feat: MODEL_CAPACITY_EXHAUSTED 使用固定1s间隔重试60次，不切换账号

MODEL_CAPACITY_EXHAUSTED (503) 表示模型容量不足，所有账号共享同一容量池，
切换账号无意义。改为固定1s间隔重试最多60次，重试耗尽后直接返回上游错误。

- 新增 antigravityModelCapacityRetryMaxAttempts=60 和 antigravityModelCapacityRetryWait=1s
- shouldTriggerAntigravitySmartRetry 新增 isModelCapacityExhausted 返回值
- handleSmartRetry 对 MODEL_CAPACITY_EXHAUSTED 使用独立重试策略
- handleModelRateLimit 对 MODEL_CAPACITY_EXHAUSTED 仅标记 Handled，不设限流
- 重试耗尽后不设置模型限流、不清除粘性会话、不切换账号

backend/internal/service/antigravity_gateway_service.go

@@ -39,6 +39,12 @@ const (
 	antigravitySmartRetryMaxAttempts    = 1                // 智能重试最大次数（仅重试 1 次，防止重复限流/长期等待）
 	antigravityDefaultRateLimitDuration = 30 * time.Second // 默认限流时间（无 retryDelay 时使用）
 
+	// MODEL_CAPACITY_EXHAUSTED 专用重试参数
+	// 模型容量不足时，所有账号共享同一容量池，切换账号无意义
+	// 使用固定 1s 间隔重试，最多重试 60 次
+	antigravityModelCapacityRetryMaxAttempts = 60
+	antigravityModelCapacityRetryWait        = 1 * time.Second
+
 	// Google RPC 状态和类型常量
 	googleRPCStatusResourceExhausted      = "RESOURCE_EXHAUSTED"
 	googleRPCStatusUnavailable            = "UNAVAILABLE"
@@ -144,7 +150,7 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 	}
 
 	// 判断是否触发智能重试
-	shouldSmartRetry, shouldRateLimitModel, waitDuration, modelName := shouldTriggerAntigravitySmartRetry(p.account, respBody)
+	shouldSmartRetry, shouldRateLimitModel, waitDuration, modelName, isModelCapacityExhausted := shouldTriggerAntigravitySmartRetry(p.account, respBody)
 
 	// 情况1: retryDelay >= 阈值，限流模型并切换账号
 	if shouldRateLimitModel {
@@ -174,14 +180,21 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 		}
 	}
 
-	// 情况2: retryDelay < 阈值，智能重试（最多 antigravitySmartRetryMaxAttempts 次）
+	// 情况2: retryDelay < 阈值（或 MODEL_CAPACITY_EXHAUSTED），智能重试
 	if shouldSmartRetry {
 		var lastRetryResp *http.Response
 		var lastRetryBody []byte
 
-		for attempt := 1; attempt <= antigravitySmartRetryMaxAttempts; attempt++ {
+		// MODEL_CAPACITY_EXHAUSTED 使用独立的重试参数（60 次，固定 1s 间隔）
+		maxAttempts := antigravitySmartRetryMaxAttempts
+		if isModelCapacityExhausted {
+			maxAttempts = antigravityModelCapacityRetryMaxAttempts
+			waitDuration = antigravityModelCapacityRetryWait
+		}
+
+		for attempt := 1; attempt <= maxAttempts; attempt++ {
 			log.Printf("%s status=%d oauth_smart_retry attempt=%d/%d delay=%v model=%s account=%d",
-				p.prefix, resp.StatusCode, attempt, antigravitySmartRetryMaxAttempts, waitDuration, modelName, p.account.ID)
+				p.prefix, resp.StatusCode, attempt, maxAttempts, waitDuration, modelName, p.account.ID)
 
 			select {
 			case <-p.ctx.Done():
@@ -207,13 +220,13 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 
 			retryResp, retryErr := p.httpUpstream.Do(retryReq, p.proxyURL, p.account.ID, p.account.Concurrency)
 			if retryErr == nil && retryResp != nil && retryResp.StatusCode != http.StatusTooManyRequests && retryResp.StatusCode != http.StatusServiceUnavailable {
-				log.Printf("%s status=%d smart_retry_success attempt=%d/%d", p.prefix, retryResp.StatusCode, attempt, antigravitySmartRetryMaxAttempts)
+				log.Printf("%s status=%d smart_retry_success attempt=%d/%d", p.prefix, retryResp.StatusCode, attempt, maxAttempts)
 				return &smartRetryResult{action: smartRetryActionBreakWithResp, resp: retryResp}
 			}
 
 			// 网络错误时，继续重试
 			if retryErr != nil || retryResp == nil {
-				log.Printf("%s status=smart_retry_network_error attempt=%d/%d error=%v", p.prefix, attempt, antigravitySmartRetryMaxAttempts, retryErr)
+				log.Printf("%s status=smart_retry_network_error attempt=%d/%d error=%v", p.prefix, attempt, maxAttempts, retryErr)
 				continue
 			}
 
@@ -227,26 +240,43 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 				_ = retryResp.Body.Close()
 			}
 
-			// 解析新的重试信息，用于下次重试的等待时间
+			// 解析新的重试信息，用于下次重试的等待时间（MODEL_CAPACITY_EXHAUSTED 使用固定循环，跳过）
-			if attempt < antigravitySmartRetryMaxAttempts && lastRetryBody != nil {
+			if !isModelCapacityExhausted && attempt < maxAttempts && lastRetryBody != nil {
-				newShouldRetry, _, newWaitDuration, _ := shouldTriggerAntigravitySmartRetry(p.account, lastRetryBody)
+				newShouldRetry, _, newWaitDuration, _, _ := shouldTriggerAntigravitySmartRetry(p.account, lastRetryBody)
 				if newShouldRetry && newWaitDuration > 0 {
 					waitDuration = newWaitDuration
 				}
 			}
 		}
 
-		// 所有重试都失败，限流当前模型并切换账号
+		// 所有重试都失败
-		rateLimitDuration := waitDuration
-		if rateLimitDuration <= 0 {
-			rateLimitDuration = antigravityDefaultRateLimitDuration
-		}
 		retryBody := lastRetryBody
 		if retryBody == nil {
 			retryBody = respBody
 		}
+
+		// MODEL_CAPACITY_EXHAUSTED：模型容量不足，切换账号无意义
+		// 直接返回上游错误响应，不设置模型限流，不切换账号
+		if isModelCapacityExhausted {
+			log.Printf("%s status=%d smart_retry_exhausted_model_capacity attempts=%d model=%s account=%d body=%s (model capacity exhausted, not switching account)",
+				p.prefix, resp.StatusCode, maxAttempts, modelName, p.account.ID, truncateForLog(retryBody, 200))
+			return &smartRetryResult{
+				action: smartRetryActionBreakWithResp,
+				resp: &http.Response{
+					StatusCode: resp.StatusCode,
+					Header:     resp.Header.Clone(),
+					Body:       io.NopCloser(bytes.NewReader(retryBody)),
+				},
+			}
+		}
+
+		// RATE_LIMIT_EXCEEDED：账号级限流，限流当前模型并切换账号
+		rateLimitDuration := waitDuration
+		if rateLimitDuration <= 0 {
+			rateLimitDuration = antigravityDefaultRateLimitDuration
+		}
 		log.Printf("%s status=%d smart_retry_exhausted attempts=%d model=%s account=%d upstream_retry_delay=%v body=%s (switch account)",
-			p.prefix, resp.StatusCode, antigravitySmartRetryMaxAttempts, modelName, p.account.ID, rateLimitDuration, truncateForLog(retryBody, 200))
+			p.prefix, resp.StatusCode, maxAttempts, modelName, p.account.ID, rateLimitDuration, truncateForLog(retryBody, 200))
 
 		resetAt := time.Now().Add(rateLimitDuration)
 		if p.accountRepo != nil && modelName != "" {
@@ -2053,8 +2083,9 @@ func antigravityFallbackCooldownSeconds() (time.Duration, bool) {
 
 // antigravitySmartRetryInfo 智能重试所需的信息
 type antigravitySmartRetryInfo struct {
-	RetryDelay time.Duration // 重试延迟时间
+	RetryDelay               time.Duration // 重试延迟时间
-	ModelName  string        // 限流的模型名称（如 "claude-sonnet-4-5"）
+	ModelName                string        // 限流的模型名称（如 "claude-sonnet-4-5"）
+	IsModelCapacityExhausted bool          // 是否为模型容量不足（MODEL_CAPACITY_EXHAUSTED）
 }
 
 // parseAntigravitySmartRetryInfo 解析 Google RPC RetryInfo 和 ErrorInfo 信息
@@ -2169,31 +2200,40 @@ func parseAntigravitySmartRetryInfo(body []byte) *antigravitySmartRetryInfo {
 	}
 
 	return &antigravitySmartRetryInfo{
-		RetryDelay: retryDelay,
+		RetryDelay:               retryDelay,
-		ModelName:  modelName,
+		ModelName:                modelName,
+		IsModelCapacityExhausted: hasModelCapacityExhausted,
 	}
 }
 
 // shouldTriggerAntigravitySmartRetry 判断是否应该触发智能重试
 // 返回：
-//   - shouldRetry: 是否应该智能重试（retryDelay < antigravityRateLimitThreshold）
+//   - shouldRetry: 是否应该智能重试（retryDelay < antigravityRateLimitThreshold，或 MODEL_CAPACITY_EXHAUSTED）
-//   - shouldRateLimitModel: 是否应该限流模型（retryDelay >= antigravityRateLimitThreshold）
+//   - shouldRateLimitModel: 是否应该限流模型并切换账号（仅 RATE_LIMIT_EXCEEDED 且 retryDelay >= 阈值）
-//   - waitDuration: 等待时间（智能重试时使用，shouldRateLimitModel=true 时为 0）
+//   - waitDuration: 等待时间
 //   - modelName: 限流的模型名称
-func shouldTriggerAntigravitySmartRetry(account *Account, respBody []byte) (shouldRetry bool, shouldRateLimitModel bool, waitDuration time.Duration, modelName string) {
+//   - isModelCapacityExhausted: 是否为模型容量不足（MODEL_CAPACITY_EXHAUSTED）
+func shouldTriggerAntigravitySmartRetry(account *Account, respBody []byte) (shouldRetry bool, shouldRateLimitModel bool, waitDuration time.Duration, modelName string, isModelCapacityExhausted bool) {
 	if account.Platform != PlatformAntigravity {
-		return false, false, 0, ""
+		return false, false, 0, "", false
 	}
 
 	info := parseAntigravitySmartRetryInfo(respBody)
 	if info == nil {
-		return false, false, 0, ""
+		return false, false, 0, "", false
 	}
 
+	// MODEL_CAPACITY_EXHAUSTED（模型容量不足）：所有账号共享同一模型容量池
+	// 切换账号无意义，使用固定 1s 间隔重试
+	if info.IsModelCapacityExhausted {
+		return true, false, antigravityModelCapacityRetryWait, info.ModelName, true
+	}
+
+	// RATE_LIMIT_EXCEEDED（账号级限流）：
 	// retryDelay >= 阈值：直接限流模型，不重试
 	// 注意：如果上游未提供 retryDelay，parseAntigravitySmartRetryInfo 已设置为默认 30s
 	if info.RetryDelay >= antigravityRateLimitThreshold {
-		return false, true, info.RetryDelay, info.ModelName
+		return false, true, info.RetryDelay, info.ModelName, false
 	}
 
 	// retryDelay < 阈值：智能重试
@@ -2202,7 +2242,7 @@ func shouldTriggerAntigravitySmartRetry(account *Account, respBody []byte) (shou
 		waitDuration = antigravitySmartRetryMinWait
 	}
 
-	return true, false, waitDuration, info.ModelName
+	return true, false, waitDuration, info.ModelName, false
 }
 
 // handleModelRateLimitParams 模型级限流处理参数
@@ -2228,8 +2268,9 @@ type handleModelRateLimitResult struct {
 
 // handleModelRateLimit 处理模型级限流（在原有逻辑之前调用）
 // 仅处理 429/503，解析模型名和 retryDelay
-// - retryDelay < antigravityRateLimitThreshold: 返回 ShouldRetry=true，由调用方等待后重试
+// - MODEL_CAPACITY_EXHAUSTED: 返回 Handled=true（实际重试由 handleSmartRetry 处理）
-// - retryDelay >= antigravityRateLimitThreshold: 设置模型限流 + 清除粘性会话 + 返回 SwitchError
+// - RATE_LIMIT_EXCEEDED + retryDelay < 阈值: 返回 ShouldRetry=true，由调用方等待后重试
+// - RATE_LIMIT_EXCEEDED + retryDelay >= 阈值: 设置模型限流 + 清除粘性会话 + 返回 SwitchError
 func (s *AntigravityGatewayService) handleModelRateLimit(p *handleModelRateLimitParams) *handleModelRateLimitResult {
 	if p.statusCode != 429 && p.statusCode != 503 {
 		return &handleModelRateLimitResult{Handled: false}
@@ -2240,7 +2281,17 @@ func (s *AntigravityGatewayService) handleModelRateLimit(p *handleModelRateLimit
 		return &handleModelRateLimitResult{Handled: false}
 	}
 
-	// < antigravityRateLimitThreshold: 等待后重试
+	// MODEL_CAPACITY_EXHAUSTED：模型容量不足，所有账号共享同一容量池
+	// 切换账号无意义，不设置模型限流（实际重试由 handleSmartRetry 处理）
+	if info.IsModelCapacityExhausted {
+		log.Printf("%s status=%d model_capacity_exhausted model=%s (not switching account, retry handled by smart retry)",
+			p.prefix, p.statusCode, info.ModelName)
+		return &handleModelRateLimitResult{
+			Handled: true,
+		}
+	}
+
+	// RATE_LIMIT_EXCEEDED: < antigravityRateLimitThreshold: 等待后重试
 	if info.RetryDelay < antigravityRateLimitThreshold {
 		log.Printf("%s status=%d model_rate_limit_wait model=%s wait=%v",
 			p.prefix, p.statusCode, info.ModelName, info.RetryDelay)
@@ -2251,7 +2302,7 @@ func (s *AntigravityGatewayService) handleModelRateLimit(p *handleModelRateLimit
 		}
 	}
 
-	// >= antigravityRateLimitThreshold: 设置限流 + 清除粘性会话 + 切换账号
+	// RATE_LIMIT_EXCEEDED: >= antigravityRateLimitThreshold: 设置限流 + 清除粘性会话 + 切换账号
 	s.setModelRateLimitAndClearSession(p, info)
 
 	return &handleModelRateLimitResult{

backend/internal/service/antigravity_rate_limit_test.go

@@ -188,13 +188,14 @@ func TestHandleUpstreamError_429_NonModelRateLimit(t *testing.T) {
 	require.Equal(t, "claude-sonnet-4-5", repo.modelRateLimitCalls[0].modelKey)
 }
 
-// TestHandleUpstreamError_503_ModelRateLimit 测试 503 模型限流场景
+// TestHandleUpstreamError_503_ModelCapacityExhausted 测试 503 模型容量不足场景
-func TestHandleUpstreamError_503_ModelRateLimit(t *testing.T) {
+// MODEL_CAPACITY_EXHAUSTED 时应等待重试，不切换账号
+func TestHandleUpstreamError_503_ModelCapacityExhausted(t *testing.T) {
 	repo := &stubAntigravityAccountRepo{}
 	svc := &AntigravityGatewayService{accountRepo: repo}
 	account := &Account{ID: 3, Name: "acc-3", Platform: PlatformAntigravity}
 
-	// 503 + MODEL_CAPACITY_EXHAUSTED → 模型限流
+	// 503 + MODEL_CAPACITY_EXHAUSTED → 等待重试，不切换账号
 	body := []byte(`{
 		"error": {
 			"status": "UNAVAILABLE",
@@ -207,13 +208,13 @@ func TestHandleUpstreamError_503_ModelRateLimit(t *testing.T) {
 
 	result := svc.handleUpstreamError(context.Background(), "[test]", account, http.StatusServiceUnavailable, http.Header{}, body, "gemini-3-pro-high", 0, "", false)
 
-	// 应该触发模型限流
+	// MODEL_CAPACITY_EXHAUSTED 应该标记为已处理，不切换账号，不设置模型限流
+	// 实际重试由 handleSmartRetry 处理
 	require.NotNil(t, result)
 	require.True(t, result.Handled)
-	require.NotNil(t, result.SwitchError)
+	require.False(t, result.ShouldRetry, "MODEL_CAPACITY_EXHAUSTED should not trigger retry from handleModelRateLimit path")
-	require.Equal(t, "gemini-3-pro-high", result.SwitchError.RateLimitedModel)
+	require.Nil(t, result.SwitchError, "MODEL_CAPACITY_EXHAUSTED should not trigger account switch")
-	require.Len(t, repo.modelRateLimitCalls, 1)
+	require.Empty(t, repo.modelRateLimitCalls, "MODEL_CAPACITY_EXHAUSTED should not set model rate limit")
-	require.Equal(t, "gemini-3-pro-high", repo.modelRateLimitCalls[0].modelKey)
 }
 
 // TestHandleUpstreamError_503_NonModelRateLimit 测试 503 非模型限流场景（不处理）
@@ -301,11 +302,12 @@ func TestParseGeminiRateLimitResetTime_QuotaResetDelay_RoundsUp(t *testing.T) {
 
 func TestParseAntigravitySmartRetryInfo(t *testing.T) {
 	tests := []struct {
-		name          string
+		name                             string
-		body          string
+		body                             string
-		expectedDelay time.Duration
+		expectedDelay                    time.Duration
-		expectedModel string
+		expectedModel                    string
-		expectedNil   bool
+		expectedNil                      bool
+		expectedIsModelCapacityExhausted bool
 	}{
 		{
 			name: "valid complete response with RATE_LIMIT_EXCEEDED",
@@ -368,8 +370,9 @@ func TestParseAntigravitySmartRetryInfo(t *testing.T) {
 					"message": "No capacity available for model gemini-3-pro-high on the server"
 				}
 			}`,
-			expectedDelay: 39 * time.Second,
+			expectedDelay:                    39 * time.Second,
-			expectedModel: "gemini-3-pro-high",
+			expectedModel:                    "gemini-3-pro-high",
+			expectedIsModelCapacityExhausted: true,
 		},
 		{
 			name: "503 UNAVAILABLE without MODEL_CAPACITY_EXHAUSTED - should return nil",
@@ -480,6 +483,9 @@ func TestParseAntigravitySmartRetryInfo(t *testing.T) {
 			if result.ModelName != tt.expectedModel {
 				t.Errorf("ModelName = %q, want %q", result.ModelName, tt.expectedModel)
 			}
+			if result.IsModelCapacityExhausted != tt.expectedIsModelCapacityExhausted {
+				t.Errorf("IsModelCapacityExhausted = %v, want %v", result.IsModelCapacityExhausted, tt.expectedIsModelCapacityExhausted)
+			}
 		})
 	}
 }
@@ -491,13 +497,14 @@ func TestShouldTriggerAntigravitySmartRetry(t *testing.T) {
 	apiKeyAccount := &Account{Type: AccountTypeAPIKey}
 
 	tests := []struct {
-		name                    string
+		name                             string
-		account                 *Account
+		account                          *Account
-		body                    string
+		body                             string
-		expectedShouldRetry     bool
+		expectedShouldRetry              bool
-		expectedShouldRateLimit bool
+		expectedShouldRateLimit          bool
-		minWait                 time.Duration
+		expectedIsModelCapacityExhausted bool
-		modelName               string
+		minWait                          time.Duration
+		modelName                        string
 	}{
 		{
 			name:    "OAuth account with short delay (< 7s) - smart retry",
@@ -611,13 +618,14 @@ func TestShouldTriggerAntigravitySmartRetry(t *testing.T) {
 					]
 				}
 			}`,
-			expectedShouldRetry:     false,
+			expectedShouldRetry:              true,
-			expectedShouldRateLimit: true,
+			expectedShouldRateLimit:          false,
-			minWait:                 39 * time.Second,
+			expectedIsModelCapacityExhausted: true,
-			modelName:               "gemini-3-pro-high",
+			minWait:                          1 * time.Second,
+			modelName:                        "gemini-3-pro-high",
 		},
 		{
-			name:    "503 UNAVAILABLE with MODEL_CAPACITY_EXHAUSTED - no retryDelay - use default rate limit",
+			name:    "503 UNAVAILABLE with MODEL_CAPACITY_EXHAUSTED - no retryDelay - use fixed wait",
 			account: oauthAccount,
 			body: `{
 				"error": {
@@ -629,10 +637,11 @@ func TestShouldTriggerAntigravitySmartRetry(t *testing.T) {
 					"message": "No capacity available for model gemini-2.5-flash on the server"
 				}
 			}`,
-			expectedShouldRetry:     false,
+			expectedShouldRetry:              true,
-			expectedShouldRateLimit: true,
+			expectedShouldRateLimit:          false,
-			minWait:                 30 * time.Second,
+			expectedIsModelCapacityExhausted: true,
-			modelName:               "gemini-2.5-flash",
+			minWait:                          1 * time.Second,
+			modelName:                        "gemini-2.5-flash",
 		},
 		{
 			name:    "429 RESOURCE_EXHAUSTED with RATE_LIMIT_EXCEEDED - no retryDelay - use default rate limit",
@@ -656,13 +665,16 @@ func TestShouldTriggerAntigravitySmartRetry(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			shouldRetry, shouldRateLimit, wait, model := shouldTriggerAntigravitySmartRetry(tt.account, []byte(tt.body))
+			shouldRetry, shouldRateLimit, wait, model, isModelCapacityExhausted := shouldTriggerAntigravitySmartRetry(tt.account, []byte(tt.body))
 			if shouldRetry != tt.expectedShouldRetry {
 				t.Errorf("shouldRetry = %v, want %v", shouldRetry, tt.expectedShouldRetry)
 			}
 			if shouldRateLimit != tt.expectedShouldRateLimit {
 				t.Errorf("shouldRateLimit = %v, want %v", shouldRateLimit, tt.expectedShouldRateLimit)
 			}
+			if isModelCapacityExhausted != tt.expectedIsModelCapacityExhausted {
+				t.Errorf("isModelCapacityExhausted = %v, want %v", isModelCapacityExhausted, tt.expectedIsModelCapacityExhausted)
+			}
 			if shouldRetry {
 				if wait < tt.minWait {
 					t.Errorf("wait = %v, want >= %v", wait, tt.minWait)

backend/internal/service/antigravity_smart_retry_test.go

@@ -294,8 +294,9 @@ func TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError(t *test
 	require.Len(t, upstream.calls, 1, "should have made one retry call (max attempts)")
 }
 
-// TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError 测试 503 MODEL_CAPACITY_EXHAUSTED 返回 switchError
+// TestHandleSmartRetry_503_ModelCapacityExhausted_RetrySuccess 测试 503 MODEL_CAPACITY_EXHAUSTED 重试成功
-func TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError(t *testing.T) {
+// MODEL_CAPACITY_EXHAUSTED 使用固定 1s 间隔重试，不切换账号
+func TestHandleSmartRetry_503_ModelCapacityExhausted_RetrySuccess(t *testing.T) {
 	repo := &stubAntigravityAccountRepo{}
 	account := &Account{
 		ID:       3,
@@ -304,7 +305,7 @@ func TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError(t *testi
 		Platform: PlatformAntigravity,
 	}
 
-	// 503 + MODEL_CAPACITY_EXHAUSTED + 39s >= 7s 阈值
+	// 503 + MODEL_CAPACITY_EXHAUSTED + 39s（上游 retryDelay 应被忽略，使用固定 1s）
 	respBody := []byte(`{
 		"error": {
 			"code": 503,
@@ -322,6 +323,14 @@ func TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError(t *testi
 		Body:       io.NopCloser(bytes.NewReader(respBody)),
 	}
 
+	// mock: 第 1 次重试返回 200 成功
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{
+			{StatusCode: http.StatusOK, Header: http.Header{}, Body: io.NopCloser(strings.NewReader(`{"ok":true}`))},
+		},
+		errors: []error{nil},
+	}
+
 	params := antigravityRetryLoopParams{
 		ctx:             context.Background(),
 		prefix:          "[test]",
@@ -330,6 +339,7 @@ func TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError(t *testi
 		action:          "generateContent",
 		body:            []byte(`{"input":"test"}`),
 		accountRepo:     repo,
+		httpUpstream:    upstream,
 		isStickySession: true,
 		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, requestedModel string, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
 			return nil
@@ -343,16 +353,67 @@ func TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError(t *testi
 
 	require.NotNil(t, result)
 	require.Equal(t, smartRetryActionBreakWithResp, result.action)
-	require.Nil(t, result.resp)
+	require.NotNil(t, result.resp, "should return successful response")
+	require.Equal(t, http.StatusOK, result.resp.StatusCode)
 	require.Nil(t, result.err)
-	require.NotNil(t, result.switchError, "should return switchError for 503 model capacity exhausted")
+	require.Nil(t, result.switchError, "MODEL_CAPACITY_EXHAUSTED should not return switchError")
-	require.Equal(t, account.ID, result.switchError.OriginalAccountID)
-	require.Equal(t, "gemini-3-pro-high", result.switchError.RateLimitedModel)
-	require.True(t, result.switchError.IsStickySession)
 
-	// 验证模型限流已设置
+	// 不应设置模型限流
-	require.Len(t, repo.modelRateLimitCalls, 1)
+	require.Empty(t, repo.modelRateLimitCalls, "MODEL_CAPACITY_EXHAUSTED should not set model rate limit")
-	require.Equal(t, "gemini-3-pro-high", repo.modelRateLimitCalls[0].modelKey)
+	require.Len(t, upstream.calls, 1, "should have made one retry call before success")
+}
+
+// TestHandleSmartRetry_503_ModelCapacityExhausted_ContextCancel 测试 MODEL_CAPACITY_EXHAUSTED 上下文取消
+func TestHandleSmartRetry_503_ModelCapacityExhausted_ContextCancel(t *testing.T) {
+	repo := &stubAntigravityAccountRepo{}
+	account := &Account{
+		ID:       3,
+		Name:     "acc-3",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"code": 503,
+			"status": "UNAVAILABLE",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro-high"}, "reason": "MODEL_CAPACITY_EXHAUSTED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "39s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusServiceUnavailable,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	// 立即取消上下文，验证重试循环能正确退出
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	params := antigravityRetryLoopParams{
+		ctx:         ctx,
+		prefix:      "[test]",
+		account:     account,
+		accessToken: "token",
+		action:      "generateContent",
+		body:        []byte(`{"input":"test"}`),
+		accountRepo: repo,
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, requestedModel string, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	svc := &AntigravityGatewayService{}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, []string{"https://ag-1.test"})
+
+	require.NotNil(t, result)
+	require.Equal(t, smartRetryActionBreakWithResp, result.action)
+	require.Error(t, result.err, "should return context error")
+	require.Nil(t, result.switchError, "should not return switchError on context cancel")
+	require.Empty(t, repo.modelRateLimitCalls, "should not set model rate limit on context cancel")
 }
 
 // TestHandleSmartRetry_NonAntigravityAccount_ContinuesDefaultLogic 测试非 Antigravity 平台账号走默认逻辑
@@ -1129,20 +1190,20 @@ func TestHandleSmartRetry_ShortDelay_NetworkError_StickySession_ClearsSession(t
 }
 
 // TestHandleSmartRetry_ShortDelay_503_StickySession_FailedRetry_ClearsSession
-// 503 + 短延迟 + 粘性会话 + 重试失败 → 清除粘性绑定
+// 429 + 短延迟 + 粘性会话 + 重试失败 → 清除粘性绑定
 func TestHandleSmartRetry_ShortDelay_503_StickySession_FailedRetry_ClearsSession(t *testing.T) {
 	failRespBody := `{
 		"error": {
-			"code": 503,
+			"code": 429,
-			"status": "UNAVAILABLE",
+			"status": "RESOURCE_EXHAUSTED",
 			"details": [
-				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "MODEL_CAPACITY_EXHAUSTED"},
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "RATE_LIMIT_EXCEEDED"},
 				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.5s"}
 			]
 		}
 	}`
 	failResp := &http.Response{
-		StatusCode: http.StatusServiceUnavailable,
+		StatusCode: http.StatusTooManyRequests,
 		Header:     http.Header{},
 		Body:       io.NopCloser(strings.NewReader(failRespBody)),
 	}
@@ -1162,16 +1223,16 @@ func TestHandleSmartRetry_ShortDelay_503_StickySession_FailedRetry_ClearsSession
 
 	respBody := []byte(`{
 		"error": {
-			"code": 503,
+			"code": 429,
-			"status": "UNAVAILABLE",
+			"status": "RESOURCE_EXHAUSTED",
 			"details": [
-				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "MODEL_CAPACITY_EXHAUSTED"},
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "RATE_LIMIT_EXCEEDED"},
 				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.5s"}
 			]
 		}
 	}`)
 	resp := &http.Response{
-		StatusCode: http.StatusServiceUnavailable,
+		StatusCode: http.StatusTooManyRequests,
 		Header:     http.Header{},
 		Body:       io.NopCloser(bytes.NewReader(respBody)),
 	}