Merge pull request #23 from router-for-me/plus

v6.6.9

README.md

@@ -11,7 +11,7 @@ The Plus release stays in lockstep with the mainline features.
 ## Differences from the Mainline
 
 - Added GitHub Copilot support (OAuth login), provided by [em4go](https://github.com/em4go/CLIProxyAPI/tree/feature/github-copilot-auth)
-- Added Kiro (AWS CodeWhisperer) support (OAuth login), provided by [fuko2935](https://github.com/fuko2935/CLIProxyAPI/tree/feature/kiro-integration)
+- Added Kiro (AWS CodeWhisperer) support (OAuth login), provided by [fuko2935](https://github.com/fuko2935/CLIProxyAPI/tree/feature/kiro-integration), [Ravens2121](https://github.com/Ravens2121/CLIProxyAPIPlus/)
 
 ## Contributing
 

README_CN.md

@@ -11,7 +11,7 @@
 ## 与主线版本版本差异
 
 - 新增 GitHub Copilot 支持（OAuth 登录），由[em4go](https://github.com/em4go/CLIProxyAPI/tree/feature/github-copilot-auth)提供
-- 新增 Kiro (AWS CodeWhisperer) 支持 (OAuth 登录), 由[fuko2935](https://github.com/fuko2935/CLIProxyAPI/tree/feature/kiro-integration)提供
+- 新增 Kiro (AWS CodeWhisperer) 支持 (OAuth 登录), 由[fuko2935](https://github.com/fuko2935/CLIProxyAPI/tree/feature/kiro-integration)、[Ravens2121](https://github.com/Ravens2121/CLIProxyAPIPlus/)提供
 
 ## 贡献
 

internal/api/handlers/management/auth_files.go

@@ -3,6 +3,9 @@ package management
 import (
 	"bytes"
 	"context"
+	"crypto/rand"
+	"crypto/sha256"
+	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -23,6 +26,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
 	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"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
@@ -2154,9 +2158,35 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 
 func (h *Handler) GetAuthStatus(c *gin.Context) {
 	state := c.Query("state")
-	if err, ok := getOAuthStatus(state); ok {
-		if err != "" {
-			c.JSON(200, gin.H{"status": "error", "error": err})
+	if statusValue, ok := getOAuthStatus(state); ok {
+		if statusValue != "" {
+			// Check for device_code prefix (Kiro AWS Builder ID flow)
+			// Format: "device_code|verification_url|user_code"
+			// Using "|" as separator because URLs contain ":"
+			if strings.HasPrefix(statusValue, "device_code|") {
+				parts := strings.SplitN(statusValue, "|", 3)
+				if len(parts) == 3 {
+					c.JSON(200, gin.H{
+						"status":           "device_code",
+						"verification_url": parts[1],
+						"user_code":        parts[2],
+					})
+					return
+				}
+			}
+			// Check for auth_url prefix (Kiro social auth flow)
+			// Format: "auth_url|url"
+			// Using "|" as separator because URLs contain ":"
+			if strings.HasPrefix(statusValue, "auth_url|") {
+				authURL := strings.TrimPrefix(statusValue, "auth_url|")
+				c.JSON(200, gin.H{
+					"status": "auth_url",
+					"url":    authURL,
+				})
+				return
+			}
+			// Otherwise treat as error
+			c.JSON(200, gin.H{"status": "error", "error": statusValue})
 		} else {
 			c.JSON(200, gin.H{"status": "wait"})
 			return
@@ -2166,3 +2196,295 @@ func (h *Handler) GetAuthStatus(c *gin.Context) {
 	}
 	deleteOAuthStatus(state)
 }
+
+const kiroCallbackPort = 9876
+
+func (h *Handler) RequestKiroToken(c *gin.Context) {
+	ctx := context.Background()
+
+	// Get the login method from query parameter (default: aws for device code flow)
+	method := strings.ToLower(strings.TrimSpace(c.Query("method")))
+	if method == "" {
+		method = "aws"
+	}
+
+	fmt.Println("Initializing Kiro authentication...")
+
+	state := fmt.Sprintf("kiro-%d", time.Now().UnixNano())
+
+	switch method {
+	case "aws", "builder-id":
+		// AWS Builder ID uses device code flow (no callback needed)
+		go func() {
+			ssoClient := kiroauth.NewSSOOIDCClient(h.cfg)
+
+			// Step 1: Register client
+			fmt.Println("Registering client...")
+			regResp, err := ssoClient.RegisterClient(ctx)
+			if err != nil {
+				log.Errorf("Failed to register client: %v", err)
+				setOAuthStatus(state, "Failed to register client")
+				return
+			}
+
+			// Step 2: Start device authorization
+			fmt.Println("Starting device authorization...")
+			authResp, err := ssoClient.StartDeviceAuthorization(ctx, regResp.ClientID, regResp.ClientSecret)
+			if err != nil {
+				log.Errorf("Failed to start device auth: %v", err)
+				setOAuthStatus(state, "Failed to start device authorization")
+				return
+			}
+
+			// Store the verification URL for the frontend to display
+			// Using "|" as separator because URLs contain ":"
+			setOAuthStatus(state, "device_code|"+authResp.VerificationURIComplete+"|"+authResp.UserCode)
+
+			// Step 3: Poll for token
+			fmt.Println("Waiting for authorization...")
+			interval := 5 * time.Second
+			if authResp.Interval > 0 {
+				interval = time.Duration(authResp.Interval) * time.Second
+			}
+			deadline := time.Now().Add(time.Duration(authResp.ExpiresIn) * time.Second)
+
+			for time.Now().Before(deadline) {
+				select {
+				case <-ctx.Done():
+					setOAuthStatus(state, "Authorization cancelled")
+					return
+				case <-time.After(interval):
+					tokenResp, err := ssoClient.CreateToken(ctx, regResp.ClientID, regResp.ClientSecret, authResp.DeviceCode)
+					if err != nil {
+						errStr := err.Error()
+						if strings.Contains(errStr, "authorization_pending") {
+							continue
+						}
+						if strings.Contains(errStr, "slow_down") {
+							interval += 5 * time.Second
+							continue
+						}
+						log.Errorf("Token creation failed: %v", err)
+						setOAuthStatus(state, "Token creation failed")
+						return
+					}
+
+					// Success! Save the token
+					expiresAt := time.Now().Add(time.Duration(tokenResp.ExpiresIn) * time.Second)
+					email := kiroauth.ExtractEmailFromJWT(tokenResp.AccessToken)
+
+					idPart := kiroauth.SanitizeEmailForFilename(email)
+					if idPart == "" {
+						idPart = fmt.Sprintf("%d", time.Now().UnixNano()%100000)
+					}
+
+					now := time.Now()
+					fileName := fmt.Sprintf("kiro-aws-%s.json", idPart)
+
+					record := &coreauth.Auth{
+						ID:       fileName,
+						Provider: "kiro",
+						FileName: fileName,
+						Metadata: map[string]any{
+							"type":          "kiro",
+							"access_token":  tokenResp.AccessToken,
+							"refresh_token": tokenResp.RefreshToken,
+							"expires_at":    expiresAt.Format(time.RFC3339),
+							"auth_method":   "builder-id",
+							"provider":      "AWS",
+							"client_id":     regResp.ClientID,
+							"client_secret": regResp.ClientSecret,
+							"email":         email,
+							"last_refresh":  now.Format(time.RFC3339),
+						},
+					}
+
+					savedPath, errSave := h.saveTokenRecord(ctx, record)
+					if errSave != nil {
+						log.Errorf("Failed to save authentication tokens: %v", errSave)
+						setOAuthStatus(state, "Failed to save authentication tokens")
+						return
+					}
+
+					fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
+					if email != "" {
+						fmt.Printf("Authenticated as: %s\n", email)
+					}
+					deleteOAuthStatus(state)
+					return
+				}
+			}
+
+			setOAuthStatus(state, "Authorization timed out")
+		}()
+
+		// Return immediately with the state for polling
+		c.JSON(200, gin.H{"status": "ok", "state": state, "method": "device_code"})
+
+	case "google", "github":
+		// Social auth uses protocol handler - for WEB UI we use a callback forwarder
+		provider := "Google"
+		if method == "github" {
+			provider = "Github"
+		}
+
+		isWebUI := isWebUIRequest(c)
+		if isWebUI {
+			targetURL, errTarget := h.managementCallbackURL("/kiro/callback")
+			if errTarget != nil {
+				log.WithError(errTarget).Error("failed to compute kiro callback target")
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
+				return
+			}
+			if _, errStart := startCallbackForwarder(kiroCallbackPort, "kiro", targetURL); errStart != nil {
+				log.WithError(errStart).Error("failed to start kiro callback forwarder")
+				c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to start callback server"})
+				return
+			}
+		}
+
+		go func() {
+			if isWebUI {
+				defer stopCallbackForwarder(kiroCallbackPort)
+			}
+
+			socialClient := kiroauth.NewSocialAuthClient(h.cfg)
+
+			// Generate PKCE codes
+			codeVerifier, codeChallenge, err := generateKiroPKCE()
+			if err != nil {
+				log.Errorf("Failed to generate PKCE: %v", err)
+				setOAuthStatus(state, "Failed to generate PKCE")
+				return
+			}
+
+			// Build login URL
+			authURL := fmt.Sprintf("%s/login?idp=%s&redirect_uri=%s&code_challenge=%s&code_challenge_method=S256&state=%s&prompt=select_account",
+				"https://prod.us-east-1.auth.desktop.kiro.dev",
+				provider,
+				url.QueryEscape(kiroauth.KiroRedirectURI),
+				codeChallenge,
+				state,
+			)
+
+			// Store auth URL for frontend
+			// Using "|" as separator because URLs contain ":"
+			setOAuthStatus(state, "auth_url|"+authURL)
+
+			// Wait for callback file
+			waitFile := filepath.Join(h.cfg.AuthDir, fmt.Sprintf(".oauth-kiro-%s.oauth", state))
+			deadline := time.Now().Add(5 * time.Minute)
+
+			for {
+				if time.Now().After(deadline) {
+					log.Error("oauth flow timed out")
+					setOAuthStatus(state, "OAuth flow timed out")
+					return
+				}
+				if data, errR := os.ReadFile(waitFile); errR == nil {
+					var m map[string]string
+					_ = json.Unmarshal(data, &m)
+					_ = os.Remove(waitFile)
+					if errStr := m["error"]; errStr != "" {
+						log.Errorf("Authentication failed: %s", errStr)
+						setOAuthStatus(state, "Authentication failed")
+						return
+					}
+					if m["state"] != state {
+						log.Errorf("State mismatch")
+						setOAuthStatus(state, "State mismatch")
+						return
+					}
+					code := m["code"]
+					if code == "" {
+						log.Error("No authorization code received")
+						setOAuthStatus(state, "No authorization code received")
+						return
+					}
+
+					// Exchange code for tokens
+					tokenReq := &kiroauth.CreateTokenRequest{
+						Code:         code,
+						CodeVerifier: codeVerifier,
+						RedirectURI:  kiroauth.KiroRedirectURI,
+					}
+
+					tokenResp, errToken := socialClient.CreateToken(ctx, tokenReq)
+					if errToken != nil {
+						log.Errorf("Failed to exchange code for tokens: %v", errToken)
+						setOAuthStatus(state, "Failed to exchange code for tokens")
+						return
+					}
+
+					// Save the token
+					expiresIn := tokenResp.ExpiresIn
+					if expiresIn <= 0 {
+						expiresIn = 3600
+					}
+					expiresAt := time.Now().Add(time.Duration(expiresIn) * time.Second)
+					email := kiroauth.ExtractEmailFromJWT(tokenResp.AccessToken)
+
+					idPart := kiroauth.SanitizeEmailForFilename(email)
+					if idPart == "" {
+						idPart = fmt.Sprintf("%d", time.Now().UnixNano()%100000)
+					}
+
+					now := time.Now()
+					fileName := fmt.Sprintf("kiro-%s-%s.json", strings.ToLower(provider), idPart)
+
+					record := &coreauth.Auth{
+						ID:       fileName,
+						Provider: "kiro",
+						FileName: fileName,
+						Metadata: map[string]any{
+							"type":          "kiro",
+							"access_token":  tokenResp.AccessToken,
+							"refresh_token": tokenResp.RefreshToken,
+							"profile_arn":   tokenResp.ProfileArn,
+							"expires_at":    expiresAt.Format(time.RFC3339),
+							"auth_method":   "social",
+							"provider":      provider,
+							"email":         email,
+							"last_refresh":  now.Format(time.RFC3339),
+						},
+					}
+
+					savedPath, errSave := h.saveTokenRecord(ctx, record)
+					if errSave != nil {
+						log.Errorf("Failed to save authentication tokens: %v", errSave)
+						setOAuthStatus(state, "Failed to save authentication tokens")
+						return
+					}
+
+					fmt.Printf("Authentication successful! Token saved to %s\n", savedPath)
+					if email != "" {
+						fmt.Printf("Authenticated as: %s\n", email)
+					}
+					deleteOAuthStatus(state)
+					return
+				}
+				time.Sleep(500 * time.Millisecond)
+			}
+		}()
+
+		setOAuthStatus(state, "")
+		c.JSON(200, gin.H{"status": "ok", "state": state, "method": "social"})
+
+	default:
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid method, use 'aws', 'google', or 'github'"})
+	}
+}
+
+// generateKiroPKCE generates PKCE code verifier and challenge for Kiro OAuth.
+func generateKiroPKCE() (verifier, challenge string, err error) {
+	b := make([]byte, 32)
+	if _, err := io.ReadFull(rand.Reader, b); err != nil {
+		return "", "", fmt.Errorf("failed to generate random bytes: %w", err)
+	}
+	verifier = base64.RawURLEncoding.EncodeToString(b)
+
+	h := sha256.Sum256([]byte(verifier))
+	challenge = base64.RawURLEncoding.EncodeToString(h[:])
+
+	return verifier, challenge, nil
+}

internal/api/modules/amp/proxy.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"net"
 	"net/http"
 	"net/http/httputil"
 	"net/url"
@@ -64,7 +65,15 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 	// Modify incoming responses to handle gzip without Content-Encoding
 	// This addresses the same issue as inline handler gzip handling, but at the proxy level
 	proxy.ModifyResponse = func(resp *http.Response) error {
-		// Only process successful responses
+		// Log upstream error responses for diagnostics (502, 503, etc.)
+		// These are NOT proxy connection errors - the upstream responded with an error status
+		if resp.StatusCode >= 500 {
+			log.Errorf("amp upstream responded with error [%d] for %s %s", resp.StatusCode, resp.Request.Method, resp.Request.URL.Path)
+		} else if resp.StatusCode >= 400 {
+			log.Warnf("amp upstream responded with client error [%d] for %s %s", resp.StatusCode, resp.Request.Method, resp.Request.URL.Path)
+		}
+
+		// Only process successful responses for gzip decompression
 		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 			return nil
 		}
@@ -148,15 +157,29 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 		return nil
 	}
 
-	// Error handler for proxy failures
+	// Error handler for proxy failures with detailed error classification for diagnostics
 	proxy.ErrorHandler = func(rw http.ResponseWriter, req *http.Request, err error) {
-		// Check if this is a client-side cancellation (normal behavior)
+		// Classify the error type for better diagnostics
+		var errType string
+		if errors.Is(err, context.DeadlineExceeded) {
+			errType = "timeout"
+		} else if errors.Is(err, context.Canceled) {
+			errType = "canceled"
+		} else if netErr, ok := err.(net.Error); ok && netErr.Timeout() {
+			errType = "dial_timeout"
+		} else if _, ok := err.(net.Error); ok {
+			errType = "network_error"
+		} else {
+			errType = "connection_error"
+		}
+
 		// Don't log as error for context canceled - it's usually client closing connection
 		if errors.Is(err, context.Canceled) {
-			log.Debugf("amp upstream proxy: client canceled request for %s %s", req.Method, req.URL.Path)
+			log.Debugf("amp upstream proxy [%s]: client canceled request for %s %s", errType, req.Method, req.URL.Path)
 		} else {
-			log.Errorf("amp upstream proxy error for %s %s: %v", req.Method, req.URL.Path, err)
+			log.Errorf("amp upstream proxy error [%s] for %s %s: %v", errType, req.Method, req.URL.Path, err)
 		}
+
 		rw.Header().Set("Content-Type", "application/json")
 		rw.WriteHeader(http.StatusBadGateway)
 		_, _ = rw.Write([]byte(`{"error":"amp_upstream_proxy_error","message":"Failed to reach Amp upstream"}`))

internal/api/server.go

@@ -421,6 +421,18 @@ func (s *Server) setupRoutes() {
 		c.String(http.StatusOK, oauthCallbackSuccessHTML)
 	})
 
+	s.engine.GET("/kiro/callback", func(c *gin.Context) {
+		code := c.Query("code")
+		state := c.Query("state")
+		errStr := c.Query("error")
+		if state != "" {
+			file := fmt.Sprintf("%s/.oauth-kiro-%s.oauth", s.cfg.AuthDir, state)
+			_ = os.WriteFile(file, []byte(fmt.Sprintf(`{"code":"%s","state":"%s","error":"%s"}`, code, state, errStr)), 0o600)
+		}
+		c.Header("Content-Type", "text/html; charset=utf-8")
+		c.String(http.StatusOK, oauthCallbackSuccessHTML)
+	})
+
 	// Management routes are registered lazily by registerManagementRoutes when a secret is configured.
 }
 
@@ -586,6 +598,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.GET("/qwen-auth-url", s.mgmt.RequestQwenToken)
 		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("/get-auth-status", s.mgmt.GetAuthStatus)
 	}
 }

internal/auth/kiro/social_auth.go

@@ -126,8 +126,8 @@ func (c *SocialAuthClient) buildLoginURL(provider, redirectURI, codeChallenge, s
 	)
 }
 
-// createToken exchanges the authorization code for tokens.
-func (c *SocialAuthClient) createToken(ctx context.Context, req *CreateTokenRequest) (*SocialTokenResponse, error) {
+// CreateToken exchanges the authorization code for tokens.
+func (c *SocialAuthClient) CreateToken(ctx context.Context, req *CreateTokenRequest) (*SocialTokenResponse, error) {
 	body, err := json.Marshal(req)
 	if err != nil {
 		return nil, fmt.Errorf("failed to marshal token request: %w", err)
@@ -326,7 +326,7 @@ func (c *SocialAuthClient) LoginWithSocial(ctx context.Context, provider SocialP
 		RedirectURI:  KiroRedirectURI,
 	}
 
-	tokenResp, err := c.createToken(ctx, tokenReq)
+	tokenResp, err := c.CreateToken(ctx, tokenReq)
 	if err != nil {
 		return nil, fmt.Errorf("failed to exchange code for tokens: %w", err)
 	}

internal/registry/model_definitions.go

@@ -895,6 +895,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Opus 4.5 via Kiro (2.2x credit)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-sonnet-4-5",
@@ -906,6 +907,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Sonnet 4.5 via Kiro (1.3x credit)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-sonnet-4",
@@ -917,6 +919,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Sonnet 4 via Kiro (1.3x credit)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-haiku-4-5",
@@ -928,6 +931,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Haiku 4.5 via Kiro (0.4x credit)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		// --- Agentic Variants (Optimized for coding agents with chunked writes) ---
 		{
@@ -940,6 +944,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Opus 4.5 optimized for coding agents (chunked writes)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-sonnet-4-5-agentic",
@@ -951,6 +956,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Sonnet 4.5 optimized for coding agents (chunked writes)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-sonnet-4-agentic",
@@ -962,6 +968,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Sonnet 4 optimized for coding agents (chunked writes)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 		{
 			ID:                  "kiro-claude-haiku-4-5-agentic",
@@ -973,6 +980,7 @@ func GetKiroModels() []*ModelInfo {
 			Description:         "Claude Haiku 4.5 optimized for coding agents (chunked writes)",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 32000, ZeroAllowed: true, DynamicAllowed: true},
 		},
 	}
 }

internal/registry/model_registry.go

@@ -748,7 +748,8 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 		}
 		return result
 
-	case "claude":
+	case "claude", "kiro", "antigravity":
+		// Claude, Kiro, and Antigravity all use Claude-compatible format for Claude Code client
 		result := map[string]any{
 			"id":       model.ID,
 			"object":   "model",
@@ -763,6 +764,19 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 		if model.DisplayName != "" {
 			result["display_name"] = model.DisplayName
 		}
+		// Add thinking support for Claude Code client
+		// Claude Code checks for "thinking" field (simple boolean) to enable tab toggle
+		// Also add "extended_thinking" for detailed budget info
+		if model.Thinking != nil {
+			result["thinking"] = true
+			result["extended_thinking"] = map[string]any{
+				"supported":       true,
+				"min":             model.Thinking.Min,
+				"max":             model.Thinking.Max,
+				"zero_allowed":    model.Thinking.ZeroAllowed,
+				"dynamic_allowed": model.Thinking.DynamicAllowed,
+			}
+		}
 		return result
 
 	case "gemini":

internal/runtime/executor/codex_executor.go

@@ -54,8 +54,8 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -152,8 +152,8 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
@@ -254,7 +254,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 
 	modelForCounting := req.Model
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
 	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)

internal/runtime/executor/iflow_executor.go

@@ -57,12 +57,12 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -148,12 +148,12 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}

internal/runtime/executor/openai_compat_executor.go

@@ -54,16 +54,18 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), opts.Stream)
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+	modelOverride := e.resolveUpstreamModel(req.Model, auth)
+	if modelOverride != "" {
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
-	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
+	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel)
+	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
 	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -148,16 +150,18 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+	modelOverride := e.resolveUpstreamModel(req.Model, auth)
+	if modelOverride != "" {
 		translated = e.overrideModel(translated, modelOverride)
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
-	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort")
+	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
+	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
-	if upstreamModel != "" {
+	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel)
+	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
 	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
@@ -323,6 +327,27 @@ func (e *OpenAICompatExecutor) resolveUpstreamModel(alias string, auth *cliproxy
 	return ""
 }
 
+func (e *OpenAICompatExecutor) allowCompatReasoningEffort(model string, auth *cliproxyauth.Auth) bool {
+	trimmed := strings.TrimSpace(model)
+	if trimmed == "" || e == nil || e.cfg == nil {
+		return false
+	}
+	compat := e.resolveCompatConfig(auth)
+	if compat == nil || len(compat.Models) == 0 {
+		return false
+	}
+	for i := range compat.Models {
+		entry := compat.Models[i]
+		if strings.EqualFold(strings.TrimSpace(entry.Alias), trimmed) {
+			return true
+		}
+		if strings.EqualFold(strings.TrimSpace(entry.Name), trimmed) {
+			return true
+		}
+	}
+	return false
+}
+
 func (e *OpenAICompatExecutor) resolveCompatConfig(auth *cliproxyauth.Auth) *config.OpenAICompatibility {
 	if auth == nil || e.cfg == nil {
 		return nil

internal/runtime/executor/payload_helpers.go

@@ -48,19 +48,35 @@ func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model str
 // applyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
 // Metadata values take precedence over any existing field when the model supports thinking, intentionally
 // overwriting caller-provided values to honor suffix/default metadata priority.
-func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string) []byte {
+func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
 	if len(metadata) == 0 {
 		return payload
 	}
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
 	if field == "" {
 		return payload
 	}
+	baseModel := util.ResolveOriginalModel(model, metadata)
+	if baseModel == "" {
+		baseModel = model
+	}
+	if !util.ModelSupportsThinking(baseModel) && !allowCompat {
+		return payload
+	}
 	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-			return updated
+		if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
+			if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+				return updated
+			}
+		}
+	}
+	// Fallback: numeric thinking_budget suffix for level-based (OpenAI-style) models.
+	if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
+		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+			if effort, ok := util.OpenAIThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
+				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+					return updated
+				}
+			}
 		}
 	}
 	return payload
@@ -219,31 +235,40 @@ func matchModelPattern(pattern, model string) bool {
 // normalizeThinkingConfig normalizes thinking-related fields in the payload
 // based on model capabilities. For models without thinking support, it strips
 // reasoning fields. For models with level-based thinking, it validates and
-// normalizes the reasoning effort level.
-func normalizeThinkingConfig(payload []byte, model string) []byte {
+// normalizes the reasoning effort level. For models with numeric budget thinking,
+// it strips the effort string fields.
+func normalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
 	if len(payload) == 0 || model == "" {
 		return payload
 	}
 
 	if !util.ModelSupportsThinking(model) {
-		return stripThinkingFields(payload)
+		if allowCompat {
+			return payload
+		}
+		return stripThinkingFields(payload, false)
 	}
 
 	if util.ModelUsesThinkingLevels(model) {
 		return normalizeReasoningEffortLevel(payload, model)
 	}
 
-	return payload
+	// Model supports thinking but uses numeric budgets, not levels.
+	// Strip effort string fields since they are not applicable.
+	return stripThinkingFields(payload, true)
 }
 
 // stripThinkingFields removes thinking-related fields from the payload for
-// models that do not support thinking.
-func stripThinkingFields(payload []byte) []byte {
+// models that do not support thinking. If effortOnly is true, only removes
+// effort string fields (for models using numeric budgets).
+func stripThinkingFields(payload []byte, effortOnly bool) []byte {
 	fieldsToRemove := []string{
-		"reasoning",
 		"reasoning_effort",
 		"reasoning.effort",
 	}
+	if !effortOnly {
+		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
+	}
 	out := payload
 	for _, field := range fieldsToRemove {
 		if gjson.GetBytes(out, field).Exists() {

internal/runtime/executor/qwen_executor.go

@@ -51,12 +51,12 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
@@ -131,12 +131,12 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
 
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort")
+	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel)
+	body = normalizeThinkingConfig(body, upstreamModel, false)
 	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}

internal/runtime/executor/token_helpers.go

@@ -2,43 +2,107 @@ package executor
 
 import (
 	"fmt"
+	"regexp"
+	"strconv"
 	"strings"
+	"sync"
 
 	"github.com/tidwall/gjson"
 	"github.com/tiktoken-go/tokenizer"
 )
 
+// tokenizerCache stores tokenizer instances to avoid repeated creation
+var tokenizerCache sync.Map
+
+// TokenizerWrapper wraps a tokenizer codec with an adjustment factor for models
+// where tiktoken may not accurately estimate token counts (e.g., Claude models)
+type TokenizerWrapper struct {
+	Codec            tokenizer.Codec
+	AdjustmentFactor float64 // 1.0 means no adjustment, >1.0 means tiktoken underestimates
+}
+
+// Count returns the token count with adjustment factor applied
+func (tw *TokenizerWrapper) Count(text string) (int, error) {
+	count, err := tw.Codec.Count(text)
+	if err != nil {
+		return 0, err
+	}
+	if tw.AdjustmentFactor != 1.0 && tw.AdjustmentFactor > 0 {
+		return int(float64(count) * tw.AdjustmentFactor), nil
+	}
+	return count, nil
+}
+
+// getTokenizer returns a cached tokenizer for the given model.
+// This improves performance by avoiding repeated tokenizer creation.
+func getTokenizer(model string) (*TokenizerWrapper, error) {
+	// Check cache first
+	if cached, ok := tokenizerCache.Load(model); ok {
+		return cached.(*TokenizerWrapper), nil
+	}
+
+	// Cache miss, create new tokenizer
+	wrapper, err := tokenizerForModel(model)
+	if err != nil {
+		return nil, err
+	}
+
+	// Store in cache (use LoadOrStore to handle race conditions)
+	actual, _ := tokenizerCache.LoadOrStore(model, wrapper)
+	return actual.(*TokenizerWrapper), nil
+}
+
 // tokenizerForModel returns a tokenizer codec suitable for an OpenAI-style model id.
-func tokenizerForModel(model string) (tokenizer.Codec, error) {
+// For Claude models, applies a 1.1 adjustment factor since tiktoken may underestimate.
+func tokenizerForModel(model string) (*TokenizerWrapper, error) {
 	sanitized := strings.ToLower(strings.TrimSpace(model))
+
+	// Claude models use cl100k_base with 1.1 adjustment factor
+	// because tiktoken may underestimate Claude's actual token count
+	if strings.Contains(sanitized, "claude") || strings.HasPrefix(sanitized, "kiro-") || strings.HasPrefix(sanitized, "amazonq-") {
+		enc, err := tokenizer.Get(tokenizer.Cl100kBase)
+		if err != nil {
+			return nil, err
+		}
+		return &TokenizerWrapper{Codec: enc, AdjustmentFactor: 1.1}, nil
+	}
+
+	var enc tokenizer.Codec
+	var err error
+
 	switch {
 	case sanitized == "":
-		return tokenizer.Get(tokenizer.Cl100kBase)
+		enc, err = tokenizer.Get(tokenizer.Cl100kBase)
 	case strings.HasPrefix(sanitized, "gpt-5"):
-		return tokenizer.ForModel(tokenizer.GPT5)
+		enc, err = tokenizer.ForModel(tokenizer.GPT5)
 	case strings.HasPrefix(sanitized, "gpt-5.1"):
-		return tokenizer.ForModel(tokenizer.GPT5)
+		enc, err = tokenizer.ForModel(tokenizer.GPT5)
 	case strings.HasPrefix(sanitized, "gpt-4.1"):
-		return tokenizer.ForModel(tokenizer.GPT41)
+		enc, err = tokenizer.ForModel(tokenizer.GPT41)
 	case strings.HasPrefix(sanitized, "gpt-4o"):
-		return tokenizer.ForModel(tokenizer.GPT4o)
+		enc, err = tokenizer.ForModel(tokenizer.GPT4o)
 	case strings.HasPrefix(sanitized, "gpt-4"):
-		return tokenizer.ForModel(tokenizer.GPT4)
+		enc, err = tokenizer.ForModel(tokenizer.GPT4)
 	case strings.HasPrefix(sanitized, "gpt-3.5"), strings.HasPrefix(sanitized, "gpt-3"):
-		return tokenizer.ForModel(tokenizer.GPT35Turbo)
+		enc, err = tokenizer.ForModel(tokenizer.GPT35Turbo)
 	case strings.HasPrefix(sanitized, "o1"):
-		return tokenizer.ForModel(tokenizer.O1)
+		enc, err = tokenizer.ForModel(tokenizer.O1)
 	case strings.HasPrefix(sanitized, "o3"):
-		return tokenizer.ForModel(tokenizer.O3)
+		enc, err = tokenizer.ForModel(tokenizer.O3)
 	case strings.HasPrefix(sanitized, "o4"):
-		return tokenizer.ForModel(tokenizer.O4Mini)
+		enc, err = tokenizer.ForModel(tokenizer.O4Mini)
 	default:
-		return tokenizer.Get(tokenizer.O200kBase)
+		enc, err = tokenizer.Get(tokenizer.O200kBase)
 	}
+
+	if err != nil {
+		return nil, err
+	}
+	return &TokenizerWrapper{Codec: enc, AdjustmentFactor: 1.0}, nil
 }
 
 // countOpenAIChatTokens approximates prompt tokens for OpenAI chat completions payloads.
-func countOpenAIChatTokens(enc tokenizer.Codec, payload []byte) (int64, error) {
+func countOpenAIChatTokens(enc *TokenizerWrapper, payload []byte) (int64, error) {
 	if enc == nil {
 		return 0, fmt.Errorf("encoder is nil")
 	}
@@ -62,11 +126,206 @@ func countOpenAIChatTokens(enc tokenizer.Codec, payload []byte) (int64, error) {
 		return 0, nil
 	}
 
+	// Count text tokens
 	count, err := enc.Count(joined)
 	if err != nil {
 		return 0, err
 	}
-	return int64(count), nil
+
+	// Extract and add image tokens from placeholders
+	imageTokens := extractImageTokens(joined)
+
+	return int64(count) + int64(imageTokens), nil
+}
+
+// countClaudeChatTokens approximates prompt tokens for Claude API chat completions payloads.
+// This handles Claude's message format with system, messages, and tools.
+// Image tokens are estimated based on image dimensions when available.
+func countClaudeChatTokens(enc *TokenizerWrapper, payload []byte) (int64, error) {
+	if enc == nil {
+		return 0, fmt.Errorf("encoder is nil")
+	}
+	if len(payload) == 0 {
+		return 0, nil
+	}
+
+	root := gjson.ParseBytes(payload)
+	segments := make([]string, 0, 32)
+
+	// Collect system prompt (can be string or array of content blocks)
+	collectClaudeSystem(root.Get("system"), &segments)
+	
+	// Collect messages
+	collectClaudeMessages(root.Get("messages"), &segments)
+	
+	// Collect tools
+	collectClaudeTools(root.Get("tools"), &segments)
+
+	joined := strings.TrimSpace(strings.Join(segments, "\n"))
+	if joined == "" {
+		return 0, nil
+	}
+
+	// Count text tokens
+	count, err := enc.Count(joined)
+	if err != nil {
+		return 0, err
+	}
+
+	// Extract and add image tokens from placeholders
+	imageTokens := extractImageTokens(joined)
+
+	return int64(count) + int64(imageTokens), nil
+}
+
+// imageTokenPattern matches [IMAGE:xxx tokens] format for extracting estimated image tokens
+var imageTokenPattern = regexp.MustCompile(`\[IMAGE:(\d+) tokens\]`)
+
+// extractImageTokens extracts image token estimates from placeholder text.
+// Placeholders are in the format [IMAGE:xxx tokens] where xxx is the estimated token count.
+func extractImageTokens(text string) int {
+	matches := imageTokenPattern.FindAllStringSubmatch(text, -1)
+	total := 0
+	for _, match := range matches {
+		if len(match) > 1 {
+			if tokens, err := strconv.Atoi(match[1]); err == nil {
+				total += tokens
+			}
+		}
+	}
+	return total
+}
+
+// estimateImageTokens calculates estimated tokens for an image based on dimensions.
+// Based on Claude's image token calculation: tokens ≈ (width * height) / 750
+// Minimum 85 tokens, maximum 1590 tokens (for 1568x1568 images).
+func estimateImageTokens(width, height float64) int {
+	if width <= 0 || height <= 0 {
+		// No valid dimensions, use default estimate (medium-sized image)
+		return 1000
+	}
+
+	tokens := int(width * height / 750)
+
+	// Apply bounds
+	if tokens < 85 {
+		tokens = 85
+	}
+	if tokens > 1590 {
+		tokens = 1590
+	}
+
+	return tokens
+}
+
+// collectClaudeSystem extracts text from Claude's system field.
+// System can be a string or an array of content blocks.
+func collectClaudeSystem(system gjson.Result, segments *[]string) {
+	if !system.Exists() {
+		return
+	}
+	if system.Type == gjson.String {
+		addIfNotEmpty(segments, system.String())
+		return
+	}
+	if system.IsArray() {
+		system.ForEach(func(_, block gjson.Result) bool {
+			blockType := block.Get("type").String()
+			if blockType == "text" || blockType == "" {
+				addIfNotEmpty(segments, block.Get("text").String())
+			}
+			// Also handle plain string blocks
+			if block.Type == gjson.String {
+				addIfNotEmpty(segments, block.String())
+			}
+			return true
+		})
+	}
+}
+
+// collectClaudeMessages extracts text from Claude's messages array.
+func collectClaudeMessages(messages gjson.Result, segments *[]string) {
+	if !messages.Exists() || !messages.IsArray() {
+		return
+	}
+	messages.ForEach(func(_, message gjson.Result) bool {
+		addIfNotEmpty(segments, message.Get("role").String())
+		collectClaudeContent(message.Get("content"), segments)
+		return true
+	})
+}
+
+// collectClaudeContent extracts text from Claude's content field.
+// Content can be a string or an array of content blocks.
+// For images, estimates token count based on dimensions when available.
+func collectClaudeContent(content gjson.Result, segments *[]string) {
+	if !content.Exists() {
+		return
+	}
+	if content.Type == gjson.String {
+		addIfNotEmpty(segments, content.String())
+		return
+	}
+	if content.IsArray() {
+		content.ForEach(func(_, part gjson.Result) bool {
+			partType := part.Get("type").String()
+			switch partType {
+			case "text":
+				addIfNotEmpty(segments, part.Get("text").String())
+			case "image":
+				// Estimate image tokens based on dimensions if available
+				source := part.Get("source")
+				if source.Exists() {
+					width := source.Get("width").Float()
+					height := source.Get("height").Float()
+					if width > 0 && height > 0 {
+						tokens := estimateImageTokens(width, height)
+						addIfNotEmpty(segments, fmt.Sprintf("[IMAGE:%d tokens]", tokens))
+					} else {
+						// No dimensions available, use default estimate
+						addIfNotEmpty(segments, "[IMAGE:1000 tokens]")
+					}
+				} else {
+					// No source info, use default estimate
+					addIfNotEmpty(segments, "[IMAGE:1000 tokens]")
+				}
+			case "tool_use":
+				addIfNotEmpty(segments, part.Get("id").String())
+				addIfNotEmpty(segments, part.Get("name").String())
+				if input := part.Get("input"); input.Exists() {
+					addIfNotEmpty(segments, input.Raw)
+				}
+			case "tool_result":
+				addIfNotEmpty(segments, part.Get("tool_use_id").String())
+				collectClaudeContent(part.Get("content"), segments)
+			case "thinking":
+				addIfNotEmpty(segments, part.Get("thinking").String())
+			default:
+				// For unknown types, try to extract any text content
+				if part.Type == gjson.String {
+					addIfNotEmpty(segments, part.String())
+				} else if part.Type == gjson.JSON {
+					addIfNotEmpty(segments, part.Raw)
+				}
+			}
+			return true
+		})
+	}
+}
+
+// collectClaudeTools extracts text from Claude's tools array.
+func collectClaudeTools(tools gjson.Result, segments *[]string) {
+	if !tools.Exists() || !tools.IsArray() {
+		return
+	}
+	tools.ForEach(func(_, tool gjson.Result) bool {
+		addIfNotEmpty(segments, tool.Get("name").String())
+		addIfNotEmpty(segments, tool.Get("description").String())
+		if inputSchema := tool.Get("input_schema"); inputSchema.Exists() {
+			addIfNotEmpty(segments, inputSchema.Raw)
+		}
+		return true
+	})
 }
 
 // buildOpenAIUsageJSON returns a minimal usage structure understood by downstream translators.

internal/util/claude_thinking.go

@@ -28,6 +28,9 @@ func ApplyClaudeThinkingConfig(body []byte, budget *int) []byte {
 // It uses the unified ResolveThinkingConfigFromMetadata and normalizes the budget.
 // Returns the normalized budget (nil if thinking should not be enabled) and whether it matched.
 func ResolveClaudeThinkingConfig(modelName string, metadata map[string]any) (*int, bool) {
+	if !ModelSupportsThinking(modelName) {
+		return nil, false
+	}
 	budget, include, matched := ResolveThinkingConfigFromMetadata(modelName, metadata)
 	if !matched {
 		return nil, false

internal/util/gemini_thinking.go

@@ -25,9 +25,15 @@ func ApplyGeminiThinkingConfig(body []byte, budget *int, includeThoughts *bool)
 			updated = rewritten
 		}
 	}
-	if includeThoughts != nil {
+	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && budget != nil && *budget != 0 {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
 		valuePath := "generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *includeThoughts)
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
 		if err == nil {
 			updated = rewritten
 		}
@@ -47,9 +53,15 @@ func ApplyGeminiCLIThinkingConfig(body []byte, budget *int, includeThoughts *boo
 			updated = rewritten
 		}
 	}
-	if includeThoughts != nil {
+	// Default to including thoughts when a budget override is present but no explicit include flag is provided.
+	incl := includeThoughts
+	if incl == nil && budget != nil && *budget != 0 {
+		defaultInclude := true
+		incl = &defaultInclude
+	}
+	if incl != nil {
 		valuePath := "request.generationConfig.thinkingConfig.include_thoughts"
-		rewritten, err := sjson.SetBytes(updated, valuePath, *includeThoughts)
+		rewritten, err := sjson.SetBytes(updated, valuePath, *incl)
 		if err == nil {
 			updated = rewritten
 		}

internal/util/openai_thinking.go

@@ -0,0 +1,34 @@
+package util
+
+// OpenAIThinkingBudgetToEffort maps a numeric thinking budget (tokens)
+// into an OpenAI-style reasoning effort level for level-based models.
+//
+// Ranges:
+//   - 0            -> "none"
+//   - 1..1024      -> "low"
+//   - 1025..8192   -> "medium"
+//   - 8193..24576  -> "high"
+//   - 24577..      -> highest supported level for the model (defaults to "xhigh")
+//
+// Negative values (except the dynamic -1 handled elsewhere) are treated as unsupported.
+func OpenAIThinkingBudgetToEffort(model string, budget int) (string, bool) {
+	switch {
+	case budget < 0:
+		return "", false
+	case budget == 0:
+		return "none", true
+	case budget > 0 && budget <= 1024:
+		return "low", true
+	case budget <= 8192:
+		return "medium", true
+	case budget <= 24576:
+		return "high", true
+	case budget > 24576:
+		if levels := GetModelThinkingLevels(model); len(levels) > 0 {
+			return levels[len(levels)-1], true
+		}
+		return "xhigh", true
+	default:
+		return "", false
+	}
+}

internal/util/thinking.go

@@ -25,33 +25,33 @@ func ModelSupportsThinking(model string) bool {
 // or min (0 if zero is allowed and mid <= 0).
 func NormalizeThinkingBudget(model string, budget int) int {
 	if budget == -1 { // dynamic
-		if found, min, max, zeroAllowed, dynamicAllowed := thinkingRangeFromRegistry(model); found {
+		if found, minBudget, maxBudget, zeroAllowed, dynamicAllowed := thinkingRangeFromRegistry(model); found {
 			if dynamicAllowed {
 				return -1
 			}
-			mid := (min + max) / 2
+			mid := (minBudget + maxBudget) / 2
 			if mid <= 0 && zeroAllowed {
 				return 0
 			}
 			if mid <= 0 {
-				return min
+				return minBudget
 			}
 			return mid
 		}
 		return -1
 	}
-	if found, min, max, zeroAllowed, _ := thinkingRangeFromRegistry(model); found {
+	if found, minBudget, maxBudget, zeroAllowed, _ := thinkingRangeFromRegistry(model); found {
 		if budget == 0 {
 			if zeroAllowed {
 				return 0
 			}
-			return min
+			return minBudget
 		}
-		if budget < min {
-			return min
+		if budget < minBudget {
+			return minBudget
 		}
-		if budget > max {
-			return max
+		if budget > maxBudget {
+			return maxBudget
 		}
 		return budget
 	}
@@ -105,3 +105,16 @@ func NormalizeReasoningEffortLevel(model, effort string) (string, bool) {
 	}
 	return "", false
 }
+
+// IsOpenAICompatibilityModel reports whether the model is registered as an OpenAI-compatibility model.
+// These models may not advertise Thinking metadata in the registry.
+func IsOpenAICompatibilityModel(model string) bool {
+	if model == "" {
+		return false
+	}
+	info := registry.GetGlobalRegistry().GetModelInfo(model)
+	if info == nil {
+		return false
+	}
+	return strings.EqualFold(strings.TrimSpace(info.Type), "openai-compatibility")
+}

internal/util/thinking_suffix.go

@@ -163,6 +163,11 @@ func ResolveThinkingConfigFromMetadata(model string, metadata map[string]any) (*
 	if !matched {
 		return nil, nil, false
 	}
+	// Level-based models (OpenAI-style) do not accept numeric thinking budgets in
+	// Claude/Gemini-style protocols, so we don't derive budgets for them here.
+	if ModelUsesThinkingLevels(model) {
+		return nil, nil, false
+	}
 
 	if budget == nil && effort != nil {
 		if derived, ok := ThinkingEffortToBudget(model, *effort); ok {

sdk/api/handlers/claude/code_handlers.go

@@ -7,7 +7,6 @@
 package claude
 
 import (
-	"bufio"
 	"bytes"
 	"compress/gzip"
 	"context"
@@ -219,52 +218,24 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 }
 
 func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.Flusher, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
-	// v6.1: Intelligent Buffered Streamer strategy
-	// Enhanced buffering with larger buffer size (16KB) and longer flush interval (120ms).
-	// Smart flush only when buffer is sufficiently filled (≥50%), dramatically reducing
-	// flush frequency from ~12.5Hz to ~5-8Hz while maintaining low latency.
-	writer := bufio.NewWriterSize(c.Writer, 16*1024) // 4KB → 16KB
-	ticker := time.NewTicker(120 * time.Millisecond) // 80ms → 120ms
-	defer ticker.Stop()
-
-	var chunkIdx int
-
+	// OpenAI-style stream forwarding: write each SSE chunk and flush immediately.
+	// This guarantees clients see incremental output even for small responses.
 	for {
 		select {
 		case <-c.Request.Context().Done():
-			// Context cancelled, flush any remaining data before exit
-			_ = writer.Flush()
 			cancel(c.Request.Context().Err())
 			return
 
-		case <-ticker.C:
-			// Smart flush: only flush when buffer has sufficient data (≥50% full)
-			// This reduces flush frequency while ensuring data flows naturally
-			buffered := writer.Buffered()
-			if buffered >= 8*1024 { // At least 8KB (50% of 16KB buffer)
-				if err := writer.Flush(); err != nil {
-					// Error flushing, cancel and return
-					cancel(err)
-					return
-				}
-				flusher.Flush() // Also flush the underlying http.ResponseWriter
-			}
-
 		case chunk, ok := <-data:
 			if !ok {
-				// Stream ended, flush remaining data
-				_ = writer.Flush()
+				flusher.Flush()
 				cancel(nil)
 				return
 			}
-
-			// Forward the complete SSE event block directly (already formatted by the translator).
-			// The translator returns a complete SSE-compliant event block, including event:, data:, and separators.
-			// The handler just needs to forward it without reassembly.
 			if len(chunk) > 0 {
-				_, _ = writer.Write(chunk)
+				_, _ = c.Writer.Write(chunk)
+				flusher.Flush()
 			}
-			chunkIdx++
 
 		case errMsg, ok := <-errs:
 			if !ok {
@@ -276,21 +247,20 @@ func (h *ClaudeCodeAPIHandler) forwardClaudeStream(c *gin.Context, flusher http.
 					status = errMsg.StatusCode
 				}
 				c.Status(status)
+
 				// An error occurred: emit as a proper SSE error event
 				errorBytes, _ := json.Marshal(h.toClaudeError(errMsg))
-				_, _ = writer.WriteString("event: error\n")
-				_, _ = writer.WriteString("data: ")
-				_, _ = writer.Write(errorBytes)
-				_, _ = writer.WriteString("\n\n")
-				_ = writer.Flush()
+				_, _ = fmt.Fprintf(c.Writer, "event: error\ndata: %s\n\n", errorBytes)
 				flusher.Flush()
 			}
+
 			var execErr error
 			if errMsg != nil {
 				execErr = errMsg.Error
 			}
 			cancel(execErr)
 			return
+		case <-time.After(500 * time.Millisecond):
 		}
 	}
 }

sdk/auth/kiro.go

@@ -47,8 +47,9 @@ func (a *KiroAuthenticator) Provider() string {
 }
 
 // RefreshLead indicates how soon before expiry a refresh should be attempted.
+// Set to 5 minutes to match Antigravity and avoid frequent refresh checks while still ensuring timely token refresh.
 func (a *KiroAuthenticator) RefreshLead() *time.Duration {
-	d := 30 * time.Minute
+	d := 5 * time.Minute
 	return &d
 }
 
@@ -103,7 +104,8 @@ func (a *KiroAuthenticator) Login(ctx context.Context, cfg *config.Config, opts
 			"source":      "aws-builder-id",
 			"email":       tokenData.Email,
 		},
-		NextRefreshAfter: expiresAt.Add(-30 * time.Minute),
+		// NextRefreshAfter is aligned with RefreshLead (5min)
+		NextRefreshAfter: expiresAt.Add(-5 * time.Minute),
 	}
 
 	if tokenData.Email != "" {
@@ -165,7 +167,8 @@ func (a *KiroAuthenticator) LoginWithGoogle(ctx context.Context, cfg *config.Con
 			"source":      "google-oauth",
 			"email":       tokenData.Email,
 		},
-		NextRefreshAfter: expiresAt.Add(-30 * time.Minute),
+		// NextRefreshAfter is aligned with RefreshLead (5min)
+		NextRefreshAfter: expiresAt.Add(-5 * time.Minute),
 	}
 
 	if tokenData.Email != "" {
@@ -227,7 +230,8 @@ func (a *KiroAuthenticator) LoginWithGitHub(ctx context.Context, cfg *config.Con
 			"source":      "github-oauth",
 			"email":       tokenData.Email,
 		},
-		NextRefreshAfter: expiresAt.Add(-30 * time.Minute),
+		// NextRefreshAfter is aligned with RefreshLead (5min)
+		NextRefreshAfter: expiresAt.Add(-5 * time.Minute),
 	}
 
 	if tokenData.Email != "" {
@@ -291,7 +295,8 @@ func (a *KiroAuthenticator) ImportFromKiroIDE(ctx context.Context, cfg *config.C
 			"source":      "kiro-ide-import",
 			"email":       tokenData.Email,
 		},
-		NextRefreshAfter: expiresAt.Add(-30 * time.Minute),
+		// NextRefreshAfter is aligned with RefreshLead (5min)
+		NextRefreshAfter: expiresAt.Add(-5 * time.Minute),
 	}
 
 	// Display the email if extracted
@@ -351,7 +356,8 @@ func (a *KiroAuthenticator) Refresh(ctx context.Context, cfg *config.Config, aut
 	updated.Metadata["refresh_token"] = tokenData.RefreshToken
 	updated.Metadata["expires_at"] = tokenData.ExpiresAt
 	updated.Metadata["last_refresh"] = now.Format(time.RFC3339) // For double-check optimization
-	updated.NextRefreshAfter = expiresAt.Add(-30 * time.Minute)
+	// NextRefreshAfter is aligned with RefreshLead (5min)
+	updated.NextRefreshAfter = expiresAt.Add(-5 * time.Minute)
 
 	return updated, nil
 }

sdk/cliproxy/auth/manager.go

@@ -40,7 +40,7 @@ type RefreshEvaluator interface {
 const (
 	refreshCheckInterval  = 5 * time.Second
 	refreshPendingBackoff = time.Minute
-	refreshFailureBackoff = 5 * time.Minute
+	refreshFailureBackoff = 1 * time.Minute
 	quotaBackoffBase      = time.Second
 	quotaBackoffMax       = 30 * time.Minute
 )
@@ -1471,7 +1471,9 @@ func (m *Manager) refreshAuth(ctx context.Context, id string) {
 		updated.Runtime = auth.Runtime
 	}
 	updated.LastRefreshedAt = now
-	updated.NextRefreshAfter = time.Time{}
+	// Preserve NextRefreshAfter set by the Authenticator
+	// If the Authenticator set a reasonable refresh time, it should not be overwritten
+	// If the Authenticator did not set it (zero value), shouldRefresh will use default logic
 	updated.LastError = nil
 	updated.UpdatedAt = now
 	_, _ = m.Update(ctx, updated)

test/thinking_conversion_test.go

@@ -0,0 +1,561 @@
+package test
+
+import (
+	"fmt"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+// statusErr mirrors executor.statusErr to keep validation behavior aligned.
+type statusErr struct {
+	code int
+	msg  string
+}
+
+func (e statusErr) Error() string { return e.msg }
+
+// registerCoreModels loads representative models across providers into the registry
+// so NormalizeThinkingBudget and level validation use real ranges.
+func registerCoreModels(t *testing.T) func() {
+	t.Helper()
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("thinking-core-%d", time.Now().UnixNano())
+	reg.RegisterClient(uid+"-gemini", "gemini", registry.GetGeminiModels())
+	reg.RegisterClient(uid+"-claude", "claude", registry.GetClaudeModels())
+	reg.RegisterClient(uid+"-openai", "codex", registry.GetOpenAIModels())
+	reg.RegisterClient(uid+"-qwen", "qwen", registry.GetQwenModels())
+	return func() {
+		reg.UnregisterClient(uid + "-gemini")
+		reg.UnregisterClient(uid + "-claude")
+		reg.UnregisterClient(uid + "-openai")
+		reg.UnregisterClient(uid + "-qwen")
+	}
+}
+
+func buildRawPayload(fromProtocol, modelWithSuffix string) []byte {
+	switch fromProtocol {
+	case "gemini":
+		return []byte(fmt.Sprintf(`{"model":"%s","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, modelWithSuffix))
+	case "openai-response":
+		return []byte(fmt.Sprintf(`{"model":"%s","input":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`, modelWithSuffix))
+	default: // openai / claude and other chat-style payloads
+		return []byte(fmt.Sprintf(`{"model":"%s","messages":[{"role":"user","content":"hi"}]}`, modelWithSuffix))
+	}
+}
+
+// applyThinkingMetadataLocal mirrors executor.applyThinkingMetadata.
+func applyThinkingMetadataLocal(payload []byte, metadata map[string]any, model string) []byte {
+	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
+	if !ok || (budgetOverride == nil && includeOverride == nil) {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if budgetOverride != nil {
+		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
+		budgetOverride = &norm
+	}
+	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
+}
+
+// applyReasoningEffortMetadataLocal mirrors executor.applyReasoningEffortMetadata.
+func applyReasoningEffortMetadataLocal(payload []byte, metadata map[string]any, model, field string) []byte {
+	if len(metadata) == 0 {
+		return payload
+	}
+	if !util.ModelSupportsThinking(model) {
+		return payload
+	}
+	if field == "" {
+		return payload
+	}
+	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
+		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+			return updated
+		}
+	}
+	if util.ModelUsesThinkingLevels(model) {
+		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+			if effort, ok := util.OpenAIThinkingBudgetToEffort(model, *budget); ok && effort != "" {
+				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
+					return updated
+				}
+			}
+		}
+	}
+	return payload
+}
+
+// normalizeThinkingConfigLocal mirrors executor.normalizeThinkingConfig.
+func normalizeThinkingConfigLocal(payload []byte, model string) []byte {
+	if len(payload) == 0 || model == "" {
+		return payload
+	}
+
+	if !util.ModelSupportsThinking(model) {
+		return stripThinkingFieldsLocal(payload, false)
+	}
+
+	if util.ModelUsesThinkingLevels(model) {
+		return normalizeReasoningEffortLevelLocal(payload, model)
+	}
+
+	// Model supports thinking but uses numeric budgets, not levels.
+	// Strip effort string fields since they are not applicable.
+	return stripThinkingFieldsLocal(payload, true)
+}
+
+// stripThinkingFieldsLocal mirrors executor.stripThinkingFields.
+func stripThinkingFieldsLocal(payload []byte, effortOnly bool) []byte {
+	fieldsToRemove := []string{
+		"reasoning_effort",
+		"reasoning.effort",
+	}
+	if !effortOnly {
+		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
+	}
+	out := payload
+	for _, field := range fieldsToRemove {
+		if gjson.GetBytes(out, field).Exists() {
+			out, _ = sjson.DeleteBytes(out, field)
+		}
+	}
+	return out
+}
+
+// normalizeReasoningEffortLevelLocal mirrors executor.normalizeReasoningEffortLevel.
+func normalizeReasoningEffortLevelLocal(payload []byte, model string) []byte {
+	out := payload
+
+	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
+		}
+	}
+
+	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
+		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
+			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
+		}
+	}
+
+	return out
+}
+
+// validateThinkingConfigLocal mirrors executor.validateThinkingConfig.
+func validateThinkingConfigLocal(payload []byte, model string) error {
+	if len(payload) == 0 || model == "" {
+		return nil
+	}
+	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
+		return nil
+	}
+
+	levels := util.GetModelThinkingLevels(model)
+	checkField := func(path string) error {
+		if effort := gjson.GetBytes(payload, path); effort.Exists() {
+			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
+				return statusErr{
+					code: http.StatusBadRequest,
+					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
+				}
+			}
+		}
+		return nil
+	}
+
+	if err := checkField("reasoning_effort"); err != nil {
+		return err
+	}
+	if err := checkField("reasoning.effort"); err != nil {
+		return err
+	}
+	return nil
+}
+
+// normalizeCodexPayload mirrors codex_executor's reasoning + streaming tweaks.
+func normalizeCodexPayload(body []byte, upstreamModel string) ([]byte, error) {
+	body = normalizeThinkingConfigLocal(body, upstreamModel)
+	if err := validateThinkingConfigLocal(body, upstreamModel); err != nil {
+		return body, err
+	}
+	body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	body, _ = sjson.SetBytes(body, "stream", true)
+	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	return body, nil
+}
+
+// buildBodyForProtocol runs a minimal request through the same translation and
+// thinking pipeline used in executors for the given target protocol.
+func buildBodyForProtocol(t *testing.T, fromProtocol, toProtocol, modelWithSuffix string) ([]byte, error) {
+	t.Helper()
+	normalizedModel, metadata := util.NormalizeThinkingModel(modelWithSuffix)
+	upstreamModel := util.ResolveOriginalModel(normalizedModel, metadata)
+	raw := buildRawPayload(fromProtocol, modelWithSuffix)
+	stream := fromProtocol != toProtocol
+
+	body := sdktranslator.TranslateRequest(
+		sdktranslator.FromString(fromProtocol),
+		sdktranslator.FromString(toProtocol),
+		normalizedModel,
+		raw,
+		stream,
+	)
+
+	var err error
+	switch toProtocol {
+	case "gemini":
+		body = applyThinkingMetadataLocal(body, metadata, normalizedModel)
+		body = util.ApplyDefaultThinkingIfNeeded(normalizedModel, body)
+		body = util.NormalizeGeminiThinkingBudget(normalizedModel, body)
+		body = util.StripThinkingConfigIfUnsupported(normalizedModel, body)
+	case "claude":
+		if budget, ok := util.ResolveClaudeThinkingConfig(normalizedModel, metadata); ok {
+			body = util.ApplyClaudeThinkingConfig(body, budget)
+		}
+	case "openai":
+		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning_effort")
+		body = normalizeThinkingConfigLocal(body, upstreamModel)
+		err = validateThinkingConfigLocal(body, upstreamModel)
+	case "codex": // OpenAI responses / codex
+		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning.effort")
+		// Mirror CodexExecutor final normalization and model override so tests log the final body.
+		body, err = normalizeCodexPayload(body, upstreamModel)
+	default:
+	}
+
+	// Mirror executor behavior: final payload uses the upstream (base) model name.
+	if upstreamModel != "" {
+		body, _ = sjson.SetBytes(body, "model", upstreamModel)
+	}
+
+	// For tests we only keep model + thinking-related fields to avoid noise.
+	body = filterThinkingBody(toProtocol, body, upstreamModel, normalizedModel)
+	return body, err
+}
+
+// filterThinkingBody projects the translated payload down to only model and
+// thinking-related fields for the given target protocol.
+func filterThinkingBody(toProtocol string, body []byte, upstreamModel, normalizedModel string) []byte {
+	if len(body) == 0 {
+		return body
+	}
+	out := []byte(`{}`)
+
+	// Preserve model if present, otherwise fall back to upstream/normalized model.
+	if m := gjson.GetBytes(body, "model"); m.Exists() {
+		out, _ = sjson.SetBytes(out, "model", m.Value())
+	} else if upstreamModel != "" {
+		out, _ = sjson.SetBytes(out, "model", upstreamModel)
+	} else if normalizedModel != "" {
+		out, _ = sjson.SetBytes(out, "model", normalizedModel)
+	}
+
+	switch toProtocol {
+	case "gemini":
+		if tc := gjson.GetBytes(body, "generationConfig.thinkingConfig"); tc.Exists() {
+			out, _ = sjson.SetRawBytes(out, "generationConfig.thinkingConfig", []byte(tc.Raw))
+		}
+	case "claude":
+		if tcfg := gjson.GetBytes(body, "thinking"); tcfg.Exists() {
+			out, _ = sjson.SetRawBytes(out, "thinking", []byte(tcfg.Raw))
+		}
+	case "openai":
+		if re := gjson.GetBytes(body, "reasoning_effort"); re.Exists() {
+			out, _ = sjson.SetBytes(out, "reasoning_effort", re.Value())
+		}
+	case "codex":
+		if re := gjson.GetBytes(body, "reasoning.effort"); re.Exists() {
+			out, _ = sjson.SetBytes(out, "reasoning.effort", re.Value())
+		}
+	}
+	return out
+}
+
+func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
+	cleanup := registerCoreModels(t)
+	defer cleanup()
+
+	models := []string{
+		"gpt-5",             // supports levels (low/medium/high)
+		"gemini-2.5-pro",    // supports numeric budget
+		"qwen3-coder-flash", // no thinking support
+	}
+	fromProtocols := []string{"openai", "claude", "gemini", "openai-response"}
+	toProtocols := []string{"gemini", "claude", "openai", "codex"}
+
+	type scenario struct {
+		name        string
+		modelSuffix string
+		expectFn    func(info *registry.ModelInfo) (present bool, budget int64)
+	}
+
+	buildBudgetFn := func(raw int) func(info *registry.ModelInfo) (bool, int64) {
+		return func(info *registry.ModelInfo) (bool, int64) {
+			if info == nil || info.Thinking == nil {
+				return false, 0
+			}
+			return true, int64(util.NormalizeThinkingBudget(info.ID, raw))
+		}
+	}
+
+	levelBudgetFn := func(level string) func(info *registry.ModelInfo) (bool, int64) {
+		return func(info *registry.ModelInfo) (bool, int64) {
+			if info == nil || info.Thinking == nil {
+				return false, 0
+			}
+			if b, ok := util.ThinkingEffortToBudget(info.ID, level); ok {
+				return true, int64(b)
+			}
+			return false, 0
+		}
+	}
+
+	for _, model := range models {
+		info := registry.GetGlobalRegistry().GetModelInfo(model)
+		min, max := 0, 0
+		if info != nil && info.Thinking != nil {
+			min = info.Thinking.Min
+			max = info.Thinking.Max
+		}
+
+		for _, from := range fromProtocols {
+			// Scenario selection follows protocol semantics:
+			// - OpenAI-style protocols (openai/openai-response) express thinking as levels.
+			// - Claude/Gemini-style protocols express thinking as numeric budgets.
+			cases := []scenario{
+				{name: "no-suffix", modelSuffix: model, expectFn: func(_ *registry.ModelInfo) (bool, int64) { return false, 0 }},
+			}
+			if from == "openai" || from == "openai-response" {
+				cases = append(cases,
+					scenario{name: "level-low", modelSuffix: fmt.Sprintf("%s(low)", model), expectFn: levelBudgetFn("low")},
+					scenario{name: "level-high", modelSuffix: fmt.Sprintf("%s(high)", model), expectFn: levelBudgetFn("high")},
+					scenario{name: "level-auto", modelSuffix: fmt.Sprintf("%s(auto)", model), expectFn: levelBudgetFn("auto")},
+				)
+			} else { // claude or gemini
+				if util.ModelUsesThinkingLevels(model) {
+					// Numeric budgets for level-based models are mapped into levels when needed.
+					cases = append(cases,
+						scenario{name: "numeric-0", modelSuffix: fmt.Sprintf("%s(0)", model), expectFn: buildBudgetFn(0)},
+						scenario{name: "numeric-1024", modelSuffix: fmt.Sprintf("%s(1024)", model), expectFn: buildBudgetFn(1024)},
+						scenario{name: "numeric-1025", modelSuffix: fmt.Sprintf("%s(1025)", model), expectFn: buildBudgetFn(1025)},
+						scenario{name: "numeric-8192", modelSuffix: fmt.Sprintf("%s(8192)", model), expectFn: buildBudgetFn(8192)},
+						scenario{name: "numeric-8193", modelSuffix: fmt.Sprintf("%s(8193)", model), expectFn: buildBudgetFn(8193)},
+						scenario{name: "numeric-24576", modelSuffix: fmt.Sprintf("%s(24576)", model), expectFn: buildBudgetFn(24576)},
+						scenario{name: "numeric-24577", modelSuffix: fmt.Sprintf("%s(24577)", model), expectFn: buildBudgetFn(24577)},
+					)
+				} else {
+					cases = append(cases,
+						scenario{name: "numeric-below-min", modelSuffix: fmt.Sprintf("%s(%d)", model, min-10), expectFn: buildBudgetFn(min - 10)},
+						scenario{name: "numeric-above-max", modelSuffix: fmt.Sprintf("%s(%d)", model, max+10), expectFn: buildBudgetFn(max + 10)},
+					)
+				}
+			}
+
+			for _, to := range toProtocols {
+				if from == to {
+					continue
+				}
+				t.Logf("─────────────────────────────────────────────────────────────────────────────────")
+				t.Logf("  %s -> %s | model: %s", from, to, model)
+				t.Logf("─────────────────────────────────────────────────────────────────────────────────")
+				for _, cs := range cases {
+					from := from
+					to := to
+					cs := cs
+					testName := fmt.Sprintf("%s->%s/%s/%s", from, to, model, cs.name)
+					t.Run(testName, func(t *testing.T) {
+						normalizedModel, metadata := util.NormalizeThinkingModel(cs.modelSuffix)
+						expectPresent, expectValue, expectErr := func() (bool, string, bool) {
+							switch to {
+							case "gemini":
+								budget, include, ok := util.ResolveThinkingConfigFromMetadata(normalizedModel, metadata)
+								if !ok || !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if include != nil && !*include {
+									return false, "", false
+								}
+								if budget == nil {
+									return false, "", false
+								}
+								norm := util.NormalizeThinkingBudget(normalizedModel, *budget)
+								return true, fmt.Sprintf("%d", norm), false
+							case "claude":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								budget, ok := util.ResolveClaudeThinkingConfig(normalizedModel, metadata)
+								if !ok || budget == nil {
+									return false, "", false
+								}
+								return true, fmt.Sprintf("%d", *budget), false
+							case "openai":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if !util.ModelUsesThinkingLevels(normalizedModel) {
+									// Non-levels models don't support effort strings in openai
+									return false, "", false
+								}
+								effort, ok := util.ReasoningEffortFromMetadata(metadata)
+								if !ok || strings.TrimSpace(effort) == "" {
+									if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+										if mapped, okMap := util.OpenAIThinkingBudgetToEffort(normalizedModel, *budget); okMap {
+											effort = mapped
+											ok = true
+										}
+									}
+								}
+								if !ok || strings.TrimSpace(effort) == "" {
+									return false, "", false
+								}
+								effort = strings.ToLower(strings.TrimSpace(effort))
+								if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, effort); okLevel {
+									return true, normalized, false
+								}
+								return false, "", true // validation would fail
+							case "codex":
+								if !util.ModelSupportsThinking(normalizedModel) {
+									return false, "", false
+								}
+								if !util.ModelUsesThinkingLevels(normalizedModel) {
+									// Non-levels models don't support effort strings in codex
+									if from != "openai-response" {
+										return false, "", false
+									}
+									return false, "", false
+								}
+								effort, ok := util.ReasoningEffortFromMetadata(metadata)
+								if ok && strings.TrimSpace(effort) != "" {
+									effort = strings.ToLower(strings.TrimSpace(effort))
+									if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, effort); okLevel {
+										return true, normalized, false
+									}
+									return false, "", true
+								}
+								if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+									if mapped, okMap := util.OpenAIThinkingBudgetToEffort(normalizedModel, *budget); okMap && mapped != "" {
+										mapped = strings.ToLower(strings.TrimSpace(mapped))
+										if normalized, okLevel := util.NormalizeReasoningEffortLevel(normalizedModel, mapped); okLevel {
+											return true, normalized, false
+										}
+										return false, "", true
+									}
+								}
+								if from != "openai-response" {
+									// Codex translators default reasoning.effort to "medium" when
+									// no explicit thinking suffix/metadata is provided.
+									return true, "medium", false
+								}
+								return false, "", false
+							default:
+								return false, "", false
+							}
+						}()
+
+						body, err := buildBodyForProtocol(t, from, to, cs.modelSuffix)
+						actualPresent, actualValue := func() (bool, string) {
+							path := ""
+							switch to {
+							case "gemini":
+								path = "generationConfig.thinkingConfig.thinkingBudget"
+							case "claude":
+								path = "thinking.budget_tokens"
+							case "openai":
+								path = "reasoning_effort"
+							case "codex":
+								path = "reasoning.effort"
+							}
+							if path == "" {
+								return false, ""
+							}
+							val := gjson.GetBytes(body, path)
+							if to == "codex" && !val.Exists() {
+								reasoning := gjson.GetBytes(body, "reasoning")
+								if reasoning.Exists() {
+									val = reasoning.Get("effort")
+								}
+							}
+							if !val.Exists() {
+								return false, ""
+							}
+							if val.Type == gjson.Number {
+								return true, fmt.Sprintf("%d", val.Int())
+							}
+							return true, val.String()
+						}()
+
+						t.Logf("from=%s to=%s model=%s suffix=%s present(expect=%v got=%v) value(expect=%s got=%s) err(expect=%v got=%v) body=%s",
+							from, to, model, cs.modelSuffix, expectPresent, actualPresent, expectValue, actualValue, expectErr, err != nil, string(body))
+
+						if expectErr {
+							if err == nil {
+								t.Fatalf("expected validation error but got none, body=%s", string(body))
+							}
+							return
+						}
+						if err != nil {
+							t.Fatalf("unexpected error: %v body=%s", err, string(body))
+						}
+
+						if expectPresent != actualPresent {
+							t.Fatalf("presence mismatch: expect %v got %v body=%s", expectPresent, actualPresent, string(body))
+						}
+						if expectPresent && expectValue != actualValue {
+							t.Fatalf("value mismatch: expect %s got %s body=%s", expectValue, actualValue, string(body))
+						}
+					})
+				}
+			}
+		}
+	}
+}
+
+func TestOpenAIThinkingBudgetToEffortRanges(t *testing.T) {
+	cleanup := registerCoreModels(t)
+	defer cleanup()
+
+	cases := []struct {
+		name   string
+		model  string
+		budget int
+		want   string
+		ok     bool
+	}{
+		{name: "zero-none", model: "gpt-5", budget: 0, want: "none", ok: true},
+		{name: "low-min", model: "gpt-5", budget: 1, want: "low", ok: true},
+		{name: "low-max", model: "gpt-5", budget: 1024, want: "low", ok: true},
+		{name: "medium-min", model: "gpt-5", budget: 1025, want: "medium", ok: true},
+		{name: "medium-max", model: "gpt-5", budget: 8192, want: "medium", ok: true},
+		{name: "high-min", model: "gpt-5", budget: 8193, want: "high", ok: true},
+		{name: "high-max", model: "gpt-5", budget: 24576, want: "high", ok: true},
+		{name: "over-max-clamps-to-highest", model: "gpt-5", budget: 24577, want: "high", ok: true},
+		{name: "over-max-xhigh-model", model: "gpt-5.2", budget: 50000, want: "xhigh", ok: true},
+		{name: "negative-unsupported", model: "gpt-5", budget: -5, want: "", ok: false},
+	}
+
+	for _, cs := range cases {
+		cs := cs
+		t.Run(cs.name, func(t *testing.T) {
+			got, ok := util.OpenAIThinkingBudgetToEffort(cs.model, cs.budget)
+			if ok != cs.ok {
+				t.Fatalf("ok mismatch for model=%s budget=%d: expect %v got %v", cs.model, cs.budget, cs.ok, ok)
+			}
+			if got != cs.want {
+				t.Fatalf("value mismatch for model=%s budget=%d: expect %q got %q", cs.model, cs.budget, cs.want, got)
+			}
+		})
+	}
+}