Merge branch 'router-for-me:main' into main

fix(translator): pass through adaptive thinking effort

config.example.yaml

@@ -219,6 +219,17 @@ nonstream-keepalive-interval: 0
 #     models: # The models supported by the provider.
 #       - name: "moonshotai/kimi-k2:free" # The actual model name.
 #         alias: "kimi-k2" # The alias used in the API.
+#       # You may repeat the same alias to build an internal model pool.
+#       # The client still sees only one alias in the model list.
+#       # Requests to that alias will round-robin across the upstream names below,
+#       # and if the chosen upstream fails before producing output, the request will
+#       # continue with the next upstream model in the same alias pool.
+#       - name: "qwen3.5-plus"
+#         alias: "claude-opus-4.66"
+#       - name: "glm-5"
+#         alias: "claude-opus-4.66"
+#       - name: "kimi-k2.5"
+#         alias: "claude-opus-4.66"
 
 # Vertex API keys (Vertex-compatible endpoints, use API key + base URL)
 # vertex-api-key:

internal/api/handlers/management/auth_files.go

@@ -1312,12 +1312,12 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			projects, errAll := onboardAllGeminiProjects(ctx, gemClient, &ts)
 			if errAll != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errAll)
-				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to complete Gemini CLI onboarding: %v", errAll))
 				return
 			}
 			if errVerify := ensureGeminiProjectsEnabled(ctx, gemClient, projects); errVerify != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errVerify)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errVerify))
 				return
 			}
 			ts.ProjectID = strings.Join(projects, ",")
@@ -1326,7 +1326,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			ts.Auto = false
 			if errSetup := performGeminiCLISetup(ctx, gemClient, &ts, ""); errSetup != nil {
 				log.Errorf("Google One auto-discovery failed: %v", errSetup)
-				SetOAuthSessionError(state, "Google One auto-discovery failed")
+				SetOAuthSessionError(state, fmt.Sprintf("Google One auto-discovery failed: %v", errSetup))
 				return
 			}
 			if strings.TrimSpace(ts.ProjectID) == "" {
@@ -1337,19 +1337,19 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
 			if errCheck != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errCheck))
 				return
 			}
 			ts.Checked = isChecked
 			if !isChecked {
 				log.Error("Cloud AI API is not enabled for the auto-discovered project")
-				SetOAuthSessionError(state, "Cloud AI API not enabled")
+				SetOAuthSessionError(state, fmt.Sprintf("Cloud AI API not enabled for project %s", ts.ProjectID))
 				return
 			}
 		} else {
 			if errEnsure := ensureGeminiProjectAndOnboard(ctx, gemClient, &ts, requestedProjectID); errEnsure != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errEnsure)
-				SetOAuthSessionError(state, "Failed to complete Gemini CLI onboarding")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to complete Gemini CLI onboarding: %v", errEnsure))
 				return
 			}
 
@@ -1362,13 +1362,13 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
 			if errCheck != nil {
 				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
-				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				SetOAuthSessionError(state, fmt.Sprintf("Failed to verify Cloud AI API status: %v", errCheck))
 				return
 			}
 			ts.Checked = isChecked
 			if !isChecked {
 				log.Error("Cloud AI API is not enabled for the selected project")
-				SetOAuthSessionError(state, "Cloud AI API not enabled")
+				SetOAuthSessionError(state, fmt.Sprintf("Cloud AI API not enabled for project %s", ts.ProjectID))
 				return
 			}
 		}

internal/registry/model_definitions_static_data.go

@@ -211,6 +211,21 @@ func GetGeminiModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-image-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-image-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Image Preview",
+			Description:                "Gemini 3.1 Flash Image Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -220,12 +235,27 @@ func GetGeminiModels() []*ModelInfo {
 			Name:                       "models/gemini-3-flash-preview",
 			Version:                    "3.0",
 			DisplayName:                "Gemini 3 Flash Preview",
-			Description:                "Gemini 3 Flash Preview",
+			Description:                "Our most intelligent model built for speed, combining frontier intelligence with superior search and grounding.",
 			InputTokenLimit:            1048576,
 			OutputTokenLimit:           65536,
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -336,6 +366,32 @@ func GetGeminiVertexModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-image-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-image-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Image Preview",
+			Description:                "Gemini 3.1 Flash Image Preview",
+		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -508,6 +564,21 @@ func GetGeminiCLIModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 	}
 }
 
@@ -604,6 +675,21 @@ func GetAIStudioModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
+		{
+			ID:                         "gemini-3.1-flash-lite-preview",
+			Object:                     "model",
+			Created:                    1776288000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-flash-lite-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Flash Lite Preview",
+			Description:                "Our smallest and most cost effective model, built for at scale usage.",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}},
+		},
 		{
 			ID:                         "gemini-pro-latest",
 			Object:                     "model",
@@ -980,6 +1066,7 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-3.1-pro-high":      {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-pro-low":       {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-flash-image":   {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}}},
+		"gemini-3.1-flash-lite-preview": {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}}},
 		"gemini-3-flash":           {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-opus-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-sonnet-4-6":        {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},

internal/registry/model_registry.go

@@ -64,6 +64,11 @@ type ModelInfo struct {
 	UserDefined bool `json:"-"`
 }
 
+type availableModelsCacheEntry struct {
+	models    []map[string]any
+	expiresAt time.Time
+}
+
 // ThinkingSupport describes a model family's supported internal reasoning budget range.
 // Values are interpreted in provider-native token units.
 type ThinkingSupport struct {
@@ -118,6 +123,8 @@ type ModelRegistry struct {
 	clientProviders map[string]string
 	// mutex ensures thread-safe access to the registry
 	mutex *sync.RWMutex
+	// availableModelsCache stores per-handler snapshots for GetAvailableModels.
+	availableModelsCache map[string]availableModelsCacheEntry
 	// hook is an optional callback sink for model registration changes
 	hook ModelRegistryHook
 }
@@ -130,15 +137,28 @@ var registryOnce sync.Once
 func GetGlobalRegistry() *ModelRegistry {
 	registryOnce.Do(func() {
 		globalRegistry = &ModelRegistry{
-			models:           make(map[string]*ModelRegistration),
-			clientModels:     make(map[string][]string),
-			clientModelInfos: make(map[string]map[string]*ModelInfo),
-			clientProviders:  make(map[string]string),
-			mutex:            &sync.RWMutex{},
+			models:               make(map[string]*ModelRegistration),
+			clientModels:         make(map[string][]string),
+			clientModelInfos:     make(map[string]map[string]*ModelInfo),
+			clientProviders:      make(map[string]string),
+			availableModelsCache: make(map[string]availableModelsCacheEntry),
+			mutex:                &sync.RWMutex{},
 		}
 	})
 	return globalRegistry
 }
+func (r *ModelRegistry) ensureAvailableModelsCacheLocked() {
+	if r.availableModelsCache == nil {
+		r.availableModelsCache = make(map[string]availableModelsCacheEntry)
+	}
+}
+
+func (r *ModelRegistry) invalidateAvailableModelsCacheLocked() {
+	if len(r.availableModelsCache) == 0 {
+		return
+	}
+	clear(r.availableModelsCache)
+}
 
 // LookupModelInfo searches dynamic registry (provider-specific > global) then static definitions.
 func LookupModelInfo(modelID string, provider ...string) *ModelInfo {
@@ -153,9 +173,9 @@ func LookupModelInfo(modelID string, provider ...string) *ModelInfo {
 	}
 
 	if info := GetGlobalRegistry().GetModelInfo(modelID, p); info != nil {
-		return info
+		return cloneModelInfo(info)
 	}
-	return LookupStaticModelInfo(modelID)
+	return cloneModelInfo(LookupStaticModelInfo(modelID))
 }
 
 // SetHook sets an optional hook for observing model registration changes.
@@ -213,6 +233,7 @@ func (r *ModelRegistry) triggerModelsUnregistered(provider, clientID string) {
 func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models []*ModelInfo) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	provider := strings.ToLower(clientProvider)
 	uniqueModelIDs := make([]string, 0, len(models))
@@ -238,6 +259,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		delete(r.clientModels, clientID)
 		delete(r.clientModelInfos, clientID)
 		delete(r.clientProviders, clientID)
+		r.invalidateAvailableModelsCacheLocked()
 		misc.LogCredentialSeparator()
 		return
 	}
@@ -265,6 +287,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		} else {
 			delete(r.clientProviders, clientID)
 		}
+		r.invalidateAvailableModelsCacheLocked()
 		r.triggerModelsRegistered(provider, clientID, models)
 		log.Debugf("Registered client %s from provider %s with %d models", clientID, clientProvider, len(rawModelIDs))
 		misc.LogCredentialSeparator()
@@ -408,6 +431,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		delete(r.clientProviders, clientID)
 	}
 
+	r.invalidateAvailableModelsCacheLocked()
 	r.triggerModelsRegistered(provider, clientID, models)
 	if len(added) == 0 && len(removed) == 0 && !providerChanged {
 		// Only metadata (e.g., display name) changed; skip separator when no log output.
@@ -511,6 +535,13 @@ func cloneModelInfo(model *ModelInfo) *ModelInfo {
 	if len(model.SupportedOutputModalities) > 0 {
 		copyModel.SupportedOutputModalities = append([]string(nil), model.SupportedOutputModalities...)
 	}
+	if model.Thinking != nil {
+		copyThinking := *model.Thinking
+		if len(model.Thinking.Levels) > 0 {
+			copyThinking.Levels = append([]string(nil), model.Thinking.Levels...)
+		}
+		copyModel.Thinking = &copyThinking
+	}
 	return &copyModel
 }
 
@@ -540,6 +571,7 @@ func (r *ModelRegistry) UnregisterClient(clientID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
 	r.unregisterClientInternal(clientID)
+	r.invalidateAvailableModelsCacheLocked()
 }
 
 // unregisterClientInternal performs the actual client unregistration (internal, no locking)
@@ -606,9 +638,12 @@ func (r *ModelRegistry) unregisterClientInternal(clientID string) {
 func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	if registration, exists := r.models[modelID]; exists {
-		registration.QuotaExceededClients[clientID] = new(time.Now())
+		now := time.Now()
+		registration.QuotaExceededClients[clientID] = &now
+		r.invalidateAvailableModelsCacheLocked()
 		log.Debugf("Marked model %s as quota exceeded for client %s", modelID, clientID)
 	}
 }
@@ -620,9 +655,11 @@ func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 func (r *ModelRegistry) ClearModelQuotaExceeded(clientID, modelID string) {
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	if registration, exists := r.models[modelID]; exists {
 		delete(registration.QuotaExceededClients, clientID)
+		r.invalidateAvailableModelsCacheLocked()
 		// log.Debugf("Cleared quota exceeded status for model %s and client %s", modelID, clientID)
 	}
 }
@@ -638,6 +675,7 @@ func (r *ModelRegistry) SuspendClientModel(clientID, modelID, reason string) {
 	}
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	registration, exists := r.models[modelID]
 	if !exists || registration == nil {
@@ -651,6 +689,7 @@ func (r *ModelRegistry) SuspendClientModel(clientID, modelID, reason string) {
 	}
 	registration.SuspendedClients[clientID] = reason
 	registration.LastUpdated = time.Now()
+	r.invalidateAvailableModelsCacheLocked()
 	if reason != "" {
 		log.Debugf("Suspended client %s for model %s: %s", clientID, modelID, reason)
 	} else {
@@ -668,6 +707,7 @@ func (r *ModelRegistry) ResumeClientModel(clientID, modelID string) {
 	}
 	r.mutex.Lock()
 	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
 
 	registration, exists := r.models[modelID]
 	if !exists || registration == nil || registration.SuspendedClients == nil {
@@ -678,6 +718,7 @@ func (r *ModelRegistry) ResumeClientModel(clientID, modelID string) {
 	}
 	delete(registration.SuspendedClients, clientID)
 	registration.LastUpdated = time.Now()
+	r.invalidateAvailableModelsCacheLocked()
 	log.Debugf("Resumed client %s for model %s", clientID, modelID)
 }
 
@@ -713,22 +754,52 @@ func (r *ModelRegistry) ClientSupportsModel(clientID, modelID string) bool {
 // Returns:
 //   - []map[string]any: List of available models in the requested format
 func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any {
-	r.mutex.RLock()
-	defer r.mutex.RUnlock()
+	now := time.Now()
 
-	models := make([]map[string]any, 0)
+	r.mutex.RLock()
+	if cache, ok := r.availableModelsCache[handlerType]; ok && (cache.expiresAt.IsZero() || now.Before(cache.expiresAt)) {
+		models := cloneModelMaps(cache.models)
+		r.mutex.RUnlock()
+		return models
+	}
+	r.mutex.RUnlock()
+
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+	r.ensureAvailableModelsCacheLocked()
+
+	if cache, ok := r.availableModelsCache[handlerType]; ok && (cache.expiresAt.IsZero() || now.Before(cache.expiresAt)) {
+		return cloneModelMaps(cache.models)
+	}
+
+	models, expiresAt := r.buildAvailableModelsLocked(handlerType, now)
+	r.availableModelsCache[handlerType] = availableModelsCacheEntry{
+		models:    cloneModelMaps(models),
+		expiresAt: expiresAt,
+	}
+
+	return models
+}
+
+func (r *ModelRegistry) buildAvailableModelsLocked(handlerType string, now time.Time) ([]map[string]any, time.Time) {
+	models := make([]map[string]any, 0, len(r.models))
 	quotaExpiredDuration := 5 * time.Minute
+	var expiresAt time.Time
 
 	for _, registration := range r.models {
-		// Check if model has any non-quota-exceeded clients
 		availableClients := registration.Count
-		now := time.Now()
 
-		// Count clients that have exceeded quota but haven't recovered yet
 		expiredClients := 0
 		for _, quotaTime := range registration.QuotaExceededClients {
-			if quotaTime != nil && now.Sub(*quotaTime) < quotaExpiredDuration {
+			if quotaTime == nil {
+				continue
+			}
+			recoveryAt := quotaTime.Add(quotaExpiredDuration)
+			if now.Before(recoveryAt) {
 				expiredClients++
+				if expiresAt.IsZero() || recoveryAt.Before(expiresAt) {
+					expiresAt = recoveryAt
+				}
 			}
 		}
 
@@ -749,7 +820,6 @@ func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any
 			effectiveClients = 0
 		}
 
-		// Include models that have available clients, or those solely cooling down.
 		if effectiveClients > 0 || (availableClients > 0 && (expiredClients > 0 || cooldownSuspended > 0) && otherSuspended == 0) {
 			model := r.convertModelToMap(registration.Info, handlerType)
 			if model != nil {
@@ -758,7 +828,44 @@ func (r *ModelRegistry) GetAvailableModels(handlerType string) []map[string]any
 		}
 	}
 
-	return models
+	return models, expiresAt
+}
+
+func cloneModelMaps(models []map[string]any) []map[string]any {
+	cloned := make([]map[string]any, 0, len(models))
+	for _, model := range models {
+		if model == nil {
+			cloned = append(cloned, nil)
+			continue
+		}
+		copyModel := make(map[string]any, len(model))
+		for key, value := range model {
+			copyModel[key] = cloneModelMapValue(value)
+		}
+		cloned = append(cloned, copyModel)
+	}
+	return cloned
+}
+
+func cloneModelMapValue(value any) any {
+	switch typed := value.(type) {
+	case map[string]any:
+		copyMap := make(map[string]any, len(typed))
+		for key, entry := range typed {
+			copyMap[key] = cloneModelMapValue(entry)
+		}
+		return copyMap
+	case []any:
+		copySlice := make([]any, len(typed))
+		for i, entry := range typed {
+			copySlice[i] = cloneModelMapValue(entry)
+		}
+		return copySlice
+	case []string:
+		return append([]string(nil), typed...)
+	default:
+		return value
+	}
 }
 
 // GetAvailableModelsByProvider returns models available for the given provider identifier.
@@ -874,11 +981,11 @@ func (r *ModelRegistry) GetAvailableModelsByProvider(provider string) []*ModelIn
 
 		if effectiveClients > 0 || (availableClients > 0 && (expiredClients > 0 || cooldownSuspended > 0) && otherSuspended == 0) {
 			if entry.info != nil {
-				result = append(result, entry.info)
+				result = append(result, cloneModelInfo(entry.info))
 				continue
 			}
 			if ok && registration != nil && registration.Info != nil {
-				result = append(result, registration.Info)
+				result = append(result, cloneModelInfo(registration.Info))
 			}
 		}
 	}
@@ -987,13 +1094,13 @@ func (r *ModelRegistry) GetModelInfo(modelID, provider string) *ModelInfo {
 			if reg.Providers != nil {
 				if count, ok := reg.Providers[provider]; ok && count > 0 {
 					if info, ok := reg.InfoByProvider[provider]; ok && info != nil {
-						return info
+						return cloneModelInfo(info)
 					}
 				}
 			}
 		}
 		// Fallback to global info (last registered)
-		return reg.Info
+		return cloneModelInfo(reg.Info)
 	}
 	return nil
 }
@@ -1033,7 +1140,7 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 			result["max_completion_tokens"] = model.MaxCompletionTokens
 		}
 		if len(model.SupportedParameters) > 0 {
-			result["supported_parameters"] = model.SupportedParameters
+			result["supported_parameters"] = append([]string(nil), model.SupportedParameters...)
 		}
 		if len(model.SupportedEndpoints) > 0 {
 			result["supported_endpoints"] = model.SupportedEndpoints
@@ -1094,13 +1201,13 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 			result["outputTokenLimit"] = model.OutputTokenLimit
 		}
 		if len(model.SupportedGenerationMethods) > 0 {
-			result["supportedGenerationMethods"] = model.SupportedGenerationMethods
+			result["supportedGenerationMethods"] = append([]string(nil), model.SupportedGenerationMethods...)
 		}
 		if len(model.SupportedInputModalities) > 0 {
-			result["supportedInputModalities"] = model.SupportedInputModalities
+			result["supportedInputModalities"] = append([]string(nil), model.SupportedInputModalities...)
 		}
 		if len(model.SupportedOutputModalities) > 0 {
-			result["supportedOutputModalities"] = model.SupportedOutputModalities
+			result["supportedOutputModalities"] = append([]string(nil), model.SupportedOutputModalities...)
 		}
 		return result
 
@@ -1130,15 +1237,20 @@ func (r *ModelRegistry) CleanupExpiredQuotas() {
 
 	now := time.Now()
 	quotaExpiredDuration := 5 * time.Minute
+	invalidated := false
 
 	for modelID, registration := range r.models {
 		for clientID, quotaTime := range registration.QuotaExceededClients {
 			if quotaTime != nil && now.Sub(*quotaTime) >= quotaExpiredDuration {
 				delete(registration.QuotaExceededClients, clientID)
+				invalidated = true
 				log.Debugf("Cleaned up expired quota tracking for model %s, client %s", modelID, clientID)
 			}
 		}
 	}
+	if invalidated {
+		r.invalidateAvailableModelsCacheLocked()
+	}
 }
 
 // GetFirstAvailableModel returns the first available model for the given handler type.
@@ -1152,8 +1264,6 @@ func (r *ModelRegistry) CleanupExpiredQuotas() {
 //   - string: The model ID of the first available model, or empty string if none available
 //   - error: An error if no models are available
 func (r *ModelRegistry) GetFirstAvailableModel(handlerType string) (string, error) {
-	r.mutex.RLock()
-	defer r.mutex.RUnlock()
 
 	// Get all available models for this handler type
 	models := r.GetAvailableModels(handlerType)
@@ -1213,13 +1323,13 @@ func (r *ModelRegistry) GetModelsForClient(clientID string) []*ModelInfo {
 		// Prefer client's own model info to preserve original type/owned_by
 		if clientInfos != nil {
 			if info, ok := clientInfos[modelID]; ok && info != nil {
-				result = append(result, info)
+				result = append(result, cloneModelInfo(info))
 				continue
 			}
 		}
 		// Fallback to global registry (for backwards compatibility)
 		if reg, ok := r.models[modelID]; ok && reg.Info != nil {
-			result = append(result, reg.Info)
+			result = append(result, cloneModelInfo(reg.Info))
 		}
 	}
 	return result

internal/registry/model_registry_cache_test.go

@@ -0,0 +1,54 @@
+package registry
+
+import "testing"
+
+func TestGetAvailableModelsReturnsClonedSnapshots(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One"}})
+
+	first := r.GetAvailableModels("openai")
+	if len(first) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(first))
+	}
+	first[0]["id"] = "mutated"
+	first[0]["display_name"] = "Mutated"
+
+	second := r.GetAvailableModels("openai")
+	if got := second[0]["id"]; got != "m1" {
+		t.Fatalf("expected cached snapshot to stay isolated, got id %v", got)
+	}
+	if got := second[0]["display_name"]; got != "Model One" {
+		t.Fatalf("expected cached snapshot to stay isolated, got display_name %v", got)
+	}
+}
+
+func TestGetAvailableModelsInvalidatesCacheOnRegistryChanges(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One"}})
+
+	models := r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(models))
+	}
+	if got := models[0]["display_name"]; got != "Model One" {
+		t.Fatalf("expected initial display_name Model One, got %v", got)
+	}
+
+	r.RegisterClient("client-1", "OpenAI", []*ModelInfo{{ID: "m1", OwnedBy: "team-a", DisplayName: "Model One Updated"}})
+	models = r.GetAvailableModels("openai")
+	if got := models[0]["display_name"]; got != "Model One Updated" {
+		t.Fatalf("expected updated display_name after cache invalidation, got %v", got)
+	}
+
+	r.SuspendClientModel("client-1", "m1", "manual")
+	models = r.GetAvailableModels("openai")
+	if len(models) != 0 {
+		t.Fatalf("expected no available models after suspension, got %d", len(models))
+	}
+
+	r.ResumeClientModel("client-1", "m1")
+	models = r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected model to reappear after resume, got %d", len(models))
+	}
+}

internal/registry/model_registry_safety_test.go

@@ -0,0 +1,149 @@
+package registry
+
+import (
+	"testing"
+	"time"
+)
+
+func TestGetModelInfoReturnsClone(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Min: 1, Max: 2, Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetModelInfo("m1", "gemini")
+	if first == nil {
+		t.Fatal("expected model info")
+	}
+	first.DisplayName = "mutated"
+	first.Thinking.Levels[0] = "mutated"
+
+	second := r.GetModelInfo("m1", "gemini")
+	if second.DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second.DisplayName)
+	}
+	if second.Thinking == nil || len(second.Thinking.Levels) == 0 || second.Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second.Thinking)
+	}
+}
+
+func TestGetModelsForClientReturnsClones(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetModelsForClient("client-1")
+	if len(first) != 1 || first[0] == nil {
+		t.Fatalf("expected one model, got %+v", first)
+	}
+	first[0].DisplayName = "mutated"
+	first[0].Thinking.Levels[0] = "mutated"
+
+	second := r.GetModelsForClient("client-1")
+	if len(second) != 1 || second[0] == nil {
+		t.Fatalf("expected one model on second fetch, got %+v", second)
+	}
+	if second[0].DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second[0].DisplayName)
+	}
+	if second[0].Thinking == nil || len(second[0].Thinking.Levels) == 0 || second[0].Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second[0].Thinking)
+	}
+}
+
+func TestGetAvailableModelsByProviderReturnsClones(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "gemini", []*ModelInfo{{
+		ID:          "m1",
+		DisplayName: "Model One",
+		Thinking:    &ThinkingSupport{Levels: []string{"low", "high"}},
+	}})
+
+	first := r.GetAvailableModelsByProvider("gemini")
+	if len(first) != 1 || first[0] == nil {
+		t.Fatalf("expected one model, got %+v", first)
+	}
+	first[0].DisplayName = "mutated"
+	first[0].Thinking.Levels[0] = "mutated"
+
+	second := r.GetAvailableModelsByProvider("gemini")
+	if len(second) != 1 || second[0] == nil {
+		t.Fatalf("expected one model on second fetch, got %+v", second)
+	}
+	if second[0].DisplayName != "Model One" {
+		t.Fatalf("expected cloned display name, got %q", second[0].DisplayName)
+	}
+	if second[0].Thinking == nil || len(second[0].Thinking.Levels) == 0 || second[0].Thinking.Levels[0] != "low" {
+		t.Fatalf("expected cloned thinking levels, got %+v", second[0].Thinking)
+	}
+}
+
+func TestCleanupExpiredQuotasInvalidatesAvailableModelsCache(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "openai", []*ModelInfo{{ID: "m1", Created: 1}})
+	r.SetModelQuotaExceeded("client-1", "m1")
+	if models := r.GetAvailableModels("openai"); len(models) != 1 {
+		t.Fatalf("expected cooldown model to remain listed before cleanup, got %d", len(models))
+	}
+
+	r.mutex.Lock()
+	quotaTime := time.Now().Add(-6 * time.Minute)
+	r.models["m1"].QuotaExceededClients["client-1"] = &quotaTime
+	r.mutex.Unlock()
+
+	r.CleanupExpiredQuotas()
+
+	if count := r.GetModelCount("m1"); count != 1 {
+		t.Fatalf("expected model count 1 after cleanup, got %d", count)
+	}
+	models := r.GetAvailableModels("openai")
+	if len(models) != 1 {
+		t.Fatalf("expected model to stay available after cleanup, got %d", len(models))
+	}
+	if got := models[0]["id"]; got != "m1" {
+		t.Fatalf("expected model id m1, got %v", got)
+	}
+}
+
+func TestGetAvailableModelsReturnsClonedSupportedParameters(t *testing.T) {
+	r := newTestModelRegistry()
+	r.RegisterClient("client-1", "openai", []*ModelInfo{{
+		ID:                  "m1",
+		DisplayName:         "Model One",
+		SupportedParameters: []string{"temperature", "top_p"},
+	}})
+
+	first := r.GetAvailableModels("openai")
+	if len(first) != 1 {
+		t.Fatalf("expected one model, got %d", len(first))
+	}
+	params, ok := first[0]["supported_parameters"].([]string)
+	if !ok || len(params) != 2 {
+		t.Fatalf("expected supported_parameters slice, got %#v", first[0]["supported_parameters"])
+	}
+	params[0] = "mutated"
+
+	second := r.GetAvailableModels("openai")
+	params, ok = second[0]["supported_parameters"].([]string)
+	if !ok || len(params) != 2 || params[0] != "temperature" {
+		t.Fatalf("expected cloned supported_parameters, got %#v", second[0]["supported_parameters"])
+	}
+}
+
+func TestLookupModelInfoReturnsCloneForStaticDefinitions(t *testing.T) {
+	first := LookupModelInfo("glm-4.6")
+	if first == nil || first.Thinking == nil || len(first.Thinking.Levels) == 0 {
+		t.Fatalf("expected static model with thinking levels, got %+v", first)
+	}
+	first.Thinking.Levels[0] = "mutated"
+
+	second := LookupModelInfo("glm-4.6")
+	if second == nil || second.Thinking == nil || len(second.Thinking.Levels) == 0 || second.Thinking.Levels[0] == "mutated" {
+		t.Fatalf("expected static lookup clone, got %+v", second)
+	}
+}

internal/runtime/executor/claude_executor.go

@@ -1485,25 +1485,27 @@ func countCacheControlsMap(root map[string]any) int {
 	return count
 }
 
-func normalizeTTLForBlock(obj map[string]any, seen5m *bool) {
+func normalizeTTLForBlock(obj map[string]any, seen5m *bool) bool {
 	ccRaw, exists := obj["cache_control"]
 	if !exists {
-		return
+		return false
 	}
 	cc, ok := asObject(ccRaw)
 	if !ok {
 		*seen5m = true
-		return
+		return false
 	}
 	ttlRaw, ttlExists := cc["ttl"]
 	ttl, ttlIsString := ttlRaw.(string)
 	if !ttlExists || !ttlIsString || ttl != "1h" {
 		*seen5m = true
-		return
+		return false
 	}
 	if *seen5m {
 		delete(cc, "ttl")
+		return true
 	}
+	return false
 }
 
 func findLastCacheControlIndex(arr []any) int {
@@ -1599,11 +1601,14 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 	}
 
 	seen5m := false
+	modified := false
 
 	if tools, ok := asArray(root["tools"]); ok {
 		for _, tool := range tools {
 			if obj, ok := asObject(tool); ok {
-				normalizeTTLForBlock(obj, &seen5m)
+				if normalizeTTLForBlock(obj, &seen5m) {
+					modified = true
+				}
 			}
 		}
 	}
@@ -1611,7 +1616,9 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 	if system, ok := asArray(root["system"]); ok {
 		for _, item := range system {
 			if obj, ok := asObject(item); ok {
-				normalizeTTLForBlock(obj, &seen5m)
+				if normalizeTTLForBlock(obj, &seen5m) {
+					modified = true
+				}
 			}
 		}
 	}
@@ -1628,12 +1635,17 @@ func normalizeCacheControlTTL(payload []byte) []byte {
 			}
 			for _, item := range content {
 				if obj, ok := asObject(item); ok {
-					normalizeTTLForBlock(obj, &seen5m)
+					if normalizeTTLForBlock(obj, &seen5m) {
+						modified = true
+					}
 				}
 			}
 		}
 	}
 
+	if !modified {
+		return payload
+	}
 	return marshalPayloadObject(payload, root)
 }
 

internal/runtime/executor/claude_executor_test.go

@@ -369,6 +369,19 @@ func TestNormalizeCacheControlTTL_DowngradesLaterOneHourBlocks(t *testing.T) {
 	}
 }
 
+func TestNormalizeCacheControlTTL_PreservesOriginalBytesWhenNoChange(t *testing.T) {
+	// Payload where no TTL normalization is needed (all blocks use 1h with no
+	// preceding 5m block). The text intentionally contains HTML chars (<, >, &)
+	// that json.Marshal would escape to \u003c etc., altering byte identity.
+	payload := []byte(`{"tools":[{"name":"t1","cache_control":{"type":"ephemeral","ttl":"1h"}}],"system":[{"type":"text","text":"<system-reminder>foo & bar</system-reminder>","cache_control":{"type":"ephemeral","ttl":"1h"}}],"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`)
+
+	out := normalizeCacheControlTTL(payload)
+
+	if !bytes.Equal(out, payload) {
+		t.Fatalf("normalizeCacheControlTTL altered bytes when no change was needed.\noriginal: %s\ngot:      %s", payload, out)
+	}
+}
+
 func TestEnforceCacheControlLimit_StripsNonLastToolBeforeMessages(t *testing.T) {
 	payload := []byte(`{
 		"tools": [

internal/runtime/executor/codex_websockets_executor.go

@@ -31,7 +31,7 @@ import (
 )
 
 const (
-	codexResponsesWebsocketBetaHeaderValue = "responses_websockets=2026-02-04"
+	codexResponsesWebsocketBetaHeaderValue = "responses_websockets=2026-02-06"
 	codexResponsesWebsocketIdleTimeout     = 5 * time.Minute
 	codexResponsesWebsocketHandshakeTO     = 30 * time.Second
 )
@@ -57,11 +57,6 @@ type codexWebsocketSession struct {
 	wsURL  string
 	authID string
 
-	// connCreateSent tracks whether a `response.create` message has been successfully sent
-	// on the current websocket connection. The upstream expects the first message on each
-	// connection to be `response.create`.
-	connCreateSent bool
-
 	writeMu sync.Mutex
 
 	activeMu     sync.Mutex
@@ -212,13 +207,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 		defer sess.reqMu.Unlock()
 	}
 
-	allowAppend := true
-	if sess != nil {
-		sess.connMu.Lock()
-		allowAppend = sess.connCreateSent
-		sess.connMu.Unlock()
-	}
-	wsReqBody := buildCodexWebsocketRequestBody(body, allowAppend)
+	wsReqBody := buildCodexWebsocketRequestBody(body)
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       wsURL,
 		Method:    "WEBSOCKET",
@@ -280,10 +269,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 			// execution session.
 			connRetry, _, errDialRetry := e.ensureUpstreamConn(ctx, auth, sess, authID, wsURL, wsHeaders)
 			if errDialRetry == nil && connRetry != nil {
-				sess.connMu.Lock()
-				allowAppend = sess.connCreateSent
-				sess.connMu.Unlock()
-				wsReqBodyRetry := buildCodexWebsocketRequestBody(body, allowAppend)
+				wsReqBodyRetry := buildCodexWebsocketRequestBody(body)
 				recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 					URL:       wsURL,
 					Method:    "WEBSOCKET",
@@ -312,7 +298,6 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 			return resp, errSend
 		}
 	}
-	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	for {
 		if ctx != nil && ctx.Err() != nil {
@@ -403,26 +388,20 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	wsHeaders = applyCodexWebsocketHeaders(ctx, wsHeaders, auth, apiKey)
 
 	var authID, authLabel, authType, authValue string
-	if auth != nil {
-		authID = auth.ID
-		authLabel = auth.Label
-		authType, authValue = auth.AccountInfo()
-	}
+	authID = auth.ID
+	authLabel = auth.Label
+	authType, authValue = auth.AccountInfo()
 
 	executionSessionID := executionSessionIDFromOptions(opts)
 	var sess *codexWebsocketSession
 	if executionSessionID != "" {
 		sess = e.getOrCreateSession(executionSessionID)
-		sess.reqMu.Lock()
+		if sess != nil {
+			sess.reqMu.Lock()
+		}
 	}
 
-	allowAppend := true
-	if sess != nil {
-		sess.connMu.Lock()
-		allowAppend = sess.connCreateSent
-		sess.connMu.Unlock()
-	}
-	wsReqBody := buildCodexWebsocketRequestBody(body, allowAppend)
+	wsReqBody := buildCodexWebsocketRequestBody(body)
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       wsURL,
 		Method:    "WEBSOCKET",
@@ -483,10 +462,7 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 				sess.reqMu.Unlock()
 				return nil, errDialRetry
 			}
-			sess.connMu.Lock()
-			allowAppend = sess.connCreateSent
-			sess.connMu.Unlock()
-			wsReqBodyRetry := buildCodexWebsocketRequestBody(body, allowAppend)
+			wsReqBodyRetry := buildCodexWebsocketRequestBody(body)
 			recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 				URL:       wsURL,
 				Method:    "WEBSOCKET",
@@ -515,7 +491,6 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 			return nil, errSend
 		}
 	}
-	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	out := make(chan cliproxyexecutor.StreamChunk)
 	go func() {
@@ -657,31 +632,14 @@ func writeCodexWebsocketMessage(sess *codexWebsocketSession, conn *websocket.Con
 	return conn.WriteMessage(websocket.TextMessage, payload)
 }
 
-func buildCodexWebsocketRequestBody(body []byte, allowAppend bool) []byte {
+func buildCodexWebsocketRequestBody(body []byte) []byte {
 	if len(body) == 0 {
 		return nil
 	}
 
-	// Codex CLI websocket v2 uses `response.create` with `previous_response_id` for incremental turns.
-	// The upstream ChatGPT Codex websocket currently rejects that with close 1008 (policy violation).
-	// Fall back to v1 `response.append` semantics on the same websocket connection to keep the session alive.
-	//
-	// NOTE: The upstream expects the first websocket event on each connection to be `response.create`,
-	// so we only use `response.append` after we have initialized the current connection.
-	if allowAppend {
-		if prev := strings.TrimSpace(gjson.GetBytes(body, "previous_response_id").String()); prev != "" {
-			inputNode := gjson.GetBytes(body, "input")
-			wsReqBody := []byte(`{}`)
-			wsReqBody, _ = sjson.SetBytes(wsReqBody, "type", "response.append")
-			if inputNode.Exists() && inputNode.IsArray() && strings.TrimSpace(inputNode.Raw) != "" {
-				wsReqBody, _ = sjson.SetRawBytes(wsReqBody, "input", []byte(inputNode.Raw))
-				return wsReqBody
-			}
-			wsReqBody, _ = sjson.SetRawBytes(wsReqBody, "input", []byte("[]"))
-			return wsReqBody
-		}
-	}
-
+	// Match codex-rs websocket v2 semantics: every request is `response.create`.
+	// Incremental follow-up turns continue on the same websocket using
+	// `previous_response_id` + incremental `input`, not `response.append`.
 	wsReqBody, errSet := sjson.SetBytes(bytes.Clone(body), "type", "response.create")
 	if errSet == nil && len(wsReqBody) > 0 {
 		return wsReqBody
@@ -725,21 +683,6 @@ func readCodexWebsocketMessage(ctx context.Context, sess *codexWebsocketSession,
 	}
 }
 
-func markCodexWebsocketCreateSent(sess *codexWebsocketSession, conn *websocket.Conn, payload []byte) {
-	if sess == nil || conn == nil || len(payload) == 0 {
-		return
-	}
-	if strings.TrimSpace(gjson.GetBytes(payload, "type").String()) != "response.create" {
-		return
-	}
-
-	sess.connMu.Lock()
-	if sess.conn == conn {
-		sess.connCreateSent = true
-	}
-	sess.connMu.Unlock()
-}
-
 func newProxyAwareWebsocketDialer(cfg *config.Config, auth *cliproxyauth.Auth) *websocket.Dialer {
 	dialer := &websocket.Dialer{
 		Proxy:             http.ProxyFromEnvironment,
@@ -1017,36 +960,6 @@ func closeHTTPResponseBody(resp *http.Response, logPrefix string) {
 	}
 }
 
-func closeOnContextDone(ctx context.Context, conn *websocket.Conn) chan struct{} {
-	done := make(chan struct{})
-	if ctx == nil || conn == nil {
-		return done
-	}
-	go func() {
-		select {
-		case <-done:
-		case <-ctx.Done():
-			_ = conn.Close()
-		}
-	}()
-	return done
-}
-
-func cancelReadOnContextDone(ctx context.Context, conn *websocket.Conn) chan struct{} {
-	done := make(chan struct{})
-	if ctx == nil || conn == nil {
-		return done
-	}
-	go func() {
-		select {
-		case <-done:
-		case <-ctx.Done():
-			_ = conn.SetReadDeadline(time.Now())
-		}
-	}()
-	return done
-}
-
 func executionSessionIDFromOptions(opts cliproxyexecutor.Options) string {
 	if len(opts.Metadata) == 0 {
 		return ""
@@ -1120,7 +1033,6 @@ func (e *CodexWebsocketsExecutor) ensureUpstreamConn(ctx context.Context, auth *
 	sess.conn = conn
 	sess.wsURL = wsURL
 	sess.authID = authID
-	sess.connCreateSent = false
 	sess.readerConn = conn
 	sess.connMu.Unlock()
 
@@ -1206,7 +1118,6 @@ func (e *CodexWebsocketsExecutor) invalidateUpstreamConn(sess *codexWebsocketSes
 		return
 	}
 	sess.conn = nil
-	sess.connCreateSent = false
 	if sess.readerConn == conn {
 		sess.readerConn = nil
 	}
@@ -1273,7 +1184,6 @@ func (e *CodexWebsocketsExecutor) closeExecutionSession(sess *codexWebsocketSess
 	authID := sess.authID
 	wsURL := sess.wsURL
 	sess.conn = nil
-	sess.connCreateSent = false
 	if sess.readerConn == conn {
 		sess.readerConn = nil
 	}

internal/runtime/executor/codex_websockets_executor_test.go

@@ -0,0 +1,36 @@
+package executor
+
+import (
+	"context"
+	"net/http"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestBuildCodexWebsocketRequestBodyPreservesPreviousResponseID(t *testing.T) {
+	body := []byte(`{"model":"gpt-5-codex","previous_response_id":"resp-1","input":[{"type":"message","id":"msg-1"}]}`)
+
+	wsReqBody := buildCodexWebsocketRequestBody(body)
+
+	if got := gjson.GetBytes(wsReqBody, "type").String(); got != "response.create" {
+		t.Fatalf("type = %s, want response.create", got)
+	}
+	if got := gjson.GetBytes(wsReqBody, "previous_response_id").String(); got != "resp-1" {
+		t.Fatalf("previous_response_id = %s, want resp-1", got)
+	}
+	if gjson.GetBytes(wsReqBody, "input.0.id").String() != "msg-1" {
+		t.Fatalf("input item id mismatch")
+	}
+	if got := gjson.GetBytes(wsReqBody, "type").String(); got == "response.append" {
+		t.Fatalf("unexpected websocket request type: %s", got)
+	}
+}
+
+func TestApplyCodexWebsocketHeadersDefaultsToCurrentResponsesBeta(t *testing.T) {
+	headers := applyCodexWebsocketHeaders(context.Background(), http.Header{}, nil, "")
+
+	if got := headers.Get("OpenAI-Beta"); got != codexResponsesWebsocketBetaHeaderValue {
+		t.Fatalf("OpenAI-Beta = %s, want %s", got, codexResponsesWebsocketBetaHeaderValue)
+	}
+}

internal/runtime/executor/gemini_vertex_executor.go

@@ -460,7 +460,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	if opts.Alt != "" && action != "countTokens" {
@@ -683,7 +683,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	action := getVertexAction(baseModel, true)
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
 	// Imagen models don't support streaming, skip SSE params
@@ -883,7 +883,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
-		baseURL = "https://generativelanguage.googleapis.com"
+		baseURL = "https://aiplatform.googleapis.com"
 	}
 	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "countTokens")
 

internal/thinking/apply.go

@@ -257,7 +257,10 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromForma
 	if suffixResult.HasSuffix {
 		config = parseSuffixToConfig(suffixResult.RawSuffix, toFormat, modelID)
 	} else {
-		config = extractThinkingConfig(body, toFormat)
+		config = extractThinkingConfig(body, fromFormat)
+		if !hasThinkingConfig(config) && fromFormat != toFormat {
+			config = extractThinkingConfig(body, toFormat)
+		}
 	}
 
 	if !hasThinkingConfig(config) {
@@ -293,6 +296,9 @@ func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat stri
 	if config.Mode != ModeLevel {
 		return config
 	}
+	if toFormat == "claude" {
+		return config
+	}
 	if !isBudgetCapableProvider(toFormat) {
 		return config
 	}

internal/thinking/apply_user_defined_test.go

@@ -0,0 +1,55 @@
+package thinking_test
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/thinking/provider/claude"
+	"github.com/tidwall/gjson"
+)
+
+func TestApplyThinking_UserDefinedClaudePreservesAdaptiveLevel(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	clientID := "test-user-defined-claude-" + t.Name()
+	modelID := "custom-claude-4-6"
+	reg.RegisterClient(clientID, "claude", []*registry.ModelInfo{{ID: modelID, UserDefined: true}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(clientID)
+	})
+
+	tests := []struct {
+		name  string
+		model string
+		body  []byte
+	}{
+		{
+			name:  "claude adaptive effort body",
+			model: modelID,
+			body:  []byte(`{"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`),
+		},
+		{
+			name:  "suffix level",
+			model: modelID + "(high)",
+			body:  []byte(`{}`),
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			out, err := thinking.ApplyThinking(tt.body, tt.model, "openai", "claude", "claude")
+			if err != nil {
+				t.Fatalf("ApplyThinking() error = %v", err)
+			}
+			if got := gjson.GetBytes(out, "thinking.type").String(); got != "adaptive" {
+				t.Fatalf("thinking.type = %q, want %q, body=%s", got, "adaptive", string(out))
+			}
+			if got := gjson.GetBytes(out, "output_config.effort").String(); got != "high" {
+				t.Fatalf("output_config.effort = %q, want %q, body=%s", got, "high", string(out))
+			}
+			if gjson.GetBytes(out, "thinking.budget_tokens").Exists() {
+				t.Fatalf("thinking.budget_tokens should be removed, body=%s", string(out))
+			}
+		})
+	}
+}

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -477,9 +477,6 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				effort = strings.ToLower(strings.TrimSpace(v.String()))
 			}
 			if effort != "" {
-				if effort == "max" {
-					effort = "high"
-				}
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
 			} else {
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -1235,64 +1235,3 @@ func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *t
 		t.Errorf("Interleaved thinking hint should be in created systemInstruction, got: %v", sysInstruction.Raw)
 	}
 }
-
-func TestConvertClaudeRequestToAntigravity_AdaptiveThinking_EffortLevels(t *testing.T) {
-	tests := []struct {
-		name     string
-		effort   string
-		expected string
-	}{
-		{"low", "low", "low"},
-		{"medium", "medium", "medium"},
-		{"high", "high", "high"},
-		{"max", "max", "high"},
-	}
-
-	for _, tt := range tests {
-		tt := tt
-		t.Run(tt.name, func(t *testing.T) {
-			inputJSON := []byte(`{
-				"model": "claude-opus-4-6-thinking",
-				"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
-				"thinking": {"type": "adaptive"},
-				"output_config": {"effort": "` + tt.effort + `"}
-			}`)
-
-			output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, false)
-			outputStr := string(output)
-
-			thinkingConfig := gjson.Get(outputStr, "request.generationConfig.thinkingConfig")
-			if !thinkingConfig.Exists() {
-				t.Fatal("thinkingConfig should exist for adaptive thinking")
-			}
-			if thinkingConfig.Get("thinkingLevel").String() != tt.expected {
-				t.Errorf("Expected thinkingLevel %q, got %q", tt.expected, thinkingConfig.Get("thinkingLevel").String())
-			}
-			if !thinkingConfig.Get("includeThoughts").Bool() {
-				t.Error("includeThoughts should be true")
-			}
-		})
-	}
-}
-
-func TestConvertClaudeRequestToAntigravity_AdaptiveThinking_NoEffort(t *testing.T) {
-	inputJSON := []byte(`{
-		"model": "claude-opus-4-6-thinking",
-		"messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
-		"thinking": {"type": "adaptive"}
-	}`)
-
-	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, false)
-	outputStr := string(output)
-
-	thinkingConfig := gjson.Get(outputStr, "request.generationConfig.thinkingConfig")
-	if !thinkingConfig.Exists() {
-		t.Fatal("thinkingConfig should exist for adaptive thinking without effort")
-	}
-	if thinkingConfig.Get("thinkingLevel").String() != "high" {
-		t.Errorf("Expected default thinkingLevel \"high\", got %q", thinkingConfig.Get("thinkingLevel").String())
-	}
-	if !thinkingConfig.Get("includeThoughts").Bool() {
-		t.Error("includeThoughts should be true")
-	}
-}

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -15,6 +15,7 @@ import (
 	"time"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	log "github.com/sirupsen/logrus"
 
 	"github.com/tidwall/gjson"
@@ -256,7 +257,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, params.ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1))))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 

internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go

@@ -212,6 +212,33 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 							} else {
 								log.Warnf("Unknown file name extension '%s' in user message, skip", ext)
 							}
+						case "input_audio":
+							audioData := item.Get("input_audio.data").String()
+							audioFormat := item.Get("input_audio.format").String()
+							if audioData != "" {
+								audioMimeMap := map[string]string{
+									"mp3":       "audio/mpeg",
+									"wav":       "audio/wav",
+									"ogg":       "audio/ogg",
+									"flac":      "audio/flac",
+									"aac":       "audio/aac",
+									"webm":      "audio/webm",
+									"pcm16":     "audio/pcm",
+									"g711_ulaw": "audio/basic",
+									"g711_alaw": "audio/basic",
+								}
+								mimeType := "audio/wav"
+								if audioFormat != "" {
+									if mapped, ok := audioMimeMap[audioFormat]; ok {
+										mimeType = mapped
+									} else {
+										mimeType = "audio/" + audioFormat
+									}
+								}
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.mime_type", mimeType)
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".inlineData.data", audioData)
+								p++
+							}
 						}
 					}
 				}

internal/translator/claude/openai/chat-completions/claude_openai_request.go

@@ -203,46 +203,9 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 					msg, _ = sjson.SetRaw(msg, "content.-1", part)
 				} else if contentResult.Exists() && contentResult.IsArray() {
 					contentResult.ForEach(func(_, part gjson.Result) bool {
-						partType := part.Get("type").String()
-
-						switch partType {
-						case "text":
-							textPart := `{"type":"text","text":""}`
-							textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
-							msg, _ = sjson.SetRaw(msg, "content.-1", textPart)
-
-						case "image_url":
-							// Convert OpenAI image format to Claude Code format
-							imageURL := part.Get("image_url.url").String()
-							if strings.HasPrefix(imageURL, "data:") {
-								// Extract base64 data and media type from data URL
-								parts := strings.Split(imageURL, ",")
-								if len(parts) == 2 {
-									mediaTypePart := strings.Split(parts[0], ";")[0]
-									mediaType := strings.TrimPrefix(mediaTypePart, "data:")
-									data := parts[1]
-
-									imagePart := `{"type":"image","source":{"type":"base64","media_type":"","data":""}}`
-									imagePart, _ = sjson.Set(imagePart, "source.media_type", mediaType)
-									imagePart, _ = sjson.Set(imagePart, "source.data", data)
-									msg, _ = sjson.SetRaw(msg, "content.-1", imagePart)
-								}
-							}
-
-						case "file":
-							fileData := part.Get("file.file_data").String()
-							if strings.HasPrefix(fileData, "data:") {
-								semicolonIdx := strings.Index(fileData, ";")
-								commaIdx := strings.Index(fileData, ",")
-								if semicolonIdx != -1 && commaIdx != -1 && commaIdx > semicolonIdx {
-									mediaType := strings.TrimPrefix(fileData[:semicolonIdx], "data:")
-									data := fileData[commaIdx+1:]
-									docPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
-									docPart, _ = sjson.Set(docPart, "source.media_type", mediaType)
-									docPart, _ = sjson.Set(docPart, "source.data", data)
-									msg, _ = sjson.SetRaw(msg, "content.-1", docPart)
-								}
-							}
+						claudePart := convertOpenAIContentPartToClaudePart(part)
+						if claudePart != "" {
+							msg, _ = sjson.SetRaw(msg, "content.-1", claudePart)
 						}
 						return true
 					})
@@ -291,11 +254,16 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 			case "tool":
 				// Handle tool result messages conversion
 				toolCallID := message.Get("tool_call_id").String()
-				content := message.Get("content").String()
+				toolContentResult := message.Get("content")
 
 				msg := `{"role":"user","content":[{"type":"tool_result","tool_use_id":"","content":""}]}`
 				msg, _ = sjson.Set(msg, "content.0.tool_use_id", toolCallID)
-				msg, _ = sjson.Set(msg, "content.0.content", content)
+				toolResultContent, toolResultContentRaw := convertOpenAIToolResultContent(toolContentResult)
+				if toolResultContentRaw {
+					msg, _ = sjson.SetRaw(msg, "content.0.content", toolResultContent)
+				} else {
+					msg, _ = sjson.Set(msg, "content.0.content", toolResultContent)
+				}
 				out, _ = sjson.SetRaw(out, "messages.-1", msg)
 				messageIndex++
 			}
@@ -358,3 +326,110 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	return []byte(out)
 }
+
+func convertOpenAIContentPartToClaudePart(part gjson.Result) string {
+	switch part.Get("type").String() {
+	case "text":
+		textPart := `{"type":"text","text":""}`
+		textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
+		return textPart
+
+	case "image_url":
+		return convertOpenAIImageURLToClaudePart(part.Get("image_url.url").String())
+
+	case "file":
+		fileData := part.Get("file.file_data").String()
+		if strings.HasPrefix(fileData, "data:") {
+			semicolonIdx := strings.Index(fileData, ";")
+			commaIdx := strings.Index(fileData, ",")
+			if semicolonIdx != -1 && commaIdx != -1 && commaIdx > semicolonIdx {
+				mediaType := strings.TrimPrefix(fileData[:semicolonIdx], "data:")
+				data := fileData[commaIdx+1:]
+				docPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
+				docPart, _ = sjson.Set(docPart, "source.media_type", mediaType)
+				docPart, _ = sjson.Set(docPart, "source.data", data)
+				return docPart
+			}
+		}
+	}
+
+	return ""
+}
+
+func convertOpenAIImageURLToClaudePart(imageURL string) string {
+	if imageURL == "" {
+		return ""
+	}
+
+	if strings.HasPrefix(imageURL, "data:") {
+		parts := strings.SplitN(imageURL, ",", 2)
+		if len(parts) != 2 {
+			return ""
+		}
+
+		mediaTypePart := strings.SplitN(parts[0], ";", 2)[0]
+		mediaType := strings.TrimPrefix(mediaTypePart, "data:")
+		if mediaType == "" {
+			mediaType = "application/octet-stream"
+		}
+
+		imagePart := `{"type":"image","source":{"type":"base64","media_type":"","data":""}}`
+		imagePart, _ = sjson.Set(imagePart, "source.media_type", mediaType)
+		imagePart, _ = sjson.Set(imagePart, "source.data", parts[1])
+		return imagePart
+	}
+
+	imagePart := `{"type":"image","source":{"type":"url","url":""}}`
+	imagePart, _ = sjson.Set(imagePart, "source.url", imageURL)
+	return imagePart
+}
+
+func convertOpenAIToolResultContent(content gjson.Result) (string, bool) {
+	if !content.Exists() {
+		return "", false
+	}
+
+	if content.Type == gjson.String {
+		return content.String(), false
+	}
+
+	if content.IsArray() {
+		claudeContent := "[]"
+		partCount := 0
+
+		content.ForEach(func(_, part gjson.Result) bool {
+			if part.Type == gjson.String {
+				textPart := `{"type":"text","text":""}`
+				textPart, _ = sjson.Set(textPart, "text", part.String())
+				claudeContent, _ = sjson.SetRaw(claudeContent, "-1", textPart)
+				partCount++
+				return true
+			}
+
+			claudePart := convertOpenAIContentPartToClaudePart(part)
+			if claudePart != "" {
+				claudeContent, _ = sjson.SetRaw(claudeContent, "-1", claudePart)
+				partCount++
+			}
+			return true
+		})
+
+		if partCount > 0 || len(content.Array()) == 0 {
+			return claudeContent, true
+		}
+
+		return content.Raw, false
+	}
+
+	if content.IsObject() {
+		claudePart := convertOpenAIContentPartToClaudePart(content)
+		if claudePart != "" {
+			claudeContent := "[]"
+			claudeContent, _ = sjson.SetRaw(claudeContent, "-1", claudePart)
+			return claudeContent, true
+		}
+		return content.Raw, false
+	}
+
+	return content.Raw, false
+}

internal/translator/claude/openai/chat-completions/claude_openai_request_test.go

@@ -0,0 +1,137 @@
+package chat_completions
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+func TestConvertOpenAIRequestToClaude_ToolResultTextAndBase64Image(t *testing.T) {
+	inputJSON := `{
+		"model": "gpt-4.1",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": "",
+				"tool_calls": [
+					{
+						"id": "call_1",
+						"type": "function",
+						"function": {
+							"name": "do_work",
+							"arguments": "{\"a\":1}"
+						}
+					}
+				]
+			},
+			{
+				"role": "tool",
+				"tool_call_id": "call_1",
+				"content": [
+					{"type": "text", "text": "tool ok"},
+					{
+						"type": "image_url",
+						"image_url": {
+							"url": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUg=="
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertOpenAIRequestToClaude("claude-sonnet-4-5", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolResult := messages[1].Get("content.0")
+	if got := toolResult.Get("type").String(); got != "tool_result" {
+		t.Fatalf("Expected content[0].type %q, got %q", "tool_result", got)
+	}
+	if got := toolResult.Get("tool_use_id").String(); got != "call_1" {
+		t.Fatalf("Expected tool_use_id %q, got %q", "call_1", got)
+	}
+
+	toolContent := toolResult.Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool_result content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "text" {
+		t.Fatalf("Expected first tool_result part type %q, got %q", "text", got)
+	}
+	if got := toolContent.Get("0.text").String(); got != "tool ok" {
+		t.Fatalf("Expected first tool_result part text %q, got %q", "tool ok", got)
+	}
+	if got := toolContent.Get("1.type").String(); got != "image" {
+		t.Fatalf("Expected second tool_result part type %q, got %q", "image", got)
+	}
+	if got := toolContent.Get("1.source.type").String(); got != "base64" {
+		t.Fatalf("Expected image source type %q, got %q", "base64", got)
+	}
+	if got := toolContent.Get("1.source.media_type").String(); got != "image/png" {
+		t.Fatalf("Expected image media type %q, got %q", "image/png", got)
+	}
+	if got := toolContent.Get("1.source.data").String(); got != "iVBORw0KGgoAAAANSUhEUg==" {
+		t.Fatalf("Unexpected base64 image data: %q", got)
+	}
+}
+
+func TestConvertOpenAIRequestToClaude_ToolResultURLImageOnly(t *testing.T) {
+	inputJSON := `{
+		"model": "gpt-4.1",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": "",
+				"tool_calls": [
+					{
+						"id": "call_1",
+						"type": "function",
+						"function": {
+							"name": "do_work",
+							"arguments": "{\"a\":1}"
+						}
+					}
+				]
+			},
+			{
+				"role": "tool",
+				"tool_call_id": "call_1",
+				"content": [
+					{
+						"type": "image_url",
+						"image_url": {
+							"url": "https://example.com/tool.png"
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertOpenAIRequestToClaude("claude-sonnet-4-5", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content.0.content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool_result content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "image" {
+		t.Fatalf("Expected tool_result part type %q, got %q", "image", got)
+	}
+	if got := toolContent.Get("0.source.type").String(); got != "url" {
+		t.Fatalf("Expected image source type %q, got %q", "url", got)
+	}
+	if got := toolContent.Get("0.source.url").String(); got != "https://example.com/tool.png" {
+		t.Fatalf("Unexpected image URL: %q", got)
+	}
+}

internal/translator/codex/claude/codex_claude_response.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"strings"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -141,7 +142,7 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 			(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = false
 			template = `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
-			template, _ = sjson.Set(template, "content_block.id", itemResult.Get("call_id").String())
+			template, _ = sjson.Set(template, "content_block.id", util.SanitizeClaudeToolID(itemResult.Get("call_id").String()))
 			{
 				// Restore original tool name if shortened
 				name := itemResult.Get("name").String()
@@ -310,7 +311,7 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 				}
 
 				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolBlock, _ = sjson.Set(toolBlock, "id", item.Get("call_id").String())
+				toolBlock, _ = sjson.Set(toolBlock, "id", util.SanitizeClaudeToolID(item.Get("call_id").String()))
 				toolBlock, _ = sjson.Set(toolBlock, "name", name)
 				inputRaw := "{}"
 				if argsStr := item.Get("arguments").String(); argsStr != "" && gjson.Valid(argsStr) {

internal/translator/gemini-cli/claude/gemini-cli_claude_response.go

@@ -14,6 +14,7 @@ import (
 	"sync/atomic"
 	"time"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -209,7 +210,7 @@ func ConvertGeminiCLIResponseToClaude(_ context.Context, _ string, originalReque
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1)))
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&toolUseIDCounter, 1))))
 				data, _ = sjson.Set(data, "content_block.name", fcName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 

internal/translator/gemini/claude/gemini_claude_response.go

@@ -224,7 +224,7 @@ func ConvertGeminiResponseToClaude(_ context.Context, _ string, originalRequestR
 
 				// Create the tool use block with unique ID and function details
 				data := fmt.Sprintf(`{"type":"content_block_start","index":%d,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`, (*param).(*Params).ResponseIndex)
-				data, _ = sjson.Set(data, "content_block.id", fmt.Sprintf("%s-%d", upstreamToolName, atomic.AddUint64(&toolUseIDCounter, 1)))
+				data, _ = sjson.Set(data, "content_block.id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d", upstreamToolName, atomic.AddUint64(&toolUseIDCounter, 1))))
 				data, _ = sjson.Set(data, "content_block.name", clientToolName)
 				output = output + fmt.Sprintf("data: %s\n\n\n", data)
 
@@ -343,7 +343,7 @@ func ConvertGeminiResponseToClaudeNonStream(_ context.Context, _ string, origina
 				clientToolName := util.MapToolName(toolNameMap, upstreamToolName)
 				toolIDCounter++
 				toolBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolBlock, _ = sjson.Set(toolBlock, "id", fmt.Sprintf("%s-%d", upstreamToolName, toolIDCounter))
+				toolBlock, _ = sjson.Set(toolBlock, "id", util.SanitizeClaudeToolID(fmt.Sprintf("%s-%d", upstreamToolName, toolIDCounter)))
 				toolBlock, _ = sjson.Set(toolBlock, "name", clientToolName)
 				inputRaw := "{}"
 				if args := functionCall.Get("args"); args.Exists() && gjson.Valid(args.Raw) && args.IsObject() {

internal/translator/gemini/openai/responses/gemini_openai-responses_request.go

@@ -237,6 +237,33 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 									partJSON, _ = sjson.Set(partJSON, "inline_data.data", data)
 								}
 							}
+						case "input_audio":
+							audioData := contentItem.Get("data").String()
+							audioFormat := contentItem.Get("format").String()
+							if audioData != "" {
+								audioMimeMap := map[string]string{
+									"mp3":       "audio/mpeg",
+									"wav":       "audio/wav",
+									"ogg":       "audio/ogg",
+									"flac":      "audio/flac",
+									"aac":       "audio/aac",
+									"webm":      "audio/webm",
+									"pcm16":     "audio/pcm",
+									"g711_ulaw": "audio/basic",
+									"g711_alaw": "audio/basic",
+								}
+								mimeType := "audio/wav"
+								if audioFormat != "" {
+									if mapped, ok := audioMimeMap[audioFormat]; ok {
+										mimeType = mapped
+									} else {
+										mimeType = "audio/" + audioFormat
+									}
+								}
+								partJSON = `{"inline_data":{"mime_type":"","data":""}}`
+								partJSON, _ = sjson.Set(partJSON, "inline_data.mime_type", mimeType)
+								partJSON, _ = sjson.Set(partJSON, "inline_data.data", audioData)
+							}
 						}
 
 						if partJSON != "" {

internal/translator/openai/claude/openai_claude_request.go

@@ -183,7 +183,12 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						// Collect tool_result to emit after the main message (ensures tool results follow tool_calls)
 						toolResultJSON := `{"role":"tool","tool_call_id":"","content":""}`
 						toolResultJSON, _ = sjson.Set(toolResultJSON, "tool_call_id", part.Get("tool_use_id").String())
-						toolResultJSON, _ = sjson.Set(toolResultJSON, "content", convertClaudeToolResultContentToString(part.Get("content")))
+						toolResultContent, toolResultContentRaw := convertClaudeToolResultContent(part.Get("content"))
+						if toolResultContentRaw {
+							toolResultJSON, _ = sjson.SetRaw(toolResultJSON, "content", toolResultContent)
+						} else {
+							toolResultJSON, _ = sjson.Set(toolResultJSON, "content", toolResultContent)
+						}
 						toolResults = append(toolResults, toolResultJSON)
 					}
 					return true
@@ -374,21 +379,41 @@ func convertClaudeContentPart(part gjson.Result) (string, bool) {
 	}
 }
 
-func convertClaudeToolResultContentToString(content gjson.Result) string {
+func convertClaudeToolResultContent(content gjson.Result) (string, bool) {
 	if !content.Exists() {
-		return ""
+		return "", false
 	}
 
 	if content.Type == gjson.String {
-		return content.String()
+		return content.String(), false
 	}
 
 	if content.IsArray() {
 		var parts []string
+		contentJSON := "[]"
+		hasImagePart := false
 		content.ForEach(func(_, item gjson.Result) bool {
 			switch {
 			case item.Type == gjson.String:
-				parts = append(parts, item.String())
+				text := item.String()
+				parts = append(parts, text)
+				textContent := `{"type":"text","text":""}`
+				textContent, _ = sjson.Set(textContent, "text", text)
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", textContent)
+			case item.IsObject() && item.Get("type").String() == "text":
+				text := item.Get("text").String()
+				parts = append(parts, text)
+				textContent := `{"type":"text","text":""}`
+				textContent, _ = sjson.Set(textContent, "text", text)
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", textContent)
+			case item.IsObject() && item.Get("type").String() == "image":
+				contentItem, ok := convertClaudeContentPart(item)
+				if ok {
+					contentJSON, _ = sjson.SetRaw(contentJSON, "-1", contentItem)
+					hasImagePart = true
+				} else {
+					parts = append(parts, item.Raw)
+				}
 			case item.IsObject() && item.Get("text").Exists() && item.Get("text").Type == gjson.String:
 				parts = append(parts, item.Get("text").String())
 			default:
@@ -397,19 +422,31 @@ func convertClaudeToolResultContentToString(content gjson.Result) string {
 			return true
 		})
 
+		if hasImagePart {
+			return contentJSON, true
+		}
+
 		joined := strings.Join(parts, "\n\n")
 		if strings.TrimSpace(joined) != "" {
-			return joined
+			return joined, false
 		}
-		return content.Raw
+		return content.Raw, false
 	}
 
 	if content.IsObject() {
-		if text := content.Get("text"); text.Exists() && text.Type == gjson.String {
-			return text.String()
+		if content.Get("type").String() == "image" {
+			contentItem, ok := convertClaudeContentPart(content)
+			if ok {
+				contentJSON := "[]"
+				contentJSON, _ = sjson.SetRaw(contentJSON, "-1", contentItem)
+				return contentJSON, true
+			}
 		}
-		return content.Raw
+		if text := content.Get("text"); text.Exists() && text.Type == gjson.String {
+			return text.String(), false
+		}
+		return content.Raw, false
 	}
 
-	return content.Raw
+	return content.Raw, false
 }

internal/translator/openai/claude/openai_claude_request_test.go

@@ -488,6 +488,114 @@ func TestConvertClaudeRequestToOpenAI_ToolResultObjectContent(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToOpenAI_ToolResultTextAndImageContent(t *testing.T) {
+	inputJSON := `{
+		"model": "claude-3-opus",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "tool_use", "id": "call_1", "name": "do_work", "input": {"a": 1}}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "call_1",
+						"content": [
+							{"type": "text", "text": "tool ok"},
+							{
+								"type": "image",
+								"source": {
+									"type": "base64",
+									"media_type": "image/png",
+									"data": "iVBORw0KGgoAAAANSUhEUg=="
+								}
+							}
+						]
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertClaudeRequestToOpenAI("test-model", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "text" {
+		t.Fatalf("Expected first tool content type %q, got %q", "text", got)
+	}
+	if got := toolContent.Get("0.text").String(); got != "tool ok" {
+		t.Fatalf("Expected first tool content text %q, got %q", "tool ok", got)
+	}
+	if got := toolContent.Get("1.type").String(); got != "image_url" {
+		t.Fatalf("Expected second tool content type %q, got %q", "image_url", got)
+	}
+	if got := toolContent.Get("1.image_url.url").String(); got != "data:image/png;base64,iVBORw0KGgoAAAANSUhEUg==" {
+		t.Fatalf("Unexpected image_url: %q", got)
+	}
+}
+
+func TestConvertClaudeRequestToOpenAI_ToolResultURLImageOnly(t *testing.T) {
+	inputJSON := `{
+		"model": "claude-3-opus",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{"type": "tool_use", "id": "call_1", "name": "do_work", "input": {"a": 1}}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "call_1",
+						"content": {
+							"type": "image",
+							"source": {
+								"type": "url",
+								"url": "https://example.com/tool.png"
+							}
+						}
+					}
+				]
+			}
+		]
+	}`
+
+	result := ConvertClaudeRequestToOpenAI("test-model", []byte(inputJSON), false)
+	resultJSON := gjson.ParseBytes(result)
+	messages := resultJSON.Get("messages").Array()
+
+	if len(messages) != 2 {
+		t.Fatalf("Expected 2 messages, got %d. Messages: %s", len(messages), resultJSON.Get("messages").Raw)
+	}
+
+	toolContent := messages[1].Get("content")
+	if !toolContent.IsArray() {
+		t.Fatalf("Expected tool content array, got %s", toolContent.Raw)
+	}
+	if got := toolContent.Get("0.type").String(); got != "image_url" {
+		t.Fatalf("Expected tool content type %q, got %q", "image_url", got)
+	}
+	if got := toolContent.Get("0.image_url.url").String(); got != "https://example.com/tool.png" {
+		t.Fatalf("Unexpected image_url: %q", got)
+	}
+}
+
 func TestConvertClaudeRequestToOpenAI_AssistantTextToolUseTextOrder(t *testing.T) {
 	inputJSON := `{
 		"model": "claude-3-opus",

internal/translator/openai/claude/openai_claude_response.go

@@ -243,7 +243,7 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 						// Send content_block_start for tool_use
 						contentBlockStartJSON := `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "index", blockIndex)
-						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.id", accumulator.ID)
+						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.id", util.SanitizeClaudeToolID(accumulator.ID))
 						contentBlockStartJSON, _ = sjson.Set(contentBlockStartJSON, "content_block.name", accumulator.Name)
 						results = append(results, "event: content_block_start\ndata: "+contentBlockStartJSON+"\n\n")
 					}
@@ -414,7 +414,7 @@ func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 		if toolCalls := choice.Get("message.tool_calls"); toolCalls.Exists() && toolCalls.IsArray() {
 			toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 				toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-				toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
+				toolUseBlock, _ = sjson.Set(toolUseBlock, "id", util.SanitizeClaudeToolID(toolCall.Get("id").String()))
 				toolUseBlock, _ = sjson.Set(toolUseBlock, "name", toolCall.Get("function.name").String())
 
 				argsStr := util.FixJSON(toolCall.Get("function.arguments").String())
@@ -612,7 +612,7 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 								toolCalls.ForEach(func(_, tc gjson.Result) bool {
 									hasToolCall = true
 									toolUse := `{"type":"tool_use","id":"","name":"","input":{}}`
-									toolUse, _ = sjson.Set(toolUse, "id", tc.Get("id").String())
+									toolUse, _ = sjson.Set(toolUse, "id", util.SanitizeClaudeToolID(tc.Get("id").String()))
 									toolUse, _ = sjson.Set(toolUse, "name", util.MapToolName(toolNameMap, tc.Get("function.name").String()))
 
 									argsStr := util.FixJSON(tc.Get("function.arguments").String())
@@ -669,7 +669,7 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 				toolCalls.ForEach(func(_, toolCall gjson.Result) bool {
 					hasToolCall = true
 					toolUseBlock := `{"type":"tool_use","id":"","name":"","input":{}}`
-					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", toolCall.Get("id").String())
+					toolUseBlock, _ = sjson.Set(toolUseBlock, "id", util.SanitizeClaudeToolID(toolCall.Get("id").String()))
 					toolUseBlock, _ = sjson.Set(toolUseBlock, "name", util.MapToolName(toolNameMap, toolCall.Get("function.name").String()))
 
 					argsStr := util.FixJSON(toolCall.Get("function.arguments").String())

internal/util/claude_tool_id.go

@@ -0,0 +1,24 @@
+package util
+
+import (
+	"fmt"
+	"regexp"
+	"sync/atomic"
+	"time"
+)
+
+var (
+	claudeToolUseIDSanitizer = regexp.MustCompile(`[^a-zA-Z0-9_-]`)
+	claudeToolUseIDCounter   uint64
+)
+
+// SanitizeClaudeToolID ensures the given id conforms to Claude's
+// tool_use.id regex ^[a-zA-Z0-9_-]+$.  Non-conforming characters are
+// replaced with '_'; an empty result gets a generated fallback.
+func SanitizeClaudeToolID(id string) string {
+	s := claudeToolUseIDSanitizer.ReplaceAllString(id, "_")
+	if s == "" {
+		s = fmt.Sprintf("toolu_%d_%d", time.Now().UnixNano(), atomic.AddUint64(&claudeToolUseIDCounter, 1))
+	}
+	return s
+}

internal/watcher/clients.go

@@ -17,6 +17,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/synthesizer"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
 )
@@ -75,6 +76,7 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 
 		w.lastAuthHashes = make(map[string]string)
 		w.lastAuthContents = make(map[string]*coreauth.Auth)
+		w.fileAuthsByPath = make(map[string]map[string]*coreauth.Auth)
 		if resolvedAuthDir, errResolveAuthDir := util.ResolveAuthDir(cfg.AuthDir); errResolveAuthDir != nil {
 			log.Errorf("failed to resolve auth directory for hash cache: %v", errResolveAuthDir)
 		} else if resolvedAuthDir != "" {
@@ -92,6 +94,17 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 						if errParse := json.Unmarshal(data, &auth); errParse == nil {
 							w.lastAuthContents[normalizedPath] = &auth
 						}
+						ctx := &synthesizer.SynthesisContext{
+							Config:      cfg,
+							AuthDir:     resolvedAuthDir,
+							Now:         time.Now(),
+							IDGenerator: synthesizer.NewStableIDGenerator(),
+						}
+						if generated := synthesizer.SynthesizeAuthFile(ctx, path, data); len(generated) > 0 {
+							if pathAuths := authSliceToMap(generated); len(pathAuths) > 0 {
+								w.fileAuthsByPath[normalizedPath] = pathAuths
+							}
+						}
 					}
 				}
 				return nil
@@ -143,13 +156,14 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	}
 
 	w.clientsMutex.Lock()
-
-	cfg := w.config
-	if cfg == nil {
+	if w.config == nil {
 		log.Error("config is nil, cannot add or update client")
 		w.clientsMutex.Unlock()
 		return
 	}
+	if w.fileAuthsByPath == nil {
+		w.fileAuthsByPath = make(map[string]map[string]*coreauth.Auth)
+	}
 	if prev, ok := w.lastAuthHashes[normalized]; ok && prev == curHash {
 		log.Debugf("auth file unchanged (hash match), skipping reload: %s", filepath.Base(path))
 		w.clientsMutex.Unlock()
@@ -177,34 +191,86 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	}
 	w.lastAuthContents[normalized] = &newAuth
 
-	w.clientsMutex.Unlock() // Unlock before the callback
-
-	w.refreshAuthState(false)
-
-	if w.reloadCallback != nil {
-		log.Debugf("triggering server update callback after add/update")
-		w.triggerServerUpdate(cfg)
+	oldByID := make(map[string]*coreauth.Auth, len(w.fileAuthsByPath[normalized]))
+	for id, a := range w.fileAuthsByPath[normalized] {
+		oldByID[id] = a
 	}
+
+	// Build synthesized auth entries for this single file only.
+	sctx := &synthesizer.SynthesisContext{
+		Config:      w.config,
+		AuthDir:     w.authDir,
+		Now:         time.Now(),
+		IDGenerator: synthesizer.NewStableIDGenerator(),
+	}
+	generated := synthesizer.SynthesizeAuthFile(sctx, path, data)
+	newByID := authSliceToMap(generated)
+	if len(newByID) > 0 {
+		w.fileAuthsByPath[normalized] = newByID
+	} else {
+		delete(w.fileAuthsByPath, normalized)
+	}
+	updates := w.computePerPathUpdatesLocked(oldByID, newByID)
+	w.clientsMutex.Unlock()
+
 	w.persistAuthAsync(fmt.Sprintf("Sync auth %s", filepath.Base(path)), path)
+	w.dispatchAuthUpdates(updates)
 }
 
 func (w *Watcher) removeClient(path string) {
 	normalized := w.normalizeAuthPath(path)
 	w.clientsMutex.Lock()
-
-	cfg := w.config
+	oldByID := make(map[string]*coreauth.Auth, len(w.fileAuthsByPath[normalized]))
+	for id, a := range w.fileAuthsByPath[normalized] {
+		oldByID[id] = a
+	}
 	delete(w.lastAuthHashes, normalized)
 	delete(w.lastAuthContents, normalized)
+	delete(w.fileAuthsByPath, normalized)
 
-	w.clientsMutex.Unlock() // Release the lock before the callback
+	updates := w.computePerPathUpdatesLocked(oldByID, map[string]*coreauth.Auth{})
+	w.clientsMutex.Unlock()
 
-	w.refreshAuthState(false)
-
-	if w.reloadCallback != nil {
-		log.Debugf("triggering server update callback after removal")
-		w.triggerServerUpdate(cfg)
-	}
 	w.persistAuthAsync(fmt.Sprintf("Remove auth %s", filepath.Base(path)), path)
+	w.dispatchAuthUpdates(updates)
+}
+
+func (w *Watcher) computePerPathUpdatesLocked(oldByID, newByID map[string]*coreauth.Auth) []AuthUpdate {
+	if w.currentAuths == nil {
+		w.currentAuths = make(map[string]*coreauth.Auth)
+	}
+	updates := make([]AuthUpdate, 0, len(oldByID)+len(newByID))
+	for id, newAuth := range newByID {
+		existing, ok := w.currentAuths[id]
+		if !ok {
+			w.currentAuths[id] = newAuth.Clone()
+			updates = append(updates, AuthUpdate{Action: AuthUpdateActionAdd, ID: id, Auth: newAuth.Clone()})
+			continue
+		}
+		if !authEqual(existing, newAuth) {
+			w.currentAuths[id] = newAuth.Clone()
+			updates = append(updates, AuthUpdate{Action: AuthUpdateActionModify, ID: id, Auth: newAuth.Clone()})
+		}
+	}
+	for id := range oldByID {
+		if _, stillExists := newByID[id]; stillExists {
+			continue
+		}
+		delete(w.currentAuths, id)
+		updates = append(updates, AuthUpdate{Action: AuthUpdateActionDelete, ID: id})
+	}
+	return updates
+}
+
+func authSliceToMap(auths []*coreauth.Auth) map[string]*coreauth.Auth {
+	byID := make(map[string]*coreauth.Auth, len(auths))
+	for _, a := range auths {
+		if a == nil || strings.TrimSpace(a.ID) == "" {
+			continue
+		}
+		byID[a.ID] = a
+	}
+	return byID
 }
 
 func (w *Watcher) loadFileClients(cfg *config.Config) int {

internal/watcher/dispatcher.go

@@ -14,6 +14,8 @@ import (
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
 
+var snapshotCoreAuthsFunc = snapshotCoreAuths
+
 func (w *Watcher) setAuthUpdateQueue(queue chan<- AuthUpdate) {
 	w.clientsMutex.Lock()
 	defer w.clientsMutex.Unlock()
@@ -76,7 +78,11 @@ func (w *Watcher) dispatchRuntimeAuthUpdate(update AuthUpdate) bool {
 }
 
 func (w *Watcher) refreshAuthState(force bool) {
-	auths := w.SnapshotCoreAuths()
+	w.clientsMutex.RLock()
+	cfg := w.config
+	authDir := w.authDir
+	w.clientsMutex.RUnlock()
+	auths := snapshotCoreAuthsFunc(cfg, authDir)
 	w.clientsMutex.Lock()
 	if len(w.runtimeAuths) > 0 {
 		for _, a := range w.runtimeAuths {

internal/watcher/synthesizer/file.go

@@ -36,9 +36,6 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 		return out, nil
 	}
 
-	now := ctx.Now
-	cfg := ctx.Config
-
 	for _, e := range entries {
 		if e.IsDir() {
 			continue
@@ -52,99 +49,120 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 		if errRead != nil || len(data) == 0 {
 			continue
 		}
-		var metadata map[string]any
-		if errUnmarshal := json.Unmarshal(data, &metadata); errUnmarshal != nil {
+		auths := synthesizeFileAuths(ctx, full, data)
+		if len(auths) == 0 {
 			continue
 		}
-		t, _ := metadata["type"].(string)
-		if t == "" {
-			continue
-		}
-		provider := strings.ToLower(t)
-		if provider == "gemini" {
-			provider = "gemini-cli"
-		}
-		label := provider
-		if email, _ := metadata["email"].(string); email != "" {
-			label = email
-		}
-		// Use relative path under authDir as ID to stay consistent with the file-based token store
-		id := full
-		if rel, errRel := filepath.Rel(ctx.AuthDir, full); errRel == nil && rel != "" {
-			id = rel
-		}
-		// On Windows, normalize ID casing to avoid duplicate auth entries caused by case-insensitive paths.
-		if runtime.GOOS == "windows" {
-			id = strings.ToLower(id)
-		}
-
-		proxyURL := ""
-		if p, ok := metadata["proxy_url"].(string); ok {
-			proxyURL = p
-		}
-
-		prefix := ""
-		if rawPrefix, ok := metadata["prefix"].(string); ok {
-			trimmed := strings.TrimSpace(rawPrefix)
-			trimmed = strings.Trim(trimmed, "/")
-			if trimmed != "" && !strings.Contains(trimmed, "/") {
-				prefix = trimmed
-			}
-		}
-
-		disabled, _ := metadata["disabled"].(bool)
-		status := coreauth.StatusActive
-		if disabled {
-			status = coreauth.StatusDisabled
-		}
-
-		// Read per-account excluded models from the OAuth JSON file
-		perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
-
-		a := &coreauth.Auth{
-			ID:       id,
-			Provider: provider,
-			Label:    label,
-			Prefix:   prefix,
-			Status:   status,
-			Disabled: disabled,
-			Attributes: map[string]string{
-				"source": full,
-				"path":   full,
-			},
-			ProxyURL:  proxyURL,
-			Metadata:  metadata,
-			CreatedAt: now,
-			UpdatedAt: now,
-		}
-		// Read priority from auth file
-		if rawPriority, ok := metadata["priority"]; ok {
-			switch v := rawPriority.(type) {
-			case float64:
-				a.Attributes["priority"] = strconv.Itoa(int(v))
-			case string:
-				priority := strings.TrimSpace(v)
-				if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
-					a.Attributes["priority"] = priority
-				}
-			}
-		}
-		ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
-		if provider == "gemini-cli" {
-			if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
-				for _, v := range virtuals {
-					ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
-				}
-				out = append(out, a)
-				out = append(out, virtuals...)
-				continue
-			}
-		}
-		out = append(out, a)
+		out = append(out, auths...)
 	}
 	return out, nil
 }
 
+// SynthesizeAuthFile generates Auth entries for one auth JSON file payload.
+// It shares exactly the same mapping behavior as FileSynthesizer.Synthesize.
+func SynthesizeAuthFile(ctx *SynthesisContext, fullPath string, data []byte) []*coreauth.Auth {
+	return synthesizeFileAuths(ctx, fullPath, data)
+}
+
+func synthesizeFileAuths(ctx *SynthesisContext, fullPath string, data []byte) []*coreauth.Auth {
+	if ctx == nil || len(data) == 0 {
+		return nil
+	}
+	now := ctx.Now
+	cfg := ctx.Config
+	var metadata map[string]any
+	if errUnmarshal := json.Unmarshal(data, &metadata); errUnmarshal != nil {
+		return nil
+	}
+	t, _ := metadata["type"].(string)
+	if t == "" {
+		return nil
+	}
+	provider := strings.ToLower(t)
+	if provider == "gemini" {
+		provider = "gemini-cli"
+	}
+	label := provider
+	if email, _ := metadata["email"].(string); email != "" {
+		label = email
+	}
+	// Use relative path under authDir as ID to stay consistent with the file-based token store.
+	id := fullPath
+	if strings.TrimSpace(ctx.AuthDir) != "" {
+		if rel, errRel := filepath.Rel(ctx.AuthDir, fullPath); errRel == nil && rel != "" {
+			id = rel
+		}
+	}
+	if runtime.GOOS == "windows" {
+		id = strings.ToLower(id)
+	}
+
+	proxyURL := ""
+	if p, ok := metadata["proxy_url"].(string); ok {
+		proxyURL = p
+	}
+
+	prefix := ""
+	if rawPrefix, ok := metadata["prefix"].(string); ok {
+		trimmed := strings.TrimSpace(rawPrefix)
+		trimmed = strings.Trim(trimmed, "/")
+		if trimmed != "" && !strings.Contains(trimmed, "/") {
+			prefix = trimmed
+		}
+	}
+
+	disabled, _ := metadata["disabled"].(bool)
+	status := coreauth.StatusActive
+	if disabled {
+		status = coreauth.StatusDisabled
+	}
+
+	// Read per-account excluded models from the OAuth JSON file.
+	perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
+
+	a := &coreauth.Auth{
+		ID:       id,
+		Provider: provider,
+		Label:    label,
+		Prefix:   prefix,
+		Status:   status,
+		Disabled: disabled,
+		Attributes: map[string]string{
+			"source": fullPath,
+			"path":   fullPath,
+		},
+		ProxyURL:  proxyURL,
+		Metadata:  metadata,
+		CreatedAt: now,
+		UpdatedAt: now,
+	}
+	// Read priority from auth file.
+	if rawPriority, ok := metadata["priority"]; ok {
+		switch v := rawPriority.(type) {
+		case float64:
+			a.Attributes["priority"] = strconv.Itoa(int(v))
+		case string:
+			priority := strings.TrimSpace(v)
+			if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
+				a.Attributes["priority"] = priority
+			}
+		}
+	}
+	ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
+	if provider == "gemini-cli" {
+		if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
+			for _, v := range virtuals {
+				ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
+			}
+			out := make([]*coreauth.Auth, 0, 1+len(virtuals))
+			out = append(out, a)
+			out = append(out, virtuals...)
+			return out
+		}
+	}
+	return []*coreauth.Auth{a}
+}
+
 // SynthesizeGeminiVirtualAuths creates virtual Auth entries for multi-project Gemini credentials.
 // It disables the primary auth and creates one virtual auth per project.
 func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]any, now time.Time) []*coreauth.Auth {

internal/watcher/watcher.go

@@ -45,6 +45,7 @@ type Watcher struct {
 	watcher           *fsnotify.Watcher
 	lastAuthHashes    map[string]string
 	lastAuthContents  map[string]*coreauth.Auth
+	fileAuthsByPath   map[string]map[string]*coreauth.Auth
 	lastRemoveTimes   map[string]time.Time
 	lastConfigHash    string
 	authQueue         chan<- AuthUpdate
@@ -92,11 +93,12 @@ func NewWatcher(configPath, authDir string, reloadCallback func(*config.Config))
 		return nil, errNewWatcher
 	}
 	w := &Watcher{
-		configPath:     configPath,
-		authDir:        authDir,
-		reloadCallback: reloadCallback,
-		watcher:        watcher,
-		lastAuthHashes: make(map[string]string),
+		configPath:      configPath,
+		authDir:         authDir,
+		reloadCallback:  reloadCallback,
+		watcher:         watcher,
+		lastAuthHashes:  make(map[string]string),
+		fileAuthsByPath: make(map[string]map[string]*coreauth.Auth),
 	}
 	w.dispatchCond = sync.NewCond(&w.dispatchMu)
 	if store := sdkAuth.GetTokenStore(); store != nil {

internal/watcher/watcher_test.go

@@ -406,8 +406,8 @@ func TestAddOrUpdateClientTriggersReloadAndHash(t *testing.T) {
 
 	w.addOrUpdateClient(authFile)
 
-	if got := atomic.LoadInt32(&reloads); got != 1 {
-		t.Fatalf("expected reload callback once, got %d", got)
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no reload callback for auth update, got %d", got)
 	}
 	// Use normalizeAuthPath to match how addOrUpdateClient stores the key
 	normalized := w.normalizeAuthPath(authFile)
@@ -436,8 +436,110 @@ func TestRemoveClientRemovesHash(t *testing.T) {
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected hash to be removed after deletion")
 	}
-	if got := atomic.LoadInt32(&reloads); got != 1 {
-		t.Fatalf("expected reload callback once, got %d", got)
+	if got := atomic.LoadInt32(&reloads); got != 0 {
+		t.Fatalf("expected no reload callback for auth removal, got %d", got)
+	}
+}
+
+func TestAuthFileEventsDoNotInvokeSnapshotCoreAuths(t *testing.T) {
+	tmpDir := t.TempDir()
+	authFile := filepath.Join(tmpDir, "sample.json")
+	if err := os.WriteFile(authFile, []byte(`{"type":"codex","email":"u@example.com"}`), 0o644); err != nil {
+		t.Fatalf("failed to create auth file: %v", err)
+	}
+
+	origSnapshot := snapshotCoreAuthsFunc
+	var snapshotCalls int32
+	snapshotCoreAuthsFunc = func(cfg *config.Config, authDir string) []*coreauth.Auth {
+		atomic.AddInt32(&snapshotCalls, 1)
+		return origSnapshot(cfg, authDir)
+	}
+	defer func() { snapshotCoreAuthsFunc = origSnapshot }()
+
+	w := &Watcher{
+		authDir:          tmpDir,
+		lastAuthHashes:   make(map[string]string),
+		lastAuthContents: make(map[string]*coreauth.Auth),
+		fileAuthsByPath:  make(map[string]map[string]*coreauth.Auth),
+	}
+	w.SetConfig(&config.Config{AuthDir: tmpDir})
+
+	w.addOrUpdateClient(authFile)
+	w.removeClient(authFile)
+
+	if got := atomic.LoadInt32(&snapshotCalls); got != 0 {
+		t.Fatalf("expected auth file events to avoid full snapshot, got %d calls", got)
+	}
+}
+
+func TestAuthSliceToMap(t *testing.T) {
+	t.Parallel()
+
+	valid1 := &coreauth.Auth{ID: "a"}
+	valid2 := &coreauth.Auth{ID: "b"}
+	dupOld := &coreauth.Auth{ID: "dup", Label: "old"}
+	dupNew := &coreauth.Auth{ID: "dup", Label: "new"}
+	empty := &coreauth.Auth{ID: "  "}
+
+	tests := []struct {
+		name string
+		in   []*coreauth.Auth
+		want map[string]*coreauth.Auth
+	}{
+		{
+			name: "nil input",
+			in:   nil,
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "empty input",
+			in:   []*coreauth.Auth{},
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "filters invalid auths",
+			in:   []*coreauth.Auth{nil, empty},
+			want: map[string]*coreauth.Auth{},
+		},
+		{
+			name: "keeps valid auths",
+			in:   []*coreauth.Auth{valid1, nil, valid2},
+			want: map[string]*coreauth.Auth{"a": valid1, "b": valid2},
+		},
+		{
+			name: "last duplicate wins",
+			in:   []*coreauth.Auth{dupOld, dupNew},
+			want: map[string]*coreauth.Auth{"dup": dupNew},
+		},
+	}
+
+	for _, tc := range tests {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+			got := authSliceToMap(tc.in)
+			if len(tc.want) == 0 {
+				if got == nil {
+					t.Fatal("expected empty map, got nil")
+				}
+				if len(got) != 0 {
+					t.Fatalf("expected empty map, got %#v", got)
+				}
+				return
+			}
+			if len(got) != len(tc.want) {
+				t.Fatalf("unexpected map length: got %d, want %d", len(got), len(tc.want))
+			}
+			for id, wantAuth := range tc.want {
+				gotAuth, ok := got[id]
+				if !ok {
+					t.Fatalf("missing id %q in result map", id)
+				}
+				if !authEqual(gotAuth, wantAuth) {
+					t.Fatalf("unexpected auth for id %q: got %#v, want %#v", id, gotAuth, wantAuth)
+				}
+			}
+		})
 	}
 }
 
@@ -695,8 +797,8 @@ func TestHandleEventRemovesAuthFile(t *testing.T) {
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Remove})
 
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected reload callback once, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected no reload callback for auth removal, got %d", reloads)
 	}
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected hash entry to be removed")
@@ -893,8 +995,8 @@ func TestHandleEventAuthWriteTriggersUpdate(t *testing.T) {
 	w.SetConfig(&config.Config{AuthDir: authDir})
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Write})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected auth write to trigger reload callback, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected auth write to avoid global reload callback, got %d", reloads)
 	}
 }
 
@@ -990,8 +1092,8 @@ func TestHandleEventAtomicReplaceChangedTriggersUpdate(t *testing.T) {
 	w.lastAuthHashes[w.normalizeAuthPath(authFile)] = hexString(oldSum[:])
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Rename})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected changed atomic replace to trigger update, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected changed atomic replace to avoid global reload, got %d", reloads)
 	}
 }
 
@@ -1045,8 +1147,8 @@ func TestHandleEventRemoveKnownFileDeletes(t *testing.T) {
 	w.lastAuthHashes[w.normalizeAuthPath(authFile)] = "hash"
 
 	w.handleEvent(fsnotify.Event{Name: authFile, Op: fsnotify.Remove})
-	if atomic.LoadInt32(&reloads) != 1 {
-		t.Fatalf("expected known remove to trigger reload, got %d", reloads)
+	if atomic.LoadInt32(&reloads) != 0 {
+		t.Fatalf("expected known remove to avoid global reload, got %d", reloads)
 	}
 	if _, ok := w.lastAuthHashes[w.normalizeAuthPath(authFile)]; ok {
 		t.Fatal("expected known auth hash to be deleted")

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -14,7 +14,11 @@ import (
 	"github.com/google/uuid"
 	"github.com/gorilla/websocket"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -26,7 +30,6 @@ const (
 	wsRequestTypeAppend  = "response.append"
 	wsEventTypeError     = "error"
 	wsEventTypeCompleted = "response.completed"
-	wsEventTypeDone      = "response.done"
 	wsDoneMarker         = "[DONE]"
 	wsTurnStateHeader    = "x-codex-turn-state"
 	wsRequestBodyKey     = "REQUEST_BODY_OVERRIDE"
@@ -101,11 +104,17 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 		// )
 		appendWebsocketEvent(&wsBodyLog, "request", payload)
 
-		allowIncrementalInputWithPreviousResponseID := websocketUpstreamSupportsIncrementalInput(nil, nil)
+		allowIncrementalInputWithPreviousResponseID := false
 		if pinnedAuthID != "" && h != nil && h.AuthManager != nil {
 			if pinnedAuth, ok := h.AuthManager.GetByID(pinnedAuthID); ok && pinnedAuth != nil {
 				allowIncrementalInputWithPreviousResponseID = websocketUpstreamSupportsIncrementalInput(pinnedAuth.Attributes, pinnedAuth.Metadata)
 			}
+		} else {
+			requestModelName := strings.TrimSpace(gjson.GetBytes(payload, "model").String())
+			if requestModelName == "" {
+				requestModelName = strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+			}
+			allowIncrementalInputWithPreviousResponseID = h.websocketUpstreamSupportsIncrementalInputForModel(requestModelName)
 		}
 
 		var requestJSON []byte
@@ -140,6 +149,22 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 			}
 			continue
 		}
+		if shouldHandleResponsesWebsocketPrewarmLocally(payload, lastRequest, allowIncrementalInputWithPreviousResponseID) {
+			if updated, errDelete := sjson.DeleteBytes(requestJSON, "generate"); errDelete == nil {
+				requestJSON = updated
+			}
+			if updated, errDelete := sjson.DeleteBytes(updatedLastRequest, "generate"); errDelete == nil {
+				updatedLastRequest = updated
+			}
+			lastRequest = updatedLastRequest
+			lastResponseOutput = []byte("[]")
+			if errWrite := writeResponsesWebsocketSyntheticPrewarm(c, conn, requestJSON, &wsBodyLog, passthroughSessionID); errWrite != nil {
+				wsTerminateErr = errWrite
+				appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errWrite.Error()))
+				return
+			}
+			continue
+		}
 		lastRequest = updatedLastRequest
 
 		modelName := gjson.GetBytes(requestJSON, "model").String()
@@ -340,6 +365,192 @@ func websocketUpstreamSupportsIncrementalInput(attributes map[string]string, met
 	return false
 }
 
+func (h *OpenAIResponsesAPIHandler) websocketUpstreamSupportsIncrementalInputForModel(modelName string) bool {
+	if h == nil || h.AuthManager == nil {
+		return false
+	}
+
+	resolvedModelName := modelName
+	initialSuffix := thinking.ParseSuffix(modelName)
+	if initialSuffix.ModelName == "auto" {
+		resolvedBase := util.ResolveAutoModel(initialSuffix.ModelName)
+		if initialSuffix.HasSuffix {
+			resolvedModelName = fmt.Sprintf("%s(%s)", resolvedBase, initialSuffix.RawSuffix)
+		} else {
+			resolvedModelName = resolvedBase
+		}
+	} else {
+		resolvedModelName = util.ResolveAutoModel(modelName)
+	}
+
+	parsed := thinking.ParseSuffix(resolvedModelName)
+	baseModel := strings.TrimSpace(parsed.ModelName)
+	providers := util.GetProviderName(baseModel)
+	if len(providers) == 0 && baseModel != resolvedModelName {
+		providers = util.GetProviderName(resolvedModelName)
+	}
+	if len(providers) == 0 {
+		return false
+	}
+
+	providerSet := make(map[string]struct{}, len(providers))
+	for i := 0; i < len(providers); i++ {
+		providerKey := strings.TrimSpace(strings.ToLower(providers[i]))
+		if providerKey == "" {
+			continue
+		}
+		providerSet[providerKey] = struct{}{}
+	}
+	if len(providerSet) == 0 {
+		return false
+	}
+
+	modelKey := baseModel
+	if modelKey == "" {
+		modelKey = strings.TrimSpace(resolvedModelName)
+	}
+	registryRef := registry.GetGlobalRegistry()
+	now := time.Now()
+	auths := h.AuthManager.List()
+	for i := 0; i < len(auths); i++ {
+		auth := auths[i]
+		if auth == nil {
+			continue
+		}
+		providerKey := strings.TrimSpace(strings.ToLower(auth.Provider))
+		if _, ok := providerSet[providerKey]; !ok {
+			continue
+		}
+		if modelKey != "" && registryRef != nil && !registryRef.ClientSupportsModel(auth.ID, modelKey) {
+			continue
+		}
+		if !responsesWebsocketAuthAvailableForModel(auth, modelKey, now) {
+			continue
+		}
+		if websocketUpstreamSupportsIncrementalInput(auth.Attributes, auth.Metadata) {
+			return true
+		}
+	}
+	return false
+}
+
+func responsesWebsocketAuthAvailableForModel(auth *coreauth.Auth, modelName string, now time.Time) bool {
+	if auth == nil {
+		return false
+	}
+	if auth.Disabled || auth.Status == coreauth.StatusDisabled {
+		return false
+	}
+	if modelName != "" && len(auth.ModelStates) > 0 {
+		state, ok := auth.ModelStates[modelName]
+		if (!ok || state == nil) && modelName != "" {
+			baseModel := strings.TrimSpace(thinking.ParseSuffix(modelName).ModelName)
+			if baseModel != "" && baseModel != modelName {
+				state, ok = auth.ModelStates[baseModel]
+			}
+		}
+		if ok && state != nil {
+			if state.Status == coreauth.StatusDisabled {
+				return false
+			}
+			if state.Unavailable && !state.NextRetryAfter.IsZero() && state.NextRetryAfter.After(now) {
+				return false
+			}
+			return true
+		}
+	}
+	if auth.Unavailable && !auth.NextRetryAfter.IsZero() && auth.NextRetryAfter.After(now) {
+		return false
+	}
+	return true
+}
+
+func shouldHandleResponsesWebsocketPrewarmLocally(rawJSON []byte, lastRequest []byte, allowIncrementalInputWithPreviousResponseID bool) bool {
+	if allowIncrementalInputWithPreviousResponseID || len(lastRequest) != 0 {
+		return false
+	}
+	if strings.TrimSpace(gjson.GetBytes(rawJSON, "type").String()) != wsRequestTypeCreate {
+		return false
+	}
+	generateResult := gjson.GetBytes(rawJSON, "generate")
+	return generateResult.Exists() && !generateResult.Bool()
+}
+
+func writeResponsesWebsocketSyntheticPrewarm(
+	c *gin.Context,
+	conn *websocket.Conn,
+	requestJSON []byte,
+	wsBodyLog *strings.Builder,
+	sessionID string,
+) error {
+	payloads, errPayloads := syntheticResponsesWebsocketPrewarmPayloads(requestJSON)
+	if errPayloads != nil {
+		return errPayloads
+	}
+	for i := 0; i < len(payloads); i++ {
+		markAPIResponseTimestamp(c)
+		appendWebsocketEvent(wsBodyLog, "response", payloads[i])
+		// log.Infof(
+		// 	"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+		// 	sessionID,
+		// 	websocket.TextMessage,
+		// 	websocketPayloadEventType(payloads[i]),
+		// 	websocketPayloadPreview(payloads[i]),
+		// )
+		if errWrite := conn.WriteMessage(websocket.TextMessage, payloads[i]); errWrite != nil {
+			log.Warnf(
+				"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+				sessionID,
+				websocketPayloadEventType(payloads[i]),
+				errWrite,
+			)
+			return errWrite
+		}
+	}
+	return nil
+}
+
+func syntheticResponsesWebsocketPrewarmPayloads(requestJSON []byte) ([][]byte, error) {
+	responseID := "resp_prewarm_" + uuid.NewString()
+	createdAt := time.Now().Unix()
+	modelName := strings.TrimSpace(gjson.GetBytes(requestJSON, "model").String())
+
+	createdPayload := []byte(`{"type":"response.created","sequence_number":0,"response":{"id":"","object":"response","created_at":0,"status":"in_progress","background":false,"error":null,"output":[]}}`)
+	var errSet error
+	createdPayload, errSet = sjson.SetBytes(createdPayload, "response.id", responseID)
+	if errSet != nil {
+		return nil, errSet
+	}
+	createdPayload, errSet = sjson.SetBytes(createdPayload, "response.created_at", createdAt)
+	if errSet != nil {
+		return nil, errSet
+	}
+	if modelName != "" {
+		createdPayload, errSet = sjson.SetBytes(createdPayload, "response.model", modelName)
+		if errSet != nil {
+			return nil, errSet
+		}
+	}
+
+	completedPayload := []byte(`{"type":"response.completed","sequence_number":1,"response":{"id":"","object":"response","created_at":0,"status":"completed","background":false,"error":null,"output":[],"usage":{"input_tokens":0,"output_tokens":0,"total_tokens":0}}}`)
+	completedPayload, errSet = sjson.SetBytes(completedPayload, "response.id", responseID)
+	if errSet != nil {
+		return nil, errSet
+	}
+	completedPayload, errSet = sjson.SetBytes(completedPayload, "response.created_at", createdAt)
+	if errSet != nil {
+		return nil, errSet
+	}
+	if modelName != "" {
+		completedPayload, errSet = sjson.SetBytes(completedPayload, "response.model", modelName)
+		if errSet != nil {
+			return nil, errSet
+		}
+	}
+
+	return [][]byte{createdPayload, completedPayload}, nil
+}
+
 func mergeJSONArrayRaw(existingRaw, appendRaw string) (string, error) {
 	existingRaw = strings.TrimSpace(existingRaw)
 	appendRaw = strings.TrimSpace(appendRaw)
@@ -469,9 +680,6 @@ func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
 			for i := range payloads {
 				eventType := gjson.GetBytes(payloads[i], "type").String()
 				if eventType == wsEventTypeCompleted {
-					// log.Infof("replace %s with %s", wsEventTypeCompleted, wsEventTypeDone)
-					payloads[i], _ = sjson.SetBytes(payloads[i], "type", wsEventTypeDone)
-
 					completed = true
 					completedOutput = responseCompletedOutputFromPayload(payloads[i])
 				}
@@ -554,47 +762,63 @@ func writeResponsesWebsocketError(conn *websocket.Conn, errMsg *interfaces.Error
 	}
 
 	body := handlers.BuildErrorResponseBody(status, errText)
-	payload := map[string]any{
-		"type":   wsEventTypeError,
-		"status": status,
+	payload := []byte(`{}`)
+	var errSet error
+	payload, errSet = sjson.SetBytes(payload, "type", wsEventTypeError)
+	if errSet != nil {
+		return nil, errSet
+	}
+	payload, errSet = sjson.SetBytes(payload, "status", status)
+	if errSet != nil {
+		return nil, errSet
 	}
 
 	if errMsg != nil && errMsg.Addon != nil {
-		headers := map[string]any{}
+		headers := []byte(`{}`)
+		hasHeaders := false
 		for key, values := range errMsg.Addon {
 			if len(values) == 0 {
 				continue
 			}
-			headers[key] = values[0]
+			headerPath := strings.ReplaceAll(strings.ReplaceAll(key, `\\`, `\\\\`), ".", `\\.`)
+			headers, errSet = sjson.SetBytes(headers, headerPath, values[0])
+			if errSet != nil {
+				return nil, errSet
+			}
+			hasHeaders = true
 		}
-		if len(headers) > 0 {
-			payload["headers"] = headers
-		}
-	}
-
-	if len(body) > 0 && json.Valid(body) {
-		var decoded map[string]any
-		if errDecode := json.Unmarshal(body, &decoded); errDecode == nil {
-			if inner, ok := decoded["error"]; ok {
-				payload["error"] = inner
-			} else {
-				payload["error"] = decoded
+		if hasHeaders {
+			payload, errSet = sjson.SetRawBytes(payload, "headers", headers)
+			if errSet != nil {
+				return nil, errSet
 			}
 		}
 	}
 
-	if _, ok := payload["error"]; !ok {
-		payload["error"] = map[string]any{
-			"type":    "server_error",
-			"message": errText,
+	if len(body) > 0 && json.Valid(body) {
+		errorNode := gjson.GetBytes(body, "error")
+		if errorNode.Exists() {
+			payload, errSet = sjson.SetRawBytes(payload, "error", []byte(errorNode.Raw))
+		} else {
+			payload, errSet = sjson.SetRawBytes(payload, "error", body)
+		}
+		if errSet != nil {
+			return nil, errSet
 		}
 	}
 
-	data, err := json.Marshal(payload)
-	if err != nil {
-		return nil, err
+	if !gjson.GetBytes(payload, "error").Exists() {
+		payload, errSet = sjson.SetBytes(payload, "error.type", "server_error")
+		if errSet != nil {
+			return nil, errSet
+		}
+		payload, errSet = sjson.SetBytes(payload, "error.message", errText)
+		if errSet != nil {
+			return nil, errSet
+		}
 	}
-	return data, conn.WriteMessage(websocket.TextMessage, data)
+
+	return payload, conn.WriteMessage(websocket.TextMessage, payload)
 }
 
 func appendWebsocketEvent(builder *strings.Builder, eventType string, payload []byte) {

sdk/api/handlers/openai/openai_responses_websocket_test.go

@@ -2,15 +2,57 @@ package openai
 
 import (
 	"bytes"
+	"context"
+	"errors"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
 
 	"github.com/gin-gonic/gin"
+	"github.com/gorilla/websocket"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	coreexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
 	"github.com/tidwall/gjson"
 )
 
+type websocketCaptureExecutor struct {
+	streamCalls int
+	payloads    [][]byte
+}
+
+func (e *websocketCaptureExecutor) Identifier() string { return "test-provider" }
+
+func (e *websocketCaptureExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, errors.New("not implemented")
+}
+
+func (e *websocketCaptureExecutor) ExecuteStream(_ context.Context, _ *coreauth.Auth, req coreexecutor.Request, _ coreexecutor.Options) (*coreexecutor.StreamResult, error) {
+	e.streamCalls++
+	e.payloads = append(e.payloads, bytes.Clone(req.Payload))
+	chunks := make(chan coreexecutor.StreamChunk, 1)
+	chunks <- coreexecutor.StreamChunk{Payload: []byte(`{"type":"response.completed","response":{"id":"resp-upstream","output":[{"type":"message","id":"out-1"}]}}`)}
+	close(chunks)
+	return &coreexecutor.StreamResult{Chunks: chunks}, nil
+}
+
+func (e *websocketCaptureExecutor) Refresh(_ context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *websocketCaptureExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, errors.New("not implemented")
+}
+
+func (e *websocketCaptureExecutor) HttpRequest(context.Context, *coreauth.Auth, *http.Request) (*http.Response, error) {
+	return nil, errors.New("not implemented")
+}
+
 func TestNormalizeResponsesWebsocketRequestCreate(t *testing.T) {
 	raw := []byte(`{"type":"response.create","model":"test-model","stream":false,"input":[{"type":"message","id":"msg-1"}]}`)
 
@@ -247,3 +289,206 @@ func TestSetWebsocketRequestBody(t *testing.T) {
 		t.Fatalf("request body = %q, want %q", string(bodyBytes), "event body")
 	}
 }
+
+func TestForwardResponsesWebsocketPreservesCompletedEvent(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	serverErrCh := make(chan error, 1)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		conn, err := responsesWebsocketUpgrader.Upgrade(w, r, nil)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		defer func() {
+			errClose := conn.Close()
+			if errClose != nil {
+				serverErrCh <- errClose
+			}
+		}()
+
+		ctx, _ := gin.CreateTestContext(httptest.NewRecorder())
+		ctx.Request = r
+
+		data := make(chan []byte, 1)
+		errCh := make(chan *interfaces.ErrorMessage)
+		data <- []byte("data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp-1\",\"output\":[{\"type\":\"message\",\"id\":\"out-1\"}]}}\n\n")
+		close(data)
+		close(errCh)
+
+		var bodyLog strings.Builder
+		completedOutput, err := (*OpenAIResponsesAPIHandler)(nil).forwardResponsesWebsocket(
+			ctx,
+			conn,
+			func(...interface{}) {},
+			data,
+			errCh,
+			&bodyLog,
+			"session-1",
+		)
+		if err != nil {
+			serverErrCh <- err
+			return
+		}
+		if gjson.GetBytes(completedOutput, "0.id").String() != "out-1" {
+			serverErrCh <- errors.New("completed output not captured")
+			return
+		}
+		serverErrCh <- nil
+	}))
+	defer server.Close()
+
+	wsURL := "ws" + strings.TrimPrefix(server.URL, "http")
+	conn, _, err := websocket.DefaultDialer.Dial(wsURL, nil)
+	if err != nil {
+		t.Fatalf("dial websocket: %v", err)
+	}
+	defer func() {
+		errClose := conn.Close()
+		if errClose != nil {
+			t.Fatalf("close websocket: %v", errClose)
+		}
+	}()
+
+	_, payload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read websocket message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(payload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("payload type = %s, want %s", gjson.GetBytes(payload, "type").String(), wsEventTypeCompleted)
+	}
+	if strings.Contains(string(payload), "response.done") {
+		t.Fatalf("payload unexpectedly rewrote completed event: %s", payload)
+	}
+
+	if errServer := <-serverErrCh; errServer != nil {
+		t.Fatalf("server error: %v", errServer)
+	}
+}
+
+func TestWebsocketUpstreamSupportsIncrementalInputForModel(t *testing.T) {
+	manager := coreauth.NewManager(nil, nil, nil)
+	auth := &coreauth.Auth{
+		ID:         "auth-ws",
+		Provider:   "test-provider",
+		Status:     coreauth.StatusActive,
+		Attributes: map[string]string{"websockets": "true"},
+	}
+	if _, err := manager.Register(context.Background(), auth); err != nil {
+		t.Fatalf("Register auth: %v", err)
+	}
+	registry.GetGlobalRegistry().RegisterClient(auth.ID, auth.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth.ID)
+	})
+
+	base := handlers.NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, manager)
+	h := NewOpenAIResponsesAPIHandler(base)
+	if !h.websocketUpstreamSupportsIncrementalInputForModel("test-model") {
+		t.Fatalf("expected websocket-capable upstream for test-model")
+	}
+}
+
+func TestResponsesWebsocketPrewarmHandledLocallyForSSEUpstream(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	executor := &websocketCaptureExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+	auth := &coreauth.Auth{ID: "auth-sse", Provider: executor.Identifier(), Status: coreauth.StatusActive}
+	if _, err := manager.Register(context.Background(), auth); err != nil {
+		t.Fatalf("Register auth: %v", err)
+	}
+	registry.GetGlobalRegistry().RegisterClient(auth.ID, auth.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth.ID)
+	})
+
+	base := handlers.NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, manager)
+	h := NewOpenAIResponsesAPIHandler(base)
+	router := gin.New()
+	router.GET("/v1/responses/ws", h.ResponsesWebsocket)
+
+	server := httptest.NewServer(router)
+	defer server.Close()
+
+	wsURL := "ws" + strings.TrimPrefix(server.URL, "http") + "/v1/responses/ws"
+	conn, _, err := websocket.DefaultDialer.Dial(wsURL, nil)
+	if err != nil {
+		t.Fatalf("dial websocket: %v", err)
+	}
+	defer func() {
+		errClose := conn.Close()
+		if errClose != nil {
+			t.Fatalf("close websocket: %v", errClose)
+		}
+	}()
+
+	errWrite := conn.WriteMessage(websocket.TextMessage, []byte(`{"type":"response.create","model":"test-model","generate":false}`))
+	if errWrite != nil {
+		t.Fatalf("write prewarm websocket message: %v", errWrite)
+	}
+
+	_, createdPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read prewarm created message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(createdPayload, "type").String() != "response.created" {
+		t.Fatalf("created payload type = %s, want response.created", gjson.GetBytes(createdPayload, "type").String())
+	}
+	prewarmResponseID := gjson.GetBytes(createdPayload, "response.id").String()
+	if prewarmResponseID == "" {
+		t.Fatalf("prewarm response id is empty")
+	}
+	if executor.streamCalls != 0 {
+		t.Fatalf("stream calls after prewarm = %d, want 0", executor.streamCalls)
+	}
+
+	_, completedPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read prewarm completed message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(completedPayload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("completed payload type = %s, want %s", gjson.GetBytes(completedPayload, "type").String(), wsEventTypeCompleted)
+	}
+	if gjson.GetBytes(completedPayload, "response.id").String() != prewarmResponseID {
+		t.Fatalf("completed response id = %s, want %s", gjson.GetBytes(completedPayload, "response.id").String(), prewarmResponseID)
+	}
+	if gjson.GetBytes(completedPayload, "response.usage.total_tokens").Int() != 0 {
+		t.Fatalf("prewarm total tokens = %d, want 0", gjson.GetBytes(completedPayload, "response.usage.total_tokens").Int())
+	}
+
+	secondRequest := fmt.Sprintf(`{"type":"response.create","previous_response_id":%q,"input":[{"type":"message","id":"msg-1"}]}`, prewarmResponseID)
+	errWrite = conn.WriteMessage(websocket.TextMessage, []byte(secondRequest))
+	if errWrite != nil {
+		t.Fatalf("write follow-up websocket message: %v", errWrite)
+	}
+
+	_, upstreamPayload, errReadMessage := conn.ReadMessage()
+	if errReadMessage != nil {
+		t.Fatalf("read upstream completed message: %v", errReadMessage)
+	}
+	if gjson.GetBytes(upstreamPayload, "type").String() != wsEventTypeCompleted {
+		t.Fatalf("upstream payload type = %s, want %s", gjson.GetBytes(upstreamPayload, "type").String(), wsEventTypeCompleted)
+	}
+	if executor.streamCalls != 1 {
+		t.Fatalf("stream calls after follow-up = %d, want 1", executor.streamCalls)
+	}
+	if len(executor.payloads) != 1 {
+		t.Fatalf("captured upstream payloads = %d, want 1", len(executor.payloads))
+	}
+	forwarded := executor.payloads[0]
+	if gjson.GetBytes(forwarded, "previous_response_id").Exists() {
+		t.Fatalf("previous_response_id leaked upstream: %s", forwarded)
+	}
+	if gjson.GetBytes(forwarded, "generate").Exists() {
+		t.Fatalf("generate leaked upstream: %s", forwarded)
+	}
+	if gjson.GetBytes(forwarded, "model").String() != "test-model" {
+		t.Fatalf("forwarded model = %s, want test-model", gjson.GetBytes(forwarded, "model").String())
+	}
+	input := gjson.GetBytes(forwarded, "input").Array()
+	if len(input) != 1 || input[0].Get("id").String() != "msg-1" {
+		t.Fatalf("unexpected forwarded input: %s", forwarded)
+	}
+}

sdk/cliproxy/auth/conductor.go

@@ -134,6 +134,7 @@ type Manager struct {
 	hook      Hook
 	mu        sync.RWMutex
 	auths     map[string]*Auth
+	scheduler *authScheduler
 	// providerOffsets tracks per-model provider rotation state for multi-provider routing.
 	providerOffsets map[string]int
 
@@ -149,6 +150,9 @@ type Manager struct {
 	// Keyed by auth.ID, value is alias(lower) -> upstream model (including suffix).
 	apiKeyModelAlias atomic.Value
 
+	// modelPoolOffsets tracks per-auth alias pool rotation state.
+	modelPoolOffsets map[string]int
+
 	// runtimeConfig stores the latest application config for request-time decisions.
 	// It is initialized in NewManager; never Load() before first Store().
 	runtimeConfig atomic.Value
@@ -176,14 +180,39 @@ func NewManager(store Store, selector Selector, hook Hook) *Manager {
 		hook:             hook,
 		auths:            make(map[string]*Auth),
 		providerOffsets:  make(map[string]int),
+		modelPoolOffsets: make(map[string]int),
 		refreshSemaphore: make(chan struct{}, refreshMaxConcurrency),
 	}
 	// atomic.Value requires non-nil initial value.
 	manager.runtimeConfig.Store(&internalconfig.Config{})
 	manager.apiKeyModelAlias.Store(apiKeyModelAliasTable(nil))
+	manager.scheduler = newAuthScheduler(selector)
 	return manager
 }
 
+func isBuiltInSelector(selector Selector) bool {
+	switch selector.(type) {
+	case *RoundRobinSelector, *FillFirstSelector:
+		return true
+	default:
+		return false
+	}
+}
+
+func (m *Manager) syncSchedulerFromSnapshot(auths []*Auth) {
+	if m == nil || m.scheduler == nil {
+		return
+	}
+	m.scheduler.rebuild(auths)
+}
+
+func (m *Manager) syncScheduler() {
+	if m == nil || m.scheduler == nil {
+		return
+	}
+	m.syncSchedulerFromSnapshot(m.snapshotAuths())
+}
+
 func (m *Manager) SetSelector(selector Selector) {
 	if m == nil {
 		return
@@ -194,6 +223,10 @@ func (m *Manager) SetSelector(selector Selector) {
 	m.mu.Lock()
 	m.selector = selector
 	m.mu.Unlock()
+	if m.scheduler != nil {
+		m.scheduler.setSelector(selector)
+		m.syncScheduler()
+	}
 }
 
 // SetStore swaps the underlying persistence store.
@@ -251,16 +284,323 @@ func (m *Manager) lookupAPIKeyUpstreamModel(authID, requestedModel string) strin
 	if resolved == "" {
 		return ""
 	}
-	// Preserve thinking suffix from the client's requested model unless config already has one.
-	requestResult := thinking.ParseSuffix(requestedModel)
-	if thinking.ParseSuffix(resolved).HasSuffix {
-		return resolved
-	}
-	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
-		return resolved + "(" + requestResult.RawSuffix + ")"
-	}
-	return resolved
+	return preserveRequestedModelSuffix(requestedModel, resolved)
+}
 
+func isAPIKeyAuth(auth *Auth) bool {
+	if auth == nil {
+		return false
+	}
+	kind, _ := auth.AccountInfo()
+	return strings.EqualFold(strings.TrimSpace(kind), "api_key")
+}
+
+func isOpenAICompatAPIKeyAuth(auth *Auth) bool {
+	if !isAPIKeyAuth(auth) {
+		return false
+	}
+	if strings.EqualFold(strings.TrimSpace(auth.Provider), "openai-compatibility") {
+		return true
+	}
+	if auth.Attributes == nil {
+		return false
+	}
+	return strings.TrimSpace(auth.Attributes["compat_name"]) != ""
+}
+
+func openAICompatProviderKey(auth *Auth) string {
+	if auth == nil {
+		return ""
+	}
+	if auth.Attributes != nil {
+		if providerKey := strings.TrimSpace(auth.Attributes["provider_key"]); providerKey != "" {
+			return strings.ToLower(providerKey)
+		}
+		if compatName := strings.TrimSpace(auth.Attributes["compat_name"]); compatName != "" {
+			return strings.ToLower(compatName)
+		}
+	}
+	return strings.ToLower(strings.TrimSpace(auth.Provider))
+}
+
+func openAICompatModelPoolKey(auth *Auth, requestedModel string) string {
+	base := strings.TrimSpace(thinking.ParseSuffix(requestedModel).ModelName)
+	if base == "" {
+		base = strings.TrimSpace(requestedModel)
+	}
+	return strings.ToLower(strings.TrimSpace(auth.ID)) + "|" + openAICompatProviderKey(auth) + "|" + strings.ToLower(base)
+}
+
+func (m *Manager) nextModelPoolOffset(key string, size int) int {
+	if m == nil || size <= 1 {
+		return 0
+	}
+	key = strings.TrimSpace(key)
+	if key == "" {
+		return 0
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if m.modelPoolOffsets == nil {
+		m.modelPoolOffsets = make(map[string]int)
+	}
+	offset := m.modelPoolOffsets[key]
+	if offset >= 2_147_483_640 {
+		offset = 0
+	}
+	m.modelPoolOffsets[key] = offset + 1
+	if size <= 0 {
+		return 0
+	}
+	return offset % size
+}
+
+func rotateStrings(values []string, offset int) []string {
+	if len(values) <= 1 {
+		return values
+	}
+	if offset <= 0 {
+		out := make([]string, len(values))
+		copy(out, values)
+		return out
+	}
+	offset = offset % len(values)
+	out := make([]string, 0, len(values))
+	out = append(out, values[offset:]...)
+	out = append(out, values[:offset]...)
+	return out
+}
+
+func (m *Manager) resolveOpenAICompatUpstreamModelPool(auth *Auth, requestedModel string) []string {
+	if m == nil || !isOpenAICompatAPIKeyAuth(auth) {
+		return nil
+	}
+	requestedModel = strings.TrimSpace(requestedModel)
+	if requestedModel == "" {
+		return nil
+	}
+	cfg, _ := m.runtimeConfig.Load().(*internalconfig.Config)
+	if cfg == nil {
+		cfg = &internalconfig.Config{}
+	}
+	providerKey := ""
+	compatName := ""
+	if auth.Attributes != nil {
+		providerKey = strings.TrimSpace(auth.Attributes["provider_key"])
+		compatName = strings.TrimSpace(auth.Attributes["compat_name"])
+	}
+	entry := resolveOpenAICompatConfig(cfg, providerKey, compatName, auth.Provider)
+	if entry == nil {
+		return nil
+	}
+	return resolveModelAliasPoolFromConfigModels(requestedModel, asModelAliasEntries(entry.Models))
+}
+
+func preserveRequestedModelSuffix(requestedModel, resolved string) string {
+	return preserveResolvedModelSuffix(resolved, thinking.ParseSuffix(requestedModel))
+}
+
+func (m *Manager) executionModelCandidates(auth *Auth, routeModel string) []string {
+	return m.prepareExecutionModels(auth, routeModel)
+}
+
+func (m *Manager) prepareExecutionModels(auth *Auth, routeModel string) []string {
+	requestedModel := rewriteModelForAuth(routeModel, auth)
+	requestedModel = m.applyOAuthModelAlias(auth, requestedModel)
+	if pool := m.resolveOpenAICompatUpstreamModelPool(auth, requestedModel); len(pool) > 0 {
+		if len(pool) == 1 {
+			return pool
+		}
+		offset := m.nextModelPoolOffset(openAICompatModelPoolKey(auth, requestedModel), len(pool))
+		return rotateStrings(pool, offset)
+	}
+	resolved := m.applyAPIKeyModelAlias(auth, requestedModel)
+	if strings.TrimSpace(resolved) == "" {
+		resolved = requestedModel
+	}
+	return []string{resolved}
+}
+
+func discardStreamChunks(ch <-chan cliproxyexecutor.StreamChunk) {
+	if ch == nil {
+		return
+	}
+	go func() {
+		for range ch {
+		}
+	}()
+}
+
+func readStreamBootstrap(ctx context.Context, ch <-chan cliproxyexecutor.StreamChunk) ([]cliproxyexecutor.StreamChunk, bool, error) {
+	if ch == nil {
+		return nil, true, nil
+	}
+	buffered := make([]cliproxyexecutor.StreamChunk, 0, 1)
+	for {
+		var (
+			chunk cliproxyexecutor.StreamChunk
+			ok    bool
+		)
+		if ctx != nil {
+			select {
+			case <-ctx.Done():
+				return nil, false, ctx.Err()
+			case chunk, ok = <-ch:
+			}
+		} else {
+			chunk, ok = <-ch
+		}
+		if !ok {
+			return buffered, true, nil
+		}
+		if chunk.Err != nil {
+			return nil, false, chunk.Err
+		}
+		buffered = append(buffered, chunk)
+		if len(chunk.Payload) > 0 {
+			return buffered, false, nil
+		}
+	}
+}
+
+func (m *Manager) wrapStreamResult(ctx context.Context, auth *Auth, provider, routeModel string, headers http.Header, buffered []cliproxyexecutor.StreamChunk, remaining <-chan cliproxyexecutor.StreamChunk) *cliproxyexecutor.StreamResult {
+	out := make(chan cliproxyexecutor.StreamChunk)
+	go func() {
+		defer close(out)
+		var failed bool
+		forward := true
+		emit := func(chunk cliproxyexecutor.StreamChunk) bool {
+			if chunk.Err != nil && !failed {
+				failed = true
+				rerr := &Error{Message: chunk.Err.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				m.MarkResult(ctx, Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr})
+			}
+			if !forward {
+				return false
+			}
+			if ctx == nil {
+				out <- chunk
+				return true
+			}
+			select {
+			case <-ctx.Done():
+				forward = false
+				return false
+			case out <- chunk:
+				return true
+			}
+		}
+		for _, chunk := range buffered {
+			if ok := emit(chunk); !ok {
+				discardStreamChunks(remaining)
+				return
+			}
+		}
+		for chunk := range remaining {
+			if ok := emit(chunk); !ok {
+				discardStreamChunks(remaining)
+				return
+			}
+		}
+		if !failed {
+			m.MarkResult(ctx, Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: true})
+		}
+	}()
+	return &cliproxyexecutor.StreamResult{Headers: headers, Chunks: out}
+}
+
+func (m *Manager) executeStreamWithModelPool(ctx context.Context, executor ProviderExecutor, auth *Auth, provider string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, routeModel string) (*cliproxyexecutor.StreamResult, error) {
+	if executor == nil {
+		return nil, &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	execModels := m.prepareExecutionModels(auth, routeModel)
+	var lastErr error
+	for idx, execModel := range execModels {
+		execReq := req
+		execReq.Model = execModel
+		streamResult, errStream := executor.ExecuteStream(ctx, auth, execReq, opts)
+		if errStream != nil {
+			if errCtx := ctx.Err(); errCtx != nil {
+				return nil, errCtx
+			}
+			rerr := &Error{Message: errStream.Error()}
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
+				rerr.HTTPStatus = se.StatusCode()
+			}
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+			result.RetryAfter = retryAfterFromError(errStream)
+			m.MarkResult(ctx, result)
+			if isRequestInvalidError(errStream) {
+				return nil, errStream
+			}
+			lastErr = errStream
+			continue
+		}
+
+		buffered, closed, bootstrapErr := readStreamBootstrap(ctx, streamResult.Chunks)
+		if bootstrapErr != nil {
+			if errCtx := ctx.Err(); errCtx != nil {
+				discardStreamChunks(streamResult.Chunks)
+				return nil, errCtx
+			}
+			if isRequestInvalidError(bootstrapErr) {
+				rerr := &Error{Message: bootstrapErr.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](bootstrapErr); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+				result.RetryAfter = retryAfterFromError(bootstrapErr)
+				m.MarkResult(ctx, result)
+				discardStreamChunks(streamResult.Chunks)
+				return nil, bootstrapErr
+			}
+			if idx < len(execModels)-1 {
+				rerr := &Error{Message: bootstrapErr.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](bootstrapErr); ok && se != nil {
+					rerr.HTTPStatus = se.StatusCode()
+				}
+				result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
+				result.RetryAfter = retryAfterFromError(bootstrapErr)
+				m.MarkResult(ctx, result)
+				discardStreamChunks(streamResult.Chunks)
+				lastErr = bootstrapErr
+				continue
+			}
+			errCh := make(chan cliproxyexecutor.StreamChunk, 1)
+			errCh <- cliproxyexecutor.StreamChunk{Err: bootstrapErr}
+			close(errCh)
+			return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, nil, errCh), nil
+		}
+
+		if closed && len(buffered) == 0 {
+			emptyErr := &Error{Code: "empty_stream", Message: "upstream stream closed before first payload", Retryable: true}
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: emptyErr}
+			m.MarkResult(ctx, result)
+			if idx < len(execModels)-1 {
+				lastErr = emptyErr
+				continue
+			}
+			errCh := make(chan cliproxyexecutor.StreamChunk, 1)
+			errCh <- cliproxyexecutor.StreamChunk{Err: emptyErr}
+			close(errCh)
+			return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, nil, errCh), nil
+		}
+
+		remaining := streamResult.Chunks
+		if closed {
+			closedCh := make(chan cliproxyexecutor.StreamChunk)
+			close(closedCh)
+			remaining = closedCh
+		}
+		return m.wrapStreamResult(ctx, auth.Clone(), provider, routeModel, streamResult.Headers, buffered, remaining), nil
+	}
+	if lastErr == nil {
+		lastErr = &Error{Code: "auth_not_found", Message: "no upstream model available"}
+	}
+	return nil, lastErr
 }
 
 func (m *Manager) rebuildAPIKeyModelAliasFromRuntimeConfig() {
@@ -448,10 +788,14 @@ func (m *Manager) Register(ctx context.Context, auth *Auth) (*Auth, error) {
 		auth.ID = uuid.NewString()
 	}
 	auth.EnsureIndex()
+	authClone := auth.Clone()
 	m.mu.Lock()
-	m.auths[auth.ID] = auth.Clone()
+	m.auths[auth.ID] = authClone
 	m.mu.Unlock()
 	m.rebuildAPIKeyModelAliasFromRuntimeConfig()
+	if m.scheduler != nil {
+		m.scheduler.upsertAuth(authClone)
+	}
 	_ = m.persist(ctx, auth)
 	m.hook.OnAuthRegistered(ctx, auth.Clone())
 	return auth.Clone(), nil
@@ -473,9 +817,13 @@ func (m *Manager) Update(ctx context.Context, auth *Auth) (*Auth, error) {
 		}
 	}
 	auth.EnsureIndex()
-	m.auths[auth.ID] = auth.Clone()
+	authClone := auth.Clone()
+	m.auths[auth.ID] = authClone
 	m.mu.Unlock()
 	m.rebuildAPIKeyModelAliasFromRuntimeConfig()
+	if m.scheduler != nil {
+		m.scheduler.upsertAuth(authClone)
+	}
 	_ = m.persist(ctx, auth)
 	m.hook.OnAuthUpdated(ctx, auth.Clone())
 	return auth.Clone(), nil
@@ -484,12 +832,13 @@ func (m *Manager) Update(ctx context.Context, auth *Auth) (*Auth, error) {
 // Load resets manager state from the backing store.
 func (m *Manager) Load(ctx context.Context) error {
 	m.mu.Lock()
-	defer m.mu.Unlock()
 	if m.store == nil {
+		m.mu.Unlock()
 		return nil
 	}
 	items, err := m.store.List(ctx)
 	if err != nil {
+		m.mu.Unlock()
 		return err
 	}
 	m.auths = make(map[string]*Auth, len(items))
@@ -505,6 +854,8 @@ func (m *Manager) Load(ctx context.Context) error {
 		cfg = &internalconfig.Config{}
 	}
 	m.rebuildAPIKeyModelAliasLocked(cfg)
+	m.mu.Unlock()
+	m.syncScheduler()
 	return nil
 }
 
@@ -634,32 +985,42 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		resp, errExec := executor.Execute(execCtx, auth, execReq, opts)
-		result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
-		if errExec != nil {
-			if errCtx := execCtx.Err(); errCtx != nil {
-				return cliproxyexecutor.Response{}, errCtx
-			}
-			result.Error = &Error{Message: errExec.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
-				result.Error.HTTPStatus = se.StatusCode()
-			}
-			if ra := retryAfterFromError(errExec); ra != nil {
-				result.RetryAfter = ra
+
+		models := m.prepareExecutionModels(auth, routeModel)
+		var authErr error
+		for _, upstreamModel := range models {
+			execReq := req
+			execReq.Model = upstreamModel
+			resp, errExec := executor.Execute(execCtx, auth, execReq, opts)
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
+			if errExec != nil {
+				if errCtx := execCtx.Err(); errCtx != nil {
+					return cliproxyexecutor.Response{}, errCtx
+				}
+				result.Error = &Error{Message: errExec.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
+					result.Error.HTTPStatus = se.StatusCode()
+				}
+				if ra := retryAfterFromError(errExec); ra != nil {
+					result.RetryAfter = ra
+				}
+				m.MarkResult(execCtx, result)
+				if isRequestInvalidError(errExec) {
+					return cliproxyexecutor.Response{}, errExec
+				}
+				authErr = errExec
+				continue
 			}
 			m.MarkResult(execCtx, result)
-			if isRequestInvalidError(errExec) {
-				return cliproxyexecutor.Response{}, errExec
+			return resp, nil
+		}
+		if authErr != nil {
+			if isRequestInvalidError(authErr) {
+				return cliproxyexecutor.Response{}, authErr
 			}
-			lastErr = errExec
+			lastErr = authErr
 			continue
 		}
-		m.MarkResult(execCtx, result)
-		return resp, nil
 	}
 }
 
@@ -696,32 +1057,42 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		resp, errExec := executor.CountTokens(execCtx, auth, execReq, opts)
-		result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
-		if errExec != nil {
-			if errCtx := execCtx.Err(); errCtx != nil {
-				return cliproxyexecutor.Response{}, errCtx
-			}
-			result.Error = &Error{Message: errExec.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
-				result.Error.HTTPStatus = se.StatusCode()
-			}
-			if ra := retryAfterFromError(errExec); ra != nil {
-				result.RetryAfter = ra
+
+		models := m.prepareExecutionModels(auth, routeModel)
+		var authErr error
+		for _, upstreamModel := range models {
+			execReq := req
+			execReq.Model = upstreamModel
+			resp, errExec := executor.CountTokens(execCtx, auth, execReq, opts)
+			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: errExec == nil}
+			if errExec != nil {
+				if errCtx := execCtx.Err(); errCtx != nil {
+					return cliproxyexecutor.Response{}, errCtx
+				}
+				result.Error = &Error{Message: errExec.Error()}
+				if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
+					result.Error.HTTPStatus = se.StatusCode()
+				}
+				if ra := retryAfterFromError(errExec); ra != nil {
+					result.RetryAfter = ra
+				}
+				m.hook.OnResult(execCtx, result)
+				if isRequestInvalidError(errExec) {
+					return cliproxyexecutor.Response{}, errExec
+				}
+				authErr = errExec
+				continue
 			}
 			m.hook.OnResult(execCtx, result)
-			if isRequestInvalidError(errExec) {
-				return cliproxyexecutor.Response{}, errExec
+			return resp, nil
+		}
+		if authErr != nil {
+			if isRequestInvalidError(authErr) {
+				return cliproxyexecutor.Response{}, authErr
 			}
-			lastErr = errExec
+			lastErr = authErr
 			continue
 		}
-		m.hook.OnResult(execCtx, result)
-		return resp, nil
 	}
 }
 
@@ -758,63 +1129,18 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 			execCtx = context.WithValue(execCtx, roundTripperContextKey{}, rt)
 			execCtx = context.WithValue(execCtx, "cliproxy.roundtripper", rt)
 		}
-		execReq := req
-		execReq.Model = rewriteModelForAuth(routeModel, auth)
-		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
-		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		streamResult, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
+		streamResult, errStream := m.executeStreamWithModelPool(execCtx, executor, auth, provider, req, opts, routeModel)
 		if errStream != nil {
 			if errCtx := execCtx.Err(); errCtx != nil {
 				return nil, errCtx
 			}
-			rerr := &Error{Message: errStream.Error()}
-			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
-				rerr.HTTPStatus = se.StatusCode()
-			}
-			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
-			result.RetryAfter = retryAfterFromError(errStream)
-			m.MarkResult(execCtx, result)
 			if isRequestInvalidError(errStream) {
 				return nil, errStream
 			}
 			lastErr = errStream
 			continue
 		}
-		out := make(chan cliproxyexecutor.StreamChunk)
-		go func(streamCtx context.Context, streamAuth *Auth, streamProvider string, streamChunks <-chan cliproxyexecutor.StreamChunk) {
-			defer close(out)
-			var failed bool
-			forward := true
-			for chunk := range streamChunks {
-				if chunk.Err != nil && !failed {
-					failed = true
-					rerr := &Error{Message: chunk.Err.Error()}
-					if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
-						rerr.HTTPStatus = se.StatusCode()
-					}
-					m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: false, Error: rerr})
-				}
-				if !forward {
-					continue
-				}
-				if streamCtx == nil {
-					out <- chunk
-					continue
-				}
-				select {
-				case <-streamCtx.Done():
-					forward = false
-				case out <- chunk:
-				}
-			}
-			if !failed {
-				m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: true})
-			}
-		}(execCtx, auth.Clone(), provider, streamResult.Chunks)
-		return &cliproxyexecutor.StreamResult{
-			Headers: streamResult.Headers,
-			Chunks:  out,
-		}, nil
+		return streamResult, nil
 	}
 }
 
@@ -1245,6 +1571,7 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 	suspendReason := ""
 	clearModelQuota := false
 	setModelQuota := false
+	var authSnapshot *Auth
 
 	m.mu.Lock()
 	if auth, ok := m.auths[result.AuthID]; ok && auth != nil {
@@ -1338,8 +1665,12 @@ func (m *Manager) MarkResult(ctx context.Context, result Result) {
 		}
 
 		_ = m.persist(ctx, auth)
+		authSnapshot = auth.Clone()
 	}
 	m.mu.Unlock()
+	if m.scheduler != nil && authSnapshot != nil {
+		m.scheduler.upsertAuth(authSnapshot)
+	}
 
 	if clearModelQuota && result.Model != "" {
 		registry.GetGlobalRegistry().ClearModelQuotaExceeded(result.AuthID, result.Model)
@@ -1533,18 +1864,22 @@ func statusCodeFromResult(err *Error) int {
 }
 
 // isRequestInvalidError returns true if the error represents a client request
-// error that should not be retried. Specifically, it checks for 400 Bad Request
-// with "invalid_request_error" in the message, indicating the request itself is
-// malformed and switching to a different auth will not help.
+// error that should not be retried. Specifically, it treats 400 responses with
+// "invalid_request_error" and all 422 responses as request-shape failures,
+// where switching auths or pooled upstream models will not help.
 func isRequestInvalidError(err error) bool {
 	if err == nil {
 		return false
 	}
 	status := statusCodeFromError(err)
-	if status != http.StatusBadRequest {
+	switch status {
+	case http.StatusBadRequest:
+		return strings.Contains(err.Error(), "invalid_request_error")
+	case http.StatusUnprocessableEntity:
+		return true
+	default:
 		return false
 	}
-	return strings.Contains(err.Error(), "invalid_request_error")
 }
 
 func applyAuthFailureState(auth *Auth, resultErr *Error, retryAfter *time.Duration, now time.Time) {
@@ -1692,7 +2027,25 @@ func (m *Manager) CloseExecutionSession(sessionID string) {
 	}
 }
 
-func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
+func (m *Manager) useSchedulerFastPath() bool {
+	if m == nil || m.scheduler == nil {
+		return false
+	}
+	return isBuiltInSelector(m.selector)
+}
+
+func shouldRetrySchedulerPick(err error) bool {
+	if err == nil {
+		return false
+	}
+	var authErr *Error
+	if !errors.As(err, &authErr) || authErr == nil {
+		return false
+	}
+	return authErr.Code == "auth_not_found" || authErr.Code == "auth_unavailable"
+}
+
+func (m *Manager) pickNextLegacy(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
 	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
 
 	m.mu.RLock()
@@ -1752,7 +2105,38 @@ func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cli
 	return authCopy, executor, nil
 }
 
-func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
+func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
+	if !m.useSchedulerFastPath() {
+		return m.pickNextLegacy(ctx, provider, model, opts, tried)
+	}
+	executor, okExecutor := m.Executor(provider)
+	if !okExecutor {
+		return nil, nil, &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	selected, errPick := m.scheduler.pickSingle(ctx, provider, model, opts, tried)
+	if errPick != nil && model != "" && shouldRetrySchedulerPick(errPick) {
+		m.syncScheduler()
+		selected, errPick = m.scheduler.pickSingle(ctx, provider, model, opts, tried)
+	}
+	if errPick != nil {
+		return nil, nil, errPick
+	}
+	if selected == nil {
+		return nil, nil, &Error{Code: "auth_not_found", Message: "selector returned no auth"}
+	}
+	authCopy := selected.Clone()
+	if !selected.indexAssigned {
+		m.mu.Lock()
+		if current := m.auths[authCopy.ID]; current != nil && !current.indexAssigned {
+			current.EnsureIndex()
+			authCopy = current.Clone()
+		}
+		m.mu.Unlock()
+	}
+	return authCopy, executor, nil
+}
+
+func (m *Manager) pickNextMixedLegacy(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
 	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
 
 	providerSet := make(map[string]struct{}, len(providers))
@@ -1835,6 +2219,58 @@ func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model s
 	return authCopy, executor, providerKey, nil
 }
 
+func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
+	if !m.useSchedulerFastPath() {
+		return m.pickNextMixedLegacy(ctx, providers, model, opts, tried)
+	}
+
+	eligibleProviders := make([]string, 0, len(providers))
+	seenProviders := make(map[string]struct{}, len(providers))
+	for _, provider := range providers {
+		providerKey := strings.TrimSpace(strings.ToLower(provider))
+		if providerKey == "" {
+			continue
+		}
+		if _, seen := seenProviders[providerKey]; seen {
+			continue
+		}
+		if _, okExecutor := m.Executor(providerKey); !okExecutor {
+			continue
+		}
+		seenProviders[providerKey] = struct{}{}
+		eligibleProviders = append(eligibleProviders, providerKey)
+	}
+	if len(eligibleProviders) == 0 {
+		return nil, nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+
+	selected, providerKey, errPick := m.scheduler.pickMixed(ctx, eligibleProviders, model, opts, tried)
+	if errPick != nil && model != "" && shouldRetrySchedulerPick(errPick) {
+		m.syncScheduler()
+		selected, providerKey, errPick = m.scheduler.pickMixed(ctx, eligibleProviders, model, opts, tried)
+	}
+	if errPick != nil {
+		return nil, nil, "", errPick
+	}
+	if selected == nil {
+		return nil, nil, "", &Error{Code: "auth_not_found", Message: "selector returned no auth"}
+	}
+	executor, okExecutor := m.Executor(providerKey)
+	if !okExecutor {
+		return nil, nil, "", &Error{Code: "executor_not_found", Message: "executor not registered"}
+	}
+	authCopy := selected.Clone()
+	if !selected.indexAssigned {
+		m.mu.Lock()
+		if current := m.auths[authCopy.ID]; current != nil && !current.indexAssigned {
+			current.EnsureIndex()
+			authCopy = current.Clone()
+		}
+		m.mu.Unlock()
+	}
+	return authCopy, executor, providerKey, nil
+}
+
 func (m *Manager) persist(ctx context.Context, auth *Auth) error {
 	if m.store == nil || auth == nil {
 		return nil
@@ -2186,6 +2622,9 @@ func (m *Manager) refreshAuth(ctx context.Context, id string) {
 			current.NextRefreshAfter = now.Add(refreshFailureBackoff)
 			current.LastError = &Error{Message: err.Error()}
 			m.auths[id] = current
+			if m.scheduler != nil {
+				m.scheduler.upsertAuth(current.Clone())
+			}
 		}
 		m.mu.Unlock()
 		return

sdk/cliproxy/auth/oauth_model_alias.go

@@ -80,54 +80,98 @@ func (m *Manager) applyOAuthModelAlias(auth *Auth, requestedModel string) string
 	return upstreamModel
 }
 
-func resolveModelAliasFromConfigModels(requestedModel string, models []modelAliasEntry) string {
+func modelAliasLookupCandidates(requestedModel string) (thinking.SuffixResult, []string) {
 	requestedModel = strings.TrimSpace(requestedModel)
 	if requestedModel == "" {
-		return ""
+		return thinking.SuffixResult{}, nil
 	}
-	if len(models) == 0 {
-		return ""
-	}
-
 	requestResult := thinking.ParseSuffix(requestedModel)
 	base := requestResult.ModelName
+	if base == "" {
+		base = requestedModel
+	}
 	candidates := []string{base}
 	if base != requestedModel {
 		candidates = append(candidates, requestedModel)
 	}
+	return requestResult, candidates
+}
 
-	preserveSuffix := func(resolved string) string {
-		resolved = strings.TrimSpace(resolved)
-		if resolved == "" {
-			return ""
-		}
-		if thinking.ParseSuffix(resolved).HasSuffix {
-			return resolved
-		}
-		if requestResult.HasSuffix && requestResult.RawSuffix != "" {
-			return resolved + "(" + requestResult.RawSuffix + ")"
-		}
+func preserveResolvedModelSuffix(resolved string, requestResult thinking.SuffixResult) string {
+	resolved = strings.TrimSpace(resolved)
+	if resolved == "" {
+		return ""
+	}
+	if thinking.ParseSuffix(resolved).HasSuffix {
 		return resolved
 	}
+	if requestResult.HasSuffix && requestResult.RawSuffix != "" {
+		return resolved + "(" + requestResult.RawSuffix + ")"
+	}
+	return resolved
+}
 
+func resolveModelAliasPoolFromConfigModels(requestedModel string, models []modelAliasEntry) []string {
+	requestedModel = strings.TrimSpace(requestedModel)
+	if requestedModel == "" {
+		return nil
+	}
+	if len(models) == 0 {
+		return nil
+	}
+
+	requestResult, candidates := modelAliasLookupCandidates(requestedModel)
+	if len(candidates) == 0 {
+		return nil
+	}
+
+	out := make([]string, 0)
+	seen := make(map[string]struct{})
 	for i := range models {
 		name := strings.TrimSpace(models[i].GetName())
 		alias := strings.TrimSpace(models[i].GetAlias())
 		for _, candidate := range candidates {
-			if candidate == "" {
+			if candidate == "" || alias == "" || !strings.EqualFold(alias, candidate) {
 				continue
 			}
-			if alias != "" && strings.EqualFold(alias, candidate) {
-				if name != "" {
-					return preserveSuffix(name)
-				}
-				return preserveSuffix(candidate)
+			resolved := candidate
+			if name != "" {
+				resolved = name
 			}
-			if name != "" && strings.EqualFold(name, candidate) {
-				return preserveSuffix(name)
+			resolved = preserveResolvedModelSuffix(resolved, requestResult)
+			key := strings.ToLower(strings.TrimSpace(resolved))
+			if key == "" {
+				break
 			}
+			if _, exists := seen[key]; exists {
+				break
+			}
+			seen[key] = struct{}{}
+			out = append(out, resolved)
+			break
 		}
 	}
+	if len(out) > 0 {
+		return out
+	}
+
+	for i := range models {
+		name := strings.TrimSpace(models[i].GetName())
+		for _, candidate := range candidates {
+			if candidate == "" || name == "" || !strings.EqualFold(name, candidate) {
+				continue
+			}
+			return []string{preserveResolvedModelSuffix(name, requestResult)}
+		}
+	}
+	return nil
+}
+
+func resolveModelAliasFromConfigModels(requestedModel string, models []modelAliasEntry) string {
+	resolved := resolveModelAliasPoolFromConfigModels(requestedModel, models)
+	if len(resolved) > 0 {
+		return resolved[0]
+	}
 	return ""
 }
 

sdk/cliproxy/auth/openai_compat_pool_test.go

@@ -0,0 +1,419 @@
+package auth
+
+import (
+	"context"
+	"net/http"
+	"sync"
+	"testing"
+
+	internalconfig "github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type openAICompatPoolExecutor struct {
+	id string
+
+	mu                sync.Mutex
+	executeModels     []string
+	countModels       []string
+	streamModels      []string
+	executeErrors     map[string]error
+	countErrors       map[string]error
+	streamFirstErrors map[string]error
+	streamPayloads    map[string][]cliproxyexecutor.StreamChunk
+}
+
+func (e *openAICompatPoolExecutor) Identifier() string { return e.id }
+
+func (e *openAICompatPoolExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.executeModels = append(e.executeModels, req.Model)
+	err := e.executeErrors[req.Model]
+	e.mu.Unlock()
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+	return cliproxyexecutor.Response{Payload: []byte(req.Model)}, nil
+}
+
+func (e *openAICompatPoolExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.streamModels = append(e.streamModels, req.Model)
+	err := e.streamFirstErrors[req.Model]
+	payloadChunks, hasCustomChunks := e.streamPayloads[req.Model]
+	chunks := append([]cliproxyexecutor.StreamChunk(nil), payloadChunks...)
+	e.mu.Unlock()
+	ch := make(chan cliproxyexecutor.StreamChunk, max(1, len(chunks)))
+	if err != nil {
+		ch <- cliproxyexecutor.StreamChunk{Err: err}
+		close(ch)
+		return &cliproxyexecutor.StreamResult{Headers: http.Header{"X-Model": {req.Model}}, Chunks: ch}, nil
+	}
+	if !hasCustomChunks {
+		ch <- cliproxyexecutor.StreamChunk{Payload: []byte(req.Model)}
+	} else {
+		for _, chunk := range chunks {
+			ch <- chunk
+		}
+	}
+	close(ch)
+	return &cliproxyexecutor.StreamResult{Headers: http.Header{"X-Model": {req.Model}}, Chunks: ch}, nil
+}
+
+func (e *openAICompatPoolExecutor) Refresh(_ context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e *openAICompatPoolExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = opts
+	e.mu.Lock()
+	e.countModels = append(e.countModels, req.Model)
+	err := e.countErrors[req.Model]
+	e.mu.Unlock()
+	if err != nil {
+		return cliproxyexecutor.Response{}, err
+	}
+	return cliproxyexecutor.Response{Payload: []byte(req.Model)}, nil
+}
+
+func (e *openAICompatPoolExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	_ = ctx
+	_ = auth
+	_ = req
+	return nil, &Error{HTTPStatus: http.StatusNotImplemented, Message: "HttpRequest not implemented"}
+}
+
+func (e *openAICompatPoolExecutor) ExecuteModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.executeModels))
+	copy(out, e.executeModels)
+	return out
+}
+
+func (e *openAICompatPoolExecutor) CountModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.countModels))
+	copy(out, e.countModels)
+	return out
+}
+
+func (e *openAICompatPoolExecutor) StreamModels() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.streamModels))
+	copy(out, e.streamModels)
+	return out
+}
+
+func newOpenAICompatPoolTestManager(t *testing.T, alias string, models []internalconfig.OpenAICompatibilityModel, executor *openAICompatPoolExecutor) *Manager {
+	t.Helper()
+	cfg := &internalconfig.Config{
+		OpenAICompatibility: []internalconfig.OpenAICompatibility{{
+			Name:   "pool",
+			Models: models,
+		}},
+	}
+	m := NewManager(nil, nil, nil)
+	m.SetConfig(cfg)
+	if executor == nil {
+		executor = &openAICompatPoolExecutor{id: "pool"}
+	}
+	m.RegisterExecutor(executor)
+
+	auth := &Auth{
+		ID:       "pool-auth-" + t.Name(),
+		Provider: "pool",
+		Status:   StatusActive,
+		Attributes: map[string]string{
+			"api_key":      "test-key",
+			"compat_name":  "pool",
+			"provider_key": "pool",
+		},
+	}
+	if _, err := m.Register(context.Background(), auth); err != nil {
+		t.Fatalf("register auth: %v", err)
+	}
+
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(auth.ID, "pool", []*registry.ModelInfo{{ID: alias}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(auth.ID)
+	})
+	return m
+}
+
+func TestManagerExecuteCount_OpenAICompatAliasPoolStopsOnInvalidRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusUnprocessableEntity, Message: "unprocessable entity"}
+	executor := &openAICompatPoolExecutor{
+		id:          "pool",
+		countErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.ExecuteCount(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute count error = %v, want %v", err, invalidErr)
+	}
+	got := executor.CountModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("count calls = %v, want only first invalid model", got)
+	}
+}
+func TestResolveModelAliasPoolFromConfigModels(t *testing.T) {
+	models := []modelAliasEntry{
+		internalconfig.OpenAICompatibilityModel{Name: "qwen3.5-plus", Alias: "claude-opus-4.66"},
+		internalconfig.OpenAICompatibilityModel{Name: "glm-5", Alias: "claude-opus-4.66"},
+		internalconfig.OpenAICompatibilityModel{Name: "kimi-k2.5", Alias: "claude-opus-4.66"},
+	}
+	got := resolveModelAliasPoolFromConfigModels("claude-opus-4.66(8192)", models)
+	want := []string{"qwen3.5-plus(8192)", "glm-5(8192)", "kimi-k2.5(8192)"}
+	if len(got) != len(want) {
+		t.Fatalf("pool len = %d, want %d (%v)", len(got), len(want), got)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("pool[%d] = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecute_OpenAICompatAliasPoolRotatesWithinAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{id: "pool"}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	for i := 0; i < 3; i++ {
+		resp, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+		if err != nil {
+			t.Fatalf("execute %d: %v", i, err)
+		}
+		if len(resp.Payload) == 0 {
+			t.Fatalf("execute %d returned empty payload", i)
+		}
+	}
+
+	got := executor.ExecuteModels()
+	want := []string{"qwen3.5-plus", "glm-5", "qwen3.5-plus"}
+	if len(got) != len(want) {
+		t.Fatalf("execute calls = %v, want %v", got, want)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("execute call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecute_OpenAICompatAliasPoolStopsOnBadRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusBadRequest, Message: "invalid_request_error: malformed payload"}
+	executor := &openAICompatPoolExecutor{
+		id:            "pool",
+		executeErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute error = %v, want %v", err, invalidErr)
+	}
+	got := executor.ExecuteModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("execute calls = %v, want only first invalid model", got)
+	}
+}
+func TestManagerExecute_OpenAICompatAliasPoolFallsBackWithinSameAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id:            "pool",
+		executeErrors: map[string]error{"qwen3.5-plus": &Error{HTTPStatus: http.StatusTooManyRequests, Message: "quota"}},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	resp, err := m.Execute(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute: %v", err)
+	}
+	if string(resp.Payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(resp.Payload), "glm-5")
+	}
+	got := executor.ExecuteModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("execute call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolRetriesOnEmptyBootstrap(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id: "pool",
+		streamPayloads: map[string][]cliproxyexecutor.StreamChunk{
+			"qwen3.5-plus": {},
+		},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute stream: %v", err)
+	}
+	var payload []byte
+	for chunk := range streamResult.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected stream error: %v", chunk.Err)
+		}
+		payload = append(payload, chunk.Payload...)
+	}
+	if string(payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(payload), "glm-5")
+	}
+	got := executor.StreamModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("stream call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolFallsBackBeforeFirstByte(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": &Error{HTTPStatus: http.StatusTooManyRequests, Message: "quota"}},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err != nil {
+		t.Fatalf("execute stream: %v", err)
+	}
+	var payload []byte
+	for chunk := range streamResult.Chunks {
+		if chunk.Err != nil {
+			t.Fatalf("unexpected stream error: %v", chunk.Err)
+		}
+		payload = append(payload, chunk.Payload...)
+	}
+	if string(payload) != "glm-5" {
+		t.Fatalf("payload = %q, want %q", string(payload), "glm-5")
+	}
+	got := executor.StreamModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("stream call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+	if gotHeader := streamResult.Headers.Get("X-Model"); gotHeader != "glm-5" {
+		t.Fatalf("header X-Model = %q, want %q", gotHeader, "glm-5")
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolStopsOnInvalidRequest(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusUnprocessableEntity, Message: "unprocessable entity"}
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	_, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil || err.Error() != invalidErr.Error() {
+		t.Fatalf("execute stream error = %v, want %v", err, invalidErr)
+	}
+	got := executor.StreamModels()
+	if len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("stream calls = %v, want only first invalid model", got)
+	}
+}
+func TestManagerExecuteCount_OpenAICompatAliasPoolRotatesWithinAuth(t *testing.T) {
+	alias := "claude-opus-4.66"
+	executor := &openAICompatPoolExecutor{id: "pool"}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	for i := 0; i < 2; i++ {
+		resp, err := m.ExecuteCount(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+		if err != nil {
+			t.Fatalf("execute count %d: %v", i, err)
+		}
+		if len(resp.Payload) == 0 {
+			t.Fatalf("execute count %d returned empty payload", i)
+		}
+	}
+
+	got := executor.CountModels()
+	want := []string{"qwen3.5-plus", "glm-5"}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("count call %d model = %q, want %q", i, got[i], want[i])
+		}
+	}
+}
+
+func TestManagerExecuteStream_OpenAICompatAliasPoolStopsOnInvalidBootstrap(t *testing.T) {
+	alias := "claude-opus-4.66"
+	invalidErr := &Error{HTTPStatus: http.StatusBadRequest, Message: "invalid_request_error: malformed payload"}
+	executor := &openAICompatPoolExecutor{
+		id:                "pool",
+		streamFirstErrors: map[string]error{"qwen3.5-plus": invalidErr},
+	}
+	m := newOpenAICompatPoolTestManager(t, alias, []internalconfig.OpenAICompatibilityModel{
+		{Name: "qwen3.5-plus", Alias: alias},
+		{Name: "glm-5", Alias: alias},
+	}, executor)
+
+	streamResult, err := m.ExecuteStream(context.Background(), []string{"pool"}, cliproxyexecutor.Request{Model: alias}, cliproxyexecutor.Options{})
+	if err == nil {
+		t.Fatal("expected invalid request error")
+	}
+	if err != invalidErr {
+		t.Fatalf("error = %v, want %v", err, invalidErr)
+	}
+	if streamResult != nil {
+		t.Fatalf("streamResult = %#v, want nil on invalid bootstrap", streamResult)
+	}
+	if got := executor.StreamModels(); len(got) != 1 || got[0] != "qwen3.5-plus" {
+		t.Fatalf("stream calls = %v, want only first upstream model", got)
+	}
+}

sdk/cliproxy/auth/scheduler.go

@@ -0,0 +1,851 @@
+package auth
+
+import (
+	"context"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+// schedulerStrategy identifies which built-in routing semantics the scheduler should apply.
+type schedulerStrategy int
+
+const (
+	schedulerStrategyCustom schedulerStrategy = iota
+	schedulerStrategyRoundRobin
+	schedulerStrategyFillFirst
+)
+
+// scheduledState describes how an auth currently participates in a model shard.
+type scheduledState int
+
+const (
+	scheduledStateReady scheduledState = iota
+	scheduledStateCooldown
+	scheduledStateBlocked
+	scheduledStateDisabled
+)
+
+// authScheduler keeps the incremental provider/model scheduling state used by Manager.
+type authScheduler struct {
+	mu            sync.Mutex
+	strategy      schedulerStrategy
+	providers     map[string]*providerScheduler
+	authProviders map[string]string
+	mixedCursors  map[string]int
+}
+
+// providerScheduler stores auth metadata and model shards for a single provider.
+type providerScheduler struct {
+	providerKey string
+	auths       map[string]*scheduledAuthMeta
+	modelShards map[string]*modelScheduler
+}
+
+// scheduledAuthMeta stores the immutable scheduling fields derived from an auth snapshot.
+type scheduledAuthMeta struct {
+	auth              *Auth
+	providerKey       string
+	priority          int
+	virtualParent     string
+	websocketEnabled  bool
+	supportedModelSet map[string]struct{}
+}
+
+// modelScheduler tracks ready and blocked auths for one provider/model combination.
+type modelScheduler struct {
+	modelKey        string
+	entries         map[string]*scheduledAuth
+	priorityOrder   []int
+	readyByPriority map[int]*readyBucket
+	blocked         cooldownQueue
+}
+
+// scheduledAuth stores the runtime scheduling state for a single auth inside a model shard.
+type scheduledAuth struct {
+	meta        *scheduledAuthMeta
+	auth        *Auth
+	state       scheduledState
+	nextRetryAt time.Time
+}
+
+// readyBucket keeps the ready views for one priority level.
+type readyBucket struct {
+	all readyView
+	ws  readyView
+}
+
+// readyView holds the selection order for flat or grouped round-robin traversal.
+type readyView struct {
+	flat         []*scheduledAuth
+	cursor       int
+	parentOrder  []string
+	parentCursor int
+	children     map[string]*childBucket
+}
+
+// childBucket keeps the per-parent rotation state for grouped Gemini virtual auths.
+type childBucket struct {
+	items  []*scheduledAuth
+	cursor int
+}
+
+// cooldownQueue is the blocked auth collection ordered by next retry time during rebuilds.
+type cooldownQueue []*scheduledAuth
+
+// newAuthScheduler constructs an empty scheduler configured for the supplied selector strategy.
+func newAuthScheduler(selector Selector) *authScheduler {
+	return &authScheduler{
+		strategy:      selectorStrategy(selector),
+		providers:     make(map[string]*providerScheduler),
+		authProviders: make(map[string]string),
+		mixedCursors:  make(map[string]int),
+	}
+}
+
+// selectorStrategy maps a selector implementation to the scheduler semantics it should emulate.
+func selectorStrategy(selector Selector) schedulerStrategy {
+	switch selector.(type) {
+	case *FillFirstSelector:
+		return schedulerStrategyFillFirst
+	case nil, *RoundRobinSelector:
+		return schedulerStrategyRoundRobin
+	default:
+		return schedulerStrategyCustom
+	}
+}
+
+// setSelector updates the active built-in strategy and resets mixed-provider cursors.
+func (s *authScheduler) setSelector(selector Selector) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.strategy = selectorStrategy(selector)
+	clear(s.mixedCursors)
+}
+
+// rebuild recreates the complete scheduler state from an auth snapshot.
+func (s *authScheduler) rebuild(auths []*Auth) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.providers = make(map[string]*providerScheduler)
+	s.authProviders = make(map[string]string)
+	s.mixedCursors = make(map[string]int)
+	now := time.Now()
+	for _, auth := range auths {
+		s.upsertAuthLocked(auth, now)
+	}
+}
+
+// upsertAuth incrementally synchronizes one auth into the scheduler.
+func (s *authScheduler) upsertAuth(auth *Auth) {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.upsertAuthLocked(auth, time.Now())
+}
+
+// removeAuth deletes one auth from every scheduler shard that references it.
+func (s *authScheduler) removeAuth(authID string) {
+	if s == nil {
+		return
+	}
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.removeAuthLocked(authID)
+}
+
+// pickSingle returns the next auth for a single provider/model request using scheduler state.
+func (s *authScheduler) pickSingle(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, error) {
+	if s == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	providerKey := strings.ToLower(strings.TrimSpace(provider))
+	modelKey := canonicalModelKey(model)
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+	preferWebsocket := cliproxyexecutor.DownstreamWebsocket(ctx) && providerKey == "codex" && pinnedAuthID == ""
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	providerState := s.providers[providerKey]
+	if providerState == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	shard := providerState.ensureModelLocked(modelKey, time.Now())
+	if shard == nil {
+		return nil, &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	predicate := func(entry *scheduledAuth) bool {
+		if entry == nil || entry.auth == nil {
+			return false
+		}
+		if pinnedAuthID != "" && entry.auth.ID != pinnedAuthID {
+			return false
+		}
+		if len(tried) > 0 {
+			if _, ok := tried[entry.auth.ID]; ok {
+				return false
+			}
+		}
+		return true
+	}
+	if picked := shard.pickReadyLocked(preferWebsocket, s.strategy, predicate); picked != nil {
+		return picked, nil
+	}
+	return nil, shard.unavailableErrorLocked(provider, model, predicate)
+}
+
+// pickMixed returns the next auth and provider for a mixed-provider request.
+func (s *authScheduler) pickMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, string, error) {
+	if s == nil {
+		return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	normalized := normalizeProviderKeys(providers)
+	if len(normalized) == 0 {
+		return nil, "", &Error{Code: "provider_not_found", Message: "no provider supplied"}
+	}
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+	modelKey := canonicalModelKey(model)
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if pinnedAuthID != "" {
+		providerKey := s.authProviders[pinnedAuthID]
+		if providerKey == "" || !containsProvider(normalized, providerKey) {
+			return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+		}
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			return nil, "", &Error{Code: "auth_not_found", Message: "no auth available"}
+		}
+		shard := providerState.ensureModelLocked(modelKey, time.Now())
+		predicate := func(entry *scheduledAuth) bool {
+			if entry == nil || entry.auth == nil || entry.auth.ID != pinnedAuthID {
+				return false
+			}
+			if len(tried) == 0 {
+				return true
+			}
+			_, ok := tried[pinnedAuthID]
+			return !ok
+		}
+		if picked := shard.pickReadyLocked(false, s.strategy, predicate); picked != nil {
+			return picked, providerKey, nil
+		}
+		return nil, "", shard.unavailableErrorLocked("mixed", model, predicate)
+	}
+
+	if s.strategy == schedulerStrategyFillFirst {
+		for _, providerKey := range normalized {
+			providerState := s.providers[providerKey]
+			if providerState == nil {
+				continue
+			}
+			shard := providerState.ensureModelLocked(modelKey, time.Now())
+			if shard == nil {
+				continue
+			}
+			picked := shard.pickReadyLocked(false, s.strategy, triedPredicate(tried))
+			if picked != nil {
+				return picked, providerKey, nil
+			}
+		}
+		return nil, "", s.mixedUnavailableErrorLocked(normalized, model, tried)
+	}
+
+	cursorKey := strings.Join(normalized, ",") + ":" + modelKey
+	start := 0
+	if len(normalized) > 0 {
+		start = s.mixedCursors[cursorKey] % len(normalized)
+	}
+	for offset := 0; offset < len(normalized); offset++ {
+		providerIndex := (start + offset) % len(normalized)
+		providerKey := normalized[providerIndex]
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			continue
+		}
+		shard := providerState.ensureModelLocked(modelKey, time.Now())
+		if shard == nil {
+			continue
+		}
+		picked := shard.pickReadyLocked(false, schedulerStrategyRoundRobin, triedPredicate(tried))
+		if picked == nil {
+			continue
+		}
+		s.mixedCursors[cursorKey] = providerIndex + 1
+		return picked, providerKey, nil
+	}
+	return nil, "", s.mixedUnavailableErrorLocked(normalized, model, tried)
+}
+
+// mixedUnavailableErrorLocked synthesizes the mixed-provider cooldown or unavailable error.
+func (s *authScheduler) mixedUnavailableErrorLocked(providers []string, model string, tried map[string]struct{}) error {
+	now := time.Now()
+	total := 0
+	cooldownCount := 0
+	earliest := time.Time{}
+	for _, providerKey := range providers {
+		providerState := s.providers[providerKey]
+		if providerState == nil {
+			continue
+		}
+		shard := providerState.ensureModelLocked(canonicalModelKey(model), now)
+		if shard == nil {
+			continue
+		}
+		localTotal, localCooldownCount, localEarliest := shard.availabilitySummaryLocked(triedPredicate(tried))
+		total += localTotal
+		cooldownCount += localCooldownCount
+		if !localEarliest.IsZero() && (earliest.IsZero() || localEarliest.Before(earliest)) {
+			earliest = localEarliest
+		}
+	}
+	if total == 0 {
+		return &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	if cooldownCount == total && !earliest.IsZero() {
+		resetIn := earliest.Sub(now)
+		if resetIn < 0 {
+			resetIn = 0
+		}
+		return newModelCooldownError(model, "", resetIn)
+	}
+	return &Error{Code: "auth_unavailable", Message: "no auth available"}
+}
+
+// triedPredicate builds a filter that excludes auths already attempted for the current request.
+func triedPredicate(tried map[string]struct{}) func(*scheduledAuth) bool {
+	if len(tried) == 0 {
+		return func(entry *scheduledAuth) bool { return entry != nil && entry.auth != nil }
+	}
+	return func(entry *scheduledAuth) bool {
+		if entry == nil || entry.auth == nil {
+			return false
+		}
+		_, ok := tried[entry.auth.ID]
+		return !ok
+	}
+}
+
+// normalizeProviderKeys lowercases, trims, and de-duplicates provider keys while preserving order.
+func normalizeProviderKeys(providers []string) []string {
+	seen := make(map[string]struct{}, len(providers))
+	out := make([]string, 0, len(providers))
+	for _, provider := range providers {
+		providerKey := strings.ToLower(strings.TrimSpace(provider))
+		if providerKey == "" {
+			continue
+		}
+		if _, ok := seen[providerKey]; ok {
+			continue
+		}
+		seen[providerKey] = struct{}{}
+		out = append(out, providerKey)
+	}
+	return out
+}
+
+// containsProvider reports whether provider is present in the normalized provider list.
+func containsProvider(providers []string, provider string) bool {
+	for _, candidate := range providers {
+		if candidate == provider {
+			return true
+		}
+	}
+	return false
+}
+
+// upsertAuthLocked updates one auth in-place while the scheduler mutex is held.
+func (s *authScheduler) upsertAuthLocked(auth *Auth, now time.Time) {
+	if auth == nil {
+		return
+	}
+	authID := strings.TrimSpace(auth.ID)
+	providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+	if authID == "" || providerKey == "" || auth.Disabled {
+		s.removeAuthLocked(authID)
+		return
+	}
+	if previousProvider := s.authProviders[authID]; previousProvider != "" && previousProvider != providerKey {
+		if previousState := s.providers[previousProvider]; previousState != nil {
+			previousState.removeAuthLocked(authID)
+		}
+	}
+	meta := buildScheduledAuthMeta(auth)
+	s.authProviders[authID] = providerKey
+	s.ensureProviderLocked(providerKey).upsertAuthLocked(meta, now)
+}
+
+// removeAuthLocked removes one auth from the scheduler while the scheduler mutex is held.
+func (s *authScheduler) removeAuthLocked(authID string) {
+	if authID == "" {
+		return
+	}
+	if providerKey := s.authProviders[authID]; providerKey != "" {
+		if providerState := s.providers[providerKey]; providerState != nil {
+			providerState.removeAuthLocked(authID)
+		}
+		delete(s.authProviders, authID)
+	}
+}
+
+// ensureProviderLocked returns the provider scheduler for providerKey, creating it when needed.
+func (s *authScheduler) ensureProviderLocked(providerKey string) *providerScheduler {
+	if s.providers == nil {
+		s.providers = make(map[string]*providerScheduler)
+	}
+	providerState := s.providers[providerKey]
+	if providerState == nil {
+		providerState = &providerScheduler{
+			providerKey: providerKey,
+			auths:       make(map[string]*scheduledAuthMeta),
+			modelShards: make(map[string]*modelScheduler),
+		}
+		s.providers[providerKey] = providerState
+	}
+	return providerState
+}
+
+// buildScheduledAuthMeta extracts the scheduling metadata needed for shard bookkeeping.
+func buildScheduledAuthMeta(auth *Auth) *scheduledAuthMeta {
+	providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+	virtualParent := ""
+	if auth.Attributes != nil {
+		virtualParent = strings.TrimSpace(auth.Attributes["gemini_virtual_parent"])
+	}
+	return &scheduledAuthMeta{
+		auth:              auth,
+		providerKey:       providerKey,
+		priority:          authPriority(auth),
+		virtualParent:     virtualParent,
+		websocketEnabled:  authWebsocketsEnabled(auth),
+		supportedModelSet: supportedModelSetForAuth(auth.ID),
+	}
+}
+
+// supportedModelSetForAuth snapshots the registry models currently registered for an auth.
+func supportedModelSetForAuth(authID string) map[string]struct{} {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return nil
+	}
+	models := registry.GetGlobalRegistry().GetModelsForClient(authID)
+	if len(models) == 0 {
+		return nil
+	}
+	set := make(map[string]struct{}, len(models))
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		modelKey := canonicalModelKey(model.ID)
+		if modelKey == "" {
+			continue
+		}
+		set[modelKey] = struct{}{}
+	}
+	return set
+}
+
+// upsertAuthLocked updates every existing model shard that can reference the auth metadata.
+func (p *providerScheduler) upsertAuthLocked(meta *scheduledAuthMeta, now time.Time) {
+	if p == nil || meta == nil || meta.auth == nil {
+		return
+	}
+	p.auths[meta.auth.ID] = meta
+	for modelKey, shard := range p.modelShards {
+		if shard == nil {
+			continue
+		}
+		if !meta.supportsModel(modelKey) {
+			shard.removeEntryLocked(meta.auth.ID)
+			continue
+		}
+		shard.upsertEntryLocked(meta, now)
+	}
+}
+
+// removeAuthLocked removes an auth from all model shards owned by the provider scheduler.
+func (p *providerScheduler) removeAuthLocked(authID string) {
+	if p == nil || authID == "" {
+		return
+	}
+	delete(p.auths, authID)
+	for _, shard := range p.modelShards {
+		if shard != nil {
+			shard.removeEntryLocked(authID)
+		}
+	}
+}
+
+// ensureModelLocked returns the shard for modelKey, building it lazily from provider auths.
+func (p *providerScheduler) ensureModelLocked(modelKey string, now time.Time) *modelScheduler {
+	if p == nil {
+		return nil
+	}
+	modelKey = canonicalModelKey(modelKey)
+	if shard, ok := p.modelShards[modelKey]; ok && shard != nil {
+		shard.promoteExpiredLocked(now)
+		return shard
+	}
+	shard := &modelScheduler{
+		modelKey:        modelKey,
+		entries:         make(map[string]*scheduledAuth),
+		readyByPriority: make(map[int]*readyBucket),
+	}
+	for _, meta := range p.auths {
+		if meta == nil || !meta.supportsModel(modelKey) {
+			continue
+		}
+		shard.upsertEntryLocked(meta, now)
+	}
+	p.modelShards[modelKey] = shard
+	return shard
+}
+
+// supportsModel reports whether the auth metadata currently supports modelKey.
+func (m *scheduledAuthMeta) supportsModel(modelKey string) bool {
+	modelKey = canonicalModelKey(modelKey)
+	if modelKey == "" {
+		return true
+	}
+	if len(m.supportedModelSet) == 0 {
+		return false
+	}
+	_, ok := m.supportedModelSet[modelKey]
+	return ok
+}
+
+// upsertEntryLocked updates or inserts one auth entry and rebuilds indexes when ordering changes.
+func (m *modelScheduler) upsertEntryLocked(meta *scheduledAuthMeta, now time.Time) {
+	if m == nil || meta == nil || meta.auth == nil {
+		return
+	}
+	entry, ok := m.entries[meta.auth.ID]
+	if !ok || entry == nil {
+		entry = &scheduledAuth{}
+		m.entries[meta.auth.ID] = entry
+	}
+	previousState := entry.state
+	previousNextRetryAt := entry.nextRetryAt
+	previousPriority := 0
+	previousParent := ""
+	previousWebsocketEnabled := false
+	if entry.meta != nil {
+		previousPriority = entry.meta.priority
+		previousParent = entry.meta.virtualParent
+		previousWebsocketEnabled = entry.meta.websocketEnabled
+	}
+
+	entry.meta = meta
+	entry.auth = meta.auth
+	entry.nextRetryAt = time.Time{}
+	blocked, reason, next := isAuthBlockedForModel(meta.auth, m.modelKey, now)
+	switch {
+	case !blocked:
+		entry.state = scheduledStateReady
+	case reason == blockReasonCooldown:
+		entry.state = scheduledStateCooldown
+		entry.nextRetryAt = next
+	case reason == blockReasonDisabled:
+		entry.state = scheduledStateDisabled
+	default:
+		entry.state = scheduledStateBlocked
+		entry.nextRetryAt = next
+	}
+
+	if ok && previousState == entry.state && previousNextRetryAt.Equal(entry.nextRetryAt) && previousPriority == meta.priority && previousParent == meta.virtualParent && previousWebsocketEnabled == meta.websocketEnabled {
+		return
+	}
+	m.rebuildIndexesLocked()
+}
+
+// removeEntryLocked deletes one auth entry and rebuilds the shard indexes if needed.
+func (m *modelScheduler) removeEntryLocked(authID string) {
+	if m == nil || authID == "" {
+		return
+	}
+	if _, ok := m.entries[authID]; !ok {
+		return
+	}
+	delete(m.entries, authID)
+	m.rebuildIndexesLocked()
+}
+
+// promoteExpiredLocked reevaluates blocked auths whose retry time has elapsed.
+func (m *modelScheduler) promoteExpiredLocked(now time.Time) {
+	if m == nil || len(m.blocked) == 0 {
+		return
+	}
+	changed := false
+	for _, entry := range m.blocked {
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		if entry.nextRetryAt.IsZero() || entry.nextRetryAt.After(now) {
+			continue
+		}
+		blocked, reason, next := isAuthBlockedForModel(entry.auth, m.modelKey, now)
+		switch {
+		case !blocked:
+			entry.state = scheduledStateReady
+			entry.nextRetryAt = time.Time{}
+		case reason == blockReasonCooldown:
+			entry.state = scheduledStateCooldown
+			entry.nextRetryAt = next
+		case reason == blockReasonDisabled:
+			entry.state = scheduledStateDisabled
+			entry.nextRetryAt = time.Time{}
+		default:
+			entry.state = scheduledStateBlocked
+			entry.nextRetryAt = next
+		}
+		changed = true
+	}
+	if changed {
+		m.rebuildIndexesLocked()
+	}
+}
+
+// pickReadyLocked selects the next ready auth from the highest available priority bucket.
+func (m *modelScheduler) pickReadyLocked(preferWebsocket bool, strategy schedulerStrategy, predicate func(*scheduledAuth) bool) *Auth {
+	if m == nil {
+		return nil
+	}
+	m.promoteExpiredLocked(time.Now())
+	for _, priority := range m.priorityOrder {
+		bucket := m.readyByPriority[priority]
+		if bucket == nil {
+			continue
+		}
+		view := &bucket.all
+		if preferWebsocket && len(bucket.ws.flat) > 0 {
+			view = &bucket.ws
+		}
+		var picked *scheduledAuth
+		if strategy == schedulerStrategyFillFirst {
+			picked = view.pickFirst(predicate)
+		} else {
+			picked = view.pickRoundRobin(predicate)
+		}
+		if picked != nil && picked.auth != nil {
+			return picked.auth
+		}
+	}
+	return nil
+}
+
+// unavailableErrorLocked returns the correct unavailable or cooldown error for the shard.
+func (m *modelScheduler) unavailableErrorLocked(provider, model string, predicate func(*scheduledAuth) bool) error {
+	now := time.Now()
+	total, cooldownCount, earliest := m.availabilitySummaryLocked(predicate)
+	if total == 0 {
+		return &Error{Code: "auth_not_found", Message: "no auth available"}
+	}
+	if cooldownCount == total && !earliest.IsZero() {
+		providerForError := provider
+		if providerForError == "mixed" {
+			providerForError = ""
+		}
+		resetIn := earliest.Sub(now)
+		if resetIn < 0 {
+			resetIn = 0
+		}
+		return newModelCooldownError(model, providerForError, resetIn)
+	}
+	return &Error{Code: "auth_unavailable", Message: "no auth available"}
+}
+
+// availabilitySummaryLocked summarizes total candidates, cooldown count, and earliest retry time.
+func (m *modelScheduler) availabilitySummaryLocked(predicate func(*scheduledAuth) bool) (int, int, time.Time) {
+	if m == nil {
+		return 0, 0, time.Time{}
+	}
+	total := 0
+	cooldownCount := 0
+	earliest := time.Time{}
+	for _, entry := range m.entries {
+		if predicate != nil && !predicate(entry) {
+			continue
+		}
+		total++
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		if entry.state != scheduledStateCooldown {
+			continue
+		}
+		cooldownCount++
+		if !entry.nextRetryAt.IsZero() && (earliest.IsZero() || entry.nextRetryAt.Before(earliest)) {
+			earliest = entry.nextRetryAt
+		}
+	}
+	return total, cooldownCount, earliest
+}
+
+// rebuildIndexesLocked reconstructs ready and blocked views from the current entry map.
+func (m *modelScheduler) rebuildIndexesLocked() {
+	m.readyByPriority = make(map[int]*readyBucket)
+	m.priorityOrder = m.priorityOrder[:0]
+	m.blocked = m.blocked[:0]
+	priorityBuckets := make(map[int][]*scheduledAuth)
+	for _, entry := range m.entries {
+		if entry == nil || entry.auth == nil {
+			continue
+		}
+		switch entry.state {
+		case scheduledStateReady:
+			priority := entry.meta.priority
+			priorityBuckets[priority] = append(priorityBuckets[priority], entry)
+		case scheduledStateCooldown, scheduledStateBlocked:
+			m.blocked = append(m.blocked, entry)
+		}
+	}
+	for priority, entries := range priorityBuckets {
+		sort.Slice(entries, func(i, j int) bool {
+			return entries[i].auth.ID < entries[j].auth.ID
+		})
+		m.readyByPriority[priority] = buildReadyBucket(entries)
+		m.priorityOrder = append(m.priorityOrder, priority)
+	}
+	sort.Slice(m.priorityOrder, func(i, j int) bool {
+		return m.priorityOrder[i] > m.priorityOrder[j]
+	})
+	sort.Slice(m.blocked, func(i, j int) bool {
+		left := m.blocked[i]
+		right := m.blocked[j]
+		if left == nil || right == nil {
+			return left != nil
+		}
+		if left.nextRetryAt.Equal(right.nextRetryAt) {
+			return left.auth.ID < right.auth.ID
+		}
+		if left.nextRetryAt.IsZero() {
+			return false
+		}
+		if right.nextRetryAt.IsZero() {
+			return true
+		}
+		return left.nextRetryAt.Before(right.nextRetryAt)
+	})
+}
+
+// buildReadyBucket prepares the general and websocket-only ready views for one priority bucket.
+func buildReadyBucket(entries []*scheduledAuth) *readyBucket {
+	bucket := &readyBucket{}
+	bucket.all = buildReadyView(entries)
+	wsEntries := make([]*scheduledAuth, 0, len(entries))
+	for _, entry := range entries {
+		if entry != nil && entry.meta != nil && entry.meta.websocketEnabled {
+			wsEntries = append(wsEntries, entry)
+		}
+	}
+	bucket.ws = buildReadyView(wsEntries)
+	return bucket
+}
+
+// buildReadyView creates either a flat view or a grouped parent/child view for rotation.
+func buildReadyView(entries []*scheduledAuth) readyView {
+	view := readyView{flat: append([]*scheduledAuth(nil), entries...)}
+	if len(entries) == 0 {
+		return view
+	}
+	groups := make(map[string][]*scheduledAuth)
+	for _, entry := range entries {
+		if entry == nil || entry.meta == nil || entry.meta.virtualParent == "" {
+			return view
+		}
+		groups[entry.meta.virtualParent] = append(groups[entry.meta.virtualParent], entry)
+	}
+	if len(groups) <= 1 {
+		return view
+	}
+	view.children = make(map[string]*childBucket, len(groups))
+	view.parentOrder = make([]string, 0, len(groups))
+	for parent := range groups {
+		view.parentOrder = append(view.parentOrder, parent)
+	}
+	sort.Strings(view.parentOrder)
+	for _, parent := range view.parentOrder {
+		view.children[parent] = &childBucket{items: append([]*scheduledAuth(nil), groups[parent]...)}
+	}
+	return view
+}
+
+// pickFirst returns the first ready entry that satisfies predicate without advancing cursors.
+func (v *readyView) pickFirst(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	for _, entry := range v.flat {
+		if predicate == nil || predicate(entry) {
+			return entry
+		}
+	}
+	return nil
+}
+
+// pickRoundRobin returns the next ready entry using flat or grouped round-robin traversal.
+func (v *readyView) pickRoundRobin(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	if len(v.parentOrder) > 1 && len(v.children) > 0 {
+		return v.pickGroupedRoundRobin(predicate)
+	}
+	if len(v.flat) == 0 {
+		return nil
+	}
+	start := 0
+	if len(v.flat) > 0 {
+		start = v.cursor % len(v.flat)
+	}
+	for offset := 0; offset < len(v.flat); offset++ {
+		index := (start + offset) % len(v.flat)
+		entry := v.flat[index]
+		if predicate != nil && !predicate(entry) {
+			continue
+		}
+		v.cursor = index + 1
+		return entry
+	}
+	return nil
+}
+
+// pickGroupedRoundRobin rotates across parents first and then within the selected parent.
+func (v *readyView) pickGroupedRoundRobin(predicate func(*scheduledAuth) bool) *scheduledAuth {
+	start := 0
+	if len(v.parentOrder) > 0 {
+		start = v.parentCursor % len(v.parentOrder)
+	}
+	for offset := 0; offset < len(v.parentOrder); offset++ {
+		parentIndex := (start + offset) % len(v.parentOrder)
+		parent := v.parentOrder[parentIndex]
+		child := v.children[parent]
+		if child == nil || len(child.items) == 0 {
+			continue
+		}
+		itemStart := child.cursor % len(child.items)
+		for itemOffset := 0; itemOffset < len(child.items); itemOffset++ {
+			itemIndex := (itemStart + itemOffset) % len(child.items)
+			entry := child.items[itemIndex]
+			if predicate != nil && !predicate(entry) {
+				continue
+			}
+			child.cursor = itemIndex + 1
+			v.parentCursor = parentIndex + 1
+			return entry
+		}
+	}
+	return nil
+}

sdk/cliproxy/auth/scheduler_benchmark_test.go

@@ -0,0 +1,197 @@
+package auth
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type schedulerBenchmarkExecutor struct {
+	id string
+}
+
+func (e schedulerBenchmarkExecutor) Identifier() string { return e.id }
+
+func (e schedulerBenchmarkExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerBenchmarkExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	return nil, nil
+}
+
+func (e schedulerBenchmarkExecutor) Refresh(ctx context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e schedulerBenchmarkExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e schedulerBenchmarkExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+func benchmarkManagerSetup(b *testing.B, total int, mixed bool, withPriority bool) (*Manager, []string, string) {
+	b.Helper()
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	providers := []string{"gemini"}
+	manager.executors["gemini"] = schedulerBenchmarkExecutor{id: "gemini"}
+	if mixed {
+		providers = []string{"gemini", "claude"}
+		manager.executors["claude"] = schedulerBenchmarkExecutor{id: "claude"}
+	}
+
+	reg := registry.GetGlobalRegistry()
+	model := "bench-model"
+	for index := 0; index < total; index++ {
+		provider := providers[0]
+		if mixed && index%2 == 1 {
+			provider = providers[1]
+		}
+		auth := &Auth{ID: fmt.Sprintf("bench-%s-%04d", provider, index), Provider: provider}
+		if withPriority {
+			priority := "0"
+			if index%2 == 0 {
+				priority = "10"
+			}
+			auth.Attributes = map[string]string{"priority": priority}
+		}
+		_, errRegister := manager.Register(context.Background(), auth)
+		if errRegister != nil {
+			b.Fatalf("Register(%s) error = %v", auth.ID, errRegister)
+		}
+		reg.RegisterClient(auth.ID, provider, []*registry.ModelInfo{{ID: model}})
+	}
+	manager.syncScheduler()
+	b.Cleanup(func() {
+		for index := 0; index < total; index++ {
+			provider := providers[0]
+			if mixed && index%2 == 1 {
+				provider = providers[1]
+			}
+			reg.UnregisterClient(fmt.Sprintf("bench-%s-%04d", provider, index))
+		}
+	})
+
+	return manager, providers, model
+}
+
+func BenchmarkManagerPickNext500(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 500, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNext1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextPriority500(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 500, false, true)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextPriority1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, true)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil {
+			b.Fatalf("pickNext failed: auth=%v exec=%v err=%v", auth, exec, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextMixed500(b *testing.B) {
+	manager, providers, model := benchmarkManagerSetup(b, 500, true, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, _, errWarm := manager.pickNextMixed(ctx, providers, model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNextMixed error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, exec, provider, errPick := manager.pickNextMixed(ctx, providers, model, opts, tried)
+		if errPick != nil || auth == nil || exec == nil || provider == "" {
+			b.Fatalf("pickNextMixed failed: auth=%v exec=%v provider=%q err=%v", auth, exec, provider, errPick)
+		}
+	}
+}
+
+func BenchmarkManagerPickNextAndMarkResult1000(b *testing.B) {
+	manager, _, model := benchmarkManagerSetup(b, 1000, false, false)
+	ctx := context.Background()
+	opts := cliproxyexecutor.Options{}
+	tried := map[string]struct{}{}
+	if _, _, errWarm := manager.pickNext(ctx, "gemini", model, opts, tried); errWarm != nil {
+		b.Fatalf("warmup pickNext error = %v", errWarm)
+	}
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		auth, _, errPick := manager.pickNext(ctx, "gemini", model, opts, tried)
+		if errPick != nil || auth == nil {
+			b.Fatalf("pickNext failed: auth=%v err=%v", auth, errPick)
+		}
+		manager.MarkResult(ctx, Result{AuthID: auth.ID, Provider: "gemini", Model: model, Success: true})
+	}
+}

sdk/cliproxy/auth/scheduler_test.go

@@ -0,0 +1,468 @@
+package auth
+
+import (
+	"context"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type schedulerTestExecutor struct{}
+
+func (schedulerTestExecutor) Identifier() string { return "test" }
+
+func (schedulerTestExecutor) Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (schedulerTestExecutor) ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
+	return nil, nil
+}
+
+func (schedulerTestExecutor) Refresh(ctx context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (schedulerTestExecutor) CountTokens(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (schedulerTestExecutor) HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+type trackingSelector struct {
+	calls      int
+	lastAuthID []string
+}
+
+func (s *trackingSelector) Pick(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, auths []*Auth) (*Auth, error) {
+	s.calls++
+	s.lastAuthID = s.lastAuthID[:0]
+	for _, auth := range auths {
+		s.lastAuthID = append(s.lastAuthID, auth.ID)
+	}
+	if len(auths) == 0 {
+		return nil, nil
+	}
+	return auths[len(auths)-1], nil
+}
+
+func newSchedulerForTest(selector Selector, auths ...*Auth) *authScheduler {
+	scheduler := newAuthScheduler(selector)
+	scheduler.rebuild(auths)
+	return scheduler
+}
+
+func registerSchedulerModels(t *testing.T, provider string, model string, authIDs ...string) {
+	t.Helper()
+	reg := registry.GetGlobalRegistry()
+	for _, authID := range authIDs {
+		reg.RegisterClient(authID, provider, []*registry.ModelInfo{{ID: model}})
+	}
+	t.Cleanup(func() {
+		for _, authID := range authIDs {
+			reg.UnregisterClient(authID)
+		}
+	})
+}
+
+func TestSchedulerPick_RoundRobinHighestPriority(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "low", Provider: "gemini", Attributes: map[string]string{"priority": "0"}},
+		&Auth{ID: "high-b", Provider: "gemini", Attributes: map[string]string{"priority": "10"}},
+		&Auth{ID: "high-a", Provider: "gemini", Attributes: map[string]string{"priority": "10"}},
+	)
+
+	want := []string{"high-a", "high-b", "high-a"}
+	for index, wantID := range want {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantID {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantID)
+		}
+	}
+}
+
+func TestSchedulerPick_FillFirstSticksToFirstReady(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&FillFirstSelector{},
+		&Auth{ID: "b", Provider: "gemini"},
+		&Auth{ID: "a", Provider: "gemini"},
+		&Auth{ID: "c", Provider: "gemini"},
+	)
+
+	for index := 0; index < 3; index++ {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != "a" {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, "a")
+		}
+	}
+}
+
+func TestSchedulerPick_PromotesExpiredCooldownBeforePick(t *testing.T) {
+	t.Parallel()
+
+	model := "gemini-2.5-pro"
+	registerSchedulerModels(t, "gemini", model, "cooldown-expired")
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{
+			ID:       "cooldown-expired",
+			Provider: "gemini",
+			ModelStates: map[string]*ModelState{
+				model: {
+					Status:         StatusError,
+					Unavailable:    true,
+					NextRetryAfter: time.Now().Add(-1 * time.Second),
+				},
+			},
+		},
+	)
+
+	got, errPick := scheduler.pickSingle(context.Background(), "gemini", model, cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("pickSingle() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickSingle() auth = nil")
+	}
+	if got.ID != "cooldown-expired" {
+		t.Fatalf("pickSingle() auth.ID = %q, want %q", got.ID, "cooldown-expired")
+	}
+}
+
+func TestSchedulerPick_GeminiVirtualParentUsesTwoLevelRotation(t *testing.T) {
+	t.Parallel()
+
+	registerSchedulerModels(t, "gemini-cli", "gemini-2.5-pro", "cred-a::proj-1", "cred-a::proj-2", "cred-b::proj-1", "cred-b::proj-2")
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "cred-a::proj-1", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-a"}},
+		&Auth{ID: "cred-a::proj-2", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-a"}},
+		&Auth{ID: "cred-b::proj-1", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-b"}},
+		&Auth{ID: "cred-b::proj-2", Provider: "gemini-cli", Attributes: map[string]string{"gemini_virtual_parent": "cred-b"}},
+	)
+
+	wantParents := []string{"cred-a", "cred-b", "cred-a", "cred-b"}
+	wantIDs := []string{"cred-a::proj-1", "cred-b::proj-1", "cred-a::proj-2", "cred-b::proj-2"}
+	for index := range wantIDs {
+		got, errPick := scheduler.pickSingle(context.Background(), "gemini-cli", "gemini-2.5-pro", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+		if got.Attributes["gemini_virtual_parent"] != wantParents[index] {
+			t.Fatalf("pickSingle() #%d parent = %q, want %q", index, got.Attributes["gemini_virtual_parent"], wantParents[index])
+		}
+	}
+}
+
+func TestSchedulerPick_CodexWebsocketPrefersWebsocketEnabledSubset(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "codex-http", Provider: "codex"},
+		&Auth{ID: "codex-ws-a", Provider: "codex", Attributes: map[string]string{"websockets": "true"}},
+		&Auth{ID: "codex-ws-b", Provider: "codex", Attributes: map[string]string{"websockets": "true"}},
+	)
+
+	ctx := cliproxyexecutor.WithDownstreamWebsocket(context.Background())
+	want := []string{"codex-ws-a", "codex-ws-b", "codex-ws-a"}
+	for index, wantID := range want {
+		got, errPick := scheduler.pickSingle(ctx, "codex", "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickSingle() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickSingle() #%d auth = nil", index)
+		}
+		if got.ID != wantID {
+			t.Fatalf("pickSingle() #%d auth.ID = %q, want %q", index, got.ID, wantID)
+		}
+	}
+}
+
+func TestSchedulerPick_MixedProvidersUsesProviderRotationOverReadyCandidates(t *testing.T) {
+	t.Parallel()
+
+	scheduler := newSchedulerForTest(
+		&RoundRobinSelector{},
+		&Auth{ID: "gemini-a", Provider: "gemini"},
+		&Auth{ID: "gemini-b", Provider: "gemini"},
+		&Auth{ID: "claude-a", Provider: "claude"},
+	)
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, provider, errPick := scheduler.pickMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManager_PickNextMixed_UsesProviderRotationBeforeCredentialRotation(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, map[string]struct{}{})
+		if errPick != nil {
+			t.Fatalf("pickNextMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickNextMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickNextMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickNextMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManagerCustomSelector_FallsBackToLegacyPath(t *testing.T) {
+	t.Parallel()
+
+	selector := &trackingSelector{}
+	manager := NewManager(nil, selector, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.auths["auth-a"] = &Auth{ID: "auth-a", Provider: "gemini"}
+	manager.auths["auth-b"] = &Auth{ID: "auth-b", Provider: "gemini"}
+
+	got, _, errPick := manager.pickNext(context.Background(), "gemini", "", cliproxyexecutor.Options{}, map[string]struct{}{})
+	if errPick != nil {
+		t.Fatalf("pickNext() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickNext() auth = nil")
+	}
+	if selector.calls != 1 {
+		t.Fatalf("selector.calls = %d, want %d", selector.calls, 1)
+	}
+	if len(selector.lastAuthID) != 2 {
+		t.Fatalf("len(selector.lastAuthID) = %d, want %d", len(selector.lastAuthID), 2)
+	}
+	if got.ID != selector.lastAuthID[len(selector.lastAuthID)-1] {
+		t.Fatalf("pickNext() auth.ID = %q, want selector-picked %q", got.ID, selector.lastAuthID[len(selector.lastAuthID)-1])
+	}
+}
+
+func TestManager_InitializesSchedulerForBuiltInSelector(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	if manager.scheduler == nil {
+		t.Fatalf("manager.scheduler = nil")
+	}
+	if manager.scheduler.strategy != schedulerStrategyRoundRobin {
+		t.Fatalf("manager.scheduler.strategy = %v, want %v", manager.scheduler.strategy, schedulerStrategyRoundRobin)
+	}
+
+	manager.SetSelector(&FillFirstSelector{})
+	if manager.scheduler.strategy != schedulerStrategyFillFirst {
+		t.Fatalf("manager.scheduler.strategy = %v, want %v", manager.scheduler.strategy, schedulerStrategyFillFirst)
+	}
+}
+
+func TestManager_SchedulerTracksRegisterAndUpdate(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-a) error = %v", errRegister)
+	}
+
+	got, errPick := manager.scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-a" {
+		t.Fatalf("scheduler.pickSingle() auth = %v, want auth-a", got)
+	}
+
+	if _, errUpdate := manager.Update(context.Background(), &Auth{ID: "auth-a", Provider: "gemini", Disabled: true}); errUpdate != nil {
+		t.Fatalf("Update(auth-a) error = %v", errUpdate)
+	}
+
+	got, errPick = manager.scheduler.pickSingle(context.Background(), "gemini", "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() after update error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-b" {
+		t.Fatalf("scheduler.pickSingle() after update auth = %v, want auth-b", got)
+	}
+}
+
+func TestManager_PickNextMixed_UsesSchedulerRotation(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["gemini"] = schedulerTestExecutor{}
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-b) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	wantProviders := []string{"gemini", "claude", "gemini", "claude"}
+	wantIDs := []string{"gemini-a", "claude-a", "gemini-b", "claude-a"}
+	for index := range wantProviders {
+		got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("pickNextMixed() #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("pickNextMixed() #%d auth = nil", index)
+		}
+		if provider != wantProviders[index] {
+			t.Fatalf("pickNextMixed() #%d provider = %q, want %q", index, provider, wantProviders[index])
+		}
+		if got.ID != wantIDs[index] {
+			t.Fatalf("pickNextMixed() #%d auth.ID = %q, want %q", index, got.ID, wantIDs[index])
+		}
+	}
+}
+
+func TestManager_PickNextMixed_SkipsProvidersWithoutExecutors(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	manager.executors["claude"] = schedulerTestExecutor{}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "gemini-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(gemini-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "claude-a", Provider: "claude"}); errRegister != nil {
+		t.Fatalf("Register(claude-a) error = %v", errRegister)
+	}
+
+	got, _, provider, errPick := manager.pickNextMixed(context.Background(), []string{"gemini", "claude"}, "", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("pickNextMixed() error = %v", errPick)
+	}
+	if got == nil {
+		t.Fatalf("pickNextMixed() auth = nil")
+	}
+	if provider != "claude" {
+		t.Fatalf("pickNextMixed() provider = %q, want %q", provider, "claude")
+	}
+	if got.ID != "claude-a" {
+		t.Fatalf("pickNextMixed() auth.ID = %q, want %q", got.ID, "claude-a")
+	}
+}
+
+func TestManager_SchedulerTracksMarkResultCooldownAndRecovery(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, &RoundRobinSelector{}, nil)
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient("auth-a", "gemini", []*registry.ModelInfo{{ID: "test-model"}})
+	reg.RegisterClient("auth-b", "gemini", []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		reg.UnregisterClient("auth-a")
+		reg.UnregisterClient("auth-b")
+	})
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-a", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-a) error = %v", errRegister)
+	}
+	if _, errRegister := manager.Register(context.Background(), &Auth{ID: "auth-b", Provider: "gemini"}); errRegister != nil {
+		t.Fatalf("Register(auth-b) error = %v", errRegister)
+	}
+
+	manager.MarkResult(context.Background(), Result{
+		AuthID:   "auth-a",
+		Provider: "gemini",
+		Model:    "test-model",
+		Success:  false,
+		Error:    &Error{HTTPStatus: 429, Message: "quota"},
+	})
+
+	got, errPick := manager.scheduler.pickSingle(context.Background(), "gemini", "test-model", cliproxyexecutor.Options{}, nil)
+	if errPick != nil {
+		t.Fatalf("scheduler.pickSingle() after cooldown error = %v", errPick)
+	}
+	if got == nil || got.ID != "auth-b" {
+		t.Fatalf("scheduler.pickSingle() after cooldown auth = %v, want auth-b", got)
+	}
+
+	manager.MarkResult(context.Background(), Result{
+		AuthID:   "auth-a",
+		Provider: "gemini",
+		Model:    "test-model",
+		Success:  true,
+	})
+
+	seen := make(map[string]struct{}, 2)
+	for index := 0; index < 2; index++ {
+		got, errPick = manager.scheduler.pickSingle(context.Background(), "gemini", "test-model", cliproxyexecutor.Options{}, nil)
+		if errPick != nil {
+			t.Fatalf("scheduler.pickSingle() after recovery #%d error = %v", index, errPick)
+		}
+		if got == nil {
+			t.Fatalf("scheduler.pickSingle() after recovery #%d auth = nil", index)
+		}
+		seen[got.ID] = struct{}{}
+	}
+	if len(seen) != 2 {
+		t.Fatalf("len(seen) = %d, want %d", len(seen), 2)
+	}
+}