Merge upstream/main (08779cc) - sync with original repo updates

internal/api/handlers/management/auth_files.go

@@ -24,6 +24,7 @@ import (
 	"github.com/gin-gonic/gin"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/claude"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot"
 	geminiAuth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/gemini"
 	iflowauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/iflow"
 	kiroauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/kiro"
@@ -1707,7 +1708,7 @@ 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",
@@ -1812,7 +1813,7 @@ 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{
@@ -1843,6 +1844,89 @@ func (h *Handler) RequestIFlowToken(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{"status": "ok", "url": authURL, "state": state})
 }
 
+func (h *Handler) RequestGitHubToken(c *gin.Context) {
+	ctx := context.Background()
+
+	fmt.Println("Initializing GitHub Copilot authentication...")
+
+	state := fmt.Sprintf("gh-%d", time.Now().UnixNano())
+
+	// Initialize Copilot auth service
+	// We need to import "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/copilot" first if not present
+	// Assuming copilot package is imported as "copilot"
+	deviceClient := copilot.NewDeviceFlowClient(h.cfg)
+
+	// Initiate device flow
+	deviceCode, err := deviceClient.RequestDeviceCode(ctx)
+	if err != nil {
+		log.Errorf("Failed to initiate device flow: %v", err)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to initiate device flow"})
+		return
+	}
+
+	authURL := deviceCode.VerificationURI
+	userCode := deviceCode.UserCode
+
+	RegisterOAuthSession(state, "github")
+
+	go func() {
+		fmt.Printf("Please visit %s and enter code: %s\n", authURL, userCode)
+
+		tokenData, errPoll := deviceClient.PollForToken(ctx, deviceCode)
+		if errPoll != nil {
+			SetOAuthSessionError(state, "Authentication failed")
+			fmt.Printf("Authentication failed: %v\n", errPoll)
+			return
+		}
+
+		username, errUser := deviceClient.FetchUserInfo(ctx, tokenData.AccessToken)
+		if errUser != nil {
+			log.Warnf("Failed to fetch user info: %v", errUser)
+			username = "github-user"
+		}
+
+		tokenStorage := &copilot.CopilotTokenStorage{
+			AccessToken: tokenData.AccessToken,
+			TokenType:   tokenData.TokenType,
+			Scope:       tokenData.Scope,
+			Username:    username,
+			Type:        "github-copilot",
+		}
+
+		fileName := fmt.Sprintf("github-%s.json", username)
+		record := &coreauth.Auth{
+			ID:       fileName,
+			Provider: "github",
+			FileName: fileName,
+			Storage:  tokenStorage,
+			Metadata: map[string]any{
+				"email":    username,
+				"username": username,
+			},
+		}
+
+		savedPath, errSave := h.saveTokenRecord(ctx, record)
+		if errSave != nil {
+			log.Errorf("Failed to save authentication tokens: %v", errSave)
+			SetOAuthSessionError(state, "Failed to save authentication tokens")
+			return
+		}
+
+		fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
+		fmt.Println("You can now use GitHub Copilot services through this CLI")
+		CompleteOAuthSession(state)
+		CompleteOAuthSessionsByProvider("github")
+	}()
+
+	c.JSON(200, gin.H{
+		"status":           "ok",
+		"url":              authURL,
+		"state":            state,
+		"user_code":        userCode,
+		"verification_uri": authURL,
+	})
+}
+
 func (h *Handler) RequestIFlowCookieToken(c *gin.Context) {
 	ctx := context.Background()
 
@@ -1897,15 +1981,17 @@ 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,

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/handlers/management/oauth_sessions.go

@@ -238,6 +238,8 @@ func NormalizeOAuthProvider(provider string) (string, error) {
 		return "qwen", nil
 	case "kiro":
 		return "kiro", nil
+	case "github":
+		return "github", nil
 	default:
 		return "", errUnsupportedOAuthFlow
 	}

internal/api/server.go

@@ -262,10 +262,7 @@ 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
 
@@ -649,6 +646,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.GET("/iflow-auth-url", s.mgmt.RequestIFlowToken)
 		mgmt.POST("/iflow-auth-url", s.mgmt.RequestIFlowCookieToken)
 		mgmt.GET("/kiro-auth-url", s.mgmt.RequestKiroToken)
+		mgmt.GET("/github-auth-url", s.mgmt.RequestGitHubToken)
 		mgmt.POST("/oauth-callback", s.mgmt.PostOAuthCallback)
 		mgmt.GET("/get-auth-status", s.mgmt.GetAuthStatus)
 	}

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) {
@@ -122,6 +122,24 @@ 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.
@@ -131,12 +149,7 @@ 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

@@ -765,21 +765,23 @@ 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 +811,9 @@ 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/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())
 	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())
 	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())
 	if err != nil {
 		return resp, err
 	}
@@ -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())
 	if err != nil {
 		return nil, err
 	}
@@ -802,7 +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())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
@@ -1005,9 +1005,6 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 			}
 			modelCfg := modelConfig[modelID]
 			modelName := modelID
-			if modelCfg != nil && modelCfg.Name != "" {
-				modelName = modelCfg.Name
-			}
 			modelInfo := &registry.ModelInfo{
 				ID:          modelID,
 				Name:        modelName,
@@ -1410,13 +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"
@@ -106,7 +105,7 @@ 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())
 	if err != nil {
 		return resp, err
 	}
@@ -119,9 +118,6 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// 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)
@@ -239,7 +235,7 @@ 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())
 	if err != nil {
 		return nil, err
 	}
@@ -250,9 +246,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	// 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 +534,6 @@ 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

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())
 	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())
 	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())
 	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())
 	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())
 	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())
 		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())
 	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())
 	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())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/gemini_vertex_executor.go

@@ -170,7 +170,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	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())
 	if err != nil {
 		return resp, err
 	}
@@ -272,7 +272,7 @@ 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())
 	if err != nil {
 		return resp, err
 	}
@@ -375,7 +375,7 @@ 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())
 	if err != nil {
 		return nil, err
 	}
@@ -494,7 +494,7 @@ 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())
 	if err != nil {
 		return nil, err
 	}
@@ -605,7 +605,7 @@ 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())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
@@ -689,7 +689,7 @@ 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())
 	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")
 	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")
 	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())
 	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())
 	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())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

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())
 	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())
 	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 +61,8 @@ 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)
 //
 // Returns:
 //   - Modified request body JSON with thinking configuration applied
@@ -76,16 +79,21 @@ 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")
 //
 //	// 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")
+func ApplyThinking(body []byte, model string, fromFormat string, toFormat string) ([]byte, error) {
+	providerFormat := strings.ToLower(strings.TrimSpace(toFormat))
+	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
@@ -100,19 +108,19 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// 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 +129,19 @@ 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 +152,17 @@ 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 +175,14 @@ 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 +236,7 @@ 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 +248,57 @@ 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 +315,11 @@ 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"
@@ -388,14 +387,11 @@ 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.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

@@ -343,8 +343,8 @@ 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")

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

@@ -132,6 +132,7 @@ 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 +142,19 @@ 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 +169,10 @@ 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 +216,10 @@ 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"
@@ -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":

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"
@@ -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/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

@@ -129,6 +129,7 @@ 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 +139,19 @@ 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

@@ -129,6 +129,7 @@ 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 +139,19 @@ 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 +166,10 @@ 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 +214,10 @@ 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/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 +77,15 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 			}
 		}
 
-		// Convert thinkingBudget to reasoning_effort
-		// Always perform conversion to support allowCompat models that may not be in registry
+		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
 		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/cliproxy/auth/types.go

@@ -227,6 +227,18 @@ func (a *Auth) AccountInfo() (string, string) {
 		}
 	}
 
+	// For GitHub provider, return username
+	if strings.ToLower(a.Provider) == "github" {
+		if a.Metadata != nil {
+			if username, ok := a.Metadata["username"].(string); ok {
+				username = strings.TrimSpace(username)
+				if username != "" {
+					return "oauth", username
+				}
+			}
+		}
+	}
+
 	// Check metadata for email first (OAuth-style auth)
 	if a.Metadata != nil {
 		if v, ok := a.Metadata["email"].(string); ok {