Merge branch 'Wei-Shaw:main' into main

backend/internal/handler/admin/antigravity_oauth_handler.go

@@ -65,3 +65,27 @@ func (h *AntigravityOAuthHandler) ExchangeCode(c *gin.Context) {
 
 	response.Success(c, tokenInfo)
 }
+
+// AntigravityRefreshTokenRequest represents the request for validating Antigravity refresh token
+type AntigravityRefreshTokenRequest struct {
+	RefreshToken string `json:"refresh_token" binding:"required"`
+	ProxyID      *int64 `json:"proxy_id"`
+}
+
+// RefreshToken validates an Antigravity refresh token and returns full token info
+// POST /api/v1/admin/antigravity/oauth/refresh-token
+func (h *AntigravityOAuthHandler) RefreshToken(c *gin.Context) {
+	var req AntigravityRefreshTokenRequest
+	if err := c.ShouldBindJSON(&req); err != nil {
+		response.BadRequest(c, "请求无效: "+err.Error())
+		return
+	}
+
+	tokenInfo, err := h.antigravityOAuthService.ValidateRefreshToken(c.Request.Context(), req.RefreshToken, req.ProxyID)
+	if err != nil {
+		response.ErrorFrom(c, err)
+		return
+	}
+
+	response.Success(c, tokenInfo)
+}

backend/internal/handler/admin/redeem_handler.go

@@ -202,7 +202,7 @@ func (h *RedeemHandler) Export(c *gin.Context) {
 	writer := csv.NewWriter(&buf)
 
 	// Write header
-	if err := writer.Write([]string{"id", "code", "type", "value", "status", "used_by", "used_at", "created_at"}); err != nil {
+	if err := writer.Write([]string{"id", "code", "type", "value", "status", "used_by", "used_by_email", "used_at", "created_at"}); err != nil {
 		response.InternalError(c, "Failed to export redeem codes: "+err.Error())
 		return
 	}
@@ -213,6 +213,10 @@ func (h *RedeemHandler) Export(c *gin.Context) {
 		if code.UsedBy != nil {
 			usedBy = fmt.Sprintf("%d", *code.UsedBy)
 		}
+		usedByEmail := ""
+		if code.User != nil {
+			usedByEmail = code.User.Email
+		}
 		usedAt := ""
 		if code.UsedAt != nil {
 			usedAt = code.UsedAt.Format("2006-01-02 15:04:05")
@@ -224,6 +228,7 @@ func (h *RedeemHandler) Export(c *gin.Context) {
 			fmt.Sprintf("%.2f", code.Value),
 			code.Status,
 			usedBy,
+			usedByEmail,
 			usedAt,
 			code.CreatedAt.Format("2006-01-02 15:04:05"),
 		}); err != nil {

backend/internal/pkg/antigravity/claude_types.go

@@ -27,7 +27,7 @@ type ClaudeMessage struct {
 
 // ThinkingConfig Thinking 配置
 type ThinkingConfig struct {
-	Type         string `json:"type"`                    // "enabled" or "disabled"
+	Type         string `json:"type"`                    // "enabled" / "adaptive" / "disabled"
 	BudgetTokens int    `json:"budget_tokens,omitempty"` // thinking budget
 }
 

backend/internal/pkg/antigravity/gemini_types.go

@@ -155,6 +155,7 @@ type GeminiUsageMetadata struct {
 	CandidatesTokenCount    int `json:"candidatesTokenCount,omitempty"`
 	CachedContentTokenCount int `json:"cachedContentTokenCount,omitempty"`
 	TotalTokenCount         int `json:"totalTokenCount,omitempty"`
+	ThoughtsTokenCount      int `json:"thoughtsTokenCount,omitempty"` // thinking tokens（按输出价格计费）
 }
 
 // GeminiGroundingMetadata Gemini grounding 元数据（Web Search）

backend/internal/pkg/antigravity/request_transformer.go

@@ -64,6 +64,10 @@ const MaxTokensBudgetPadding = 1000
 // Gemini 2.5 Flash thinking budget 上限
 const Gemini25FlashThinkingBudgetLimit = 24576
 
+// 对于 Antigravity 的 Claude（budget-only）模型，该语义最终等价为 thinkingBudget=24576。
+// 这里复用相同数值以保持行为一致。
+const ClaudeAdaptiveHighThinkingBudgetTokens = Gemini25FlashThinkingBudgetLimit
+
 // ensureMaxTokensGreaterThanBudget 确保 max_tokens > budget_tokens
 // Claude API 要求启用 thinking 时，max_tokens 必须大于 thinking.budget_tokens
 // 返回调整后的 maxTokens 和是否进行了调整
@@ -96,7 +100,7 @@ func TransformClaudeToGeminiWithOptions(claudeReq *ClaudeRequest, projectID, map
 	}
 
 	// 检测是否启用 thinking
-	isThinkingEnabled := claudeReq.Thinking != nil && claudeReq.Thinking.Type == "enabled"
+	isThinkingEnabled := claudeReq.Thinking != nil && (claudeReq.Thinking.Type == "enabled" || claudeReq.Thinking.Type == "adaptive")
 
 	// 只有 Gemini 模型支持 dummy thought workaround
 	// Claude 模型通过 Vertex/Google API 需要有效的 thought signatures
@@ -198,8 +202,7 @@ type modelInfo struct {
 
 // modelInfoMap 模型前缀 → 模型信息映射
 // 只有在此映射表中的模型才会注入身份提示词
-// 注意：当前 claude-opus-4-6 会被映射到 claude-opus-4-5-thinking，
-// 但保留此条目以便后续 Antigravity 上游支持 4.6 时快速切换
+// 注意：模型映射逻辑在网关层完成；这里仅用于按模型前缀判断是否注入身份提示词。
 var modelInfoMap = map[string]modelInfo{
 	"claude-opus-4-5":   {DisplayName: "Claude Opus 4.5", CanonicalID: "claude-opus-4-5-20250929"},
 	"claude-opus-4-6":   {DisplayName: "Claude Opus 4.6", CanonicalID: "claude-opus-4-6"},
@@ -593,6 +596,10 @@ func maxOutputTokensLimit(model string) int {
 	return maxOutputTokensUpperBound
 }
 
+func isAntigravityOpus46Model(model string) bool {
+	return strings.HasPrefix(strings.ToLower(model), "claude-opus-4-6")
+}
+
 func buildGenerationConfig(req *ClaudeRequest) *GeminiGenerationConfig {
 	maxLimit := maxOutputTokensLimit(req.Model)
 	config := &GeminiGenerationConfig{
@@ -606,25 +613,36 @@ func buildGenerationConfig(req *ClaudeRequest) *GeminiGenerationConfig {
 	}
 
 	// Thinking 配置
-	if req.Thinking != nil && req.Thinking.Type == "enabled" {
+	if req.Thinking != nil && (req.Thinking.Type == "enabled" || req.Thinking.Type == "adaptive") {
 		config.ThinkingConfig = &GeminiThinkingConfig{
 			IncludeThoughts: true,
 		}
+
+		// - thinking.type=enabled：budget_tokens>0 用显式预算
+		// - thinking.type=adaptive：仅在 Antigravity 的 Opus 4.6 上覆写为 （24576）
+		budget := -1
 		if req.Thinking.BudgetTokens > 0 {
-			budget := req.Thinking.BudgetTokens
+			budget = req.Thinking.BudgetTokens
+		}
+		if req.Thinking.Type == "adaptive" && isAntigravityOpus46Model(req.Model) {
+			budget = ClaudeAdaptiveHighThinkingBudgetTokens
+		}
+
+		// 正预算需要做上限与 max_tokens 约束；动态预算（-1）直接透传给上游。
+		if budget > 0 {
 			// gemini-2.5-flash 上限
 			if strings.Contains(req.Model, "gemini-2.5-flash") && budget > Gemini25FlashThinkingBudgetLimit {
 				budget = Gemini25FlashThinkingBudgetLimit
 			}
-			config.ThinkingConfig.ThinkingBudget = budget
 
-			// 自动修正：max_tokens 必须大于 budget_tokens
+			// 自动修正：max_tokens 必须大于 budget_tokens（Claude 上游要求）
 			if adjusted, ok := ensureMaxTokensGreaterThanBudget(config.MaxOutputTokens, budget); ok {
 				log.Printf("[Antigravity] Auto-adjusted max_tokens from %d to %d (must be > budget_tokens=%d)",
 					config.MaxOutputTokens, adjusted, budget)
 				config.MaxOutputTokens = adjusted
 			}
 		}
+		config.ThinkingConfig.ThinkingBudget = budget
 	}
 
 	if config.MaxOutputTokens > maxLimit {

backend/internal/pkg/antigravity/request_transformer_test.go

@@ -259,3 +259,93 @@ func TestBuildTools_CustomTypeTools(t *testing.T) {
 		})
 	}
 }
+
+func TestBuildGenerationConfig_ThinkingDynamicBudget(t *testing.T) {
+	tests := []struct {
+		name        string
+		model       string
+		thinking    *ThinkingConfig
+		wantBudget  int
+		wantPresent bool
+	}{
+		{
+			name:        "enabled without budget defaults to dynamic (-1)",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    &ThinkingConfig{Type: "enabled"},
+			wantBudget:  -1,
+			wantPresent: true,
+		},
+		{
+			name:        "enabled with budget uses the provided value",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    &ThinkingConfig{Type: "enabled", BudgetTokens: 1024},
+			wantBudget:  1024,
+			wantPresent: true,
+		},
+		{
+			name:        "enabled with -1 budget uses dynamic (-1)",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    &ThinkingConfig{Type: "enabled", BudgetTokens: -1},
+			wantBudget:  -1,
+			wantPresent: true,
+		},
+		{
+			name:        "adaptive on opus4.6 maps to high budget (24576)",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    &ThinkingConfig{Type: "adaptive", BudgetTokens: 20000},
+			wantBudget:  ClaudeAdaptiveHighThinkingBudgetTokens,
+			wantPresent: true,
+		},
+		{
+			name:        "adaptive on non-opus model keeps default dynamic (-1)",
+			model:       "claude-sonnet-4-5-thinking",
+			thinking:    &ThinkingConfig{Type: "adaptive"},
+			wantBudget:  -1,
+			wantPresent: true,
+		},
+		{
+			name:        "disabled does not emit thinkingConfig",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    &ThinkingConfig{Type: "disabled", BudgetTokens: 1024},
+			wantBudget:  0,
+			wantPresent: false,
+		},
+		{
+			name:        "nil thinking does not emit thinkingConfig",
+			model:       "claude-opus-4-6-thinking",
+			thinking:    nil,
+			wantBudget:  0,
+			wantPresent: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			req := &ClaudeRequest{
+				Model:    tt.model,
+				Thinking: tt.thinking,
+			}
+			cfg := buildGenerationConfig(req)
+			if cfg == nil {
+				t.Fatalf("expected non-nil generationConfig")
+			}
+
+			if tt.wantPresent {
+				if cfg.ThinkingConfig == nil {
+					t.Fatalf("expected thinkingConfig to be present")
+				}
+				if !cfg.ThinkingConfig.IncludeThoughts {
+					t.Fatalf("expected includeThoughts=true")
+				}
+				if cfg.ThinkingConfig.ThinkingBudget != tt.wantBudget {
+					t.Fatalf("expected thinkingBudget=%d, got %d", tt.wantBudget, cfg.ThinkingConfig.ThinkingBudget)
+				}
+				return
+			}
+
+			if cfg.ThinkingConfig != nil {
+				t.Fatalf("expected thinkingConfig to be nil, got %+v", cfg.ThinkingConfig)
+			}
+		})
+	}
+}

backend/internal/pkg/antigravity/response_transformer.go

@@ -280,7 +280,7 @@ func (p *NonStreamingProcessor) buildResponse(geminiResp *GeminiResponse, respon
 	if geminiResp.UsageMetadata != nil {
 		cached := geminiResp.UsageMetadata.CachedContentTokenCount
 		usage.InputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
-		usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount
+		usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
 		usage.CacheReadInputTokens = cached
 	}
 

backend/internal/pkg/antigravity/stream_transformer.go

@@ -85,7 +85,7 @@ func (p *StreamingProcessor) ProcessLine(line string) []byte {
 	if geminiResp.UsageMetadata != nil {
 		cached := geminiResp.UsageMetadata.CachedContentTokenCount
 		p.inputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
-		p.outputTokens = geminiResp.UsageMetadata.CandidatesTokenCount
+		p.outputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
 		p.cacheReadTokens = cached
 	}
 
@@ -146,7 +146,7 @@ func (p *StreamingProcessor) emitMessageStart(v1Resp *V1InternalResponse) []byte
 	if v1Resp.Response.UsageMetadata != nil {
 		cached := v1Resp.Response.UsageMetadata.CachedContentTokenCount
 		usage.InputTokens = v1Resp.Response.UsageMetadata.PromptTokenCount - cached
-		usage.OutputTokens = v1Resp.Response.UsageMetadata.CandidatesTokenCount
+		usage.OutputTokens = v1Resp.Response.UsageMetadata.CandidatesTokenCount + v1Resp.Response.UsageMetadata.ThoughtsTokenCount
 		usage.CacheReadInputTokens = cached
 	}
 

backend/internal/repository/account_repo.go

@@ -448,7 +448,12 @@ func (r *accountRepository) ListWithFilters(ctx context.Context, params paginati
 		q = q.Where(dbaccount.TypeEQ(accountType))
 	}
 	if status != "" {
-		q = q.Where(dbaccount.StatusEQ(status))
+		switch status {
+		case "rate_limited":
+			q = q.Where(dbaccount.RateLimitResetAtGT(time.Now()))
+		default:
+			q = q.Where(dbaccount.StatusEQ(status))
+		}
 	}
 	if search != "" {
 		q = q.Where(dbaccount.NameContainsFold(search))

backend/internal/repository/redeem_code_repo.go

@@ -6,6 +6,7 @@ import (
 
 	dbent "github.com/Wei-Shaw/sub2api/ent"
 	"github.com/Wei-Shaw/sub2api/ent/redeemcode"
+	"github.com/Wei-Shaw/sub2api/ent/user"
 	"github.com/Wei-Shaw/sub2api/internal/pkg/pagination"
 	"github.com/Wei-Shaw/sub2api/internal/service"
 )
@@ -106,7 +107,12 @@ func (r *redeemCodeRepository) ListWithFilters(ctx context.Context, params pagin
 		q = q.Where(redeemcode.StatusEQ(status))
 	}
 	if search != "" {
-		q = q.Where(redeemcode.CodeContainsFold(search))
+		q = q.Where(
+			redeemcode.Or(
+				redeemcode.CodeContainsFold(search),
+				redeemcode.HasUserWith(user.EmailContainsFold(search)),
+			),
+		)
 	}
 
 	total, err := q.Count(ctx)

backend/internal/repository/user_repo.go

@@ -10,6 +10,7 @@ import (
 	"time"
 
 	dbent "github.com/Wei-Shaw/sub2api/ent"
+	"github.com/Wei-Shaw/sub2api/ent/apikey"
 	dbuser "github.com/Wei-Shaw/sub2api/ent/user"
 	"github.com/Wei-Shaw/sub2api/ent/userallowedgroup"
 	"github.com/Wei-Shaw/sub2api/ent/usersubscription"
@@ -191,6 +192,7 @@ func (r *userRepository) ListWithFilters(ctx context.Context, params pagination.
 				dbuser.EmailContainsFold(filters.Search),
 				dbuser.UsernameContainsFold(filters.Search),
 				dbuser.NotesContainsFold(filters.Search),
+				dbuser.HasAPIKeysWith(apikey.KeyContainsFold(filters.Search)),
 			),
 		)
 	}

backend/internal/server/routes/admin.go

@@ -281,6 +281,7 @@ func registerAntigravityOAuthRoutes(admin *gin.RouterGroup, h *handler.Handlers)
 	{
 		antigravity.POST("/oauth/auth-url", h.Admin.AntigravityOAuth.GenerateAuthURL)
 		antigravity.POST("/oauth/exchange-code", h.Admin.AntigravityOAuth.ExchangeCode)
+		antigravity.POST("/oauth/refresh-token", h.Admin.AntigravityOAuth.RefreshToken)
 	}
 }
 

backend/internal/service/antigravity_gateway_service.go

@@ -1309,7 +1309,7 @@ func (s *AntigravityGatewayService) Forward(ctx context.Context, c *gin.Context,
 		return nil, s.writeClaudeError(c, http.StatusForbidden, "permission_error", fmt.Sprintf("model %s not in whitelist", claudeReq.Model))
 	}
 	// 应用 thinking 模式自动后缀：如果 thinking 开启且目标是 claude-sonnet-4-5，自动改为 thinking 版本
-	thinkingEnabled := claudeReq.Thinking != nil && claudeReq.Thinking.Type == "enabled"
+	thinkingEnabled := claudeReq.Thinking != nil && (claudeReq.Thinking.Type == "enabled" || claudeReq.Thinking.Type == "adaptive")
 	mappedModel = applyThinkingModelSuffix(mappedModel, thinkingEnabled)
 
 	// 获取 access_token

backend/internal/service/antigravity_gateway_service_test.go

@@ -591,6 +591,75 @@ func TestHandleClaudeStreamingResponse_NormalComplete(t *testing.T) {
 	require.NotContains(t, body, "event: error")
 }
 
+// TestHandleGeminiStreamingResponse_ThoughtsTokenCount
+// 验证：Gemini 流式转发时 thoughtsTokenCount 被计入 OutputTokens
+func TestHandleGeminiStreamingResponse_ThoughtsTokenCount(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	svc := newAntigravityTestService(&config.Config{
+		Gateway: config.GatewayConfig{MaxLineSize: defaultMaxLineSize},
+	})
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/", nil)
+
+	pr, pw := io.Pipe()
+	resp := &http.Response{StatusCode: http.StatusOK, Body: pr, Header: http.Header{}}
+
+	go func() {
+		defer func() { _ = pw.Close() }()
+		fmt.Fprintln(pw, `data: {"candidates":[{"content":{"parts":[{"text":"Hello"}]}}],"usageMetadata":{"promptTokenCount":100,"candidatesTokenCount":20,"thoughtsTokenCount":50}}`)
+		fmt.Fprintln(pw, "")
+		fmt.Fprintln(pw, `data: {"candidates":[{"content":{"parts":[{"text":" world"}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":100,"candidatesTokenCount":30,"thoughtsTokenCount":80,"cachedContentTokenCount":10}}`)
+		fmt.Fprintln(pw, "")
+	}()
+
+	result, err := svc.handleGeminiStreamingResponse(c, resp, time.Now())
+	_ = pr.Close()
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.NotNil(t, result.usage)
+	// promptTokenCount=100, cachedContentTokenCount=10 → InputTokens=90
+	require.Equal(t, 90, result.usage.InputTokens)
+	// candidatesTokenCount=30 + thoughtsTokenCount=80 → OutputTokens=110
+	require.Equal(t, 110, result.usage.OutputTokens)
+	require.Equal(t, 10, result.usage.CacheReadInputTokens)
+}
+
+// TestHandleClaudeStreamingResponse_ThoughtsTokenCount
+// 验证：Gemini→Claude 流式转换时 thoughtsTokenCount 被计入 OutputTokens
+func TestHandleClaudeStreamingResponse_ThoughtsTokenCount(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	svc := newAntigravityTestService(&config.Config{
+		Gateway: config.GatewayConfig{MaxLineSize: defaultMaxLineSize},
+	})
+
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/", nil)
+
+	pr, pw := io.Pipe()
+	resp := &http.Response{StatusCode: http.StatusOK, Body: pr, Header: http.Header{}}
+
+	go func() {
+		defer func() { _ = pw.Close() }()
+		fmt.Fprintln(pw, `data: {"response":{"candidates":[{"content":{"parts":[{"text":"Hi"}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":50,"candidatesTokenCount":10,"thoughtsTokenCount":25}}}`)
+		fmt.Fprintln(pw, "")
+	}()
+
+	result, err := svc.handleClaudeStreamingResponse(c, resp, time.Now(), "gemini-2.5-pro")
+	_ = pr.Close()
+
+	require.NoError(t, err)
+	require.NotNil(t, result)
+	require.NotNil(t, result.usage)
+	// promptTokenCount=50 → InputTokens=50
+	require.Equal(t, 50, result.usage.InputTokens)
+	// candidatesTokenCount=10 + thoughtsTokenCount=25 → OutputTokens=35
+	require.Equal(t, 35, result.usage.OutputTokens)
+}
+
 // --- 流式客户端断开检测测试 ---
 
 // TestStreamUpstreamResponse_ClientDisconnectDrainsUsage

backend/internal/service/antigravity_oauth_service.go

@@ -192,6 +192,43 @@ func (s *AntigravityOAuthService) RefreshToken(ctx context.Context, refreshToken
 	return nil, fmt.Errorf("token 刷新失败 (重试后): %w", lastErr)
 }
 
+// ValidateRefreshToken 用 refresh token 验证并获取完整的 token 信息（含 email 和 project_id）
+func (s *AntigravityOAuthService) ValidateRefreshToken(ctx context.Context, refreshToken string, proxyID *int64) (*AntigravityTokenInfo, error) {
+	var proxyURL string
+	if proxyID != nil {
+		proxy, err := s.proxyRepo.GetByID(ctx, *proxyID)
+		if err == nil && proxy != nil {
+			proxyURL = proxy.URL()
+		}
+	}
+
+	// 刷新 token
+	tokenInfo, err := s.RefreshToken(ctx, refreshToken, proxyURL)
+	if err != nil {
+		return nil, err
+	}
+
+	// 获取用户信息（email）
+	client := antigravity.NewClient(proxyURL)
+	userInfo, err := client.GetUserInfo(ctx, tokenInfo.AccessToken)
+	if err != nil {
+		fmt.Printf("[AntigravityOAuth] 警告: 获取用户信息失败: %v\n", err)
+	} else {
+		tokenInfo.Email = userInfo.Email
+	}
+
+	// 获取 project_id（容错，失败不阻塞）
+	projectID, loadErr := s.loadProjectIDWithRetry(ctx, tokenInfo.AccessToken, proxyURL, 3)
+	if loadErr != nil {
+		fmt.Printf("[AntigravityOAuth] 警告: 获取 project_id 失败（重试后）: %v\n", loadErr)
+		tokenInfo.ProjectIDMissing = true
+	} else {
+		tokenInfo.ProjectID = projectID
+	}
+
+	return tokenInfo, nil
+}
+
 func isNonRetryableAntigravityOAuthError(err error) bool {
 	msg := err.Error()
 	nonRetryable := []string{

backend/internal/service/gateway_request.go

@@ -101,9 +101,9 @@ func ParseGatewayRequest(body []byte, protocol string) (*ParsedRequest, error) {
 		}
 	}
 
-	// thinking: {type: "enabled"}
+	// thinking: {type: "enabled" | "adaptive"}
 	if rawThinking, ok := req["thinking"].(map[string]any); ok {
-		if t, ok := rawThinking["type"].(string); ok && t == "enabled" {
+		if t, ok := rawThinking["type"].(string); ok && (t == "enabled" || t == "adaptive") {
 			parsed.ThinkingEnabled = true
 		}
 	}
@@ -161,9 +161,9 @@ func parseIntegralNumber(raw any) (int, bool) {
 // Returns filtered body or original body if filtering fails (fail-safe)
 // This prevents 400 errors from invalid thinking block signatures
 //
-// Strategy:
-//   - When thinking.type != "enabled": Remove all thinking blocks
-//   - When thinking.type == "enabled": Only remove thinking blocks without valid signatures
+// 策略：
+//   - 当 thinking.type 不是 "enabled"/"adaptive"：移除所有 thinking 相关块
+//   - 当 thinking.type 是 "enabled"/"adaptive"：仅移除缺失/无效 signature 的 thinking 块（避免 400）
 //     (blocks with missing/empty/dummy signatures that would cause 400 errors)
 func FilterThinkingBlocks(body []byte) []byte {
 	return filterThinkingBlocksInternal(body, false)
@@ -489,9 +489,9 @@ func FilterSignatureSensitiveBlocksForRetry(body []byte) []byte {
 }
 
 // filterThinkingBlocksInternal removes invalid thinking blocks from request
-// Strategy:
-//   - When thinking.type != "enabled": Remove all thinking blocks
-//   - When thinking.type == "enabled": Only remove thinking blocks without valid signatures
+// 策略：
+//   - 当 thinking.type 不是 "enabled"/"adaptive"：移除所有 thinking 相关块
+//   - 当 thinking.type 是 "enabled"/"adaptive"：仅移除缺失/无效 signature 的 thinking 块
 func filterThinkingBlocksInternal(body []byte, _ bool) []byte {
 	// Fast path: if body doesn't contain "thinking", skip parsing
 	if !bytes.Contains(body, []byte(`"type":"thinking"`)) &&
@@ -511,7 +511,7 @@ func filterThinkingBlocksInternal(body []byte, _ bool) []byte {
 	// Check if thinking is enabled
 	thinkingEnabled := false
 	if thinking, ok := req["thinking"].(map[string]any); ok {
-		if thinkType, ok := thinking["type"].(string); ok && thinkType == "enabled" {
+		if thinkType, ok := thinking["type"].(string); ok && (thinkType == "enabled" || thinkType == "adaptive") {
 			thinkingEnabled = true
 		}
 	}

backend/internal/service/gateway_request_test.go

@@ -29,6 +29,14 @@ func TestParseGatewayRequest_ThinkingEnabled(t *testing.T) {
 	require.True(t, parsed.ThinkingEnabled)
 }
 
+func TestParseGatewayRequest_ThinkingAdaptiveEnabled(t *testing.T) {
+	body := []byte(`{"model":"claude-sonnet-4-5","thinking":{"type":"adaptive"},"messages":[{"content":"hi"}]}`)
+	parsed, err := ParseGatewayRequest(body, "")
+	require.NoError(t, err)
+	require.Equal(t, "claude-sonnet-4-5", parsed.Model)
+	require.True(t, parsed.ThinkingEnabled)
+}
+
 func TestParseGatewayRequest_MaxTokens(t *testing.T) {
 	body := []byte(`{"model":"claude-haiku-4-5","max_tokens":1}`)
 	parsed, err := ParseGatewayRequest(body, "")
@@ -209,6 +217,16 @@ func TestFilterThinkingBlocks(t *testing.T) {
 			input:        `{"model":"claude-3-5-sonnet-20241022","messages":[{"role":"user","content":[{"type":"text","text":"Hello"},{"type":"thinking","thinking":"internal","signature":"invalid"},{"type":"text","text":"World"}]}]}`,
 			shouldFilter: true,
 		},
+		{
+			name:         "does not filter signed thinking blocks when thinking adaptive",
+			input:        `{"thinking":{"type":"adaptive"},"messages":[{"role":"assistant","content":[{"type":"thinking","thinking":"ok","signature":"sig_real_123"},{"type":"text","text":"B"}]}]}`,
+			shouldFilter: false,
+		},
+		{
+			name:         "filters unsigned thinking blocks when thinking adaptive",
+			input:        `{"thinking":{"type":"adaptive"},"messages":[{"role":"assistant","content":[{"type":"thinking","thinking":"internal","signature":""},{"type":"text","text":"B"}]}]}`,
+			shouldFilter: true,
+		},
 		{
 			name:         "handles no thinking blocks",
 			input:        `{"model":"claude-3-5-sonnet-20241022","messages":[{"role":"user","content":[{"type":"text","text":"Hello"}]}]}`,

backend/internal/service/gateway_service.go

@@ -3631,7 +3631,8 @@ func requestNeedsBetaFeatures(body []byte) bool {
 	if tools.Exists() && tools.IsArray() && len(tools.Array()) > 0 {
 		return true
 	}
-	if strings.EqualFold(gjson.GetBytes(body, "thinking.type").String(), "enabled") {
+	thinkingType := gjson.GetBytes(body, "thinking.type").String()
+	if strings.EqualFold(thinkingType, "enabled") || strings.EqualFold(thinkingType, "adaptive") {
 		return true
 	}
 	return false

backend/internal/service/gemini_messages_compat_service.go

@@ -2663,11 +2663,12 @@ func extractGeminiUsage(geminiResp map[string]any) *ClaudeUsage {
 	prompt, _ := asInt(usageMeta["promptTokenCount"])
 	cand, _ := asInt(usageMeta["candidatesTokenCount"])
 	cached, _ := asInt(usageMeta["cachedContentTokenCount"])
+	thoughts, _ := asInt(usageMeta["thoughtsTokenCount"])
 	// 注意：Gemini 的 promptTokenCount 包含 cachedContentTokenCount，
 	// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens，需要减去
 	return &ClaudeUsage{
 		InputTokens:          prompt - cached,
-		OutputTokens:         cand,
+		OutputTokens:         cand + thoughts,
 		CacheReadInputTokens: cached,
 	}
 }

backend/internal/service/gemini_messages_compat_service_test.go

@@ -4,6 +4,8 @@ import (
 	"encoding/json"
 	"strings"
 	"testing"
+
+	"github.com/stretchr/testify/require"
 )
 
 // TestConvertClaudeToolsToGeminiTools_CustomType 测试custom类型工具转换
@@ -203,3 +205,70 @@ func TestEnsureGeminiFunctionCallThoughtSignatures_InsertsWhenMissing(t *testing
 		t.Fatalf("expected injected thoughtSignature %q, got: %s", geminiDummyThoughtSignature, s)
 	}
 }
+
+func TestExtractGeminiUsage_ThoughtsTokenCount(t *testing.T) {
+	tests := []struct {
+		name          string
+		resp          map[string]any
+		wantInput     int
+		wantOutput    int
+		wantCacheRead int
+		wantNil       bool
+	}{
+		{
+			name: "with thoughtsTokenCount",
+			resp: map[string]any{
+				"usageMetadata": map[string]any{
+					"promptTokenCount":     float64(100),
+					"candidatesTokenCount": float64(20),
+					"thoughtsTokenCount":   float64(50),
+				},
+			},
+			wantInput:  100,
+			wantOutput: 70,
+		},
+		{
+			name: "with thoughtsTokenCount and cache",
+			resp: map[string]any{
+				"usageMetadata": map[string]any{
+					"promptTokenCount":        float64(100),
+					"candidatesTokenCount":    float64(20),
+					"cachedContentTokenCount": float64(30),
+					"thoughtsTokenCount":      float64(50),
+				},
+			},
+			wantInput:     70,
+			wantOutput:    70,
+			wantCacheRead: 30,
+		},
+		{
+			name: "without thoughtsTokenCount (old model)",
+			resp: map[string]any{
+				"usageMetadata": map[string]any{
+					"promptTokenCount":     float64(100),
+					"candidatesTokenCount": float64(20),
+				},
+			},
+			wantInput:  100,
+			wantOutput: 20,
+		},
+		{
+			name:    "no usageMetadata",
+			resp:    map[string]any{},
+			wantNil: true,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			usage := extractGeminiUsage(tt.resp)
+			if tt.wantNil {
+				require.Nil(t, usage)
+				return
+			}
+			require.NotNil(t, usage)
+			require.Equal(t, tt.wantInput, usage.InputTokens)
+			require.Equal(t, tt.wantOutput, usage.OutputTokens)
+			require.Equal(t, tt.wantCacheRead, usage.CacheReadInputTokens)
+		})
+	}
+}