**fix(auth): properly handle callback forwarder instance in WebUI requests**

v6.8.40

.dockerignore

@@ -31,6 +31,7 @@ bin/*
 .agent/*
 .agents/*
 .opencode/*
+.idea/*
 .bmad/*
 _bmad/*
 _bmad-output/*

.gitignore

@@ -44,6 +44,7 @@ GEMINI.md
 .agents/*
 .agents/*
 .opencode/*
+.idea/*
 .bmad/*
 _bmad/*
 _bmad-output/*

README.md

@@ -10,20 +10,11 @@ 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), [Ravens2121](https://github.com/Ravens2121/CLIProxyAPIPlus/)
+[![z.ai](https://assets.router-for.me/english-5.png)](https://z.ai/subscribe?ic=8JVLJQFSKB)
 
 ## New Features (Plus Enhanced)
 
-- **OAuth Web Authentication**: Browser-based OAuth login for Kiro with beautiful web UI
-- **Rate Limiter**: Built-in request rate limiting to prevent API abuse
-- **Background Token Refresh**: Automatic token refresh 10 minutes before expiration
-- **Metrics & Monitoring**: Request metrics collection for monitoring and debugging
-- **Device Fingerprint**: Device fingerprint generation for enhanced security
-- **Cooldown Management**: Smart cooldown mechanism for API rate limits
-- **Usage Checker**: Real-time usage monitoring and quota management
-- **Model Converter**: Unified model name conversion across providers
-- **UTF-8 Stream Processing**: Improved streaming response handling
+GLM CODING PLAN is a subscription service designed for AI coding, starting at just $10/month. It provides access to their flagship GLM-4.7 & （GLM-5 Only Available  for Pro Users）model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
 
 ## Kiro Authentication
 

README_CN.md

@@ -10,22 +10,13 @@
 
 ## 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)、[Ravens2121](https://github.com/Ravens2121/CLIProxyAPIPlus/)提供
+[![bigmodel.cn](https://assets.router-for.me/chinese-5.png)](https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII)
 
 ## 新增功能 (Plus 增强版)
 
-- **OAuth Web 认证**: 基于浏览器的 Kiro OAuth 登录，提供美观的 Web UI
-- **请求限流器**: 内置请求限流，防止 API 滥用
-- **后台令牌刷新**: 过期前 10 分钟自动刷新令牌
-- **监控指标**: 请求指标收集，用于监控和调试
-- **设备指纹**: 设备指纹生成，增强安全性
-- **冷却管理**: 智能冷却机制，应对 API 速率限制
-- **用量检查器**: 实时用量监控和配额管理
-- **模型转换器**: 跨供应商的统一模型名称转换
-- **UTF-8 流处理**: 改进的流式响应处理
+GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如 Claude Code、Cline、Roo Code 中畅享智谱旗舰模型GLM-4.7（受限于算力，目前仅限Pro用户开放），为开发者提供顶尖的编码体验。
 
-## Kiro 认证
+智谱AI为本产品提供了特别优惠，使用以下链接购买可以享受九折优惠：https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII
 
 ### 命令行登录
 

internal/api/handlers/management/auth_files.go

@@ -48,14 +48,11 @@ import (
 var lastRefreshKeys = []string{"last_refresh", "lastRefresh", "last_refreshed_at", "lastRefreshedAt"}
 
 const (
-	anthropicCallbackPort   = 54545
-	geminiCallbackPort      = 8085
-	codexCallbackPort       = 1455
-	geminiCLIEndpoint       = "https://cloudcode-pa.googleapis.com"
-	geminiCLIVersion        = "v1internal"
-	geminiCLIUserAgent      = "google-api-nodejs-client/9.15.1"
-	geminiCLIApiClient      = "gl-node/22.17.0"
-	geminiCLIClientMetadata = "ideType=IDE_UNSPECIFIED,platform=PLATFORM_UNSPECIFIED,pluginType=GEMINI"
+	anthropicCallbackPort = 54545
+	geminiCallbackPort    = 8085
+	codexCallbackPort     = 1455
+	geminiCLIEndpoint     = "https://cloudcode-pa.googleapis.com"
+	geminiCLIVersion      = "v1internal"
 )
 
 type callbackForwarder struct {
@@ -195,17 +192,6 @@ func startCallbackForwarder(port int, provider, targetBase string) (*callbackFor
 	return forwarder, nil
 }
 
-func stopCallbackForwarder(port int) {
-	callbackForwardersMu.Lock()
-	forwarder := callbackForwarders[port]
-	if forwarder != nil {
-		delete(callbackForwarders, port)
-	}
-	callbackForwardersMu.Unlock()
-
-	stopForwarderInstance(port, forwarder)
-}
-
 func stopCallbackForwarderInstance(port int, forwarder *callbackForwarder) {
 	if forwarder == nil {
 		return
@@ -647,28 +633,66 @@ func (h *Handler) DeleteAuthFile(c *gin.Context) {
 		c.JSON(400, gin.H{"error": "invalid name"})
 		return
 	}
-	full := filepath.Join(h.cfg.AuthDir, filepath.Base(name))
-	if !filepath.IsAbs(full) {
-		if abs, errAbs := filepath.Abs(full); errAbs == nil {
-			full = abs
+
+	targetPath := filepath.Join(h.cfg.AuthDir, filepath.Base(name))
+	targetID := ""
+	if targetAuth := h.findAuthForDelete(name); targetAuth != nil {
+		targetID = strings.TrimSpace(targetAuth.ID)
+		if path := strings.TrimSpace(authAttribute(targetAuth, "path")); path != "" {
+			targetPath = path
 		}
 	}
-	if err := os.Remove(full); err != nil {
-		if os.IsNotExist(err) {
+	if !filepath.IsAbs(targetPath) {
+		if abs, errAbs := filepath.Abs(targetPath); errAbs == nil {
+			targetPath = abs
+		}
+	}
+	if errRemove := os.Remove(targetPath); errRemove != nil {
+		if os.IsNotExist(errRemove) {
 			c.JSON(404, gin.H{"error": "file not found"})
 		} else {
-			c.JSON(500, gin.H{"error": fmt.Sprintf("failed to remove file: %v", err)})
+			c.JSON(500, gin.H{"error": fmt.Sprintf("failed to remove file: %v", errRemove)})
 		}
 		return
 	}
-	if err := h.deleteTokenRecord(ctx, full); err != nil {
-		c.JSON(500, gin.H{"error": err.Error()})
+	if errDeleteRecord := h.deleteTokenRecord(ctx, targetPath); errDeleteRecord != nil {
+		c.JSON(500, gin.H{"error": errDeleteRecord.Error()})
 		return
 	}
-	h.disableAuth(ctx, full)
+	if targetID != "" {
+		h.disableAuth(ctx, targetID)
+	} else {
+		h.disableAuth(ctx, targetPath)
+	}
 	c.JSON(200, gin.H{"status": "ok"})
 }
 
+func (h *Handler) findAuthForDelete(name string) *coreauth.Auth {
+	if h == nil || h.authManager == nil {
+		return nil
+	}
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return nil
+	}
+	if auth, ok := h.authManager.GetByID(name); ok {
+		return auth
+	}
+	auths := h.authManager.List()
+	for _, auth := range auths {
+		if auth == nil {
+			continue
+		}
+		if strings.TrimSpace(auth.FileName) == name {
+			return auth
+		}
+		if filepath.Base(strings.TrimSpace(authAttribute(auth, "path"))) == name {
+			return auth
+		}
+	}
+	return nil
+}
+
 func (h *Handler) authIDForPath(path string) string {
 	path = strings.TrimSpace(path)
 	if path == "" {
@@ -902,10 +926,19 @@ func (h *Handler) disableAuth(ctx context.Context, id string) {
 	if h == nil || h.authManager == nil {
 		return
 	}
-	authID := h.authIDForPath(id)
-	if authID == "" {
-		authID = strings.TrimSpace(id)
+	id = strings.TrimSpace(id)
+	if id == "" {
+		return
 	}
+	if auth, ok := h.authManager.GetByID(id); ok {
+		auth.Disabled = true
+		auth.Status = coreauth.StatusDisabled
+		auth.StatusMessage = "removed via management API"
+		auth.UpdatedAt = time.Now()
+		_, _ = h.authManager.Update(ctx, auth)
+		return
+	}
+	authID := h.authIDForPath(id)
 	if authID == "" {
 		return
 	}
@@ -2384,9 +2417,7 @@ func callGeminiCLI(ctx context.Context, httpClient *http.Client, endpoint string
 		return fmt.Errorf("create request: %w", errRequest)
 	}
 	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("User-Agent", geminiCLIUserAgent)
-	req.Header.Set("X-Goog-Api-Client", geminiCLIApiClient)
-	req.Header.Set("Client-Metadata", geminiCLIClientMetadata)
+	req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 
 	resp, errDo := httpClient.Do(req)
 	if errDo != nil {
@@ -2456,7 +2487,7 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 			return false, fmt.Errorf("failed to create request: %w", errRequest)
 		}
 		req.Header.Set("Content-Type", "application/json")
-		req.Header.Set("User-Agent", geminiCLIUserAgent)
+		req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 		resp, errDo := httpClient.Do(req)
 		if errDo != nil {
 			return false, fmt.Errorf("failed to execute request: %w", errDo)
@@ -2477,7 +2508,7 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 			return false, fmt.Errorf("failed to create request: %w", errRequest)
 		}
 		req.Header.Set("Content-Type", "application/json")
-		req.Header.Set("User-Agent", geminiCLIUserAgent)
+		req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 		resp, errDo = httpClient.Do(req)
 		if errDo != nil {
 			return false, fmt.Errorf("failed to execute request: %w", errDo)
@@ -2554,6 +2585,7 @@ func PopulateAuthContext(ctx context.Context, c *gin.Context) context.Context {
 	}
 	return coreauth.WithRequestInfo(ctx, info)
 }
+
 const kiroCallbackPort = 9876
 
 func (h *Handler) RequestKiroToken(c *gin.Context) {
@@ -2690,6 +2722,7 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 		}
 
 		isWebUI := isWebUIRequest(c)
+		var forwarder *callbackForwarder
 		if isWebUI {
 			targetURL, errTarget := h.managementCallbackURL("/kiro/callback")
 			if errTarget != nil {
@@ -2697,7 +2730,8 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 				c.JSON(http.StatusInternalServerError, gin.H{"error": "callback server unavailable"})
 				return
 			}
-			if _, errStart := startCallbackForwarder(kiroCallbackPort, "kiro", targetURL); errStart != nil {
+			var errStart error
+			if forwarder, 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
@@ -2706,7 +2740,7 @@ func (h *Handler) RequestKiroToken(c *gin.Context) {
 
 		go func() {
 			if isWebUI {
-				defer stopCallbackForwarder(kiroCallbackPort)
+				defer stopCallbackForwarderInstance(kiroCallbackPort, forwarder)
 			}
 
 			socialClient := kiroauth.NewSocialAuthClient(h.cfg)
@@ -2909,7 +2943,7 @@ func (h *Handler) RequestKiloToken(c *gin.Context) {
 			Metadata: map[string]any{
 				"email":           status.UserEmail,
 				"organization_id": orgID,
-				"model":          defaults.Model,
+				"model":           defaults.Model,
 			},
 		}
 

internal/api/handlers/management/auth_files_delete_test.go

@@ -0,0 +1,129 @@
+package management
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+func TestDeleteAuthFile_UsesAuthPathFromManager(t *testing.T) {
+	t.Setenv("MANAGEMENT_PASSWORD", "")
+	gin.SetMode(gin.TestMode)
+
+	tempDir := t.TempDir()
+	authDir := filepath.Join(tempDir, "auth")
+	externalDir := filepath.Join(tempDir, "external")
+	if errMkdirAuth := os.MkdirAll(authDir, 0o700); errMkdirAuth != nil {
+		t.Fatalf("failed to create auth dir: %v", errMkdirAuth)
+	}
+	if errMkdirExternal := os.MkdirAll(externalDir, 0o700); errMkdirExternal != nil {
+		t.Fatalf("failed to create external dir: %v", errMkdirExternal)
+	}
+
+	fileName := "codex-user@example.com-plus.json"
+	shadowPath := filepath.Join(authDir, fileName)
+	realPath := filepath.Join(externalDir, fileName)
+	if errWriteShadow := os.WriteFile(shadowPath, []byte(`{"type":"codex","email":"shadow@example.com"}`), 0o600); errWriteShadow != nil {
+		t.Fatalf("failed to write shadow file: %v", errWriteShadow)
+	}
+	if errWriteReal := os.WriteFile(realPath, []byte(`{"type":"codex","email":"real@example.com"}`), 0o600); errWriteReal != nil {
+		t.Fatalf("failed to write real file: %v", errWriteReal)
+	}
+
+	manager := coreauth.NewManager(nil, nil, nil)
+	record := &coreauth.Auth{
+		ID:          "legacy/" + fileName,
+		FileName:    fileName,
+		Provider:    "codex",
+		Status:      coreauth.StatusError,
+		Unavailable: true,
+		Attributes: map[string]string{
+			"path": realPath,
+		},
+		Metadata: map[string]any{
+			"type":  "codex",
+			"email": "real@example.com",
+		},
+	}
+	if _, errRegister := manager.Register(context.Background(), record); errRegister != nil {
+		t.Fatalf("failed to register auth record: %v", errRegister)
+	}
+
+	h := NewHandlerWithoutConfigFilePath(&config.Config{AuthDir: authDir}, manager)
+	h.tokenStore = &memoryAuthStore{}
+
+	deleteRec := httptest.NewRecorder()
+	deleteCtx, _ := gin.CreateTestContext(deleteRec)
+	deleteReq := httptest.NewRequest(http.MethodDelete, "/v0/management/auth-files?name="+url.QueryEscape(fileName), nil)
+	deleteCtx.Request = deleteReq
+	h.DeleteAuthFile(deleteCtx)
+
+	if deleteRec.Code != http.StatusOK {
+		t.Fatalf("expected delete status %d, got %d with body %s", http.StatusOK, deleteRec.Code, deleteRec.Body.String())
+	}
+	if _, errStatReal := os.Stat(realPath); !os.IsNotExist(errStatReal) {
+		t.Fatalf("expected managed auth file to be removed, stat err: %v", errStatReal)
+	}
+	if _, errStatShadow := os.Stat(shadowPath); errStatShadow != nil {
+		t.Fatalf("expected shadow auth file to remain, stat err: %v", errStatShadow)
+	}
+
+	listRec := httptest.NewRecorder()
+	listCtx, _ := gin.CreateTestContext(listRec)
+	listReq := httptest.NewRequest(http.MethodGet, "/v0/management/auth-files", nil)
+	listCtx.Request = listReq
+	h.ListAuthFiles(listCtx)
+
+	if listRec.Code != http.StatusOK {
+		t.Fatalf("expected list status %d, got %d with body %s", http.StatusOK, listRec.Code, listRec.Body.String())
+	}
+	var listPayload map[string]any
+	if errUnmarshal := json.Unmarshal(listRec.Body.Bytes(), &listPayload); errUnmarshal != nil {
+		t.Fatalf("failed to decode list payload: %v", errUnmarshal)
+	}
+	filesRaw, ok := listPayload["files"].([]any)
+	if !ok {
+		t.Fatalf("expected files array, payload: %#v", listPayload)
+	}
+	if len(filesRaw) != 0 {
+		t.Fatalf("expected removed auth to be hidden from list, got %d entries", len(filesRaw))
+	}
+}
+
+func TestDeleteAuthFile_FallbackToAuthDirPath(t *testing.T) {
+	t.Setenv("MANAGEMENT_PASSWORD", "")
+	gin.SetMode(gin.TestMode)
+
+	authDir := t.TempDir()
+	fileName := "fallback-user.json"
+	filePath := filepath.Join(authDir, fileName)
+	if errWrite := os.WriteFile(filePath, []byte(`{"type":"codex"}`), 0o600); errWrite != nil {
+		t.Fatalf("failed to write auth file: %v", errWrite)
+	}
+
+	manager := coreauth.NewManager(nil, nil, nil)
+	h := NewHandlerWithoutConfigFilePath(&config.Config{AuthDir: authDir}, manager)
+	h.tokenStore = &memoryAuthStore{}
+
+	deleteRec := httptest.NewRecorder()
+	deleteCtx, _ := gin.CreateTestContext(deleteRec)
+	deleteReq := httptest.NewRequest(http.MethodDelete, "/v0/management/auth-files?name="+url.QueryEscape(fileName), nil)
+	deleteCtx.Request = deleteReq
+	h.DeleteAuthFile(deleteCtx)
+
+	if deleteRec.Code != http.StatusOK {
+		t.Fatalf("expected delete status %d, got %d with body %s", http.StatusOK, deleteRec.Code, deleteRec.Body.String())
+	}
+	if _, errStat := os.Stat(filePath); !os.IsNotExist(errStat) {
+		t.Fatalf("expected auth file to be removed from auth dir, stat err: %v", errStat)
+	}
+}

internal/api/modules/amp/proxy.go

@@ -15,6 +15,7 @@ import (
 	"strings"
 
 	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	log "github.com/sirupsen/logrus"
 )
 
@@ -77,6 +78,9 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 		req.Header.Del("X-Api-Key")
 		req.Header.Del("X-Goog-Api-Key")
 
+		// Remove proxy, client identity, and browser fingerprint headers
+		misc.ScrubProxyAndFingerprintHeaders(req)
+
 		// Remove query-based credentials if they match the authenticated client API key.
 		// This prevents leaking client auth material to the Amp upstream while avoiding
 		// breaking unrelated upstream query parameters.

internal/cmd/login.go

@@ -20,6 +20,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/gemini"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
 	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
@@ -27,11 +28,8 @@ import (
 )
 
 const (
-	geminiCLIEndpoint       = "https://cloudcode-pa.googleapis.com"
-	geminiCLIVersion        = "v1internal"
-	geminiCLIUserAgent      = "google-api-nodejs-client/9.15.1"
-	geminiCLIApiClient      = "gl-node/22.17.0"
-	geminiCLIClientMetadata = "ideType=IDE_UNSPECIFIED,platform=PLATFORM_UNSPECIFIED,pluginType=GEMINI"
+	geminiCLIEndpoint = "https://cloudcode-pa.googleapis.com"
+	geminiCLIVersion  = "v1internal"
 )
 
 type projectSelectionRequiredError struct{}
@@ -409,9 +407,7 @@ func callGeminiCLI(ctx context.Context, httpClient *http.Client, endpoint string
 		return fmt.Errorf("create request: %w", errRequest)
 	}
 	req.Header.Set("Content-Type", "application/json")
-	req.Header.Set("User-Agent", geminiCLIUserAgent)
-	req.Header.Set("X-Goog-Api-Client", geminiCLIApiClient)
-	req.Header.Set("Client-Metadata", geminiCLIClientMetadata)
+	req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 
 	resp, errDo := httpClient.Do(req)
 	if errDo != nil {
@@ -630,7 +626,7 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 			return false, fmt.Errorf("failed to create request: %w", errRequest)
 		}
 		req.Header.Set("Content-Type", "application/json")
-		req.Header.Set("User-Agent", geminiCLIUserAgent)
+		req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 		resp, errDo := httpClient.Do(req)
 		if errDo != nil {
 			return false, fmt.Errorf("failed to execute request: %w", errDo)
@@ -651,7 +647,7 @@ func checkCloudAPIIsEnabled(ctx context.Context, httpClient *http.Client, projec
 			return false, fmt.Errorf("failed to create request: %w", errRequest)
 		}
 		req.Header.Set("Content-Type", "application/json")
-		req.Header.Set("User-Agent", geminiCLIUserAgent)
+		req.Header.Set("User-Agent", misc.GeminiCLIUserAgent(""))
 		resp, errDo = httpClient.Do(req)
 		if errDo != nil {
 			return false, fmt.Errorf("failed to execute request: %w", errDo)

internal/config/config.go

@@ -579,16 +579,6 @@ func LoadConfig(configFile string) (*Config, error) {
 // If optional is true and the file is missing, it returns an empty Config.
 // If optional is true and the file is empty or invalid, it returns an empty Config.
 func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
-	// NOTE: Startup oauth-model-alias migration is intentionally disabled.
-	// Reason: avoid mutating config.yaml during server startup.
-	// Re-enable the block below if automatic startup migration is needed again.
-	// if migrated, err := MigrateOAuthModelAlias(configFile); err != nil {
-	// 	// Log warning but don't fail - config loading should still work
-	// 	fmt.Printf("Warning: oauth-model-alias migration failed: %v\n", err)
-	// } else if migrated {
-	// 	fmt.Println("Migrated oauth-model-mappings to oauth-model-alias")
-	// }
-
 	// Read the entire configuration file into memory.
 	data, err := os.ReadFile(configFile)
 	if err != nil {
@@ -1676,9 +1666,6 @@ func pruneMappingToGeneratedKeys(dstRoot, srcRoot *yaml.Node, key string) {
 	srcIdx := findMapKeyIndex(srcRoot, key)
 	if srcIdx < 0 {
 		// Keep an explicit empty mapping for oauth-model-alias when it was previously present.
-		//
-		// Rationale: LoadConfig runs MigrateOAuthModelAlias before unmarshalling. If the
-		// oauth-model-alias key is missing, migration will add the default antigravity aliases.
 		// When users delete the last channel from oauth-model-alias via the management API,
 		// we want that deletion to persist across hot reloads and restarts.
 		if key == "oauth-model-alias" {

internal/config/oauth_model_alias_defaults.go

@@ -0,0 +1,37 @@
+package config
+
+// defaultKiroAliases returns default oauth-model-alias entries for Kiro.
+// These aliases expose standard Claude IDs for Kiro-prefixed upstream models.
+func defaultKiroAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		// Sonnet 4.6
+		{Name: "kiro-claude-sonnet-4-6", Alias: "claude-sonnet-4-6", Fork: true},
+		// Sonnet 4.5
+		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5-20250929", Fork: true},
+		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5", Fork: true},
+		// Sonnet 4
+		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4-20250514", Fork: true},
+		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4", Fork: true},
+		// Opus 4.6
+		{Name: "kiro-claude-opus-4-6", Alias: "claude-opus-4-6", Fork: true},
+		// Opus 4.5
+		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5-20251101", Fork: true},
+		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5", Fork: true},
+		// Haiku 4.5
+		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5-20251001", Fork: true},
+		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5", Fork: true},
+	}
+}
+
+// defaultGitHubCopilotAliases returns default oauth-model-alias entries for
+// GitHub Copilot Claude models. It exposes hyphen-style IDs used by clients.
+func defaultGitHubCopilotAliases() []OAuthModelAlias {
+	return []OAuthModelAlias{
+		{Name: "claude-haiku-4.5", Alias: "claude-haiku-4-5", Fork: true},
+		{Name: "claude-opus-4.1", Alias: "claude-opus-4-1", Fork: true},
+		{Name: "claude-opus-4.5", Alias: "claude-opus-4-5", Fork: true},
+		{Name: "claude-opus-4.6", Alias: "claude-opus-4-6", Fork: true},
+		{Name: "claude-sonnet-4.5", Alias: "claude-sonnet-4-5", Fork: true},
+		{Name: "claude-sonnet-4.6", Alias: "claude-sonnet-4-6", Fork: true},
+	}
+}

internal/config/oauth_model_alias_migration.go

@@ -1,316 +0,0 @@
-package config
-
-import (
-	"os"
-	"strings"
-
-	"gopkg.in/yaml.v3"
-)
-
-// antigravityModelConversionTable maps old built-in aliases to actual model names
-// for the antigravity channel during migration.
-var antigravityModelConversionTable = map[string]string{
-	"gemini-2.5-computer-use-preview-10-2025": "rev19-uic3-1p",
-	"gemini-3-pro-image-preview":              "gemini-3-pro-image",
-	"gemini-3-pro-preview":                    "gemini-3-pro-high",
-	"gemini-3-flash-preview":                  "gemini-3-flash",
-	"gemini-claude-sonnet-4-5":                "claude-sonnet-4-5",
-	"gemini-claude-sonnet-4-5-thinking":       "claude-sonnet-4-5-thinking",
-	"gemini-claude-opus-4-5-thinking":         "claude-opus-4-5-thinking",
-	"gemini-claude-opus-4-6-thinking":         "claude-opus-4-6-thinking",
-}
-
-// defaultKiroAliases returns the default oauth-model-alias configuration
-// for the kiro channel. Maps kiro-prefixed model names to standard Claude model
-// names so that clients like Claude Code can use standard names directly.
-func defaultKiroAliases() []OAuthModelAlias {
-	return []OAuthModelAlias{
-		// Sonnet 4.6
-		{Name: "kiro-claude-sonnet-4-6", Alias: "claude-sonnet-4-6", Fork: true},
-		// Sonnet 4.5
-		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5-20250929", Fork: true},
-		{Name: "kiro-claude-sonnet-4-5", Alias: "claude-sonnet-4-5", Fork: true},
-		// Sonnet 4
-		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4-20250514", Fork: true},
-		{Name: "kiro-claude-sonnet-4", Alias: "claude-sonnet-4", Fork: true},
-		// Opus 4.6
-		{Name: "kiro-claude-opus-4-6", Alias: "claude-opus-4-6", Fork: true},
-		// Opus 4.5
-		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5-20251101", Fork: true},
-		{Name: "kiro-claude-opus-4-5", Alias: "claude-opus-4-5", Fork: true},
-		// Haiku 4.5
-		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5-20251001", Fork: true},
-		{Name: "kiro-claude-haiku-4-5", Alias: "claude-haiku-4-5", Fork: true},
-	}
-}
-
-// defaultGitHubCopilotAliases returns default oauth-model-alias entries that
-// expose Claude hyphen-style IDs for GitHub Copilot Claude models.
-// This keeps compatibility with clients (e.g. Claude Code) that use
-// Anthropic-style model IDs like "claude-opus-4-6".
-func defaultGitHubCopilotAliases() []OAuthModelAlias {
-	return []OAuthModelAlias{
-		{Name: "claude-haiku-4.5", Alias: "claude-haiku-4-5", Fork: true},
-		{Name: "claude-opus-4.1", Alias: "claude-opus-4-1", Fork: true},
-		{Name: "claude-opus-4.5", Alias: "claude-opus-4-5", Fork: true},
-		{Name: "claude-opus-4.6", Alias: "claude-opus-4-6", Fork: true},
-		{Name: "claude-sonnet-4.5", Alias: "claude-sonnet-4-5", Fork: true},
-		{Name: "claude-sonnet-4.6", Alias: "claude-sonnet-4-6", Fork: true},
-	}
-}
-
-// defaultAntigravityAliases returns the default oauth-model-alias configuration
-// for the antigravity channel when neither field exists.
-func defaultAntigravityAliases() []OAuthModelAlias {
-	return []OAuthModelAlias{
-		{Name: "rev19-uic3-1p", Alias: "gemini-2.5-computer-use-preview-10-2025"},
-		{Name: "gemini-3-pro-image", Alias: "gemini-3-pro-image-preview"},
-		{Name: "gemini-3-pro-high", Alias: "gemini-3-pro-preview"},
-		{Name: "gemini-3-flash", Alias: "gemini-3-flash-preview"},
-		{Name: "claude-sonnet-4-5", Alias: "gemini-claude-sonnet-4-5"},
-		{Name: "claude-sonnet-4-5-thinking", Alias: "gemini-claude-sonnet-4-5-thinking"},
-		{Name: "claude-opus-4-5-thinking", Alias: "gemini-claude-opus-4-5-thinking"},
-		{Name: "claude-opus-4-6-thinking", Alias: "gemini-claude-opus-4-6-thinking"},
-	}
-}
-
-// MigrateOAuthModelAlias checks for and performs migration from oauth-model-mappings
-// to oauth-model-alias at startup. Returns true if migration was performed.
-//
-// Migration flow:
-// 1. Check if oauth-model-alias exists -> skip migration
-// 2. Check if oauth-model-mappings exists -> convert and migrate
-//   - For antigravity channel, convert old built-in aliases to actual model names
-//
-// 3. Neither exists -> add default antigravity config
-func MigrateOAuthModelAlias(configFile string) (bool, error) {
-	data, err := os.ReadFile(configFile)
-	if err != nil {
-		if os.IsNotExist(err) {
-			return false, nil
-		}
-		return false, err
-	}
-	if len(data) == 0 {
-		return false, nil
-	}
-
-	// Parse YAML into node tree to preserve structure
-	var root yaml.Node
-	if err := yaml.Unmarshal(data, &root); err != nil {
-		return false, nil
-	}
-	if root.Kind != yaml.DocumentNode || len(root.Content) == 0 {
-		return false, nil
-	}
-	rootMap := root.Content[0]
-	if rootMap == nil || rootMap.Kind != yaml.MappingNode {
-		return false, nil
-	}
-
-	// Check if oauth-model-alias already exists
-	if findMapKeyIndex(rootMap, "oauth-model-alias") >= 0 {
-		return false, nil
-	}
-
-	// Check if oauth-model-mappings exists
-	oldIdx := findMapKeyIndex(rootMap, "oauth-model-mappings")
-	if oldIdx >= 0 {
-		// Migrate from old field
-		return migrateFromOldField(configFile, &root, rootMap, oldIdx)
-	}
-
-	// Neither field exists - add default antigravity config
-	return addDefaultAntigravityConfig(configFile, &root, rootMap)
-}
-
-// migrateFromOldField converts oauth-model-mappings to oauth-model-alias
-func migrateFromOldField(configFile string, root *yaml.Node, rootMap *yaml.Node, oldIdx int) (bool, error) {
-	if oldIdx+1 >= len(rootMap.Content) {
-		return false, nil
-	}
-	oldValue := rootMap.Content[oldIdx+1]
-	if oldValue == nil || oldValue.Kind != yaml.MappingNode {
-		return false, nil
-	}
-
-	// Parse the old aliases
-	oldAliases := parseOldAliasNode(oldValue)
-	if len(oldAliases) == 0 {
-		// Remove the old field and write
-		removeMapKeyByIndex(rootMap, oldIdx)
-		return writeYAMLNode(configFile, root)
-	}
-
-	// Convert model names for antigravity channel
-	newAliases := make(map[string][]OAuthModelAlias, len(oldAliases))
-	for channel, entries := range oldAliases {
-		converted := make([]OAuthModelAlias, 0, len(entries))
-		for _, entry := range entries {
-			newEntry := OAuthModelAlias{
-				Name:  entry.Name,
-				Alias: entry.Alias,
-				Fork:  entry.Fork,
-			}
-			// Convert model names for antigravity channel
-			if strings.EqualFold(channel, "antigravity") {
-				if actual, ok := antigravityModelConversionTable[entry.Name]; ok {
-					newEntry.Name = actual
-				}
-			}
-			converted = append(converted, newEntry)
-		}
-		newAliases[channel] = converted
-	}
-
-	// For antigravity channel, supplement missing default aliases
-	if antigravityEntries, exists := newAliases["antigravity"]; exists {
-		// Build a set of already configured model names (upstream names)
-		configuredModels := make(map[string]bool, len(antigravityEntries))
-		for _, entry := range antigravityEntries {
-			configuredModels[entry.Name] = true
-		}
-
-		// Add missing default aliases
-		for _, defaultAlias := range defaultAntigravityAliases() {
-			if !configuredModels[defaultAlias.Name] {
-				antigravityEntries = append(antigravityEntries, defaultAlias)
-			}
-		}
-		newAliases["antigravity"] = antigravityEntries
-	}
-
-	// Build new node
-	newNode := buildOAuthModelAliasNode(newAliases)
-
-	// Replace old key with new key and value
-	rootMap.Content[oldIdx].Value = "oauth-model-alias"
-	rootMap.Content[oldIdx+1] = newNode
-
-	return writeYAMLNode(configFile, root)
-}
-
-// addDefaultAntigravityConfig adds the default antigravity configuration
-func addDefaultAntigravityConfig(configFile string, root *yaml.Node, rootMap *yaml.Node) (bool, error) {
-	defaults := map[string][]OAuthModelAlias{
-		"antigravity": defaultAntigravityAliases(),
-	}
-	newNode := buildOAuthModelAliasNode(defaults)
-
-	// Add new key-value pair
-	keyNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "oauth-model-alias"}
-	rootMap.Content = append(rootMap.Content, keyNode, newNode)
-
-	return writeYAMLNode(configFile, root)
-}
-
-// parseOldAliasNode parses the old oauth-model-mappings node structure
-func parseOldAliasNode(node *yaml.Node) map[string][]OAuthModelAlias {
-	if node == nil || node.Kind != yaml.MappingNode {
-		return nil
-	}
-	result := make(map[string][]OAuthModelAlias)
-	for i := 0; i+1 < len(node.Content); i += 2 {
-		channelNode := node.Content[i]
-		entriesNode := node.Content[i+1]
-		if channelNode == nil || entriesNode == nil {
-			continue
-		}
-		channel := strings.ToLower(strings.TrimSpace(channelNode.Value))
-		if channel == "" || entriesNode.Kind != yaml.SequenceNode {
-			continue
-		}
-		entries := make([]OAuthModelAlias, 0, len(entriesNode.Content))
-		for _, entryNode := range entriesNode.Content {
-			if entryNode == nil || entryNode.Kind != yaml.MappingNode {
-				continue
-			}
-			entry := parseAliasEntry(entryNode)
-			if entry.Name != "" && entry.Alias != "" {
-				entries = append(entries, entry)
-			}
-		}
-		if len(entries) > 0 {
-			result[channel] = entries
-		}
-	}
-	return result
-}
-
-// parseAliasEntry parses a single alias entry node
-func parseAliasEntry(node *yaml.Node) OAuthModelAlias {
-	var entry OAuthModelAlias
-	for i := 0; i+1 < len(node.Content); i += 2 {
-		keyNode := node.Content[i]
-		valNode := node.Content[i+1]
-		if keyNode == nil || valNode == nil {
-			continue
-		}
-		switch strings.ToLower(strings.TrimSpace(keyNode.Value)) {
-		case "name":
-			entry.Name = strings.TrimSpace(valNode.Value)
-		case "alias":
-			entry.Alias = strings.TrimSpace(valNode.Value)
-		case "fork":
-			entry.Fork = strings.ToLower(strings.TrimSpace(valNode.Value)) == "true"
-		}
-	}
-	return entry
-}
-
-// buildOAuthModelAliasNode creates a YAML node for oauth-model-alias
-func buildOAuthModelAliasNode(aliases map[string][]OAuthModelAlias) *yaml.Node {
-	node := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
-	for channel, entries := range aliases {
-		channelNode := &yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: channel}
-		entriesNode := &yaml.Node{Kind: yaml.SequenceNode, Tag: "!!seq"}
-		for _, entry := range entries {
-			entryNode := &yaml.Node{Kind: yaml.MappingNode, Tag: "!!map"}
-			entryNode.Content = append(entryNode.Content,
-				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "name"},
-				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Name},
-				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "alias"},
-				&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: entry.Alias},
-			)
-			if entry.Fork {
-				entryNode.Content = append(entryNode.Content,
-					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!str", Value: "fork"},
-					&yaml.Node{Kind: yaml.ScalarNode, Tag: "!!bool", Value: "true"},
-				)
-			}
-			entriesNode.Content = append(entriesNode.Content, entryNode)
-		}
-		node.Content = append(node.Content, channelNode, entriesNode)
-	}
-	return node
-}
-
-// removeMapKeyByIndex removes a key-value pair from a mapping node by index
-func removeMapKeyByIndex(mapNode *yaml.Node, keyIdx int) {
-	if mapNode == nil || mapNode.Kind != yaml.MappingNode {
-		return
-	}
-	if keyIdx < 0 || keyIdx+1 >= len(mapNode.Content) {
-		return
-	}
-	mapNode.Content = append(mapNode.Content[:keyIdx], mapNode.Content[keyIdx+2:]...)
-}
-
-// writeYAMLNode writes the YAML node tree back to file
-func writeYAMLNode(configFile string, root *yaml.Node) (bool, error) {
-	f, err := os.Create(configFile)
-	if err != nil {
-		return false, err
-	}
-	defer f.Close()
-
-	enc := yaml.NewEncoder(f)
-	enc.SetIndent(2)
-	if err := enc.Encode(root); err != nil {
-		return false, err
-	}
-	if err := enc.Close(); err != nil {
-		return false, err
-	}
-	return true, nil
-}

internal/config/oauth_model_alias_migration_test.go

@@ -1,245 +0,0 @@
-package config
-
-import (
-	"os"
-	"path/filepath"
-	"strings"
-	"testing"
-
-	"gopkg.in/yaml.v3"
-)
-
-func TestMigrateOAuthModelAlias_SkipsIfNewFieldExists(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	content := `oauth-model-alias:
-  gemini-cli:
-    - name: "gemini-2.5-pro"
-      alias: "g2.5p"
-`
-	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if migrated {
-		t.Fatal("expected no migration when oauth-model-alias already exists")
-	}
-
-	// Verify file unchanged
-	data, _ := os.ReadFile(configFile)
-	if !strings.Contains(string(data), "oauth-model-alias:") {
-		t.Fatal("file should still contain oauth-model-alias")
-	}
-}
-
-func TestMigrateOAuthModelAlias_MigratesOldField(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	content := `oauth-model-mappings:
-  gemini-cli:
-    - name: "gemini-2.5-pro"
-      alias: "g2.5p"
-      fork: true
-`
-	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !migrated {
-		t.Fatal("expected migration to occur")
-	}
-
-	// Verify new field exists and old field removed
-	data, _ := os.ReadFile(configFile)
-	if strings.Contains(string(data), "oauth-model-mappings:") {
-		t.Fatal("old field should be removed")
-	}
-	if !strings.Contains(string(data), "oauth-model-alias:") {
-		t.Fatal("new field should exist")
-	}
-
-	// Parse and verify structure
-	var root yaml.Node
-	if err := yaml.Unmarshal(data, &root); err != nil {
-		t.Fatal(err)
-	}
-}
-
-func TestMigrateOAuthModelAlias_ConvertsAntigravityModels(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	// Use old model names that should be converted
-	content := `oauth-model-mappings:
-  antigravity:
-    - name: "gemini-2.5-computer-use-preview-10-2025"
-      alias: "computer-use"
-    - name: "gemini-3-pro-preview"
-      alias: "g3p"
-`
-	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !migrated {
-		t.Fatal("expected migration to occur")
-	}
-
-	// Verify model names were converted
-	data, _ := os.ReadFile(configFile)
-	content = string(data)
-	if !strings.Contains(content, "rev19-uic3-1p") {
-		t.Fatal("expected gemini-2.5-computer-use-preview-10-2025 to be converted to rev19-uic3-1p")
-	}
-	if !strings.Contains(content, "gemini-3-pro-high") {
-		t.Fatal("expected gemini-3-pro-preview to be converted to gemini-3-pro-high")
-	}
-
-	// Verify missing default aliases were supplemented
-	if !strings.Contains(content, "gemini-3-pro-image") {
-		t.Fatal("expected missing default alias gemini-3-pro-image to be added")
-	}
-	if !strings.Contains(content, "gemini-3-flash") {
-		t.Fatal("expected missing default alias gemini-3-flash to be added")
-	}
-	if !strings.Contains(content, "claude-sonnet-4-5") {
-		t.Fatal("expected missing default alias claude-sonnet-4-5 to be added")
-	}
-	if !strings.Contains(content, "claude-sonnet-4-5-thinking") {
-		t.Fatal("expected missing default alias claude-sonnet-4-5-thinking to be added")
-	}
-	if !strings.Contains(content, "claude-opus-4-5-thinking") {
-		t.Fatal("expected missing default alias claude-opus-4-5-thinking to be added")
-	}
-	if !strings.Contains(content, "claude-opus-4-6-thinking") {
-		t.Fatal("expected missing default alias claude-opus-4-6-thinking to be added")
-	}
-}
-
-func TestMigrateOAuthModelAlias_AddsDefaultIfNeitherExists(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	content := `debug: true
-port: 8080
-`
-	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !migrated {
-		t.Fatal("expected migration to add default config")
-	}
-
-	// Verify default antigravity config was added
-	data, _ := os.ReadFile(configFile)
-	content = string(data)
-	if !strings.Contains(content, "oauth-model-alias:") {
-		t.Fatal("expected oauth-model-alias to be added")
-	}
-	if !strings.Contains(content, "antigravity:") {
-		t.Fatal("expected antigravity channel to be added")
-	}
-	if !strings.Contains(content, "rev19-uic3-1p") {
-		t.Fatal("expected default antigravity aliases to include rev19-uic3-1p")
-	}
-}
-
-func TestMigrateOAuthModelAlias_PreservesOtherConfig(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	content := `debug: true
-port: 8080
-oauth-model-mappings:
-  gemini-cli:
-    - name: "test"
-      alias: "t"
-api-keys:
-  - "key1"
-  - "key2"
-`
-	if err := os.WriteFile(configFile, []byte(content), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if !migrated {
-		t.Fatal("expected migration to occur")
-	}
-
-	// Verify other config preserved
-	data, _ := os.ReadFile(configFile)
-	content = string(data)
-	if !strings.Contains(content, "debug: true") {
-		t.Fatal("expected debug field to be preserved")
-	}
-	if !strings.Contains(content, "port: 8080") {
-		t.Fatal("expected port field to be preserved")
-	}
-	if !strings.Contains(content, "api-keys:") {
-		t.Fatal("expected api-keys field to be preserved")
-	}
-}
-
-func TestMigrateOAuthModelAlias_NonexistentFile(t *testing.T) {
-	t.Parallel()
-
-	migrated, err := MigrateOAuthModelAlias("/nonexistent/path/config.yaml")
-	if err != nil {
-		t.Fatalf("unexpected error for nonexistent file: %v", err)
-	}
-	if migrated {
-		t.Fatal("expected no migration for nonexistent file")
-	}
-}
-
-func TestMigrateOAuthModelAlias_EmptyFile(t *testing.T) {
-	t.Parallel()
-
-	dir := t.TempDir()
-	configFile := filepath.Join(dir, "config.yaml")
-
-	if err := os.WriteFile(configFile, []byte(""), 0644); err != nil {
-		t.Fatal(err)
-	}
-
-	migrated, err := MigrateOAuthModelAlias(configFile)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
-	}
-	if migrated {
-		t.Fatal("expected no migration for empty file")
-	}
-}

internal/misc/header_utils.go

@@ -4,10 +4,98 @@
 package misc
 
 import (
+	"fmt"
 	"net/http"
+	"runtime"
 	"strings"
 )
 
+const (
+	// GeminiCLIVersion is the version string reported in the User-Agent for upstream requests.
+	GeminiCLIVersion = "0.31.0"
+
+	// GeminiCLIApiClientHeader is the value for the X-Goog-Api-Client header sent to the Gemini CLI upstream.
+	GeminiCLIApiClientHeader = "google-genai-sdk/1.41.0 gl-node/v22.19.0"
+)
+
+// geminiCLIOS maps Go runtime OS names to the Node.js-style platform strings used by Gemini CLI.
+func geminiCLIOS() string {
+	switch runtime.GOOS {
+	case "windows":
+		return "win32"
+	default:
+		return runtime.GOOS
+	}
+}
+
+// geminiCLIArch maps Go runtime architecture names to the Node.js-style arch strings used by Gemini CLI.
+func geminiCLIArch() string {
+	switch runtime.GOARCH {
+	case "amd64":
+		return "x64"
+	case "386":
+		return "x86"
+	default:
+		return runtime.GOARCH
+	}
+}
+
+// GeminiCLIUserAgent returns a User-Agent string that matches the Gemini CLI format.
+// The model parameter is included in the UA; pass "" or "unknown" when the model is not applicable.
+func GeminiCLIUserAgent(model string) string {
+	if model == "" {
+		model = "unknown"
+	}
+	return fmt.Sprintf("GeminiCLI/%s/%s (%s; %s)", GeminiCLIVersion, model, geminiCLIOS(), geminiCLIArch())
+}
+
+// ScrubProxyAndFingerprintHeaders removes all headers that could reveal
+// proxy infrastructure, client identity, or browser fingerprints from an
+// outgoing request. This ensures requests to upstream services look like they
+// originate directly from a native client rather than a third-party client
+// behind a reverse proxy.
+func ScrubProxyAndFingerprintHeaders(req *http.Request) {
+	if req == nil {
+		return
+	}
+
+	// --- Proxy tracing headers ---
+	req.Header.Del("X-Forwarded-For")
+	req.Header.Del("X-Forwarded-Host")
+	req.Header.Del("X-Forwarded-Proto")
+	req.Header.Del("X-Forwarded-Port")
+	req.Header.Del("X-Real-IP")
+	req.Header.Del("Forwarded")
+	req.Header.Del("Via")
+
+	// --- Client identity headers ---
+	req.Header.Del("X-Title")
+	req.Header.Del("X-Stainless-Lang")
+	req.Header.Del("X-Stainless-Package-Version")
+	req.Header.Del("X-Stainless-Os")
+	req.Header.Del("X-Stainless-Arch")
+	req.Header.Del("X-Stainless-Runtime")
+	req.Header.Del("X-Stainless-Runtime-Version")
+	req.Header.Del("Http-Referer")
+	req.Header.Del("Referer")
+
+	// --- Browser / Chromium fingerprint headers ---
+	// These are sent by Electron-based clients (e.g. CherryStudio) using the
+	// Fetch API, but NOT by Node.js https module (which Antigravity uses).
+	req.Header.Del("Sec-Ch-Ua")
+	req.Header.Del("Sec-Ch-Ua-Mobile")
+	req.Header.Del("Sec-Ch-Ua-Platform")
+	req.Header.Del("Sec-Fetch-Mode")
+	req.Header.Del("Sec-Fetch-Site")
+	req.Header.Del("Sec-Fetch-Dest")
+	req.Header.Del("Priority")
+
+	// --- Encoding negotiation ---
+	// Antigravity (Node.js) sends "gzip, deflate, br" by default;
+	// Electron-based clients may add "zstd" which is a fingerprint mismatch.
+	req.Header.Del("Accept-Encoding")
+}
+
 // EnsureHeader ensures that a header exists in the target header map by checking
 // multiple sources in order of priority: source headers, existing target headers,
 // and finally the default value. It only sets the header if it's not already present

internal/registry/model_definitions_static_data.go

@@ -37,7 +37,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.6 Sonnet",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high"}},
 		},
 		{
 			ID:                  "claude-opus-4-6",
@@ -49,7 +49,7 @@ func GetClaudeModels() []*ModelInfo {
 			Description:         "Premium model combining maximum intelligence with practical performance",
 			ContextLength:       1000000,
 			MaxCompletionTokens: 128000,
-			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high", "max"}},
 		},
 		{
 			ID:                  "claude-sonnet-4-6",
@@ -959,18 +959,17 @@ type AntigravityModelConfig struct {
 // Keys use upstream model names returned by the Antigravity models endpoint.
 func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 	return map[string]*AntigravityModelConfig{
-		// "rev19-uic3-1p":              {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true}},
 		"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}},
 		"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-pro-low":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-pro-high":      {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3.1-pro-low":       {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3.1-flash-image":   {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "high"}}},
 		"gemini-3-flash":           {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
-		"claude-opus-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}},
-		"claude-sonnet-4-6":        {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}},
+		"claude-opus-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-sonnet-4-6":        {Thinking: &ThinkingSupport{Min: 1024, Max: 64000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"gpt-oss-120b-medium":      {},
-		"tab_flash_lite_preview":   {},
 	}
 }
 

internal/registry/model_registry.go

@@ -49,6 +49,10 @@ type ModelInfo struct {
 	SupportedParameters []string `json:"supported_parameters,omitempty"`
 	// SupportedEndpoints lists supported API endpoints (e.g., "/chat/completions", "/responses").
 	SupportedEndpoints []string `json:"supported_endpoints,omitempty"`
+	// SupportedInputModalities lists supported input modalities (e.g., TEXT, IMAGE, VIDEO, AUDIO)
+	SupportedInputModalities []string `json:"supportedInputModalities,omitempty"`
+	// SupportedOutputModalities lists supported output modalities (e.g., TEXT, IMAGE)
+	SupportedOutputModalities []string `json:"supportedOutputModalities,omitempty"`
 
 	// Thinking holds provider-specific reasoning/thinking budget capabilities.
 	// This is optional and currently used for Gemini thinking budget normalization.
@@ -501,8 +505,11 @@ func cloneModelInfo(model *ModelInfo) *ModelInfo {
 	if len(model.SupportedParameters) > 0 {
 		copyModel.SupportedParameters = append([]string(nil), model.SupportedParameters...)
 	}
-	if len(model.SupportedEndpoints) > 0 {
-		copyModel.SupportedEndpoints = append([]string(nil), model.SupportedEndpoints...)
+	if len(model.SupportedInputModalities) > 0 {
+		copyModel.SupportedInputModalities = append([]string(nil), model.SupportedInputModalities...)
+	}
+	if len(model.SupportedOutputModalities) > 0 {
+		copyModel.SupportedOutputModalities = append([]string(nil), model.SupportedOutputModalities...)
 	}
 	return &copyModel
 }
@@ -1089,6 +1096,12 @@ func (r *ModelRegistry) convertModelToMap(model *ModelInfo, handlerType string)
 		if len(model.SupportedGenerationMethods) > 0 {
 			result["supportedGenerationMethods"] = model.SupportedGenerationMethods
 		}
+		if len(model.SupportedInputModalities) > 0 {
+			result["supportedInputModalities"] = model.SupportedInputModalities
+		}
+		if len(model.SupportedOutputModalities) > 0 {
+			result["supportedOutputModalities"] = model.SupportedOutputModalities
+		}
 		return result
 
 	default:

internal/runtime/executor/antigravity_executor.go

@@ -8,6 +8,7 @@ import (
 	"bytes"
 	"context"
 	"crypto/sha256"
+	"crypto/tls"
 	"encoding/binary"
 	"encoding/json"
 	"errors"
@@ -45,10 +46,10 @@ const (
 	antigravityModelsPath          = "/v1internal:fetchAvailableModels"
 	antigravityClientID            = "1071006060591-tmhssin2h21lcre235vtolojh4g403ep.apps.googleusercontent.com"
 	antigravityClientSecret        = "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"
-	defaultAntigravityAgent        = "antigravity/1.104.0 darwin/arm64"
+	defaultAntigravityAgent        = "antigravity/1.19.6 darwin/arm64"
 	antigravityAuthType            = "antigravity"
 	refreshSkew                    = 3000 * time.Second
-	systemInstruction              = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.**Absolute paths only****Proactiveness**"
+	// systemInstruction              = "You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team working on Advanced Agentic Coding.You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question.**Absolute paths only****Proactiveness**"
 )
 
 var (
@@ -142,6 +143,62 @@ func NewAntigravityExecutor(cfg *config.Config) *AntigravityExecutor {
 	return &AntigravityExecutor{cfg: cfg}
 }
 
+// antigravityTransport is a singleton HTTP/1.1 transport shared by all Antigravity requests.
+// It is initialized once via antigravityTransportOnce to avoid leaking a new connection pool
+// (and the goroutines managing it) on every request.
+var (
+	antigravityTransport     *http.Transport
+	antigravityTransportOnce sync.Once
+)
+
+func cloneTransportWithHTTP11(base *http.Transport) *http.Transport {
+	if base == nil {
+		return nil
+	}
+
+	clone := base.Clone()
+	clone.ForceAttemptHTTP2 = false
+	// Wipe TLSNextProto to prevent implicit HTTP/2 upgrade.
+	clone.TLSNextProto = make(map[string]func(authority string, c *tls.Conn) http.RoundTripper)
+	if clone.TLSClientConfig == nil {
+		clone.TLSClientConfig = &tls.Config{}
+	} else {
+		clone.TLSClientConfig = clone.TLSClientConfig.Clone()
+	}
+	// Actively advertise only HTTP/1.1 in the ALPN handshake.
+	clone.TLSClientConfig.NextProtos = []string{"http/1.1"}
+	return clone
+}
+
+// initAntigravityTransport creates the shared HTTP/1.1 transport exactly once.
+func initAntigravityTransport() {
+	base, ok := http.DefaultTransport.(*http.Transport)
+	if !ok {
+		base = &http.Transport{}
+	}
+	antigravityTransport = cloneTransportWithHTTP11(base)
+}
+
+// newAntigravityHTTPClient creates an HTTP client specifically for Antigravity,
+// enforcing HTTP/1.1 by disabling HTTP/2 to perfectly mimic Node.js https defaults.
+// The underlying Transport is a singleton to avoid leaking connection pools.
+func newAntigravityHTTPClient(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, timeout time.Duration) *http.Client {
+	antigravityTransportOnce.Do(initAntigravityTransport)
+
+	client := newProxyAwareHTTPClient(ctx, cfg, auth, timeout)
+	// If no transport is set, use the shared HTTP/1.1 transport.
+	if client.Transport == nil {
+		client.Transport = antigravityTransport
+		return client
+	}
+
+	// Preserve proxy settings from proxy-aware transports while forcing HTTP/1.1.
+	if transport, ok := client.Transport.(*http.Transport); ok {
+		client.Transport = cloneTransportWithHTTP11(transport)
+	}
+	return client
+}
+
 // Identifier returns the executor identifier.
 func (e *AntigravityExecutor) Identifier() string { return antigravityAuthType }
 
@@ -162,6 +219,8 @@ func (e *AntigravityExecutor) PrepareRequest(req *http.Request, auth *cliproxyau
 }
 
 // HttpRequest injects Antigravity credentials into the request and executes it.
+// It uses a whitelist approach: all incoming headers are stripped and only
+// the minimum set required by the Antigravity protocol is explicitly set.
 func (e *AntigravityExecutor) HttpRequest(ctx context.Context, auth *cliproxyauth.Auth, req *http.Request) (*http.Response, error) {
 	if req == nil {
 		return nil, fmt.Errorf("antigravity executor: request is nil")
@@ -170,10 +229,29 @@ func (e *AntigravityExecutor) HttpRequest(ctx context.Context, auth *cliproxyaut
 		ctx = req.Context()
 	}
 	httpReq := req.WithContext(ctx)
+
+	// --- Whitelist: save only the headers we need from the original request ---
+	contentType := httpReq.Header.Get("Content-Type")
+
+	// Wipe ALL incoming headers
+	for k := range httpReq.Header {
+		delete(httpReq.Header, k)
+	}
+
+	// --- Set only the headers Antigravity actually sends ---
+	if contentType != "" {
+		httpReq.Header.Set("Content-Type", contentType)
+	}
+	// Content-Length is managed automatically by Go's http.Client from the Body
+	httpReq.Header.Set("User-Agent", resolveUserAgent(auth))
+	httpReq.Close = true // sends Connection: close
+
+	// Inject Authorization: Bearer <token>
 	if err := e.PrepareRequest(httpReq, auth); err != nil {
 		return nil, err
 	}
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 	return httpClient.Do(httpReq)
 }
 
@@ -185,7 +263,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 	isClaude := strings.Contains(strings.ToLower(baseModel), "claude")
 
-	if isClaude || strings.Contains(baseModel, "gemini-3-pro") {
+	if isClaude || strings.Contains(baseModel, "gemini-3-pro") || strings.Contains(baseModel, "gemini-3.1-flash-image") {
 		return e.executeClaudeNonStream(ctx, auth, req, opts)
 	}
 
@@ -220,7 +298,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
 	attempts := antigravityRetryAttempts(auth, e.cfg)
 
@@ -362,7 +440,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
 	attempts := antigravityRetryAttempts(auth, e.cfg)
 
@@ -754,7 +832,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, "antigravity", "request", translated, originalTranslated, requestedModel)
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
 	attempts := antigravityRetryAttempts(auth, e.cfg)
 
@@ -956,7 +1034,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	payload = deleteJSONField(payload, "request.safetySettings")
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 
 	var authID, authLabel, authType, authValue string
 	if auth != nil {
@@ -987,10 +1065,10 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		if errReq != nil {
 			return cliproxyexecutor.Response{}, errReq
 		}
+		httpReq.Close = true
 		httpReq.Header.Set("Content-Type", "application/json")
 		httpReq.Header.Set("Authorization", "Bearer "+token)
 		httpReq.Header.Set("User-Agent", resolveUserAgent(auth))
-		httpReq.Header.Set("Accept", "application/json")
 		if host := resolveHost(base); host != "" {
 			httpReq.Host = host
 		}
@@ -1084,14 +1162,26 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 	}
 
 	baseURLs := antigravityBaseURLFallbackOrder(auth)
-	httpClient := newProxyAwareHTTPClient(ctx, cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, cfg, auth, 0)
 
 	for idx, baseURL := range baseURLs {
 		modelsURL := baseURL + antigravityModelsPath
-		httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, modelsURL, bytes.NewReader([]byte(`{}`)))
+
+		var payload []byte
+		if auth != nil && auth.Metadata != nil {
+			if pid, ok := auth.Metadata["project_id"].(string); ok && strings.TrimSpace(pid) != "" {
+				payload = []byte(fmt.Sprintf(`{"project": "%s"}`, strings.TrimSpace(pid)))
+			}
+		}
+		if len(payload) == 0 {
+			payload = []byte(`{}`)
+		}
+
+		httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, modelsURL, bytes.NewReader(payload))
 		if errReq != nil {
 			return fallbackAntigravityPrimaryModels()
 		}
+		httpReq.Close = true
 		httpReq.Header.Set("Content-Type", "application/json")
 		httpReq.Header.Set("Authorization", "Bearer "+token)
 		httpReq.Header.Set("User-Agent", resolveUserAgent(auth))
@@ -1152,7 +1242,7 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				continue
 			}
 			switch modelID {
-			case "chat_20706", "chat_23310", "gemini-2.5-flash-thinking", "gemini-3-pro-low", "gemini-2.5-pro":
+			case "chat_20706", "chat_23310", "tab_flash_lite_preview", "tab_jump_flash_lite_preview", "gemini-2.5-flash-thinking", "gemini-2.5-pro":
 				continue
 			}
 			modelCfg := modelConfig[modelID]
@@ -1174,6 +1264,29 @@ func FetchAntigravityModels(ctx context.Context, auth *cliproxyauth.Auth, cfg *c
 				OwnedBy:     antigravityAuthType,
 				Type:        antigravityAuthType,
 			}
+
+			// Build input modalities from upstream capability flags.
+			inputModalities := []string{"TEXT"}
+			if modelData.Get("supportsImages").Bool() {
+				inputModalities = append(inputModalities, "IMAGE")
+			}
+			if modelData.Get("supportsVideo").Bool() {
+				inputModalities = append(inputModalities, "VIDEO")
+			}
+			modelInfo.SupportedInputModalities = inputModalities
+			modelInfo.SupportedOutputModalities = []string{"TEXT"}
+
+			// Token limits from upstream.
+			if maxTok := modelData.Get("maxTokens").Int(); maxTok > 0 {
+				modelInfo.InputTokenLimit = int(maxTok)
+			}
+			if maxOut := modelData.Get("maxOutputTokens").Int(); maxOut > 0 {
+				modelInfo.OutputTokenLimit = int(maxOut)
+			}
+
+			// Supported generation methods (Gemini v1beta convention).
+			modelInfo.SupportedGenerationMethods = []string{"generateContent", "countTokens"}
+
 			// Look up Thinking support from static config using upstream model name.
 			if modelCfg != nil {
 				if modelCfg.Thinking != nil {
@@ -1241,10 +1354,11 @@ func (e *AntigravityExecutor) refreshToken(ctx context.Context, auth *cliproxyau
 		return auth, errReq
 	}
 	httpReq.Header.Set("Host", "oauth2.googleapis.com")
-	httpReq.Header.Set("User-Agent", defaultAntigravityAgent)
 	httpReq.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	// Real Antigravity uses Go's default User-Agent for OAuth token refresh
+	httpReq.Header.Set("User-Agent", "Go-http-client/2.0")
 
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 	httpResp, errDo := httpClient.Do(httpReq)
 	if errDo != nil {
 		return auth, errDo
@@ -1315,7 +1429,7 @@ func (e *AntigravityExecutor) ensureAntigravityProjectID(ctx context.Context, au
 		return nil
 	}
 
-	httpClient := newProxyAwareHTTPClient(ctx, e.cfg, auth, 0)
+	httpClient := newAntigravityHTTPClient(ctx, e.cfg, auth, 0)
 	projectID, errFetch := sdkAuth.FetchAntigravityProjectID(ctx, token, httpClient)
 	if errFetch != nil {
 		return errFetch
@@ -1369,7 +1483,7 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 	payload = geminiToAntigravity(modelName, payload, projectID)
 	payload, _ = sjson.SetBytes(payload, "model", modelName)
 
-	useAntigravitySchema := strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high")
+	useAntigravitySchema := strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro") || strings.Contains(modelName, "gemini-3.1-pro")
 	payloadStr := string(payload)
 	paths := make([]string, 0)
 	util.Walk(gjson.Parse(payloadStr), "", "parametersJsonSchema", &paths)
@@ -1383,18 +1497,18 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		payloadStr = util.CleanJSONSchemaForGemini(payloadStr)
 	}
 
-	if useAntigravitySchema {
-		systemInstructionPartsResult := gjson.Get(payloadStr, "request.systemInstruction.parts")
-		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.role", "user")
-		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.0.text", systemInstruction)
-		payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
+	// if useAntigravitySchema {
+	// 	systemInstructionPartsResult := gjson.Get(payloadStr, "request.systemInstruction.parts")
+	// 	payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.role", "user")
+	// 	payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.0.text", systemInstruction)
+	// 	payloadStr, _ = sjson.Set(payloadStr, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
 
-		if systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
-			for _, partResult := range systemInstructionPartsResult.Array() {
-				payloadStr, _ = sjson.SetRaw(payloadStr, "request.systemInstruction.parts.-1", partResult.Raw)
-			}
-		}
-	}
+	// 	if systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
+	// 		for _, partResult := range systemInstructionPartsResult.Array() {
+	// 			payloadStr, _ = sjson.SetRaw(payloadStr, "request.systemInstruction.parts.-1", partResult.Raw)
+	// 		}
+	// 	}
+	// }
 
 	if strings.Contains(modelName, "claude") {
 		payloadStr, _ = sjson.Set(payloadStr, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
@@ -1406,14 +1520,10 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 	if errReq != nil {
 		return nil, errReq
 	}
+	httpReq.Close = true
 	httpReq.Header.Set("Content-Type", "application/json")
 	httpReq.Header.Set("Authorization", "Bearer "+token)
 	httpReq.Header.Set("User-Agent", resolveUserAgent(auth))
-	if stream {
-		httpReq.Header.Set("Accept", "text/event-stream")
-	} else {
-		httpReq.Header.Set("Accept", "application/json")
-	}
 	if host := resolveHost(base); host != "" {
 		httpReq.Host = host
 	}
@@ -1625,7 +1735,16 @@ func resolveCustomAntigravityBaseURL(auth *cliproxyauth.Auth) string {
 func geminiToAntigravity(modelName string, payload []byte, projectID string) []byte {
 	template, _ := sjson.Set(string(payload), "model", modelName)
 	template, _ = sjson.Set(template, "userAgent", "antigravity")
-	template, _ = sjson.Set(template, "requestType", "agent")
+
+	isImageModel := strings.Contains(modelName, "image")
+
+	var reqType string
+	if isImageModel {
+		reqType = "image_gen"
+	} else {
+		reqType = "agent"
+	}
+	template, _ = sjson.Set(template, "requestType", reqType)
 
 	// Use real project ID from auth if available, otherwise generate random (legacy fallback)
 	if projectID != "" {
@@ -1633,8 +1752,13 @@ func geminiToAntigravity(modelName string, payload []byte, projectID string) []b
 	} else {
 		template, _ = sjson.Set(template, "project", generateProjectID())
 	}
-	template, _ = sjson.Set(template, "requestId", generateRequestID())
-	template, _ = sjson.Set(template, "request.sessionId", generateStableSessionID(payload))
+
+	if isImageModel {
+		template, _ = sjson.Set(template, "requestId", generateImageGenRequestID())
+	} else {
+		template, _ = sjson.Set(template, "requestId", generateRequestID())
+		template, _ = sjson.Set(template, "request.sessionId", generateStableSessionID(payload))
+	}
 
 	template, _ = sjson.Delete(template, "request.safetySettings")
 	if toolConfig := gjson.Get(template, "toolConfig"); toolConfig.Exists() && !gjson.Get(template, "request.toolConfig").Exists() {
@@ -1648,6 +1772,10 @@ func generateRequestID() string {
 	return "agent-" + uuid.NewString()
 }
 
+func generateImageGenRequestID() string {
+	return fmt.Sprintf("image_gen/%d/%s/12", time.Now().UnixMilli(), uuid.NewString())
+}
+
 func generateSessionID() string {
 	randSourceMutex.Lock()
 	n := randSource.Int63n(9_000_000_000_000_000_000)

internal/runtime/executor/claude_executor.go

@@ -9,9 +9,11 @@ import (
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"fmt"
 	"io"
 	"net/http"
+	"net/textproto"
 	"runtime"
 	"strings"
 	"time"
@@ -135,6 +137,15 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		body = ensureCacheControl(body)
 	}
 
+	// Enforce Anthropic's cache_control block limit (max 4 breakpoints per request).
+	// Cloaking and ensureCacheControl may push the total over 4 when the client
+	// (e.g. Amp CLI) already sends multiple cache_control blocks.
+	body = enforceCacheControlLimit(body, 4)
+
+	// Normalize TTL values to prevent ordering violations under prompt-caching-scope-2026-01-05.
+	// A 1h-TTL block must not appear after a 5m-TTL block in evaluation order (tools→system→messages).
+	body = normalizeCacheControlTTL(body)
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -176,11 +187,29 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		b, _ := io.ReadAll(httpResp.Body)
+		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
+		errBody := httpResp.Body
+		if ce := httpResp.Header.Get("Content-Encoding"); ce != "" {
+			var decErr error
+			errBody, decErr = decodeResponseBody(httpResp.Body, ce)
+			if decErr != nil {
+				recordAPIResponseError(ctx, e.cfg, decErr)
+				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
+				logWithRequestID(ctx).Warn(msg)
+				return resp, statusErr{code: httpResp.StatusCode, msg: msg}
+			}
+		}
+		b, readErr := io.ReadAll(errBody)
+		if readErr != nil {
+			recordAPIResponseError(ctx, e.cfg, readErr)
+			msg := fmt.Sprintf("failed to read error response body: %v", readErr)
+			logWithRequestID(ctx).Warn(msg)
+			b = []byte(msg)
+		}
 		appendAPIResponseChunk(ctx, e.cfg, b)
 		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
-		if errClose := httpResp.Body.Close(); errClose != nil {
+		if errClose := errBody.Close(); errClose != nil {
 			log.Errorf("response body close error: %v", errClose)
 		}
 		return resp, err
@@ -276,6 +305,12 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		body = ensureCacheControl(body)
 	}
 
+	// Enforce Anthropic's cache_control block limit (max 4 breakpoints per request).
+	body = enforceCacheControlLimit(body, 4)
+
+	// Normalize TTL values to prevent ordering violations under prompt-caching-scope-2026-01-05.
+	body = normalizeCacheControlTTL(body)
+
 	// Extract betas from body and convert to header
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -317,10 +352,28 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		b, _ := io.ReadAll(httpResp.Body)
+		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
+		errBody := httpResp.Body
+		if ce := httpResp.Header.Get("Content-Encoding"); ce != "" {
+			var decErr error
+			errBody, decErr = decodeResponseBody(httpResp.Body, ce)
+			if decErr != nil {
+				recordAPIResponseError(ctx, e.cfg, decErr)
+				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
+				logWithRequestID(ctx).Warn(msg)
+				return nil, statusErr{code: httpResp.StatusCode, msg: msg}
+			}
+		}
+		b, readErr := io.ReadAll(errBody)
+		if readErr != nil {
+			recordAPIResponseError(ctx, e.cfg, readErr)
+			msg := fmt.Sprintf("failed to read error response body: %v", readErr)
+			logWithRequestID(ctx).Warn(msg)
+			b = []byte(msg)
+		}
 		appendAPIResponseChunk(ctx, e.cfg, b)
 		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
-		if errClose := httpResp.Body.Close(); errClose != nil {
+		if errClose := errBody.Close(); errClose != nil {
 			log.Errorf("response body close error: %v", errClose)
 		}
 		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
@@ -425,6 +478,10 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 		body = checkSystemInstructions(body)
 	}
 
+	// Keep count_tokens requests compatible with Anthropic cache-control constraints too.
+	body = enforceCacheControlLimit(body, 4)
+	body = normalizeCacheControlTTL(body)
+
 	// Extract betas from body and convert to header (for count_tokens too)
 	var extraBetas []string
 	extraBetas, body = extractAndRemoveBetas(body)
@@ -464,9 +521,27 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, resp.StatusCode, resp.Header.Clone())
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		b, _ := io.ReadAll(resp.Body)
+		// Decompress error responses (e.g. gzip-compressed 400 errors from Anthropic API).
+		errBody := resp.Body
+		if ce := resp.Header.Get("Content-Encoding"); ce != "" {
+			var decErr error
+			errBody, decErr = decodeResponseBody(resp.Body, ce)
+			if decErr != nil {
+				recordAPIResponseError(ctx, e.cfg, decErr)
+				msg := fmt.Sprintf("failed to decode error response body (encoding=%s): %v", ce, decErr)
+				logWithRequestID(ctx).Warn(msg)
+				return cliproxyexecutor.Response{}, statusErr{code: resp.StatusCode, msg: msg}
+			}
+		}
+		b, readErr := io.ReadAll(errBody)
+		if readErr != nil {
+			recordAPIResponseError(ctx, e.cfg, readErr)
+			msg := fmt.Sprintf("failed to read error response body: %v", readErr)
+			logWithRequestID(ctx).Warn(msg)
+			b = []byte(msg)
+		}
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		if errClose := resp.Body.Close(); errClose != nil {
+		if errClose := errBody.Close(); errClose != nil {
 			log.Errorf("response body close error: %v", errClose)
 		}
 		return cliproxyexecutor.Response{}, statusErr{code: resp.StatusCode, msg: string(b)}
@@ -559,6 +634,12 @@ func disableThinkingIfToolChoiceForced(body []byte) []byte {
 	if toolChoiceType == "any" || toolChoiceType == "tool" {
 		// Remove thinking configuration entirely to avoid API error
 		body, _ = sjson.DeleteBytes(body, "thinking")
+		// Adaptive thinking may also set output_config.effort; remove it to avoid
+		// leaking thinking controls when tool_choice forces tool use.
+		body, _ = sjson.DeleteBytes(body, "output_config.effort")
+		if oc := gjson.GetBytes(body, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			body, _ = sjson.DeleteBytes(body, "output_config")
+		}
 	}
 	return body
 }
@@ -709,11 +790,21 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		}
 	}
 
-	// Merge extra betas from request body
-	if len(extraBetas) > 0 {
+	hasClaude1MHeader := false
+	if ginHeaders != nil {
+		if _, ok := ginHeaders[textproto.CanonicalMIMEHeaderKey("X-CPA-CLAUDE-1M")]; ok {
+			hasClaude1MHeader = true
+		}
+	}
+
+	// Merge extra betas from request body and request flags.
+	if len(extraBetas) > 0 || hasClaude1MHeader {
 		existingSet := make(map[string]bool)
 		for _, b := range strings.Split(baseBetas, ",") {
-			existingSet[strings.TrimSpace(b)] = true
+			betaName := strings.TrimSpace(b)
+			if betaName != "" {
+				existingSet[betaName] = true
+			}
 		}
 		for _, beta := range extraBetas {
 			beta = strings.TrimSpace(beta)
@@ -722,6 +813,9 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 				existingSet[beta] = true
 			}
 		}
+		if hasClaude1MHeader && !existingSet["context-1m-2025-08-07"] {
+			baseBetas += ",context-1m-2025-08-07"
+		}
 	}
 	r.Header.Set("Anthropic-Beta", baseBetas)
 
@@ -1073,17 +1167,22 @@ func generateBillingHeader(payload []byte) string {
 	return fmt.Sprintf("x-anthropic-billing-header: cc_version=2.1.63.%s; cc_entrypoint=cli; cch=%s;", buildHash, cch)
 }
 
-// checkSystemInstructionsWithMode injects Claude Code system prompt to match
-// the real Claude Code request format:
-//   system[0]: billing header (no cache_control)
-//   system[1]: "You are a Claude agent, built on Anthropic's Claude Agent SDK." (with cache_control)
-//   system[2..]: user's system messages (with cache_control on last)
+// checkSystemInstructionsWithMode injects Claude Code-style system blocks:
+//
+//	system[0]: billing header (no cache_control)
+//	system[1]: agent identifier (no cache_control)
+//	system[2..]: user system messages (cache_control added when missing)
 func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
 	system := gjson.GetBytes(payload, "system")
 
 	billingText := generateBillingHeader(payload)
 	billingBlock := fmt.Sprintf(`{"type":"text","text":"%s"}`, billingText)
-	agentBlock := `{"type":"text","text":"You are a Claude agent, built on Anthropic's Claude Agent SDK.","cache_control":{"type":"ephemeral","ttl":"1h"}}`
+	// No cache_control on the agent block. It is a cloaking artifact with zero cache
+	// value (the last system block is what actually triggers caching of all system content).
+	// Including any cache_control here creates an intra-system TTL ordering violation
+	// when the client's system blocks use ttl='1h' (prompt-caching-scope-2026-01-05 beta
+	// forbids 1h blocks after 5m blocks, and a no-TTL block defaults to 5m).
+	agentBlock := `{"type":"text","text":"You are a Claude agent, built on Anthropic's Claude Agent SDK."}`
 
 	if strictMode {
 		// Strict mode: billing header + agent identifier only
@@ -1103,11 +1202,12 @@ func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
 	if system.IsArray() {
 		system.ForEach(func(_, part gjson.Result) bool {
 			if part.Get("type").String() == "text" {
-				// Add cache_control with ttl to user system messages if not present
+				// Add cache_control to user system messages if not present.
+				// Do NOT add ttl — let it inherit the default (5m) to avoid
+				// TTL ordering violations with the prompt-caching-scope-2026-01-05 beta.
 				partJSON := part.Raw
 				if !part.Get("cache_control").Exists() {
 					partJSON, _ = sjson.Set(partJSON, "cache_control.type", "ephemeral")
-					partJSON, _ = sjson.Set(partJSON, "cache_control.ttl", "1h")
 				}
 				result += "," + partJSON
 			}
@@ -1254,6 +1354,313 @@ func countCacheControls(payload []byte) int {
 	return count
 }
 
+func parsePayloadObject(payload []byte) (map[string]any, bool) {
+	if len(payload) == 0 {
+		return nil, false
+	}
+	var root map[string]any
+	if err := json.Unmarshal(payload, &root); err != nil {
+		return nil, false
+	}
+	return root, true
+}
+
+func marshalPayloadObject(original []byte, root map[string]any) []byte {
+	if root == nil {
+		return original
+	}
+	out, err := json.Marshal(root)
+	if err != nil {
+		return original
+	}
+	return out
+}
+
+func asObject(v any) (map[string]any, bool) {
+	obj, ok := v.(map[string]any)
+	return obj, ok
+}
+
+func asArray(v any) ([]any, bool) {
+	arr, ok := v.([]any)
+	return arr, ok
+}
+
+func countCacheControlsMap(root map[string]any) int {
+	count := 0
+
+	if system, ok := asArray(root["system"]); ok {
+		for _, item := range system {
+			if obj, ok := asObject(item); ok {
+				if _, exists := obj["cache_control"]; exists {
+					count++
+				}
+			}
+		}
+	}
+
+	if tools, ok := asArray(root["tools"]); ok {
+		for _, item := range tools {
+			if obj, ok := asObject(item); ok {
+				if _, exists := obj["cache_control"]; exists {
+					count++
+				}
+			}
+		}
+	}
+
+	if messages, ok := asArray(root["messages"]); ok {
+		for _, msg := range messages {
+			msgObj, ok := asObject(msg)
+			if !ok {
+				continue
+			}
+			content, ok := asArray(msgObj["content"])
+			if !ok {
+				continue
+			}
+			for _, item := range content {
+				if obj, ok := asObject(item); ok {
+					if _, exists := obj["cache_control"]; exists {
+						count++
+					}
+				}
+			}
+		}
+	}
+
+	return count
+}
+
+func normalizeTTLForBlock(obj map[string]any, seen5m *bool) {
+	ccRaw, exists := obj["cache_control"]
+	if !exists {
+		return
+	}
+	cc, ok := asObject(ccRaw)
+	if !ok {
+		*seen5m = true
+		return
+	}
+	ttlRaw, ttlExists := cc["ttl"]
+	ttl, ttlIsString := ttlRaw.(string)
+	if !ttlExists || !ttlIsString || ttl != "1h" {
+		*seen5m = true
+		return
+	}
+	if *seen5m {
+		delete(cc, "ttl")
+	}
+}
+
+func findLastCacheControlIndex(arr []any) int {
+	last := -1
+	for idx, item := range arr {
+		obj, ok := asObject(item)
+		if !ok {
+			continue
+		}
+		if _, exists := obj["cache_control"]; exists {
+			last = idx
+		}
+	}
+	return last
+}
+
+func stripCacheControlExceptIndex(arr []any, preserveIdx int, excess *int) {
+	for idx, item := range arr {
+		if *excess <= 0 {
+			return
+		}
+		obj, ok := asObject(item)
+		if !ok {
+			continue
+		}
+		if _, exists := obj["cache_control"]; exists && idx != preserveIdx {
+			delete(obj, "cache_control")
+			*excess--
+		}
+	}
+}
+
+func stripAllCacheControl(arr []any, excess *int) {
+	for _, item := range arr {
+		if *excess <= 0 {
+			return
+		}
+		obj, ok := asObject(item)
+		if !ok {
+			continue
+		}
+		if _, exists := obj["cache_control"]; exists {
+			delete(obj, "cache_control")
+			*excess--
+		}
+	}
+}
+
+func stripMessageCacheControl(messages []any, excess *int) {
+	for _, msg := range messages {
+		if *excess <= 0 {
+			return
+		}
+		msgObj, ok := asObject(msg)
+		if !ok {
+			continue
+		}
+		content, ok := asArray(msgObj["content"])
+		if !ok {
+			continue
+		}
+		for _, item := range content {
+			if *excess <= 0 {
+				return
+			}
+			obj, ok := asObject(item)
+			if !ok {
+				continue
+			}
+			if _, exists := obj["cache_control"]; exists {
+				delete(obj, "cache_control")
+				*excess--
+			}
+		}
+	}
+}
+
+// normalizeCacheControlTTL ensures cache_control TTL values don't violate the
+// prompt-caching-scope-2026-01-05 ordering constraint: a 1h-TTL block must not
+// appear after a 5m-TTL block anywhere in the evaluation order.
+//
+// Anthropic evaluates blocks in order: tools → system (index 0..N) → messages.
+// Within each section, blocks are evaluated in array order. A 5m (default) block
+// followed by a 1h block at ANY later position is an error — including within
+// the same section (e.g. system[1]=5m then system[3]=1h).
+//
+// Strategy: walk all cache_control blocks in evaluation order. Once a 5m block
+// is seen, strip ttl from ALL subsequent 1h blocks (downgrading them to 5m).
+func normalizeCacheControlTTL(payload []byte) []byte {
+	root, ok := parsePayloadObject(payload)
+	if !ok {
+		return payload
+	}
+
+	seen5m := false
+
+	if tools, ok := asArray(root["tools"]); ok {
+		for _, tool := range tools {
+			if obj, ok := asObject(tool); ok {
+				normalizeTTLForBlock(obj, &seen5m)
+			}
+		}
+	}
+
+	if system, ok := asArray(root["system"]); ok {
+		for _, item := range system {
+			if obj, ok := asObject(item); ok {
+				normalizeTTLForBlock(obj, &seen5m)
+			}
+		}
+	}
+
+	if messages, ok := asArray(root["messages"]); ok {
+		for _, msg := range messages {
+			msgObj, ok := asObject(msg)
+			if !ok {
+				continue
+			}
+			content, ok := asArray(msgObj["content"])
+			if !ok {
+				continue
+			}
+			for _, item := range content {
+				if obj, ok := asObject(item); ok {
+					normalizeTTLForBlock(obj, &seen5m)
+				}
+			}
+		}
+	}
+
+	return marshalPayloadObject(payload, root)
+}
+
+// enforceCacheControlLimit removes excess cache_control blocks from a payload
+// so the total does not exceed the Anthropic API limit (currently 4).
+//
+// Anthropic evaluates cache breakpoints in order: tools → system → messages.
+// The most valuable breakpoints are:
+//  1. Last tool         — caches ALL tool definitions
+//  2. Last system block — caches ALL system content
+//  3. Recent messages   — cache conversation context
+//
+// Removal priority (strip lowest-value first):
+//
+//	Phase 1: system blocks earliest-first, preserving the last one.
+//	Phase 2: tool blocks earliest-first, preserving the last one.
+//	Phase 3: message content blocks earliest-first.
+//	Phase 4: remaining system blocks (last system).
+//	Phase 5: remaining tool blocks (last tool).
+func enforceCacheControlLimit(payload []byte, maxBlocks int) []byte {
+	root, ok := parsePayloadObject(payload)
+	if !ok {
+		return payload
+	}
+
+	total := countCacheControlsMap(root)
+	if total <= maxBlocks {
+		return payload
+	}
+
+	excess := total - maxBlocks
+
+	var system []any
+	if arr, ok := asArray(root["system"]); ok {
+		system = arr
+	}
+	var tools []any
+	if arr, ok := asArray(root["tools"]); ok {
+		tools = arr
+	}
+	var messages []any
+	if arr, ok := asArray(root["messages"]); ok {
+		messages = arr
+	}
+
+	if len(system) > 0 {
+		stripCacheControlExceptIndex(system, findLastCacheControlIndex(system), &excess)
+	}
+	if excess <= 0 {
+		return marshalPayloadObject(payload, root)
+	}
+
+	if len(tools) > 0 {
+		stripCacheControlExceptIndex(tools, findLastCacheControlIndex(tools), &excess)
+	}
+	if excess <= 0 {
+		return marshalPayloadObject(payload, root)
+	}
+
+	if len(messages) > 0 {
+		stripMessageCacheControl(messages, &excess)
+	}
+	if excess <= 0 {
+		return marshalPayloadObject(payload, root)
+	}
+
+	if len(system) > 0 {
+		stripAllCacheControl(system, &excess)
+	}
+	if excess <= 0 {
+		return marshalPayloadObject(payload, root)
+	}
+
+	if len(tools) > 0 {
+		stripAllCacheControl(tools, &excess)
+	}
+
+	return marshalPayloadObject(payload, root)
+}
+
 // injectMessagesCacheControl adds cache_control to the second-to-last user turn for multi-turn caching.
 // Per Anthropic docs: "Place cache_control on the second-to-last User message to let the model reuse the earlier cache."
 // This enables caching of conversation history, which is especially beneficial for long multi-turn conversations.

internal/runtime/executor/claude_executor_test.go

@@ -6,6 +6,7 @@ import (
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"testing"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
@@ -348,3 +349,237 @@ func TestApplyClaudeToolPrefix_SkipsBuiltinToolReference(t *testing.T) {
 		t.Fatalf("built-in tool_reference should not be prefixed, got %q", got)
 	}
 }
+
+func TestNormalizeCacheControlTTL_DowngradesLaterOneHourBlocks(t *testing.T) {
+	payload := []byte(`{
+		"tools": [{"name":"t1","cache_control":{"type":"ephemeral","ttl":"1h"}}],
+		"system": [{"type":"text","text":"s1","cache_control":{"type":"ephemeral"}}],
+		"messages": [{"role":"user","content":[{"type":"text","text":"u1","cache_control":{"type":"ephemeral","ttl":"1h"}}]}]
+	}`)
+
+	out := normalizeCacheControlTTL(payload)
+
+	if got := gjson.GetBytes(out, "tools.0.cache_control.ttl").String(); got != "1h" {
+		t.Fatalf("tools.0.cache_control.ttl = %q, want %q", got, "1h")
+	}
+	if gjson.GetBytes(out, "messages.0.content.0.cache_control.ttl").Exists() {
+		t.Fatalf("messages.0.content.0.cache_control.ttl should be removed after a default-5m block")
+	}
+}
+
+func TestEnforceCacheControlLimit_StripsNonLastToolBeforeMessages(t *testing.T) {
+	payload := []byte(`{
+		"tools": [
+			{"name":"t1","cache_control":{"type":"ephemeral"}},
+			{"name":"t2","cache_control":{"type":"ephemeral"}}
+		],
+		"system": [{"type":"text","text":"s1","cache_control":{"type":"ephemeral"}}],
+		"messages": [
+			{"role":"user","content":[{"type":"text","text":"u1","cache_control":{"type":"ephemeral"}}]},
+			{"role":"user","content":[{"type":"text","text":"u2","cache_control":{"type":"ephemeral"}}]}
+		]
+	}`)
+
+	out := enforceCacheControlLimit(payload, 4)
+
+	if got := countCacheControls(out); got != 4 {
+		t.Fatalf("cache_control count = %d, want 4", got)
+	}
+	if gjson.GetBytes(out, "tools.0.cache_control").Exists() {
+		t.Fatalf("tools.0.cache_control should be removed first (non-last tool)")
+	}
+	if !gjson.GetBytes(out, "tools.1.cache_control").Exists() {
+		t.Fatalf("tools.1.cache_control (last tool) should be preserved")
+	}
+	if !gjson.GetBytes(out, "messages.0.content.0.cache_control").Exists() || !gjson.GetBytes(out, "messages.1.content.0.cache_control").Exists() {
+		t.Fatalf("message cache_control blocks should be preserved when non-last tool removal is enough")
+	}
+}
+
+func TestEnforceCacheControlLimit_ToolOnlyPayloadStillRespectsLimit(t *testing.T) {
+	payload := []byte(`{
+		"tools": [
+			{"name":"t1","cache_control":{"type":"ephemeral"}},
+			{"name":"t2","cache_control":{"type":"ephemeral"}},
+			{"name":"t3","cache_control":{"type":"ephemeral"}},
+			{"name":"t4","cache_control":{"type":"ephemeral"}},
+			{"name":"t5","cache_control":{"type":"ephemeral"}}
+		]
+	}`)
+
+	out := enforceCacheControlLimit(payload, 4)
+
+	if got := countCacheControls(out); got != 4 {
+		t.Fatalf("cache_control count = %d, want 4", got)
+	}
+	if gjson.GetBytes(out, "tools.0.cache_control").Exists() {
+		t.Fatalf("tools.0.cache_control should be removed to satisfy max=4")
+	}
+	if !gjson.GetBytes(out, "tools.4.cache_control").Exists() {
+		t.Fatalf("last tool cache_control should be preserved when possible")
+	}
+}
+
+func TestClaudeExecutor_CountTokens_AppliesCacheControlGuards(t *testing.T) {
+	var seenBody []byte
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		seenBody = bytes.Clone(body)
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"input_tokens":42}`))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+
+	payload := []byte(`{
+		"tools": [
+			{"name":"t1","cache_control":{"type":"ephemeral","ttl":"1h"}},
+			{"name":"t2","cache_control":{"type":"ephemeral"}}
+		],
+		"system": [
+			{"type":"text","text":"s1","cache_control":{"type":"ephemeral","ttl":"1h"}},
+			{"type":"text","text":"s2","cache_control":{"type":"ephemeral","ttl":"1h"}}
+		],
+		"messages": [
+			{"role":"user","content":[{"type":"text","text":"u1","cache_control":{"type":"ephemeral","ttl":"1h"}}]},
+			{"role":"user","content":[{"type":"text","text":"u2","cache_control":{"type":"ephemeral","ttl":"1h"}}]}
+		]
+	}`)
+
+	_, err := executor.CountTokens(context.Background(), auth, cliproxyexecutor.Request{
+		Model:   "claude-3-5-haiku-20241022",
+		Payload: payload,
+	}, cliproxyexecutor.Options{SourceFormat: sdktranslator.FromString("claude")})
+	if err != nil {
+		t.Fatalf("CountTokens error: %v", err)
+	}
+
+	if len(seenBody) == 0 {
+		t.Fatal("expected count_tokens request body to be captured")
+	}
+	if got := countCacheControls(seenBody); got > 4 {
+		t.Fatalf("count_tokens body has %d cache_control blocks, want <= 4", got)
+	}
+	if hasTTLOrderingViolation(seenBody) {
+		t.Fatalf("count_tokens body still has ttl ordering violations: %s", string(seenBody))
+	}
+}
+
+func hasTTLOrderingViolation(payload []byte) bool {
+	seen5m := false
+	violates := false
+
+	checkCC := func(cc gjson.Result) {
+		if !cc.Exists() || violates {
+			return
+		}
+		ttl := cc.Get("ttl").String()
+		if ttl != "1h" {
+			seen5m = true
+			return
+		}
+		if seen5m {
+			violates = true
+		}
+	}
+
+	tools := gjson.GetBytes(payload, "tools")
+	if tools.IsArray() {
+		tools.ForEach(func(_, tool gjson.Result) bool {
+			checkCC(tool.Get("cache_control"))
+			return !violates
+		})
+	}
+
+	system := gjson.GetBytes(payload, "system")
+	if system.IsArray() {
+		system.ForEach(func(_, item gjson.Result) bool {
+			checkCC(item.Get("cache_control"))
+			return !violates
+		})
+	}
+
+	messages := gjson.GetBytes(payload, "messages")
+	if messages.IsArray() {
+		messages.ForEach(func(_, msg gjson.Result) bool {
+			content := msg.Get("content")
+			if content.IsArray() {
+				content.ForEach(func(_, item gjson.Result) bool {
+					checkCC(item.Get("cache_control"))
+					return !violates
+				})
+			}
+			return !violates
+		})
+	}
+
+	return violates
+}
+
+func TestClaudeExecutor_Execute_InvalidGzipErrorBodyReturnsDecodeMessage(t *testing.T) {
+	testClaudeExecutorInvalidCompressedErrorBody(t, func(executor *ClaudeExecutor, auth *cliproxyauth.Auth, payload []byte) error {
+		_, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   "claude-3-5-sonnet-20241022",
+			Payload: payload,
+		}, cliproxyexecutor.Options{SourceFormat: sdktranslator.FromString("claude")})
+		return err
+	})
+}
+
+func TestClaudeExecutor_ExecuteStream_InvalidGzipErrorBodyReturnsDecodeMessage(t *testing.T) {
+	testClaudeExecutorInvalidCompressedErrorBody(t, func(executor *ClaudeExecutor, auth *cliproxyauth.Auth, payload []byte) error {
+		_, err := executor.ExecuteStream(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   "claude-3-5-sonnet-20241022",
+			Payload: payload,
+		}, cliproxyexecutor.Options{SourceFormat: sdktranslator.FromString("claude")})
+		return err
+	})
+}
+
+func TestClaudeExecutor_CountTokens_InvalidGzipErrorBodyReturnsDecodeMessage(t *testing.T) {
+	testClaudeExecutorInvalidCompressedErrorBody(t, func(executor *ClaudeExecutor, auth *cliproxyauth.Auth, payload []byte) error {
+		_, err := executor.CountTokens(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   "claude-3-5-sonnet-20241022",
+			Payload: payload,
+		}, cliproxyexecutor.Options{SourceFormat: sdktranslator.FromString("claude")})
+		return err
+	})
+}
+
+func testClaudeExecutorInvalidCompressedErrorBody(
+	t *testing.T,
+	invoke func(executor *ClaudeExecutor, auth *cliproxyauth.Auth, payload []byte) error,
+) {
+	t.Helper()
+
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.Header().Set("Content-Encoding", "gzip")
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = w.Write([]byte("not-a-valid-gzip-stream"))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	err := invoke(executor, auth, payload)
+	if err == nil {
+		t.Fatal("expected error, got nil")
+	}
+	if !strings.Contains(err.Error(), "failed to decode error response body") {
+		t.Fatalf("expected decode failure message, got: %v", err)
+	}
+	if statusProvider, ok := err.(interface{ StatusCode() int }); !ok || statusProvider.StatusCode() != http.StatusBadRequest {
+		t.Fatalf("expected status code 400, got: %v", err)
+	}
+}

internal/runtime/executor/gemini_cli_executor.go

@@ -16,7 +16,6 @@ import (
 	"strings"
 	"time"
 
-	"github.com/gin-gonic/gin"
 	"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/runtime/geminicli"
@@ -81,7 +80,7 @@ func (e *GeminiCLIExecutor) PrepareRequest(req *http.Request, auth *cliproxyauth
 		return statusErr{code: http.StatusUnauthorized, msg: "missing access token"}
 	}
 	req.Header.Set("Authorization", "Bearer "+tok.AccessToken)
-	applyGeminiCLIHeaders(req)
+	applyGeminiCLIHeaders(req, "unknown")
 	return nil
 }
 
@@ -189,7 +188,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 		}
 		reqHTTP.Header.Set("Content-Type", "application/json")
 		reqHTTP.Header.Set("Authorization", "Bearer "+tok.AccessToken)
-		applyGeminiCLIHeaders(reqHTTP)
+		applyGeminiCLIHeaders(reqHTTP, attemptModel)
 		reqHTTP.Header.Set("Accept", "application/json")
 		recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 			URL:       url,
@@ -334,7 +333,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 		}
 		reqHTTP.Header.Set("Content-Type", "application/json")
 		reqHTTP.Header.Set("Authorization", "Bearer "+tok.AccessToken)
-		applyGeminiCLIHeaders(reqHTTP)
+		applyGeminiCLIHeaders(reqHTTP, attemptModel)
 		reqHTTP.Header.Set("Accept", "text/event-stream")
 		recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 			URL:       url,
@@ -515,7 +514,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 		}
 		reqHTTP.Header.Set("Content-Type", "application/json")
 		reqHTTP.Header.Set("Authorization", "Bearer "+tok.AccessToken)
-		applyGeminiCLIHeaders(reqHTTP)
+		applyGeminiCLIHeaders(reqHTTP, baseModel)
 		reqHTTP.Header.Set("Accept", "application/json")
 		recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 			URL:       url,
@@ -738,21 +737,11 @@ func stringValue(m map[string]any, key string) string {
 }
 
 // applyGeminiCLIHeaders sets required headers for the Gemini CLI upstream.
-func applyGeminiCLIHeaders(r *http.Request) {
-	var ginHeaders http.Header
-	if ginCtx, ok := r.Context().Value("gin").(*gin.Context); ok && ginCtx != nil && ginCtx.Request != nil {
-		ginHeaders = ginCtx.Request.Header
-	}
-
-	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", "google-api-nodejs-client/9.15.1")
-	misc.EnsureHeader(r.Header, ginHeaders, "X-Goog-Api-Client", "gl-node/22.17.0")
-	misc.EnsureHeader(r.Header, ginHeaders, "Client-Metadata", geminiCLIClientMetadata())
-}
-
-// geminiCLIClientMetadata returns a compact metadata string required by upstream.
-func geminiCLIClientMetadata() string {
-	// Keep parity with CLI client defaults
-	return "ideType=IDE_UNSPECIFIED,platform=PLATFORM_UNSPECIFIED,pluginType=GEMINI"
+// User-Agent is always forced to the GeminiCLI format regardless of the client's value,
+// so that upstream identifies the request as a native GeminiCLI client.
+func applyGeminiCLIHeaders(r *http.Request, model string) {
+	r.Header.Set("User-Agent", misc.GeminiCLIUserAgent(model))
+	r.Header.Set("X-Goog-Api-Client", misc.GeminiCLIApiClientHeader)
 }
 
 // cliPreviewFallbackOrder returns preview model candidates for a base model.

internal/thinking/apply.go

@@ -293,7 +293,7 @@ func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat stri
 	if config.Mode != ModeLevel {
 		return config
 	}
-	if !isBudgetBasedProvider(toFormat) || !isLevelBasedProvider(fromFormat) {
+	if !isBudgetCapableProvider(toFormat) {
 		return config
 	}
 	budget, ok := ConvertLevelToBudget(string(config.Level))
@@ -353,6 +353,26 @@ func extractClaudeConfig(body []byte) ThinkingConfig {
 	if thinkingType == "disabled" {
 		return ThinkingConfig{Mode: ModeNone, Budget: 0}
 	}
+	if thinkingType == "adaptive" || thinkingType == "auto" {
+		// Claude adaptive thinking uses output_config.effort (low/medium/high/max).
+		// We only treat it as a thinking config when effort is explicitly present;
+		// otherwise we passthrough and let upstream defaults apply.
+		if effort := gjson.GetBytes(body, "output_config.effort"); effort.Exists() && effort.Type == gjson.String {
+			value := strings.ToLower(strings.TrimSpace(effort.String()))
+			if value == "" {
+				return ThinkingConfig{}
+			}
+			switch value {
+			case "none":
+				return ThinkingConfig{Mode: ModeNone, Budget: 0}
+			case "auto":
+				return ThinkingConfig{Mode: ModeAuto, Budget: -1}
+			default:
+				return ThinkingConfig{Mode: ModeLevel, Level: ThinkingLevel(value)}
+			}
+		}
+		return ThinkingConfig{}
+	}
 
 	// Check budget_tokens
 	if budget := gjson.GetBytes(body, "thinking.budget_tokens"); budget.Exists() {

internal/thinking/convert.go

@@ -16,6 +16,9 @@ var levelToBudgetMap = map[string]int{
 	"medium":  8192,
 	"high":    24576,
 	"xhigh":   32768,
+	// "max" is used by Claude adaptive thinking effort. We map it to a large budget
+	// and rely on per-model clamping when converting to budget-only providers.
+	"max": 128000,
 }
 
 // ConvertLevelToBudget converts a thinking level to a budget value.
@@ -31,6 +34,7 @@ var levelToBudgetMap = map[string]int{
 //   - medium  → 8192
 //   - high    → 24576
 //   - xhigh   → 32768
+//   - max     → 128000
 //
 // Returns:
 //   - budget: The converted budget value
@@ -92,6 +96,43 @@ func ConvertBudgetToLevel(budget int) (string, bool) {
 	}
 }
 
+// HasLevel reports whether the given target level exists in the levels slice.
+// Matching is case-insensitive with leading/trailing whitespace trimmed.
+func HasLevel(levels []string, target string) bool {
+	for _, level := range levels {
+		if strings.EqualFold(strings.TrimSpace(level), target) {
+			return true
+		}
+	}
+	return false
+}
+
+// MapToClaudeEffort maps a generic thinking level string to a Claude adaptive
+// thinking effort value (low/medium/high/max).
+//
+// supportsMax indicates whether the target model supports "max" effort.
+// Returns the mapped effort and true if the level is valid, or ("", false) otherwise.
+func MapToClaudeEffort(level string, supportsMax bool) (string, bool) {
+	level = strings.ToLower(strings.TrimSpace(level))
+	switch level {
+	case "":
+		return "", false
+	case "minimal":
+		return "low", true
+	case "low", "medium", "high":
+		return level, true
+	case "xhigh", "max":
+		if supportsMax {
+			return "max", true
+		}
+		return "high", true
+	case "auto":
+		return "high", true
+	default:
+		return "", false
+	}
+}
+
 // ModelCapability describes the thinking format support of a model.
 type ModelCapability int
 

internal/thinking/provider/claude/apply.go

@@ -1,8 +1,10 @@
 // Package claude implements thinking configuration scaffolding for Claude models.
 //
-// Claude models use the thinking.budget_tokens format with values in the range
-// 1024-128000. Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5),
-// while older models do not.
+// Claude models support two thinking control styles:
+//   - Manual thinking: thinking.type="enabled" with thinking.budget_tokens (token budget)
+//   - Adaptive thinking (Claude 4.6): thinking.type="adaptive" with output_config.effort (low/medium/high/max)
+//
+// Some Claude models support ZeroAllowed (sonnet-4-5, opus-4-5), while older models do not.
 // See: _bmad-output/planning-artifacts/architecture.md#Epic-6
 package claude
 
@@ -34,7 +36,11 @@ func init() {
 //   - Budget clamping to model range
 //   - ZeroAllowed constraint enforcement
 //
-// Apply only processes ModeBudget and ModeNone; other modes are passed through unchanged.
+// Apply processes:
+//   - ModeBudget: manual thinking budget_tokens
+//   - ModeLevel: adaptive thinking effort (Claude 4.6)
+//   - ModeAuto: provider default adaptive/manual behavior
+//   - ModeNone: disabled
 //
 // Expected output format when enabled:
 //
@@ -45,6 +51,17 @@ func init() {
 //	  }
 //	}
 //
+// Expected output format for adaptive:
+//
+//	{
+//	  "thinking": {
+//	    "type": "adaptive"
+//	  },
+//	  "output_config": {
+//	    "effort": "high"
+//	  }
+//	}
+//
 // Expected output format when disabled:
 //
 //	{
@@ -60,30 +77,91 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 		return body, nil
 	}
 
-	// Only process ModeBudget and ModeNone; other modes pass through
-	// (caller should use ValidateConfig first to normalize modes)
-	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone {
-		return body, nil
-	}
-
 	if len(body) == 0 || !gjson.ValidBytes(body) {
 		body = []byte(`{}`)
 	}
 
-	// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced)
-	// Decide enabled/disabled based on budget value
-	if config.Budget == 0 {
+	supportsAdaptive := modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0
+
+	switch config.Mode {
+	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
+
+	case thinking.ModeLevel:
+		// Adaptive thinking effort is only valid when the model advertises discrete levels.
+		// (Claude 4.6 uses output_config.effort.)
+		if supportsAdaptive && config.Level != "" {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
+			return result, nil
+		}
+
+		// Fallback for non-adaptive Claude models: convert level to budget_tokens.
+		if budget, ok := thinking.ConvertLevelToBudget(string(config.Level)); ok {
+			config.Mode = thinking.ModeBudget
+			config.Budget = budget
+			config.Level = ""
+		} else {
+			return body, nil
+		}
+		fallthrough
+
+	case thinking.ModeBudget:
+		// Budget is expected to be pre-validated by ValidateConfig (clamped, ZeroAllowed enforced).
+		// Decide enabled/disabled based on budget value.
+		if config.Budget == 0 {
+			result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			result, _ = sjson.DeleteBytes(result, "output_config.effort")
+			if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+				result, _ = sjson.DeleteBytes(result, "output_config")
+			}
+			return result, nil
+		}
+
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+
+		// Ensure max_tokens > thinking.budget_tokens (Anthropic API constraint).
+		result = a.normalizeClaudeBudget(result, config.Budget, modelInfo)
+		return result, nil
+
+	case thinking.ModeAuto:
+		// For Claude 4.6 models, auto maps to adaptive thinking with upstream defaults.
+		if supportsAdaptive {
+			result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+			result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+			// Explicit effort is optional for adaptive thinking; omit it to allow upstream default.
+			result, _ = sjson.DeleteBytes(result, "output_config.effort")
+			if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+				result, _ = sjson.DeleteBytes(result, "output_config")
+			}
+			return result, nil
+		}
+
+		// Legacy fallback: enable thinking without specifying budget_tokens.
+		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+		return result, nil
+
+	default:
+		return body, nil
 	}
-
-	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.
@@ -141,7 +219,7 @@ func (a *Applier) effectiveMaxTokens(body []byte, modelInfo *registry.ModelInfo)
 }
 
 func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto {
+	if config.Mode != thinking.ModeBudget && config.Mode != thinking.ModeNone && config.Mode != thinking.ModeAuto && config.Mode != thinking.ModeLevel {
 		return body, nil
 	}
 
@@ -153,14 +231,36 @@ func applyCompatibleClaude(body []byte, config thinking.ThinkingConfig) ([]byte,
 	case thinking.ModeNone:
 		result, _ := sjson.SetBytes(body, "thinking.type", "disabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
 	case thinking.ModeAuto:
 		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+		return result, nil
+	case thinking.ModeLevel:
+		// For user-defined models, interpret ModeLevel as Claude adaptive thinking effort.
+		// Upstream is responsible for validating whether the target model supports it.
+		if config.Level == "" {
+			return body, nil
+		}
+		result, _ := sjson.SetBytes(body, "thinking.type", "adaptive")
+		result, _ = sjson.DeleteBytes(result, "thinking.budget_tokens")
+		result, _ = sjson.SetBytes(result, "output_config.effort", string(config.Level))
 		return result, nil
 	default:
 		result, _ := sjson.SetBytes(body, "thinking.type", "enabled")
 		result, _ = sjson.SetBytes(result, "thinking.budget_tokens", config.Budget)
+		result, _ = sjson.DeleteBytes(result, "output_config.effort")
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
 		return result, nil
 	}
 }

internal/thinking/provider/codex/apply.go

@@ -7,8 +7,6 @@
 package codex
 
 import (
-	"strings"
-
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
@@ -68,7 +66,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	effort := ""
 	support := modelInfo.Thinking
 	if config.Budget == 0 {
-		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
 			effort = string(thinking.LevelNone)
 		}
 	}
@@ -120,12 +118,3 @@ func applyCompatibleCodex(body []byte, config thinking.ThinkingConfig) ([]byte,
 	result, _ := sjson.SetBytes(body, "reasoning.effort", effort)
 	return result, nil
 }
-
-func hasLevel(levels []string, target string) bool {
-	for _, level := range levels {
-		if strings.EqualFold(strings.TrimSpace(level), target) {
-			return true
-		}
-	}
-	return false
-}

internal/thinking/provider/openai/apply.go

@@ -6,8 +6,6 @@
 package openai
 
 import (
-	"strings"
-
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
@@ -65,7 +63,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	effort := ""
 	support := modelInfo.Thinking
 	if config.Budget == 0 {
-		if support.ZeroAllowed || hasLevel(support.Levels, string(thinking.LevelNone)) {
+		if support.ZeroAllowed || thinking.HasLevel(support.Levels, string(thinking.LevelNone)) {
 			effort = string(thinking.LevelNone)
 		}
 	}
@@ -117,12 +115,3 @@ func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte,
 	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
 	return result, nil
 }
-
-func hasLevel(levels []string, target string) bool {
-	for _, level := range levels {
-		if strings.EqualFold(strings.TrimSpace(level), target) {
-			return true
-		}
-	}
-	return false
-}

internal/thinking/strip.go

@@ -30,7 +30,7 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 	var paths []string
 	switch provider {
 	case "claude":
-		paths = []string{"thinking"}
+		paths = []string{"thinking", "output_config.effort"}
 	case "gemini":
 		paths = []string{"generationConfig.thinkingConfig"}
 	case "gemini-cli", "antigravity":
@@ -59,5 +59,12 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 	for _, path := range paths {
 		result, _ = sjson.DeleteBytes(result, path)
 	}
+
+	// Avoid leaving an empty output_config object for Claude when effort was the only field.
+	if provider == "claude" {
+		if oc := gjson.GetBytes(result, "output_config"); oc.Exists() && oc.IsObject() && len(oc.Map()) == 0 {
+			result, _ = sjson.DeleteBytes(result, "output_config")
+		}
+	}
 	return result
 }

internal/thinking/suffix.go

@@ -109,7 +109,7 @@ func ParseSpecialSuffix(rawSuffix string) (mode ThinkingMode, ok bool) {
 // ParseLevelSuffix attempts to parse a raw suffix as a discrete thinking level.
 //
 // This function parses the raw suffix content (from ParseSuffix.RawSuffix) as a level.
-// Only discrete effort levels are valid: minimal, low, medium, high, xhigh.
+// Only discrete effort levels are valid: minimal, low, medium, high, xhigh, max.
 // Level matching is case-insensitive.
 //
 // Special values (none, auto) are NOT handled by this function; use ParseSpecialSuffix
@@ -140,6 +140,8 @@ func ParseLevelSuffix(rawSuffix string) (level ThinkingLevel, ok bool) {
 		return LevelHigh, true
 	case "xhigh":
 		return LevelXHigh, true
+	case "max":
+		return LevelMax, true
 	default:
 		return "", false
 	}

internal/thinking/types.go

@@ -54,6 +54,9 @@ const (
 	LevelHigh ThinkingLevel = "high"
 	// LevelXHigh sets extra-high thinking effort
 	LevelXHigh ThinkingLevel = "xhigh"
+	// LevelMax sets maximum thinking effort.
+	// This is currently used by Claude 4.6 adaptive thinking (opus supports "max").
+	LevelMax ThinkingLevel = "max"
 )
 
 // ThinkingConfig represents a unified thinking configuration.

internal/thinking/validate.go

@@ -53,7 +53,17 @@ func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, fromFo
 		return &config, nil
 	}
 
-	allowClampUnsupported := isBudgetBasedProvider(fromFormat) && isLevelBasedProvider(toFormat)
+	// allowClampUnsupported determines whether to clamp unsupported levels instead of returning an error.
+	// This applies when crossing provider families (e.g., openai→gemini, claude→gemini) and the target
+	// model supports discrete levels. Same-family conversions require strict validation.
+	toCapability := detectModelCapability(modelInfo)
+	toHasLevelSupport := toCapability == CapabilityLevelOnly || toCapability == CapabilityHybrid
+	allowClampUnsupported := toHasLevelSupport && !isSameProviderFamily(fromFormat, toFormat)
+
+	// strictBudget determines whether to enforce strict budget range validation.
+	// This applies when: (1) config comes from request body (not suffix), (2) source format is known,
+	// and (3) source and target are in the same provider family. Cross-family or suffix-based configs
+	// are clamped instead of rejected to improve interoperability.
 	strictBudget := !fromSuffix && fromFormat != "" && isSameProviderFamily(fromFormat, toFormat)
 	budgetDerivedFromLevel := false
 
@@ -201,7 +211,7 @@ func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupp
 }
 
 // standardLevelOrder defines the canonical ordering of thinking levels from lowest to highest.
-var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh}
+var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh, LevelMax}
 
 // clampLevel clamps the given level to the nearest supported level.
 // On tie, prefers the lower level.
@@ -325,7 +335,9 @@ func normalizeLevels(levels []string) []string {
 	return out
 }
 
-func isBudgetBasedProvider(provider string) bool {
+// isBudgetCapableProvider returns true if the provider supports budget-based thinking.
+// These providers may also support level-based thinking (hybrid models).
+func isBudgetCapableProvider(provider string) bool {
 	switch provider {
 	case "gemini", "gemini-cli", "antigravity", "claude":
 		return true
@@ -334,15 +346,6 @@ func isBudgetBasedProvider(provider string) bool {
 	}
 }
 
-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":
@@ -352,11 +355,21 @@ func isGeminiFamily(provider string) bool {
 	}
 }
 
+func isOpenAIFamily(provider string) bool {
+	switch provider {
+	case "openai", "openai-response", "codex":
+		return true
+	default:
+		return false
+	}
+}
+
 func isSameProviderFamily(from, to string) bool {
 	if from == to {
 		return true
 	}
-	return isGeminiFamily(from) && isGeminiFamily(to)
+	return (isGeminiFamily(from) && isGeminiFamily(to)) ||
+		(isOpenAIFamily(from) && isOpenAIFamily(to))
 }
 
 func abs(x int) int {

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -440,14 +440,8 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
-		case "auto":
-			// Amp sends thinking.type="auto" — use max budget from model config
-			// Antigravity API for Claude models requires a concrete positive budget,
-			// not -1. Use a high default that ApplyThinking will cap to model max.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", 64000)
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
-		case "adaptive":
-			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+		case "adaptive", "auto":
+			// Keep adaptive/auto as a high level sentinel; ApplyThinking resolves it
 			// to model-specific max capability.
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)

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

@@ -34,6 +34,11 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
+	// Let user-provided generationConfig pass through
+	if genConfig := gjson.GetBytes(rawJSON, "generationConfig"); genConfig.Exists() {
+		out, _ = sjson.SetRawBytes(out, "request.generationConfig", []byte(genConfig.Raw))
+	}
+
 	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini CLI thinkingConfig.
 	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")

internal/translator/claude/gemini/claude_gemini_request.go

@@ -14,6 +14,7 @@ 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"
@@ -115,24 +116,47 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 		// Include thoughts configuration for reasoning process visibility
 		// Translator only does format conversion, ApplyThinking handles model capability validation.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
 			thinkingLevel := thinkingConfig.Get("thinkingLevel")
 			if !thinkingLevel.Exists() {
 				thinkingLevel = thinkingConfig.Get("thinking_level")
 			}
 			if 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 {
+				if supportsAdaptive {
+					switch level {
+					case "":
+					case "none":
+						out, _ = sjson.Set(out, "thinking.type", "disabled")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						out, _ = sjson.Delete(out, "output_config.effort")
+					default:
+						if mapped, ok := thinking.MapToClaudeEffort(level, supportsMax); ok {
+							level = mapped
+						}
+						out, _ = sjson.Set(out, "thinking.type", "adaptive")
+						out, _ = sjson.Delete(out, "thinking.budget_tokens")
+						out, _ = sjson.Set(out, "output_config.effort", level)
+					}
+				} else {
+					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.Set(out, "thinking.budget_tokens", budget)
+						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 {
@@ -142,16 +166,35 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 				}
 				if 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)
+					if supportsAdaptive {
+						switch budget {
+						case 0:
+							out, _ = sjson.Set(out, "thinking.type", "disabled")
+							out, _ = sjson.Delete(out, "thinking.budget_tokens")
+							out, _ = sjson.Delete(out, "output_config.effort")
+						default:
+							level, ok := thinking.ConvertBudgetToLevel(budget)
+							if ok {
+								if mapped, okM := thinking.MapToClaudeEffort(level, supportsMax); okM {
+									level = mapped
+								}
+								out, _ = sjson.Set(out, "thinking.type", "adaptive")
+								out, _ = sjson.Delete(out, "thinking.budget_tokens")
+								out, _ = sjson.Set(out, "output_config.effort", level)
+							}
+						}
+					} else {
+						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")

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

@@ -14,6 +14,7 @@ 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/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -68,17 +69,45 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 	if v := root.Get("reasoning_effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if ok {
-				switch budget {
-				case 0:
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
+			if supportsAdaptive {
+				switch effort {
+				case "none":
 					out, _ = sjson.Set(out, "thinking.type", "disabled")
-				case -1:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
+				case "auto":
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					if budget > 0 {
+					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
+						effort = mapped
+					}
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Set(out, "output_config.effort", effort)
+				}
+			} else {
+				// Legacy/manual thinking (budget_tokens).
+				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")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					default:
+						if budget > 0 {
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				}
 			}

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

@@ -9,6 +9,7 @@ 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/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -56,17 +57,45 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 	if v := root.Get("reasoning.effort"); v.Exists() {
 		effort := strings.ToLower(strings.TrimSpace(v.String()))
 		if effort != "" {
-			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if ok {
-				switch budget {
-				case 0:
+			mi := registry.LookupModelInfo(modelName, "claude")
+			supportsAdaptive := mi != nil && mi.Thinking != nil && len(mi.Thinking.Levels) > 0
+			supportsMax := supportsAdaptive && thinking.HasLevel(mi.Thinking.Levels, string(thinking.LevelMax))
+
+			// Claude 4.6 supports adaptive thinking with output_config.effort.
+			// MapToClaudeEffort normalizes levels (e.g. minimal→low, xhigh→high) to avoid
+			// validation errors since validate treats same-provider unsupported levels as errors.
+			if supportsAdaptive {
+				switch effort {
+				case "none":
 					out, _ = sjson.Set(out, "thinking.type", "disabled")
-				case -1:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
+				case "auto":
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Delete(out, "output_config.effort")
 				default:
-					if budget > 0 {
+					if mapped, ok := thinking.MapToClaudeEffort(effort, supportsMax); ok {
+						effort = mapped
+					}
+					out, _ = sjson.Set(out, "thinking.type", "adaptive")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+					out, _ = sjson.Set(out, "output_config.effort", effort)
+				}
+			} else {
+				// Legacy/manual thinking (budget_tokens).
+				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")
-						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					default:
+						if budget > 0 {
+							out, _ = sjson.Set(out, "thinking.type", "enabled")
+							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+						}
 					}
 				}
 			}

internal/translator/codex/claude/codex_claude_request.go

@@ -160,7 +160,51 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 						flushMessage()
 						functionCallOutputMessage := `{"type":"function_call_output"}`
 						functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "call_id", messageContentResult.Get("tool_use_id").String())
-						functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+
+						contentResult := messageContentResult.Get("content")
+						if contentResult.IsArray() {
+							toolResultContentIndex := 0
+							toolResultContent := `[]`
+							contentResults := contentResult.Array()
+							for k := 0; k < len(contentResults); k++ {
+								toolResultContentType := contentResults[k].Get("type").String()
+								if toolResultContentType == "image" {
+									sourceResult := contentResults[k].Get("source")
+									if sourceResult.Exists() {
+										data := sourceResult.Get("data").String()
+										if data == "" {
+											data = sourceResult.Get("base64").String()
+										}
+										if data != "" {
+											mediaType := sourceResult.Get("media_type").String()
+											if mediaType == "" {
+												mediaType = sourceResult.Get("mime_type").String()
+											}
+											if mediaType == "" {
+												mediaType = "application/octet-stream"
+											}
+											dataURL := fmt.Sprintf("data:%s;base64,%s", mediaType, data)
+
+											toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.type", toolResultContentIndex), "input_image")
+											toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.image_url", toolResultContentIndex), dataURL)
+											toolResultContentIndex++
+										}
+									}
+								} else if toolResultContentType == "text" {
+									toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.type", toolResultContentIndex), "input_text")
+									toolResultContent, _ = sjson.Set(toolResultContent, fmt.Sprintf("%d.text", toolResultContentIndex), contentResults[k].Get("text").String())
+									toolResultContentIndex++
+								}
+							}
+							if toolResultContent != `[]` {
+								functionCallOutputMessage, _ = sjson.SetRaw(functionCallOutputMessage, "output", toolResultContent)
+							} else {
+								functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+							}
+						} else {
+							functionCallOutputMessage, _ = sjson.Set(functionCallOutputMessage, "output", messageContentResult.Get("content").String())
+						}
+
 						template, _ = sjson.SetRaw(template, "input.-1", functionCallOutputMessage)
 					}
 				}
@@ -230,10 +274,18 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 					reasoningEffort = effort
 				}
 			}
-		case "adaptive":
-			// Claude adaptive means "enable with max capacity"; keep it as highest level
-			// and let ApplyThinking normalize per target model capability.
-			reasoningEffort = string(thinking.LevelXHigh)
+		case "adaptive", "auto":
+			// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
+			// Pass through directly; ApplyThinking handles clamping to target model's levels.
+			effort := ""
+			if v := rootResult.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				reasoningEffort = effort
+			} else {
+				reasoningEffort = string(thinking.LevelXHigh)
+			}
 		case "disabled":
 			if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 				reasoningEffort = effort

internal/translator/codex/claude/codex_claude_response.go

@@ -22,8 +22,8 @@ var (
 
 // ConvertCodexResponseToClaudeParams holds parameters for response conversion.
 type ConvertCodexResponseToClaudeParams struct {
-	HasToolCall              bool
-	BlockIndex               int
+	HasToolCall               bool
+	BlockIndex                int
 	HasReceivedArgumentsDelta bool
 }
 

internal/translator/codex/openai/responses/codex_openai-responses_request_test.go

@@ -264,18 +264,18 @@ func TestConvertSystemRoleToDeveloper_AssistantRole(t *testing.T) {
 	}
 }
 
-func TestUserFieldDeletion(t *testing.T) {  
+func TestUserFieldDeletion(t *testing.T) {
 	inputJSON := []byte(`{  
 		"model": "gpt-5.2",  
 		"user": "test-user",  
 		"input": [{"role": "user", "content": "Hello"}]  
-	}`)  
-	  
-	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)  
-	outputStr := string(output)  
-	  
-	// Verify user field is deleted  
-	userField := gjson.Get(outputStr, "user")  
+	}`)
+
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)
+	outputStr := string(output)
+
+	// Verify user field is deleted
+	userField := gjson.Get(outputStr, "user")
 	if userField.Exists() {
 		t.Errorf("user field should be deleted, but it was found with value: %s", userField.Raw)
 	}

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

@@ -171,7 +171,8 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 		}
 	}
 
-	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
+	// Map Anthropic thinking -> Gemini CLI thinkingConfig when enabled
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
 		case "enabled":
@@ -180,10 +181,20 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
-		case "adaptive":
-			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+		case "adaptive", "auto":
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit high.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			}
 			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

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

@@ -34,6 +34,11 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
+	// Let user-provided generationConfig pass through
+	if genConfig := gjson.GetBytes(rawJSON, "generationConfig"); genConfig.Exists() {
+		out, _ = sjson.SetRawBytes(out, "request.generationConfig", []byte(genConfig.Raw))
+	}
+
 	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini CLI thinkingConfig.
 	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")

internal/translator/gemini/claude/gemini_claude_request.go

@@ -9,6 +9,7 @@ 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"
@@ -151,7 +152,7 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 		}
 	}
 
-	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
+	// Map Anthropic thinking -> Gemini thinking config when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
 		switch t.Get("type").String() {
@@ -161,10 +162,28 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
-		case "adaptive":
-			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
-			// to model-specific max capability.
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+		case "adaptive", "auto":
+			// For adaptive thinking:
+			// - If output_config.effort is explicitly present, pass through as thinkingLevel.
+			// - Otherwise, treat it as "enabled with target-model maximum" and emit thinkingBudget=max.
+			// ApplyThinking handles clamping to target model's supported levels.
+			effort := ""
+			if v := gjson.GetBytes(rawJSON, "output_config.effort"); v.Exists() && v.Type == gjson.String {
+				effort = strings.ToLower(strings.TrimSpace(v.String()))
+			}
+			if effort != "" {
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", effort)
+			} else {
+				maxBudget := 0
+				if mi := registry.LookupModelInfo(modelName, "gemini"); mi != nil && mi.Thinking != nil {
+					maxBudget = mi.Thinking.Max
+				}
+				if maxBudget > 0 {
+					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", maxBudget)
+				} else {
+					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+				}
+			}
 			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}

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

@@ -34,6 +34,11 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	// Model
 	out, _ = sjson.SetBytes(out, "model", modelName)
 
+	// Let user-provided generationConfig pass through
+	if genConfig := gjson.GetBytes(rawJSON, "generationConfig"); genConfig.Exists() {
+		out, _ = sjson.SetRawBytes(out, "generationConfig", []byte(genConfig.Raw))
+	}
+
 	// Apply thinking configuration: convert OpenAI reasoning_effort to Gemini thinkingConfig.
 	// Inline translation-only mapping; capability checks happen later in ApplyThinking.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")

internal/translator/openai/claude/openai_claude_request.go

@@ -75,10 +75,18 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Set(out, "reasoning_effort", effort)
 					}
 				}
-			case "adaptive":
-				// Claude adaptive means "enable with max capacity"; keep it as highest level
-				// and let ApplyThinking normalize per target model capability.
-				out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
+			case "adaptive", "auto":
+				// Adaptive thinking can carry an explicit effort in output_config.effort (Claude 4.6).
+				// Pass through directly; ApplyThinking handles clamping to target model's levels.
+				effort := ""
+				if v := root.Get("output_config.effort"); v.Exists() && v.Type == gjson.String {
+					effort = strings.ToLower(strings.TrimSpace(v.String()))
+				}
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				} else {
+					out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
+				}
 			case "disabled":
 				if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)

sdk/cliproxy/auth/conductor_overrides_test.go

@@ -7,6 +7,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
 )
 
@@ -115,8 +117,19 @@ func newCredentialRetryLimitTestManager(t *testing.T, maxRetryCredentials int) (
 	executor := &credentialRetryLimitExecutor{id: "claude"}
 	m.RegisterExecutor(executor)
 
-	auth1 := &Auth{ID: "auth-1", Provider: "claude"}
-	auth2 := &Auth{ID: "auth-2", Provider: "claude"}
+	baseID := uuid.NewString()
+	auth1 := &Auth{ID: baseID + "-auth-1", Provider: "claude"}
+	auth2 := &Auth{ID: baseID + "-auth-2", Provider: "claude"}
+
+	// Auth selection requires that the global model registry knows each credential supports the model.
+	reg := registry.GetGlobalRegistry()
+	reg.RegisterClient(auth1.ID, "claude", []*registry.ModelInfo{{ID: "test-model"}})
+	reg.RegisterClient(auth2.ID, "claude", []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		reg.UnregisterClient(auth1.ID)
+		reg.UnregisterClient(auth2.ID)
+	})
+
 	if _, errRegister := m.Register(context.Background(), auth1); errRegister != nil {
 		t.Fatalf("register auth1: %v", errRegister)
 	}

test/thinking_conversion_test.go

@@ -34,6 +34,8 @@ type thinkingTestCase struct {
 	inputJSON       string
 	expectField     string
 	expectValue     string
+	expectField2    string
+	expectValue2    string
 	includeThoughts string
 	expectErr       bool
 }
@@ -384,15 +386,17 @@ func TestThinkingE2EMatrix_Suffix(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// Case 30: Effort xhigh → not in low/high → error
+		// Case 30: Effort xhigh → clamped to high
 		{
-			name:        "30",
-			from:        "openai",
-			to:          "gemini",
-			model:       "gemini-mixed-model(xhigh)",
-			inputJSON:   `{"model":"gemini-mixed-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
-			expectField: "",
-			expectErr:   true,
+			name:            "30",
+			from:            "openai",
+			to:              "gemini",
+			model:           "gemini-mixed-model(xhigh)",
+			inputJSON:       `{"model":"gemini-mixed-model(xhigh)","messages":[{"role":"user","content":"hi"}]}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
 		},
 		// Case 31: Effort none → clamped to low (min supported) → includeThoughts=false
 		{
@@ -1782,15 +1786,17 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// Case 30: reasoning_effort=xhigh → error (not in low/high)
+		// Case 30: reasoning_effort=xhigh → clamped to high
 		{
-			name:        "30",
-			from:        "openai",
-			to:          "gemini",
-			model:       "gemini-mixed-model",
-			inputJSON:   `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
-			expectField: "",
-			expectErr:   true,
+			name:            "30",
+			from:            "openai",
+			to:              "gemini",
+			model:           "gemini-mixed-model",
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
 		},
 		// Case 31: reasoning_effort=none → clamped to low → includeThoughts=false
 		{
@@ -2822,9 +2828,8 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 	runThinkingTests(t, cases)
 }
 
-// TestThinkingE2EClaudeAdaptive_Body tests Claude thinking.type=adaptive extended body-only cases.
-// These cases validate that adaptive means "thinking enabled without explicit budget", and
-// cross-protocol conversion should resolve to target-model maximum thinking capability.
+// TestThinkingE2EClaudeAdaptive_Body covers Group 3 cases in docs/thinking-e2e-test-cases.md.
+// It focuses on Claude 4.6 adaptive thinking and effort/level cross-protocol semantics (body-only).
 func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 	reg := registry.GetGlobalRegistry()
 	uid := fmt.Sprintf("thinking-e2e-claude-adaptive-%d", time.Now().UnixNano())
@@ -2833,32 +2838,347 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 	defer reg.UnregisterClient(uid)
 
 	cases := []thinkingTestCase{
-		// A1: Claude adaptive to OpenAI level model -> highest supported level
+		// A subgroup: OpenAI -> Claude (reasoning_effort -> output_config.effort)
 		{
 			name:        "A1",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"minimal"}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "A2",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"low"}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "A3",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"medium"}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "A4",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"high"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "A5",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "A6",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"xhigh"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "A7",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"max"}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "A8",
+			from:        "openai",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"reasoning_effort":"max"}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		// B subgroup: Gemini -> Claude (thinkingLevel/thinkingBudget -> output_config.effort)
+		{
+			name:        "B1",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"minimal"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B2",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"low"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B3",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"medium"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "B4",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"high"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B5",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"xhigh"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "B6",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"xhigh"}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B7",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":512}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B8",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":1024}}}`,
+			expectField: "output_config.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "B9",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":8192}}}`,
+			expectField: "output_config.effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "B10",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":24576}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B11",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-opus-4-6-model",
+			inputJSON:   `{"model":"claude-opus-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":32768}}}`,
+			expectField: "output_config.effort",
+			expectValue: "max",
+			expectErr:   false,
+		},
+		{
+			name:        "B12",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":32768}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "B13",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":0}}}`,
+			expectField: "thinking.type",
+			expectValue: "disabled",
+			expectErr:   false,
+		},
+		{
+			name:        "B14",
+			from:        "gemini",
+			to:          "claude",
+			model:       "claude-sonnet-4-6-model",
+			inputJSON:   `{"model":"claude-sonnet-4-6-model","contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"thinkingConfig":{"thinkingBudget":-1}}}`,
+			expectField: "output_config.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		// C subgroup: Claude adaptive + effort cross-protocol conversion
+		{
+			name:        "C1",
 			from:        "claude",
 			to:          "openai",
 			model:       "level-model",
-			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "minimal",
+			expectErr:   false,
+		},
+		{
+			name:        "C2",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "C3",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "medium",
+			expectErr:   false,
+		},
+		{
+			name:        "C4",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField: "reasoning_effort",
 			expectValue: "high",
 			expectErr:   false,
 		},
-		// A2: Claude adaptive to Gemini level subset model -> highest supported level
 		{
-			name:            "A2",
+			name:        "C5",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C6",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C7",
+			from:        "claude",
+			to:          "openai",
+			model:       "no-thinking-model",
+			inputJSON:   `{"model":"no-thinking-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "",
+			expectErr:   false,
+		},
+
+		{
+			name:            "C8",
 			from:            "claude",
 			to:              "gemini",
 			model:           "level-subset-model",
-			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A3: Claude adaptive to Gemini budget model -> max budget
 		{
-			name:            "A3",
+			name:            "C9",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "1024",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C10",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "8192",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C11",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		{
+			name:            "C12",
 			from:            "claude",
 			to:              "gemini",
 			model:           "gemini-budget-model",
@@ -2868,32 +3188,91 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A4: Claude adaptive to Gemini mixed model -> highest supported level
 		{
-			name:            "A4",
+			name:            "C13",
 			from:            "claude",
 			to:              "gemini",
 			model:           "gemini-mixed-model",
-			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
 			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
 			expectValue:     "high",
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A5: Claude adaptive passthrough for same protocol
+
 		{
-			name:        "A5",
+			name:        "C14",
 			from:        "claude",
-			to:          "claude",
-			model:       "claude-budget-model",
-			inputJSON:   `{"model":"claude-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
-			expectField: "thinking.type",
-			expectValue: "adaptive",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "minimal",
 			expectErr:   false,
 		},
-		// A6: Claude adaptive to Antigravity budget model -> max budget
 		{
-			name:            "A6",
+			name:        "C15",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"low"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "low",
+			expectErr:   false,
+		},
+		{
+			name:        "C16",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C17",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		{
+			name:        "C18",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+
+		{
+			name:        "C19",
+			from:        "claude",
+			to:          "iflow",
+			model:       "glm-test",
+			inputJSON:   `{"model":"glm-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"minimal"}}`,
+			expectField: "chat_template_kwargs.enable_thinking",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		{
+			name:        "C20",
+			from:        "claude",
+			to:          "iflow",
+			model:       "minimax-test",
+			inputJSON:   `{"model":"minimax-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField: "reasoning_split",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		{
+			name:            "C21",
 			from:            "claude",
 			to:              "antigravity",
 			model:           "antigravity-budget-model",
@@ -2903,48 +3282,66 @@ func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
 			includeThoughts: "true",
 			expectErr:       false,
 		},
-		// A7: Claude adaptive to iFlow GLM -> enabled boolean
+
 		{
-			name:        "A7",
-			from:        "claude",
-			to:          "iflow",
-			model:       "glm-test",
-			inputJSON:   `{"model":"glm-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
-			expectField: "chat_template_kwargs.enable_thinking",
-			expectValue: "true",
-			expectErr:   false,
+			name:         "C22",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-sonnet-4-6-model",
+			inputJSON:    `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"medium"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "medium",
+			expectErr:    false,
 		},
-		// A8: Claude adaptive to iFlow MiniMax -> enabled boolean
 		{
-			name:        "A8",
-			from:        "claude",
-			to:          "iflow",
-			model:       "minimax-test",
-			inputJSON:   `{"model":"minimax-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
-			expectField: "reasoning_split",
-			expectValue: "true",
-			expectErr:   false,
+			name:         "C23",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-opus-4-6-model",
+			inputJSON:    `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "max",
+			expectErr:    false,
 		},
-		// A9: Claude adaptive to Codex level model -> highest supported level
 		{
-			name:        "A9",
-			from:        "claude",
-			to:          "codex",
-			model:       "level-model",
-			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
-			expectField: "reasoning.effort",
-			expectValue: "high",
-			expectErr:   false,
+			name:      "C24",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-opus-4-6-model",
+			inputJSON: `{"model":"claude-opus-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectErr: true,
 		},
-		// A10: Claude adaptive on non-thinking model should still be stripped
 		{
-			name:        "A10",
-			from:        "claude",
-			to:          "openai",
-			model:       "no-thinking-model",
-			inputJSON:   `{"model":"no-thinking-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
-			expectField: "",
-			expectErr:   false,
+			name:         "C25",
+			from:         "claude",
+			to:           "claude",
+			model:        "claude-sonnet-4-6-model",
+			inputJSON:    `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"high"}}`,
+			expectField:  "thinking.type",
+			expectValue:  "adaptive",
+			expectField2: "output_config.effort",
+			expectValue2: "high",
+			expectErr:    false,
+		},
+		{
+			name:      "C26",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-sonnet-4-6-model",
+			inputJSON: `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"max"}}`,
+			expectErr: true,
+		},
+		{
+			name:      "C27",
+			from:      "claude",
+			to:        "claude",
+			model:     "claude-sonnet-4-6-model",
+			inputJSON: `{"model":"claude-sonnet-4-6-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"},"output_config":{"effort":"xhigh"}}`,
+			expectErr: true,
 		},
 	}
 
@@ -2999,6 +3396,29 @@ func getTestModels() []*registry.ModelInfo {
 			DisplayName: "Claude Budget Model",
 			Thinking:    &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-sonnet-4-6-model",
+			Object:              "model",
+			Created:             1771372800, // 2026-02-17
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high"}},
+		},
+		{
+			ID:                  "claude-opus-4-6-model",
+			Object:              "model",
+			Created:             1770318000, // 2026-02-05
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Opus",
+			Description:         "Premium model combining maximum intelligence with practical performance",
+			ContextLength:       1000000,
+			MaxCompletionTokens: 128000,
+			Thinking:            &registry.ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false, Levels: []string{"low", "medium", "high", "max"}},
+		},
 		{
 			ID:          "antigravity-budget-model",
 			Object:      "model",
@@ -3165,17 +3585,23 @@ func runThinkingTests(t *testing.T, cases []thinkingTestCase) {
 				return
 			}
 
-			val := gjson.GetBytes(body, tc.expectField)
-			if !val.Exists() {
-				t.Fatalf("expected field %s not found, body=%s", tc.expectField, string(body))
+			assertField := func(fieldPath, expected string) {
+				val := gjson.GetBytes(body, fieldPath)
+				if !val.Exists() {
+					t.Fatalf("expected field %s not found, body=%s", fieldPath, string(body))
+				}
+				actualValue := val.String()
+				if val.Type == gjson.Number {
+					actualValue = fmt.Sprintf("%d", val.Int())
+				}
+				if actualValue != expected {
+					t.Fatalf("field %s: expected %q, got %q, body=%s", fieldPath, expected, actualValue, string(body))
+				}
 			}
 
-			actualValue := val.String()
-			if val.Type == gjson.Number {
-				actualValue = fmt.Sprintf("%d", val.Int())
-			}
-			if actualValue != tc.expectValue {
-				t.Fatalf("field %s: expected %q, got %q, body=%s", tc.expectField, tc.expectValue, actualValue, string(body))
+			assertField(tc.expectField, tc.expectValue)
+			if tc.expectField2 != "" {
+				assertField(tc.expectField2, tc.expectValue2)
 			}
 
 			if tc.includeThoughts != "" && (tc.to == "gemini" || tc.to == "gemini-cli" || tc.to == "antigravity") {