Merge branch 'main' into fix/enc_coot

README.md

@@ -39,6 +39,16 @@ Sub2API is an AI API gateway platform designed to distribute and manage API quot
 - **Concurrency Control** - Per-user and per-account concurrency limits
 - **Rate Limiting** - Configurable request and token rate limits
 - **Admin Dashboard** - Web interface for monitoring and management
+- **External System Integration** - Embed external systems (e.g. payment, ticketing) via iframe to extend the admin dashboard
+
+## Ecosystem
+
+Community projects that extend or integrate with Sub2API:
+
+| Project | Description | Features |
+|---------|-------------|----------|
+| [Sub2ApiPay](https://github.com/touwaeriol/sub2apipay) | Self-service payment system | Self-service top-up and subscription purchase; supports YiPay protocol, WeChat Pay, Alipay, Stripe; embeddable via iframe |
+| [sub2api-mobile](https://github.com/ckken/sub2api-mobile) | Mobile admin console | Cross-platform app (iOS/Android/Web) for user management, account management, monitoring dashboard, and multi-backend switching; built with Expo + React Native |
 
 ## Tech Stack
 

README_CN.md

@@ -39,6 +39,16 @@ Sub2API 是一个 AI API 网关平台，用于分发和管理 AI 产品订阅（
 - **并发控制** - 用户级和账号级并发限制
 - **速率限制** - 可配置的请求和 Token 速率限制
 - **管理后台** - Web 界面进行监控和管理
+- **外部系统集成** - 支持通过 iframe 嵌入外部系统（如支付、工单等），扩展管理后台功能
+
+## 生态项目
+
+围绕 Sub2API 的社区扩展与集成项目：
+
+| 项目 | 说明 | 功能 |
+|------|------|------|
+| [Sub2ApiPay](https://github.com/touwaeriol/sub2apipay) | 自助支付系统 | 用户自助充值、自助订阅购买；兼容易支付协议、微信官方支付、支付宝官方支付、Stripe；支持 iframe 嵌入管理后台 |
+| [sub2api-mobile](https://github.com/ckken/sub2api-mobile) | 移动端管理控制台 | 跨平台应用（iOS/Android/Web），支持用户管理、账号管理、监控看板、多后端切换；基于 Expo + React Native 构建 |
 
 ## 技术栈
 

backend/internal/domain/constants.go

@@ -31,8 +31,7 @@ const (
 	AccountTypeSetupToken = "setup-token" // Setup Token类型账号（inference only scope）
 	AccountTypeAPIKey     = "apikey"      // API Key类型账号
 	AccountTypeUpstream   = "upstream"    // 上游透传类型账号（通过 Base URL + API Key 连接上游）
-	AccountTypeBedrock       = "bedrock"        // AWS Bedrock 类型账号（通过 SigV4 签名连接 Bedrock）
-	AccountTypeBedrockAPIKey = "bedrock-apikey" // AWS Bedrock API Key 类型账号（通过 Bearer Token 连接 Bedrock）
+	AccountTypeBedrock    = "bedrock"     // AWS Bedrock 类型账号（通过 SigV4 签名或 API Key 连接 Bedrock，由 credentials.auth_mode 区分）
 )
 
 // Redeem type constants

backend/internal/handler/admin/account_handler.go

@@ -97,7 +97,7 @@ type CreateAccountRequest struct {
 	Name                    string         `json:"name" binding:"required"`
 	Notes                   *string        `json:"notes"`
 	Platform                string         `json:"platform" binding:"required"`
-	Type                    string         `json:"type" binding:"required,oneof=oauth setup-token apikey upstream bedrock bedrock-apikey"`
+	Type                    string         `json:"type" binding:"required,oneof=oauth setup-token apikey upstream bedrock"`
 	Credentials             map[string]any `json:"credentials" binding:"required"`
 	Extra                   map[string]any `json:"extra"`
 	ProxyID                 *int64         `json:"proxy_id"`
@@ -116,7 +116,7 @@ type CreateAccountRequest struct {
 type UpdateAccountRequest struct {
 	Name                    string         `json:"name"`
 	Notes                   *string        `json:"notes"`
-	Type                    string         `json:"type" binding:"omitempty,oneof=oauth setup-token apikey upstream bedrock bedrock-apikey"`
+	Type                    string         `json:"type" binding:"omitempty,oneof=oauth setup-token apikey upstream bedrock"`
 	Credentials             map[string]any `json:"credentials"`
 	Extra                   map[string]any `json:"extra"`
 	ProxyID                 *int64         `json:"proxy_id"`
@@ -1718,13 +1718,12 @@ func (h *AccountHandler) GetAvailableModels(c *gin.Context) {
 
 	// Handle OpenAI accounts
 	if account.IsOpenAI() {
-		// For OAuth accounts: return default OpenAI models
-		if account.IsOAuth() {
+		// OpenAI 自动透传会绕过常规模型改写，测试/模型列表也应回落到默认模型集。
+		if account.IsOpenAIPassthroughEnabled() {
 			response.Success(c, openai.DefaultModels)
 			return
 		}
 
-		// For API Key accounts: check model_mapping
 		mapping := account.GetModelMapping()
 		if len(mapping) == 0 {
 			response.Success(c, openai.DefaultModels)

backend/internal/handler/admin/account_handler_available_models_test.go

+package admin
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/service"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+type availableModelsAdminService struct {
+	*stubAdminService
+	account service.Account
+}
+
+func (s *availableModelsAdminService) GetAccount(_ context.Context, id int64) (*service.Account, error) {
+	if s.account.ID == id {
+		acc := s.account
+		return &acc, nil
+	}
+	return s.stubAdminService.GetAccount(context.Background(), id)
+}
+
+func setupAvailableModelsRouter(adminSvc service.AdminService) *gin.Engine {
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	handler := NewAccountHandler(adminSvc, nil, nil, nil, nil, nil, nil, nil, nil, nil, nil, nil, nil)
+	router.GET("/api/v1/admin/accounts/:id/models", handler.GetAvailableModels)
+	return router
+}
+
+func TestAccountHandlerGetAvailableModels_OpenAIOAuthUsesExplicitModelMapping(t *testing.T) {
+	svc := &availableModelsAdminService{
+		stubAdminService: newStubAdminService(),
+		account: service.Account{
+			ID:       42,
+			Name:     "openai-oauth",
+			Platform: service.PlatformOpenAI,
+			Type:     service.AccountTypeOAuth,
+			Status:   service.StatusActive,
+			Credentials: map[string]any{
+				"model_mapping": map[string]any{
+					"gpt-5": "gpt-5.1",
+				},
+			},
+		},
+	}
+	router := setupAvailableModelsRouter(svc)
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/api/v1/admin/accounts/42/models", nil)
+	router.ServeHTTP(rec, req)
+
+	require.Equal(t, http.StatusOK, rec.Code)
+
+	var resp struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &resp))
+	require.Len(t, resp.Data, 1)
+	require.Equal(t, "gpt-5", resp.Data[0].ID)
+}
+
+func TestAccountHandlerGetAvailableModels_OpenAIOAuthPassthroughFallsBackToDefaults(t *testing.T) {
+	svc := &availableModelsAdminService{
+		stubAdminService: newStubAdminService(),
+		account: service.Account{
+			ID:       43,
+			Name:     "openai-oauth-passthrough",
+			Platform: service.PlatformOpenAI,
+			Type:     service.AccountTypeOAuth,
+			Status:   service.StatusActive,
+			Credentials: map[string]any{
+				"model_mapping": map[string]any{
+					"gpt-5": "gpt-5.1",
+				},
+			},
+			Extra: map[string]any{
+				"openai_passthrough": true,
+			},
+		},
+	}
+	router := setupAvailableModelsRouter(svc)
+
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/api/v1/admin/accounts/43/models", nil)
+	router.ServeHTTP(rec, req)
+
+	require.Equal(t, http.StatusOK, rec.Code)
+
+	var resp struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &resp))
+	require.NotEmpty(t, resp.Data)
+	require.NotEqual(t, "gpt-5", resp.Data[0].ID)
+}

backend/internal/handler/admin/setting_handler.go

@@ -125,6 +125,7 @@ func (h *SettingHandler) GetSettings(c *gin.Context) {
 		OpsMetricsIntervalSeconds:            settings.OpsMetricsIntervalSeconds,
 		MinClaudeCodeVersion:                 settings.MinClaudeCodeVersion,
 		AllowUngroupedKeyScheduling:          settings.AllowUngroupedKeyScheduling,
+		BackendModeEnabled:                   settings.BackendModeEnabled,
 	})
 }
 
@@ -199,6 +200,9 @@ type UpdateSettingsRequest struct {
 
 	// 分组隔离
 	AllowUngroupedKeyScheduling bool `json:"allow_ungrouped_key_scheduling"`
+
+	// Backend Mode
+	BackendModeEnabled bool `json:"backend_mode_enabled"`
 }
 
 // UpdateSettings 更新系统设置
@@ -473,6 +477,7 @@ func (h *SettingHandler) UpdateSettings(c *gin.Context) {
 		IdentityPatchPrompt:              req.IdentityPatchPrompt,
 		MinClaudeCodeVersion:             req.MinClaudeCodeVersion,
 		AllowUngroupedKeyScheduling:      req.AllowUngroupedKeyScheduling,
+		BackendModeEnabled:               req.BackendModeEnabled,
 		OpsMonitoringEnabled: func() bool {
 			if req.OpsMonitoringEnabled != nil {
 				return *req.OpsMonitoringEnabled
@@ -571,6 +576,7 @@ func (h *SettingHandler) UpdateSettings(c *gin.Context) {
 		OpsMetricsIntervalSeconds:            updatedSettings.OpsMetricsIntervalSeconds,
 		MinClaudeCodeVersion:                 updatedSettings.MinClaudeCodeVersion,
 		AllowUngroupedKeyScheduling:          updatedSettings.AllowUngroupedKeyScheduling,
+		BackendModeEnabled:                   updatedSettings.BackendModeEnabled,
 	})
 }
 
@@ -725,6 +731,9 @@ func diffSettings(before *service.SystemSettings, after *service.SystemSettings,
 	if before.AllowUngroupedKeyScheduling != after.AllowUngroupedKeyScheduling {
 		changed = append(changed, "allow_ungrouped_key_scheduling")
 	}
+	if before.BackendModeEnabled != after.BackendModeEnabled {
+		changed = append(changed, "backend_mode_enabled")
+	}
 	if before.PurchaseSubscriptionEnabled != after.PurchaseSubscriptionEnabled {
 		changed = append(changed, "purchase_subscription_enabled")
 	}

backend/internal/handler/auth_handler.go

@@ -194,6 +194,12 @@ func (h *AuthHandler) Login(c *gin.Context) {
 		return
 	}
 
+	// Backend mode: only admin can login
+	if h.settingSvc.IsBackendModeEnabled(c.Request.Context()) && !user.IsAdmin() {
+		response.Forbidden(c, "Backend mode is active. Only admin login is allowed.")
+		return
+	}
+
 	h.respondWithTokenPair(c, user)
 }
 
@@ -250,16 +256,22 @@ func (h *AuthHandler) Login2FA(c *gin.Context) {
 		return
 	}
 
-	// Delete the login session
-	_ = h.totpService.DeleteLoginSession(c.Request.Context(), req.TempToken)
-
-	// Get the user
+	// Get the user (before session deletion so we can check backend mode)
 	user, err := h.userService.GetByID(c.Request.Context(), session.UserID)
 	if err != nil {
 		response.ErrorFrom(c, err)
 		return
 	}
 
+	// Backend mode: only admin can login (check BEFORE deleting session)
+	if h.settingSvc.IsBackendModeEnabled(c.Request.Context()) && !user.IsAdmin() {
+		response.Forbidden(c, "Backend mode is active. Only admin login is allowed.")
+		return
+	}
+
+	// Delete the login session (only after all checks pass)
+	_ = h.totpService.DeleteLoginSession(c.Request.Context(), req.TempToken)
+
 	h.respondWithTokenPair(c, user)
 }
 
@@ -522,16 +534,22 @@ func (h *AuthHandler) RefreshToken(c *gin.Context) {
 		return
 	}
 
-	tokenPair, err := h.authService.RefreshTokenPair(c.Request.Context(), req.RefreshToken)
+	result, err := h.authService.RefreshTokenPair(c.Request.Context(), req.RefreshToken)
 	if err != nil {
 		response.ErrorFrom(c, err)
 		return
 	}
 
+	// Backend mode: block non-admin token refresh
+	if h.settingSvc.IsBackendModeEnabled(c.Request.Context()) && result.UserRole != "admin" {
+		response.Forbidden(c, "Backend mode is active. Only admin login is allowed.")
+		return
+	}
+
 	response.Success(c, RefreshTokenResponse{
-		AccessToken:  tokenPair.AccessToken,
-		RefreshToken: tokenPair.RefreshToken,
-		ExpiresIn:    tokenPair.ExpiresIn,
+		AccessToken:  result.AccessToken,
+		RefreshToken: result.RefreshToken,
+		ExpiresIn:    result.ExpiresIn,
 		TokenType:    "Bearer",
 	})
 }

backend/internal/handler/dto/mappers.go

@@ -264,8 +264,8 @@ func AccountFromServiceShallow(a *service.Account) *Account {
 		}
 	}
 
-	// 提取 API Key 账号配额限制（仅 apikey 类型有效）
-	if a.Type == service.AccountTypeAPIKey {
+	// 提取账号配额限制（apikey / bedrock 类型有效）
+	if a.IsAPIKeyOrBedrock() {
 		if limit := a.GetQuotaLimit(); limit > 0 {
 			out.QuotaLimit = &limit
 			used := a.GetQuotaUsed()
@@ -281,6 +281,31 @@ func AccountFromServiceShallow(a *service.Account) *Account {
 			used := a.GetQuotaWeeklyUsed()
 			out.QuotaWeeklyUsed = &used
 		}
+		// 固定时间重置配置
+		if mode := a.GetQuotaDailyResetMode(); mode == "fixed" {
+			out.QuotaDailyResetMode = &mode
+			hour := a.GetQuotaDailyResetHour()
+			out.QuotaDailyResetHour = &hour
+		}
+		if mode := a.GetQuotaWeeklyResetMode(); mode == "fixed" {
+			out.QuotaWeeklyResetMode = &mode
+			day := a.GetQuotaWeeklyResetDay()
+			out.QuotaWeeklyResetDay = &day
+			hour := a.GetQuotaWeeklyResetHour()
+			out.QuotaWeeklyResetHour = &hour
+		}
+		if a.GetQuotaDailyResetMode() == "fixed" || a.GetQuotaWeeklyResetMode() == "fixed" {
+			tz := a.GetQuotaResetTimezone()
+			out.QuotaResetTimezone = &tz
+		}
+		if a.Extra != nil {
+			if v, ok := a.Extra["quota_daily_reset_at"].(string); ok && v != "" {
+				out.QuotaDailyResetAt = &v
+			}
+			if v, ok := a.Extra["quota_weekly_reset_at"].(string); ok && v != "" {
+				out.QuotaWeeklyResetAt = &v
+			}
+		}
 	}
 
 	return out

backend/internal/handler/dto/settings.go

@@ -81,6 +81,9 @@ type SystemSettings struct {
 
 	// 分组隔离
 	AllowUngroupedKeyScheduling bool `json:"allow_ungrouped_key_scheduling"`
+
+	// Backend Mode
+	BackendModeEnabled bool `json:"backend_mode_enabled"`
 }
 
 type DefaultSubscriptionSetting struct {
@@ -111,6 +114,7 @@ type PublicSettings struct {
 	CustomMenuItems                  []CustomMenuItem `json:"custom_menu_items"`
 	LinuxDoOAuthEnabled              bool             `json:"linuxdo_oauth_enabled"`
 	SoraClientEnabled                bool             `json:"sora_client_enabled"`
+	BackendModeEnabled               bool             `json:"backend_mode_enabled"`
 	Version                          string           `json:"version"`
 }
 

backend/internal/handler/dto/types.go

@@ -203,6 +203,16 @@ type Account struct {
 	QuotaWeeklyLimit *float64 `json:"quota_weekly_limit,omitempty"`
 	QuotaWeeklyUsed  *float64 `json:"quota_weekly_used,omitempty"`
 
+	// 配额固定时间重置配置
+	QuotaDailyResetMode  *string `json:"quota_daily_reset_mode,omitempty"`
+	QuotaDailyResetHour  *int    `json:"quota_daily_reset_hour,omitempty"`
+	QuotaWeeklyResetMode *string `json:"quota_weekly_reset_mode,omitempty"`
+	QuotaWeeklyResetDay  *int    `json:"quota_weekly_reset_day,omitempty"`
+	QuotaWeeklyResetHour *int    `json:"quota_weekly_reset_hour,omitempty"`
+	QuotaResetTimezone   *string `json:"quota_reset_timezone,omitempty"`
+	QuotaDailyResetAt    *string `json:"quota_daily_reset_at,omitempty"`
+	QuotaWeeklyResetAt   *string `json:"quota_weekly_reset_at,omitempty"`
+
 	Proxy         *Proxy         `json:"proxy,omitempty"`
 	AccountGroups []AccountGroup `json:"account_groups,omitempty"`
 

backend/internal/handler/openai_chat_completions.go

@@ -181,13 +181,7 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) {
 		service.SetOpsLatencyMs(c, service.OpsRoutingLatencyMsKey, time.Since(routingStart).Milliseconds())
 		forwardStart := time.Now()
 
-		defaultMappedModel := ""
-		if apiKey.Group != nil {
-			defaultMappedModel = apiKey.Group.DefaultMappedModel
-		}
-		if fallbackModel := c.GetString("openai_chat_completions_fallback_model"); fallbackModel != "" {
-			defaultMappedModel = fallbackModel
-		}
+		defaultMappedModel := c.GetString("openai_chat_completions_fallback_model")
 		result, err := h.gatewayService.ForwardAsChatCompletions(c.Request.Context(), c, account, body, promptCacheKey, defaultMappedModel)
 
 		forwardDurationMs := time.Since(forwardStart).Milliseconds()

backend/internal/handler/openai_gateway_handler.go

@@ -655,14 +655,9 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) {
 		service.SetOpsLatencyMs(c, service.OpsRoutingLatencyMsKey, time.Since(routingStart).Milliseconds())
 		forwardStart := time.Now()
 
-		defaultMappedModel := ""
-		if apiKey.Group != nil {
-			defaultMappedModel = apiKey.Group.DefaultMappedModel
-		}
-		// 如果使用了降级模型调度，强制使用降级模型
-		if fallbackModel := c.GetString("openai_messages_fallback_model"); fallbackModel != "" {
-			defaultMappedModel = fallbackModel
-		}
+		// 仅在调度时实际触发了降级（原模型无可用账号、改用默认模型重试成功）时，
+		// 才将降级模型传给 Forward 层做模型替换；否则保持用户请求的原始模型。
+		defaultMappedModel := c.GetString("openai_messages_fallback_model")
 		result, err := h.gatewayService.ForwardAsAnthropic(c.Request.Context(), c, account, body, promptCacheKey, defaultMappedModel)
 
 		forwardDurationMs := time.Since(forwardStart).Milliseconds()

backend/internal/handler/setting_handler.go

@@ -54,6 +54,7 @@ func (h *SettingHandler) GetPublicSettings(c *gin.Context) {
 		CustomMenuItems:                  dto.ParseUserVisibleMenuItems(settings.CustomMenuItems),
 		LinuxDoOAuthEnabled:              settings.LinuxDoOAuthEnabled,
 		SoraClientEnabled:                settings.SoraClientEnabled,
+		BackendModeEnabled:               settings.BackendModeEnabled,
 		Version:                          h.version,
 	})
 }

backend/internal/pkg/antigravity/client.go

@@ -19,6 +19,16 @@ import (
 	"github.com/Wei-Shaw/sub2api/internal/pkg/proxyutil"
 )
 
+// ForbiddenError 表示上游返回 403 Forbidden
+type ForbiddenError struct {
+	StatusCode int
+	Body       string
+}
+
+func (e *ForbiddenError) Error() string {
+	return fmt.Sprintf("fetchAvailableModels 失败 (HTTP %d): %s", e.StatusCode, e.Body)
+}
+
 // NewAPIRequestWithURL 使用指定的 base URL 创建 Antigravity API 请求（v1internal 端点）
 func NewAPIRequestWithURL(ctx context.Context, baseURL, action, accessToken string, body []byte) (*http.Request, error) {
 	// 构建 URL，流式请求添加 ?alt=sse 参数
@@ -515,6 +525,19 @@ type ModelQuotaInfo struct {
 // ModelInfo 模型信息
 type ModelInfo struct {
 	QuotaInfo          *ModelQuotaInfo `json:"quotaInfo,omitempty"`
+	DisplayName        string          `json:"displayName,omitempty"`
+	SupportsImages     *bool           `json:"supportsImages,omitempty"`
+	SupportsThinking   *bool           `json:"supportsThinking,omitempty"`
+	ThinkingBudget     *int            `json:"thinkingBudget,omitempty"`
+	Recommended        *bool           `json:"recommended,omitempty"`
+	MaxTokens          *int            `json:"maxTokens,omitempty"`
+	MaxOutputTokens    *int            `json:"maxOutputTokens,omitempty"`
+	SupportedMimeTypes map[string]bool `json:"supportedMimeTypes,omitempty"`
+}
+
+// DeprecatedModelInfo 废弃模型转发信息
+type DeprecatedModelInfo struct {
+	NewModelID string `json:"newModelId"`
 }
 
 // FetchAvailableModelsRequest fetchAvailableModels 请求
@@ -525,6 +548,7 @@ type FetchAvailableModelsRequest struct {
 // FetchAvailableModelsResponse fetchAvailableModels 响应
 type FetchAvailableModelsResponse struct {
 	Models             map[string]ModelInfo           `json:"models"`
+	DeprecatedModelIDs map[string]DeprecatedModelInfo `json:"deprecatedModelIds,omitempty"`
 }
 
 // FetchAvailableModels 获取可用模型和配额信息，返回解析后的结构体和原始 JSON
@@ -573,6 +597,13 @@ func (c *Client) FetchAvailableModels(ctx context.Context, accessToken, projectI
 			continue
 		}
 
+		if resp.StatusCode == http.StatusForbidden {
+			return nil, nil, &ForbiddenError{
+				StatusCode: resp.StatusCode,
+				Body:       string(respBodyBytes),
+			}
+		}
+
 		if resp.StatusCode != http.StatusOK {
 			return nil, nil, fmt.Errorf("fetchAvailableModels 失败 (HTTP %d): %s", resp.StatusCode, string(respBodyBytes))
 		}

backend/internal/pkg/apicompat/anthropic_responses_test.go

@@ -105,6 +105,7 @@ func TestAnthropicToResponses_ToolUse(t *testing.T) {
 	assert.Equal(t, "assistant", items[1].Role)
 	assert.Equal(t, "function_call", items[2].Type)
 	assert.Equal(t, "fc_call_1", items[2].CallID)
+	assert.Empty(t, items[2].ID)
 	assert.Equal(t, "function_call_output", items[3].Type)
 	assert.Equal(t, "fc_call_1", items[3].CallID)
 	assert.Equal(t, "Sunny, 72°F", items[3].Output)

backend/internal/pkg/apicompat/anthropic_to_responses.go

@@ -277,7 +277,6 @@ func anthropicAssistantToResponses(raw json.RawMessage) ([]ResponsesInputItem, e
 			CallID:    fcID,
 			Name:      b.Name,
 			Arguments: args,
-			ID:        fcID,
 		})
 	}
 

backend/internal/pkg/apicompat/chatcompletions_responses_test.go

@@ -99,6 +99,7 @@ func TestChatCompletionsToResponses_ToolCalls(t *testing.T) {
 	// Check function_call item
 	assert.Equal(t, "function_call", items[1].Type)
 	assert.Equal(t, "call_1", items[1].CallID)
+	assert.Empty(t, items[1].ID)
 	assert.Equal(t, "ping", items[1].Name)
 
 	// Check function_call_output item
@@ -252,6 +253,55 @@ func TestChatCompletionsToResponses_AssistantWithTextAndToolCalls(t *testing.T)
 	assert.Equal(t, "user", items[0].Role)
 	assert.Equal(t, "assistant", items[1].Role)
 	assert.Equal(t, "function_call", items[2].Type)
+	assert.Empty(t, items[2].ID)
+}
+
+func TestChatCompletionsToResponses_AssistantArrayContentPreserved(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+			{Role: "assistant", Content: json.RawMessage(`[{"type":"text","text":"A"},{"type":"text","text":"B"}]`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+	assert.Equal(t, "assistant", items[1].Role)
+
+	var parts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[1].Content, &parts))
+	require.Len(t, parts, 1)
+	assert.Equal(t, "output_text", parts[0].Type)
+	assert.Equal(t, "AB", parts[0].Text)
+}
+
+func TestChatCompletionsToResponses_AssistantThinkingTagPreserved(t *testing.T) {
+	req := &ChatCompletionsRequest{
+		Model: "gpt-4o",
+		Messages: []ChatMessage{
+			{Role: "user", Content: json.RawMessage(`"Hi"`)},
+			{Role: "assistant", Content: json.RawMessage(`[{"type":"thinking","thinking":"internal plan"},{"type":"text","text":"final answer"}]`)},
+		},
+	}
+
+	resp, err := ChatCompletionsToResponses(req)
+	require.NoError(t, err)
+
+	var items []ResponsesInputItem
+	require.NoError(t, json.Unmarshal(resp.Input, &items))
+	require.Len(t, items, 2)
+
+	var parts []ResponsesContentPart
+	require.NoError(t, json.Unmarshal(items[1].Content, &parts))
+	require.Len(t, parts, 1)
+	assert.Equal(t, "output_text", parts[0].Type)
+	assert.Contains(t, parts[0].Text, "<thinking>internal plan</thinking>")
+	assert.Contains(t, parts[0].Text, "final answer")
 }
 
 // ---------------------------------------------------------------------------
@@ -344,8 +394,8 @@ func TestResponsesToChatCompletions_Reasoning(t *testing.T) {
 
 	var content string
 	require.NoError(t, json.Unmarshal(chat.Choices[0].Message.Content, &content))
-	// Reasoning summary is prepended to text
-	assert.Equal(t, "I thought about it.The answer is 42.", content)
+	assert.Equal(t, "The answer is 42.", content)
+	assert.Equal(t, "I thought about it.", chat.Choices[0].Message.ReasoningContent)
 }
 
 func TestResponsesToChatCompletions_Incomplete(t *testing.T) {
@@ -582,8 +632,35 @@ func TestResponsesEventToChatChunks_ReasoningDelta(t *testing.T) {
 		Delta: "Thinking...",
 	}, state)
 	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.ReasoningContent)
+	assert.Equal(t, "Thinking...", *chunks[0].Choices[0].Delta.ReasoningContent)
+
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.reasoning_summary_text.done",
+	}, state)
+	require.Len(t, chunks, 0)
+}
+
+func TestResponsesEventToChatChunks_ReasoningThenTextAutoCloseTag(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.SentRole = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.reasoning_summary_text.delta",
+		Delta: "plan",
+	}, state)
+	require.Len(t, chunks, 1)
+	require.NotNil(t, chunks[0].Choices[0].Delta.ReasoningContent)
+	assert.Equal(t, "plan", *chunks[0].Choices[0].Delta.ReasoningContent)
+
+	chunks = ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type:  "response.output_text.delta",
+		Delta: "answer",
+	}, state)
+	require.Len(t, chunks, 1)
 	require.NotNil(t, chunks[0].Choices[0].Delta.Content)
-	assert.Equal(t, "Thinking...", *chunks[0].Choices[0].Delta.Content)
+	assert.Equal(t, "answer", *chunks[0].Choices[0].Delta.Content)
 }
 
 func TestFinalizeResponsesChatStream(t *testing.T) {

backend/internal/pkg/apicompat/chatcompletions_to_responses.go

@@ -3,6 +3,7 @@ package apicompat
 import (
 	"encoding/json"
 	"fmt"
+	"strings"
 )
 
 // ChatCompletionsToResponses converts a Chat Completions request into a
@@ -174,8 +175,11 @@ func chatAssistantToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
 
 	// Emit assistant message with output_text if content is non-empty.
 	if len(m.Content) > 0 {
-		var s string
-		if err := json.Unmarshal(m.Content, &s); err == nil && s != "" {
+		s, err := parseAssistantContent(m.Content)
+		if err != nil {
+			return nil, err
+		}
+		if s != "" {
 			parts := []ResponsesContentPart{{Type: "output_text", Text: s}}
 			partsJSON, err := json.Marshal(parts)
 			if err != nil {
@@ -196,13 +200,82 @@ func chatAssistantToResponses(m ChatMessage) ([]ResponsesInputItem, error) {
 			CallID:    tc.ID,
 			Name:      tc.Function.Name,
 			Arguments: args,
-			ID:        tc.ID,
 		})
 	}
 
 	return items, nil
 }
 
+// parseAssistantContent returns assistant content as plain text.
+//
+// Supported formats:
+// - JSON string
+// - JSON array of typed parts (e.g. [{"type":"text","text":"..."}])
+//
+// For structured thinking/reasoning parts, it preserves semantics by wrapping
+// the text in explicit tags so downstream can still distinguish it from normal text.
+func parseAssistantContent(raw json.RawMessage) (string, error) {
+	if len(raw) == 0 {
+		return "", nil
+	}
+
+	var s string
+	if err := json.Unmarshal(raw, &s); err == nil {
+		return s, nil
+	}
+
+	var parts []map[string]any
+	if err := json.Unmarshal(raw, &parts); err != nil {
+		// Keep compatibility with prior behavior: unsupported assistant content
+		// formats are ignored instead of failing the whole request conversion.
+		return "", nil
+	}
+
+	var b strings.Builder
+	write := func(v string) error {
+		_, err := b.WriteString(v)
+		return err
+	}
+	for _, p := range parts {
+		typ, _ := p["type"].(string)
+		text, _ := p["text"].(string)
+		thinking, _ := p["thinking"].(string)
+
+		switch typ {
+		case "thinking", "reasoning":
+			if thinking != "" {
+				if err := write("<thinking>"); err != nil {
+					return "", err
+				}
+				if err := write(thinking); err != nil {
+					return "", err
+				}
+				if err := write("</thinking>"); err != nil {
+					return "", err
+				}
+			} else if text != "" {
+				if err := write("<thinking>"); err != nil {
+					return "", err
+				}
+				if err := write(text); err != nil {
+					return "", err
+				}
+				if err := write("</thinking>"); err != nil {
+					return "", err
+				}
+			}
+		default:
+			if text != "" {
+				if err := write(text); err != nil {
+					return "", err
+				}
+			}
+		}
+	}
+
+	return b.String(), nil
+}
+
 // chatToolToResponses converts a tool result message (role=tool) into a
 // function_call_output item.
 func chatToolToResponses(m ChatMessage) ([]ResponsesInputItem, error) {

backend/internal/pkg/apicompat/responses_to_chatcompletions.go

@@ -29,6 +29,7 @@ func ResponsesToChatCompletions(resp *ResponsesResponse, model string) *ChatComp
 	}
 
 	var contentText string
+	var reasoningText string
 	var toolCalls []ChatToolCall
 
 	for _, item := range resp.Output {
@@ -51,7 +52,7 @@ func ResponsesToChatCompletions(resp *ResponsesResponse, model string) *ChatComp
 		case "reasoning":
 			for _, s := range item.Summary {
 				if s.Type == "summary_text" && s.Text != "" {
-					contentText += s.Text
+					reasoningText += s.Text
 				}
 			}
 		case "web_search_call":
@@ -67,6 +68,9 @@ func ResponsesToChatCompletions(resp *ResponsesResponse, model string) *ChatComp
 		raw, _ := json.Marshal(contentText)
 		msg.Content = raw
 	}
+	if reasoningText != "" {
+		msg.ReasoningContent = reasoningText
+	}
 
 	finishReason := responsesStatusToChatFinishReason(resp.Status, resp.IncompleteDetails, toolCalls)
 
@@ -153,6 +157,8 @@ func ResponsesEventToChatChunks(evt *ResponsesStreamEvent, state *ResponsesEvent
 		return resToChatHandleFuncArgsDelta(evt, state)
 	case "response.reasoning_summary_text.delta":
 		return resToChatHandleReasoningDelta(evt, state)
+	case "response.reasoning_summary_text.done":
+		return nil
 	case "response.completed", "response.incomplete", "response.failed":
 		return resToChatHandleCompleted(evt, state)
 	default:
@@ -276,8 +282,8 @@ func resToChatHandleReasoningDelta(evt *ResponsesStreamEvent, state *ResponsesEv
 	if evt.Delta == "" {
 		return nil
 	}
-	content := evt.Delta
-	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{Content: &content})}
+	reasoning := evt.Delta
+	return []ChatCompletionsChunk{makeChatDeltaChunk(state, ChatDelta{ReasoningContent: &reasoning})}
 }
 
 func resToChatHandleCompleted(evt *ResponsesStreamEvent, state *ResponsesEventToChatState) []ChatCompletionsChunk {

backend/internal/pkg/apicompat/types.go

@@ -363,6 +363,7 @@ type ChatStreamOptions struct {
 type ChatMessage struct {
 	Role             string          `json:"role"` // "system" | "user" | "assistant" | "tool" | "function"
 	Content          json.RawMessage `json:"content,omitempty"`
+	ReasoningContent string          `json:"reasoning_content,omitempty"`
 	Name             string          `json:"name,omitempty"`
 	ToolCalls        []ChatToolCall  `json:"tool_calls,omitempty"`
 	ToolCallID       string          `json:"tool_call_id,omitempty"`
@@ -468,6 +469,7 @@ type ChatChunkChoice struct {
 type ChatDelta struct {
 	Role             string         `json:"role,omitempty"`
 	Content          *string        `json:"content,omitempty"` // pointer: omit when not present, null vs "" matters
+	ReasoningContent *string        `json:"reasoning_content,omitempty"`
 	ToolCalls        []ChatToolCall `json:"tool_calls,omitempty"`
 }