feat(translator): improve signature handling by associating with model name in cache functions

feat(translator): enhance session ID derivation with user_id parsing in Claude

README.md

@@ -130,6 +130,10 @@ Windows-native CLIProxyAPI fork with TUI, system tray, and multi-provider OAuth
 
 VSCode extension for quick switching between Claude Code models, featuring integrated CLIProxyAPI as its backend with automatic background lifecycle management.
 
+### [ZeroLimit](https://github.com/0xtbug/zero-limit)
+
+Windows desktop app built with Tauri + React for monitoring AI coding assistant quotas via CLIProxyAPI. Track usage across Gemini, Claude, OpenAI Codex, and Antigravity accounts with real-time dashboard, system tray integration, and one-click proxy control - no API keys needed.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 

README_CN.md

@@ -129,6 +129,10 @@ CLI 封装器，用于通过 CLIProxyAPI OAuth 即时切换多个 Claude 账户
 
 一款 VSCode 扩展，提供了在 VSCode 中快速切换 Claude Code 模型的功能，内置 CLIProxyAPI 作为其后端，支持后台自动启动和关闭。
 
+### [ZeroLimit](https://github.com/0xtbug/zero-limit)
+
+Windows 桌面应用，基于 Tauri + React 构建，用于通过 CLIProxyAPI 监控 AI 编程助手配额。支持跨 Gemini、Claude、OpenAI Codex 和 Antigravity 账户的使用量追踪，提供实时仪表盘、系统托盘集成和一键代理控制，无需 API 密钥。
+
 > [!NOTE]  
 > 如果你开发了基于 CLIProxyAPI 的项目，请提交一个 PR（拉取请求）将其添加到此列表中。
 

config.example.yaml

@@ -141,6 +141,15 @@ codex-instructions-enabled: false
 #       - "claude-3-*"               # wildcard matching prefix (e.g. claude-3-7-sonnet-20250219)
 #       - "*-thinking"               # wildcard matching suffix (e.g. claude-opus-4-5-thinking)
 #       - "*haiku*"                  # wildcard matching substring (e.g. claude-3-5-haiku-20241022)
+#     cloak:                         # optional: request cloaking for non-Claude-Code clients
+#       mode: "auto"                 # "auto" (default): cloak only when client is not Claude Code
+#                                    # "always": always apply cloaking
+#                                    # "never": never apply cloaking
+#       strict-mode: false           # false (default): prepend Claude Code prompt to user system messages
+#                                    # true: strip all user system messages, keep only Claude Code prompt
+#       sensitive-words:             # optional: words to obfuscate with zero-width characters
+#         - "API"
+#         - "proxy"
 
 # OpenAI compatibility providers
 # openai-compatibility:

internal/api/handlers/management/auth_files.go

@@ -3,6 +3,8 @@ package management
 import (
 	"bytes"
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -1383,9 +1385,16 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 		claims, _ := codex.ParseJWTToken(tokenResp.IDToken)
 		email := ""
 		accountID := ""
+		planType := ""
 		if claims != nil {
 			email = claims.GetUserEmail()
 			accountID = claims.GetAccountID()
+			planType = strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType)
+		}
+		hashAccountID := ""
+		if accountID != "" {
+			digest := sha256.Sum256([]byte(accountID))
+			hashAccountID = hex.EncodeToString(digest[:])[:8]
 		}
 		// Build bundle compatible with existing storage
 		bundle := &codex.CodexAuthBundle{
@@ -1402,10 +1411,11 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 
 		// Create token storage and persist
 		tokenStorage := openaiAuth.CreateTokenStorage(bundle)
+		fileName := codex.CredentialFileName(tokenStorage.Email, planType, hashAccountID, true)
 		record := &coreauth.Auth{
-			ID:       fmt.Sprintf("codex-%s.json", tokenStorage.Email),
+			ID:       fileName,
 			Provider: "codex",
-			FileName: fmt.Sprintf("codex-%s.json", tokenStorage.Email),
+			FileName: fileName,
 			Storage:  tokenStorage,
 			Metadata: map[string]any{
 				"email":      tokenStorage.Email,
@@ -1703,7 +1713,7 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 		// Create token storage
 		tokenStorage := qwenAuth.CreateTokenStorage(tokenData)
 
-		tokenStorage.Email = fmt.Sprintf("qwen-%d", time.Now().UnixMilli())
+		tokenStorage.Email = fmt.Sprintf("%d", time.Now().UnixMilli())
 		record := &coreauth.Auth{
 			ID:       fmt.Sprintf("qwen-%s.json", tokenStorage.Email),
 			Provider: "qwen",
@@ -1808,7 +1818,7 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 		tokenStorage := authSvc.CreateTokenStorage(tokenData)
 		identifier := strings.TrimSpace(tokenStorage.Email)
 		if identifier == "" {
-			identifier = fmt.Sprintf("iflow-%d", time.Now().UnixMilli())
+			identifier = fmt.Sprintf("%d", time.Now().UnixMilli())
 			tokenStorage.Email = identifier
 		}
 		record := &coreauth.Auth{
@@ -1893,15 +1903,17 @@ func (h *Handler) RequestIFlowCookieToken(c *gin.Context) {
 	fileName := iflowauth.SanitizeIFlowFileName(email)
 	if fileName == "" {
 		fileName = fmt.Sprintf("iflow-%d", time.Now().UnixMilli())
+	} else {
+		fileName = fmt.Sprintf("iflow-%s", fileName)
 	}
 
 	tokenStorage.Email = email
 	timestamp := time.Now().Unix()
 
 	record := &coreauth.Auth{
-		ID:       fmt.Sprintf("iflow-%s-%d.json", fileName, timestamp),
+		ID:       fmt.Sprintf("%s-%d.json", fileName, timestamp),
 		Provider: "iflow",
-		FileName: fmt.Sprintf("iflow-%s-%d.json", fileName, timestamp),
+		FileName: fmt.Sprintf("%s-%d.json", fileName, timestamp),
 		Storage:  tokenStorage,
 		Metadata: map[string]any{
 			"email":        email,
@@ -2108,7 +2120,20 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 			finalProjectID := projectID
 			if responseProjectID != "" {
 				if explicitProject && !strings.EqualFold(responseProjectID, projectID) {
-					log.Warnf("Gemini onboarding returned project %s instead of requested %s; keeping requested project ID.", responseProjectID, projectID)
+					// Check if this is a free user (gen-lang-client projects or free/legacy tier)
+					isFreeUser := strings.HasPrefix(projectID, "gen-lang-client-") ||
+						strings.EqualFold(tierID, "FREE") ||
+						strings.EqualFold(tierID, "LEGACY")
+
+					if isFreeUser {
+						// For free users, use backend project ID for preview model access
+						log.Infof("Gemini onboarding: frontend project %s maps to backend project %s", projectID, responseProjectID)
+						log.Infof("Using backend project ID: %s (recommended for preview model access)", responseProjectID)
+						finalProjectID = responseProjectID
+					} else {
+						// Pro users: keep requested project ID (original behavior)
+						log.Warnf("Gemini onboarding returned project %s instead of requested %s; keeping requested project ID.", responseProjectID, projectID)
+					}
 				} else {
 					finalProjectID = responseProjectID
 				}

internal/api/handlers/management/logs.go

@@ -13,7 +13,7 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 )
 
 const (
@@ -360,16 +360,7 @@ func (h *Handler) logDirectory() string {
 	if h.logDir != "" {
 		return h.logDir
 	}
-	if base := util.WritablePath(); base != "" {
-		return filepath.Join(base, "logs")
-	}
-	if h.configFilePath != "" {
-		dir := filepath.Dir(h.configFilePath)
-		if dir != "" && dir != "." {
-			return filepath.Join(dir, "logs")
-		}
-	}
-	return "logs"
+	return logging.ResolveLogDirectory(h.cfg)
 }
 
 func (h *Handler) collectLogFiles(dir string) ([]string, error) {

internal/api/server.go

@@ -261,10 +261,7 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	if optionState.localPassword != "" {
 		s.mgmt.SetLocalPassword(optionState.localPassword)
 	}
-	logDir := filepath.Join(s.currentPath, "logs")
-	if base := util.WritablePath(); base != "" {
-		logDir = filepath.Join(base, "logs")
-	}
+	logDir := logging.ResolveLogDirectory(cfg)
 	s.mgmt.SetLogDirectory(logDir)
 	s.localPassword = optionState.localPassword
 

internal/auth/codex/filename.go

@@ -0,0 +1,57 @@
+package codex
+
+import (
+	"fmt"
+	"strings"
+	"unicode"
+
+	"golang.org/x/text/cases"
+	"golang.org/x/text/language"
+)
+
+// CredentialFileName returns the filename used to persist Codex OAuth credentials.
+// When planType is available (e.g. "plus", "team"), it is appended after the email
+// as a suffix to disambiguate subscriptions.
+func CredentialFileName(email, planType, hashAccountID string, includeProviderPrefix bool) string {
+	email = strings.TrimSpace(email)
+	plan := normalizePlanTypeForFilename(planType)
+
+	prefix := ""
+	if includeProviderPrefix {
+		prefix = "codex"
+	}
+
+	if plan == "" {
+		return fmt.Sprintf("%s-%s.json", prefix, email)
+	} else if plan == "team" {
+		return fmt.Sprintf("%s-%s-%s-%s.json", prefix, hashAccountID, email, plan)
+	}
+	return fmt.Sprintf("%s-%s-%s.json", prefix, email, plan)
+}
+
+func normalizePlanTypeForFilename(planType string) string {
+	planType = strings.TrimSpace(planType)
+	if planType == "" {
+		return ""
+	}
+
+	parts := strings.FieldsFunc(planType, func(r rune) bool {
+		return !unicode.IsLetter(r) && !unicode.IsDigit(r)
+	})
+	if len(parts) == 0 {
+		return ""
+	}
+
+	for i, part := range parts {
+		parts[i] = titleToken(part)
+	}
+	return strings.Join(parts, "-")
+}
+
+func titleToken(token string) string {
+	token = strings.TrimSpace(token)
+	if token == "" {
+		return ""
+	}
+	return cases.Title(language.English).String(token)
+}

internal/cache/signature_cache.go

@@ -3,6 +3,7 @@ package cache
 import (
 	"crypto/sha256"
 	"encoding/hex"
+	"fmt"
 	"sync"
 	"time"
 )
@@ -94,7 +95,7 @@ func purgeExpiredSessions() {
 
 // CacheSignature stores a thinking signature for a given session and text.
 // Used for Claude models that require signed thinking blocks in multi-turn conversations.
-func CacheSignature(sessionID, text, signature string) {
+func CacheSignature(modelName, sessionID, text, signature string) {
 	if sessionID == "" || text == "" || signature == "" {
 		return
 	}
@@ -102,7 +103,7 @@ func CacheSignature(sessionID, text, signature string) {
 		return
 	}
 
-	sc := getOrCreateSession(sessionID)
+	sc := getOrCreateSession(fmt.Sprintf("%s#%s", modelName, sessionID))
 	textHash := hashText(text)
 
 	sc.mu.Lock()
@@ -116,12 +117,12 @@ func CacheSignature(sessionID, text, signature string) {
 
 // GetCachedSignature retrieves a cached signature for a given session and text.
 // Returns empty string if not found or expired.
-func GetCachedSignature(sessionID, text string) string {
+func GetCachedSignature(modelName, sessionID, text string) string {
 	if sessionID == "" || text == "" {
 		return ""
 	}
 
-	val, ok := signatureCache.Load(sessionID)
+	val, ok := signatureCache.Load(fmt.Sprintf("%s#%s", modelName, sessionID))
 	if !ok {
 		return ""
 	}

internal/cmd/login.go

@@ -118,6 +118,7 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 	}
 
 	activatedProjects := make([]string, 0, len(projectSelections))
+	seenProjects := make(map[string]bool)
 	for _, candidateID := range projectSelections {
 		log.Infof("Activating project %s", candidateID)
 		if errSetup := performGeminiCLISetup(ctx, httpClient, storage, candidateID); errSetup != nil {
@@ -134,6 +135,13 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 		if finalID == "" {
 			finalID = candidateID
 		}
+
+		// Skip duplicates
+		if seenProjects[finalID] {
+			log.Infof("Project %s already activated, skipping", finalID)
+			continue
+		}
+		seenProjects[finalID] = true
 		activatedProjects = append(activatedProjects, finalID)
 	}
 
@@ -261,7 +269,39 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 			finalProjectID := projectID
 			if responseProjectID != "" {
 				if explicitProject && !strings.EqualFold(responseProjectID, projectID) {
-					log.Warnf("Gemini onboarding returned project %s instead of requested %s; keeping requested project ID.", responseProjectID, projectID)
+					// Check if this is a free user (gen-lang-client projects or free/legacy tier)
+					isFreeUser := strings.HasPrefix(projectID, "gen-lang-client-") ||
+						strings.EqualFold(tierID, "FREE") ||
+						strings.EqualFold(tierID, "LEGACY")
+
+					if isFreeUser {
+						// Interactive prompt for free users
+						fmt.Printf("\nGoogle returned a different project ID:\n")
+						fmt.Printf("  Requested (frontend): %s\n", projectID)
+						fmt.Printf("  Returned (backend):   %s\n\n", responseProjectID)
+						fmt.Printf("  Backend project IDs have access to preview models (gemini-3-*).\n")
+						fmt.Printf("  This is normal for free tier users.\n\n")
+						fmt.Printf("Which project ID would you like to use?\n")
+						fmt.Printf("  [1] Backend (recommended): %s\n", responseProjectID)
+						fmt.Printf("  [2] Frontend: %s\n\n", projectID)
+						fmt.Printf("Enter choice [1]: ")
+
+						reader := bufio.NewReader(os.Stdin)
+						choice, _ := reader.ReadString('\n')
+						choice = strings.TrimSpace(choice)
+
+						if choice == "2" {
+							log.Infof("Using frontend project ID: %s", projectID)
+							fmt.Println(". Warning: Frontend project IDs may not have access to preview models.")
+							finalProjectID = projectID
+						} else {
+							log.Infof("Using backend project ID: %s (recommended)", responseProjectID)
+							finalProjectID = responseProjectID
+						}
+					} else {
+						// Pro users: keep requested project ID (original behavior)
+						log.Warnf("Gemini onboarding returned project %s instead of requested %s; keeping requested project ID.", responseProjectID, projectID)
+					}
 				} else {
 					finalProjectID = responseProjectID
 				}

internal/config/config.go

@@ -248,6 +248,25 @@ type PayloadModelRule struct {
 	Protocol string `yaml:"protocol" json:"protocol"`
 }
 
+// CloakConfig configures request cloaking for non-Claude-Code clients.
+// Cloaking disguises API requests to appear as originating from the official Claude Code CLI.
+type CloakConfig struct {
+	// Mode controls cloaking behavior: "auto" (default), "always", or "never".
+	// - "auto": cloak only when client is not Claude Code (based on User-Agent)
+	// - "always": always apply cloaking regardless of client
+	// - "never": never apply cloaking
+	Mode string `yaml:"mode,omitempty" json:"mode,omitempty"`
+
+	// StrictMode controls how system prompts are handled when cloaking.
+	// - false (default): prepend Claude Code prompt to user system messages
+	// - true: strip all user system messages, keep only Claude Code prompt
+	StrictMode bool `yaml:"strict-mode,omitempty" json:"strict-mode,omitempty"`
+
+	// SensitiveWords is a list of words to obfuscate with zero-width characters.
+	// This can help bypass certain content filters.
+	SensitiveWords []string `yaml:"sensitive-words,omitempty" json:"sensitive-words,omitempty"`
+}
+
 // ClaudeKey represents the configuration for a Claude API key,
 // including the API key itself and an optional base URL for the API endpoint.
 type ClaudeKey struct {
@@ -276,6 +295,9 @@ type ClaudeKey struct {
 
 	// ExcludedModels lists model IDs that should be excluded for this provider.
 	ExcludedModels []string `yaml:"excluded-models,omitempty" json:"excluded-models,omitempty"`
+
+	// Cloak configures request cloaking for non-Claude-Code clients.
+	Cloak *CloakConfig `yaml:"cloak,omitempty" json:"cloak,omitempty"`
 }
 
 func (k ClaudeKey) GetAPIKey() string  { return k.APIKey }

internal/logging/global_logger.go

@@ -30,7 +30,7 @@ var (
 type LogFormatter struct{}
 
 // logFieldOrder defines the display order for common log fields.
-var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_value", "min", "max", "clamped_to", "error"}
+var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_mode", "original_value", "min", "max", "clamped_to", "error"}
 
 // Format renders a single log entry with custom formatting.
 func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {
@@ -121,6 +121,24 @@ func isDirWritable(dir string) bool {
 	return true
 }
 
+// ResolveLogDirectory determines the directory used for application logs.
+func ResolveLogDirectory(cfg *config.Config) string {
+	logDir := "logs"
+	if base := util.WritablePath(); base != "" {
+		return filepath.Join(base, "logs")
+	}
+	if cfg == nil {
+		return logDir
+	}
+	if !isDirWritable(logDir) {
+		authDir := strings.TrimSpace(cfg.AuthDir)
+		if authDir != "" {
+			logDir = filepath.Join(authDir, "logs")
+		}
+	}
+	return logDir
+}
+
 // ConfigureLogOutput switches the global log destination between rotating files and stdout.
 // When logsMaxTotalSizeMB > 0, a background cleaner removes the oldest log files in the logs directory
 // until the total size is within the limit.
@@ -130,12 +148,7 @@ func ConfigureLogOutput(cfg *config.Config) error {
 	writerMu.Lock()
 	defer writerMu.Unlock()
 
-	logDir := "logs"
-	if base := util.WritablePath(); base != "" {
-		logDir = filepath.Join(base, "logs")
-	} else if !isDirWritable(logDir) {
-		logDir = filepath.Join(cfg.AuthDir, "logs")
-	}
+	logDir := ResolveLogDirectory(cfg)
 
 	protectedPath := ""
 	if cfg.LoggingToFile {

internal/registry/model_definitions.go

@@ -287,6 +287,67 @@ func GetGeminiVertexModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		// Imagen image generation models - use :predict action
+		{
+			ID:                         "imagen-4.0-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Generate",
+			Description:                "Imagen 4.0 image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-4.0-ultra-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-ultra-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Ultra Generate",
+			Description:                "Imagen 4.0 Ultra high-quality image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-3.0-generate-002",
+			Object:                     "model",
+			Created:                    1740000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-3.0-generate-002",
+			Version:                    "3.0",
+			DisplayName:                "Imagen 3.0 Generate",
+			Description:                "Imagen 3.0 image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-3.0-fast-generate-001",
+			Object:                     "model",
+			Created:                    1740000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-3.0-fast-generate-001",
+			Version:                    "3.0",
+			DisplayName:                "Imagen 3.0 Fast Generate",
+			Description:                "Imagen 3.0 fast image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
+		{
+			ID:                         "imagen-4.0-fast-generate-001",
+			Object:                     "model",
+			Created:                    1750000000,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/imagen-4.0-fast-generate-001",
+			Version:                    "4.0",
+			DisplayName:                "Imagen 4.0 Fast Generate",
+			Description:                "Imagen 4.0 fast image generation model",
+			SupportedGenerationMethods: []string{"predict"},
+		},
 	}
 }
 
@@ -765,21 +826,23 @@ func GetIFlowModels() []*ModelInfo {
 type AntigravityModelConfig struct {
 	Thinking            *ThinkingSupport
 	MaxCompletionTokens int
-	Name                string
 }
 
 // GetAntigravityModelConfig returns static configuration for antigravity models.
 // Keys use upstream model names returned by the Antigravity models endpoint.
 func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 	return map[string]*AntigravityModelConfig{
-		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash"},
-		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}, Name: "models/gemini-2.5-flash-lite"},
-		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}, Name: "models/rev19-uic3-1p"},
-		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-high"},
-		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}, Name: "models/gemini-3-pro-image"},
-		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}, Name: "models/gemini-3-flash"},
+		"gemini-2.5-flash":           {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
+		"rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}},
+		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
+		"gpt-oss-120b-medium":        {},
+		"tab_flash_lite_preview":     {},
 	}
 }
 
@@ -809,10 +872,9 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 	}
 
 	// Check Antigravity static config
-	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil && cfg.Thinking != nil {
+	if cfg := GetAntigravityModelConfig()[modelID]; cfg != nil {
 		return &ModelInfo{
 			ID:                  modelID,
-			Name:                cfg.Name,
 			Thinking:            cfg.Thinking,
 			MaxCompletionTokens: cfg.MaxCompletionTokens,
 		}

internal/registry/model_registry.go

@@ -78,6 +78,8 @@ type ThinkingSupport struct {
 type ModelRegistration struct {
 	// Info contains the model metadata
 	Info *ModelInfo
+	// InfoByProvider maps provider identifiers to specific ModelInfo to support differing capabilities.
+	InfoByProvider map[string]*ModelInfo
 	// Count is the number of active clients that can provide this model
 	Count int
 	// LastUpdated tracks when this registration was last modified
@@ -132,16 +134,19 @@ func GetGlobalRegistry() *ModelRegistry {
 	return globalRegistry
 }
 
-// LookupModelInfo searches the dynamic registry first, then falls back to static model definitions.
-//
-// This helper exists because some code paths only have a model ID and still need Thinking and
-// max completion token metadata even when the dynamic registry hasn't been populated.
-func LookupModelInfo(modelID string) *ModelInfo {
+// LookupModelInfo searches dynamic registry (provider-specific > global) then static definitions.
+func LookupModelInfo(modelID string, provider ...string) *ModelInfo {
 	modelID = strings.TrimSpace(modelID)
 	if modelID == "" {
 		return nil
 	}
-	if info := GetGlobalRegistry().GetModelInfo(modelID); info != nil {
+
+	p := ""
+	if len(provider) > 0 {
+		p = strings.ToLower(strings.TrimSpace(provider[0]))
+	}
+
+	if info := GetGlobalRegistry().GetModelInfo(modelID, p); info != nil {
 		return info
 	}
 	return LookupStaticModelInfo(modelID)
@@ -297,6 +302,9 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 				if count, okProv := reg.Providers[oldProvider]; okProv {
 					if count <= toRemove {
 						delete(reg.Providers, oldProvider)
+						if reg.InfoByProvider != nil {
+							delete(reg.InfoByProvider, oldProvider)
+						}
 					} else {
 						reg.Providers[oldProvider] = count - toRemove
 					}
@@ -346,6 +354,12 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		model := newModels[id]
 		if reg, ok := r.models[id]; ok {
 			reg.Info = cloneModelInfo(model)
+			if provider != "" {
+				if reg.InfoByProvider == nil {
+					reg.InfoByProvider = make(map[string]*ModelInfo)
+				}
+				reg.InfoByProvider[provider] = cloneModelInfo(model)
+			}
 			reg.LastUpdated = now
 			if reg.QuotaExceededClients != nil {
 				delete(reg.QuotaExceededClients, clientID)
@@ -409,11 +423,15 @@ func (r *ModelRegistry) addModelRegistration(modelID, provider string, model *Mo
 		if existing.SuspendedClients == nil {
 			existing.SuspendedClients = make(map[string]string)
 		}
+		if existing.InfoByProvider == nil {
+			existing.InfoByProvider = make(map[string]*ModelInfo)
+		}
 		if provider != "" {
 			if existing.Providers == nil {
 				existing.Providers = make(map[string]int)
 			}
 			existing.Providers[provider]++
+			existing.InfoByProvider[provider] = cloneModelInfo(model)
 		}
 		log.Debugf("Incremented count for model %s, now %d clients", modelID, existing.Count)
 		return
@@ -421,6 +439,7 @@ func (r *ModelRegistry) addModelRegistration(modelID, provider string, model *Mo
 
 	registration := &ModelRegistration{
 		Info:                 cloneModelInfo(model),
+		InfoByProvider:       make(map[string]*ModelInfo),
 		Count:                1,
 		LastUpdated:          now,
 		QuotaExceededClients: make(map[string]*time.Time),
@@ -428,6 +447,7 @@ func (r *ModelRegistry) addModelRegistration(modelID, provider string, model *Mo
 	}
 	if provider != "" {
 		registration.Providers = map[string]int{provider: 1}
+		registration.InfoByProvider[provider] = cloneModelInfo(model)
 	}
 	r.models[modelID] = registration
 	log.Debugf("Registered new model %s from provider %s", modelID, provider)
@@ -453,6 +473,9 @@ func (r *ModelRegistry) removeModelRegistration(clientID, modelID, provider stri
 		if count, ok := registration.Providers[provider]; ok {
 			if count <= 1 {
 				delete(registration.Providers, provider)
+				if registration.InfoByProvider != nil {
+					delete(registration.InfoByProvider, provider)
+				}
 			} else {
 				registration.Providers[provider] = count - 1
 			}
@@ -534,6 +557,9 @@ func (r *ModelRegistry) unregisterClientInternal(clientID string) {
 				if count, ok := registration.Providers[provider]; ok {
 					if count <= 1 {
 						delete(registration.Providers, provider)
+						if registration.InfoByProvider != nil {
+							delete(registration.InfoByProvider, provider)
+						}
 					} else {
 						registration.Providers[provider] = count - 1
 					}
@@ -940,12 +966,22 @@ func (r *ModelRegistry) GetModelProviders(modelID string) []string {
 	return result
 }
 
-// GetModelInfo returns the registered ModelInfo for the given model ID, if present.
-// Returns nil if the model is unknown to the registry.
-func (r *ModelRegistry) GetModelInfo(modelID string) *ModelInfo {
+// GetModelInfo returns ModelInfo, prioritizing provider-specific definition if available.
+func (r *ModelRegistry) GetModelInfo(modelID, provider string) *ModelInfo {
 	r.mutex.RLock()
 	defer r.mutex.RUnlock()
 	if reg, ok := r.models[modelID]; ok && reg != nil {
+		// Try provider specific definition first
+		if provider != "" && reg.InfoByProvider != nil {
+			if reg.Providers != nil {
+				if count, ok := reg.Providers[provider]; ok && count > 0 {
+					if info, ok := reg.InfoByProvider[provider]; ok && info != nil {
+						return info
+					}
+				}
+			}
+		}
+		// Fallback to global info (last registered)
 		return reg.Info
 	}
 	return nil

internal/runtime/executor/aistudio_executor.go

@@ -393,7 +393,7 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	}
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
-	payload, err := thinking.ApplyThinking(payload, req.Model, "gemini")
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, translatedPayload{}, err
 	}

internal/runtime/executor/antigravity_executor.go

@@ -137,7 +137,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -256,7 +256,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -517,8 +517,8 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 		}
 		if usageResult := responseNode.Get("usageMetadata"); usageResult.Exists() {
 			usageRaw = usageResult.Raw
-		} else if usageResult := root.Get("usageMetadata"); usageResult.Exists() {
-			usageRaw = usageResult.Raw
+		} else if usageMetadataResult := root.Get("usageMetadata"); usageMetadataResult.Exists() {
+			usageRaw = usageMetadataResult.Raw
 		}
 
 		if partsResult := responseNode.Get("candidates.0.content.parts"); partsResult.IsArray() {
@@ -622,7 +622,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -642,7 +642,6 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 			err = errReq
 			return nil, err
 		}
-
 		httpResp, errDo := httpClient.Do(httpReq)
 		if errDo != nil {
 			recordAPIResponseError(ctx, e.cfg, errDo)
@@ -803,7 +802,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	// Prepare payload once (doesn't depend on baseURL)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	payload, err := thinking.ApplyThinking(payload, req.Model, "antigravity")
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
@@ -1004,11 +1003,8 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 			case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
 				continue
 			}
-			cfg := modelConfig[modelID]
+			modelCfg := modelConfig[modelID]
 			modelName := modelID
-			if cfg != nil && cfg.Name != "" {
-				modelName = cfg.Name
-			}
 			modelInfo := &registry.ModelInfo{
 				ID:          modelID,
 				Name:        modelName,
@@ -1021,12 +1017,12 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				Type:        antigravityAuthType,
 			}
 			// Look up Thinking support from static config using upstream model name.
-			if cfg != nil {
-				if cfg.Thinking != nil {
-					modelInfo.Thinking = cfg.Thinking
+			if modelCfg != nil {
+				if modelCfg.Thinking != nil {
+					modelInfo.Thinking = modelCfg.Thinking
 				}
-				if cfg.MaxCompletionTokens > 0 {
-					modelInfo.MaxCompletionTokens = cfg.MaxCompletionTokens
+				if modelCfg.MaxCompletionTokens > 0 {
+					modelInfo.MaxCompletionTokens = modelCfg.MaxCompletionTokens
 				}
 			}
 			models = append(models, modelInfo)
@@ -1411,13 +1407,6 @@ func geminiToAntigravity(modelName string, payload []byte, projectID string) []b
 	template, _ = sjson.Delete(template, "request.safetySettings")
 	template, _ = sjson.Set(template, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
 
-	if !strings.HasPrefix(modelName, "gemini-3-") {
-		if thinkingLevel := gjson.Get(template, "request.generationConfig.thinkingConfig.thinkingLevel"); thinkingLevel.Exists() {
-			template, _ = sjson.Delete(template, "request.generationConfig.thinkingConfig.thinkingLevel")
-			template, _ = sjson.Set(template, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-		}
-	}
-
 	if strings.Contains(modelName, "claude") {
 		gjson.Get(template, "request.tools").ForEach(func(key, tool gjson.Result) bool {
 			tool.Get("functionDeclarations").ForEach(func(funKey, funcDecl gjson.Result) bool {

internal/runtime/executor/claude_executor.go

@@ -17,7 +17,6 @@ import (
 	claudeauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/claude"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"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"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
@@ -106,22 +105,20 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
 
-	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {
-		body = checkSystemInstructions(body)
-	}
+	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
+	// based on client type and configuration.
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(baseModel, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -239,20 +236,20 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
 
-	body = checkSystemInstructions(body)
+	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
+	// based on client type and configuration.
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 
 	// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
 	body = disableThinkingIfToolChoiceForced(body)
 
-	// Ensure max_tokens > thinking.budget_tokens when thinking is enabled
-	body = ensureMaxTokensForThinking(baseModel, body)
-
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -541,81 +538,6 @@ func disableThinkingIfToolChoiceForced(body []byte) []byte {
 	return body
 }
 
-// ensureMaxTokensForThinking ensures max_tokens > thinking.budget_tokens when thinking is enabled.
-// Anthropic API requires this constraint; violating it returns a 400 error.
-// This function should be called after all thinking configuration is finalized.
-// It looks up the model's MaxCompletionTokens from the registry to use as the cap.
-func ensureMaxTokensForThinking(modelName string, body []byte) []byte {
-	thinkingType := gjson.GetBytes(body, "thinking.type").String()
-	if thinkingType != "enabled" {
-		return body
-	}
-
-	budgetTokens := gjson.GetBytes(body, "thinking.budget_tokens").Int()
-	if budgetTokens <= 0 {
-		return body
-	}
-
-	maxTokens := gjson.GetBytes(body, "max_tokens").Int()
-
-	// Look up the model's max completion tokens from the registry
-	maxCompletionTokens := 0
-	if modelInfo := registry.LookupModelInfo(modelName); modelInfo != nil {
-		maxCompletionTokens = modelInfo.MaxCompletionTokens
-	}
-
-	// Fall back to budget + buffer if registry lookup fails or returns 0
-	const fallbackBuffer = 4000
-	requiredMaxTokens := budgetTokens + fallbackBuffer
-	if maxCompletionTokens > 0 {
-		requiredMaxTokens = int64(maxCompletionTokens)
-	}
-
-	if maxTokens < requiredMaxTokens {
-		body, _ = sjson.SetBytes(body, "max_tokens", requiredMaxTokens)
-	}
-	return body
-}
-
-func (e *ClaudeExecutor) resolveClaudeConfig(auth *cliproxyauth.Auth) *config.ClaudeKey {
-	if auth == nil || e.cfg == nil {
-		return nil
-	}
-	var attrKey, attrBase string
-	if auth.Attributes != nil {
-		attrKey = strings.TrimSpace(auth.Attributes["api_key"])
-		attrBase = strings.TrimSpace(auth.Attributes["base_url"])
-	}
-	for i := range e.cfg.ClaudeKey {
-		entry := &e.cfg.ClaudeKey[i]
-		cfgKey := strings.TrimSpace(entry.APIKey)
-		cfgBase := strings.TrimSpace(entry.BaseURL)
-		if attrKey != "" && attrBase != "" {
-			if strings.EqualFold(cfgKey, attrKey) && strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-			continue
-		}
-		if attrKey != "" && strings.EqualFold(cfgKey, attrKey) {
-			if cfgBase == "" || strings.EqualFold(cfgBase, attrBase) {
-				return entry
-			}
-		}
-		if attrKey == "" && attrBase != "" && strings.EqualFold(cfgBase, attrBase) {
-			return entry
-		}
-	}
-	if attrKey != "" {
-		for i := range e.cfg.ClaudeKey {
-			entry := &e.cfg.ClaudeKey[i]
-			if strings.EqualFold(strings.TrimSpace(entry.APIKey), attrKey) {
-				return entry
-			}
-		}
-	}
-	return nil
-}
-
 type compositeReadCloser struct {
 	io.Reader
 	closers []func() error
@@ -901,3 +823,163 @@ func stripClaudeToolPrefixFromStreamLine(line []byte, prefix string) []byte {
 	}
 	return updated
 }
+
+// getClientUserAgent extracts the client User-Agent from the gin context.
+func getClientUserAgent(ctx context.Context) string {
+	if ginCtx, ok := ctx.Value("gin").(*gin.Context); ok && ginCtx != nil && ginCtx.Request != nil {
+		return ginCtx.GetHeader("User-Agent")
+	}
+	return ""
+}
+
+// getCloakConfigFromAuth extracts cloak configuration from auth attributes.
+// Returns (cloakMode, strictMode, sensitiveWords).
+func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string) {
+	if auth == nil || auth.Attributes == nil {
+		return "auto", false, nil
+	}
+
+	cloakMode := auth.Attributes["cloak_mode"]
+	if cloakMode == "" {
+		cloakMode = "auto"
+	}
+
+	strictMode := strings.ToLower(auth.Attributes["cloak_strict_mode"]) == "true"
+
+	var sensitiveWords []string
+	if wordsStr := auth.Attributes["cloak_sensitive_words"]; wordsStr != "" {
+		sensitiveWords = strings.Split(wordsStr, ",")
+		for i := range sensitiveWords {
+			sensitiveWords[i] = strings.TrimSpace(sensitiveWords[i])
+		}
+	}
+
+	return cloakMode, strictMode, sensitiveWords
+}
+
+// resolveClaudeKeyCloakConfig finds the matching ClaudeKey config and returns its CloakConfig.
+func resolveClaudeKeyCloakConfig(cfg *config.Config, auth *cliproxyauth.Auth) *config.CloakConfig {
+	if cfg == nil || auth == nil {
+		return nil
+	}
+
+	apiKey, baseURL := claudeCreds(auth)
+	if apiKey == "" {
+		return nil
+	}
+
+	for i := range cfg.ClaudeKey {
+		entry := &cfg.ClaudeKey[i]
+		cfgKey := strings.TrimSpace(entry.APIKey)
+		cfgBase := strings.TrimSpace(entry.BaseURL)
+
+		// Match by API key
+		if strings.EqualFold(cfgKey, apiKey) {
+			// If baseURL is specified, also check it
+			if baseURL != "" && cfgBase != "" && !strings.EqualFold(cfgBase, baseURL) {
+				continue
+			}
+			return entry.Cloak
+		}
+	}
+
+	return nil
+}
+
+// injectFakeUserID generates and injects a fake user ID into the request metadata.
+func injectFakeUserID(payload []byte) []byte {
+	metadata := gjson.GetBytes(payload, "metadata")
+	if !metadata.Exists() {
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+		return payload
+	}
+
+	existingUserID := gjson.GetBytes(payload, "metadata.user_id").String()
+	if existingUserID == "" || !isValidUserID(existingUserID) {
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+	}
+	return payload
+}
+
+// checkSystemInstructionsWithMode injects Claude Code system prompt.
+// In strict mode, it replaces all user system messages.
+// In non-strict mode (default), it prepends to existing system messages.
+func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
+	system := gjson.GetBytes(payload, "system")
+	claudeCodeInstructions := `[{"type":"text","text":"You are Claude Code, Anthropic's official CLI for Claude."}]`
+
+	if strictMode {
+		// Strict mode: replace all system messages with Claude Code prompt only
+		payload, _ = sjson.SetRawBytes(payload, "system", []byte(claudeCodeInstructions))
+		return payload
+	}
+
+	// Non-strict mode (default): prepend Claude Code prompt to existing system messages
+	if system.IsArray() {
+		if gjson.GetBytes(payload, "system.0.text").String() != "You are Claude Code, Anthropic's official CLI for Claude." {
+			system.ForEach(func(_, part gjson.Result) bool {
+				if part.Get("type").String() == "text" {
+					claudeCodeInstructions, _ = sjson.SetRaw(claudeCodeInstructions, "-1", part.Raw)
+				}
+				return true
+			})
+			payload, _ = sjson.SetRawBytes(payload, "system", []byte(claudeCodeInstructions))
+		}
+	} else {
+		payload, _ = sjson.SetRawBytes(payload, "system", []byte(claudeCodeInstructions))
+	}
+	return payload
+}
+
+// applyCloaking applies cloaking transformations to the payload based on config and client.
+// Cloaking includes: system prompt injection, fake user ID, and sensitive word obfuscation.
+func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte, model string) []byte {
+	clientUserAgent := getClientUserAgent(ctx)
+
+	// Get cloak config from ClaudeKey configuration
+	cloakCfg := resolveClaudeKeyCloakConfig(cfg, auth)
+
+	// Determine cloak settings
+	var cloakMode string
+	var strictMode bool
+	var sensitiveWords []string
+
+	if cloakCfg != nil {
+		cloakMode = cloakCfg.Mode
+		strictMode = cloakCfg.StrictMode
+		sensitiveWords = cloakCfg.SensitiveWords
+	}
+
+	// Fallback to auth attributes if no config found
+	if cloakMode == "" {
+		attrMode, attrStrict, attrWords := getCloakConfigFromAuth(auth)
+		cloakMode = attrMode
+		if !strictMode {
+			strictMode = attrStrict
+		}
+		if len(sensitiveWords) == 0 {
+			sensitiveWords = attrWords
+		}
+	}
+
+	// Determine if cloaking should be applied
+	if !shouldCloak(cloakMode, clientUserAgent) {
+		return payload
+	}
+
+	// Skip system instructions for claude-3-5-haiku models
+	if !strings.HasPrefix(model, "claude-3-5-haiku") {
+		payload = checkSystemInstructionsWithMode(payload, strictMode)
+	}
+
+	// Inject fake user ID
+	payload = injectFakeUserID(payload)
+
+	// Apply sensitive word obfuscation
+	if len(sensitiveWords) > 0 {
+		matcher := buildSensitiveWordMatcher(sensitiveWords)
+		payload = obfuscateSensitiveWords(payload, matcher)
+	}
+
+	return payload
+}

internal/runtime/executor/cloak_obfuscate.go

@@ -0,0 +1,176 @@
+package executor
+
+import (
+	"regexp"
+	"sort"
+	"strings"
+	"unicode/utf8"
+
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// zeroWidthSpace is the Unicode zero-width space character used for obfuscation.
+const zeroWidthSpace = "\u200B"
+
+// SensitiveWordMatcher holds the compiled regex for matching sensitive words.
+type SensitiveWordMatcher struct {
+	regex *regexp.Regexp
+}
+
+// buildSensitiveWordMatcher compiles a regex from the word list.
+// Words are sorted by length (longest first) for proper matching.
+func buildSensitiveWordMatcher(words []string) *SensitiveWordMatcher {
+	if len(words) == 0 {
+		return nil
+	}
+
+	// Filter and normalize words
+	var validWords []string
+	for _, w := range words {
+		w = strings.TrimSpace(w)
+		if utf8.RuneCountInString(w) >= 2 && !strings.Contains(w, zeroWidthSpace) {
+			validWords = append(validWords, w)
+		}
+	}
+
+	if len(validWords) == 0 {
+		return nil
+	}
+
+	// Sort by length (longest first) for proper matching
+	sort.Slice(validWords, func(i, j int) bool {
+		return len(validWords[i]) > len(validWords[j])
+	})
+
+	// Escape and join
+	escaped := make([]string, len(validWords))
+	for i, w := range validWords {
+		escaped[i] = regexp.QuoteMeta(w)
+	}
+
+	pattern := "(?i)" + strings.Join(escaped, "|")
+	re, err := regexp.Compile(pattern)
+	if err != nil {
+		return nil
+	}
+
+	return &SensitiveWordMatcher{regex: re}
+}
+
+// obfuscateWord inserts a zero-width space after the first grapheme.
+func obfuscateWord(word string) string {
+	if strings.Contains(word, zeroWidthSpace) {
+		return word
+	}
+
+	// Get first rune
+	r, size := utf8.DecodeRuneInString(word)
+	if r == utf8.RuneError || size >= len(word) {
+		return word
+	}
+
+	return string(r) + zeroWidthSpace + word[size:]
+}
+
+// obfuscateText replaces all sensitive words in the text.
+func (m *SensitiveWordMatcher) obfuscateText(text string) string {
+	if m == nil || m.regex == nil {
+		return text
+	}
+	return m.regex.ReplaceAllStringFunc(text, obfuscateWord)
+}
+
+// obfuscateSensitiveWords processes the payload and obfuscates sensitive words
+// in system blocks and message content.
+func obfuscateSensitiveWords(payload []byte, matcher *SensitiveWordMatcher) []byte {
+	if matcher == nil || matcher.regex == nil {
+		return payload
+	}
+
+	// Obfuscate in system blocks
+	payload = obfuscateSystemBlocks(payload, matcher)
+
+	// Obfuscate in messages
+	payload = obfuscateMessages(payload, matcher)
+
+	return payload
+}
+
+// obfuscateSystemBlocks obfuscates sensitive words in system blocks.
+func obfuscateSystemBlocks(payload []byte, matcher *SensitiveWordMatcher) []byte {
+	system := gjson.GetBytes(payload, "system")
+	if !system.Exists() {
+		return payload
+	}
+
+	if system.IsArray() {
+		modified := false
+		system.ForEach(func(key, value gjson.Result) bool {
+			if value.Get("type").String() == "text" {
+				text := value.Get("text").String()
+				obfuscated := matcher.obfuscateText(text)
+				if obfuscated != text {
+					path := "system." + key.String() + ".text"
+					payload, _ = sjson.SetBytes(payload, path, obfuscated)
+					modified = true
+				}
+			}
+			return true
+		})
+		if modified {
+			return payload
+		}
+	} else if system.Type == gjson.String {
+		text := system.String()
+		obfuscated := matcher.obfuscateText(text)
+		if obfuscated != text {
+			payload, _ = sjson.SetBytes(payload, "system", obfuscated)
+		}
+	}
+
+	return payload
+}
+
+// obfuscateMessages obfuscates sensitive words in message content.
+func obfuscateMessages(payload []byte, matcher *SensitiveWordMatcher) []byte {
+	messages := gjson.GetBytes(payload, "messages")
+	if !messages.Exists() || !messages.IsArray() {
+		return payload
+	}
+
+	messages.ForEach(func(msgKey, msg gjson.Result) bool {
+		content := msg.Get("content")
+		if !content.Exists() {
+			return true
+		}
+
+		msgPath := "messages." + msgKey.String()
+
+		if content.Type == gjson.String {
+			// Simple string content
+			text := content.String()
+			obfuscated := matcher.obfuscateText(text)
+			if obfuscated != text {
+				payload, _ = sjson.SetBytes(payload, msgPath+".content", obfuscated)
+			}
+		} else if content.IsArray() {
+			// Array of content blocks
+			content.ForEach(func(blockKey, block gjson.Result) bool {
+				if block.Get("type").String() == "text" {
+					text := block.Get("text").String()
+					obfuscated := matcher.obfuscateText(text)
+					if obfuscated != text {
+						path := msgPath + ".content." + blockKey.String() + ".text"
+						payload, _ = sjson.SetBytes(payload, path, obfuscated)
+					}
+				}
+				return true
+			})
+		}
+
+		return true
+	})
+
+	return payload
+}

internal/runtime/executor/cloak_utils.go

@@ -0,0 +1,47 @@
+package executor
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"regexp"
+	"strings"
+
+	"github.com/google/uuid"
+)
+
+// userIDPattern matches Claude Code format: user_[64-hex]_account__session_[uuid-v4]
+var userIDPattern = regexp.MustCompile(`^user_[a-fA-F0-9]{64}_account__session_[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$`)
+
+// generateFakeUserID generates a fake user ID in Claude Code format.
+// Format: user_[64-hex-chars]_account__session_[UUID-v4]
+func generateFakeUserID() string {
+	hexBytes := make([]byte, 32)
+	_, _ = rand.Read(hexBytes)
+	hexPart := hex.EncodeToString(hexBytes)
+	uuidPart := uuid.New().String()
+	return "user_" + hexPart + "_account__session_" + uuidPart
+}
+
+// isValidUserID checks if a user ID matches Claude Code format.
+func isValidUserID(userID string) bool {
+	return userIDPattern.MatchString(userID)
+}
+
+// shouldCloak determines if request should be cloaked based on config and client User-Agent.
+// Returns true if cloaking should be applied.
+func shouldCloak(cloakMode string, userAgent string) bool {
+	switch strings.ToLower(cloakMode) {
+	case "always":
+		return true
+	case "never":
+		return false
+	default: // "auto" or empty
+		// If client is Claude Code, don't cloak
+		return !strings.HasPrefix(userAgent, "claude-cli")
+	}
+}
+
+// isClaudeCodeClient checks if the User-Agent indicates a Claude Code client.
+func isClaudeCodeClient(userAgent string) bool {
+	return strings.HasPrefix(userAgent, "claude-cli")
+}

internal/runtime/executor/codex_executor.go

@@ -96,7 +96,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -208,7 +208,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, true)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -316,7 +316,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err := thinking.ApplyThinking(body, req.Model, "codex")
+	body, err := thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/gemini_cli_executor.go

@@ -123,7 +123,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -272,7 +272,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -479,7 +479,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	for range models {
 		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-		payload, err = thinking.ApplyThinking(payload, req.Model, "gemini-cli")
+		payload, err = thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 		if err != nil {
 			return cliproxyexecutor.Response{}, err
 		}

internal/runtime/executor/gemini_executor.go

@@ -120,7 +120,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -222,7 +222,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -338,7 +338,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/gemini_vertex_executor.go

@@ -12,6 +12,7 @@ import (
 	"io"
 	"net/http"
 	"strings"
+	"time"
 
 	vertexauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/vertex"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
@@ -31,6 +32,143 @@ const (
 	vertexAPIVersion = "v1"
 )
 
+// isImagenModel checks if the model name is an Imagen image generation model.
+// Imagen models use the :predict action instead of :generateContent.
+func isImagenModel(model string) bool {
+	lowerModel := strings.ToLower(model)
+	return strings.Contains(lowerModel, "imagen")
+}
+
+// getVertexAction returns the appropriate action for the given model.
+// Imagen models use "predict", while Gemini models use "generateContent".
+func getVertexAction(model string, isStream bool) string {
+	if isImagenModel(model) {
+		return "predict"
+	}
+	if isStream {
+		return "streamGenerateContent"
+	}
+	return "generateContent"
+}
+
+// convertImagenToGeminiResponse converts Imagen API response to Gemini format
+// so it can be processed by the standard translation pipeline.
+// This ensures Imagen models return responses in the same format as gemini-3-pro-image-preview.
+func convertImagenToGeminiResponse(data []byte, model string) []byte {
+	predictions := gjson.GetBytes(data, "predictions")
+	if !predictions.Exists() || !predictions.IsArray() {
+		return data
+	}
+
+	// Build Gemini-compatible response with inlineData
+	parts := make([]map[string]any, 0)
+	for _, pred := range predictions.Array() {
+		imageData := pred.Get("bytesBase64Encoded").String()
+		mimeType := pred.Get("mimeType").String()
+		if mimeType == "" {
+			mimeType = "image/png"
+		}
+		if imageData != "" {
+			parts = append(parts, map[string]any{
+				"inlineData": map[string]any{
+					"mimeType": mimeType,
+					"data":     imageData,
+				},
+			})
+		}
+	}
+
+	// Generate unique response ID using timestamp
+	responseId := fmt.Sprintf("imagen-%d", time.Now().UnixNano())
+
+	response := map[string]any{
+		"candidates": []map[string]any{{
+			"content": map[string]any{
+				"parts": parts,
+				"role":  "model",
+			},
+			"finishReason": "STOP",
+		}},
+		"responseId":   responseId,
+		"modelVersion": model,
+		// Imagen API doesn't return token counts, set to 0 for tracking purposes
+		"usageMetadata": map[string]any{
+			"promptTokenCount":     0,
+			"candidatesTokenCount": 0,
+			"totalTokenCount":      0,
+		},
+	}
+
+	result, err := json.Marshal(response)
+	if err != nil {
+		return data
+	}
+	return result
+}
+
+// convertToImagenRequest converts a Gemini-style request to Imagen API format.
+// Imagen API uses a different structure: instances[].prompt instead of contents[].
+func convertToImagenRequest(payload []byte) ([]byte, error) {
+	// Extract prompt from Gemini-style contents
+	prompt := ""
+
+	// Try to get prompt from contents[0].parts[0].text
+	contentsText := gjson.GetBytes(payload, "contents.0.parts.0.text")
+	if contentsText.Exists() {
+		prompt = contentsText.String()
+	}
+
+	// If no contents, try messages format (OpenAI-compatible)
+	if prompt == "" {
+		messagesText := gjson.GetBytes(payload, "messages.#.content")
+		if messagesText.Exists() && messagesText.IsArray() {
+			for _, msg := range messagesText.Array() {
+				if msg.String() != "" {
+					prompt = msg.String()
+					break
+				}
+			}
+		}
+	}
+
+	// If still no prompt, try direct prompt field
+	if prompt == "" {
+		directPrompt := gjson.GetBytes(payload, "prompt")
+		if directPrompt.Exists() {
+			prompt = directPrompt.String()
+		}
+	}
+
+	if prompt == "" {
+		return nil, fmt.Errorf("imagen: no prompt found in request")
+	}
+
+	// Build Imagen API request
+	imagenReq := map[string]any{
+		"instances": []map[string]any{
+			{
+				"prompt": prompt,
+			},
+		},
+		"parameters": map[string]any{
+			"sampleCount": 1,
+		},
+	}
+
+	// Extract optional parameters
+	if aspectRatio := gjson.GetBytes(payload, "aspectRatio"); aspectRatio.Exists() {
+		imagenReq["parameters"].(map[string]any)["aspectRatio"] = aspectRatio.String()
+	}
+	if sampleCount := gjson.GetBytes(payload, "sampleCount"); sampleCount.Exists() {
+		imagenReq["parameters"].(map[string]any)["sampleCount"] = int(sampleCount.Int())
+	}
+	if negativePrompt := gjson.GetBytes(payload, "negativePrompt"); negativePrompt.Exists() {
+		imagenReq["instances"].([]map[string]any)[0]["negativePrompt"] = negativePrompt.String()
+	}
+
+	return json.Marshal(imagenReq)
+}
+
 // GeminiVertexExecutor sends requests to Vertex AI Gemini endpoints using service account credentials.
 type GeminiVertexExecutor struct {
 	cfg *config.Config
@@ -160,26 +298,38 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
 	defer reporter.trackFailure(ctx, &err)
 
-	from := opts.SourceFormat
-	to := sdktranslator.FromString("gemini")
+	var body []byte
 
-	originalPayload := bytes.Clone(req.Payload)
-	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
-	}
-	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	// Handle Imagen models with special request format
+	if isImagenModel(baseModel) {
+		imagenBody, errImagen := convertToImagenRequest(req.Payload)
+		if errImagen != nil {
+			return resp, errImagen
+		}
+		body = imagenBody
+	} else {
+		// Standard Gemini translation flow
+		from := opts.SourceFormat
+		to := sdktranslator.FromString("gemini")
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
-	if err != nil {
-		return resp, err
+		originalPayload := bytes.Clone(req.Payload)
+		if len(opts.OriginalRequest) > 0 {
+			originalPayload = bytes.Clone(opts.OriginalRequest)
+		}
+		originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
+		body = sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+
+		body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
+		if err != nil {
+			return resp, err
+		}
+
+		body = fixGeminiImageAspectRatio(baseModel, body)
+		body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
+		body, _ = sjson.SetBytes(body, "model", baseModel)
 	}
 
-	body = fixGeminiImageAspectRatio(baseModel, body)
-	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
-	body, _ = sjson.SetBytes(body, "model", baseModel)
-
-	action := "generateContent"
+	action := getVertexAction(baseModel, false)
 	if req.Metadata != nil {
 		if a, _ := req.Metadata["action"].(string); a == "countTokens" {
 			action = "countTokens"
@@ -249,6 +399,16 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	}
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	reporter.publish(ctx, parseGeminiUsage(data))
+
+	// For Imagen models, convert response to Gemini format before translation
+	// This ensures Imagen responses use the same format as gemini-3-pro-image-preview
+	if isImagenModel(baseModel) {
+		data = convertImagenToGeminiResponse(data, baseModel)
+	}
+
+	// Standard Gemini translation (works for both Gemini and converted Imagen responses)
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("gemini")
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
@@ -272,7 +432,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -281,7 +441,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	action := "generateContent"
+	action := getVertexAction(baseModel, false)
 	if req.Metadata != nil {
 		if a, _ := req.Metadata["action"].(string); a == "countTokens" {
 			action = "countTokens"
@@ -375,7 +535,7 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -384,12 +544,16 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
+	action := getVertexAction(baseModel, true)
 	baseURL := vertexBaseURL(location)
-	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, "streamGenerateContent")
-	if opts.Alt == "" {
-		url = url + "?alt=sse"
-	} else {
-		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
+	url := fmt.Sprintf("%s/%s/projects/%s/locations/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, projectID, location, baseModel, action)
+	// Imagen models don't support streaming, skip SSE params
+	if !isImagenModel(baseModel) {
+		if opts.Alt == "" {
+			url = url + "?alt=sse"
+		} else {
+			url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
+		}
 	}
 	body, _ = sjson.DeleteBytes(body, "session_id")
 
@@ -494,7 +658,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -503,15 +667,19 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
+	action := getVertexAction(baseModel, true)
 	// For API key auth, use simpler URL format without project/location
 	if baseURL == "" {
 		baseURL = "https://generativelanguage.googleapis.com"
 	}
-	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, "streamGenerateContent")
-	if opts.Alt == "" {
-		url = url + "?alt=sse"
-	} else {
-		url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
+	url := fmt.Sprintf("%s/%s/publishers/google/models/%s:%s", baseURL, vertexAPIVersion, baseModel, action)
+	// Imagen models don't support streaming, skip SSE params
+	if !isImagenModel(baseModel) {
+		if opts.Alt == "" {
+			url = url + "?alt=sse"
+		} else {
+			url = url + fmt.Sprintf("?$alt=%s", opts.Alt)
+		}
 	}
 	body, _ = sjson.DeleteBytes(body, "session_id")
 
@@ -605,7 +773,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
@@ -689,7 +857,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/iflow_executor.go

@@ -92,7 +92,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow", e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -190,7 +190,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow", e.Identifier())
 	if err != nil {
 		return nil, err
 	}

internal/runtime/executor/openai_compat_executor.go

@@ -92,7 +92,7 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -187,7 +187,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}
@@ -297,7 +297,7 @@ func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyau
 
 	modelForCounting := baseModel
 
-	translated, err := thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err := thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/payload_helpers.go

@@ -25,6 +25,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	if model == "" {
 		return payload
 	}
+	candidates := payloadModelCandidates(cfg, model, protocol)
 	out := payload
 	source := original
 	if len(source) == 0 {
@@ -34,7 +35,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply default rules: first write wins per field across all matching rules.
 	for i := range rules.Default {
 		rule := &rules.Default[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -59,7 +60,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply default raw rules: first write wins per field across all matching rules.
 	for i := range rules.DefaultRaw {
 		rule := &rules.DefaultRaw[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -88,7 +89,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply override rules: last write wins per field across all matching rules.
 	for i := range rules.Override {
 		rule := &rules.Override[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -106,7 +107,7 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	// Apply override raw rules: last write wins per field across all matching rules.
 	for i := range rules.OverrideRaw {
 		rule := &rules.OverrideRaw[i]
-		if !payloadRuleMatchesModel(rule, model, protocol) {
+		if !payloadRuleMatchesModels(rule, protocol, candidates) {
 			continue
 		}
 		for path, value := range rule.Params {
@@ -128,6 +129,18 @@ func applyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string
 	return out
 }
 
+func payloadRuleMatchesModels(rule *config.PayloadRule, protocol string, models []string) bool {
+	if rule == nil || len(models) == 0 {
+		return false
+	}
+	for _, model := range models {
+		if payloadRuleMatchesModel(rule, model, protocol) {
+			return true
+		}
+	}
+	return false
+}
+
 func payloadRuleMatchesModel(rule *config.PayloadRule, model, protocol string) bool {
 	if rule == nil {
 		return false
@@ -150,6 +163,65 @@ func payloadRuleMatchesModel(rule *config.PayloadRule, model, protocol string) b
 	return false
 }
 
+func payloadModelCandidates(cfg *config.Config, model, protocol string) []string {
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return nil
+	}
+	candidates := []string{model}
+	if cfg == nil {
+		return candidates
+	}
+	aliases := payloadModelAliases(cfg, model, protocol)
+	if len(aliases) == 0 {
+		return candidates
+	}
+	seen := map[string]struct{}{strings.ToLower(model): struct{}{}}
+	for _, alias := range aliases {
+		alias = strings.TrimSpace(alias)
+		if alias == "" {
+			continue
+		}
+		key := strings.ToLower(alias)
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		candidates = append(candidates, alias)
+	}
+	return candidates
+}
+
+func payloadModelAliases(cfg *config.Config, model, protocol string) []string {
+	if cfg == nil {
+		return nil
+	}
+	model = strings.TrimSpace(model)
+	if model == "" {
+		return nil
+	}
+	channel := strings.ToLower(strings.TrimSpace(protocol))
+	if channel == "" {
+		return nil
+	}
+	entries := cfg.OAuthModelAlias[channel]
+	if len(entries) == 0 {
+		return nil
+	}
+	aliases := make([]string, 0, 2)
+	for _, entry := range entries {
+		if !strings.EqualFold(strings.TrimSpace(entry.Name), model) {
+			continue
+		}
+		alias := strings.TrimSpace(entry.Alias)
+		if alias == "" {
+			continue
+		}
+		aliases = append(aliases, alias)
+	}
+	return aliases
+}
+
 // buildPayloadPath combines an optional root path with a relative parameter path.
 // When root is empty, the parameter path is used as-is. When root is non-empty,
 // the parameter path is treated as relative to root.

internal/runtime/executor/qwen_executor.go

@@ -86,7 +86,7 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return resp, err
 	}
@@ -172,7 +172,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, err
 	}

internal/thinking/apply.go

@@ -2,6 +2,8 @@
 package thinking
 
 import (
+	"strings"
+
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -59,7 +61,9 @@ func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
 // Parameters:
 //   - body: Original request body JSON
 //   - model: Model name, optionally with thinking suffix (e.g., "claude-sonnet-4-5(16384)")
-//   - provider: Provider name (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
+//   - fromFormat: Source request format (e.g., openai, codex, gemini)
+//   - toFormat: Target provider format for the request body (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
+//   - providerKey: Provider identifier used for registry model lookups (may differ from toFormat, e.g., openrouter -> openai)
 //
 // Returns:
 //   - Modified request body JSON with thinking configuration applied
@@ -76,16 +80,25 @@ func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
 // Example:
 //
 //	// With suffix - suffix config takes priority
-//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini")
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini", "gemini", "gemini")
 //
 //	// Without suffix - uses body config
-//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini")
-func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini", "gemini", "gemini")
+func ApplyThinking(body []byte, model string, fromFormat string, toFormat string, providerKey string) ([]byte, error) {
+	providerFormat := strings.ToLower(strings.TrimSpace(toFormat))
+	providerKey = strings.ToLower(strings.TrimSpace(providerKey))
+	if providerKey == "" {
+		providerKey = providerFormat
+	}
+	fromFormat = strings.ToLower(strings.TrimSpace(fromFormat))
+	if fromFormat == "" {
+		fromFormat = providerFormat
+	}
 	// 1. Route check: Get provider applier
-	applier := GetProviderApplier(provider)
+	applier := GetProviderApplier(providerFormat)
 	if applier == nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    model,
 		}).Debug("thinking: unknown provider, passthrough |")
 		return body, nil
@@ -94,25 +107,26 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// 2. Parse suffix and get modelInfo
 	suffixResult := ParseSuffix(model)
 	baseModel := suffixResult.ModelName
-	modelInfo := registry.LookupModelInfo(baseModel)
+	// Use provider-specific lookup to handle capability differences across providers.
+	modelInfo := registry.LookupModelInfo(baseModel, providerKey)
 
 	// 3. Model capability check
 	// Unknown models are treated as user-defined so thinking config can still be applied.
 	// The upstream service is responsible for validating the configuration.
 	if IsUserDefinedModel(modelInfo) {
-		return applyUserDefinedModel(body, modelInfo, provider, suffixResult)
+		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult)
 	}
 	if modelInfo.Thinking == nil {
-		config := extractThinkingConfig(body, provider)
+		config := extractThinkingConfig(body, providerFormat)
 		if hasThinkingConfig(config) {
 			log.WithFields(log.Fields{
 				"model":    baseModel,
-				"provider": provider,
+				"provider": providerFormat,
 			}).Debug("thinking: model does not support thinking, stripping config |")
-			return StripThinkingConfig(body, provider), nil
+			return StripThinkingConfig(body, providerFormat), nil
 		}
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    baseModel,
 		}).Debug("thinking: model does not support thinking, passthrough |")
 		return body, nil
@@ -121,19 +135,19 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// 4. Get config: suffix priority over body
 	var config ThinkingConfig
 	if suffixResult.HasSuffix {
-		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, model)
+		config = parseSuffixToConfig(suffixResult.RawSuffix, providerFormat, model)
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    model,
 			"mode":     config.Mode,
 			"budget":   config.Budget,
 			"level":    config.Level,
 		}).Debug("thinking: config from model suffix |")
 	} else {
-		config = extractThinkingConfig(body, provider)
+		config = extractThinkingConfig(body, providerFormat)
 		if hasThinkingConfig(config) {
 			log.WithFields(log.Fields{
-				"provider": provider,
+				"provider": providerFormat,
 				"model":    modelInfo.ID,
 				"mode":     config.Mode,
 				"budget":   config.Budget,
@@ -144,17 +158,17 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 
 	if !hasThinkingConfig(config) {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Debug("thinking: no config found, passthrough |")
 		return body, nil
 	}
 
 	// 5. Validate and normalize configuration
-	validated, err := ValidateConfig(config, modelInfo, provider)
+	validated, err := ValidateConfig(config, modelInfo, fromFormat, providerFormat, suffixResult.HasSuffix)
 	if err != nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 			"error":    err.Error(),
 		}).Warn("thinking: validation failed |")
@@ -167,14 +181,14 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// Defensive check: ValidateConfig should never return (nil, nil)
 	if validated == nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Warn("thinking: ValidateConfig returned nil config without error, passthrough |")
 		return body, nil
 	}
 
 	log.WithFields(log.Fields{
-		"provider": provider,
+		"provider": providerFormat,
 		"model":    modelInfo.ID,
 		"mode":     validated.Mode,
 		"budget":   validated.Budget,
@@ -228,7 +242,7 @@ func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
 
 // applyUserDefinedModel applies thinking configuration for user-defined models
 // without ThinkingSupport validation.
-func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, provider string, suffixResult SuffixResult) ([]byte, error) {
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult) ([]byte, error) {
 	// Get model ID for logging
 	modelID := ""
 	if modelInfo != nil {
@@ -240,39 +254,57 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, provider
 	// Get config: suffix priority over body
 	var config ThinkingConfig
 	if suffixResult.HasSuffix {
-		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, modelID)
+		config = parseSuffixToConfig(suffixResult.RawSuffix, toFormat, modelID)
 	} else {
-		config = extractThinkingConfig(body, provider)
+		config = extractThinkingConfig(body, toFormat)
 	}
 
 	if !hasThinkingConfig(config) {
 		log.WithFields(log.Fields{
 			"model":    modelID,
-			"provider": provider,
+			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (no config) |")
 		return body, nil
 	}
 
-	applier := GetProviderApplier(provider)
+	applier := GetProviderApplier(toFormat)
 	if applier == nil {
 		log.WithFields(log.Fields{
 			"model":    modelID,
-			"provider": provider,
+			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (unknown provider) |")
 		return body, nil
 	}
 
 	log.WithFields(log.Fields{
-		"provider": provider,
+		"provider": toFormat,
 		"model":    modelID,
 		"mode":     config.Mode,
 		"budget":   config.Budget,
 		"level":    config.Level,
 	}).Debug("thinking: applying config for user-defined model (skip validation)")
 
+	config = normalizeUserDefinedConfig(config, fromFormat, toFormat)
 	return applier.Apply(body, config, modelInfo)
 }
 
+func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat string) ThinkingConfig {
+	if config.Mode != ModeLevel {
+		return config
+	}
+	if !isBudgetBasedProvider(toFormat) || !isLevelBasedProvider(fromFormat) {
+		return config
+	}
+	budget, ok := ConvertLevelToBudget(string(config.Level))
+	if !ok {
+		return config
+	}
+	config.Mode = ModeBudget
+	config.Budget = budget
+	config.Level = ""
+	return config
+}
+
 // extractThinkingConfig extracts provider-specific thinking config from request body.
 func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
 	if len(body) == 0 || !gjson.ValidBytes(body) {
@@ -289,7 +321,11 @@ func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
 	case "codex":
 		return extractCodexConfig(body)
 	case "iflow":
-		return extractIFlowConfig(body)
+		config := extractIFlowConfig(body)
+		if hasThinkingConfig(config) {
+			return config
+		}
+		return extractOpenAIConfig(body)
 	default:
 		return ThinkingConfig{}
 	}

internal/thinking/errors.go

@@ -24,6 +24,10 @@ const (
 	// Example: using level with a budget-only model
 	ErrLevelNotSupported ErrorCode = "LEVEL_NOT_SUPPORTED"
 
+	// ErrBudgetOutOfRange indicates the budget value is outside model range.
+	// Example: budget 64000 exceeds max 20000
+	ErrBudgetOutOfRange ErrorCode = "BUDGET_OUT_OF_RANGE"
+
 	// ErrProviderMismatch indicates the provider does not match the model.
 	// Example: applying Claude format to a Gemini model
 	ErrProviderMismatch ErrorCode = "PROVIDER_MISMATCH"

internal/thinking/provider/claude/apply.go

@@ -80,9 +80,66 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 
 	result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 	result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+
+	// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint)
+	result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
 	return result, nil
 }
 
+// normalizeClaudeBudget applies Claude-specific constraints to ensure max_tokens > budget_tokens.
+// Anthropic API requires this constraint; violating it returns a 400 error.
+func (a *Applier) normalizeClaudeBudget(body []byte, budgetTokens int, modelInfo *registry.ModelInfo) []byte {
+	if budgetTokens <= 0 {
+		return body
+	}
+
+	// Ensure the request satisfies Claude constraints:
+	//  1) Determine effective max_tokens (request overrides model default)
+	//  2) If budget_tokens >= max_tokens, reduce budget_tokens to max_tokens-1
+	//  3) If the adjusted budget falls below the model minimum, leave the request unchanged
+	//  4) If max_tokens came from model default, write it back into the request
+
+	effectiveMax, setDefaultMax := a.effectiveMaxTokens(body, modelInfo)
+	if setDefaultMax && effectiveMax > 0 {
+		body, _ = sjson.SetBytes(body, "max_tokens", effectiveMax)
+	}
+
+	// Compute the budget we would apply after enforcing budget_tokens < max_tokens.
+	adjustedBudget := budgetTokens
+	if effectiveMax > 0 && adjustedBudget >= effectiveMax {
+		adjustedBudget = effectiveMax - 1
+	}
+
+	minBudget := 0
+	if modelInfo != nil && modelInfo.Thinking != nil {
+		minBudget = modelInfo.Thinking.Min
+	}
+	if minBudget > 0 && adjustedBudget > 0 && adjustedBudget < minBudget {
+		// If enforcing the max_tokens constraint would push the budget below the model minimum,
+		// leave the request unchanged.
+		return body
+	}
+
+	if adjustedBudget != budgetTokens {
+		body, _ = sjson.SetBytes(body, "thinking.budget_tokens", adjustedBudget)
+	}
+
+	return body
+}
+
+// effectiveMaxTokens returns the max tokens to cap thinking:
+// prefer request-provided max_tokens; otherwise fall back to model default.
+// The boolean indicates whether the value came from the model default (and thus should be written back).
+func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo) (max int, fromModel bool) {
+	if maxTok := gjson.GetBytes(body, "max_tokens"); maxTok.Exists() && maxTok.Int() > 0 {
+		return int(maxTok.Int()), false
+	}
+	if modelInfo != nil && modelInfo.MaxCompletionTokens > 0 {
+		return modelInfo.MaxCompletionTokens, true
+	}
+	return 0, false
+}
+
 func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
 	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
 		return body, nil

internal/thinking/strip.go

@@ -27,28 +27,32 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 		return body
 	}
 
+	var paths []string
 	switch provider {
 	case "claude":
-		result, _ := sjson.DeleteBytes(body, "thinking")
-		return result
+		paths = []string{"thinking"}
 	case "gemini":
-		result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig")
-		return result
+		paths = []string{"generationConfig.thinkingConfig"}
 	case "gemini-cli", "antigravity":
-		result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig")
-		return result
+		paths = []string{"request.generationConfig.thinkingConfig"}
 	case "openai":
-		result, _ := sjson.DeleteBytes(body, "reasoning_effort")
-		return result
+		paths = []string{"reasoning_effort"}
 	case "codex":
-		result, _ := sjson.DeleteBytes(body, "reasoning.effort")
-		return result
+		paths = []string{"reasoning.effort"}
 	case "iflow":
-		result, _ := sjson.DeleteBytes(body, "chat_template_kwargs.enable_thinking")
-		result, _ = sjson.DeleteBytes(result, "chat_template_kwargs.clear_thinking")
-		result, _ = sjson.DeleteBytes(result, "reasoning_split")
-		return result
+		paths = []string{
+			"chat_template_kwargs.enable_thinking",
+			"chat_template_kwargs.clear_thinking",
+			"reasoning_split",
+			"reasoning_effort",
+		}
 	default:
 		return body
 	}
+
+	result := body
+	for _, path := range paths {
+		result, _ = sjson.DeleteBytes(result, path)
+	}
+	return result
 }

internal/thinking/validate.go

@@ -9,64 +9,6 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
-// ClampBudget clamps a budget value to the model's supported range.
-//
-// Logging:
-//   - Warn when value=0 but ZeroAllowed=false
-//   - Debug when value is clamped to min/max
-//
-// Fields: provider, model, original_value, clamped_to, min, max
-func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
-	model := "unknown"
-	support := (*registry.ThinkingSupport)(nil)
-	if modelInfo != nil {
-		if modelInfo.ID != "" {
-			model = modelInfo.ID
-		}
-		support = modelInfo.Thinking
-	}
-	if support == nil {
-		return value
-	}
-
-	// Auto value (-1) passes through without clamping.
-	if value == -1 {
-		return value
-	}
-
-	min := support.Min
-	max := support.Max
-	if value == 0 && !support.ZeroAllowed {
-		log.WithFields(log.Fields{
-			"provider":       provider,
-			"model":          model,
-			"original_value": value,
-			"clamped_to":     min,
-			"min":            min,
-			"max":            max,
-		}).Warn("thinking: budget zero not allowed |")
-		return min
-	}
-
-	// Some models are level-only and do not define numeric budget ranges.
-	if min == 0 && max == 0 {
-		return value
-	}
-
-	if value < min {
-		if value == 0 && support.ZeroAllowed {
-			return 0
-		}
-		logClamp(provider, model, value, min, min, max)
-		return min
-	}
-	if value > max {
-		logClamp(provider, model, value, max, min, max)
-		return max
-	}
-	return value
-}
-
 // ValidateConfig validates a thinking configuration against model capabilities.
 //
 // This function performs comprehensive validation:
@@ -74,10 +16,16 @@ func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int
 //   - Auto-converts between Budget and Level formats based on model capability
 //   - Validates that requested level is in the model's supported levels list
 //   - Clamps budget values to model's allowed range
+//   - When converting Budget -> Level for level-only models, clamps the derived standard level to the nearest supported level
+//     (special values none/auto are preserved)
+//   - When config comes from a model suffix, strict budget validation is disabled (we clamp instead of error)
 //
 // Parameters:
 //   - config: The thinking configuration to validate
 //   - support: Model's ThinkingSupport properties (nil means no thinking support)
+//   - fromFormat: Source provider format (used to determine strict validation rules)
+//   - toFormat: Target provider format
+//   - fromSuffix: Whether config was sourced from model suffix
 //
 // Returns:
 //   - Normalized ThinkingConfig with clamped values
@@ -87,9 +35,8 @@ func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int
 //   - Budget-only model + Level config → Level converted to Budget
 //   - Level-only model + Budget config → Budget converted to Level
 //   - Hybrid model → preserve original format
-func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, provider string) (*ThinkingConfig, error) {
-	normalized := config
-
+func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, fromFormat, toFormat string, fromSuffix bool) (*ThinkingConfig, error) {
+	fromFormat, toFormat = strings.ToLower(strings.TrimSpace(fromFormat)), strings.ToLower(strings.TrimSpace(toFormat))
 	model := "unknown"
 	support := (*registry.ThinkingSupport)(nil)
 	if modelInfo != nil {
@@ -103,101 +50,108 @@ func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, provid
 		if config.Mode != ModeNone {
 			return nil, NewThinkingErrorWithModel(ErrThinkingNotSupported, "thinking not supported for this model", model)
 		}
-		return &normalized, nil
+		return &config, nil
 	}
 
+	allowClampUnsupported := isBudgetBasedProvider(fromFormat) && isLevelBasedProvider(toFormat)
+	strictBudget := !fromSuffix && fromFormat != "" && isSameProviderFamily(fromFormat, toFormat)
+	budgetDerivedFromLevel := false
+
 	capability := detectModelCapability(modelInfo)
 	switch capability {
 	case CapabilityBudgetOnly:
-		if normalized.Mode == ModeLevel {
-			if normalized.Level == LevelAuto {
+		if config.Mode == ModeLevel {
+			if config.Level == LevelAuto {
 				break
 			}
-			budget, ok := ConvertLevelToBudget(string(normalized.Level))
+			budget, ok := ConvertLevelToBudget(string(config.Level))
 			if !ok {
-				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", normalized.Level))
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", config.Level))
 			}
-			normalized.Mode = ModeBudget
-			normalized.Budget = budget
-			normalized.Level = ""
+			config.Mode = ModeBudget
+			config.Budget = budget
+			config.Level = ""
+			budgetDerivedFromLevel = true
 		}
 	case CapabilityLevelOnly:
-		if normalized.Mode == ModeBudget {
-			level, ok := ConvertBudgetToLevel(normalized.Budget)
+		if config.Mode == ModeBudget {
+			level, ok := ConvertBudgetToLevel(config.Budget)
 			if !ok {
-				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", normalized.Budget))
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", config.Budget))
 			}
-			normalized.Mode = ModeLevel
-			normalized.Level = ThinkingLevel(level)
-			normalized.Budget = 0
+			// When converting Budget -> Level for level-only models, clamp the derived standard level
+			// to the nearest supported level. Special values (none/auto) are preserved.
+			config.Mode = ModeLevel
+			config.Level = clampLevel(ThinkingLevel(level), modelInfo, toFormat)
+			config.Budget = 0
 		}
 	case CapabilityHybrid:
 	}
 
-	if normalized.Mode == ModeLevel && normalized.Level == LevelNone {
-		normalized.Mode = ModeNone
-		normalized.Budget = 0
-		normalized.Level = ""
+	if config.Mode == ModeLevel && config.Level == LevelNone {
+		config.Mode = ModeNone
+		config.Budget = 0
+		config.Level = ""
 	}
-	if normalized.Mode == ModeLevel && normalized.Level == LevelAuto {
-		normalized.Mode = ModeAuto
-		normalized.Budget = -1
-		normalized.Level = ""
+	if config.Mode == ModeLevel && config.Level == LevelAuto {
+		config.Mode = ModeAuto
+		config.Budget = -1
+		config.Level = ""
 	}
-	if normalized.Mode == ModeBudget && normalized.Budget == 0 {
-		normalized.Mode = ModeNone
-		normalized.Level = ""
+	if config.Mode == ModeBudget && config.Budget == 0 {
+		config.Mode = ModeNone
+		config.Level = ""
 	}
 
-	if len(support.Levels) > 0 && normalized.Mode == ModeLevel {
-		if !isLevelSupported(string(normalized.Level), support.Levels) {
-			validLevels := normalizeLevels(support.Levels)
-			message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(normalized.Level)), strings.Join(validLevels, ", "))
-			return nil, NewThinkingError(ErrLevelNotSupported, message)
+	if len(support.Levels) > 0 && config.Mode == ModeLevel {
+		if !isLevelSupported(string(config.Level), support.Levels) {
+			if allowClampUnsupported {
+				config.Level = clampLevel(config.Level, modelInfo, toFormat)
+			}
+			if !isLevelSupported(string(config.Level), support.Levels) {
+				// User explicitly specified an unsupported level - return error
+				// (budget-derived levels may be clamped based on source format)
+				validLevels := normalizeLevels(support.Levels)
+				message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(config.Level)), strings.Join(validLevels, ", "))
+				return nil, NewThinkingError(ErrLevelNotSupported, message)
+			}
+		}
+	}
+
+	if strictBudget && config.Mode == ModeBudget && !budgetDerivedFromLevel {
+		min, max := support.Min, support.Max
+		if min != 0 || max != 0 {
+			if config.Budget < min || config.Budget > max || (config.Budget == 0 && !support.ZeroAllowed) {
+				message := fmt.Sprintf("budget %d out of range [%d,%d]", config.Budget, min, max)
+				return nil, NewThinkingError(ErrBudgetOutOfRange, message)
+			}
 		}
 	}
 
 	// Convert ModeAuto to mid-range if dynamic not allowed
-	if normalized.Mode == ModeAuto && !support.DynamicAllowed {
-		normalized = convertAutoToMidRange(normalized, support, provider, model)
+	if config.Mode == ModeAuto && !support.DynamicAllowed {
+		config = convertAutoToMidRange(config, support, toFormat, model)
 	}
 
-	if normalized.Mode == ModeNone && provider == "claude" {
+	if config.Mode == ModeNone && toFormat == "claude" {
 		// Claude supports explicit disable via thinking.type="disabled".
 		// Keep Budget=0 so applier can omit budget_tokens.
-		normalized.Budget = 0
-		normalized.Level = ""
+		config.Budget = 0
+		config.Level = ""
 	} else {
-		switch normalized.Mode {
+		switch config.Mode {
 		case ModeBudget, ModeAuto, ModeNone:
-			normalized.Budget = ClampBudget(normalized.Budget, modelInfo, provider)
+			config.Budget = clampBudget(config.Budget, modelInfo, toFormat)
 		}
 
 		// ModeNone with clamped Budget > 0: set Level to lowest for Level-only/Hybrid models
 		// This ensures Apply layer doesn't need to access support.Levels
-		if normalized.Mode == ModeNone && normalized.Budget > 0 && len(support.Levels) > 0 {
-			normalized.Level = ThinkingLevel(support.Levels[0])
+		if config.Mode == ModeNone && config.Budget > 0 && len(support.Levels) > 0 {
+			config.Level = ThinkingLevel(support.Levels[0])
 		}
 	}
 
-	return &normalized, nil
-}
-
-func isLevelSupported(level string, supported []string) bool {
-	for _, candidate := range supported {
-		if strings.EqualFold(level, strings.TrimSpace(candidate)) {
-			return true
-		}
-	}
-	return false
-}
-
-func normalizeLevels(levels []string) []string {
-	normalized := make([]string, 0, len(levels))
-	for _, level := range levels {
-		normalized = append(normalized, strings.ToLower(strings.TrimSpace(level)))
-	}
-	return normalized
+	return &config, nil
 }
 
 // convertAutoToMidRange converts ModeAuto to a mid-range value when dynamic is not allowed.
@@ -246,7 +200,172 @@ func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupp
 	return config
 }
 
-// logClamp logs a debug message when budget clamping occurs.
+// standardLevelOrder defines the canonical ordering of thinking levels from lowest to highest.
+var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh}
+
+// clampLevel clamps the given level to the nearest supported level.
+// On tie, prefers the lower level.
+func clampLevel(level ThinkingLevel, modelInfo *registry.ModelInfo, provider string) ThinkingLevel {
+	model := "unknown"
+	var supported []string
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		if modelInfo.Thinking != nil {
+			supported = modelInfo.Thinking.Levels
+		}
+	}
+
+	if len(supported) == 0 || isLevelSupported(string(level), supported) {
+		return level
+	}
+
+	pos := levelIndex(string(level))
+	if pos == -1 {
+		return level
+	}
+	bestIdx, bestDist := -1, len(standardLevelOrder)+1
+
+	for _, s := range supported {
+		if idx := levelIndex(strings.TrimSpace(s)); idx != -1 {
+			if dist := abs(pos - idx); dist < bestDist || (dist == bestDist && idx < bestIdx) {
+				bestIdx, bestDist = idx, dist
+			}
+		}
+	}
+
+	if bestIdx >= 0 {
+		clamped := standardLevelOrder[bestIdx]
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": string(level),
+			"clamped_to":     string(clamped),
+		}).Debug("thinking: level clamped |")
+		return clamped
+	}
+	return level
+}
+
+// clampBudget clamps a budget value to the model's supported range.
+func clampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+	if support == nil {
+		return value
+	}
+
+	// Auto value (-1) passes through without clamping.
+	if value == -1 {
+		return value
+	}
+
+	min, max := support.Min, support.Max
+	if value == 0 && !support.ZeroAllowed {
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": value,
+			"clamped_to":     min,
+			"min":            min,
+			"max":            max,
+		}).Warn("thinking: budget zero not allowed |")
+		return min
+	}
+
+	// Some models are level-only and do not define numeric budget ranges.
+	if min == 0 && max == 0 {
+		return value
+	}
+
+	if value < min {
+		if value == 0 && support.ZeroAllowed {
+			return 0
+		}
+		logClamp(provider, model, value, min, min, max)
+		return min
+	}
+	if value > max {
+		logClamp(provider, model, value, max, min, max)
+		return max
+	}
+	return value
+}
+
+func isLevelSupported(level string, supported []string) bool {
+	for _, s := range supported {
+		if strings.EqualFold(level, strings.TrimSpace(s)) {
+			return true
+		}
+	}
+	return false
+}
+
+func levelIndex(level string) int {
+	for i, l := range standardLevelOrder {
+		if strings.EqualFold(level, string(l)) {
+			return i
+		}
+	}
+	return -1
+}
+
+func normalizeLevels(levels []string) []string {
+	out := make([]string, len(levels))
+	for i, l := range levels {
+		out[i] = strings.ToLower(strings.TrimSpace(l))
+	}
+	return out
+}
+
+func isBudgetBasedProvider(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity", "claude":
+		return true
+	default:
+		return false
+	}
+}
+
+func isLevelBasedProvider(provider string) bool {
+	switch provider {
+	case "openai", "openai-response", "codex":
+		return true
+	default:
+		return false
+	}
+}
+
+func isGeminiFamily(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity":
+		return true
+	default:
+		return false
+	}
+}
+
+func isSameProviderFamily(from, to string) bool {
+	if from == to {
+		return true
+	}
+	return isGeminiFamily(from) && isGeminiFamily(to)
+}
+
+func abs(x int) int {
+	if x < 0 {
+		return -x
+	}
+	return x
+}
+
 func logClamp(provider, model string, original, clampedTo, min, max int) {
 	log.WithFields(log.Fields{
 		"provider":       provider,

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -12,7 +12,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
-	"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/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
@@ -23,6 +22,14 @@ import (
 // deriveSessionID generates a stable session ID from the request.
 // Uses the hash of the first user message to identify the conversation.
 func deriveSessionID(rawJSON []byte) string {
+	userIDResult := gjson.GetBytes(rawJSON, "metadata.user_id")
+	if userIDResult.Exists() {
+		userID := userIDResult.String()
+		idx := strings.Index(userID, "session_")
+		if idx != -1 {
+			return userID[idx+8:]
+		}
+	}
 	messages := gjson.GetBytes(rawJSON, "messages")
 	if !messages.IsArray() {
 		return ""
@@ -62,6 +69,7 @@ func deriveSessionID(rawJSON []byte) string {
 // Returns:
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
+	enableThoughtTranslate := true
 	rawJSON := bytes.Clone(inputRawJSON)
 
 	// Derive session ID for signature caching
@@ -125,25 +133,32 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					if contentTypeResult.Type == gjson.String && contentTypeResult.String() == "thinking" {
 						// Use GetThinkingText to handle wrapped thinking objects
 						thinkingText := thinking.GetThinkingText(contentResult)
-						signatureResult := contentResult.Get("signature")
-						clientSignature := ""
-						if signatureResult.Exists() && signatureResult.String() != "" {
-							clientSignature = signatureResult.String()
-						}
 
 						// Always try cached signature first (more reliable than client-provided)
 						// Client may send stale or invalid signatures from different sessions
 						signature := ""
 						if sessionID != "" && thinkingText != "" {
-							if cachedSig := cache.GetCachedSignature(sessionID, thinkingText); cachedSig != "" {
+							if cachedSig := cache.GetCachedSignature(modelName, sessionID, thinkingText); cachedSig != "" {
 								signature = cachedSig
 								// log.Debugf("Using cached signature for thinking block")
 							}
 						}
 
 						// Fallback to client signature only if cache miss and client signature is valid
-						if signature == "" && cache.HasValidSignature(clientSignature) {
-							signature = clientSignature
+						if signature == "" {
+							signatureResult := contentResult.Get("signature")
+							clientSignature := ""
+							if signatureResult.Exists() && signatureResult.String() != "" {
+								arrayClientSignatures := strings.SplitN(signatureResult.String(), "#", 2)
+								if len(arrayClientSignatures) == 2 {
+									if modelName == arrayClientSignatures[0] {
+										clientSignature = arrayClientSignatures[1]
+									}
+								}
+							}
+							if cache.HasValidSignature(clientSignature) {
+								signature = clientSignature
+							}
 							// log.Debugf("Using client-provided signature for thinking block")
 						}
 
@@ -160,6 +175,7 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 						// Converting to text would break this requirement
 						if isUnsigned {
 							// log.Debugf("Dropping unsigned thinking block (no valid signature)")
+							enableThoughtTranslate = false
 							continue
 						}
 
@@ -387,15 +403,12 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
-		if modelInfo != nil && modelInfo.Thinking != nil {
-			if t.Get("type").String() == "enabled" {
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-					budget := int(b.Int())
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-				}
+	if t := gjson.GetBytes(rawJSON, "thinking"); enableThoughtTranslate && t.Exists() && t.IsObject() {
+		if t.Get("type").String() == "enabled" {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				budget := int(b.Int())
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -4,6 +4,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
 	"github.com/tidwall/gjson"
 )
 
@@ -75,28 +76,42 @@ func TestConvertClaudeRequestToAntigravity_RoleMapping(t *testing.T) {
 func TestConvertClaudeRequestToAntigravity_ThinkingBlocks(t *testing.T) {
 	// Valid signature must be at least 50 characters
 	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	thinkingText := "Let me think..."
+
+	// Pre-cache the signature (simulating a response from the same session)
+	// The session ID is derived from the first user message hash
+	// Since there's no user message in this test, we need to add one
 	inputJSON := []byte(`{
 		"model": "claude-sonnet-4-5-thinking",
 		"messages": [
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Test user message"}]
+			},
 			{
 				"role": "assistant",
 				"content": [
-					{"type": "thinking", "thinking": "Let me think...", "signature": "` + validSignature + `"},
+					{"type": "thinking", "thinking": "` + thinkingText + `", "signature": "` + validSignature + `"},
 					{"type": "text", "text": "Answer"}
 				]
 			}
 		]
 	}`)
 
+	// Derive session ID and cache the signature
+	sessionID := deriveSessionID(inputJSON)
+	cache.CacheSignature(sessionID, thinkingText, validSignature)
+	defer cache.ClearSignatureCache(sessionID)
+
 	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
 	outputStr := string(output)
 
-	// Check thinking block conversion
-	firstPart := gjson.Get(outputStr, "request.contents.0.parts.0")
+	// Check thinking block conversion (now in contents.1 due to user message)
+	firstPart := gjson.Get(outputStr, "request.contents.1.parts.0")
 	if !firstPart.Get("thought").Bool() {
 		t.Error("thinking block should have thought: true")
 	}
-	if firstPart.Get("text").String() != "Let me think..." {
+	if firstPart.Get("text").String() != thinkingText {
 		t.Error("thinking text mismatch")
 	}
 	if firstPart.Get("thoughtSignature").String() != validSignature {
@@ -227,13 +242,19 @@ func TestConvertClaudeRequestToAntigravity_ToolUse(t *testing.T) {
 
 func TestConvertClaudeRequestToAntigravity_ToolUse_WithSignature(t *testing.T) {
 	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	thinkingText := "Let me think..."
+
 	inputJSON := []byte(`{
 		"model": "claude-sonnet-4-5-thinking",
 		"messages": [
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Test user message"}]
+			},
 			{
 				"role": "assistant",
 				"content": [
-					{"type": "thinking", "thinking": "Let me think...", "signature": "` + validSignature + `"},
+					{"type": "thinking", "thinking": "` + thinkingText + `", "signature": "` + validSignature + `"},
 					{
 						"type": "tool_use",
 						"id": "call_123",
@@ -245,11 +266,16 @@ func TestConvertClaudeRequestToAntigravity_ToolUse_WithSignature(t *testing.T) {
 		]
 	}`)
 
+	// Derive session ID and cache the signature
+	sessionID := deriveSessionID(inputJSON)
+	cache.CacheSignature(sessionID, thinkingText, validSignature)
+	defer cache.ClearSignatureCache(sessionID)
+
 	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
 	outputStr := string(output)
 
-	// Check function call has the signature from the preceding thinking block
-	part := gjson.Get(outputStr, "request.contents.0.parts.1")
+	// Check function call has the signature from the preceding thinking block (now in contents.1)
+	part := gjson.Get(outputStr, "request.contents.1.parts.1")
 	if part.Get("functionCall.name").String() != "get_weather" {
 		t.Errorf("Expected functionCall, got %s", part.Raw)
 	}
@@ -261,24 +287,35 @@ func TestConvertClaudeRequestToAntigravity_ToolUse_WithSignature(t *testing.T) {
 func TestConvertClaudeRequestToAntigravity_ReorderThinking(t *testing.T) {
 	// Case: text block followed by thinking block -> should be reordered to thinking first
 	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	thinkingText := "Planning..."
+
 	inputJSON := []byte(`{
 		"model": "claude-sonnet-4-5-thinking",
 		"messages": [
+			{
+				"role": "user",
+				"content": [{"type": "text", "text": "Test user message"}]
+			},
 			{
 				"role": "assistant",
 				"content": [
 					{"type": "text", "text": "Here is the plan."},
-					{"type": "thinking", "thinking": "Planning...", "signature": "` + validSignature + `"}
+					{"type": "thinking", "thinking": "` + thinkingText + `", "signature": "` + validSignature + `"}
 				]
 			}
 		]
 	}`)
 
+	// Derive session ID and cache the signature
+	sessionID := deriveSessionID(inputJSON)
+	cache.CacheSignature(sessionID, thinkingText, validSignature)
+	defer cache.ClearSignatureCache(sessionID)
+
 	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
 	outputStr := string(output)
 
-	// Verify order: Thinking block MUST be first
-	parts := gjson.Get(outputStr, "request.contents.0.parts").Array()
+	// Verify order: Thinking block MUST be first (now in contents.1 due to user message)
+	parts := gjson.Get(outputStr, "request.contents.1.parts").Array()
 	if len(parts) != 2 {
 		t.Fatalf("Expected 2 parts, got %d", len(parts))
 	}
@@ -343,8 +380,8 @@ func TestConvertClaudeRequestToAntigravity_ThinkingConfig(t *testing.T) {
 		if thinkingConfig.Get("thinkingBudget").Int() != 8000 {
 			t.Errorf("Expected thinkingBudget 8000, got %d", thinkingConfig.Get("thinkingBudget").Int())
 		}
-		if !thinkingConfig.Get("include_thoughts").Bool() {
-			t.Error("include_thoughts should be true")
+		if !thinkingConfig.Get("includeThoughts").Bool() {
+			t.Error("includeThoughts should be true")
 		}
 	} else {
 		t.Log("thinkingConfig not present - model may not be registered in test registry")
@@ -460,6 +497,9 @@ func TestConvertClaudeRequestToAntigravity_TrailingUnsignedThinking_Removed(t *t
 
 func TestConvertClaudeRequestToAntigravity_TrailingSignedThinking_Kept(t *testing.T) {
 	// Last assistant message ends with signed thinking block - should be kept
+	validSignature := "abc123validSignature1234567890123456789012345678901234567890"
+	thinkingText := "Valid thinking..."
+
 	inputJSON := []byte(`{
 		"model": "claude-sonnet-4-5-thinking",
 		"messages": [
@@ -471,12 +511,17 @@ func TestConvertClaudeRequestToAntigravity_TrailingSignedThinking_Kept(t *testin
 				"role": "assistant",
 				"content": [
 					{"type": "text", "text": "Here is my answer"},
-					{"type": "thinking", "thinking": "Valid thinking...", "signature": "abc123validSignature1234567890123456789012345678901234567890"}
+					{"type": "thinking", "thinking": "` + thinkingText + `", "signature": "` + validSignature + `"}
 				]
 			}
 		]
 	}`)
 
+	// Derive session ID and cache the signature
+	sessionID := deriveSessionID(inputJSON)
+	cache.CacheSignature(sessionID, thinkingText, validSignature)
+	defer cache.ClearSignatureCache(sessionID)
+
 	output := ConvertClaudeRequestToAntigravity("claude-sonnet-4-5-thinking", inputJSON, false)
 	outputStr := string(output)
 

internal/translator/antigravity/claude/antigravity_claude_response.go

@@ -73,6 +73,7 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 			SessionID:        deriveSessionID(originalRequestRawJSON),
 		}
 	}
+	modelName := gjson.GetBytes(requestRawJSON, "model").String()
 
 	params := (*param).(*Params)
 
@@ -139,13 +140,13 @@ func ConvertAntigravityResponseToClaude(_ context.Context, _ string, originalReq
 						// log.Debug("Branch: signature_delta")
 
 						if params.SessionID != "" && params.CurrentThinkingText.Len() > 0 {
-							cache.CacheSignature(params.SessionID, params.CurrentThinkingText.String(), thoughtSignature.String())
+							cache.CacheSignature(modelName, params.SessionID, params.CurrentThinkingText.String(), thoughtSignature.String())
 							// log.Debugf("Cached signature for thinking block (sessionID=%s, textLen=%d)", params.SessionID, params.CurrentThinkingText.Len())
 							params.CurrentThinkingText.Reset()
 						}
 
 						output = output + "event: content_block_delta\n"
-						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":""}}`, params.ResponseIndex), "delta.signature", thoughtSignature.String())
+						data, _ := sjson.Set(fmt.Sprintf(`{"type":"content_block_delta","index":%d,"delta":{"type":"signature_delta","signature":""}}`, params.ResponseIndex), "delta.signature", fmt.Sprintf("%s#%s", modelName, thoughtSignature.String()))
 						output = output + fmt.Sprintf("data: %s\n\n\n", data)
 						params.HasContent = true
 					} else if params.ResponseType == 2 { // Continue existing thinking block if already in thinking state
@@ -372,7 +373,7 @@ func resolveStopReason(params *Params) string {
 //   - string: A Claude-compatible JSON response.
 func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
 	_ = originalRequestRawJSON
-	_ = requestRawJSON
+	modelName := gjson.GetBytes(requestRawJSON, "model").String()
 
 	root := gjson.ParseBytes(rawJSON)
 	promptTokens := root.Get("response.usageMetadata.promptTokenCount").Int()
@@ -437,7 +438,7 @@ func ConvertAntigravityResponseToClaudeNonStream(_ context.Context, _ string, or
 		block := `{"type":"thinking","thinking":""}`
 		block, _ = sjson.Set(block, "thinking", thinkingBuilder.String())
 		if thinkingSignature != "" {
-			block, _ = sjson.Set(block, "signature", thinkingSignature)
+			block, _ = sjson.Set(block, "signature", fmt.Sprintf("%s#%s", modelName, thinkingSignature))
 		}
 		responseJSON, _ = sjson.SetRaw(responseJSON, "content.-1", block)
 		thinkingBuilder.Reset()

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

@@ -66,6 +66,13 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 		out, _ = sjson.SetBytes(out, "request.generationConfig.maxOutputTokens", maxTok.Num)
 	}
 
+	// Candidate count (OpenAI 'n' parameter)
+	if n := gjson.GetBytes(rawJSON, "n"); n.Exists() && n.Type == gjson.Number {
+		if val := n.Int(); val > 1 {
+			out, _ = sjson.SetBytes(out, "request.generationConfig.candidateCount", val)
+		}
+	}
+
 	// Map OpenAI modalities -> Gemini CLI request.generationConfig.responseModalities
 	// e.g. "modalities": ["image", "text"] -> ["IMAGE", "TEXT"]
 	if mods := gjson.GetBytes(rawJSON, "modalities"); mods.Exists() && mods.IsArray() {
@@ -132,6 +139,7 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
@@ -141,16 +149,19 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				// system -> request.systemInstruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.Get("text").String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} else if content.IsArray() {
 					contents := content.Array()
 					if len(contents) > 0 {
 						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
 						for j := 0; j < len(contents); j++ {
-							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}
@@ -165,7 +176,10 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					for _, item := range items {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							imageURL := item.Get("image_url.url").String()
@@ -209,7 +223,10 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 					for _, item := range content.Array() {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							// If the assistant returned an inline data URL, preserve it for history fidelity.

internal/translator/claude/gemini/claude_gemini_request.go

@@ -15,7 +15,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"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/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -115,18 +115,41 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			}
 		}
 		// Include thoughts configuration for reasoning process visibility
-		// Only apply for models that support thinking and use numeric budgets, not discrete levels.
+		// Translator only does format conversion, ApplyThinking handles model capability validation.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			modelInfo := registry.LookupModelInfo(modelName)
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
-				// Check for thinkingBudget first - if present, enable thinking with budget
-				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() && thinkingBudget.Int() > 0 {
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					out, _ = sjson.Set(out, "thinking.budget_tokens", thinkingBudget.Int())
-				} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-					// Fallback to include_thoughts if no budget specified
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				level := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				switch level {
+				case "":
+				case "none":
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				case "auto":
 					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					if budget, ok := thinking.ConvertLevelToBudget(level); ok {
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
 				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				budget := int(thinkingBudget.Int())
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+				}
+			} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
+			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}
 	}

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

@@ -15,7 +15,6 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -66,23 +65,21 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	root := gjson.ParseBytes(rawJSON)
 
+	// Convert OpenAI reasoning_effort to Claude thinking config.
 	if v := root.Get("reasoning_effort"); v.Exists() {
-		modelInfo := registry.LookupModelInfo(modelName)
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
-			effort := strings.ToLower(strings.TrimSpace(v.String()))
-			if effort != "" {
-				budget, ok := thinking.ConvertLevelToBudget(effort)
-				if ok {
-					switch budget {
-					case 0:
-						out, _ = sjson.Set(out, "thinking.type", "disabled")
-					case -1:
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
+		if effort != "" {
+			budget, ok := thinking.ConvertLevelToBudget(effort)
+			if ok {
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				default:
+					if budget > 0 {
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-					default:
-						if budget > 0 {
-							out, _ = sjson.Set(out, "thinking.type", "enabled")
-							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						}
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
 					}
 				}
 			}
@@ -141,17 +138,35 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	// Process messages and transform them to Claude Code format
 	if messages := root.Get("messages"); messages.Exists() && messages.IsArray() {
+		messageIndex := 0
+		systemMessageIndex := -1
 		messages.ForEach(func(_, message gjson.Result) bool {
 			role := message.Get("role").String()
 			contentResult := message.Get("content")
 
 			switch role {
-			case "system", "user", "assistant":
-				// Create Claude Code message with appropriate role mapping
-				if role == "system" {
-					role = "user"
+			case "system":
+				if systemMessageIndex == -1 {
+					systemMsg := `{"role":"user","content":[]}`
+					out, _ = sjson.SetRaw(out, "messages.-1", systemMsg)
+					systemMessageIndex = messageIndex
+					messageIndex++
 				}
-
+				if contentResult.Exists() && contentResult.Type == gjson.String && contentResult.String() != "" {
+					textPart := `{"type":"text","text":""}`
+					textPart, _ = sjson.Set(textPart, "text", contentResult.String())
+					out, _ = sjson.SetRaw(out, fmt.Sprintf("messages.%d.content.-1", systemMessageIndex), textPart)
+				} else if contentResult.Exists() && contentResult.IsArray() {
+					contentResult.ForEach(func(_, part gjson.Result) bool {
+						if part.Get("type").String() == "text" {
+							textPart := `{"type":"text","text":""}`
+							textPart, _ = sjson.Set(textPart, "text", part.Get("text").String())
+							out, _ = sjson.SetRaw(out, fmt.Sprintf("messages.%d.content.-1", systemMessageIndex), textPart)
+						}
+						return true
+					})
+				}
+			case "user", "assistant":
 				msg := `{"role":"","content":[]}`
 				msg, _ = sjson.Set(msg, "role", role)
 
@@ -230,6 +245,7 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 				}
 
 				out, _ = sjson.SetRaw(out, "messages.-1", msg)
+				messageIndex++
 
 			case "tool":
 				// Handle tool result messages conversion
@@ -240,6 +256,7 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 				msg, _ = sjson.Set(msg, "content.0.tool_use_id", toolCallID)
 				msg, _ = sjson.Set(msg, "content.0.content", content)
 				out, _ = sjson.SetRaw(out, "messages.-1", msg)
+				messageIndex++
 			}
 			return true
 		})

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -10,7 +10,6 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -54,23 +53,21 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 
 	root := gjson.ParseBytes(rawJSON)
 
+	// Convert OpenAI Responses reasoning.effort to Claude thinking config.
 	if v := root.Get("reasoning.effort"); v.Exists() {
-		modelInfo := registry.LookupModelInfo(modelName)
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
-			effort := strings.ToLower(strings.TrimSpace(v.String()))
-			if effort != "" {
-				budget, ok := thinking.ConvertLevelToBudget(effort)
-				if ok {
-					switch budget {
-					case 0:
-						out, _ = sjson.Set(out, "thinking.type", "disabled")
-					case -1:
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
+		if effort != "" {
+			budget, ok := thinking.ConvertLevelToBudget(effort)
+			if ok {
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				default:
+					if budget > 0 {
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-					default:
-						if budget > 0 {
-							out, _ = sjson.Set(out, "thinking.type", "enabled")
-							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						}
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
 					}
 				}
 			}

internal/translator/codex/claude/codex_claude_request.go

@@ -12,7 +12,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -52,7 +51,7 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	systemsResult := rootResult.Get("system")
 	if systemsResult.IsArray() {
 		systemResults := systemsResult.Array()
-		message := `{"type":"message","role":"user","content":[]}`
+		message := `{"type":"message","role":"developer","content":[]}`
 		for i := 0; i < len(systemResults); i++ {
 			systemResult := systemResults[i]
 			systemTypeResult := systemResult.Get("type")
@@ -218,18 +217,15 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Add additional configuration parameters for the Codex API.
 	template, _ = sjson.Set(template, "parallel_tool_calls", true)
 
-	// Convert thinking.budget_tokens to reasoning.effort for level-based models
-	reasoningEffort := "medium" // default
+	// Convert thinking.budget_tokens to reasoning.effort.
+	reasoningEffort := "medium"
 	if thinkingConfig := rootResult.Get("thinking"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
 		switch thinkingConfig.Get("type").String() {
 		case "enabled":
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0 {
-				if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
-					budget := int(budgetTokens.Int())
-					if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
-						reasoningEffort = effort
-					}
+			if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
+				budget := int(budgetTokens.Int())
+				if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+					reasoningEffort = effort
 				}
 			}
 		case "disabled":
@@ -245,21 +241,23 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	template, _ = sjson.Set(template, "include", []string{"reasoning.encrypted_content"})
 
 	// Add a first message to ignore system instructions and ensure proper execution.
-	inputResult := gjson.Get(template, "input")
-	if inputResult.Exists() && inputResult.IsArray() {
-		inputResults := inputResult.Array()
-		newInput := "[]"
-		for i := 0; i < len(inputResults); i++ {
-			if i == 0 {
-				firstText := inputResults[i].Get("content.0.text")
-				firstInstructions := "EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"
-				if firstText.Exists() && firstText.String() != firstInstructions {
-					newInput, _ = sjson.SetRaw(newInput, "-1", `{"type":"message","role":"user","content":[{"type":"input_text","text":"EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"}]}`)
+	if misc.GetCodexInstructionsEnabled() {
+		inputResult := gjson.Get(template, "input")
+		if inputResult.Exists() && inputResult.IsArray() {
+			inputResults := inputResult.Array()
+			newInput := "[]"
+			for i := 0; i < len(inputResults); i++ {
+				if i == 0 {
+					firstText := inputResults[i].Get("content.0.text")
+					firstInstructions := "EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"
+					if firstText.Exists() && firstText.String() != firstInstructions {
+						newInput, _ = sjson.SetRaw(newInput, "-1", `{"type":"message","role":"user","content":[{"type":"input_text","text":"EXECUTE ACCORDING TO THE FOLLOWING INSTRUCTIONS!!!"}]}`)
+					}
 				}
+				newInput, _ = sjson.SetRaw(newInput, "-1", inputResults[i].Raw)
 			}
-			newInput, _ = sjson.SetRaw(newInput, "-1", inputResults[i].Raw)
+			template, _ = sjson.SetRaw(template, "input", newInput)
 		}
-		template, _ = sjson.SetRaw(template, "input", newInput)
 	}
 
 	return []byte(template)

internal/translator/codex/claude/codex_claude_response.go

@@ -117,8 +117,12 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		} else {
 			template, _ = sjson.Set(template, "delta.stop_reason", "end_turn")
 		}
-		template, _ = sjson.Set(template, "usage.input_tokens", rootResult.Get("response.usage.input_tokens").Int())
-		template, _ = sjson.Set(template, "usage.output_tokens", rootResult.Get("response.usage.output_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractResponsesUsage(rootResult.Get("response.usage"))
+		template, _ = sjson.Set(template, "usage.input_tokens", inputTokens)
+		template, _ = sjson.Set(template, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			template, _ = sjson.Set(template, "usage.cache_read_input_tokens", cachedTokens)
+		}
 
 		output = "event: message_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
@@ -204,8 +208,12 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 	out := `{"id":"","type":"message","role":"assistant","model":"","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}`
 	out, _ = sjson.Set(out, "id", responseData.Get("id").String())
 	out, _ = sjson.Set(out, "model", responseData.Get("model").String())
-	out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
-	out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
+	inputTokens, outputTokens, cachedTokens := extractResponsesUsage(responseData.Get("usage"))
+	out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+	out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+	if cachedTokens > 0 {
+		out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+	}
 
 	hasToolCall := false
 
@@ -308,12 +316,27 @@ func ConvertCodexResponseToClaudeNonStream(_ context.Context, _ string, original
 		out, _ = sjson.SetRaw(out, "stop_sequence", stopSequence.Raw)
 	}
 
-	if responseData.Get("usage.input_tokens").Exists() || responseData.Get("usage.output_tokens").Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", responseData.Get("usage.input_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", responseData.Get("usage.output_tokens").Int())
+	return out
+}
+
+func extractResponsesUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
 	}
 
-	return out
+	inputTokens := usage.Get("input_tokens").Int()
+	outputTokens := usage.Get("output_tokens").Int()
+	cachedTokens := usage.Get("input_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
 }
 
 // buildReverseMapFromClaudeOriginalShortToOriginal builds a map[short]original from original Claude request tools.

internal/translator/codex/gemini/codex_gemini_request.go

@@ -14,7 +14,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"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/tidwall/gjson"
@@ -95,7 +94,7 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// System instruction -> as a user message with input_text parts
 	sysParts := root.Get("system_instruction.parts")
 	if sysParts.IsArray() {
-		msg := `{"type":"message","role":"user","content":[]}`
+		msg := `{"type":"message","role":"developer","content":[]}`
 		arr := sysParts.Array()
 		for i := 0; i < len(arr); i++ {
 			p := arr[i]
@@ -249,22 +248,28 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Fixed flags aligning with Codex expectations
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
 
-	// Convert thinkingBudget to reasoning.effort for level-based models
-	reasoningEffort := "medium" // default
+	// Convert Gemini thinkingConfig to Codex reasoning.effort.
+	effortSet := false
 	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			modelInfo := registry.LookupModelInfo(modelName)
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0 {
-				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-					budget := int(thinkingBudget.Int())
-					if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
-						reasoningEffort = effort
-					}
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
+					effortSet = true
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
+					effortSet = true
 				}
 			}
 		}
 	}
-	out, _ = sjson.Set(out, "reasoning.effort", reasoningEffort)
+	if !effortSet {
+		// No thinking config, set default effort
+		out, _ = sjson.Set(out, "reasoning.effort", "medium")
+	}
 	out, _ = sjson.Set(out, "reasoning.summary", "auto")
 	out, _ = sjson.Set(out, "stream", true)
 	out, _ = sjson.Set(out, "store", false)

internal/translator/codex/openai/chat-completions/codex_openai_request.go

@@ -33,7 +33,7 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 	rawJSON := bytes.Clone(inputRawJSON)
 	userAgent := misc.ExtractCodexUserAgent(rawJSON)
 	// Start with empty JSON object
-	out := `{}`
+	out := `{"instructions":""}`
 
 	// Stream must be set to true
 	out, _ = sjson.Set(out, "stream", stream)
@@ -98,7 +98,9 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 	// Extract system instructions from first system message (string or text object)
 	messages := gjson.GetBytes(rawJSON, "messages")
 	_, instructions := misc.CodexInstructionsForModel(modelName, "", userAgent)
-	out, _ = sjson.Set(out, "instructions", instructions)
+	if misc.GetCodexInstructionsEnabled() {
+		out, _ = sjson.Set(out, "instructions", instructions)
+	}
 	// if messages.IsArray() {
 	// 	arr := messages.Array()
 	// 	for i := 0; i < len(arr); i++ {
@@ -141,7 +143,7 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 				msg := `{}`
 				msg, _ = sjson.Set(msg, "type", "message")
 				if role == "system" {
-					msg, _ = sjson.Set(msg, "role", "user")
+					msg, _ = sjson.Set(msg, "role", "developer")
 				} else {
 					msg, _ = sjson.Set(msg, "role", role)
 				}

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

@@ -9,7 +9,6 @@ import (
 	"bytes"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -161,14 +160,11 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
-		if modelInfo != nil && modelInfo.Thinking != nil {
-			if t.Get("type").String() == "enabled" {
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-					budget := int(b.Int())
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-				}
+		if t.Get("type").String() == "enabled" {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				budget := int(b.Int())
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_request.go

@@ -63,6 +63,13 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 		out, _ = sjson.SetBytes(out, "request.generationConfig.topK", tkr.Num)
 	}
 
+	// Candidate count (OpenAI 'n' parameter)
+	if n := gjson.GetBytes(rawJSON, "n"); n.Exists() && n.Type == gjson.Number {
+		if val := n.Int(); val > 1 {
+			out, _ = sjson.SetBytes(out, "request.generationConfig.candidateCount", val)
+		}
+	}
+
 	// Map OpenAI modalities -> Gemini CLI request.generationConfig.responseModalities
 	// e.g. "modalities": ["image", "text"] -> ["IMAGE", "TEXT"]
 	if mods := gjson.GetBytes(rawJSON, "modalities"); mods.Exists() && mods.IsArray() {
@@ -129,6 +136,7 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
@@ -138,16 +146,19 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 				// system -> request.systemInstruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
-					out, _ = sjson.SetBytes(out, "request.systemInstruction.parts.0.text", content.Get("text").String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} else if content.IsArray() {
 					contents := content.Array()
 					if len(contents) > 0 {
 						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
 						for j := 0; j < len(contents); j++ {
-							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -9,7 +9,6 @@ import (
 	"bytes"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -153,16 +152,13 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
-	// Only apply for models that use numeric budgets, not discrete levels.
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
-			if t.Get("type").String() == "enabled" {
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-					budget := int(b.Int())
-					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-				}
+		if t.Get("type").String() == "enabled" {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				budget := int(b.Int())
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
 		}
 	}

internal/translator/gemini/openai/chat-completions/gemini_openai_request.go

@@ -63,6 +63,13 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		out, _ = sjson.SetBytes(out, "generationConfig.topK", tkr.Num)
 	}
 
+	// Candidate count (OpenAI 'n' parameter)
+	if n := gjson.GetBytes(rawJSON, "n"); n.Exists() && n.Type == gjson.Number {
+		if val := n.Int(); val > 1 {
+			out, _ = sjson.SetBytes(out, "generationConfig.candidateCount", val)
+		}
+	}
+
 	// Map OpenAI modalities -> Gemini generationConfig.responseModalities
 	// e.g. "modalities": ["image", "text"] -> ["IMAGE", "TEXT"]
 	if mods := gjson.GetBytes(rawJSON, "modalities"); mods.Exists() && mods.IsArray() {
@@ -129,6 +136,7 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 			}
 		}
 
+		systemPartIndex := 0
 		for i := 0; i < len(arr); i++ {
 			m := arr[i]
 			role := m.Get("role").String()
@@ -138,16 +146,19 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				// system -> system_instruction as a user message style
 				if content.Type == gjson.String {
 					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
-					out, _ = sjson.SetBytes(out, "system_instruction.parts.0.text", content.String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.String())
+					systemPartIndex++
 				} else if content.IsObject() && content.Get("type").String() == "text" {
 					out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
-					out, _ = sjson.SetBytes(out, "system_instruction.parts.0.text", content.Get("text").String())
+					out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), content.Get("text").String())
+					systemPartIndex++
 				} else if content.IsArray() {
 					contents := content.Array()
 					if len(contents) > 0 {
-						out, _ = sjson.SetBytes(out, "request.systemInstruction.role", "user")
+						out, _ = sjson.SetBytes(out, "system_instruction.role", "user")
 						for j := 0; j < len(contents); j++ {
-							out, _ = sjson.SetBytes(out, fmt.Sprintf("request.systemInstruction.parts.%d.text", j), contents[j].Get("text").String())
+							out, _ = sjson.SetBytes(out, fmt.Sprintf("system_instruction.parts.%d.text", systemPartIndex), contents[j].Get("text").String())
+							systemPartIndex++
 						}
 					}
 				}
@@ -162,7 +173,10 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 					for _, item := range items {
 						switch item.Get("type").String() {
 						case "text":
-							node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", item.Get("text").String())
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							imageURL := item.Get("image_url.url").String()
@@ -207,6 +221,10 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 					for _, item := range content.Array() {
 						switch item.Get("type").String() {
 						case "text":
+							text := item.Get("text").String()
+							if text != "" {
+								node, _ = sjson.SetBytes(node, "parts."+itoa(p)+".text", text)
+							}
 							p++
 						case "image_url":
 							// If the assistant returned an inline data URL, preserve it for history fidelity.

internal/translator/gemini/openai/chat-completions/gemini_openai_response.go

@@ -21,7 +21,8 @@ import (
 // convertGeminiResponseToOpenAIChatParams holds parameters for response conversion.
 type convertGeminiResponseToOpenAIChatParams struct {
 	UnixTimestamp int64
-	FunctionIndex int
+	// FunctionIndex tracks tool call indices per candidate index to support multiple candidates.
+	FunctionIndex map[int]int
 }
 
 // functionCallIDCounter provides a process-wide unique counter for function call identifiers.
@@ -42,13 +43,20 @@ var functionCallIDCounter uint64
 // Returns:
 //   - []string: A slice of strings, each containing an OpenAI-compatible JSON response
 func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, param *any) []string {
+	// Initialize parameters if nil.
 	if *param == nil {
 		*param = &convertGeminiResponseToOpenAIChatParams{
 			UnixTimestamp: 0,
-			FunctionIndex: 0,
+			FunctionIndex: make(map[int]int),
 		}
 	}
 
+	// Ensure the Map is initialized (handling cases where param might be reused from older context).
+	p := (*param).(*convertGeminiResponseToOpenAIChatParams)
+	if p.FunctionIndex == nil {
+		p.FunctionIndex = make(map[int]int)
+	}
+
 	if bytes.HasPrefix(rawJSON, []byte("data:")) {
 		rawJSON = bytes.TrimSpace(rawJSON[5:])
 	}
@@ -57,151 +65,179 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 		return []string{}
 	}
 
-	// Initialize the OpenAI SSE template.
-	template := `{"id":"","object":"chat.completion.chunk","created":12345,"model":"model","choices":[{"index":0,"delta":{"role":null,"content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}]}`
+	// Initialize the OpenAI SSE base template.
+	// We use a base template and clone it for each candidate to support multiple candidates.
+	baseTemplate := `{"id":"","object":"chat.completion.chunk","created":12345,"model":"model","choices":[{"index":0,"delta":{"role":null,"content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}]}`
 
 	// Extract and set the model version.
 	if modelVersionResult := gjson.GetBytes(rawJSON, "modelVersion"); modelVersionResult.Exists() {
-		template, _ = sjson.Set(template, "model", modelVersionResult.String())
+		baseTemplate, _ = sjson.Set(baseTemplate, "model", modelVersionResult.String())
 	}
 
 	// Extract and set the creation timestamp.
 	if createTimeResult := gjson.GetBytes(rawJSON, "createTime"); createTimeResult.Exists() {
 		t, err := time.Parse(time.RFC3339Nano, createTimeResult.String())
 		if err == nil {
-			(*param).(*convertGeminiResponseToOpenAIChatParams).UnixTimestamp = t.Unix()
+			p.UnixTimestamp = t.Unix()
 		}
-		template, _ = sjson.Set(template, "created", (*param).(*convertGeminiResponseToOpenAIChatParams).UnixTimestamp)
+		baseTemplate, _ = sjson.Set(baseTemplate, "created", p.UnixTimestamp)
 	} else {
-		template, _ = sjson.Set(template, "created", (*param).(*convertGeminiResponseToOpenAIChatParams).UnixTimestamp)
+		baseTemplate, _ = sjson.Set(baseTemplate, "created", p.UnixTimestamp)
 	}
 
 	// Extract and set the response ID.
 	if responseIDResult := gjson.GetBytes(rawJSON, "responseId"); responseIDResult.Exists() {
-		template, _ = sjson.Set(template, "id", responseIDResult.String())
-	}
-
-	// Extract and set the finish reason.
-	if finishReasonResult := gjson.GetBytes(rawJSON, "candidates.0.finishReason"); finishReasonResult.Exists() {
-		template, _ = sjson.Set(template, "choices.0.finish_reason", strings.ToLower(finishReasonResult.String()))
-		template, _ = sjson.Set(template, "choices.0.native_finish_reason", strings.ToLower(finishReasonResult.String()))
+		baseTemplate, _ = sjson.Set(baseTemplate, "id", responseIDResult.String())
 	}
 
 	// Extract and set usage metadata (token counts).
+	// Usage is applied to the base template so it appears in the chunks.
 	if usageResult := gjson.GetBytes(rawJSON, "usageMetadata"); usageResult.Exists() {
 		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
-			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
+			baseTemplate, _ = sjson.Set(baseTemplate, "usage.completion_tokens", candidatesTokenCountResult.Int())
 		}
 		if totalTokenCountResult := usageResult.Get("totalTokenCount"); totalTokenCountResult.Exists() {
-			template, _ = sjson.Set(template, "usage.total_tokens", totalTokenCountResult.Int())
+			baseTemplate, _ = sjson.Set(baseTemplate, "usage.total_tokens", totalTokenCountResult.Int())
 		}
 		promptTokenCount := usageResult.Get("promptTokenCount").Int() - cachedTokenCount
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
+		baseTemplate, _ = sjson.Set(baseTemplate, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
 		if thoughtsTokenCount > 0 {
-			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
+			baseTemplate, _ = sjson.Set(baseTemplate, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
 		// Include cached token count if present (indicates prompt caching is working)
 		if cachedTokenCount > 0 {
 			var err error
-			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			baseTemplate, err = sjson.Set(baseTemplate, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
 			if err != nil {
 				log.Warnf("gemini openai response: failed to set cached_tokens in streaming: %v", err)
 			}
 		}
 	}
 
-	// Process the main content part of the response.
-	partsResult := gjson.GetBytes(rawJSON, "candidates.0.content.parts")
-	hasFunctionCall := false
-	if partsResult.IsArray() {
-		partResults := partsResult.Array()
-		for i := 0; i < len(partResults); i++ {
-			partResult := partResults[i]
-			partTextResult := partResult.Get("text")
-			functionCallResult := partResult.Get("functionCall")
-			inlineDataResult := partResult.Get("inlineData")
-			if !inlineDataResult.Exists() {
-				inlineDataResult = partResult.Get("inline_data")
-			}
-			thoughtSignatureResult := partResult.Get("thoughtSignature")
-			if !thoughtSignatureResult.Exists() {
-				thoughtSignatureResult = partResult.Get("thought_signature")
+	var responseStrings []string
+	candidates := gjson.GetBytes(rawJSON, "candidates")
+
+	// Iterate over all candidates to support candidate_count > 1.
+	if candidates.IsArray() {
+		candidates.ForEach(func(_, candidate gjson.Result) bool {
+			// Clone the template for the current candidate.
+			template := baseTemplate
+
+			// Set the specific index for this candidate.
+			candidateIndex := int(candidate.Get("index").Int())
+			template, _ = sjson.Set(template, "choices.0.index", candidateIndex)
+
+			// Extract and set the finish reason.
+			if finishReasonResult := candidate.Get("finishReason"); finishReasonResult.Exists() {
+				template, _ = sjson.Set(template, "choices.0.finish_reason", strings.ToLower(finishReasonResult.String()))
+				template, _ = sjson.Set(template, "choices.0.native_finish_reason", strings.ToLower(finishReasonResult.String()))
 			}
 
-			hasThoughtSignature := thoughtSignatureResult.Exists() && thoughtSignatureResult.String() != ""
-			hasContentPayload := partTextResult.Exists() || functionCallResult.Exists() || inlineDataResult.Exists()
+			partsResult := candidate.Get("content.parts")
+			hasFunctionCall := false
 
-			// Skip pure thoughtSignature parts but keep any actual payload in the same part.
-			if hasThoughtSignature && !hasContentPayload {
-				continue
+			if partsResult.IsArray() {
+				partResults := partsResult.Array()
+				for i := 0; i < len(partResults); i++ {
+					partResult := partResults[i]
+					partTextResult := partResult.Get("text")
+					functionCallResult := partResult.Get("functionCall")
+					inlineDataResult := partResult.Get("inlineData")
+					if !inlineDataResult.Exists() {
+						inlineDataResult = partResult.Get("inline_data")
+					}
+					thoughtSignatureResult := partResult.Get("thoughtSignature")
+					if !thoughtSignatureResult.Exists() {
+						thoughtSignatureResult = partResult.Get("thought_signature")
+					}
+
+					hasThoughtSignature := thoughtSignatureResult.Exists() && thoughtSignatureResult.String() != ""
+					hasContentPayload := partTextResult.Exists() || functionCallResult.Exists() || inlineDataResult.Exists()
+
+					// Skip pure thoughtSignature parts but keep any actual payload in the same part.
+					if hasThoughtSignature && !hasContentPayload {
+						continue
+					}
+
+					if partTextResult.Exists() {
+						text := partTextResult.String()
+						// Handle text content, distinguishing between regular content and reasoning/thoughts.
+						if partResult.Get("thought").Bool() {
+							template, _ = sjson.Set(template, "choices.0.delta.reasoning_content", text)
+						} else {
+							template, _ = sjson.Set(template, "choices.0.delta.content", text)
+						}
+						template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+					} else if functionCallResult.Exists() {
+						// Handle function call content.
+						hasFunctionCall = true
+						toolCallsResult := gjson.Get(template, "choices.0.delta.tool_calls")
+
+						// Retrieve the function index for this specific candidate.
+						functionCallIndex := p.FunctionIndex[candidateIndex]
+						p.FunctionIndex[candidateIndex]++
+
+						if toolCallsResult.Exists() && toolCallsResult.IsArray() {
+							functionCallIndex = len(toolCallsResult.Array())
+						} else {
+							template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
+						}
+
+						functionCallTemplate := `{"id": "","index": 0,"type": "function","function": {"name": "","arguments": ""}}`
+						fcName := functionCallResult.Get("name").String()
+						functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
+						functionCallTemplate, _ = sjson.Set(functionCallTemplate, "index", functionCallIndex)
+						functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
+						if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
+							functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.arguments", fcArgsResult.Raw)
+						}
+						template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+						template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallTemplate)
+					} else if inlineDataResult.Exists() {
+						data := inlineDataResult.Get("data").String()
+						if data == "" {
+							continue
+						}
+						mimeType := inlineDataResult.Get("mimeType").String()
+						if mimeType == "" {
+							mimeType = inlineDataResult.Get("mime_type").String()
+						}
+						if mimeType == "" {
+							mimeType = "image/png"
+						}
+						imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
+						imagesResult := gjson.Get(template, "choices.0.delta.images")
+						if !imagesResult.Exists() || !imagesResult.IsArray() {
+							template, _ = sjson.SetRaw(template, "choices.0.delta.images", `[]`)
+						}
+						imageIndex := len(gjson.Get(template, "choices.0.delta.images").Array())
+						imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+						imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+						imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
+						template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
+						template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", imagePayload)
+					}
+				}
 			}
 
-			if partTextResult.Exists() {
-				text := partTextResult.String()
-				// Handle text content, distinguishing between regular content and reasoning/thoughts.
-				if partResult.Get("thought").Bool() {
-					template, _ = sjson.Set(template, "choices.0.delta.reasoning_content", text)
-				} else {
-					template, _ = sjson.Set(template, "choices.0.delta.content", text)
-				}
-				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-			} else if functionCallResult.Exists() {
-				// Handle function call content.
-				hasFunctionCall = true
-				toolCallsResult := gjson.Get(template, "choices.0.delta.tool_calls")
-				functionCallIndex := (*param).(*convertGeminiResponseToOpenAIChatParams).FunctionIndex
-				(*param).(*convertGeminiResponseToOpenAIChatParams).FunctionIndex++
-				if toolCallsResult.Exists() && toolCallsResult.IsArray() {
-					functionCallIndex = len(toolCallsResult.Array())
-				} else {
-					template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls", `[]`)
-				}
-
-				functionCallTemplate := `{"id": "","index": 0,"type": "function","function": {"name": "","arguments": ""}}`
-				fcName := functionCallResult.Get("name").String()
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "index", functionCallIndex)
-				functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.name", fcName)
-				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
-					functionCallTemplate, _ = sjson.Set(functionCallTemplate, "function.arguments", fcArgsResult.Raw)
-				}
-				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.delta.tool_calls.-1", functionCallTemplate)
-			} else if inlineDataResult.Exists() {
-				data := inlineDataResult.Get("data").String()
-				if data == "" {
-					continue
-				}
-				mimeType := inlineDataResult.Get("mimeType").String()
-				if mimeType == "" {
-					mimeType = inlineDataResult.Get("mime_type").String()
-				}
-				if mimeType == "" {
-					mimeType = "image/png"
-				}
-				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagesResult := gjson.Get(template, "choices.0.delta.images")
-				if !imagesResult.Exists() || !imagesResult.IsArray() {
-					template, _ = sjson.SetRaw(template, "choices.0.delta.images", `[]`)
-				}
-				imageIndex := len(gjson.Get(template, "choices.0.delta.images").Array())
-				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
-				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
-				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
-				template, _ = sjson.Set(template, "choices.0.delta.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.delta.images.-1", imagePayload)
+			if hasFunctionCall {
+				template, _ = sjson.Set(template, "choices.0.finish_reason", "tool_calls")
+				template, _ = sjson.Set(template, "choices.0.native_finish_reason", "tool_calls")
 			}
+
+			responseStrings = append(responseStrings, template)
+			return true // continue loop
+		})
+	} else {
+		// If there are no candidates (e.g., a pure usageMetadata chunk), return the usage chunk if present.
+		if gjson.GetBytes(rawJSON, "usageMetadata").Exists() && len(responseStrings) == 0 {
+			responseStrings = append(responseStrings, baseTemplate)
 		}
 	}
 
-	if hasFunctionCall {
-		template, _ = sjson.Set(template, "choices.0.finish_reason", "tool_calls")
-		template, _ = sjson.Set(template, "choices.0.native_finish_reason", "tool_calls")
-	}
-
-	return []string{template}
+	return responseStrings
 }
 
 // ConvertGeminiResponseToOpenAINonStream converts a non-streaming Gemini response to a non-streaming OpenAI response.
@@ -219,7 +255,9 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 //   - string: An OpenAI-compatible JSON response containing all message content and metadata
 func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, originalRequestRawJSON, requestRawJSON, rawJSON []byte, _ *any) string {
 	var unixTimestamp int64
-	template := `{"id":"","object":"chat.completion","created":123456,"model":"model","choices":[{"index":0,"message":{"role":"assistant","content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}]}`
+	// Initialize template with an empty choices array to support multiple candidates.
+	template := `{"id":"","object":"chat.completion","created":123456,"model":"model","choices":[]}`
+
 	if modelVersionResult := gjson.GetBytes(rawJSON, "modelVersion"); modelVersionResult.Exists() {
 		template, _ = sjson.Set(template, "model", modelVersionResult.String())
 	}
@@ -238,11 +276,6 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 		template, _ = sjson.Set(template, "id", responseIDResult.String())
 	}
 
-	if finishReasonResult := gjson.GetBytes(rawJSON, "candidates.0.finishReason"); finishReasonResult.Exists() {
-		template, _ = sjson.Set(template, "choices.0.finish_reason", strings.ToLower(finishReasonResult.String()))
-		template, _ = sjson.Set(template, "choices.0.native_finish_reason", strings.ToLower(finishReasonResult.String()))
-	}
-
 	if usageResult := gjson.GetBytes(rawJSON, "usageMetadata"); usageResult.Exists() {
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
@@ -267,74 +300,96 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 		}
 	}
 
-	// Process the main content part of the response.
-	partsResult := gjson.GetBytes(rawJSON, "candidates.0.content.parts")
-	hasFunctionCall := false
-	if partsResult.IsArray() {
-		partsResults := partsResult.Array()
-		for i := 0; i < len(partsResults); i++ {
-			partResult := partsResults[i]
-			partTextResult := partResult.Get("text")
-			functionCallResult := partResult.Get("functionCall")
-			inlineDataResult := partResult.Get("inlineData")
-			if !inlineDataResult.Exists() {
-				inlineDataResult = partResult.Get("inline_data")
+	// Process the main content part of the response for all candidates.
+	candidates := gjson.GetBytes(rawJSON, "candidates")
+	if candidates.IsArray() {
+		candidates.ForEach(func(_, candidate gjson.Result) bool {
+			// Construct a single Choice object.
+			choiceTemplate := `{"index":0,"message":{"role":"assistant","content":null,"reasoning_content":null,"tool_calls":null},"finish_reason":null,"native_finish_reason":null}`
+
+			// Set the index for this choice.
+			choiceTemplate, _ = sjson.Set(choiceTemplate, "index", candidate.Get("index").Int())
+
+			// Set finish reason.
+			if finishReasonResult := candidate.Get("finishReason"); finishReasonResult.Exists() {
+				choiceTemplate, _ = sjson.Set(choiceTemplate, "finish_reason", strings.ToLower(finishReasonResult.String()))
+				choiceTemplate, _ = sjson.Set(choiceTemplate, "native_finish_reason", strings.ToLower(finishReasonResult.String()))
 			}
 
-			if partTextResult.Exists() {
-				// Append text content, distinguishing between regular content and reasoning.
-				if partResult.Get("thought").Bool() {
-					template, _ = sjson.Set(template, "choices.0.message.reasoning_content", partTextResult.String())
-				} else {
-					template, _ = sjson.Set(template, "choices.0.message.content", partTextResult.String())
-				}
-				template, _ = sjson.Set(template, "choices.0.message.role", "assistant")
-			} else if functionCallResult.Exists() {
-				// Append function call content to the tool_calls array.
-				hasFunctionCall = true
-				toolCallsResult := gjson.Get(template, "choices.0.message.tool_calls")
-				if !toolCallsResult.Exists() || !toolCallsResult.IsArray() {
-					template, _ = sjson.SetRaw(template, "choices.0.message.tool_calls", `[]`)
-				}
-				functionCallItemTemplate := `{"id": "","type": "function","function": {"name": "","arguments": ""}}`
-				fcName := functionCallResult.Get("name").String()
-				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
-				functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", fcName)
-				if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
-					functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", fcArgsResult.Raw)
-				}
-				template, _ = sjson.Set(template, "choices.0.message.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.message.tool_calls.-1", functionCallItemTemplate)
-			} else if inlineDataResult.Exists() {
-				data := inlineDataResult.Get("data").String()
-				if data == "" {
-					continue
-				}
-				mimeType := inlineDataResult.Get("mimeType").String()
-				if mimeType == "" {
-					mimeType = inlineDataResult.Get("mime_type").String()
-				}
-				if mimeType == "" {
-					mimeType = "image/png"
-				}
-				imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
-				imagesResult := gjson.Get(template, "choices.0.message.images")
-				if !imagesResult.Exists() || !imagesResult.IsArray() {
-					template, _ = sjson.SetRaw(template, "choices.0.message.images", `[]`)
-				}
-				imageIndex := len(gjson.Get(template, "choices.0.message.images").Array())
-				imagePayload := `{"type":"image_url","image_url":{"url":""}}`
-				imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
-				imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
-				template, _ = sjson.Set(template, "choices.0.message.role", "assistant")
-				template, _ = sjson.SetRaw(template, "choices.0.message.images.-1", imagePayload)
-			}
-		}
-	}
+			partsResult := candidate.Get("content.parts")
+			hasFunctionCall := false
+			if partsResult.IsArray() {
+				partsResults := partsResult.Array()
+				for i := 0; i < len(partsResults); i++ {
+					partResult := partsResults[i]
+					partTextResult := partResult.Get("text")
+					functionCallResult := partResult.Get("functionCall")
+					inlineDataResult := partResult.Get("inlineData")
+					if !inlineDataResult.Exists() {
+						inlineDataResult = partResult.Get("inline_data")
+					}
 
-	if hasFunctionCall {
-		template, _ = sjson.Set(template, "choices.0.finish_reason", "tool_calls")
-		template, _ = sjson.Set(template, "choices.0.native_finish_reason", "tool_calls")
+					if partTextResult.Exists() {
+						// Append text content, distinguishing between regular content and reasoning.
+						if partResult.Get("thought").Bool() {
+							oldVal := gjson.Get(choiceTemplate, "message.reasoning_content").String()
+							choiceTemplate, _ = sjson.Set(choiceTemplate, "message.reasoning_content", oldVal+partTextResult.String())
+						} else {
+							oldVal := gjson.Get(choiceTemplate, "message.content").String()
+							choiceTemplate, _ = sjson.Set(choiceTemplate, "message.content", oldVal+partTextResult.String())
+						}
+						choiceTemplate, _ = sjson.Set(choiceTemplate, "message.role", "assistant")
+					} else if functionCallResult.Exists() {
+						// Append function call content to the tool_calls array.
+						hasFunctionCall = true
+						toolCallsResult := gjson.Get(choiceTemplate, "message.tool_calls")
+						if !toolCallsResult.Exists() || !toolCallsResult.IsArray() {
+							choiceTemplate, _ = sjson.SetRaw(choiceTemplate, "message.tool_calls", `[]`)
+						}
+						functionCallItemTemplate := `{"id": "","type": "function","function": {"name": "","arguments": ""}}`
+						fcName := functionCallResult.Get("name").String()
+						functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "id", fmt.Sprintf("%s-%d-%d", fcName, time.Now().UnixNano(), atomic.AddUint64(&functionCallIDCounter, 1)))
+						functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.name", fcName)
+						if fcArgsResult := functionCallResult.Get("args"); fcArgsResult.Exists() {
+							functionCallItemTemplate, _ = sjson.Set(functionCallItemTemplate, "function.arguments", fcArgsResult.Raw)
+						}
+						choiceTemplate, _ = sjson.Set(choiceTemplate, "message.role", "assistant")
+						choiceTemplate, _ = sjson.SetRaw(choiceTemplate, "message.tool_calls.-1", functionCallItemTemplate)
+					} else if inlineDataResult.Exists() {
+						data := inlineDataResult.Get("data").String()
+						if data != "" {
+							mimeType := inlineDataResult.Get("mimeType").String()
+							if mimeType == "" {
+								mimeType = inlineDataResult.Get("mime_type").String()
+							}
+							if mimeType == "" {
+								mimeType = "image/png"
+							}
+							imageURL := fmt.Sprintf("data:%s;base64,%s", mimeType, data)
+							imagesResult := gjson.Get(choiceTemplate, "message.images")
+							if !imagesResult.Exists() || !imagesResult.IsArray() {
+								choiceTemplate, _ = sjson.SetRaw(choiceTemplate, "message.images", `[]`)
+							}
+							imageIndex := len(gjson.Get(choiceTemplate, "message.images").Array())
+							imagePayload := `{"type":"image_url","image_url":{"url":""}}`
+							imagePayload, _ = sjson.Set(imagePayload, "index", imageIndex)
+							imagePayload, _ = sjson.Set(imagePayload, "image_url.url", imageURL)
+							choiceTemplate, _ = sjson.Set(choiceTemplate, "message.role", "assistant")
+							choiceTemplate, _ = sjson.SetRaw(choiceTemplate, "message.images.-1", imagePayload)
+						}
+					}
+				}
+			}
+
+			if hasFunctionCall {
+				choiceTemplate, _ = sjson.Set(choiceTemplate, "finish_reason", "tool_calls")
+				choiceTemplate, _ = sjson.Set(choiceTemplate, "native_finish_reason", "tool_calls")
+			}
+
+			// Append the constructed choice to the main choices array.
+			template, _ = sjson.SetRaw(template, "choices.-1", choiceTemplate)
+			return true
+		})
 	}
 
 	return template

internal/translator/openai/claude/openai_claude_request.go

@@ -88,7 +88,7 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	var messagesJSON = "[]"
 
 	// Handle system message first
-	systemMsgJSON := `{"role":"system","content":[{"type":"text","text":"Use ANY tool, the parameters MUST accord with RFC 8259 (The JavaScript Object Notation (JSON) Data Interchange Format), the keys and value MUST be enclosed in double quotes."}]}`
+	systemMsgJSON := `{"role":"system","content":[]}`
 	if system := root.Get("system"); system.Exists() {
 		if system.Type == gjson.String {
 			if system.String() != "" {

internal/translator/openai/claude/openai_claude_response.go

@@ -289,21 +289,17 @@ func convertOpenAIStreamingChunkToAnthropic(rawJSON []byte, param *ConvertOpenAI
 	// Only process if usage has actual values (not null)
 	if param.FinishReason != "" {
 		usage := root.Get("usage")
-		var inputTokens, outputTokens int64
+		var inputTokens, outputTokens, cachedTokens int64
 		if usage.Exists() && usage.Type != gjson.Null {
-			// Check if usage has actual token counts
-			promptTokens := usage.Get("prompt_tokens")
-			completionTokens := usage.Get("completion_tokens")
-
-			if promptTokens.Exists() && completionTokens.Exists() {
-				inputTokens = promptTokens.Int()
-				outputTokens = completionTokens.Int()
-			}
+			inputTokens, outputTokens, cachedTokens = extractOpenAIUsage(usage)
 			// Send message_delta with usage
 			messageDeltaJSON := `{"type":"message_delta","delta":{"stop_reason":"","stop_sequence":null},"usage":{"input_tokens":0,"output_tokens":0}}`
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "delta.stop_reason", mapOpenAIFinishReasonToAnthropic(param.FinishReason))
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.input_tokens", inputTokens)
 			messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.output_tokens", outputTokens)
+			if cachedTokens > 0 {
+				messageDeltaJSON, _ = sjson.Set(messageDeltaJSON, "usage.cache_read_input_tokens", cachedTokens)
+			}
 			results = append(results, "event: message_delta\ndata: "+messageDeltaJSON+"\n\n")
 			param.MessageDeltaSent = true
 
@@ -423,13 +419,12 @@ func convertOpenAINonStreamingToAnthropic(rawJSON []byte) []string {
 
 	// Set usage information
 	if usage := root.Get("usage"); usage.Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", usage.Get("prompt_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", usage.Get("completion_tokens").Int())
-		reasoningTokens := int64(0)
-		if v := usage.Get("completion_tokens_details.reasoning_tokens"); v.Exists() {
-			reasoningTokens = v.Int()
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(usage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
 		}
-		out, _ = sjson.Set(out, "usage.reasoning_tokens", reasoningTokens)
 	}
 
 	return []string{out}
@@ -674,8 +669,12 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 	}
 
 	if respUsage := root.Get("usage"); respUsage.Exists() {
-		out, _ = sjson.Set(out, "usage.input_tokens", respUsage.Get("prompt_tokens").Int())
-		out, _ = sjson.Set(out, "usage.output_tokens", respUsage.Get("completion_tokens").Int())
+		inputTokens, outputTokens, cachedTokens := extractOpenAIUsage(respUsage)
+		out, _ = sjson.Set(out, "usage.input_tokens", inputTokens)
+		out, _ = sjson.Set(out, "usage.output_tokens", outputTokens)
+		if cachedTokens > 0 {
+			out, _ = sjson.Set(out, "usage.cache_read_input_tokens", cachedTokens)
+		}
 	}
 
 	if !stopReasonSet {
@@ -692,3 +691,23 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 func ClaudeTokenCount(ctx context.Context, count int64) string {
 	return fmt.Sprintf(`{"input_tokens":%d}`, count)
 }
+
+func extractOpenAIUsage(usage gjson.Result) (int64, int64, int64) {
+	if !usage.Exists() || usage.Type == gjson.Null {
+		return 0, 0, 0
+	}
+
+	inputTokens := usage.Get("prompt_tokens").Int()
+	outputTokens := usage.Get("completion_tokens").Int()
+	cachedTokens := usage.Get("prompt_tokens_details.cached_tokens").Int()
+
+	if cachedTokens > 0 {
+		if inputTokens >= cachedTokens {
+			inputTokens -= cachedTokens
+		} else {
+			inputTokens = 0
+		}
+	}
+
+	return inputTokens, outputTokens, cachedTokens
+}

internal/translator/openai/gemini/openai_gemini_request.go

@@ -77,12 +77,21 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 			}
 		}
 
-		// Convert thinkingBudget to reasoning_effort
+		// Candidate count (OpenAI 'n' parameter)
+		if candidateCount := genConfig.Get("candidateCount"); candidateCount.Exists() {
+			out, _ = sjson.Set(out, "n", candidateCount.Int())
+		}
+
+		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
 		// Always perform conversion to support allowCompat models that may not be in registry
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-				budget := int(thinkingBudget.Int())
-				if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}

internal/translator/openai/openai/responses/openai_openai-responses_response.go

@@ -12,6 +12,10 @@ import (
 	"github.com/tidwall/sjson"
 )
 
+type oaiToResponsesStateReasoning struct {
+	ReasoningID   string
+	ReasoningData string
+}
 type oaiToResponsesState struct {
 	Seq            int
 	ResponseID     string
@@ -23,6 +27,7 @@ type oaiToResponsesState struct {
 	// Per-output message text buffers by index
 	MsgTextBuf   map[int]*strings.Builder
 	ReasoningBuf strings.Builder
+	Reasonings   []oaiToResponsesStateReasoning
 	FuncArgsBuf  map[int]*strings.Builder // index -> args
 	FuncNames    map[int]string           // index -> name
 	FuncCallIDs  map[int]string           // index -> call_id
@@ -63,6 +68,7 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 			MsgItemDone:     make(map[int]bool),
 			FuncArgsDone:    make(map[int]bool),
 			FuncItemDone:    make(map[int]bool),
+			Reasonings:      make([]oaiToResponsesStateReasoning, 0),
 		}
 	}
 	st := (*param).(*oaiToResponsesState)
@@ -157,6 +163,31 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 		st.Started = true
 	}
 
+	stopReasoning := func(text string) {
+		// Emit reasoning done events
+		textDone := `{"type":"response.reasoning_summary_text.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
+		textDone, _ = sjson.Set(textDone, "sequence_number", nextSeq())
+		textDone, _ = sjson.Set(textDone, "item_id", st.ReasoningID)
+		textDone, _ = sjson.Set(textDone, "output_index", st.ReasoningIndex)
+		textDone, _ = sjson.Set(textDone, "text", text)
+		out = append(out, emitRespEvent("response.reasoning_summary_text.done", textDone))
+		partDone := `{"type":"response.reasoning_summary_part.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`
+		partDone, _ = sjson.Set(partDone, "sequence_number", nextSeq())
+		partDone, _ = sjson.Set(partDone, "item_id", st.ReasoningID)
+		partDone, _ = sjson.Set(partDone, "output_index", st.ReasoningIndex)
+		partDone, _ = sjson.Set(partDone, "part.text", text)
+		out = append(out, emitRespEvent("response.reasoning_summary_part.done", partDone))
+		outputItemDone := `{"type":"response.output_item.done","item":{"id":"","type":"reasoning","encrypted_content":"","summary":[{"type":"summary_text","text":""}]},"output_index":0,"sequence_number":0}`
+		outputItemDone, _ = sjson.Set(outputItemDone, "sequence_number", nextSeq())
+		outputItemDone, _ = sjson.Set(outputItemDone, "item.id", st.ReasoningID)
+		outputItemDone, _ = sjson.Set(outputItemDone, "output_index", st.ReasoningIndex)
+		outputItemDone, _ = sjson.Set(outputItemDone, "item.summary.text", text)
+		out = append(out, emitRespEvent("response.output_item.done", outputItemDone))
+
+		st.Reasonings = append(st.Reasonings, oaiToResponsesStateReasoning{ReasoningID: st.ReasoningID, ReasoningData: text})
+		st.ReasoningID = ""
+	}
+
 	// choices[].delta content / tool_calls / reasoning_content
 	if choices := root.Get("choices"); choices.Exists() && choices.IsArray() {
 		choices.ForEach(func(_, choice gjson.Result) bool {
@@ -165,6 +196,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 			if delta.Exists() {
 				if c := delta.Get("content"); c.Exists() && c.String() != "" {
 					// Ensure the message item and its first content part are announced before any text deltas
+					if st.ReasoningID != "" {
+						stopReasoning(st.ReasoningBuf.String())
+						st.ReasoningBuf.Reset()
+					}
 					if !st.MsgItemAdded[idx] {
 						item := `{"type":"response.output_item.added","sequence_number":0,"output_index":0,"item":{"id":"","type":"message","status":"in_progress","content":[],"role":"assistant"}}`
 						item, _ = sjson.Set(item, "sequence_number", nextSeq())
@@ -226,6 +261,10 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 
 				// tool calls
 				if tcs := delta.Get("tool_calls"); tcs.Exists() && tcs.IsArray() {
+					if st.ReasoningID != "" {
+						stopReasoning(st.ReasoningBuf.String())
+						st.ReasoningBuf.Reset()
+					}
 					// Before emitting any function events, if a message is open for this index,
 					// close its text/content to match Codex expected ordering.
 					if st.MsgItemAdded[idx] && !st.MsgItemDone[idx] {
@@ -361,17 +400,8 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 				}
 
 				if st.ReasoningID != "" {
-					// Emit reasoning done events
-					textDone := `{"type":"response.reasoning_summary_text.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"text":""}`
-					textDone, _ = sjson.Set(textDone, "sequence_number", nextSeq())
-					textDone, _ = sjson.Set(textDone, "item_id", st.ReasoningID)
-					textDone, _ = sjson.Set(textDone, "output_index", st.ReasoningIndex)
-					out = append(out, emitRespEvent("response.reasoning_summary_text.done", textDone))
-					partDone := `{"type":"response.reasoning_summary_part.done","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`
-					partDone, _ = sjson.Set(partDone, "sequence_number", nextSeq())
-					partDone, _ = sjson.Set(partDone, "item_id", st.ReasoningID)
-					partDone, _ = sjson.Set(partDone, "output_index", st.ReasoningIndex)
-					out = append(out, emitRespEvent("response.reasoning_summary_part.done", partDone))
+					stopReasoning(st.ReasoningBuf.String())
+					st.ReasoningBuf.Reset()
 				}
 
 				// Emit function call done events for any active function calls
@@ -485,11 +515,13 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
 				}
 				// Build response.output using aggregated buffers
 				outputsWrapper := `{"arr":[]}`
-				if st.ReasoningBuf.Len() > 0 {
-					item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
-					item, _ = sjson.Set(item, "id", st.ReasoningID)
-					item, _ = sjson.Set(item, "summary.0.text", st.ReasoningBuf.String())
-					outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
+				if len(st.Reasonings) > 0 {
+					for _, r := range st.Reasonings {
+						item := `{"id":"","type":"reasoning","summary":[{"type":"summary_text","text":""}]}`
+						item, _ = sjson.Set(item, "id", r.ReasoningID)
+						item, _ = sjson.Set(item, "summary.0.text", r.ReasoningData)
+						outputsWrapper, _ = sjson.SetRaw(outputsWrapper, "arr.-1", item)
+					}
 				}
 				// Append message items in ascending index order
 				if len(st.MsgItemAdded) > 0 {

internal/util/gemini_schema.go

@@ -19,6 +19,7 @@ func CleanJSONSchemaForAntigravity(jsonStr string) string {
 	// Phase 1: Convert and add hints
 	jsonStr = convertRefsToHints(jsonStr)
 	jsonStr = convertConstToEnum(jsonStr)
+	jsonStr = convertEnumValuesToStrings(jsonStr)
 	jsonStr = addEnumHints(jsonStr)
 	jsonStr = addAdditionalPropertiesHints(jsonStr)
 	jsonStr = moveConstraintsToDescription(jsonStr)
@@ -77,6 +78,33 @@ func convertConstToEnum(jsonStr string) string {
 	return jsonStr
 }
 
+// convertEnumValuesToStrings ensures all enum values are strings.
+// Gemini API requires enum values to be of type string, not numbers or booleans.
+func convertEnumValuesToStrings(jsonStr string) string {
+	for _, p := range findPaths(jsonStr, "enum") {
+		arr := gjson.Get(jsonStr, p)
+		if !arr.IsArray() {
+			continue
+		}
+
+		var stringVals []string
+		needsConversion := false
+		for _, item := range arr.Array() {
+			// Check if any value is not a string
+			if item.Type != gjson.String {
+				needsConversion = true
+			}
+			stringVals = append(stringVals, item.String())
+		}
+
+		// Only update if we found non-string values
+		if needsConversion {
+			jsonStr, _ = sjson.Set(jsonStr, p, stringVals)
+		}
+	}
+	return jsonStr
+}
+
 func addEnumHints(jsonStr string) string {
 	for _, p := range findPaths(jsonStr, "enum") {
 		arr := gjson.Get(jsonStr, p)

internal/util/gemini_schema_test.go

@@ -818,3 +818,54 @@ func TestCleanJSONSchemaForAntigravity_MultipleFormats(t *testing.T) {
 		t.Errorf("date-time format hint should be added, got: %s", result)
 	}
 }
+
+func TestCleanJSONSchemaForAntigravity_NumericEnumToString(t *testing.T) {
+	// Gemini API requires enum values to be strings, not numbers
+	input := `{
+		"type": "object",
+		"properties": {
+			"priority": {"type": "integer", "enum": [0, 1, 2]},
+			"level": {"type": "number", "enum": [1.5, 2.5, 3.5]},
+			"status": {"type": "string", "enum": ["active", "inactive"]}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Numeric enum values should be converted to strings
+	if strings.Contains(result, `"enum":[0,1,2]`) {
+		t.Errorf("Integer enum values should be converted to strings, got: %s", result)
+	}
+	if strings.Contains(result, `"enum":[1.5,2.5,3.5]`) {
+		t.Errorf("Float enum values should be converted to strings, got: %s", result)
+	}
+	// Should contain string versions
+	if !strings.Contains(result, `"0"`) || !strings.Contains(result, `"1"`) || !strings.Contains(result, `"2"`) {
+		t.Errorf("Integer enum values should be converted to string format, got: %s", result)
+	}
+	// String enum values should remain unchanged
+	if !strings.Contains(result, `"active"`) || !strings.Contains(result, `"inactive"`) {
+		t.Errorf("String enum values should remain unchanged, got: %s", result)
+	}
+}
+
+func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
+	// Boolean enum values should also be converted to strings
+	input := `{
+		"type": "object",
+		"properties": {
+			"enabled": {"type": "boolean", "enum": [true, false]}
+		}
+	}`
+
+	result := CleanJSONSchemaForAntigravity(input)
+
+	// Boolean enum values should be converted to strings
+	if strings.Contains(result, `"enum":[true,false]`) {
+		t.Errorf("Boolean enum values should be converted to strings, got: %s", result)
+	}
+	// Should contain string versions "true" and "false"
+	if !strings.Contains(result, `"true"`) || !strings.Contains(result, `"false"`) {
+		t.Errorf("Boolean enum values should be converted to string format, got: %s", result)
+	}
+}

sdk/api/handlers/gemini/gemini_handlers.go

@@ -56,8 +56,12 @@ func (h *GeminiAPIHandler) GeminiModels(c *gin.Context) {
 		for k, v := range model {
 			normalizedModel[k] = v
 		}
-		if name, ok := normalizedModel["name"].(string); ok && name != "" && !strings.HasPrefix(name, "models/") {
-			normalizedModel["name"] = "models/" + name
+		if name, ok := normalizedModel["name"].(string); ok && name != "" {
+			if !strings.HasPrefix(name, "models/") {
+				normalizedModel["name"] = "models/" + name
+			}
+			normalizedModel["displayName"] = name
+			normalizedModel["description"] = name
 		}
 		if _, ok := normalizedModel["supportedGenerationMethods"]; !ok {
 			normalizedModel["supportedGenerationMethods"] = defaultMethods

sdk/auth/codex.go

@@ -2,6 +2,8 @@ package auth
 
 import (
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
 	"net/http"
 	"strings"
@@ -191,7 +193,19 @@ waitForCallback:
 		return nil, fmt.Errorf("codex token storage missing account information")
 	}
 
-	fileName := fmt.Sprintf("codex-%s.json", tokenStorage.Email)
+	planType := ""
+	hashAccountID := ""
+	if tokenStorage.IDToken != "" {
+		if claims, errParse := codex.ParseJWTToken(tokenStorage.IDToken); errParse == nil && claims != nil {
+			planType = strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType)
+			accountID := strings.TrimSpace(claims.CodexAuthInfo.ChatgptAccountID)
+			if accountID != "" {
+				digest := sha256.Sum256([]byte(accountID))
+				hashAccountID = hex.EncodeToString(digest[:])[:8]
+			}
+		}
+	}
+	fileName := codex.CredentialFileName(tokenStorage.Email, planType, hashAccountID, true)
 	metadata := map[string]any{
 		"email": tokenStorage.Email,
 	}

sdk/auth/filestore.go

@@ -8,7 +8,6 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
-	"reflect"
 	"strings"
 	"sync"
 	"time"
@@ -76,7 +75,7 @@ func (s *FileTokenStore) Save(ctx context.Context, auth *cliproxyauth.Auth) (str
 		if existing, errRead := os.ReadFile(path); errRead == nil {
 			// Use metadataEqualIgnoringTimestamps to skip writes when only timestamp fields change.
 			// This prevents the token refresh loop caused by timestamp/expired/expires_in changes.
-			if metadataEqualIgnoringTimestamps(existing, raw) {
+			if metadataEqualIgnoringTimestamps(existing, raw, auth.Provider) {
 				return path, nil
 			}
 			file, errOpen := os.OpenFile(path, os.O_WRONLY|os.O_TRUNC, 0o600)
@@ -300,28 +299,101 @@ func (s *FileTokenStore) baseDirSnapshot() string {
 	return s.baseDir
 }
 
-// metadataEqualIgnoringTimestamps compares two metadata JSON blobs, ignoring volatile fields that
-// change on every refresh but don't affect authentication logic.
-func metadataEqualIgnoringTimestamps(a, b []byte) bool {
-	var objA map[string]any
-	var objB map[string]any
-	if errUnmarshalA := json.Unmarshal(a, &objA); errUnmarshalA != nil {
+// DEPRECATED: Use metadataEqualIgnoringTimestamps for comparing auth metadata.
+// This function is kept for backward compatibility but can cause refresh loops.
+func jsonEqual(a, b []byte) bool {
+	var objA any
+	var objB any
+	if err := json.Unmarshal(a, &objA); err != nil {
 		return false
 	}
-	if errUnmarshalB := json.Unmarshal(b, &objB); errUnmarshalB != nil {
+	if err := json.Unmarshal(b, &objB); err != nil {
 		return false
 	}
-	stripVolatileMetadataFields(objA)
-	stripVolatileMetadataFields(objB)
-	return reflect.DeepEqual(objA, objB)
+	return deepEqualJSON(objA, objB)
 }
 
-func stripVolatileMetadataFields(metadata map[string]any) {
-	if metadata == nil {
-		return
+// metadataEqualIgnoringTimestamps compares two metadata JSON blobs,
+// ignoring fields that change on every refresh but don't affect functionality.
+// This prevents unnecessary file writes that would trigger watcher events and
+// create refresh loops.
+// The provider parameter controls whether access_token is ignored: providers like
+// Google OAuth (gemini, gemini-cli) can re-fetch tokens when needed, while others
+// like iFlow require the refreshed token to be persisted.
+func metadataEqualIgnoringTimestamps(a, b []byte, provider string) bool {
+	var objA, objB map[string]any
+	if err := json.Unmarshal(a, &objA); err != nil {
+		return false
 	}
-	// These fields change on refresh and would otherwise trigger watcher reload loops.
-	for _, field := range []string{"timestamp", "expired", "expires_in", "last_refresh", "access_token"} {
-		delete(metadata, field)
+	if err := json.Unmarshal(b, &objB); err != nil {
+		return false
+	}
+
+	// Fields to ignore: these change on every refresh but don't affect authentication logic.
+	// - timestamp, expired, expires_in, last_refresh: time-related fields that change on refresh
+	ignoredFields := []string{"timestamp", "expired", "expires_in", "last_refresh"}
+
+	// For providers that can re-fetch tokens when needed (e.g., Google OAuth),
+	// we ignore access_token to avoid unnecessary file writes.
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity":
+		ignoredFields = append(ignoredFields, "access_token")
+	}
+
+	for _, field := range ignoredFields {
+		delete(objA, field)
+		delete(objB, field)
+	}
+
+	return deepEqualJSON(objA, objB)
+}
+
+func deepEqualJSON(a, b any) bool {
+	switch valA := a.(type) {
+	case map[string]any:
+		valB, ok := b.(map[string]any)
+		if !ok || len(valA) != len(valB) {
+			return false
+		}
+		for key, subA := range valA {
+			subB, ok1 := valB[key]
+			if !ok1 || !deepEqualJSON(subA, subB) {
+				return false
+			}
+		}
+		return true
+	case []any:
+		sliceB, ok := b.([]any)
+		if !ok || len(valA) != len(sliceB) {
+			return false
+		}
+		for i := range valA {
+			if !deepEqualJSON(valA[i], sliceB[i]) {
+				return false
+			}
+		}
+		return true
+	case float64:
+		valB, ok := b.(float64)
+		if !ok {
+			return false
+		}
+		return valA == valB
+	case string:
+		valB, ok := b.(string)
+		if !ok {
+			return false
+		}
+		return valA == valB
+	case bool:
+		valB, ok := b.(bool)
+		if !ok {
+			return false
+		}
+		return valA == valB
+	case nil:
+		return b == nil
+	default:
+		return false
 	}
 }

sdk/cliproxy/service.go

@@ -1212,6 +1212,9 @@ func rewriteModelInfoName(name, oldID, newID string) string {
 	if strings.EqualFold(oldID, newID) {
 		return name
 	}
+	if strings.EqualFold(trimmed, oldID) {
+		return newID
+	}
 	if strings.HasSuffix(trimmed, "/"+oldID) {
 		prefix := strings.TrimSuffix(trimmed, oldID)
 		return prefix + newID