Merge pull request #1435 from tianyicui/fix/haiku-4-5-thinking-support

fix: Enable extended thinking support for Claude Haiku 4.5

README.md

@@ -30,6 +30,10 @@ Get 10% OFF GLM CODING PLAN：https://z.ai/subscribe?ic=8JVLJQFSKB
 <td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
 <td>Thanks to Cubence for sponsoring this project! Cubence is a reliable and efficient API relay service provider, offering relay services for Claude Code, Codex, Gemini, and more. Cubence provides special discounts for our software users: register using <a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">this link</a> and enter the "CLIPROXYAPI" promo code during recharge to get 10% off.</td>
 </tr>
+<tr>
+<td width="180"><a href="https://www.aicodemirror.com/register?invitecode=TJNAIF"><img src="./assets/aicodemirror.png" alt="AICodeMirror" width="150"></a></td>
+<td>Thanks to AICodeMirror for sponsoring this project! AICodeMirror provides official high-stability relay services for Claude Code / Codex / Gemini CLI, with enterprise-grade concurrency, fast invoicing, and 24/7 dedicated technical support. Claude Code / Codex / Gemini official channels at 38% / 2% / 9% of original price, with extra discounts on top-ups! AICodeMirror offers special benefits for CLIProxyAPI users: register via <a href="https://www.aicodemirror.com/register?invitecode=TJNAIF">this link</a> to enjoy 20% off your first top-up, and enterprise customers can get up to 25% off!</td>
+</tr>
 </tbody>
 </table>
 
@@ -142,6 +146,10 @@ A lightweight web admin panel for CLIProxyAPI with health checks, resource monit
 
 A Windows tray application implemented using PowerShell scripts, without relying on any third-party libraries. The main features include: automatic creation of shortcuts, silent running, password management, channel switching (Main / Plus), and automatic downloading and updating.
 
+### [霖君](https://github.com/wangdabaoqq/LinJun)
+
+霖君 is a cross-platform desktop application for managing AI programming assistants, supporting macOS, Windows, and Linux systems. Unified management of Claude Code, Gemini CLI, OpenAI Codex, Qwen Code, and other AI coding tools, with local proxy for multi-account quota tracking and one-click configuration.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 

README_CN.md

@@ -30,6 +30,10 @@ GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元
 <td width="180"><a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa"><img src="./assets/cubence.png" alt="Cubence" width="150"></a></td>
 <td>感谢 Cubence 对本项目的赞助！Cubence 是一家可靠高效的 API 中转服务商，提供 Claude Code、Codex、Gemini 等多种服务的中转。Cubence 为本软件用户提供了特别优惠：使用<a href="https://cubence.com/signup?code=CLIPROXYAPI&source=cpa">此链接</a>注册，并在充值时输入 "CLIPROXYAPI" 优惠码即可享受九折优惠。</td>
 </tr>
+<tr>
+<td width="180"><a href="https://www.aicodemirror.com/register?invitecode=TJNAIF"><img src="./assets/aicodemirror.png" alt="AICodeMirror" width="150"></a></td>
+<td>感谢 AICodeMirror 赞助了本项目！AICodeMirror 提供 Claude Code / Codex / Gemini CLI 官方高稳定中转服务，支持企业级高并发、极速开票、7×24 专属技术支持。 Claude Code / Codex / Gemini 官方渠道低至 3.8 / 0.2 / 0.9 折，充值更有折上折！AICodeMirror 为 CLIProxyAPI 的用户提供了特别福利，通过<a href="https://www.aicodemirror.com/register?invitecode=TJNAIF">此链接</a>注册的用户，可享受首充8折，企业客户最高可享 7.5 折！</td>
+</tr>
 </tbody>
 </table>
 
@@ -137,6 +141,14 @@ Windows 桌面应用，基于 Tauri + React 构建，用于通过 CLIProxyAPI
 
 面向 CLIProxyAPI 的 Web 管理面板，提供健康检查、资源监控、日志查看、自动更新、请求统计与定价展示，支持一键安装与 systemd 服务。
 
+### [CLIProxyAPI Tray](https://github.com/kitephp/CLIProxyAPI_Tray)
+
+Windows 托盘应用，基于 PowerShell 脚本实现，不依赖任何第三方库。主要功能包括：自动创建快捷方式、静默运行、密码管理、通道切换（Main / Plus）以及自动下载与更新。
+
+### [霖君](https://github.com/wangdabaoqq/LinJun)
+
+霖君是一款用于管理AI编程助手的跨平台桌面应用，支持macOS、Windows、Linux系统。统一管理Claude Code、Gemini CLI、OpenAI Codex、Qwen Code等AI编程工具，本地代理实现多账户配额跟踪和一键配置。
+
 > [!NOTE]  
 > 如果你开发了基于 CLIProxyAPI 的项目，请提交一个 PR（拉取请求）将其添加到此列表中。
 
@@ -148,10 +160,6 @@ Windows 桌面应用，基于 Tauri + React 构建，用于通过 CLIProxyAPI
 
 基于 Next.js 的实现，灵感来自 CLIProxyAPI，易于安装使用；自研格式转换（OpenAI/Claude/Gemini/Ollama）、组合系统与自动回退、多账户管理（指数退避）、Next.js Web 控制台，并支持 Cursor、Claude Code、Cline、RooCode 等 CLI 工具，无需 API 密钥。
 
-### [CLIProxyAPI Tray](https://github.com/kitephp/CLIProxyAPI_Tray)
-
-Windows 托盘应用，基于 PowerShell 脚本实现，不依赖任何第三方库。主要功能包括：自动创建快捷方式、静默运行、密码管理、通道切换（Main / Plus）以及自动下载与更新。
-
 > [!NOTE]  
 > 如果你开发了 CLIProxyAPI 的移植或衍生项目，请提交 PR 将其添加到此列表中。
 

config.example.yaml

@@ -40,6 +40,11 @@ api-keys:
 # Enable debug logging
 debug: false
 
+# Enable pprof HTTP debug server (host:port). Keep it bound to localhost for safety.
+pprof:
+  enable: false
+  addr: "127.0.0.1:8316"
+
 # When true, disable high-overhead HTTP middleware features to reduce per-request memory usage under high concurrency.
 commercial-mode: false
 

internal/config/config.go

@@ -18,7 +18,10 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
-const DefaultPanelGitHubRepository = "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+const (
+	DefaultPanelGitHubRepository = "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"
+	DefaultPprofAddr             = "127.0.0.1:8316"
+)
 
 // Config represents the application's configuration, loaded from a YAML file.
 type Config struct {
@@ -41,6 +44,9 @@ type Config struct {
 	// Debug enables or disables debug-level logging and other debug features.
 	Debug bool `yaml:"debug" json:"debug"`
 
+	// Pprof config controls the optional pprof HTTP debug server.
+	Pprof PprofConfig `yaml:"pprof" json:"pprof"`
+
 	// CommercialMode disables high-overhead HTTP middleware features to minimize per-request memory usage.
 	CommercialMode bool `yaml:"commercial-mode" json:"commercial-mode"`
 
@@ -121,6 +127,14 @@ type TLSConfig struct {
 	Key string `yaml:"key" json:"key"`
 }
 
+// PprofConfig holds pprof HTTP server settings.
+type PprofConfig struct {
+	// Enable toggles the pprof HTTP debug server.
+	Enable bool `yaml:"enable" json:"enable"`
+	// Addr is the host:port address for the pprof HTTP server.
+	Addr string `yaml:"addr" json:"addr"`
+}
+
 // RemoteManagement holds management API configuration under 'remote-management'.
 type RemoteManagement struct {
 	// AllowRemote toggles remote (non-localhost) access to management API.
@@ -514,6 +528,8 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	cfg.ErrorLogsMaxFiles = 10
 	cfg.UsageStatisticsEnabled = false
 	cfg.DisableCooling = false
+	cfg.Pprof.Enable = false
+	cfg.Pprof.Addr = DefaultPprofAddr
 	cfg.AmpCode.RestrictManagementToLocalhost = false // Default to false: API key auth is sufficient
 	cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
 	if err = yaml.Unmarshal(data, &cfg); err != nil {
@@ -556,6 +572,11 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 		cfg.RemoteManagement.PanelGitHubRepository = DefaultPanelGitHubRepository
 	}
 
+	cfg.Pprof.Addr = strings.TrimSpace(cfg.Pprof.Addr)
+	if cfg.Pprof.Addr == "" {
+		cfg.Pprof.Addr = DefaultPprofAddr
+	}
+
 	if cfg.LogsMaxTotalSizeMB < 0 {
 		cfg.LogsMaxTotalSizeMB = 0
 	}

internal/config/oauth_model_alias_migration.go

@@ -17,6 +17,7 @@ var antigravityModelConversionTable = map[string]string{
 	"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",
 }
 
 // defaultAntigravityAliases returns the default oauth-model-alias configuration
@@ -30,6 +31,7 @@ func defaultAntigravityAliases() []OAuthModelAlias {
 		{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"},
 	}
 }
 

internal/config/oauth_model_alias_migration_test.go

@@ -131,6 +131,9 @@ func TestMigrateOAuthModelAlias_ConvertsAntigravityModels(t *testing.T) {
 	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) {

internal/logging/global_logger.go

@@ -131,7 +131,10 @@ func ResolveLogDirectory(cfg *config.Config) string {
 		return logDir
 	}
 	if !isDirWritable(logDir) {
-		authDir := strings.TrimSpace(cfg.AuthDir)
+		authDir, err := util.ResolveAuthDir(cfg.AuthDir)
+		if err != nil {
+			log.Warnf("Failed to resolve auth-dir %q for log directory: %v", cfg.AuthDir, err)
+		}
 		if authDir != "" {
 			logDir = filepath.Join(authDir, "logs")
 		}

internal/registry/model_definitions_static_data.go

@@ -15,7 +15,7 @@ func GetClaudeModels() []*ModelInfo {
 			DisplayName:         "Claude 4.5 Haiku",
 			ContextLength:       200000,
 			MaxCompletionTokens: 64000,
-			// Thinking: not supported for Haiku models
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
 		{
 			ID:                  "claude-sonnet-4-5-20250929",
@@ -28,6 +28,18 @@ func GetClaudeModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-opus-4-6",
+			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:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
 		{
 			ID:                  "claude-opus-4-5-20251101",
 			Object:              "model",
@@ -716,6 +728,20 @@ func GetOpenAIModels() []*ModelInfo {
 			SupportedParameters: []string{"tools"},
 			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
+		{
+			ID:                  "gpt-5.3-codex",
+			Object:              "model",
+			Created:             1770307200,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.3",
+			DisplayName:         "GPT 5.3 Codex",
+			Description:         "Stable version of GPT 5.3 Codex, The best model for coding and agentic tasks across domains.",
+			ContextLength:       400000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
 	}
 }
 
@@ -803,6 +829,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: iFlowThinkingSupport},
 		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
 		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
+		{ID: "kimi-k2.5", DisplayName: "Kimi-K2.5", Description: "Moonshot Kimi K2.5", Created: 1769443200, Thinking: iFlowThinkingSupport},
 	}
 	models := make([]*ModelInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -839,6 +866,7 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
+		"claude-opus-4-6-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 128000},
 		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
 		"gpt-oss-120b-medium":        {},
 		"tab_flash_lite_preview":     {},

internal/runtime/executor/aistudio_executor.go

@@ -141,7 +141,7 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 		URL:       endpoint,
 		Method:    http.MethodPost,
 		Headers:   wsReq.Headers.Clone(),
-		Body:      bytes.Clone(body.payload),
+		Body:      body.payload,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -156,14 +156,14 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, wsResp.Status, wsResp.Headers.Clone())
 	if len(wsResp.Body) > 0 {
-		appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(wsResp.Body))
+		appendAPIResponseChunk(ctx, e.cfg, wsResp.Body)
 	}
 	if wsResp.Status < 200 || wsResp.Status >= 300 {
 		return resp, statusErr{code: wsResp.Status, msg: string(wsResp.Body)}
 	}
 	reporter.publish(ctx, parseGeminiUsage(wsResp.Body))
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, body.toFormat, opts.SourceFormat, req.Model, bytes.Clone(opts.OriginalRequest), bytes.Clone(translatedReq), bytes.Clone(wsResp.Body), &param)
+	out := sdktranslator.TranslateNonStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, wsResp.Body, &param)
 	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out))}
 	return resp, nil
 }
@@ -199,7 +199,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		URL:       endpoint,
 		Method:    http.MethodPost,
 		Headers:   wsReq.Headers.Clone(),
-		Body:      bytes.Clone(body.payload),
+		Body:      body.payload,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -225,7 +225,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		}
 		var body bytes.Buffer
 		if len(firstEvent.Payload) > 0 {
-			appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(firstEvent.Payload))
+			appendAPIResponseChunk(ctx, e.cfg, firstEvent.Payload)
 			body.Write(firstEvent.Payload)
 		}
 		if firstEvent.Type == wsrelay.MessageTypeStreamEnd {
@@ -244,7 +244,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 				metadataLogged = true
 			}
 			if len(event.Payload) > 0 {
-				appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(event.Payload))
+				appendAPIResponseChunk(ctx, e.cfg, event.Payload)
 				body.Write(event.Payload)
 			}
 			if event.Type == wsrelay.MessageTypeStreamEnd {
@@ -274,12 +274,12 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 				}
 			case wsrelay.MessageTypeStreamChunk:
 				if len(event.Payload) > 0 {
-					appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(event.Payload))
+					appendAPIResponseChunk(ctx, e.cfg, event.Payload)
 					filtered := FilterSSEUsageMetadata(event.Payload)
 					if detail, ok := parseGeminiStreamUsage(filtered); ok {
 						reporter.publish(ctx, detail)
 					}
-					lines := sdktranslator.TranslateStream(ctx, body.toFormat, opts.SourceFormat, req.Model, bytes.Clone(opts.OriginalRequest), translatedReq, bytes.Clone(filtered), &param)
+					lines := sdktranslator.TranslateStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, filtered, &param)
 					for i := range lines {
 						out <- cliproxyexecutor.StreamChunk{Payload: ensureColonSpacedJSON([]byte(lines[i]))}
 					}
@@ -293,9 +293,9 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 					metadataLogged = true
 				}
 				if len(event.Payload) > 0 {
-					appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(event.Payload))
+					appendAPIResponseChunk(ctx, e.cfg, event.Payload)
 				}
-				lines := sdktranslator.TranslateStream(ctx, body.toFormat, opts.SourceFormat, req.Model, bytes.Clone(opts.OriginalRequest), translatedReq, bytes.Clone(event.Payload), &param)
+				lines := sdktranslator.TranslateStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, event.Payload, &param)
 				for i := range lines {
 					out <- cliproxyexecutor.StreamChunk{Payload: ensureColonSpacedJSON([]byte(lines[i]))}
 				}
@@ -350,7 +350,7 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 		URL:       endpoint,
 		Method:    http.MethodPost,
 		Headers:   wsReq.Headers.Clone(),
-		Body:      bytes.Clone(body.payload),
+		Body:      body.payload,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,
@@ -364,7 +364,7 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 	}
 	recordAPIResponseMetadata(ctx, e.cfg, resp.Status, resp.Headers.Clone())
 	if len(resp.Body) > 0 {
-		appendAPIResponseChunk(ctx, e.cfg, bytes.Clone(resp.Body))
+		appendAPIResponseChunk(ctx, e.cfg, resp.Body)
 	}
 	if resp.Status < 200 || resp.Status >= 300 {
 		return cliproxyexecutor.Response{}, statusErr{code: resp.Status, msg: string(resp.Body)}
@@ -373,7 +373,7 @@ func (e *AIStudioExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.A
 	if totalTokens <= 0 {
 		return cliproxyexecutor.Response{}, fmt.Errorf("wsrelay: totalTokens missing in response")
 	}
-	translated := sdktranslator.TranslateTokenCount(ctx, body.toFormat, opts.SourceFormat, totalTokens, bytes.Clone(resp.Body))
+	translated := sdktranslator.TranslateTokenCount(ctx, body.toFormat, opts.SourceFormat, totalTokens, resp.Body)
 	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }
 
@@ -393,12 +393,13 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
-	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, stream)
 	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
 		return nil, translatedPayload{}, err

internal/runtime/executor/antigravity_executor.go

@@ -133,12 +133,13 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -230,7 +231,7 @@ attemptLoop:
 
 			reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
 			var param any
-			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bodyBytes, &param)
+			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bodyBytes, &param)
 			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
 			reporter.ensurePublished(ctx)
 			return resp, nil
@@ -274,12 +275,13 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -433,7 +435,7 @@ attemptLoop:
 
 			reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
 			var param any
-			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, resp.Payload, &param)
+			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, resp.Payload, &param)
 			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
 			reporter.ensurePublished(ctx)
 
@@ -665,12 +667,13 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("antigravity")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -800,12 +803,12 @@ attemptLoop:
 						reporter.publish(ctx, detail)
 					}
 
-					chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(payload), &param)
+					chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bytes.Clone(payload), &param)
 					for i := range chunks {
 						out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 					}
 				}
-				tail := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, []byte("[DONE]"), &param)
+				tail := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, []byte("[DONE]"), &param)
 				for i := range tail {
 					out <- cliproxyexecutor.StreamChunk{Payload: []byte(tail[i])}
 				}
@@ -872,7 +875,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
 
 	// Prepare payload once (doesn't depend on baseURL)
-	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	payload := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -1280,51 +1283,40 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 	payload = geminiToAntigravity(modelName, payload, projectID)
 	payload, _ = sjson.SetBytes(payload, "model", modelName)
 
-	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high") {
-		strJSON := string(payload)
-		paths := make([]string, 0)
-		util.Walk(gjson.ParseBytes(payload), "", "parametersJsonSchema", &paths)
-		for _, p := range paths {
-			strJSON, _ = util.RenameKey(strJSON, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
-		}
-
-		// Use the centralized schema cleaner to handle unsupported keywords,
-		// const->enum conversion, and flattening of types/anyOf.
-		strJSON = util.CleanJSONSchemaForAntigravity(strJSON)
-		payload = []byte(strJSON)
-	} else {
-		strJSON := string(payload)
-		paths := make([]string, 0)
-		util.Walk(gjson.Parse(strJSON), "", "parametersJsonSchema", &paths)
-		for _, p := range paths {
-			strJSON, _ = util.RenameKey(strJSON, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
-		}
-		// Clean tool schemas for Gemini to remove unsupported JSON Schema keywords
-		// without adding empty-schema placeholders.
-		strJSON = util.CleanJSONSchemaForGemini(strJSON)
-		payload = []byte(strJSON)
+	useAntigravitySchema := strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high")
+	payloadStr := string(payload)
+	paths := make([]string, 0)
+	util.Walk(gjson.Parse(payloadStr), "", "parametersJsonSchema", &paths)
+	for _, p := range paths {
+		payloadStr, _ = util.RenameKey(payloadStr, p, p[:len(p)-len("parametersJsonSchema")]+"parameters")
 	}
 
-	if strings.Contains(modelName, "claude") || strings.Contains(modelName, "gemini-3-pro-high") {
-		systemInstructionPartsResult := gjson.GetBytes(payload, "request.systemInstruction.parts")
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.role", "user")
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.0.text", systemInstruction)
-		payload, _ = sjson.SetBytes(payload, "request.systemInstruction.parts.1.text", fmt.Sprintf("Please ignore following [ignore]%s[/ignore]", systemInstruction))
+	if useAntigravitySchema {
+		payloadStr = util.CleanJSONSchemaForAntigravity(payloadStr)
+	} else {
+		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 systemInstructionPartsResult.Exists() && systemInstructionPartsResult.IsArray() {
 			for _, partResult := range systemInstructionPartsResult.Array() {
-				payload, _ = sjson.SetRawBytes(payload, "request.systemInstruction.parts.-1", []byte(partResult.Raw))
+				payloadStr, _ = sjson.SetRaw(payloadStr, "request.systemInstruction.parts.-1", partResult.Raw)
 			}
 		}
 	}
 
 	if strings.Contains(modelName, "claude") {
-		payload, _ = sjson.SetBytes(payload, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
+		payloadStr, _ = sjson.Set(payloadStr, "request.toolConfig.functionCallingConfig.mode", "VALIDATED")
 	} else {
-		payload, _ = sjson.DeleteBytes(payload, "request.generationConfig.maxOutputTokens")
+		payloadStr, _ = sjson.Delete(payloadStr, "request.generationConfig.maxOutputTokens")
 	}
 
-	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), bytes.NewReader(payload))
+	httpReq, errReq := http.NewRequestWithContext(ctx, http.MethodPost, requestURL.String(), strings.NewReader(payloadStr))
 	if errReq != nil {
 		return nil, errReq
 	}
@@ -1346,11 +1338,15 @@ func (e *AntigravityExecutor) buildRequest(ctx context.Context, auth *cliproxyau
 		authLabel = auth.Label
 		authType, authValue = auth.AccountInfo()
 	}
+	var payloadLog []byte
+	if e.cfg != nil && e.cfg.RequestLog {
+		payloadLog = []byte(payloadStr)
+	}
 	recordAPIRequest(ctx, e.cfg, upstreamRequestLog{
 		URL:       requestURL.String(),
 		Method:    http.MethodPost,
 		Headers:   httpReq.Header.Clone(),
-		Body:      payload,
+		Body:      payloadLog,
 		Provider:  e.Identifier(),
 		AuthID:    authID,
 		AuthLabel: authLabel,

internal/runtime/executor/claude_executor.go

@@ -100,12 +100,13 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, stream)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
@@ -216,7 +217,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		to,
 		from,
 		req.Model,
-		bytes.Clone(opts.OriginalRequest),
+		opts.OriginalRequest,
 		bodyForTranslation,
 		data,
 		&param,
@@ -240,12 +241,13 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	defer reporter.trackFailure(ctx, &err)
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("claude")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
@@ -381,7 +383,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 				to,
 				from,
 				req.Model,
-				bytes.Clone(opts.OriginalRequest),
+				opts.OriginalRequest,
 				bodyForTranslation,
 				bytes.Clone(line),
 				&param,
@@ -411,7 +413,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("claude")
 	// Use streaming translation to preserve function calling, except for claude.
 	stream := from != to
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, stream)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	if !strings.HasPrefix(baseModel, "claude-3-5-haiku") {

internal/runtime/executor/codex_executor.go

@@ -27,6 +27,11 @@ import (
 	"github.com/google/uuid"
 )
 
+const (
+	codexClientVersion = "0.98.0"
+	codexUserAgent     = "codex_cli_rs/0.98.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464"
+)
+
 var dataTag = []byte("data:")
 
 // CodexExecutor is a stateless executor for Codex (OpenAI Responses API entrypoint).
@@ -88,12 +93,13 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -176,7 +182,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 		}
 
 		var param any
-		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, line, &param)
+		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, line, &param)
 		resp = cliproxyexecutor.Response{Payload: []byte(out)}
 		return resp, nil
 	}
@@ -197,12 +203,13 @@ func (e *CodexExecutor) executeCompact(ctx context.Context, auth *cliproxyauth.A
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai-response")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -265,7 +272,7 @@ func (e *CodexExecutor) executeCompact(ctx context.Context, auth *cliproxyauth.A
 	reporter.publish(ctx, parseOpenAIUsage(data))
 	reporter.ensurePublished(ctx)
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -286,12 +293,13 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -378,7 +386,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 				}
 			}
 
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(originalPayload), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, originalPayload, body, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -397,7 +405,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	body, err := thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -634,10 +642,10 @@ func applyCodexHeaders(r *http.Request, auth *cliproxyauth.Auth, token string, s
 		ginHeaders = ginCtx.Request.Header
 	}
 
-	misc.EnsureHeader(r.Header, ginHeaders, "Version", "0.21.0")
+	misc.EnsureHeader(r.Header, ginHeaders, "Version", codexClientVersion)
 	misc.EnsureHeader(r.Header, ginHeaders, "Openai-Beta", "responses=experimental")
 	misc.EnsureHeader(r.Header, ginHeaders, "Session_id", uuid.NewString())
-	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", "codex_cli_rs/0.50.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464")
+	misc.EnsureHeader(r.Header, ginHeaders, "User-Agent", codexUserAgent)
 
 	if stream {
 		r.Header.Set("Accept", "text/event-stream")

internal/runtime/executor/gemini_cli_executor.go

@@ -119,12 +119,13 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -223,7 +224,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 		if httpResp.StatusCode >= 200 && httpResp.StatusCode < 300 {
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
-			out := sdktranslator.TranslateNonStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), payload, data, &param)
+			out := sdktranslator.TranslateNonStream(respCtx, to, from, attemptModel, opts.OriginalRequest, payload, data, &param)
 			resp = cliproxyexecutor.Response{Payload: []byte(out)}
 			return resp, nil
 		}
@@ -272,12 +273,13 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini-cli")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -399,14 +401,14 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 						reporter.publish(ctx, detail)
 					}
 					if bytes.HasPrefix(line, dataTag) {
-						segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone(line), &param)
+						segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, opts.OriginalRequest, reqBody, bytes.Clone(line), &param)
 						for i := range segments {
 							out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 						}
 					}
 				}
 
-				segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
+				segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, opts.OriginalRequest, reqBody, []byte("[DONE]"), &param)
 				for i := range segments {
 					out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 				}
@@ -428,12 +430,12 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			appendAPIResponseChunk(ctx, e.cfg, data)
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
-			segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, data, &param)
+			segments := sdktranslator.TranslateStream(respCtx, to, from, attemptModel, opts.OriginalRequest, reqBody, data, &param)
 			for i := range segments {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 			}
 
-			segments = sdktranslator.TranslateStream(respCtx, to, from, attemptModel, bytes.Clone(opts.OriginalRequest), reqBody, bytes.Clone([]byte("[DONE]")), &param)
+			segments = sdktranslator.TranslateStream(respCtx, to, from, attemptModel, opts.OriginalRequest, reqBody, []byte("[DONE]"), &param)
 			for i := range segments {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(segments[i])}
 			}
@@ -485,7 +487,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	// The loop variable attemptModel is only used as the concrete model id sent to the upstream
 	// Gemini CLI endpoint when iterating fallback variants.
 	for range models {
-		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+		payload := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 		payload, err = thinking.ApplyThinking(payload, req.Model, from.String(), to.String(), e.Identifier())
 		if err != nil {

internal/runtime/executor/gemini_executor.go

@@ -116,12 +116,13 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	// Official Gemini API via API key or OAuth bearer
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -203,7 +204,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -222,12 +223,13 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -318,12 +320,12 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			if detail, ok := parseGeminiStreamUsage(payload); ok {
 				reporter.publish(ctx, detail)
 			}
-			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(payload), &param)
+			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(payload), &param)
 			for i := range lines {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 			}
 		}
-		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone([]byte("[DONE]")), &param)
+		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, []byte("[DONE]"), &param)
 		for i := range lines {
 			out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 		}
@@ -344,7 +346,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
-	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {

internal/runtime/executor/gemini_vertex_executor.go

@@ -318,12 +318,13 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 		from := opts.SourceFormat
 		to := sdktranslator.FromString("gemini")
 
-		originalPayload := bytes.Clone(req.Payload)
+		originalPayloadSource := req.Payload
 		if len(opts.OriginalRequest) > 0 {
-			originalPayload = bytes.Clone(opts.OriginalRequest)
+			originalPayloadSource = opts.OriginalRequest
 		}
+		originalPayload := originalPayloadSource
 		originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-		body = sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+		body = sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 		body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 		if err != nil {
@@ -417,7 +418,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -432,12 +433,13 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -521,7 +523,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -536,12 +538,13 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -632,12 +635,12 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 			if detail, ok := parseGeminiStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(line), &param)
 			for i := range lines {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 			}
 		}
-		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, []byte("[DONE]"), &param)
+		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, []byte("[DONE]"), &param)
 		for i := range lines {
 			out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 		}
@@ -660,12 +663,13 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -756,12 +760,12 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 			if detail, ok := parseGeminiStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(line), &param)
 			for i := range lines {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 			}
 		}
-		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, []byte("[DONE]"), &param)
+		lines := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, []byte("[DONE]"), &param)
 		for i := range lines {
 			out <- cliproxyexecutor.StreamChunk{Payload: []byte(lines[i])}
 		}
@@ -781,7 +785,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 
-	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {
@@ -865,7 +869,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 
-	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String(), e.Identifier())
 	if err != nil {

internal/runtime/executor/iflow_executor.go

@@ -87,12 +87,13 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow", e.Identifier())
@@ -163,7 +164,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	var param any
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -189,12 +190,13 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow", e.Identifier())
@@ -274,7 +276,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			if detail, ok := parseOpenAIStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -296,7 +298,7 @@ func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	enc, err := tokenizerForModel(baseModel)
 	if err != nil {

internal/runtime/executor/logging_helpers.go

@@ -80,7 +80,7 @@ func recordAPIRequest(ctx context.Context, cfg *config.Config, info upstreamRequ
 	writeHeaders(builder, info.Headers)
 	builder.WriteString("\nBody:\n")
 	if len(info.Body) > 0 {
-		builder.WriteString(string(bytes.Clone(info.Body)))
+		builder.WriteString(string(info.Body))
 	} else {
 		builder.WriteString("<empty>")
 	}
@@ -152,7 +152,7 @@ func appendAPIResponseChunk(ctx context.Context, cfg *config.Config, chunk []byt
 	if cfg == nil || !cfg.RequestLog {
 		return
 	}
-	data := bytes.TrimSpace(bytes.Clone(chunk))
+	data := bytes.TrimSpace(chunk)
 	if len(data) == 0 {
 		return
 	}

internal/runtime/executor/openai_compat_executor.go

@@ -88,12 +88,13 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 		to = sdktranslator.FromString("openai-response")
 		endpoint = "/responses/compact"
 	}
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, opts.Stream)
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, opts.Stream)
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated, requestedModel)
 	if opts.Alt == "responses/compact" {
@@ -170,7 +171,7 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	reporter.ensurePublished(ctx)
 	// Translate response back to source format when needed
 	var param any
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, body, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, body, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -189,12 +190,13 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated, requestedModel)
 
@@ -283,7 +285,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 
 			// OpenAI-compatible streams are SSE: lines typically prefixed with "data: ".
 			// Pass through translator; it yields one or more chunks for the target schema.
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), translated, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
@@ -304,7 +306,7 @@ func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyau
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	translated := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	modelForCounting := baseModel
 

internal/runtime/executor/qwen_executor.go

@@ -81,12 +81,13 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
@@ -150,7 +151,7 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	var param any
 	// Note: TranslateNonStream uses req.Model (original with suffix) to preserve
 	// the original model name in the response for client compatibility.
-	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, data, &param)
+	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
 	resp = cliproxyexecutor.Response{Payload: []byte(out)}
 	return resp, nil
 }
@@ -171,12 +172,13 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	originalPayload := bytes.Clone(req.Payload)
+	originalPayloadSource := req.Payload
 	if len(opts.OriginalRequest) > 0 {
-		originalPayload = bytes.Clone(opts.OriginalRequest)
+		originalPayloadSource = opts.OriginalRequest
 	}
+	originalPayload := originalPayloadSource
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
 	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String(), e.Identifier())
@@ -253,12 +255,12 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			if detail, ok := parseOpenAIStreamUsage(line); ok {
 				reporter.publish(ctx, detail)
 			}
-			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone(line), &param)
+			chunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, bytes.Clone(line), &param)
 			for i := range chunks {
 				out <- cliproxyexecutor.StreamChunk{Payload: []byte(chunks[i])}
 			}
 		}
-		doneChunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, bytes.Clone(opts.OriginalRequest), body, bytes.Clone([]byte("[DONE]")), &param)
+		doneChunks := sdktranslator.TranslateStream(ctx, to, from, req.Model, opts.OriginalRequest, body, []byte("[DONE]"), &param)
 		for i := range doneChunks {
 			out <- cliproxyexecutor.StreamChunk{Payload: []byte(doneChunks[i])}
 		}
@@ -276,7 +278,7 @@ func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth,
 
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
-	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
+	body := sdktranslator.TranslateRequest(from, to, baseModel, req.Payload, false)
 
 	modelName := gjson.GetBytes(body, "model").String()
 	if strings.TrimSpace(modelName) == "" {

internal/thinking/apply.go

@@ -388,7 +388,12 @@ func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
 	}
 
 	// Check thinkingLevel first (Gemini 3 format takes precedence)
-	if level := gjson.GetBytes(body, prefix+".thinkingLevel"); level.Exists() {
+	level := gjson.GetBytes(body, prefix+".thinkingLevel")
+	if !level.Exists() {
+		// Google official Gemini Python SDK sends snake_case field names
+		level = gjson.GetBytes(body, prefix+".thinking_level")
+	}
+	if level.Exists() {
 		value := level.String()
 		switch value {
 		case "none":
@@ -401,7 +406,12 @@ func extractGeminiConfig(body []byte, provider string) ThinkingConfig {
 	}
 
 	// Check thinkingBudget (Gemini 2.5 format)
-	if budget := gjson.GetBytes(body, prefix+".thinkingBudget"); budget.Exists() {
+	budget := gjson.GetBytes(body, prefix+".thinkingBudget")
+	if !budget.Exists() {
+		// Google official Gemini Python SDK sends snake_case field names
+		budget = gjson.GetBytes(body, prefix+".thinking_budget")
+	}
+	if budget.Exists() {
 		value := int(budget.Int())
 		switch value {
 		case 0:

internal/thinking/provider/antigravity/apply.go

@@ -94,8 +94,10 @@ func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig, m
 }
 
 func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
@@ -114,28 +116,30 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig, modelInfo *registry.ModelInfo, isClaude bool) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	includeThoughts := false
-	switch config.Mode {
-	case thinking.ModeNone:
-		includeThoughts = false
-	case thinking.ModeAuto:
-		includeThoughts = true
-	default:
-		includeThoughts = budget > 0
-	}
 
-	// Apply Claude-specific constraints
+	// Apply Claude-specific constraints first to get the final budget value
 	if isClaude && modelInfo != nil {
 		budget, result = a.normalizeClaudeBudget(budget, result, modelInfo)
 		// Check if budget was removed entirely
@@ -144,6 +148,37 @@ func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig,
 		}
 	}
 
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
+	// This ensures that when user requests budget=0 (disable thinking output),
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+		return result, nil
+	}
+
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
+	}
+
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil

internal/thinking/provider/gemini/apply.go

@@ -118,8 +118,10 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
 	// ValidateConfig sets config.Level to the lowest level when ModeNone + Budget > 0.
 
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
@@ -138,29 +140,58 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	// ModeNone semantics:
-	//   - ModeNone + Budget=0: completely disable thinking
-	//   - ModeNone + Budget>0: forced to think but hide output (includeThoughts=false)
-	// When ZeroAllowed=false, ValidateConfig clamps Budget to Min while preserving ModeNone.
-	includeThoughts := false
-	switch config.Mode {
-	case thinking.ModeNone:
-		includeThoughts = false
-	case thinking.ModeAuto:
-		includeThoughts = true
-	default:
-		includeThoughts = budget > 0
+
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
+	// This ensures that when user requests budget=0 (disable thinking output),
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)
+		result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.includeThoughts", false)
+		return result, nil
+	}
+
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
 	}
 
 	result, _ = sjson.SetBytes(result, "generationConfig.thinkingConfig.thinkingBudget", budget)

internal/thinking/provider/geminicli/apply.go

@@ -79,8 +79,10 @@ func (a *Applier) applyCompatible(body []byte, config thinking.ThinkingConfig) (
 }
 
 func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingBudget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
@@ -99,25 +101,58 @@ func (a *Applier) applyLevelFormat(body []byte, config thinking.ThinkingConfig)
 
 	level := string(config.Level)
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingLevel", level)
-	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", true)
+
+	// Respect user's explicit includeThoughts setting from original body; default to true if not set
+	// Support both camelCase and snake_case variants
+	includeThoughts := true
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+	}
+	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", includeThoughts)
 	return result, nil
 }
 
 func (a *Applier) applyBudgetFormat(body []byte, config thinking.ThinkingConfig) ([]byte, error) {
-	// Remove conflicting field to avoid both thinkingLevel and thinkingBudget in output
+	// Remove conflicting fields to avoid both thinkingLevel and thinkingBudget in output
 	result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig.thinkingLevel")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_level")
+	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.thinking_budget")
 	// Normalize includeThoughts field name to avoid oneof conflicts in upstream JSON parsing.
 	result, _ = sjson.DeleteBytes(result, "request.generationConfig.thinkingConfig.include_thoughts")
 
 	budget := config.Budget
-	includeThoughts := false
-	switch config.Mode {
-	case thinking.ModeNone:
-		includeThoughts = false
-	case thinking.ModeAuto:
-		includeThoughts = true
-	default:
-		includeThoughts = budget > 0
+
+	// For ModeNone, always set includeThoughts to false regardless of user setting.
+	// This ensures that when user requests budget=0 (disable thinking output),
+	// the includeThoughts is correctly set to false even if budget is clamped to min.
+	if config.Mode == thinking.ModeNone {
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
+		result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.includeThoughts", false)
+		return result, nil
+	}
+
+	// Determine includeThoughts: respect user's explicit setting from original body if provided
+	// Support both camelCase and snake_case variants
+	var includeThoughts bool
+	var userSetIncludeThoughts bool
+	if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.includeThoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	} else if inc := gjson.GetBytes(body, "request.generationConfig.thinkingConfig.include_thoughts"); inc.Exists() {
+		includeThoughts = inc.Bool()
+		userSetIncludeThoughts = true
+	}
+
+	if !userSetIncludeThoughts {
+		// No explicit setting, use default logic based on mode
+		switch config.Mode {
+		case thinking.ModeAuto:
+			includeThoughts = true
+		default:
+			includeThoughts = budget > 0
+		}
 	}
 
 	result, _ = sjson.SetBytes(result, "request.generationConfig.thinkingConfig.thinkingBudget", budget)

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -6,7 +6,6 @@
 package claude
 
 import (
-	"bytes"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
@@ -37,7 +36,7 @@ import (
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
 	enableThoughtTranslate := true
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	// system instruction
 	systemInstructionJSON := ""
@@ -115,7 +114,7 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 							if signatureResult.Exists() && signatureResult.String() != "" {
 								arrayClientSignatures := strings.SplitN(signatureResult.String(), "#", 2)
 								if len(arrayClientSignatures) == 2 {
-									if modelName == arrayClientSignatures[0] {
+									if cache.GetModelGroup(modelName) == arrayClientSignatures[0] {
 										clientSignature = arrayClientSignatures[1]
 									}
 								}

internal/translator/antigravity/gemini/antigravity_gemini_request.go

@@ -6,7 +6,6 @@
 package gemini
 
 import (
-	"bytes"
 	"fmt"
 	"strings"
 
@@ -34,7 +33,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in Gemini API format
 func ConvertGeminiRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	template := ""
 	template = `{"project":"","request":{},"model":""}`
 	template, _ = sjson.SetRaw(template, "request", string(rawJSON))

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

@@ -3,7 +3,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"fmt"
 	"strings"
 
@@ -28,7 +27,7 @@ const geminiCLIFunctionThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base envelope (no default thinkingConfig)
 	out := []byte(`{"project":"","request":{"contents":[]},"model":"gemini-2.5-pro"}`)
 

internal/translator/antigravity/openai/responses/antigravity_openai-responses_request.go

@@ -1,14 +1,12 @@
 package responses
 
 import (
-	"bytes"
-
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/antigravity/gemini"
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/responses"
 )
 
 func ConvertOpenAIResponsesRequestToAntigravity(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	rawJSON = ConvertOpenAIResponsesRequestToGemini(modelName, rawJSON, stream)
 	return ConvertGeminiRequestToAntigravity(modelName, rawJSON, stream)
 }

internal/translator/claude/gemini-cli/claude_gemini-cli_request.go

@@ -6,8 +6,6 @@
 package geminiCLI
 
 import (
-	"bytes"
-
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/claude/gemini"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -30,7 +28,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in Claude Code API format
 func ConvertGeminiCLIRequestToClaude(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	modelResult := gjson.GetBytes(rawJSON, "model")
 	// Extract the inner request object and promote it to the top level

internal/translator/claude/gemini/claude_gemini_request.go

@@ -6,7 +6,6 @@
 package gemini
 
 import (
-	"bytes"
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
@@ -46,7 +45,7 @@ var (
 // Returns:
 //   - []byte: The transformed request data in Claude Code API format
 func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	if account == "" {
 		u, _ := uuid.NewRandom()
@@ -116,7 +115,11 @@ 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() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			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 "":
@@ -132,23 +135,29 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
 					}
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-				budget := int(thinkingBudget.Int())
-				switch budget {
-				case 0:
-					out, _ = sjson.Set(out, "thinking.type", "disabled")
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
-				case -1:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					out, _ = sjson.Delete(out, "thinking.budget_tokens")
-				default:
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+			} else {
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
+				if !thinkingBudget.Exists() {
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
+				}
+				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)
+					}
+				} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
 				}
-			} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-				out, _ = sjson.Set(out, "thinking.type", "enabled")
-			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
-				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}
 	}

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

@@ -6,7 +6,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
@@ -44,7 +43,7 @@ var (
 // Returns:
 //   - []byte: The transformed request data in Claude Code API format
 func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	if account == "" {
 		u, _ := uuid.NewRandom()

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

@@ -1,7 +1,6 @@
 package responses
 
 import (
-	"bytes"
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
@@ -32,7 +31,7 @@ var (
 // - max_output_tokens -> max_tokens
 // - stream passthrough via parameter
 func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	if account == "" {
 		u, _ := uuid.NewRandom()

internal/translator/codex/claude/codex_claude_request.go

@@ -6,7 +6,6 @@
 package claude
 
 import (
-	"bytes"
 	"fmt"
 	"strconv"
 	"strings"
@@ -35,7 +34,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in internal client format
 func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	template := `{"model":"","instructions":"","input":[]}`
 

internal/translator/codex/gemini-cli/codex_gemini-cli_request.go

@@ -6,8 +6,6 @@
 package geminiCLI
 
 import (
-	"bytes"
-
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/codex/gemini"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -30,7 +28,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in Codex API format
 func ConvertGeminiCLIRequestToCodex(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	rawJSON = []byte(gjson.GetBytes(rawJSON, "request").Raw)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "model", modelName)

internal/translator/codex/gemini/codex_gemini_request.go

@@ -6,7 +6,6 @@
 package gemini
 
 import (
-	"bytes"
 	"crypto/rand"
 	"fmt"
 	"math/big"
@@ -37,7 +36,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in Codex API format
 func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base template
 	out := `{"model":"","instructions":"","input":[]}`
 
@@ -243,19 +242,30 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
 
 	// Convert Gemini thinkingConfig to Codex reasoning.effort.
+	// Note: Google official Python SDK sends snake_case fields (thinking_level/thinking_budget).
 	effortSet := false
 	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			thinkingLevel := thinkingConfig.Get("thinkingLevel")
+			if !thinkingLevel.Exists() {
+				thinkingLevel = thinkingConfig.Get("thinking_level")
+			}
+			if thinkingLevel.Exists() {
 				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
 				if effort != "" {
 					out, _ = sjson.Set(out, "reasoning.effort", effort)
 					effortSet = true
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
-					out, _ = sjson.Set(out, "reasoning.effort", effort)
-					effortSet = true
+			} else {
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
+				if !thinkingBudget.Exists() {
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
+				}
+				if thinkingBudget.Exists() {
+					if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+						out, _ = sjson.Set(out, "reasoning.effort", effort)
+						effortSet = true
+					}
 				}
 			}
 		}

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

@@ -7,8 +7,6 @@
 package chat_completions
 
 import (
-	"bytes"
-
 	"strconv"
 	"strings"
 
@@ -29,7 +27,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in OpenAI Responses API format
 func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Start with empty JSON object
 	out := `{"instructions":""}`
 

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

@@ -1,7 +1,6 @@
 package responses
 
 import (
-	"bytes"
 	"fmt"
 
 	"github.com/tidwall/gjson"
@@ -9,7 +8,7 @@ import (
 )
 
 func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	inputResult := gjson.GetBytes(rawJSON, "input")
 	if inputResult.Type == gjson.String {

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

@@ -35,7 +35,7 @@ const geminiCLIClaudeThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	rawJSON = bytes.Replace(rawJSON, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
 
 	// Build output Gemini CLI request JSON
@@ -116,6 +116,19 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 						part, _ = sjson.Set(part, "functionResponse.name", funcName)
 						part, _ = sjson.Set(part, "functionResponse.response.result", responseData)
 						contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+
+					case "image":
+						source := contentResult.Get("source")
+						if source.Get("type").String() == "base64" {
+							mimeType := source.Get("media_type").String()
+							data := source.Get("data").String()
+							if mimeType != "" && data != "" {
+								part := `{"inlineData":{"mime_type":"","data":""}}`
+								part, _ = sjson.Set(part, "inlineData.mime_type", mimeType)
+								part, _ = sjson.Set(part, "inlineData.data", data)
+								contentJSON, _ = sjson.SetRaw(contentJSON, "parts.-1", part)
+							}
+						}
 					}
 					return true
 				})

internal/translator/gemini-cli/gemini/gemini-cli_gemini_request.go

@@ -6,7 +6,6 @@
 package gemini
 
 import (
-	"bytes"
 	"fmt"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
@@ -33,7 +32,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in Gemini API format
 func ConvertGeminiRequestToGeminiCLI(_ string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	template := ""
 	template = `{"project":"","request":{},"model":""}`
 	template, _ = sjson.SetRaw(template, "request", string(rawJSON))

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

@@ -3,7 +3,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"fmt"
 	"strings"
 
@@ -28,7 +27,7 @@ const geminiCLIFunctionThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request data in Gemini CLI API format
 func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base envelope (no default thinkingConfig)
 	out := []byte(`{"project":"","request":{"contents":[]},"model":"gemini-2.5-pro"}`)
 

internal/translator/gemini-cli/openai/responses/gemini-cli_openai-responses_request.go

@@ -1,14 +1,12 @@
 package responses
 
 import (
-	"bytes"
-
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini-cli/gemini"
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/responses"
 )
 
 func ConvertOpenAIResponsesRequestToGeminiCLI(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	rawJSON = ConvertOpenAIResponsesRequestToGemini(modelName, rawJSON, stream)
 	return ConvertGeminiRequestToGeminiCLI(modelName, rawJSON, stream)
 }

internal/translator/gemini/claude/gemini_claude_request.go

@@ -28,7 +28,7 @@ const geminiClaudeThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request in Gemini CLI format.
 func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	rawJSON = bytes.Replace(rawJSON, []byte(`"url":{"type":"string","format":"uri",`), []byte(`"url":{"type":"string",`), -1)
 
 	// Build output Gemini CLI request JSON

internal/translator/gemini/gemini-cli/gemini_gemini-cli_request.go

@@ -6,7 +6,6 @@
 package geminiCLI
 
 import (
-	"bytes"
 	"fmt"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
@@ -19,7 +18,7 @@ import (
 // It extracts the model name, system instruction, message contents, and tool declarations
 // from the raw JSON request and returns them in the format expected by the internal client.
 func ConvertGeminiCLIRequestToGemini(_ string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	modelResult := gjson.GetBytes(rawJSON, "model")
 	rawJSON = []byte(gjson.GetBytes(rawJSON, "request").Raw)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "model", modelResult.String())

internal/translator/gemini/gemini/gemini_gemini_request.go

@@ -4,7 +4,6 @@
 package gemini
 
 import (
-	"bytes"
 	"fmt"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
@@ -19,7 +18,7 @@ import (
 //
 // It keeps the payload otherwise unchanged.
 func ConvertGeminiRequestToGemini(_ string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Fast path: if no contents field, only attach safety settings
 	contents := gjson.GetBytes(rawJSON, "contents")
 	if !contents.Exists() {

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

@@ -3,7 +3,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"fmt"
 	"strings"
 
@@ -28,7 +27,7 @@ const geminiFunctionThoughtSignature = "skip_thought_signature_validator"
 // Returns:
 //   - []byte: The transformed request data in Gemini API format
 func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base envelope (no default thinkingConfig)
 	out := []byte(`{"contents":[]}`)
 

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

@@ -1,7 +1,6 @@
 package responses
 
 import (
-	"bytes"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
@@ -12,7 +11,7 @@ import (
 const geminiResponsesThoughtSignature = "skip_thought_signature_validator"
 
 func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 
 	// Note: modelName and stream parameters are part of the fixed method signature
 	_ = modelName // Unused but required by interface

internal/translator/openai/claude/openai_claude_request.go

@@ -6,7 +6,6 @@
 package claude
 
 import (
-	"bytes"
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
@@ -18,7 +17,7 @@ import (
 // It extracts the model name, system instruction, message contents, and tool declarations
 // from the raw JSON request and returns them in the format expected by the OpenAI API.
 func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base OpenAI Chat Completions API template
 	out := `{"model":"","messages":[]}`
 

internal/translator/openai/gemini-cli/openai_gemini_request.go

@@ -6,8 +6,6 @@
 package geminiCLI
 
 import (
-	"bytes"
-
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/openai/gemini"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -17,7 +15,7 @@ import (
 // It extracts the model name, generation config, message contents, and tool declarations
 // from the raw JSON request and returns them in the format expected by the OpenAI API.
 func ConvertGeminiCLIRequestToOpenAI(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	rawJSON = []byte(gjson.GetBytes(rawJSON, "request").Raw)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "model", modelName)
 	if gjson.GetBytes(rawJSON, "systemInstruction").Exists() {

internal/translator/openai/gemini/openai_gemini_request.go

@@ -6,7 +6,6 @@
 package gemini
 
 import (
-	"bytes"
 	"crypto/rand"
 	"fmt"
 	"math/big"
@@ -21,7 +20,7 @@ import (
 // It extracts the model name, generation config, message contents, and tool declarations
 // from the raw JSON request and returns them in the format expected by the OpenAI API.
 func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base OpenAI Chat Completions API template
 	out := `{"model":"","messages":[]}`
 
@@ -83,16 +82,27 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 		}
 
 		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
-		// Always perform conversion to support allowCompat models that may not be in registry
+		// Always perform conversion to support allowCompat models that may not be in registry.
+		// Note: Google official Python SDK sends snake_case fields (thinking_level/thinking_budget).
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
+			thinkingLevel := thinkingConfig.Get("thinkingLevel")
+			if !thinkingLevel.Exists() {
+				thinkingLevel = thinkingConfig.Get("thinking_level")
+			}
+			if thinkingLevel.Exists() {
 				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
 				if effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
-			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
-					out, _ = sjson.Set(out, "reasoning_effort", effort)
+			} else {
+				thinkingBudget := thinkingConfig.Get("thinkingBudget")
+				if !thinkingBudget.Exists() {
+					thinkingBudget = thinkingConfig.Get("thinking_budget")
+				}
+				if thinkingBudget.Exists() {
+					if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+						out, _ = sjson.Set(out, "reasoning_effort", effort)
+					}
 				}
 			}
 		}

internal/translator/openai/openai/chat-completions/openai_openai_request.go

@@ -3,7 +3,6 @@
 package chat_completions
 
 import (
-	"bytes"
 	"github.com/tidwall/sjson"
 )
 
@@ -25,7 +24,7 @@ func ConvertOpenAIRequestToOpenAI(modelName string, inputRawJSON []byte, _ bool)
 		// If there's an error, return the original JSON or handle the error appropriately.
 		// For now, we'll return the original, but in a real scenario, logging or a more robust error
 		// handling mechanism would be needed.
-		return bytes.Clone(inputRawJSON)
+		return inputRawJSON
 	}
 	return updatedJSON
 }

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

@@ -1,7 +1,6 @@
 package responses
 
 import (
-	"bytes"
 	"strings"
 
 	"github.com/tidwall/gjson"
@@ -28,7 +27,7 @@ import (
 // Returns:
 //   - []byte: The transformed request data in OpenAI chat completions format
 func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inputRawJSON []byte, stream bool) []byte {
-	rawJSON := bytes.Clone(inputRawJSON)
+	rawJSON := inputRawJSON
 	// Base OpenAI chat completions template with default values
 	out := `{"model":"","messages":[],"stream":false}`
 

internal/util/claude_model_test.go

@@ -11,6 +11,7 @@ func TestIsClaudeThinkingModel(t *testing.T) {
 		// Claude thinking models - should return true
 		{"claude-sonnet-4-5-thinking", "claude-sonnet-4-5-thinking", true},
 		{"claude-opus-4-5-thinking", "claude-opus-4-5-thinking", true},
+		{"claude-opus-4-6-thinking", "claude-opus-4-6-thinking", true},
 		{"Claude-Sonnet-Thinking uppercase", "Claude-Sonnet-4-5-Thinking", true},
 		{"claude thinking mixed case", "Claude-THINKING-Model", true},
 

internal/util/gemini_schema.go

@@ -61,14 +61,20 @@ func cleanJSONSchema(jsonStr string, addPlaceholder bool) string {
 
 // removeKeywords removes all occurrences of specified keywords from the JSON schema.
 func removeKeywords(jsonStr string, keywords []string) string {
+	deletePaths := make([]string, 0)
+	pathsByField := findPathsByFields(jsonStr, keywords)
 	for _, key := range keywords {
-		for _, p := range findPaths(jsonStr, key) {
+		for _, p := range pathsByField[key] {
 			if isPropertyDefinition(trimSuffix(p, "."+key)) {
 				continue
 			}
-			jsonStr, _ = sjson.Delete(jsonStr, p)
+			deletePaths = append(deletePaths, p)
 		}
 	}
+	sortByDepth(deletePaths)
+	for _, p := range deletePaths {
+		jsonStr, _ = sjson.Delete(jsonStr, p)
+	}
 	return jsonStr
 }
 
@@ -235,8 +241,9 @@ var unsupportedConstraints = []string{
 }
 
 func moveConstraintsToDescription(jsonStr string) string {
+	pathsByField := findPathsByFields(jsonStr, unsupportedConstraints)
 	for _, key := range unsupportedConstraints {
-		for _, p := range findPaths(jsonStr, key) {
+		for _, p := range pathsByField[key] {
 			val := gjson.Get(jsonStr, p)
 			if !val.Exists() || val.IsObject() || val.IsArray() {
 				continue
@@ -424,14 +431,21 @@ func removeUnsupportedKeywords(jsonStr string) string {
 		"$schema", "$defs", "definitions", "const", "$ref", "additionalProperties",
 		"propertyNames", // Gemini doesn't support property name validation
 	)
+
+	deletePaths := make([]string, 0)
+	pathsByField := findPathsByFields(jsonStr, keywords)
 	for _, key := range keywords {
-		for _, p := range findPaths(jsonStr, key) {
+		for _, p := range pathsByField[key] {
 			if isPropertyDefinition(trimSuffix(p, "."+key)) {
 				continue
 			}
-			jsonStr, _ = sjson.Delete(jsonStr, p)
+			deletePaths = append(deletePaths, p)
 		}
 	}
+	sortByDepth(deletePaths)
+	for _, p := range deletePaths {
+		jsonStr, _ = sjson.Delete(jsonStr, p)
+	}
 	// Remove x-* extension fields (e.g., x-google-enum-descriptions) that are not supported by Gemini API
 	jsonStr = removeExtensionFields(jsonStr)
 	return jsonStr
@@ -581,6 +595,42 @@ func findPaths(jsonStr, field string) []string {
 	return paths
 }
 
+func findPathsByFields(jsonStr string, fields []string) map[string][]string {
+	set := make(map[string]struct{}, len(fields))
+	for _, field := range fields {
+		set[field] = struct{}{}
+	}
+	paths := make(map[string][]string, len(set))
+	walkForFields(gjson.Parse(jsonStr), "", set, paths)
+	return paths
+}
+
+func walkForFields(value gjson.Result, path string, fields map[string]struct{}, paths map[string][]string) {
+	switch value.Type {
+	case gjson.JSON:
+		value.ForEach(func(key, val gjson.Result) bool {
+			keyStr := key.String()
+			safeKey := escapeGJSONPathKey(keyStr)
+
+			var childPath string
+			if path == "" {
+				childPath = safeKey
+			} else {
+				childPath = path + "." + safeKey
+			}
+
+			if _, ok := fields[keyStr]; ok {
+				paths[keyStr] = append(paths[keyStr], childPath)
+			}
+
+			walkForFields(val, childPath, fields, paths)
+			return true
+		})
+	case gjson.String, gjson.Number, gjson.True, gjson.False, gjson.Null:
+		// Terminal types - no further traversal needed
+	}
+}
+
 func sortByDepth(paths []string) {
 	sort.Slice(paths, func(i, j int) bool { return len(paths[i]) > len(paths[j]) })
 }
@@ -667,6 +717,9 @@ func orDefault(val, def string) string {
 }
 
 func escapeGJSONPathKey(key string) string {
+	if strings.IndexAny(key, ".*?") == -1 {
+		return key
+	}
 	return gjsonPathKeyReplacer.Replace(key)
 }
 

internal/util/translator.go

@@ -6,7 +6,6 @@ package util
 import (
 	"bytes"
 	"fmt"
-	"strings"
 
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -33,15 +32,15 @@ func Walk(value gjson.Result, path, field string, paths *[]string) {
 			// . -> \.
 			// * -> \*
 			// ? -> \?
-			var keyReplacer = strings.NewReplacer(".", "\\.", "*", "\\*", "?", "\\?")
-			safeKey := keyReplacer.Replace(key.String())
+			keyStr := key.String()
+			safeKey := escapeGJSONPathKey(keyStr)
 
 			if path == "" {
 				childPath = safeKey
 			} else {
 				childPath = path + "." + safeKey
 			}
-			if key.String() == field {
+			if keyStr == field {
 				*paths = append(*paths, childPath)
 			}
 			Walk(val, childPath, field, paths)
@@ -87,15 +86,6 @@ func RenameKey(jsonStr, oldKeyPath, newKeyPath string) (string, error) {
 	return finalJson, nil
 }
 
-func DeleteKey(jsonStr, keyName string) string {
-	paths := make([]string, 0)
-	Walk(gjson.Parse(jsonStr), "", keyName, &paths)
-	for _, p := range paths {
-		jsonStr, _ = sjson.Delete(jsonStr, p)
-	}
-	return jsonStr
-}
-
 // FixJSON converts non-standard JSON that uses single quotes for strings into
 // RFC 8259-compliant JSON by converting those single-quoted strings to
 // double-quoted strings with proper escaping.

internal/watcher/clients.go

@@ -6,6 +6,7 @@ import (
 	"context"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"fmt"
 	"io/fs"
 	"os"
@@ -15,6 +16,7 @@ import (
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 	log "github.com/sirupsen/logrus"
 )
@@ -72,6 +74,7 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 		w.clientsMutex.Lock()
 
 		w.lastAuthHashes = make(map[string]string)
+		w.lastAuthContents = make(map[string]*coreauth.Auth)
 		if resolvedAuthDir, errResolveAuthDir := util.ResolveAuthDir(cfg.AuthDir); errResolveAuthDir != nil {
 			log.Errorf("failed to resolve auth directory for hash cache: %v", errResolveAuthDir)
 		} else if resolvedAuthDir != "" {
@@ -84,6 +87,11 @@ func (w *Watcher) reloadClients(rescanAuth bool, affectedOAuthProviders []string
 						sum := sha256.Sum256(data)
 						normalizedPath := w.normalizeAuthPath(path)
 						w.lastAuthHashes[normalizedPath] = hex.EncodeToString(sum[:])
+						// Parse and cache auth content for future diff comparisons
+						var auth coreauth.Auth
+						if errParse := json.Unmarshal(data, &auth); errParse == nil {
+							w.lastAuthContents[normalizedPath] = &auth
+						}
 					}
 				}
 				return nil
@@ -127,6 +135,13 @@ func (w *Watcher) addOrUpdateClient(path string) {
 	curHash := hex.EncodeToString(sum[:])
 	normalized := w.normalizeAuthPath(path)
 
+	// Parse new auth content for diff comparison
+	var newAuth coreauth.Auth
+	if errParse := json.Unmarshal(data, &newAuth); errParse != nil {
+		log.Errorf("failed to parse auth file %s: %v", filepath.Base(path), errParse)
+		return
+	}
+
 	w.clientsMutex.Lock()
 
 	cfg := w.config
@@ -141,7 +156,26 @@ func (w *Watcher) addOrUpdateClient(path string) {
 		return
 	}
 
+	// Get old auth for diff comparison
+	var oldAuth *coreauth.Auth
+	if w.lastAuthContents != nil {
+		oldAuth = w.lastAuthContents[normalized]
+	}
+
+	// Compute and log field changes
+	if changes := diff.BuildAuthChangeDetails(oldAuth, &newAuth); len(changes) > 0 {
+		log.Debugf("auth field changes for %s:", filepath.Base(path))
+		for _, c := range changes {
+			log.Debugf("  %s", c)
+		}
+	}
+
+	// Update caches
 	w.lastAuthHashes[normalized] = curHash
+	if w.lastAuthContents == nil {
+		w.lastAuthContents = make(map[string]*coreauth.Auth)
+	}
+	w.lastAuthContents[normalized] = &newAuth
 
 	w.clientsMutex.Unlock() // Unlock before the callback
 
@@ -160,6 +194,7 @@ func (w *Watcher) removeClient(path string) {
 
 	cfg := w.config
 	delete(w.lastAuthHashes, normalized)
+	delete(w.lastAuthContents, normalized)
 
 	w.clientsMutex.Unlock() // Release the lock before the callback
 

internal/watcher/diff/auth_diff.go

@@ -0,0 +1,44 @@
+// auth_diff.go computes human-readable diffs for auth file field changes.
+package diff
+
+import (
+	"fmt"
+	"strings"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+// BuildAuthChangeDetails computes a redacted, human-readable list of auth field changes.
+// Only prefix, proxy_url, and disabled fields are tracked; sensitive data is never printed.
+func BuildAuthChangeDetails(oldAuth, newAuth *coreauth.Auth) []string {
+	changes := make([]string, 0, 3)
+
+	// Handle nil cases by using empty Auth as default
+	if oldAuth == nil {
+		oldAuth = &coreauth.Auth{}
+	}
+	if newAuth == nil {
+		return changes
+	}
+
+	// Compare prefix
+	oldPrefix := strings.TrimSpace(oldAuth.Prefix)
+	newPrefix := strings.TrimSpace(newAuth.Prefix)
+	if oldPrefix != newPrefix {
+		changes = append(changes, fmt.Sprintf("prefix: %s -> %s", oldPrefix, newPrefix))
+	}
+
+	// Compare proxy_url (redacted)
+	oldProxy := strings.TrimSpace(oldAuth.ProxyURL)
+	newProxy := strings.TrimSpace(newAuth.ProxyURL)
+	if oldProxy != newProxy {
+		changes = append(changes, fmt.Sprintf("proxy_url: %s -> %s", formatProxyURL(oldProxy), formatProxyURL(newProxy)))
+	}
+
+	// Compare disabled
+	if oldAuth.Disabled != newAuth.Disabled {
+		changes = append(changes, fmt.Sprintf("disabled: %t -> %t", oldAuth.Disabled, newAuth.Disabled))
+	}
+
+	return changes
+}

internal/watcher/diff/config_diff.go

@@ -27,6 +27,12 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 	if oldCfg.Debug != newCfg.Debug {
 		changes = append(changes, fmt.Sprintf("debug: %t -> %t", oldCfg.Debug, newCfg.Debug))
 	}
+	if oldCfg.Pprof.Enable != newCfg.Pprof.Enable {
+		changes = append(changes, fmt.Sprintf("pprof.enable: %t -> %t", oldCfg.Pprof.Enable, newCfg.Pprof.Enable))
+	}
+	if strings.TrimSpace(oldCfg.Pprof.Addr) != strings.TrimSpace(newCfg.Pprof.Addr) {
+		changes = append(changes, fmt.Sprintf("pprof.addr: %s -> %s", strings.TrimSpace(oldCfg.Pprof.Addr), strings.TrimSpace(newCfg.Pprof.Addr)))
+	}
 	if oldCfg.LoggingToFile != newCfg.LoggingToFile {
 		changes = append(changes, fmt.Sprintf("logging-to-file: %t -> %t", oldCfg.LoggingToFile, newCfg.LoggingToFile))
 	}

internal/watcher/watcher.go

@@ -38,6 +38,7 @@ type Watcher struct {
 	reloadCallback    func(*config.Config)
 	watcher           *fsnotify.Watcher
 	lastAuthHashes    map[string]string
+	lastAuthContents  map[string]*coreauth.Auth
 	lastRemoveTimes   map[string]time.Time
 	lastConfigHash    string
 	authQueue         chan<- AuthUpdate

sdk/api/handlers/handlers.go

@@ -155,20 +155,6 @@ func requestExecutionMetadata(ctx context.Context) map[string]any {
 	return map[string]any{idempotencyKeyMetadataKey: key}
 }
 
-func mergeMetadata(base, overlay map[string]any) map[string]any {
-	if len(base) == 0 && len(overlay) == 0 {
-		return nil
-	}
-	out := make(map[string]any, len(base)+len(overlay))
-	for k, v := range base {
-		out[k] = v
-	}
-	for k, v := range overlay {
-		out[k] = v
-	}
-	return out
-}
-
 // BaseAPIHandler contains the handlers for API endpoints.
 // It holds a pool of clients to interact with the backend service and manages
 // load balancing, client selection, and configuration.
@@ -391,14 +377,18 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
+	payload := rawJSON
+	if len(payload) == 0 {
+		payload = nil
+	}
 	req := coreexecutor.Request{
 		Model:   normalizedModel,
-		Payload: cloneBytes(rawJSON),
+		Payload: payload,
 	}
 	opts := coreexecutor.Options{
 		Stream:          false,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -418,7 +408,7 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 		}
 		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return cloneBytes(resp.Payload), nil
+	return resp.Payload, nil
 }
 
 // ExecuteCountWithAuthManager executes a non-streaming request via the core auth manager.
@@ -430,14 +420,18 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
+	payload := rawJSON
+	if len(payload) == 0 {
+		payload = nil
+	}
 	req := coreexecutor.Request{
 		Model:   normalizedModel,
-		Payload: cloneBytes(rawJSON),
+		Payload: payload,
 	}
 	opts := coreexecutor.Options{
 		Stream:          false,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -457,7 +451,7 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 		}
 		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return cloneBytes(resp.Payload), nil
+	return resp.Payload, nil
 }
 
 // ExecuteStreamWithAuthManager executes a streaming request via the core auth manager.
@@ -472,14 +466,18 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
+	payload := rawJSON
+	if len(payload) == 0 {
+		payload = nil
+	}
 	req := coreexecutor.Request{
 		Model:   normalizedModel,
-		Payload: cloneBytes(rawJSON),
+		Payload: payload,
 	}
 	opts := coreexecutor.Options{
 		Stream:          true,
 		Alt:             alt,
-		OriginalRequest: cloneBytes(rawJSON),
+		OriginalRequest: rawJSON,
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
@@ -668,17 +666,6 @@ func cloneBytes(src []byte) []byte {
 	return dst
 }
 
-func cloneMetadata(src map[string]any) map[string]any {
-	if len(src) == 0 {
-		return nil
-	}
-	dst := make(map[string]any, len(src))
-	for k, v := range src {
-		dst[k] = v
-	}
-	return dst
-}
-
 // WriteErrorResponse writes an error message to the response writer using the HTTP status embedded in the message.
 func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.ErrorMessage) {
 	status := http.StatusInternalServerError
@@ -709,7 +696,7 @@ func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.Erro
 	var previous []byte
 	if existing, exists := c.Get("API_RESPONSE"); exists {
 		if existingBytes, ok := existing.([]byte); ok && len(existingBytes) > 0 {
-			previous = bytes.Clone(existingBytes)
+			previous = existingBytes
 		}
 	}
 	appendAPIResponse(c, body)

sdk/cliproxy/pprof_server.go

@@ -0,0 +1,163 @@
+package cliproxy
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/pprof"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	log "github.com/sirupsen/logrus"
+)
+
+type pprofServer struct {
+	mu      sync.Mutex
+	server  *http.Server
+	addr    string
+	enabled bool
+}
+
+func newPprofServer() *pprofServer {
+	return &pprofServer{}
+}
+
+func (s *Service) applyPprofConfig(cfg *config.Config) {
+	if s == nil || cfg == nil {
+		return
+	}
+	if s.pprofServer == nil {
+		s.pprofServer = newPprofServer()
+	}
+	s.pprofServer.Apply(cfg)
+}
+
+func (s *Service) shutdownPprof(ctx context.Context) error {
+	if s == nil || s.pprofServer == nil {
+		return nil
+	}
+	return s.pprofServer.Shutdown(ctx)
+}
+
+func (p *pprofServer) Apply(cfg *config.Config) {
+	if p == nil || cfg == nil {
+		return
+	}
+	addr := strings.TrimSpace(cfg.Pprof.Addr)
+	if addr == "" {
+		addr = config.DefaultPprofAddr
+	}
+	enabled := cfg.Pprof.Enable
+
+	p.mu.Lock()
+	currentServer := p.server
+	currentAddr := p.addr
+	p.addr = addr
+	p.enabled = enabled
+	if !enabled {
+		p.server = nil
+		p.mu.Unlock()
+		if currentServer != nil {
+			p.stopServer(currentServer, currentAddr, "disabled")
+		}
+		return
+	}
+	if currentServer != nil && currentAddr == addr {
+		p.mu.Unlock()
+		return
+	}
+	p.server = nil
+	p.mu.Unlock()
+
+	if currentServer != nil {
+		p.stopServer(currentServer, currentAddr, "restarted")
+	}
+
+	p.startServer(addr)
+}
+
+func (p *pprofServer) Shutdown(ctx context.Context) error {
+	if p == nil {
+		return nil
+	}
+	p.mu.Lock()
+	currentServer := p.server
+	currentAddr := p.addr
+	p.server = nil
+	p.enabled = false
+	p.mu.Unlock()
+
+	if currentServer == nil {
+		return nil
+	}
+	return p.stopServerWithContext(ctx, currentServer, currentAddr, "shutdown")
+}
+
+func (p *pprofServer) startServer(addr string) {
+	mux := newPprofMux()
+	server := &http.Server{
+		Addr:              addr,
+		Handler:           mux,
+		ReadHeaderTimeout: 5 * time.Second,
+	}
+
+	p.mu.Lock()
+	if !p.enabled || p.addr != addr || p.server != nil {
+		p.mu.Unlock()
+		return
+	}
+	p.server = server
+	p.mu.Unlock()
+
+	log.Infof("pprof server starting on %s", addr)
+	go func() {
+		if errServe := server.ListenAndServe(); errServe != nil && !errors.Is(errServe, http.ErrServerClosed) {
+			log.Errorf("pprof server failed on %s: %v", addr, errServe)
+			p.mu.Lock()
+			if p.server == server {
+				p.server = nil
+			}
+			p.mu.Unlock()
+		}
+	}()
+}
+
+func (p *pprofServer) stopServer(server *http.Server, addr string, reason string) {
+	_ = p.stopServerWithContext(context.Background(), server, addr, reason)
+}
+
+func (p *pprofServer) stopServerWithContext(ctx context.Context, server *http.Server, addr string, reason string) error {
+	if server == nil {
+		return nil
+	}
+	stopCtx := ctx
+	if stopCtx == nil {
+		stopCtx = context.Background()
+	}
+	stopCtx, cancel := context.WithTimeout(stopCtx, 5*time.Second)
+	defer cancel()
+	if errStop := server.Shutdown(stopCtx); errStop != nil {
+		log.Errorf("pprof server stop failed on %s: %v", addr, errStop)
+		return errStop
+	}
+	log.Infof("pprof server stopped on %s (%s)", addr, reason)
+	return nil
+}
+
+func newPprofMux() *http.ServeMux {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/debug/pprof/", pprof.Index)
+	mux.HandleFunc("/debug/pprof/cmdline", pprof.Cmdline)
+	mux.HandleFunc("/debug/pprof/profile", pprof.Profile)
+	mux.HandleFunc("/debug/pprof/symbol", pprof.Symbol)
+	mux.HandleFunc("/debug/pprof/trace", pprof.Trace)
+	mux.Handle("/debug/pprof/allocs", pprof.Handler("allocs"))
+	mux.Handle("/debug/pprof/block", pprof.Handler("block"))
+	mux.Handle("/debug/pprof/goroutine", pprof.Handler("goroutine"))
+	mux.Handle("/debug/pprof/heap", pprof.Handler("heap"))
+	mux.Handle("/debug/pprof/mutex", pprof.Handler("mutex"))
+	mux.Handle("/debug/pprof/threadcreate", pprof.Handler("threadcreate"))
+	return mux
+}

sdk/cliproxy/service.go

@@ -57,6 +57,9 @@ type Service struct {
 	// server is the HTTP API server instance.
 	server *api.Server
 
+	// pprofServer manages the optional pprof HTTP debug server.
+	pprofServer *pprofServer
+
 	// serverErr channel for server startup/shutdown errors.
 	serverErr chan error
 
@@ -270,27 +273,42 @@ func (s *Service) wsOnDisconnected(channelID string, reason error) {
 }
 
 func (s *Service) applyCoreAuthAddOrUpdate(ctx context.Context, auth *coreauth.Auth) {
-	if s == nil || auth == nil || auth.ID == "" {
-		return
-	}
-	if s.coreManager == nil {
+	if s == nil || s.coreManager == nil || auth == nil || auth.ID == "" {
 		return
 	}
 	auth = auth.Clone()
 	s.ensureExecutorsForAuth(auth)
-	s.registerModelsForAuth(auth)
-	if existing, ok := s.coreManager.GetByID(auth.ID); ok && existing != nil {
+
+	// IMPORTANT: Update coreManager FIRST, before model registration.
+	// This ensures that configuration changes (proxy_url, prefix, etc.) take effect
+	// immediately for API calls, rather than waiting for model registration to complete.
+	// Model registration may involve network calls (e.g., FetchAntigravityModels) that
+	// could timeout if the new proxy_url is unreachable.
+	op := "register"
+	var err error
+	if existing, ok := s.coreManager.GetByID(auth.ID); ok {
 		auth.CreatedAt = existing.CreatedAt
 		auth.LastRefreshedAt = existing.LastRefreshedAt
 		auth.NextRefreshAfter = existing.NextRefreshAfter
-		if _, err := s.coreManager.Update(ctx, auth); err != nil {
-			log.Errorf("failed to update auth %s: %v", auth.ID, err)
+		op = "update"
+		_, err = s.coreManager.Update(ctx, auth)
+	} else {
+		_, err = s.coreManager.Register(ctx, auth)
+	}
+	if err != nil {
+		log.Errorf("failed to %s auth %s: %v", op, auth.ID, err)
+		current, ok := s.coreManager.GetByID(auth.ID)
+		if !ok || current.Disabled {
+			GlobalModelRegistry().UnregisterClient(auth.ID)
+			return
 		}
-		return
-	}
-	if _, err := s.coreManager.Register(ctx, auth); err != nil {
-		log.Errorf("failed to register auth %s: %v", auth.ID, err)
+		auth = current
 	}
+
+	// Register models after auth is updated in coreManager.
+	// This operation may block on network calls, but the auth configuration
+	// is already effective at this point.
+	s.registerModelsForAuth(auth)
 }
 
 func (s *Service) applyCoreAuthRemoval(ctx context.Context, id string) {
@@ -501,6 +519,8 @@ func (s *Service) Run(ctx context.Context) error {
 	time.Sleep(100 * time.Millisecond)
 	fmt.Printf("API server started successfully on: %s:%d\n", s.cfg.Host, s.cfg.Port)
 
+	s.applyPprofConfig(s.cfg)
+
 	if s.hooks.OnAfterStart != nil {
 		s.hooks.OnAfterStart(s)
 	}
@@ -546,6 +566,7 @@ func (s *Service) Run(ctx context.Context) error {
 		}
 
 		s.applyRetryConfig(newCfg)
+		s.applyPprofConfig(newCfg)
 		if s.server != nil {
 			s.server.UpdateClients(newCfg)
 		}
@@ -639,6 +660,13 @@ func (s *Service) Shutdown(ctx context.Context) error {
 			s.authQueueStop = nil
 		}
 
+		if errShutdownPprof := s.shutdownPprof(ctx); errShutdownPprof != nil {
+			log.Errorf("failed to stop pprof server: %v", errShutdownPprof)
+			if shutdownErr == nil {
+				shutdownErr = errShutdownPprof
+			}
+		}
+
 		// no legacy clients to persist
 
 		if s.server != nil {

test/builtin_tools_translation_test.go

@@ -33,7 +33,7 @@ func TestOpenAIToCodex_PreservesBuiltinTools(t *testing.T) {
 	}
 }
 
-func TestOpenAIResponsesToOpenAI_PreservesBuiltinTools(t *testing.T) {
+func TestOpenAIResponsesToOpenAI_IgnoresBuiltinTools(t *testing.T) {
 	in := []byte(`{
 		"model":"gpt-5",
 		"input":[{"role":"user","content":[{"type":"input_text","text":"hi"}]}],
@@ -42,13 +42,7 @@ func TestOpenAIResponsesToOpenAI_PreservesBuiltinTools(t *testing.T) {
 
 	out := sdktranslator.TranslateRequest(sdktranslator.FormatOpenAIResponse, sdktranslator.FormatOpenAI, "gpt-5", in, false)
 
-	if got := gjson.GetBytes(out, "tools.#").Int(); got != 1 {
-		t.Fatalf("expected 1 tool, got %d: %s", got, string(out))
-	}
-	if got := gjson.GetBytes(out, "tools.0.type").String(); got != "web_search" {
-		t.Fatalf("expected tools[0].type=web_search, got %q: %s", got, string(out))
-	}
-	if got := gjson.GetBytes(out, "tools.0.search_context_size").String(); got != "low" {
-		t.Fatalf("expected tools[0].search_context_size=low, got %q: %s", got, string(out))
+	if got := gjson.GetBytes(out, "tools.#").Int(); got != 0 {
+		t.Fatalf("expected 0 tools (builtin tools not supported in Chat Completions), got %d: %s", got, string(out))
 	}
 }