
                     if affinity.get("model_name") and global_model_map.get(affinity.get("model_name"))
@@ -817,6 +842,65 @@ class AdminClearUserCacheAdapter(AdminApiAdapter):
             raise HTTPException(status_code=500, detail=f"清除失败: {exc}")
 
 
+@dataclass
+class AdminClearSingleAffinityAdapter(AdminApiAdapter):
+    affinity_key: str
+    endpoint_id: str
+    model_id: str
+    api_format: str
+
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
+        db = context.db
+        try:
+            redis_client = get_redis_client_sync()
+            affinity_mgr = await get_affinity_manager(redis_client)
+
+            # 直接获取指定的亲和性记录（无需遍历全部）
+            existing_affinity = await affinity_mgr.get_affinity(
+                self.affinity_key, self.api_format, self.model_id
+            )
+
+            if not existing_affinity:
+                raise HTTPException(status_code=404, detail="未找到指定的缓存亲和性记录")
+
+            # 验证 endpoint_id 是否匹配
+            if existing_affinity.endpoint_id != self.endpoint_id:
+                raise HTTPException(status_code=404, detail="未找到指定的缓存亲和性记录")
+
+            # 失效单条记录
+            await affinity_mgr.invalidate_affinity(
+                self.affinity_key, self.api_format, self.model_id, endpoint_id=self.endpoint_id
+            )
+
+            # 获取用于日志的信息
+            api_key = db.query(ApiKey).filter(ApiKey.id == self.affinity_key).first()
+            api_key_name = api_key.name if api_key else None
+
+            logger.info(
+                f"已清除单条缓存亲和性: affinity_key={self.affinity_key[:8]}..., "
+                f"endpoint_id={self.endpoint_id[:8]}..., model_id={self.model_id[:8]}..."
+            )
+
+            context.add_audit_metadata(
+                action="cache_clear_single",
+                affinity_key=self.affinity_key,
+                endpoint_id=self.endpoint_id,
+                model_id=self.model_id,
+            )
+            return {
+                "status": "ok",
+                "message": f"已清除缓存亲和性: {api_key_name or self.affinity_key[:8]}",
+                "affinity_key": self.affinity_key,
+                "endpoint_id": self.endpoint_id,
+                "model_id": self.model_id,
+            }
+        except HTTPException:
+            raise
+        except Exception as exc:
+            logger.exception(f"清除单条缓存亲和性失败: {exc}")
+            raise HTTPException(status_code=500, detail=f"清除失败: {exc}")
+
+
 class AdminClearAllCacheAdapter(AdminApiAdapter):
     async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         try:

src/api/admin/providers/models.py

@@ -9,6 +9,7 @@ from fastapi import APIRouter, Depends, Request
 from sqlalchemy.orm import Session, joinedload
 
 from src.api.base.admin_adapter import AdminApiAdapter
+from src.api.base.models_service import invalidate_models_list_cache
 from src.api.base.pipeline import ApiRequestPipeline
 from src.core.exceptions import InvalidRequestException, NotFoundException
 from src.core.logger import logger
@@ -419,4 +420,8 @@ class AdminBatchAssignModelsToProviderAdapter(AdminApiAdapter):
             f"Batch assigned {len(success)} GlobalModels to provider {provider.name} by {context.user.username}"
         )
 
+        # 清除 /v1/models 列表缓存
+        if success:
+            await invalidate_models_list_cache()
+
         return BatchAssignModelsToProviderResponse(success=success, errors=errors)

src/api/base/models_service.py

@@ -55,6 +55,23 @@ async def _set_cached_models(api_formats: list[str], models: list["ModelInfo"])
         logger.warning(f"[ModelsService] 缓存写入失败: {e}")
 
 
+async def invalidate_models_list_cache() -> None:
+    """
+    清除所有 /v1/models 列表缓存
+
+    在模型创建、更新、删除时调用，确保模型列表实时更新
+    """
+    # 清除所有格式的缓存
+    all_formats = ["CLAUDE", "OPENAI", "GEMINI"]
+    for fmt in all_formats:
+        cache_key = f"{_CACHE_KEY_PREFIX}:{fmt}"
+        try:
+            await CacheService.delete(cache_key)
+            logger.debug(f"[ModelsService] 已清除缓存: {cache_key}")
+        except Exception as e:
+            logger.warning(f"[ModelsService] 清除缓存失败 {cache_key}: {e}")
+
+
 @dataclass
 class ModelInfo:
     """统一的模型信息结构"""

src/api/handlers/base/cli_handler_base.py

@@ -1114,8 +1114,10 @@ class CliMessageHandlerBase(BaseMessageHandler):
             async for chunk in stream_generator:
                 yield chunk
         except asyncio.CancelledError:
-            ctx.status_code = 499
+            # 如果响应已完成，不标记为失败
-            ctx.error_message = "Client disconnected"
+            if not ctx.has_completion:
+                ctx.status_code = 499
+                ctx.error_message = "Client disconnected"
             raise
         except httpx.TimeoutException as e:
             ctx.status_code = 504

src/api/handlers/base/content_extractors.py

@@ -0,0 +1,274 @@
+"""
+流式内容提取器 - 策略模式实现
+
+为不同 API 格式（OpenAI、Claude、Gemini）提供内容提取和 chunk 构造的抽象。
+StreamSmoother 使用这些提取器来处理不同格式的 SSE 事件。
+"""
+
+import copy
+import json
+from abc import ABC, abstractmethod
+from typing import Optional
+
+
+class ContentExtractor(ABC):
+    """
+    流式内容提取器抽象基类
+
+    定义从 SSE 事件中提取文本内容和构造新 chunk 的接口。
+    每种 API 格式（OpenAI、Claude、Gemini）需要实现自己的提取器。
+    """
+
+    @abstractmethod
+    def extract_content(self, data: dict) -> Optional[str]:
+        """
+        从 SSE 数据中提取可拆分的文本内容
+
+        Args:
+            data: 解析后的 JSON 数据
+
+        Returns:
+            提取的文本内容，如果无法提取则返回 None
+        """
+        pass
+
+    @abstractmethod
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        """
+        使用新内容构造 SSE chunk
+
+        Args:
+            original_data: 原始 JSON 数据
+            new_content: 新的文本内容
+            event_type: SSE 事件类型（某些格式需要）
+            is_first: 是否是第一个 chunk（用于保留 role 等字段）
+
+        Returns:
+            编码后的 SSE 字节数据
+        """
+        pass
+
+
+class OpenAIContentExtractor(ContentExtractor):
+    """
+    OpenAI 格式内容提取器
+
+    处理 OpenAI Chat Completions API 的流式响应格式：
+    - 数据结构: choices[0].delta.content
+    - 只在 delta 仅包含 role/content 时允许拆分，避免破坏 tool_calls 等结构
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        choices = data.get("choices")
+        if not isinstance(choices, list) or len(choices) != 1:
+            return None
+
+        first_choice = choices[0]
+        if not isinstance(first_choice, dict):
+            return None
+
+        delta = first_choice.get("delta")
+        if not isinstance(delta, dict):
+            return None
+
+        content = delta.get("content")
+        if not isinstance(content, str):
+            return None
+
+        # 只有 delta 仅包含 role/content 时才允许拆分
+        # 避免破坏 tool_calls、function_call 等复杂结构
+        allowed_keys = {"role", "content"}
+        if not all(key in allowed_keys for key in delta.keys()):
+            return None
+
+        return content
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = original_data.copy()
+
+        if "choices" in new_data and new_data["choices"]:
+            new_choices = []
+            for choice in new_data["choices"]:
+                new_choice = choice.copy()
+                if "delta" in new_choice:
+                    new_delta = {}
+                    # 只有第一个 chunk 保留 role
+                    if is_first and "role" in new_choice["delta"]:
+                        new_delta["role"] = new_choice["delta"]["role"]
+                    new_delta["content"] = new_content
+                    new_choice["delta"] = new_delta
+                new_choices.append(new_choice)
+            new_data["choices"] = new_choices
+
+        return f"data: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode("utf-8")
+
+
+class ClaudeContentExtractor(ContentExtractor):
+    """
+    Claude 格式内容提取器
+
+    处理 Claude Messages API 的流式响应格式：
+    - 事件类型: content_block_delta
+    - 数据结构: delta.type=text_delta, delta.text
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        # 检查事件类型
+        if data.get("type") != "content_block_delta":
+            return None
+
+        delta = data.get("delta", {})
+        if not isinstance(delta, dict):
+            return None
+
+        # 检查 delta 类型
+        if delta.get("type") != "text_delta":
+            return None
+
+        text = delta.get("text")
+        if not isinstance(text, str):
+            return None
+
+        return text
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = original_data.copy()
+
+        if "delta" in new_data:
+            new_delta = new_data["delta"].copy()
+            new_delta["text"] = new_content
+            new_data["delta"] = new_delta
+
+        # Claude 格式需要 event: 前缀
+        event_name = event_type or "content_block_delta"
+        return f"event: {event_name}\ndata: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode(
+            "utf-8"
+        )
+
+
+class GeminiContentExtractor(ContentExtractor):
+    """
+    Gemini 格式内容提取器
+
+    处理 Gemini API 的流式响应格式：
+    - 数据结构: candidates[0].content.parts[0].text
+    - 只有纯文本块才拆分
+    """
+
+    def extract_content(self, data: dict) -> Optional[str]:
+        if not isinstance(data, dict):
+            return None
+
+        candidates = data.get("candidates")
+        if not isinstance(candidates, list) or len(candidates) != 1:
+            return None
+
+        first_candidate = candidates[0]
+        if not isinstance(first_candidate, dict):
+            return None
+
+        content = first_candidate.get("content", {})
+        if not isinstance(content, dict):
+            return None
+
+        parts = content.get("parts", [])
+        if not isinstance(parts, list) or len(parts) != 1:
+            return None
+
+        first_part = parts[0]
+        if not isinstance(first_part, dict):
+            return None
+
+        text = first_part.get("text")
+        # 只有纯文本块（只有 text 字段）才拆分
+        if not isinstance(text, str) or len(first_part) != 1:
+            return None
+
+        return text
+
+    def create_chunk(
+        self,
+        original_data: dict,
+        new_content: str,
+        event_type: str = "",
+        is_first: bool = False,
+    ) -> bytes:
+        new_data = copy.deepcopy(original_data)
+
+        if "candidates" in new_data and new_data["candidates"]:
+            first_candidate = new_data["candidates"][0]
+            if "content" in first_candidate:
+                content = first_candidate["content"]
+                if "parts" in content and content["parts"]:
+                    content["parts"][0]["text"] = new_content
+
+        return f"data: {json.dumps(new_data, ensure_ascii=False)}\n\n".encode("utf-8")
+
+
+# 提取器注册表
+_EXTRACTORS: dict[str, type[ContentExtractor]] = {
+    "openai": OpenAIContentExtractor,
+    "claude": ClaudeContentExtractor,
+    "gemini": GeminiContentExtractor,
+}
+
+
+def get_extractor(format_name: str) -> Optional[ContentExtractor]:
+    """
+    根据格式名获取对应的内容提取器实例
+
+    Args:
+        format_name: 格式名称（openai, claude, gemini）
+
+    Returns:
+        对应的提取器实例，如果格式不支持则返回 None
+    """
+    extractor_class = _EXTRACTORS.get(format_name.lower())
+    if extractor_class:
+        return extractor_class()
+    return None
+
+
+def register_extractor(format_name: str, extractor_class: type[ContentExtractor]) -> None:
+    """
+    注册新的内容提取器
+
+    Args:
+        format_name: 格式名称
+        extractor_class: 提取器类
+    """
+    _EXTRACTORS[format_name.lower()] = extractor_class
+
+
+def get_extractor_formats() -> list[str]:
+    """
+    获取所有已注册的格式名称列表
+
+    Returns:
+        格式名称列表
+    """
+    return list(_EXTRACTORS.keys())

src/api/handlers/base/stream_processor.py

@@ -6,16 +6,22 @@
 2. 响应流生成
 3. 预读和嵌套错误检测
 4. 客户端断开检测
+5. 流式平滑输出
 """
 
 import asyncio
 import codecs
 import json
-import time
+from dataclasses import dataclass
 from typing import Any, AsyncGenerator, Callable, Optional
 
 import httpx
 
+from src.api.handlers.base.content_extractors import (
+    ContentExtractor,
+    get_extractor,
+    get_extractor_formats,
+)
 from src.api.handlers.base.parsers import get_parser_for_format
 from src.api.handlers.base.response_parser import ResponseParser
 from src.api.handlers.base.stream_context import StreamContext
@@ -25,11 +31,20 @@ from src.models.database import Provider, ProviderEndpoint
 from src.utils.sse_parser import SSEEventParser
 
 
+@dataclass
+class StreamSmoothingConfig:
+    """流式平滑输出配置"""
+
+    enabled: bool = False
+    chunk_size: int = 20
+    delay_ms: int = 8
+
+
 class StreamProcessor:
     """
     流式响应处理器
 
-    负责处理 SSE 流的解析、错误检测和响应生成。
+    负责处理 SSE 流的解析、错误检测、响应生成和平滑输出。
     从 ChatHandlerBase 中提取，使其职责更加单一。
     """
 
@@ -40,6 +55,7 @@ class StreamProcessor:
         on_streaming_start: Optional[Callable[[], None]] = None,
         *,
         collect_text: bool = False,
+        smoothing_config: Optional[StreamSmoothingConfig] = None,
     ):
         """
         初始化流处理器
@@ -48,11 +64,17 @@ class StreamProcessor:
             request_id: 请求 ID（用于日志）
             default_parser: 默认响应解析器
             on_streaming_start: 流开始时的回调（用于更新状态）
+            collect_text: 是否收集文本内容
+            smoothing_config: 流式平滑输出配置
         """
         self.request_id = request_id
         self.default_parser = default_parser
         self.on_streaming_start = on_streaming_start
         self.collect_text = collect_text
+        self.smoothing_config = smoothing_config or StreamSmoothingConfig()
+
+        # 内容提取器缓存
+        self._extractors: dict[str, ContentExtractor] = {}
 
     def get_parser_for_provider(self, ctx: StreamContext) -> ResponseParser:
         """
@@ -127,6 +149,13 @@ class StreamProcessor:
         if event_type in ("response.completed", "message_stop"):
             ctx.has_completion = True
 
+        # 检查 OpenAI 格式的 finish_reason
+        choices = data.get("choices", [])
+        if choices and isinstance(choices, list) and len(choices) > 0:
+            finish_reason = choices[0].get("finish_reason")
+            if finish_reason is not None:
+                ctx.has_completion = True
+
     async def prefetch_and_check_error(
         self,
         byte_iterator: Any,
@@ -369,7 +398,7 @@ class StreamProcessor:
             sse_parser: SSE 解析器
             line: 原始行数据
         """
-        # SSEEventParser 以“去掉换行符”的单行文本作为输入；这里统一剔除 CR/LF，
+        # SSEEventParser 以"去掉换行符"的单行文本作为输入；这里统一剔除 CR/LF，
         # 避免把空行误判成 "\n" 并导致事件边界解析错误。
         normalized_line = line.rstrip("\r\n")
         events = sse_parser.feed_line(normalized_line)
@@ -400,32 +429,201 @@ class StreamProcessor:
             响应数据块
         """
         try:
-            # 断连检查频率：每次 await 都会引入调度开销，过于频繁会让流式"发一段停一段"
+            # 使用后台任务检查断连，完全不阻塞流式传输
-            # 这里按时间间隔节流，兼顾及时停止上游读取与吞吐平滑性。
+            disconnected = False
-            next_disconnect_check_at = 0.0
-            disconnect_check_interval_s = 0.25
 
-            async for chunk in stream_generator:
+            async def check_disconnect_background() -> None:
-                now = time.monotonic()
+                nonlocal disconnected
-                if now >= next_disconnect_check_at:
+                while not disconnected and not ctx.has_completion:
-                    next_disconnect_check_at = now + disconnect_check_interval_s
+                    await asyncio.sleep(0.5)
                     if await is_disconnected():
-                        logger.warning(f"ID:{self.request_id} | Client disconnected")
+                        disconnected = True
-                        ctx.status_code = 499  # Client Closed Request
-                        ctx.error_message = "client_disconnected"
-
                         break
-                yield chunk
-        except asyncio.CancelledError:
-            ctx.status_code = 499
-            ctx.error_message = "client_disconnected"
 
+            # 启动后台检查任务
+            check_task = asyncio.create_task(check_disconnect_background())
+
+            try:
+                async for chunk in stream_generator:
+                    if disconnected:
+                        # 如果响应已完成，客户端断开不算失败
+                        if ctx.has_completion:
+                            logger.info(
+                                f"ID:{self.request_id} | Client disconnected after completion"
+                            )
+                        else:
+                            logger.warning(f"ID:{self.request_id} | Client disconnected")
+                            ctx.status_code = 499
+                            ctx.error_message = "client_disconnected"
+                        break
+                    yield chunk
+            finally:
+                check_task.cancel()
+                try:
+                    await check_task
+                except asyncio.CancelledError:
+                    pass
+        except asyncio.CancelledError:
+            # 如果响应已完成，不标记为失败
+            if not ctx.has_completion:
+                ctx.status_code = 499
+                ctx.error_message = "client_disconnected"
             raise
         except Exception as e:
             ctx.status_code = 500
             ctx.error_message = str(e)
             raise
 
+    async def create_smoothed_stream(
+        self,
+        stream_generator: AsyncGenerator[bytes, None],
+    ) -> AsyncGenerator[bytes, None]:
+        """
+        创建平滑输出的流生成器
+
+        如果启用了平滑输出，将大 chunk 拆分成小块并添加微小延迟。
+        否则直接透传原始流。
+
+        Args:
+            stream_generator: 原始流生成器
+
+        Yields:
+            平滑处理后的响应数据块
+        """
+        if not self.smoothing_config.enabled:
+            # 未启用平滑输出，直接透传
+            async for chunk in stream_generator:
+                yield chunk
+            return
+
+        # 启用平滑输出
+        buffer = b""
+        is_first_content = True
+
+        async for chunk in stream_generator:
+            buffer += chunk
+
+            # 按双换行分割 SSE 事件（标准 SSE 格式）
+            while b"\n\n" in buffer:
+                event_block, buffer = buffer.split(b"\n\n", 1)
+                event_str = event_block.decode("utf-8", errors="replace")
+
+                # 解析事件块
+                lines = event_str.strip().split("\n")
+                data_str = None
+                event_type = ""
+
+                for line in lines:
+                    line = line.rstrip("\r")
+                    if line.startswith("event: "):
+                        event_type = line[7:].strip()
+                    elif line.startswith("data: "):
+                        data_str = line[6:]
+
+                # 没有 data 行，直接透传
+                if data_str is None:
+                    yield event_block + b"\n\n"
+                    continue
+
+                # [DONE] 直接透传
+                if data_str.strip() == "[DONE]":
+                    yield event_block + b"\n\n"
+                    continue
+
+                # 尝试解析 JSON
+                try:
+                    data = json.loads(data_str)
+                except json.JSONDecodeError:
+                    yield event_block + b"\n\n"
+                    continue
+
+                # 检测格式并提取内容
+                content, extractor = self._detect_format_and_extract(data)
+
+                # 只有内容长度大于 1 才需要平滑处理
+                if content and len(content) > 1 and extractor:
+                    # 获取配置的延迟
+                    delay_seconds = self._calculate_delay()
+
+                    # 拆分内容
+                    content_chunks = self._split_content(content)
+
+                    for i, sub_content in enumerate(content_chunks):
+                        is_first = is_first_content and i == 0
+
+                        # 使用提取器创建新 chunk
+                        sse_chunk = extractor.create_chunk(
+                            data,
+                            sub_content,
+                            event_type=event_type,
+                            is_first=is_first,
+                        )
+
+                        yield sse_chunk
+
+                        # 除了最后一个块，其他块之间加延迟
+                        if i < len(content_chunks) - 1:
+                            await asyncio.sleep(delay_seconds)
+
+                    is_first_content = False
+                else:
+                    # 不需要拆分，直接透传
+                    yield event_block + b"\n\n"
+                    if content:
+                        is_first_content = False
+
+        # 处理剩余数据
+        if buffer:
+            yield buffer
+
+    def _get_extractor(self, format_name: str) -> Optional[ContentExtractor]:
+        """获取或创建格式对应的提取器（带缓存）"""
+        if format_name not in self._extractors:
+            extractor = get_extractor(format_name)
+            if extractor:
+                self._extractors[format_name] = extractor
+        return self._extractors.get(format_name)
+
+    def _detect_format_and_extract(
+        self, data: dict
+    ) -> tuple[Optional[str], Optional[ContentExtractor]]:
+        """
+        检测数据格式并提取内容
+
+        依次尝试各格式的提取器，返回第一个成功提取内容的结果。
+
+        Returns:
+            (content, extractor): 提取的内容和对应的提取器
+        """
+        for format_name in get_extractor_formats():
+            extractor = self._get_extractor(format_name)
+            if extractor:
+                content = extractor.extract_content(data)
+                if content is not None:
+                    return content, extractor
+
+        return None, None
+
+    def _calculate_delay(self) -> float:
+        """获取配置的延迟（秒）"""
+        return self.smoothing_config.delay_ms / 1000.0
+
+    def _split_content(self, content: str) -> list[str]:
+        """
+        按块拆分文本
+        """
+        chunk_size = self.smoothing_config.chunk_size
+        text_length = len(content)
+
+        if text_length <= chunk_size:
+            return [content]
+
+        # 按块拆分
+        chunks = []
+        for i in range(0, text_length, chunk_size):
+            chunks.append(content[i : i + chunk_size])
+        return chunks
+
     async def _cleanup(
         self,
         response_ctx: Any,
@@ -440,3 +638,128 @@ class StreamProcessor:
             await http_client.aclose()
         except Exception:
             pass
+
+
+async def create_smoothed_stream(
+    stream_generator: AsyncGenerator[bytes, None],
+    chunk_size: int = 20,
+    delay_ms: int = 8,
+) -> AsyncGenerator[bytes, None]:
+    """
+    独立的平滑流生成函数
+
+    供 CLI handler 等场景使用，无需创建完整的 StreamProcessor 实例。
+
+    Args:
+        stream_generator: 原始流生成器
+        chunk_size: 每块字符数
+        delay_ms: 每块之间的延迟毫秒数
+
+    Yields:
+        平滑处理后的响应数据块
+    """
+    processor = _LightweightSmoother(chunk_size=chunk_size, delay_ms=delay_ms)
+    async for chunk in processor.smooth(stream_generator):
+        yield chunk
+
+
+class _LightweightSmoother:
+    """
+    轻量级平滑处理器
+
+    只包含平滑输出所需的最小逻辑，不依赖 StreamProcessor 的其他功能。
+    """
+
+    def __init__(self, chunk_size: int = 20, delay_ms: int = 8) -> None:
+        self.chunk_size = chunk_size
+        self.delay_ms = delay_ms
+        self._extractors: dict[str, ContentExtractor] = {}
+
+    def _get_extractor(self, format_name: str) -> Optional[ContentExtractor]:
+        if format_name not in self._extractors:
+            extractor = get_extractor(format_name)
+            if extractor:
+                self._extractors[format_name] = extractor
+        return self._extractors.get(format_name)
+
+    def _detect_format_and_extract(
+        self, data: dict
+    ) -> tuple[Optional[str], Optional[ContentExtractor]]:
+        for format_name in get_extractor_formats():
+            extractor = self._get_extractor(format_name)
+            if extractor:
+                content = extractor.extract_content(data)
+                if content is not None:
+                    return content, extractor
+        return None, None
+
+    def _calculate_delay(self) -> float:
+        return self.delay_ms / 1000.0
+
+    def _split_content(self, content: str) -> list[str]:
+        text_length = len(content)
+        if text_length <= self.chunk_size:
+            return [content]
+        return [content[i : i + self.chunk_size] for i in range(0, text_length, self.chunk_size)]
+
+    async def smooth(
+        self, stream_generator: AsyncGenerator[bytes, None]
+    ) -> AsyncGenerator[bytes, None]:
+        buffer = b""
+        is_first_content = True
+
+        async for chunk in stream_generator:
+            buffer += chunk
+
+            while b"\n\n" in buffer:
+                event_block, buffer = buffer.split(b"\n\n", 1)
+                event_str = event_block.decode("utf-8", errors="replace")
+
+                lines = event_str.strip().split("\n")
+                data_str = None
+                event_type = ""
+
+                for line in lines:
+                    line = line.rstrip("\r")
+                    if line.startswith("event: "):
+                        event_type = line[7:].strip()
+                    elif line.startswith("data: "):
+                        data_str = line[6:]
+
+                if data_str is None:
+                    yield event_block + b"\n\n"
+                    continue
+
+                if data_str.strip() == "[DONE]":
+                    yield event_block + b"\n\n"
+                    continue
+
+                try:
+                    data = json.loads(data_str)
+                except json.JSONDecodeError:
+                    yield event_block + b"\n\n"
+                    continue
+
+                content, extractor = self._detect_format_and_extract(data)
+
+                if content and len(content) > 1 and extractor:
+                    delay_seconds = self._calculate_delay()
+                    content_chunks = self._split_content(content)
+
+                    for i, sub_content in enumerate(content_chunks):
+                        is_first = is_first_content and i == 0
+                        sse_chunk = extractor.create_chunk(
+                            data, sub_content, event_type=event_type, is_first=is_first
+                        )
+                        yield sse_chunk
+                        if i < len(content_chunks) - 1:
+                            await asyncio.sleep(delay_seconds)
+
+                    is_first_content = False
+                else:
+                    yield event_block + b"\n\n"
+                    if content:
+                        is_first_content = False
+
+        if buffer:
+            yield buffer

src/core/enums.py

@@ -10,8 +10,8 @@ class APIFormat(Enum):
     """API格式枚举 - 决定请求/响应的处理方式"""
 
     CLAUDE = "CLAUDE"  # Claude API 格式
-    OPENAI = "OPENAI"  # OpenAI API 格式
     CLAUDE_CLI = "CLAUDE_CLI"  # Claude CLI API 格式（使用 authorization: Bearer）
+    OPENAI = "OPENAI"  # OpenAI API 格式
     OPENAI_CLI = "OPENAI_CLI"  # OpenAI CLI/Responses API 格式（用于 Claude Code 等客户端）
     GEMINI = "GEMINI"  # Google Gemini API 格式
     GEMINI_CLI = "GEMINI_CLI"  # Gemini CLI API 格式

src/core/exceptions.py

@@ -442,6 +442,36 @@ class EmbeddedErrorException(ProviderException):
         self.error_status = error_status
 
 
+class ProviderCompatibilityException(ProviderException):
+    """Provider 兼容性错误异常 - 应该触发故障转移
+
+    用于处理因 Provider 不支持某些参数或功能导致的错误。
+    这类错误不是用户请求本身的问题，换一个 Provider 可能就能成功，应该触发故障转移。
+
+    常见场景：
+    - Unsupported parameter（不支持的参数）
+    - Unsupported model（不支持的模型）
+    - Unsupported feature（不支持的功能）
+    """
+
+    def __init__(
+        self,
+        message: str,
+        provider_name: Optional[str] = None,
+        status_code: int = 400,
+        upstream_error: Optional[str] = None,
+        request_metadata: Optional[Any] = None,
+    ):
+        self.upstream_error = upstream_error
+        super().__init__(
+            message=message,
+            provider_name=provider_name,
+            request_metadata=request_metadata,
+        )
+        # 覆盖状态码为 400（保持与上游一致）
+        self.status_code = status_code
+
+
 class UpstreamClientException(ProxyException):
     """上游返回的客户端错误异常 - HTTP 4xx 错误，不应该重试
 

src/services/model/service.py

@@ -13,6 +13,7 @@ from src.core.exceptions import InvalidRequestException, NotFoundException
 from src.core.logger import logger
 from src.models.api import ModelCreate, ModelResponse, ModelUpdate
 from src.models.database import Model, Provider
+from src.api.base.models_service import invalidate_models_list_cache
 from src.services.cache.invalidation import get_cache_invalidation_service
 from src.services.cache.model_cache import ModelCacheService
 
@@ -75,6 +76,10 @@ class ModelService:
                 )
 
             logger.info(f"创建模型成功: provider={provider.name}, model={model.provider_model_name}, global_model_id={model.global_model_id}")
+
+            # 清除 /v1/models 列表缓存
+            asyncio.create_task(invalidate_models_list_cache())
+
             return model
 
         except IntegrityError as e:
@@ -197,6 +202,9 @@ class ModelService:
                 cache_service = get_cache_invalidation_service()
                 cache_service.on_model_changed(model.provider_id, model.global_model_id)
 
+            # 清除 /v1/models 列表缓存
+            asyncio.create_task(invalidate_models_list_cache())
+
             logger.info(f"更新模型成功: id={model_id}, 最终 supports_vision: {model.supports_vision}, supports_function_calling: {model.supports_function_calling}, supports_extended_thinking: {model.supports_extended_thinking}")
             return model
         except IntegrityError as e:
@@ -261,6 +269,9 @@ class ModelService:
                 cache_service = get_cache_invalidation_service()
                 cache_service.on_model_changed(cache_info["provider_id"], cache_info["global_model_id"])
 
+            # 清除 /v1/models 列表缓存
+            asyncio.create_task(invalidate_models_list_cache())
+
             logger.info(f"删除模型成功: id={model_id}, provider_model_name={cache_info['provider_model_name']}, "
                 f"global_model_id={cache_info['global_model_id'][:8] if cache_info['global_model_id'] else 'None'}...")
         except Exception as e:
@@ -295,6 +306,9 @@ class ModelService:
             cache_service = get_cache_invalidation_service()
             cache_service.on_model_changed(model.provider_id, model.global_model_id)
 
+        # 清除 /v1/models 列表缓存
+        asyncio.create_task(invalidate_models_list_cache())
+
         status = "可用" if is_available else "不可用"
         logger.info(f"更新模型可用状态: id={model_id}, status={status}")
         return model
@@ -358,6 +372,9 @@ class ModelService:
                 for model in created_models:
                     db.refresh(model)
                 logger.info(f"批量创建 {len(created_models)} 个模型成功")
+
+                # 清除 /v1/models 列表缓存
+                asyncio.create_task(invalidate_models_list_cache())
             except IntegrityError as e:
                 db.rollback()
                 logger.error(f"批量创建模型失败: {str(e)}")

src/services/orchestration/error_classifier.py

@@ -15,6 +15,7 @@ from src.core.enums import APIFormat
 from src.core.exceptions import (
     ConcurrencyLimitError,
     ProviderAuthException,
+    ProviderCompatibilityException,
     ProviderException,
     ProviderNotAvailableException,
     ProviderRateLimitException,
@@ -81,7 +82,9 @@ class ErrorClassifier:
         "context_length_exceeded",  # 上下文长度超限
         "content_length_limit",  # 请求内容长度超限 (Claude API)
         "content_length_exceeds",  # 内容长度超限变体 (AWS CodeWhisperer)
-        "max_tokens",  # token 数超限
+        # 注意：移除了 "max_tokens"，因为 max_tokens 相关错误可能是 Provider 兼容性问题
+        # 如 "Unsupported parameter: 'max_tokens' is not supported with this model"
+        # 这类错误应由 COMPATIBILITY_ERROR_PATTERNS 处理
         "invalid_prompt",  # 无效的提示词
         "content too long",  # 内容过长
         "input is too long",  # 输入过长 (AWS)
@@ -136,6 +139,19 @@ class ErrorClassifier:
         "CONTENT_POLICY_VIOLATION",
     )
 
+    # Provider 兼容性错误模式 - 这类错误应该触发故障转移
+    # 因为换一个 Provider 可能就能成功
+    COMPATIBILITY_ERROR_PATTERNS: Tuple[str, ...] = (
+        "unsupported parameter",  # 不支持的参数
+        "unsupported model",  # 不支持的模型
+        "unsupported feature",  # 不支持的功能
+        "not supported with this model",  # 此模型不支持
+        "model does not support",  # 模型不支持
+        "parameter is not supported",  # 参数不支持
+        "feature is not supported",  # 功能不支持
+        "not available for this model",  # 此模型不可用
+    )
+
     def _parse_error_response(self, error_text: Optional[str]) -> Dict[str, Any]:
         """
         解析错误响应为结构化数据
@@ -261,6 +277,25 @@ class ErrorClassifier:
         search_text = f"{parsed['message']} {parsed['raw']}".lower()
         return any(pattern.lower() in search_text for pattern in self.CLIENT_ERROR_PATTERNS)
 
+    def _is_compatibility_error(self, error_text: Optional[str]) -> bool:
+        """
+        检测错误响应是否为 Provider 兼容性错误（应触发故障转移）
+
+        这类错误是因为 Provider 不支持某些参数或功能导致的，
+        换一个 Provider 可能就能成功。
+
+        Args:
+            error_text: 错误响应文本
+
+        Returns:
+            是否为兼容性错误
+        """
+        if not error_text:
+            return False
+
+        search_text = error_text.lower()
+        return any(pattern.lower() in search_text for pattern in self.COMPATIBILITY_ERROR_PATTERNS)
+
     def _extract_error_message(self, error_text: Optional[str]) -> Optional[str]:
         """
         从错误响应中提取错误消息
@@ -425,6 +460,16 @@ class ErrorClassifier:
                 ),
             )
 
+        # 400 错误：先检查是否为 Provider 兼容性错误（应触发故障转移）
+        if status == 400 and self._is_compatibility_error(error_response_text):
+            logger.info(f"检测到 Provider 兼容性错误，将触发故障转移: {extracted_message}")
+            return ProviderCompatibilityException(
+                message=extracted_message or "Provider 不支持此请求",
+                provider_name=provider_name,
+                status_code=400,
+                upstream_error=error_response_text,
+            )
+
         # 400 错误：检查是否为客户端请求错误（不应重试）
         if status == 400 and self._is_client_error(error_response_text):
             logger.info(f"检测到客户端请求错误，不进行重试: {extracted_message}")

src/services/system/config.py

@@ -12,7 +12,6 @@ from src.core.logger import logger
 from src.models.database import Provider, SystemConfig
 
 
-
 class LogLevel(str, Enum):
     """日志记录级别"""
 
@@ -94,6 +93,35 @@ class SystemConfigService:
 
         return default
 
+    @classmethod
+    def get_configs(cls, db: Session, keys: List[str]) -> Dict[str, Any]:
+        """
+        批量获取系统配置值
+
+        Args:
+            db: 数据库会话
+            keys: 配置键列表
+
+        Returns:
+            配置键值字典
+        """
+        result = {}
+
+        # 一次查询获取所有配置
+        configs = db.query(SystemConfig).filter(SystemConfig.key.in_(keys)).all()
+        config_map = {c.key: c.value for c in configs}
+
+        # 填充结果，不存在的使用默认值
+        for key in keys:
+            if key in config_map:
+                result[key] = config_map[key]
+            elif key in cls.DEFAULT_CONFIGS:
+                result[key] = cls.DEFAULT_CONFIGS[key]["value"]
+            else:
+                result[key] = None
+
+        return result
+
     @staticmethod
     def set_config(db: Session, key: str, value: Any, description: str = None) -> SystemConfig:
         """设置系统配置值"""
@@ -111,6 +139,7 @@ class SystemConfigService:
 
         db.commit()
         db.refresh(config)
+
         return config
 
     @staticmethod
@@ -153,8 +182,8 @@ class SystemConfigService:
             for config in configs
         ]
 
-    @staticmethod
+    @classmethod
-    def delete_config(db: Session, key: str) -> bool:
+    def delete_config(cls, db: Session, key: str) -> bool:
         """删除系统配置"""
         config = db.query(SystemConfig).filter(SystemConfig.key == key).first()
         if config: