Merge pull request #1664 from ciberponk/pr/responses-compaction-compat

feat: add codex responses compatibility for compaction payloads

README.md

@@ -161,6 +161,12 @@ Those projects are ports of CLIProxyAPI or inspired by it:
 
 A Next.js implementation inspired by CLIProxyAPI, easy to install and use, built from scratch with format translation (OpenAI/Claude/Gemini/Ollama), combo system with auto-fallback, multi-account management with exponential backoff, a Next.js web dashboard, and support for CLI tools (Cursor, Claude Code, Cline, RooCode) - no API keys needed.
 
+### [OmniRoute](https://github.com/diegosouzapw/OmniRoute)
+
+Never stop coding. Smart routing to FREE & low-cost AI models with automatic fallback.
+
+OmniRoute is an AI gateway for multi-provider LLMs: an OpenAI-compatible endpoint with smart routing, load balancing, retries, and fallbacks. Add policies, rate limits, caching, and observability for reliable, cost-aware inference.
+
 > [!NOTE]  
 > If you have developed a port of CLIProxyAPI or a project inspired by it, please open a PR to add it to this list.
 

README_CN.md

@@ -160,6 +160,12 @@ Windows 托盘应用，基于 PowerShell 脚本实现，不依赖任何第三方
 
 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。
 
+### [OmniRoute](https://github.com/diegosouzapw/OmniRoute)
+
+代码不止，创新不停。智能路由至免费及低成本 AI 模型，并支持自动故障转移。
+
+OmniRoute 是一个面向多供应商大语言模型的 AI 网关：它提供兼容 OpenAI 的端点，具备智能路由、负载均衡、重试及回退机制。通过添加策略、速率限制、缓存和可观测性，确保推理过程既可靠又具备成本意识。
+
 > [!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。
 

cmd/server/main.go

@@ -58,6 +58,7 @@ func main() {
 	// Command-line flags to control the application's behavior.
 	var login bool
 	var codexLogin bool
+	var codexDeviceLogin bool
 	var claudeLogin bool
 	var qwenLogin bool
 	var iflowLogin bool
@@ -76,6 +77,7 @@ func main() {
 	// Define command-line flags for different operation modes.
 	flag.BoolVar(&login, "login", false, "Login Google Account")
 	flag.BoolVar(&codexLogin, "codex-login", false, "Login to Codex using OAuth")
+	flag.BoolVar(&codexDeviceLogin, "codex-device-login", false, "Login to Codex using device code flow")
 	flag.BoolVar(&claudeLogin, "claude-login", false, "Login to Claude using OAuth")
 	flag.BoolVar(&qwenLogin, "qwen-login", false, "Login to Qwen using OAuth")
 	flag.BoolVar(&iflowLogin, "iflow-login", false, "Login to iFlow using OAuth")
@@ -467,6 +469,9 @@ func main() {
 	} else if codexLogin {
 		// Handle Codex login
 		cmd.DoCodexLogin(cfg, options)
+	} else if codexDeviceLogin {
+		// Handle Codex device-code login
+		cmd.DoCodexDeviceLogin(cfg, options)
 	} else if claudeLogin {
 		// Handle Claude login
 		cmd.DoClaudeLogin(cfg, options)

config.example.yaml

@@ -68,6 +68,10 @@ proxy-url: ""
 # When true, unprefixed model requests only use credentials without a prefix (except when prefix == model name).
 force-model-prefix: false
 
+# When true, forward filtered upstream response headers to downstream clients.
+# Default is false (disabled).
+passthrough-headers: false
+
 # Number of times to retry a request. Retries will occur if the HTTP response code is 403, 408, 500, 502, 503, or 504.
 request-retry: 3
 
@@ -155,6 +159,7 @@ nonstream-keepalive-interval: 0
 #       sensitive-words:             # optional: words to obfuscate with zero-width characters
 #         - "API"
 #         - "proxy"
+#       cache-user-id: true          # optional: default is false; set true to reuse cached user_id per API key instead of generating a random one each request
 
 # Default headers for Claude API requests. Update when Claude Code releases new versions.
 # These are used as fallbacks when the client does not send its own headers.

examples/custom-provider/main.go

@@ -159,13 +159,13 @@ func (MyExecutor) CountTokens(context.Context, *coreauth.Auth, clipexec.Request,
 	return clipexec.Response{}, errors.New("count tokens not implemented")
 }
 
-func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (MyExecutor) ExecuteStream(ctx context.Context, a *coreauth.Auth, req clipexec.Request, opts clipexec.Options) (*clipexec.StreamResult, error) {
 	ch := make(chan clipexec.StreamChunk, 1)
 	go func() {
 		defer close(ch)
 		ch <- clipexec.StreamChunk{Payload: []byte("data: {\"ok\":true}\n\n")}
 	}()
-	return ch, nil
+	return &clipexec.StreamResult{Chunks: ch}, nil
 }
 
 func (MyExecutor) Refresh(ctx context.Context, a *coreauth.Auth) (*coreauth.Auth, error) {

examples/http-request/main.go

@@ -58,7 +58,7 @@ func (EchoExecutor) Execute(context.Context, *coreauth.Auth, clipexec.Request, c
 	return clipexec.Response{}, errors.New("echo executor: Execute not implemented")
 }
 
-func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (<-chan clipexec.StreamChunk, error) {
+func (EchoExecutor) ExecuteStream(context.Context, *coreauth.Auth, clipexec.Request, clipexec.Options) (*clipexec.StreamResult, error) {
 	return nil, errors.New("echo executor: ExecuteStream not implemented")
 }
 

internal/api/handlers/management/auth_files.go

@@ -945,11 +945,17 @@ func (h *Handler) saveTokenRecord(ctx context.Context, record *coreauth.Auth) (s
 	if store == nil {
 		return "", fmt.Errorf("token store unavailable")
 	}
+	if h.postAuthHook != nil {
+		if err := h.postAuthHook(ctx, record); err != nil {
+			return "", fmt.Errorf("post-auth hook failed: %w", err)
+		}
+	}
 	return store.Save(ctx, record)
 }
 
 func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing Claude authentication...")
 
@@ -1094,6 +1100,7 @@ func (h *Handler) RequestAnthropicToken(c *gin.Context) {
 
 func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 	proxyHTTPClient := util.SetProxy(&h.cfg.SDKConfig, &http.Client{})
 	ctx = context.WithValue(ctx, oauth2.HTTPClient, proxyHTTPClient)
 
@@ -1352,6 +1359,7 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 
 func (h *Handler) RequestCodexToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing Codex authentication...")
 
@@ -1497,6 +1505,7 @@ func (h *Handler) RequestCodexToken(c *gin.Context) {
 
 func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing Antigravity authentication...")
 
@@ -1661,6 +1670,7 @@ func (h *Handler) RequestAntigravityToken(c *gin.Context) {
 
 func (h *Handler) RequestQwenToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing Qwen authentication...")
 
@@ -1716,6 +1726,7 @@ func (h *Handler) RequestQwenToken(c *gin.Context) {
 
 func (h *Handler) RequestKimiToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing Kimi authentication...")
 
@@ -1792,6 +1803,7 @@ func (h *Handler) RequestKimiToken(c *gin.Context) {
 
 func (h *Handler) RequestIFlowToken(c *gin.Context) {
 	ctx := context.Background()
+	ctx = PopulateAuthContext(ctx, c)
 
 	fmt.Println("Initializing iFlow authentication...")
 
@@ -2412,3 +2424,12 @@ func (h *Handler) GetAuthStatus(c *gin.Context) {
 	}
 	c.JSON(http.StatusOK, gin.H{"status": "wait"})
 }
+
+// PopulateAuthContext extracts request info and adds it to the context
+func PopulateAuthContext(ctx context.Context, c *gin.Context) context.Context {
+	info := &coreauth.RequestInfo{
+		Query:   c.Request.URL.Query(),
+		Headers: c.Request.Header,
+	}
+	return coreauth.WithRequestInfo(ctx, info)
+}

internal/api/handlers/management/handler.go

@@ -47,6 +47,7 @@ type Handler struct {
 	allowRemoteOverride bool
 	envSecret           string
 	logDir              string
+	postAuthHook        coreauth.PostAuthHook
 }
 
 // NewHandler creates a new management handler instance.
@@ -128,6 +129,11 @@ func (h *Handler) SetLogDirectory(dir string) {
 	h.logDir = dir
 }
 
+// SetPostAuthHook registers a hook to be called after auth record creation but before persistence.
+func (h *Handler) SetPostAuthHook(hook coreauth.PostAuthHook) {
+	h.postAuthHook = hook
+}
+
 // Middleware enforces access control for management endpoints.
 // All requests (local and remote) require a valid management key.
 // Additionally, remote access requires allow-remote-management=true.

internal/api/modules/amp/proxy.go

@@ -3,6 +3,8 @@ package amp
 import (
 	"bytes"
 	"compress/gzip"
+	"context"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -188,6 +190,10 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 
 	// Error handler for proxy failures
 	proxy.ErrorHandler = func(rw http.ResponseWriter, req *http.Request, err error) {
+		// Client-side cancellations are common during polling; suppress logging in this case
+		if errors.Is(err, context.Canceled) {
+			return
+		}
 		log.Errorf("amp upstream proxy error for %s %s: %v", req.Method, req.URL.Path, err)
 		rw.Header().Set("Content-Type", "application/json")
 		rw.WriteHeader(http.StatusBadGateway)

internal/api/modules/amp/proxy_test.go

@@ -493,6 +493,30 @@ func TestReverseProxy_ErrorHandler(t *testing.T) {
 	}
 }
 
+func TestReverseProxy_ErrorHandler_ContextCanceled(t *testing.T) {
+	// Test that context.Canceled errors return 499 without generic error response
+	proxy, err := createReverseProxy("http://example.com", NewStaticSecretSource(""))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Create a canceled context to trigger the cancellation path
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // Cancel immediately
+
+	req := httptest.NewRequest(http.MethodGet, "/test", nil).WithContext(ctx)
+	rr := httptest.NewRecorder()
+
+	// Directly invoke the ErrorHandler with context.Canceled
+	proxy.ErrorHandler(rr, req, context.Canceled)
+
+	// Body should be empty for canceled requests (no JSON error response)
+	body := rr.Body.Bytes()
+	if len(body) > 0 {
+		t.Fatalf("expected empty body for canceled context, got: %s", body)
+	}
+}
+
 func TestReverseProxy_FullRoundTrip_Gzip(t *testing.T) {
 	// Upstream returns gzipped JSON without Content-Encoding header
 	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

internal/api/server.go

@@ -51,6 +51,7 @@ type serverOptionConfig struct {
 	keepAliveEnabled     bool
 	keepAliveTimeout     time.Duration
 	keepAliveOnTimeout   func()
+	postAuthHook         auth.PostAuthHook
 }
 
 // ServerOption customises HTTP server construction.
@@ -111,6 +112,13 @@ func WithRequestLoggerFactory(factory func(*config.Config, string) logging.Reque
 	}
 }
 
+// WithPostAuthHook registers a hook to be called after auth record creation.
+func WithPostAuthHook(hook auth.PostAuthHook) ServerOption {
+	return func(cfg *serverOptionConfig) {
+		cfg.postAuthHook = hook
+	}
+}
+
 // Server represents the main API server.
 // It encapsulates the Gin engine, HTTP server, handlers, and configuration.
 type Server struct {
@@ -262,6 +270,9 @@ func NewServer(cfg *config.Config, authManager *auth.Manager, accessManager *sdk
 	}
 	logDir := logging.ResolveLogDirectory(cfg)
 	s.mgmt.SetLogDirectory(logDir)
+	if optionState.postAuthHook != nil {
+		s.mgmt.SetPostAuthHook(optionState.postAuthHook)
+	}
 	s.localPassword = optionState.localPassword
 
 	// Setup routes

internal/auth/claude/anthropic_auth.go

@@ -20,7 +20,7 @@ import (
 // OAuth configuration constants for Claude/Anthropic
 const (
 	AuthURL     = "https://claude.ai/oauth/authorize"
-	TokenURL    = "https://console.anthropic.com/v1/oauth/token"
+	TokenURL    = "https://api.anthropic.com/v1/oauth/token"
 	ClientID    = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
 	RedirectURI = "http://localhost:54545/callback"
 )

internal/auth/claude/token.go

@@ -36,11 +36,21 @@ type ClaudeTokenStorage struct {
 
 	// Expire is the timestamp when the current access token expires.
 	Expire string `json:"expired"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *ClaudeTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // SaveTokenToFile serializes the Claude token storage to a JSON file.
 // This method creates the necessary directory structure and writes the token
 // data in JSON format to the specified file path for persistent storage.
+// It merges any injected metadata into the top-level JSON object.
 //
 // Parameters:
 //   - authFilePath: The full path where the token file should be saved
@@ -65,8 +75,14 @@ func (ts *ClaudeTokenStorage) SaveTokenToFile(authFilePath string) error {
 		_ = f.Close()
 	}()
 
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
+
 	// Encode and write the token data as JSON
-	if err = json.NewEncoder(f).Encode(ts); err != nil {
+	if err = json.NewEncoder(f).Encode(data); err != nil {
 		return fmt.Errorf("failed to write token to file: %w", err)
 	}
 	return nil

internal/auth/codex/openai_auth.go

@@ -71,16 +71,26 @@ func (o *CodexAuth) GenerateAuthURL(state string, pkceCodes *PKCECodes) (string,
 // It performs an HTTP POST request to the OpenAI token endpoint with the provided
 // authorization code and PKCE verifier.
 func (o *CodexAuth) ExchangeCodeForTokens(ctx context.Context, code string, pkceCodes *PKCECodes) (*CodexAuthBundle, error) {
+	return o.ExchangeCodeForTokensWithRedirect(ctx, code, RedirectURI, pkceCodes)
+}
+
+// ExchangeCodeForTokensWithRedirect exchanges an authorization code for tokens using
+// a caller-provided redirect URI. This supports alternate auth flows such as device
+// login while preserving the existing token parsing and storage behavior.
+func (o *CodexAuth) ExchangeCodeForTokensWithRedirect(ctx context.Context, code, redirectURI string, pkceCodes *PKCECodes) (*CodexAuthBundle, error) {
 	if pkceCodes == nil {
 		return nil, fmt.Errorf("PKCE codes are required for token exchange")
 	}
+	if strings.TrimSpace(redirectURI) == "" {
+		return nil, fmt.Errorf("redirect URI is required for token exchange")
+	}
 
 	// Prepare token exchange request
 	data := url.Values{
 		"grant_type":    {"authorization_code"},
 		"client_id":     {ClientID},
 		"code":          {code},
-		"redirect_uri":  {RedirectURI},
+		"redirect_uri":  {strings.TrimSpace(redirectURI)},
 		"code_verifier": {pkceCodes.CodeVerifier},
 	}
 
@@ -266,6 +276,10 @@ func (o *CodexAuth) RefreshTokensWithRetry(ctx context.Context, refreshToken str
 		if err == nil {
 			return tokenData, nil
 		}
+		if isNonRetryableRefreshErr(err) {
+			log.Warnf("Token refresh attempt %d failed with non-retryable error: %v", attempt+1, err)
+			return nil, err
+		}
 
 		lastErr = err
 		log.Warnf("Token refresh attempt %d failed: %v", attempt+1, err)
@@ -274,6 +288,14 @@ func (o *CodexAuth) RefreshTokensWithRetry(ctx context.Context, refreshToken str
 	return nil, fmt.Errorf("token refresh failed after %d attempts: %w", maxRetries, lastErr)
 }
 
+func isNonRetryableRefreshErr(err error) bool {
+	if err == nil {
+		return false
+	}
+	raw := strings.ToLower(err.Error())
+	return strings.Contains(raw, "refresh_token_reused")
+}
+
 // UpdateTokenStorage updates an existing CodexTokenStorage with new token data.
 // This is typically called after a successful token refresh to persist the new credentials.
 func (o *CodexAuth) UpdateTokenStorage(storage *CodexTokenStorage, tokenData *CodexTokenData) {

internal/auth/codex/openai_auth_test.go

@@ -0,0 +1,44 @@
+package codex
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"sync/atomic"
+	"testing"
+)
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+func TestRefreshTokensWithRetry_NonRetryableOnlyAttemptsOnce(t *testing.T) {
+	var calls int32
+	auth := &CodexAuth{
+		httpClient: &http.Client{
+			Transport: roundTripFunc(func(req *http.Request) (*http.Response, error) {
+				atomic.AddInt32(&calls, 1)
+				return &http.Response{
+					StatusCode: http.StatusBadRequest,
+					Body:       io.NopCloser(strings.NewReader(`{"error":"invalid_grant","code":"refresh_token_reused"}`)),
+					Header:     make(http.Header),
+					Request:    req,
+				}, nil
+			}),
+		},
+	}
+
+	_, err := auth.RefreshTokensWithRetry(context.Background(), "dummy_refresh_token", 3)
+	if err == nil {
+		t.Fatalf("expected error for non-retryable refresh failure")
+	}
+	if !strings.Contains(strings.ToLower(err.Error()), "refresh_token_reused") {
+		t.Fatalf("expected refresh_token_reused in error, got: %v", err)
+	}
+	if got := atomic.LoadInt32(&calls); got != 1 {
+		t.Fatalf("expected 1 refresh attempt, got %d", got)
+	}
+}

internal/auth/codex/token.go

@@ -32,11 +32,21 @@ type CodexTokenStorage struct {
 	Type string `json:"type"`
 	// Expire is the timestamp when the current access token expires.
 	Expire string `json:"expired"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *CodexTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // SaveTokenToFile serializes the Codex token storage to a JSON file.
 // This method creates the necessary directory structure and writes the token
 // data in JSON format to the specified file path for persistent storage.
+// It merges any injected metadata into the top-level JSON object.
 //
 // Parameters:
 //   - authFilePath: The full path where the token file should be saved
@@ -58,7 +68,13 @@ func (ts *CodexTokenStorage) SaveTokenToFile(authFilePath string) error {
 		_ = f.Close()
 	}()
 
-	if err = json.NewEncoder(f).Encode(ts); err != nil {
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
+
+	if err = json.NewEncoder(f).Encode(data); err != nil {
 		return fmt.Errorf("failed to write token to file: %w", err)
 	}
 	return nil

internal/auth/gemini/gemini_token.go

@@ -35,11 +35,21 @@ type GeminiTokenStorage struct {
 
 	// Type indicates the authentication provider type, always "gemini" for this storage.
 	Type string `json:"type"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *GeminiTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // SaveTokenToFile serializes the Gemini token storage to a JSON file.
 // This method creates the necessary directory structure and writes the token
 // data in JSON format to the specified file path for persistent storage.
+// It merges any injected metadata into the top-level JSON object.
 //
 // Parameters:
 //   - authFilePath: The full path where the token file should be saved
@@ -49,6 +59,11 @@ type GeminiTokenStorage struct {
 func (ts *GeminiTokenStorage) SaveTokenToFile(authFilePath string) error {
 	misc.LogSavingCredentials(authFilePath)
 	ts.Type = "gemini"
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
 	if err := os.MkdirAll(filepath.Dir(authFilePath), 0700); err != nil {
 		return fmt.Errorf("failed to create directory: %v", err)
 	}
@@ -63,7 +78,9 @@ func (ts *GeminiTokenStorage) SaveTokenToFile(authFilePath string) error {
 		}
 	}()
 
-	if err = json.NewEncoder(f).Encode(ts); err != nil {
+	enc := json.NewEncoder(f)
+	enc.SetIndent("", "  ")
+	if err := enc.Encode(data); err != nil {
 		return fmt.Errorf("failed to write token to file: %w", err)
 	}
 	return nil

internal/auth/iflow/iflow_token.go

@@ -21,6 +21,15 @@ type IFlowTokenStorage struct {
 	Scope        string `json:"scope"`
 	Cookie       string `json:"cookie"`
 	Type         string `json:"type"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *IFlowTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // SaveTokenToFile serialises the token storage to disk.
@@ -37,7 +46,13 @@ func (ts *IFlowTokenStorage) SaveTokenToFile(authFilePath string) error {
 	}
 	defer func() { _ = f.Close() }()
 
-	if err = json.NewEncoder(f).Encode(ts); err != nil {
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
+
+	if err = json.NewEncoder(f).Encode(data); err != nil {
 		return fmt.Errorf("iflow token: encode token failed: %w", err)
 	}
 	return nil

internal/auth/kimi/token.go

@@ -29,6 +29,15 @@ type KimiTokenStorage struct {
 	Expired string `json:"expired,omitempty"`
 	// Type indicates the authentication provider type, always "kimi" for this storage.
 	Type string `json:"type"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *KimiTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // KimiTokenData holds the raw OAuth token response from Kimi.
@@ -86,9 +95,15 @@ func (ts *KimiTokenStorage) SaveTokenToFile(authFilePath string) error {
 		_ = f.Close()
 	}()
 
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
+
 	encoder := json.NewEncoder(f)
 	encoder.SetIndent("", "  ")
-	if err = encoder.Encode(ts); err != nil {
+	if err = encoder.Encode(data); err != nil {
 		return fmt.Errorf("failed to write token to file: %w", err)
 	}
 	return nil

internal/auth/qwen/qwen_token.go

@@ -30,11 +30,21 @@ type QwenTokenStorage struct {
 	Type string `json:"type"`
 	// Expire is the timestamp when the current access token expires.
 	Expire string `json:"expired"`
+
+	// Metadata holds arbitrary key-value pairs injected via hooks.
+	// It is not exported to JSON directly to allow flattening during serialization.
+	Metadata map[string]any `json:"-"`
+}
+
+// SetMetadata allows external callers to inject metadata into the storage before saving.
+func (ts *QwenTokenStorage) SetMetadata(meta map[string]any) {
+	ts.Metadata = meta
 }
 
 // SaveTokenToFile serializes the Qwen token storage to a JSON file.
 // This method creates the necessary directory structure and writes the token
 // data in JSON format to the specified file path for persistent storage.
+// It merges any injected metadata into the top-level JSON object.
 //
 // Parameters:
 //   - authFilePath: The full path where the token file should be saved
@@ -56,7 +66,13 @@ func (ts *QwenTokenStorage) SaveTokenToFile(authFilePath string) error {
 		_ = f.Close()
 	}()
 
-	if err = json.NewEncoder(f).Encode(ts); err != nil {
+	// Merge metadata using helper
+	data, errMerge := misc.MergeMetadata(ts, ts.Metadata)
+	if errMerge != nil {
+		return fmt.Errorf("failed to merge metadata: %w", errMerge)
+	}
+
+	if err = json.NewEncoder(f).Encode(data); err != nil {
 		return fmt.Errorf("failed to write token to file: %w", err)
 	}
 	return nil

internal/cmd/openai_device_login.go

@@ -0,0 +1,60 @@
+package cmd
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	codexLoginModeMetadataKey = "codex_login_mode"
+	codexLoginModeDevice      = "device"
+)
+
+// DoCodexDeviceLogin triggers the Codex device-code flow while keeping the
+// existing codex-login OAuth callback flow intact.
+func DoCodexDeviceLogin(cfg *config.Config, options *LoginOptions) {
+	if options == nil {
+		options = &LoginOptions{}
+	}
+
+	promptFn := options.Prompt
+	if promptFn == nil {
+		promptFn = defaultProjectPrompt()
+	}
+
+	manager := newAuthManager()
+
+	authOpts := &sdkAuth.LoginOptions{
+		NoBrowser:    options.NoBrowser,
+		CallbackPort: options.CallbackPort,
+		Metadata: map[string]string{
+			codexLoginModeMetadataKey: codexLoginModeDevice,
+		},
+		Prompt: promptFn,
+	}
+
+	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)
+	if err != nil {
+		if authErr, ok := errors.AsType[*codex.AuthenticationError](err); ok {
+			log.Error(codex.GetUserFriendlyMessage(authErr))
+			if authErr.Type == codex.ErrPortInUse.Type {
+				os.Exit(codex.ErrPortInUse.Code)
+			}
+			return
+		}
+		fmt.Printf("Codex device authentication failed: %v\n", err)
+		return
+	}
+
+	if savedPath != "" {
+		fmt.Printf("Authentication saved to %s\n", savedPath)
+	}
+	fmt.Println("Codex device authentication successful!")
+}

internal/config/config.go

@@ -301,6 +301,10 @@ type CloakConfig struct {
 	// SensitiveWords is a list of words to obfuscate with zero-width characters.
 	// This can help bypass certain content filters.
 	SensitiveWords []string `yaml:"sensitive-words,omitempty" json:"sensitive-words,omitempty"`
+
+	// CacheUserID controls whether Claude user_id values are cached per API key.
+	// When false, a fresh random user_id is generated for every request.
+	CacheUserID *bool `yaml:"cache-user-id,omitempty" json:"cache-user-id,omitempty"`
 }
 
 // ClaudeKey represents the configuration for a Claude API key,

internal/config/sdk_config.go

@@ -20,6 +20,10 @@ type SDKConfig struct {
 	// APIKeys is a list of keys for authenticating clients to this proxy server.
 	APIKeys []string `yaml:"api-keys" json:"api-keys"`
 
+	// PassthroughHeaders controls whether upstream response headers are forwarded to downstream clients.
+	// Default is false (disabled).
+	PassthroughHeaders bool `yaml:"passthrough-headers" json:"passthrough-headers"`
+
 	// Streaming configures server-side streaming behavior (keep-alives and safe bootstrap retries).
 	Streaming StreamingConfig `yaml:"streaming" json:"streaming"`
 

internal/misc/credentials.go

@@ -1,6 +1,7 @@
 package misc
 
 import (
+	"encoding/json"
 	"fmt"
 	"path/filepath"
 	"strings"
@@ -24,3 +25,37 @@ func LogSavingCredentials(path string) {
 func LogCredentialSeparator() {
 	log.Debug(credentialSeparator)
 }
+
+// MergeMetadata serializes the source struct into a map and merges the provided metadata into it.
+func MergeMetadata(source any, metadata map[string]any) (map[string]any, error) {
+	var data map[string]any
+
+	// Fast path: if source is already a map, just copy it to avoid mutation of original
+	if srcMap, ok := source.(map[string]any); ok {
+		data = make(map[string]any, len(srcMap)+len(metadata))
+		for k, v := range srcMap {
+			data[k] = v
+		}
+	} else {
+		// Slow path: marshal to JSON and back to map to respect JSON tags
+		temp, err := json.Marshal(source)
+		if err != nil {
+			return nil, fmt.Errorf("failed to marshal source: %w", err)
+		}
+		if err := json.Unmarshal(temp, &data); err != nil {
+			return nil, fmt.Errorf("failed to unmarshal to map: %w", err)
+		}
+	}
+
+	// Merge extra metadata
+	if metadata != nil {
+		if data == nil {
+			data = make(map[string]any)
+		}
+		for k, v := range metadata {
+			data[k] = v
+		}
+	}
+
+	return data, nil
+}

internal/registry/model_definitions_static_data.go

@@ -184,6 +184,21 @@ func GetGeminiModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -294,6 +309,21 @@ func GetGeminiVertexModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-pro-image-preview",
 			Object:                     "model",
@@ -436,6 +466,21 @@ func GetGeminiCLIModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -517,6 +562,21 @@ func GetAIStudioModels() []*ModelInfo {
 			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
 			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
 		},
+		{
+			ID:                         "gemini-3.1-pro-preview",
+			Object:                     "model",
+			Created:                    1771459200,
+			OwnedBy:                    "google",
+			Type:                       "gemini",
+			Name:                       "models/gemini-3.1-pro-preview",
+			Version:                    "3.1",
+			DisplayName:                "Gemini 3.1 Pro Preview",
+			Description:                "Gemini 3.1 Pro Preview",
+			InputTokenLimit:            1048576,
+			OutputTokenLimit:           65536,
+			SupportedGenerationMethods: []string{"generateContent", "countTokens", "createCachedContent", "batchGenerateContent"},
+			Thinking:                   &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true},
+		},
 		{
 			ID:                         "gemini-3-flash-preview",
 			Object:                     "model",
@@ -903,6 +963,7 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-2.5-flash-lite":      {Thinking: &ThinkingSupport{Min: 0, Max: 24576, ZeroAllowed: true, DynamicAllowed: true}},
 		"gemini-3-pro-high":          {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3-pro-image":         {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
+		"gemini-3.1-pro-high":        {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"low", "high"}}},
 		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},

internal/runtime/executor/aistudio_executor.go

@@ -164,12 +164,12 @@ func (e *AIStudioExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth,
 	reporter.publish(ctx, parseGeminiUsage(wsResp.Body))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, body.toFormat, opts.SourceFormat, req.Model, opts.OriginalRequest, translatedReq, wsResp.Body, &param)
-	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out))}
+	resp = cliproxyexecutor.Response{Payload: ensureColonSpacedJSON([]byte(out)), Headers: wsResp.Headers.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming request to the AI Studio API.
-func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -254,7 +254,6 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 		return nil, statusErr{code: firstEvent.Status, msg: body.String()}
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func(first wsrelay.StreamEvent) {
 		defer close(out)
 		var param any
@@ -318,7 +317,7 @@ func (e *AIStudioExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth
 			}
 		}
 	}(firstEvent)
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: firstEvent.Headers.Clone(), Chunks: out}, nil
 }
 
 // CountTokens counts tokens for the given request using the AI Studio API.

internal/runtime/executor/antigravity_executor.go

@@ -232,7 +232,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(bodyBytes))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, bodyBytes, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)
 			return resp, nil
 		}
@@ -436,7 +436,7 @@ attemptLoop:
 			reporter.publish(ctx, parseAntigravityUsage(resp.Payload))
 			var param any
 			converted := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, resp.Payload, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(converted)}
+			resp = cliproxyexecutor.Response{Payload: []byte(converted), Headers: httpResp.Header.Clone()}
 			reporter.ensurePublished(ctx)
 
 			return resp, nil
@@ -645,7 +645,7 @@ func (e *AntigravityExecutor) convertStreamToNonStream(stream []byte) []byte {
 }
 
 // ExecuteStream performs a streaming request to the Antigravity API.
-func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -775,7 +775,6 @@ attemptLoop:
 			}
 
 			out := make(chan cliproxyexecutor.StreamChunk)
-			stream = out
 			go func(resp *http.Response) {
 				defer close(out)
 				defer func() {
@@ -820,7 +819,7 @@ attemptLoop:
 					reporter.ensurePublished(ctx)
 				}
 			}(httpResp)
-			return stream, nil
+			return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 		}
 
 		switch {
@@ -968,7 +967,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 		if httpResp.StatusCode >= http.StatusOK && httpResp.StatusCode < http.StatusMultipleChoices {
 			count := gjson.GetBytes(bodyBytes, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, bodyBytes)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: httpResp.Header.Clone()}, nil
 		}
 
 		lastStatus = httpResp.StatusCode

internal/runtime/executor/claude_executor.go

@@ -117,7 +117,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 
 	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
 	// based on client type and configuration.
-	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel, apiKey)
 
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
@@ -223,11 +223,11 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 		data,
 		&param,
 	)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -258,7 +258,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 
 	// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
 	// based on client type and configuration.
-	body = applyCloaking(ctx, e.cfg, auth, body, baseModel)
+	body = applyCloaking(ctx, e.cfg, auth, body, baseModel, apiKey)
 
 	requestedModel := payloadRequestedModel(opts, req.Model)
 	body = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", body, originalTranslated, requestedModel)
@@ -330,7 +330,6 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -399,7 +398,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
@@ -488,7 +487,7 @@ func (e *ClaudeExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "input_tokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: resp.Header.Clone()}, nil
 }
 
 func (e *ClaudeExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
@@ -983,10 +982,10 @@ func getClientUserAgent(ctx context.Context) string {
 }
 
 // getCloakConfigFromAuth extracts cloak configuration from auth attributes.
-// Returns (cloakMode, strictMode, sensitiveWords).
-func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string) {
+// Returns (cloakMode, strictMode, sensitiveWords, cacheUserID).
+func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string, bool) {
 	if auth == nil || auth.Attributes == nil {
-		return "auto", false, nil
+		return "auto", false, nil, false
 	}
 
 	cloakMode := auth.Attributes["cloak_mode"]
@@ -1004,7 +1003,9 @@ func getCloakConfigFromAuth(auth *cliproxyauth.Auth) (string, bool, []string) {
 		}
 	}
 
-	return cloakMode, strictMode, sensitiveWords
+	cacheUserID := strings.EqualFold(strings.TrimSpace(auth.Attributes["cloak_cache_user_id"]), "true")
+
+	return cloakMode, strictMode, sensitiveWords, cacheUserID
 }
 
 // resolveClaudeKeyCloakConfig finds the matching ClaudeKey config and returns its CloakConfig.
@@ -1037,16 +1038,24 @@ func resolveClaudeKeyCloakConfig(cfg *config.Config, auth *cliproxyauth.Auth) *c
 }
 
 // injectFakeUserID generates and injects a fake user ID into the request metadata.
-func injectFakeUserID(payload []byte) []byte {
+// When useCache is false, a new user ID is generated for every call.
+func injectFakeUserID(payload []byte, apiKey string, useCache bool) []byte {
+	generateID := func() string {
+		if useCache {
+			return cachedUserID(apiKey)
+		}
+		return generateFakeUserID()
+	}
+
 	metadata := gjson.GetBytes(payload, "metadata")
 	if !metadata.Exists() {
-		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateID())
 		return payload
 	}
 
 	existingUserID := gjson.GetBytes(payload, "metadata.user_id").String()
 	if existingUserID == "" || !isValidUserID(existingUserID) {
-		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateFakeUserID())
+		payload, _ = sjson.SetBytes(payload, "metadata.user_id", generateID())
 	}
 	return payload
 }
@@ -1083,7 +1092,7 @@ func checkSystemInstructionsWithMode(payload []byte, strictMode bool) []byte {
 
 // applyCloaking applies cloaking transformations to the payload based on config and client.
 // Cloaking includes: system prompt injection, fake user ID, and sensitive word obfuscation.
-func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte, model string) []byte {
+func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte, model string, apiKey string) []byte {
 	clientUserAgent := getClientUserAgent(ctx)
 
 	// Get cloak config from ClaudeKey configuration
@@ -1093,16 +1102,20 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 	var cloakMode string
 	var strictMode bool
 	var sensitiveWords []string
+	var cacheUserID bool
 
 	if cloakCfg != nil {
 		cloakMode = cloakCfg.Mode
 		strictMode = cloakCfg.StrictMode
 		sensitiveWords = cloakCfg.SensitiveWords
+		if cloakCfg.CacheUserID != nil {
+			cacheUserID = *cloakCfg.CacheUserID
+		}
 	}
 
 	// Fallback to auth attributes if no config found
 	if cloakMode == "" {
-		attrMode, attrStrict, attrWords := getCloakConfigFromAuth(auth)
+		attrMode, attrStrict, attrWords, attrCache := getCloakConfigFromAuth(auth)
 		cloakMode = attrMode
 		if !strictMode {
 			strictMode = attrStrict
@@ -1110,6 +1123,12 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 		if len(sensitiveWords) == 0 {
 			sensitiveWords = attrWords
 		}
+		if cloakCfg == nil || cloakCfg.CacheUserID == nil {
+			cacheUserID = attrCache
+		}
+	} else if cloakCfg == nil || cloakCfg.CacheUserID == nil {
+		_, _, _, attrCache := getCloakConfigFromAuth(auth)
+		cacheUserID = attrCache
 	}
 
 	// Determine if cloaking should be applied
@@ -1123,7 +1142,7 @@ func applyCloaking(ctx context.Context, cfg *config.Config, auth *cliproxyauth.A
 	}
 
 	// Inject fake user ID
-	payload = injectFakeUserID(payload)
+	payload = injectFakeUserID(payload, apiKey, cacheUserID)
 
 	// Apply sensitive word obfuscation
 	if len(sensitiveWords) > 0 {

internal/runtime/executor/claude_executor_test.go

@@ -2,9 +2,18 @@ package executor
 
 import (
 	"bytes"
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
 	"testing"
 
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
 )
 
 func TestApplyClaudeToolPrefix(t *testing.T) {
@@ -199,6 +208,119 @@ func TestApplyClaudeToolPrefix_NestedToolReference(t *testing.T) {
 	}
 }
 
+func TestClaudeExecutor_ReusesUserIDAcrossModelsWhenCacheEnabled(t *testing.T) {
+	resetUserIDCache()
+
+	var userIDs []string
+	var requestModels []string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		userID := gjson.GetBytes(body, "metadata.user_id").String()
+		model := gjson.GetBytes(body, "model").String()
+		userIDs = append(userIDs, userID)
+		requestModels = append(requestModels, model)
+		t.Logf("HTTP Server received request: model=%s, user_id=%s, url=%s", model, userID, r.URL.String())
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet","role":"assistant","content":[{"type":"text","text":"ok"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	t.Logf("End-to-end test: Fake HTTP server started at %s", server.URL)
+
+	cacheEnabled := true
+	executor := NewClaudeExecutor(&config.Config{
+		ClaudeKey: []config.ClaudeKey{
+			{
+				APIKey:  "key-123",
+				BaseURL: server.URL,
+				Cloak: &config.CloakConfig{
+					CacheUserID: &cacheEnabled,
+				},
+			},
+		},
+	})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+	models := []string{"claude-3-5-sonnet", "claude-3-5-haiku"}
+	for _, model := range models {
+		t.Logf("Sending request for model: %s", model)
+		modelPayload, _ := sjson.SetBytes(payload, "model", model)
+		if _, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   model,
+			Payload: modelPayload,
+		}, cliproxyexecutor.Options{
+			SourceFormat: sdktranslator.FromString("claude"),
+		}); err != nil {
+			t.Fatalf("Execute(%s) error: %v", model, err)
+		}
+	}
+
+	if len(userIDs) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(userIDs))
+	}
+	if userIDs[0] == "" || userIDs[1] == "" {
+		t.Fatal("expected user_id to be populated")
+	}
+	t.Logf("user_id[0] (model=%s): %s", requestModels[0], userIDs[0])
+	t.Logf("user_id[1] (model=%s): %s", requestModels[1], userIDs[1])
+	if userIDs[0] != userIDs[1] {
+		t.Fatalf("expected user_id to be reused across models, got %q and %q", userIDs[0], userIDs[1])
+	}
+	if !isValidUserID(userIDs[0]) {
+		t.Fatalf("user_id %q is not valid", userIDs[0])
+	}
+	t.Logf("✓ End-to-end test passed: Same user_id (%s) was used for both models", userIDs[0])
+}
+
+func TestClaudeExecutor_GeneratesNewUserIDByDefault(t *testing.T) {
+	resetUserIDCache()
+
+	var userIDs []string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		userIDs = append(userIDs, gjson.GetBytes(body, "metadata.user_id").String())
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"msg_1","type":"message","model":"claude-3-5-sonnet","role":"assistant","content":[{"type":"text","text":"ok"}],"usage":{"input_tokens":1,"output_tokens":1}}`))
+	}))
+	defer server.Close()
+
+	executor := NewClaudeExecutor(&config.Config{})
+	auth := &cliproxyauth.Auth{Attributes: map[string]string{
+		"api_key":  "key-123",
+		"base_url": server.URL,
+	}}
+
+	payload := []byte(`{"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	for i := 0; i < 2; i++ {
+		if _, err := executor.Execute(context.Background(), auth, cliproxyexecutor.Request{
+			Model:   "claude-3-5-sonnet",
+			Payload: payload,
+		}, cliproxyexecutor.Options{
+			SourceFormat: sdktranslator.FromString("claude"),
+		}); err != nil {
+			t.Fatalf("Execute call %d error: %v", i, err)
+		}
+	}
+
+	if len(userIDs) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(userIDs))
+	}
+	if userIDs[0] == "" || userIDs[1] == "" {
+		t.Fatal("expected user_id to be populated")
+	}
+	if userIDs[0] == userIDs[1] {
+		t.Fatalf("expected user_id to change when caching is not enabled, got identical values %q", userIDs[0])
+	}
+	if !isValidUserID(userIDs[0]) || !isValidUserID(userIDs[1]) {
+		t.Fatalf("user_ids should be valid, got %q and %q", userIDs[0], userIDs[1])
+	}
+}
+
 func TestStripClaudeToolPrefixFromResponse_NestedToolReference(t *testing.T) {
 	input := []byte(`{"content":[{"type":"tool_result","tool_use_id":"toolu_123","content":[{"type":"tool_reference","tool_name":"proxy_mcp__nia__manage_resource"}]}]}`)
 	out := stripClaudeToolPrefixFromResponse(input, "proxy_")

internal/runtime/executor/codex_executor.go

@@ -183,7 +183,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 
 		var param any
 		out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, line, &param)
-		resp = cliproxyexecutor.Response{Payload: []byte(out)}
+		resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 		return resp, nil
 	}
 	err = statusErr{code: 408, msg: "stream error: stream disconnected before completion: stream closed before response.completed"}
@@ -273,11 +273,11 @@ func (e *CodexExecutor) executeCompact(ctx context.Context, auth *cliproxyauth.A
 	reporter.ensurePublished(ctx)
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, originalPayload, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusBadRequest, msg: "streaming not supported for /responses/compact"}
 	}
@@ -362,7 +362,6 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -397,7 +396,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/codex_websockets_executor.go

@@ -363,7 +363,7 @@ func (e *CodexWebsocketsExecutor) Execute(ctx context.Context, auth *cliproxyaut
 	}
 }
 
-func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	log.Debugf("Executing Codex Websockets stream request with auth ID: %s, model: %s", auth.ID, req.Model)
 	if ctx == nil {
 		ctx = context.Background()
@@ -436,7 +436,9 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	})
 
 	conn, respHS, errDial := e.ensureUpstreamConn(ctx, auth, sess, authID, wsURL, wsHeaders)
+	var upstreamHeaders http.Header
 	if respHS != nil {
+		upstreamHeaders = respHS.Header.Clone()
 		recordAPIResponseMetadata(ctx, e.cfg, respHS.StatusCode, respHS.Header.Clone())
 	}
 	if errDial != nil {
@@ -516,7 +518,6 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 	markCodexWebsocketCreateSent(sess, conn, wsReqBody)
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		terminateReason := "completed"
 		var terminateErr error
@@ -627,7 +628,7 @@ func (e *CodexWebsocketsExecutor) ExecuteStream(ctx context.Context, auth *clipr
 		}
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: upstreamHeaders, Chunks: out}, nil
 }
 
 func (e *CodexWebsocketsExecutor) dialCodexWebsocket(ctx context.Context, auth *cliproxyauth.Auth, wsURL string, headers http.Header) (*websocket.Conn, *http.Response, error) {
@@ -1343,7 +1344,7 @@ func (e *CodexAutoExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	return e.httpExec.Execute(ctx, auth, req, opts)
 }
 
-func (e *CodexAutoExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (e *CodexAutoExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if e == nil || e.httpExec == nil || e.wsExec == nil {
 		return nil, fmt.Errorf("codex auto executor: executor is nil")
 	}

internal/runtime/executor/gemini_cli_executor.go

@@ -225,7 +225,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 			reporter.publish(ctx, parseGeminiCLIUsage(data))
 			var param any
 			out := sdktranslator.TranslateNonStream(respCtx, to, from, attemptModel, opts.OriginalRequest, payload, data, &param)
-			resp = cliproxyexecutor.Response{Payload: []byte(out)}
+			resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 			return resp, nil
 		}
 
@@ -256,7 +256,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 }
 
 // ExecuteStream performs a streaming request to the Gemini CLI API.
-func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -382,7 +382,6 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 		}
 
 		out := make(chan cliproxyexecutor.StreamChunk)
-		stream = out
 		go func(resp *http.Response, reqBody []byte, attemptModel string) {
 			defer close(out)
 			defer func() {
@@ -441,7 +440,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 			}
 		}(httpResp, append([]byte(nil), payload...), attemptModel)
 
-		return stream, nil
+		return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 	}
 
 	if len(lastBody) > 0 {
@@ -546,7 +545,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 		if resp.StatusCode >= 200 && resp.StatusCode < 300 {
 			count := gjson.GetBytes(data, "totalTokens").Int()
 			translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-			return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+			return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 		}
 		lastStatus = resp.StatusCode
 		lastBody = append([]byte(nil), data...)

internal/runtime/executor/gemini_executor.go

@@ -205,12 +205,12 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming request to the Gemini API.
-func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -298,7 +298,6 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -335,7 +334,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // CountTokens counts tokens for the given request using the Gemini API.
@@ -416,7 +415,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	translated := sdktranslator.TranslateTokenCount(respCtx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(translated), Headers: resp.Header.Clone()}, nil
 }
 
 // Refresh refreshes the authentication credentials (no-op for Gemini API key).

internal/runtime/executor/gemini_vertex_executor.go

@@ -253,7 +253,7 @@ func (e *GeminiVertexExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 }
 
 // ExecuteStream performs a streaming request to the Vertex AI API.
-func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -419,7 +419,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	to := sdktranslator.FromString("gemini")
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
@@ -524,12 +524,12 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	reporter.publish(ctx, parseGeminiUsage(data))
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // executeStreamWithServiceAccount handles streaming authentication using service account credentials.
-func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, projectID, location string, saJSON []byte) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -618,7 +618,6 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -650,11 +649,11 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // executeStreamWithAPIKey handles streaming authentication using API key credentials.
-func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, apiKey, baseURL string) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -743,7 +742,6 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -775,7 +773,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // countTokensWithServiceAccount counts tokens using service account credentials.
@@ -859,7 +857,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }
 
 // countTokensWithAPIKey handles token counting using API key credentials.
@@ -943,7 +941,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 	appendAPIResponseChunk(ctx, e.cfg, data)
 	count := gjson.GetBytes(data, "totalTokens").Int()
 	out := sdktranslator.TranslateTokenCount(ctx, to, from, count, data)
-	return cliproxyexecutor.Response{Payload: []byte(out)}, nil
+	return cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}, nil
 }
 
 // vertexCreds extracts project, location and raw service account JSON from auth metadata.

internal/runtime/executor/iflow_executor.go

@@ -169,12 +169,12 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	// 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, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming chat completion request.
-func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
@@ -262,7 +262,6 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	}
 
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -294,7 +293,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 		reporter.ensurePublished(ctx)
 	}()
 
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/kimi_executor.go

@@ -161,12 +161,12 @@ func (e *KimiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// 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, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
 // ExecuteStream performs a streaming chat completion request to Kimi.
-func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	from := opts.SourceFormat
 	if from.String() == "claude" {
 		auth.Attributes["base_url"] = kimiauth.KimiAPIBaseURL
@@ -253,7 +253,6 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -285,7 +284,7 @@ func (e *KimiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 // CountTokens estimates token count for Kimi requests.

internal/runtime/executor/openai_compat_executor.go

@@ -172,11 +172,11 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	// Translate response back to source format when needed
 	var param any
 	out := sdktranslator.TranslateNonStream(ctx, to, from, req.Model, opts.OriginalRequest, translated, body, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	reporter := newUsageReporter(ctx, e.Identifier(), baseModel, auth)
@@ -258,7 +258,6 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -298,7 +297,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 		// Ensure we record the request if no usage chunk was ever seen
 		reporter.ensurePublished(ctx)
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/qwen_executor.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"net/http"
 	"strings"
+	"sync"
 	"time"
 
 	qwenauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
@@ -22,9 +23,151 @@ import (
 )
 
 const (
-	qwenUserAgent = "QwenCode/0.10.3 (darwin; arm64)"
+	qwenUserAgent       = "QwenCode/0.10.3 (darwin; arm64)"
+	qwenRateLimitPerMin = 60          // 60 requests per minute per credential
+	qwenRateLimitWindow = time.Minute // sliding window duration
 )
 
+// qwenBeijingLoc caches the Beijing timezone to avoid repeated LoadLocation syscalls.
+var qwenBeijingLoc = func() *time.Location {
+	loc, err := time.LoadLocation("Asia/Shanghai")
+	if err != nil || loc == nil {
+		log.Warnf("qwen: failed to load Asia/Shanghai timezone: %v, using fixed UTC+8", err)
+		return time.FixedZone("CST", 8*3600)
+	}
+	return loc
+}()
+
+// qwenQuotaCodes is a package-level set of error codes that indicate quota exhaustion.
+var qwenQuotaCodes = map[string]struct{}{
+	"insufficient_quota": {},
+	"quota_exceeded":     {},
+}
+
+// qwenRateLimiter tracks request timestamps per credential for rate limiting.
+// Qwen has a limit of 60 requests per minute per account.
+var qwenRateLimiter = struct {
+	sync.Mutex
+	requests map[string][]time.Time // authID -> request timestamps
+}{
+	requests: make(map[string][]time.Time),
+}
+
+// redactAuthID returns a redacted version of the auth ID for safe logging.
+// Keeps a small prefix/suffix to allow correlation across events.
+func redactAuthID(id string) string {
+	if id == "" {
+		return ""
+	}
+	if len(id) <= 8 {
+		return id
+	}
+	return id[:4] + "..." + id[len(id)-4:]
+}
+
+// checkQwenRateLimit checks if the credential has exceeded the rate limit.
+// Returns nil if allowed, or a statusErr with retryAfter if rate limited.
+func checkQwenRateLimit(authID string) error {
+	if authID == "" {
+		// Empty authID should not bypass rate limiting in production
+		// Use debug level to avoid log spam for certain auth flows
+		log.Debug("qwen rate limit check: empty authID, skipping rate limit")
+		return nil
+	}
+
+	now := time.Now()
+	windowStart := now.Add(-qwenRateLimitWindow)
+
+	qwenRateLimiter.Lock()
+	defer qwenRateLimiter.Unlock()
+
+	// Get and filter timestamps within the window
+	timestamps := qwenRateLimiter.requests[authID]
+	var validTimestamps []time.Time
+	for _, ts := range timestamps {
+		if ts.After(windowStart) {
+			validTimestamps = append(validTimestamps, ts)
+		}
+	}
+
+	// Always prune expired entries to prevent memory leak
+	// Delete empty entries, otherwise update with pruned slice
+	if len(validTimestamps) == 0 {
+		delete(qwenRateLimiter.requests, authID)
+	}
+
+	// Check if rate limit exceeded
+	if len(validTimestamps) >= qwenRateLimitPerMin {
+		// Calculate when the oldest request will expire
+		oldestInWindow := validTimestamps[0]
+		retryAfter := oldestInWindow.Add(qwenRateLimitWindow).Sub(now)
+		if retryAfter < time.Second {
+			retryAfter = time.Second
+		}
+		retryAfterSec := int(retryAfter.Seconds())
+		return statusErr{
+			code:       http.StatusTooManyRequests,
+			msg:        fmt.Sprintf(`{"error":{"code":"rate_limit_exceeded","message":"Qwen rate limit: %d requests/minute exceeded, retry after %ds","type":"rate_limit_exceeded"}}`, qwenRateLimitPerMin, retryAfterSec),
+			retryAfter: &retryAfter,
+		}
+	}
+
+	// Record this request and update the map with pruned timestamps
+	validTimestamps = append(validTimestamps, now)
+	qwenRateLimiter.requests[authID] = validTimestamps
+
+	return nil
+}
+
+// isQwenQuotaError checks if the error response indicates a quota exceeded error.
+// Qwen returns HTTP 403 with error.code="insufficient_quota" when daily quota is exhausted.
+func isQwenQuotaError(body []byte) bool {
+	code := strings.ToLower(gjson.GetBytes(body, "error.code").String())
+	errType := strings.ToLower(gjson.GetBytes(body, "error.type").String())
+
+	// Primary check: exact match on error.code or error.type (most reliable)
+	if _, ok := qwenQuotaCodes[code]; ok {
+		return true
+	}
+	if _, ok := qwenQuotaCodes[errType]; ok {
+		return true
+	}
+
+	// Fallback: check message only if code/type don't match (less reliable)
+	msg := strings.ToLower(gjson.GetBytes(body, "error.message").String())
+	if strings.Contains(msg, "insufficient_quota") || strings.Contains(msg, "quota exceeded") ||
+		strings.Contains(msg, "free allocated quota exceeded") {
+		return true
+	}
+
+	return false
+}
+
+// wrapQwenError wraps an HTTP error response, detecting quota errors and mapping them to 429.
+// Returns the appropriate status code and retryAfter duration for statusErr.
+// Only checks for quota errors when httpCode is 403 or 429 to avoid false positives.
+func wrapQwenError(ctx context.Context, httpCode int, body []byte) (errCode int, retryAfter *time.Duration) {
+	errCode = httpCode
+	// Only check quota errors for expected status codes to avoid false positives
+	// Qwen returns 403 for quota errors, 429 for rate limits
+	if (httpCode == http.StatusForbidden || httpCode == http.StatusTooManyRequests) && isQwenQuotaError(body) {
+		errCode = http.StatusTooManyRequests // Map to 429 to trigger quota logic
+		cooldown := timeUntilNextDay()
+		retryAfter = &cooldown
+		logWithRequestID(ctx).Warnf("qwen quota exceeded (http %d -> %d), cooling down until tomorrow (%v)", httpCode, errCode, cooldown)
+	}
+	return errCode, retryAfter
+}
+
+// timeUntilNextDay returns duration until midnight Beijing time (UTC+8).
+// Qwen's daily quota resets at 00:00 Beijing time.
+func timeUntilNextDay() time.Duration {
+	now := time.Now()
+	nowLocal := now.In(qwenBeijingLoc)
+	tomorrow := time.Date(nowLocal.Year(), nowLocal.Month(), nowLocal.Day()+1, 0, 0, 0, 0, qwenBeijingLoc)
+	return tomorrow.Sub(now)
+}
+
 // QwenExecutor is a stateless executor for Qwen Code using OpenAI-compatible chat completions.
 // If access token is unavailable, it falls back to legacy via ClientAdapter.
 type QwenExecutor struct {
@@ -67,6 +210,17 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	if opts.Alt == "responses/compact" {
 		return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
+
+	// Check rate limit before proceeding
+	var authID string
+	if auth != nil {
+		authID = auth.ID
+	}
+	if err := checkQwenRateLimit(authID); err != nil {
+		logWithRequestID(ctx).Warnf("qwen rate limit exceeded for credential %s", redactAuthID(authID))
+		return resp, err
+	}
+
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	token, baseURL := qwenCreds(auth)
@@ -102,9 +256,8 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 		return resp, err
 	}
 	applyQwenHeaders(httpReq, token, false)
-	var authID, authLabel, authType, authValue string
+	var authLabel, authType, authValue string
 	if auth != nil {
-		authID = auth.ID
 		authLabel = auth.Label
 		authType, authValue = auth.AccountInfo()
 	}
@@ -135,8 +288,10 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
-		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+
+		errCode, retryAfter := wrapQwenError(ctx, httpResp.StatusCode, b)
+		logWithRequestID(ctx).Debugf("request error, error status: %d (mapped: %d), error message: %s", httpResp.StatusCode, errCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+		err = statusErr{code: errCode, msg: string(b), retryAfter: retryAfter}
 		return resp, err
 	}
 	data, err := io.ReadAll(httpResp.Body)
@@ -150,14 +305,25 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	// 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, opts.OriginalRequest, body, data, &param)
-	resp = cliproxyexecutor.Response{Payload: []byte(out)}
+	resp = cliproxyexecutor.Response{Payload: []byte(out), Headers: httpResp.Header.Clone()}
 	return resp, nil
 }
 
-func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (stream <-chan cliproxyexecutor.StreamChunk, err error) {
+func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (_ *cliproxyexecutor.StreamResult, err error) {
 	if opts.Alt == "responses/compact" {
 		return nil, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
 	}
+
+	// Check rate limit before proceeding
+	var authID string
+	if auth != nil {
+		authID = auth.ID
+	}
+	if err := checkQwenRateLimit(authID); err != nil {
+		logWithRequestID(ctx).Warnf("qwen rate limit exceeded for credential %s", redactAuthID(authID))
+		return nil, err
+	}
+
 	baseModel := thinking.ParseSuffix(req.Model).ModelName
 
 	token, baseURL := qwenCreds(auth)
@@ -200,9 +366,8 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 		return nil, err
 	}
 	applyQwenHeaders(httpReq, token, true)
-	var authID, authLabel, authType, authValue string
+	var authLabel, authType, authValue string
 	if auth != nil {
-		authID = auth.ID
 		authLabel = auth.Label
 		authType, authValue = auth.AccountInfo()
 	}
@@ -228,15 +393,16 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
 		b, _ := io.ReadAll(httpResp.Body)
 		appendAPIResponseChunk(ctx, e.cfg, b)
-		logWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
+
+		errCode, retryAfter := wrapQwenError(ctx, httpResp.StatusCode, b)
+		logWithRequestID(ctx).Debugf("request error, error status: %d (mapped: %d), error message: %s", httpResp.StatusCode, errCode, summarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
 		if errClose := httpResp.Body.Close(); errClose != nil {
 			log.Errorf("qwen executor: close response body error: %v", errClose)
 		}
-		err = statusErr{code: httpResp.StatusCode, msg: string(b)}
+		err = statusErr{code: errCode, msg: string(b), retryAfter: retryAfter}
 		return nil, err
 	}
 	out := make(chan cliproxyexecutor.StreamChunk)
-	stream = out
 	go func() {
 		defer close(out)
 		defer func() {
@@ -268,7 +434,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 			out <- cliproxyexecutor.StreamChunk{Err: errScan}
 		}
 	}()
-	return stream, nil
+	return &cliproxyexecutor.StreamResult{Headers: httpResp.Header.Clone(), Chunks: out}, nil
 }
 
 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {

internal/runtime/executor/user_id_cache.go

@@ -0,0 +1,89 @@
+package executor
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"sync"
+	"time"
+)
+
+type userIDCacheEntry struct {
+	value  string
+	expire time.Time
+}
+
+var (
+	userIDCache            = make(map[string]userIDCacheEntry)
+	userIDCacheMu          sync.RWMutex
+	userIDCacheCleanupOnce sync.Once
+)
+
+const (
+	userIDTTL                = time.Hour
+	userIDCacheCleanupPeriod = 15 * time.Minute
+)
+
+func startUserIDCacheCleanup() {
+	go func() {
+		ticker := time.NewTicker(userIDCacheCleanupPeriod)
+		defer ticker.Stop()
+		for range ticker.C {
+			purgeExpiredUserIDs()
+		}
+	}()
+}
+
+func purgeExpiredUserIDs() {
+	now := time.Now()
+	userIDCacheMu.Lock()
+	for key, entry := range userIDCache {
+		if !entry.expire.After(now) {
+			delete(userIDCache, key)
+		}
+	}
+	userIDCacheMu.Unlock()
+}
+
+func userIDCacheKey(apiKey string) string {
+	sum := sha256.Sum256([]byte(apiKey))
+	return hex.EncodeToString(sum[:])
+}
+
+func cachedUserID(apiKey string) string {
+	if apiKey == "" {
+		return generateFakeUserID()
+	}
+
+	userIDCacheCleanupOnce.Do(startUserIDCacheCleanup)
+
+	key := userIDCacheKey(apiKey)
+	now := time.Now()
+
+	userIDCacheMu.RLock()
+	entry, ok := userIDCache[key]
+	valid := ok && entry.value != "" && entry.expire.After(now) && isValidUserID(entry.value)
+	userIDCacheMu.RUnlock()
+	if valid {
+		userIDCacheMu.Lock()
+		entry = userIDCache[key]
+		if entry.value != "" && entry.expire.After(now) && isValidUserID(entry.value) {
+			entry.expire = now.Add(userIDTTL)
+			userIDCache[key] = entry
+			userIDCacheMu.Unlock()
+			return entry.value
+		}
+		userIDCacheMu.Unlock()
+	}
+
+	newID := generateFakeUserID()
+
+	userIDCacheMu.Lock()
+	entry, ok = userIDCache[key]
+	if !ok || entry.value == "" || !entry.expire.After(now) || !isValidUserID(entry.value) {
+		entry.value = newID
+	}
+	entry.expire = now.Add(userIDTTL)
+	userIDCache[key] = entry
+	userIDCacheMu.Unlock()
+	return entry.value
+}

internal/runtime/executor/user_id_cache_test.go

@@ -0,0 +1,86 @@
+package executor
+
+import (
+	"testing"
+	"time"
+)
+
+func resetUserIDCache() {
+	userIDCacheMu.Lock()
+	userIDCache = make(map[string]userIDCacheEntry)
+	userIDCacheMu.Unlock()
+}
+
+func TestCachedUserID_ReusesWithinTTL(t *testing.T) {
+	resetUserIDCache()
+
+	first := cachedUserID("api-key-1")
+	second := cachedUserID("api-key-1")
+
+	if first == "" {
+		t.Fatal("expected generated user_id to be non-empty")
+	}
+	if first != second {
+		t.Fatalf("expected cached user_id to be reused, got %q and %q", first, second)
+	}
+}
+
+func TestCachedUserID_ExpiresAfterTTL(t *testing.T) {
+	resetUserIDCache()
+
+	expiredID := cachedUserID("api-key-expired")
+	cacheKey := userIDCacheKey("api-key-expired")
+	userIDCacheMu.Lock()
+	userIDCache[cacheKey] = userIDCacheEntry{
+		value:  expiredID,
+		expire: time.Now().Add(-time.Minute),
+	}
+	userIDCacheMu.Unlock()
+
+	newID := cachedUserID("api-key-expired")
+	if newID == expiredID {
+		t.Fatalf("expected expired user_id to be replaced, got %q", newID)
+	}
+	if newID == "" {
+		t.Fatal("expected regenerated user_id to be non-empty")
+	}
+}
+
+func TestCachedUserID_IsScopedByAPIKey(t *testing.T) {
+	resetUserIDCache()
+
+	first := cachedUserID("api-key-1")
+	second := cachedUserID("api-key-2")
+
+	if first == second {
+		t.Fatalf("expected different API keys to have different user_ids, got %q", first)
+	}
+}
+
+func TestCachedUserID_RenewsTTLOnHit(t *testing.T) {
+	resetUserIDCache()
+
+	key := "api-key-renew"
+	id := cachedUserID(key)
+	cacheKey := userIDCacheKey(key)
+
+	soon := time.Now()
+	userIDCacheMu.Lock()
+	userIDCache[cacheKey] = userIDCacheEntry{
+		value:  id,
+		expire: soon.Add(2 * time.Second),
+	}
+	userIDCacheMu.Unlock()
+
+	if refreshed := cachedUserID(key); refreshed != id {
+		t.Fatalf("expected cached user_id to be reused before expiry, got %q", refreshed)
+	}
+
+	userIDCacheMu.RLock()
+	entry := userIDCache[cacheKey]
+	userIDCacheMu.RUnlock()
+
+	if entry.expire.Sub(soon) < 30*time.Minute {
+		t.Fatalf("expected TTL to renew, got %v remaining", entry.expire.Sub(soon))
+	}
+}

internal/thinking/provider/openai/apply.go

@@ -10,53 +10,10 @@ import (
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
-	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
 
-// validReasoningEffortLevels contains the standard values accepted by the
-// OpenAI reasoning_effort field. Provider-specific extensions (xhigh, minimal,
-// auto) are NOT in this set and must be clamped before use.
-var validReasoningEffortLevels = map[string]struct{}{
-	"none":   {},
-	"low":    {},
-	"medium": {},
-	"high":   {},
-}
-
-// clampReasoningEffort maps any thinking level string to a value that is safe
-// to send as OpenAI reasoning_effort. Non-standard CPA-internal values are
-// mapped to the nearest standard equivalent.
-//
-// Mapping rules:
-//   - none / low / medium / high  → returned as-is (already valid)
-//   - xhigh                       → "high" (nearest lower standard level)
-//   - minimal                     → "low" (nearest higher standard level)
-//   - auto                        → "medium" (reasonable default)
-//   - anything else               → "medium" (safe default)
-func clampReasoningEffort(level string) string {
-	if _, ok := validReasoningEffortLevels[level]; ok {
-		return level
-	}
-	var clamped string
-	switch level {
-	case string(thinking.LevelXHigh):
-		clamped = string(thinking.LevelHigh)
-	case string(thinking.LevelMinimal):
-		clamped = string(thinking.LevelLow)
-	case string(thinking.LevelAuto):
-		clamped = string(thinking.LevelMedium)
-	default:
-		clamped = string(thinking.LevelMedium)
-	}
-	log.WithFields(log.Fields{
-		"original": level,
-		"clamped":  clamped,
-	}).Debug("openai: reasoning_effort clamped to nearest valid standard value")
-	return clamped
-}
-
 // Applier implements thinking.ProviderApplier for OpenAI models.
 //
 // OpenAI-specific behavior:
@@ -101,7 +58,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 	}
 
 	if config.Mode == thinking.ModeLevel {
-		result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(string(config.Level)))
+		result, _ := sjson.SetBytes(body, "reasoning_effort", string(config.Level))
 		return result, nil
 	}
 
@@ -122,7 +79,7 @@ func (a *Applier) Apply(body []byte, config thinking.ThinkingConfig, modelInfo *
 		return body, nil
 	}
 
-	result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(effort))
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
 	return result, nil
 }
 
@@ -157,7 +114,7 @@ func applyCompatibleOpenAI(body []byte, config thinking.ThinkingConfig) ([]byte,
 		return body, nil
 	}
 
-	result, _ := sjson.SetBytes(body, "reasoning_effort", clampReasoningEffort(effort))
+	result, _ := sjson.SetBytes(body, "reasoning_effort", effort)
 	return result, nil
 }
 

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -231,8 +231,12 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 							} else if functionResponseResult.IsObject() {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
-							} else {
+							} else if functionResponseResult.Raw != "" {
 								functionResponseJSON, _ = sjson.SetRaw(functionResponseJSON, "response.result", functionResponseResult.Raw)
+							} else {
+								// Content field is missing entirely — .Raw is empty which
+								// causes sjson.SetRaw to produce invalid JSON (e.g. "result":}).
+								functionResponseJSON, _ = sjson.Set(functionResponseJSON, "response.result", "")
 							}
 
 							partJSON := `{}`

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -661,6 +661,85 @@ func TestConvertClaudeRequestToAntigravity_ThinkingOnly_NoHint(t *testing.T) {
 	}
 }
 
+func TestConvertClaudeRequestToAntigravity_ToolResultNoContent(t *testing.T) {
+	// Bug repro: tool_result with no content field produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456"
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+
+	// Verify the functionResponse has a valid result value
+	fr := gjson.Get(outputStr, "request.contents.1.parts.0.functionResponse.response.result")
+	if !fr.Exists() {
+		t.Error("functionResponse.response.result should exist")
+	}
+}
+
+func TestConvertClaudeRequestToAntigravity_ToolResultNullContent(t *testing.T) {
+	// Bug repro: tool_result with null content produces invalid JSON
+	inputJSON := []byte(`{
+		"model": "claude-opus-4-6-thinking",
+		"messages": [
+			{
+				"role": "assistant",
+				"content": [
+					{
+						"type": "tool_use",
+						"id": "MyTool-123-456",
+						"name": "MyTool",
+						"input": {"key": "value"}
+					}
+				]
+			},
+			{
+				"role": "user",
+				"content": [
+					{
+						"type": "tool_result",
+						"tool_use_id": "MyTool-123-456",
+						"content": null
+					}
+				]
+			}
+		]
+	}`)
+
+	output := ConvertClaudeRequestToAntigravity("claude-opus-4-6-thinking", inputJSON, true)
+	outputStr := string(output)
+
+	if !gjson.Valid(outputStr) {
+		t.Errorf("Result is not valid JSON:\n%s", outputStr)
+	}
+}
+
 func TestConvertClaudeRequestToAntigravity_ToolAndThinking_NoExistingSystem(t *testing.T) {
 	// When tools + thinking but no system instruction, should create one with hint
 	inputJSON := []byte(`{

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

@@ -95,9 +95,9 @@ func ConvertAntigravityResponseToOpenAI(_ context.Context, _ string, originalReq
 		if totalTokenCountResult := usageResult.Get("totalTokenCount"); totalTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.total_tokens", totalTokenCountResult.Int())
 		}
-		promptTokenCount := usageResult.Get("promptTokenCount").Int() - cachedTokenCount
+		promptTokenCount := usageResult.Get("promptTokenCount").Int()
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
+		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}

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

@@ -199,6 +199,21 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 									msg, _ = sjson.SetRaw(msg, "content.-1", imagePart)
 								}
 							}
+
+						case "file":
+							fileData := part.Get("file.file_data").String()
+							if strings.HasPrefix(fileData, "data:") {
+								semicolonIdx := strings.Index(fileData, ";")
+								commaIdx := strings.Index(fileData, ",")
+								if semicolonIdx != -1 && commaIdx != -1 && commaIdx > semicolonIdx {
+									mediaType := strings.TrimPrefix(fileData[:semicolonIdx], "data:")
+									data := fileData[commaIdx+1:]
+									docPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
+									docPart, _ = sjson.Set(docPart, "source.media_type", mediaType)
+									docPart, _ = sjson.Set(docPart, "source.data", data)
+									msg, _ = sjson.SetRaw(msg, "content.-1", docPart)
+								}
+							}
 						}
 						return true
 					})

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

@@ -155,6 +155,7 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 				var textAggregate strings.Builder
 				var partsJSON []string
 				hasImage := false
+				hasFile := false
 				if parts := item.Get("content"); parts.Exists() && parts.IsArray() {
 					parts.ForEach(func(_, part gjson.Result) bool {
 						ptype := part.Get("type").String()
@@ -207,6 +208,30 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 									hasImage = true
 								}
 							}
+						case "input_file":
+							fileData := part.Get("file_data").String()
+							if fileData != "" {
+								mediaType := "application/octet-stream"
+								data := fileData
+								if strings.HasPrefix(fileData, "data:") {
+									trimmed := strings.TrimPrefix(fileData, "data:")
+									mediaAndData := strings.SplitN(trimmed, ";base64,", 2)
+									if len(mediaAndData) == 2 {
+										if mediaAndData[0] != "" {
+											mediaType = mediaAndData[0]
+										}
+										data = mediaAndData[1]
+									}
+								}
+								contentPart := `{"type":"document","source":{"type":"base64","media_type":"","data":""}}`
+								contentPart, _ = sjson.Set(contentPart, "source.media_type", mediaType)
+								contentPart, _ = sjson.Set(contentPart, "source.data", data)
+								partsJSON = append(partsJSON, contentPart)
+								if role == "" {
+									role = "user"
+								}
+								hasFile = true
+							}
 						}
 						return true
 					})
@@ -228,7 +253,7 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 				if len(partsJSON) > 0 {
 					msg := `{"role":"","content":[]}`
 					msg, _ = sjson.Set(msg, "role", role)
-					if len(partsJSON) == 1 && !hasImage {
+					if len(partsJSON) == 1 && !hasImage && !hasFile {
 						// Preserve legacy behavior for single text content
 						msg, _ = sjson.Delete(msg, "content")
 						textPart := gjson.Parse(partsJSON[0])

internal/translator/codex/claude/codex_claude_response.go

@@ -22,8 +22,9 @@ var (
 
 // ConvertCodexResponseToClaudeParams holds parameters for response conversion.
 type ConvertCodexResponseToClaudeParams struct {
-	HasToolCall bool
-	BlockIndex  int
+	HasToolCall              bool
+	BlockIndex               int
+	HasReceivedArgumentsDelta bool
 }
 
 // ConvertCodexResponseToClaude performs sophisticated streaming response format conversion.
@@ -137,6 +138,7 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 		itemType := itemResult.Get("type").String()
 		if itemType == "function_call" {
 			(*param).(*ConvertCodexResponseToClaudeParams).HasToolCall = true
+			(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = false
 			template = `{"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"","name":"","input":{}}}`
 			template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 			template, _ = sjson.Set(template, "content_block.id", itemResult.Get("call_id").String())
@@ -171,12 +173,29 @@ func ConvertCodexResponseToClaude(_ context.Context, _ string, originalRequestRa
 			output += fmt.Sprintf("data: %s\n\n", template)
 		}
 	} else if typeStr == "response.function_call_arguments.delta" {
+		(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta = true
 		template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
 		template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
 		template, _ = sjson.Set(template, "delta.partial_json", rootResult.Get("delta").String())
 
 		output += "event: content_block_delta\n"
 		output += fmt.Sprintf("data: %s\n\n", template)
+	} else if typeStr == "response.function_call_arguments.done" {
+		// Some models (e.g. gpt-5.3-codex-spark) send function call arguments
+		// in a single "done" event without preceding "delta" events.
+		// Emit the full arguments as a single input_json_delta so the
+		// downstream Claude client receives the complete tool input.
+		// When delta events were already received, skip to avoid duplicating arguments.
+		if !(*param).(*ConvertCodexResponseToClaudeParams).HasReceivedArgumentsDelta {
+			if args := rootResult.Get("arguments").String(); args != "" {
+				template = `{"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""}}`
+				template, _ = sjson.Set(template, "index", (*param).(*ConvertCodexResponseToClaudeParams).BlockIndex)
+				template, _ = sjson.Set(template, "delta.partial_json", args)
+
+				output += "event: content_block_delta\n"
+				output += fmt.Sprintf("data: %s\n\n", template)
+			}
+		}
 	}
 
 	return []string{output}

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

@@ -180,7 +180,19 @@ func ConvertOpenAIRequestToCodex(modelName string, inputRawJSON []byte, stream b
 								msg, _ = sjson.SetRaw(msg, "content.-1", part)
 							}
 						case "file":
-							// Files are not specified in examples; skip for now
+							if role == "user" {
+								fileData := it.Get("file.file_data").String()
+								filename := it.Get("file.filename").String()
+								if fileData != "" {
+									part := `{}`
+									part, _ = sjson.Set(part, "type", "input_file")
+									part, _ = sjson.Set(part, "file_data", fileData)
+									if filename != "" {
+										part, _ = sjson.Set(part, "filename", filename)
+									}
+									msg, _ = sjson.SetRaw(msg, "content.-1", part)
+								}
+							}
 						}
 					}
 				}

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

@@ -26,6 +26,8 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "temperature")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
+	rawJSON, _ = sjson.DeleteBytes(rawJSON, "truncation")
+	rawJSON = applyResponsesCompactionCompatibility(rawJSON)
 
 	// Delete the user field as it is not supported by the Codex upstream.
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "user")
@@ -36,6 +38,23 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	return rawJSON
 }
 
+// applyResponsesCompactionCompatibility handles OpenAI Responses context_management.compaction
+// for Codex upstream compatibility.
+//
+// Codex /responses currently rejects context_management with:
+// {"detail":"Unsupported parameter: context_management"}.
+//
+// Compatibility strategy:
+// 1) Remove context_management before forwarding to Codex upstream.
+func applyResponsesCompactionCompatibility(rawJSON []byte) []byte {
+	if !gjson.GetBytes(rawJSON, "context_management").Exists() {
+		return rawJSON
+	}
+
+	rawJSON, _ = sjson.DeleteBytes(rawJSON, "context_management")
+	return rawJSON
+}
+
 // convertSystemRoleToDeveloper traverses the input array and converts any message items
 // with role "system" to role "developer". This is necessary because Codex API does not
 // accept "system" role in the input array.

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

@@ -280,3 +280,41 @@ func TestUserFieldDeletion(t *testing.T) {
 		t.Errorf("user field should be deleted, but it was found with value: %s", userField.Raw)
 	}
 }
+
+func TestContextManagementCompactionCompatibility(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gpt-5.2",
+		"context_management": [
+			{
+				"type": "compaction",
+				"compact_threshold": 12000
+			}
+		],
+		"input": [{"role":"user","content":"hello"}]
+	}`)
+
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)
+	outputStr := string(output)
+
+	if gjson.Get(outputStr, "context_management").Exists() {
+		t.Fatalf("context_management should be removed for Codex compatibility")
+	}
+	if gjson.Get(outputStr, "truncation").Exists() {
+		t.Fatalf("truncation should be removed for Codex compatibility")
+	}
+}
+
+func TestTruncationRemovedForCodexCompatibility(t *testing.T) {
+	inputJSON := []byte(`{
+		"model": "gpt-5.2",
+		"truncation": "disabled",
+		"input": [{"role":"user","content":"hello"}]
+	}`)
+
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)
+	outputStr := string(output)
+
+	if gjson.Get(outputStr, "truncation").Exists() {
+		t.Fatalf("truncation should be removed for Codex compatibility")
+	}
+}

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

@@ -100,7 +100,7 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 		}
 		promptTokenCount := usageResult.Get("promptTokenCount").Int()
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
+		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}

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

@@ -100,9 +100,9 @@ func ConvertGeminiResponseToOpenAI(_ context.Context, _ string, originalRequestR
 		if totalTokenCountResult := usageResult.Get("totalTokenCount"); totalTokenCountResult.Exists() {
 			baseTemplate, _ = sjson.Set(baseTemplate, "usage.total_tokens", totalTokenCountResult.Int())
 		}
-		promptTokenCount := usageResult.Get("promptTokenCount").Int() - cachedTokenCount
+		promptTokenCount := usageResult.Get("promptTokenCount").Int()
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
-		baseTemplate, _ = sjson.Set(baseTemplate, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
+		baseTemplate, _ = sjson.Set(baseTemplate, "usage.prompt_tokens", promptTokenCount)
 		if thoughtsTokenCount > 0 {
 			baseTemplate, _ = sjson.Set(baseTemplate, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
@@ -297,7 +297,7 @@ func ConvertGeminiResponseToOpenAINonStream(_ context.Context, _ string, origina
 		promptTokenCount := usageResult.Get("promptTokenCount").Int()
 		thoughtsTokenCount := usageResult.Get("thoughtsTokenCount").Int()
 		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
-		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount+thoughtsTokenCount)
+		template, _ = sjson.Set(template, "usage.prompt_tokens", promptTokenCount)
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}

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

@@ -531,8 +531,8 @@ func ConvertGeminiResponseToOpenAIResponses(_ context.Context, modelName string,
 
 		// usage mapping
 		if um := root.Get("usageMetadata"); um.Exists() {
-			// input tokens = prompt + thoughts
-			input := um.Get("promptTokenCount").Int() + um.Get("thoughtsTokenCount").Int()
+			// input tokens = prompt only (thoughts go to output)
+			input := um.Get("promptTokenCount").Int()
 			completed, _ = sjson.Set(completed, "response.usage.input_tokens", input)
 			// cached token details: align with OpenAI "cached_tokens" semantics.
 			completed, _ = sjson.Set(completed, "response.usage.input_tokens_details.cached_tokens", um.Get("cachedContentTokenCount").Int())
@@ -737,8 +737,8 @@ func ConvertGeminiResponseToOpenAIResponsesNonStream(_ context.Context, _ string
 
 	// usage mapping
 	if um := root.Get("usageMetadata"); um.Exists() {
-		// input tokens = prompt + thoughts
-		input := um.Get("promptTokenCount").Int() + um.Get("thoughtsTokenCount").Int()
+		// input tokens = prompt only (thoughts go to output)
+		input := um.Get("promptTokenCount").Int()
 		resp, _ = sjson.Set(resp, "usage.input_tokens", input)
 		// cached token details: align with OpenAI "cached_tokens" semantics.
 		resp, _ = sjson.Set(resp, "usage.input_tokens_details.cached_tokens", um.Get("cachedContentTokenCount").Int())

sdk/api/handlers/claude/code_handlers.go

@@ -112,12 +112,13 @@ func (h *ClaudeCodeAPIHandler) ClaudeCountTokens(c *gin.Context) {
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 
-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -165,7 +166,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
@@ -194,6 +195,7 @@ func (h *ClaudeCodeAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSO
 		}
 	}
 
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -225,7 +227,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 	// This allows proper cleanup and cancellation of ongoing requests
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
 		c.Header("Cache-Control", "no-cache")
@@ -257,6 +259,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -264,6 +267,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 
 			// Success! Set headers now.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write the first chunk
 			if len(chunk) > 0 {

sdk/api/handlers/gemini/gemini-cli_handlers.go

@@ -159,7 +159,8 @@ func (h *GeminiCLIAPIHandler) handleInternalStreamGenerateContent(c *gin.Context
 	modelName := modelResult.String()
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	h.forwardCLIStream(c, flusher, "", func(err error) { cliCancel(err) }, dataChan, errChan)
 	return
 }
@@ -172,12 +173,13 @@ func (h *GeminiCLIAPIHandler) handleInternalGenerateContent(c *gin.Context, rawJ
 	modelName := modelResult.String()
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }

sdk/api/handlers/gemini/gemini_handlers.go

@@ -188,7 +188,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 	}
 
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -223,6 +223,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 				if alt == "" {
 					setSSEHeaders()
 				}
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				flusher.Flush()
 				cliCancel(nil)
 				return
@@ -232,6 +233,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 			if alt == "" {
 				setSSEHeaders()
 			}
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write first chunk
 			if alt == "" {
@@ -262,12 +264,13 @@ func (h *GeminiAPIHandler) handleCountTokens(c *gin.Context, modelName string, r
 	c.Header("Content-Type", "application/json")
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteCountWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -286,13 +289,14 @@ func (h *GeminiAPIHandler) handleGenerateContent(c *gin.Context, modelName strin
 	alt := h.GetAlt(c)
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, alt)
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }

sdk/api/handlers/handlers.go

@@ -179,6 +179,12 @@ func StreamingBootstrapRetries(cfg *config.SDKConfig) int {
 	return retries
 }
 
+// PassthroughHeadersEnabled returns whether upstream response headers should be forwarded to clients.
+// Default is false.
+func PassthroughHeadersEnabled(cfg *config.SDKConfig) bool {
+	return cfg != nil && cfg.PassthroughHeaders
+}
+
 func requestExecutionMetadata(ctx context.Context) map[string]any {
 	// Idempotency-Key is an optional client-supplied header used to correlate retries.
 	// It is forwarded as execution metadata; when absent we generate a UUID.
@@ -461,10 +467,10 @@ func appendAPIResponse(c *gin.Context, data []byte) {
 
 // ExecuteWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -497,17 +503,20 @@ func (h *BaseAPIHandler) ExecuteWithAuthManager(ctx context.Context, handlerType
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }
 
 // ExecuteCountWithAuthManager executes a non-streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, *interfaces.ErrorMessage) {
+func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) ([]byte, http.Header, *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
-		return nil, errMsg
+		return nil, nil, errMsg
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -540,20 +549,24 @@ func (h *BaseAPIHandler) ExecuteCountWithAuthManager(ctx context.Context, handle
 				addon = hdr.Clone()
 			}
 		}
-		return nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
+		return nil, nil, &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 	}
-	return resp.Payload, nil
+	if !PassthroughHeadersEnabled(h.Cfg) {
+		return resp.Payload, nil, nil
+	}
+	return resp.Payload, FilterUpstreamHeaders(resp.Headers), nil
 }
 
 // ExecuteStreamWithAuthManager executes a streaming request via the core auth manager.
 // This path is the only supported execution route.
-func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, <-chan *interfaces.ErrorMessage) {
+// The returned http.Header carries upstream response headers captured before streaming begins.
+func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handlerType, modelName string, rawJSON []byte, alt string) (<-chan []byte, http.Header, <-chan *interfaces.ErrorMessage) {
 	providers, normalizedModel, errMsg := h.getRequestDetails(modelName)
 	if errMsg != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		errChan <- errMsg
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
 	reqMeta := requestExecutionMetadata(ctx)
 	reqMeta[coreexecutor.RequestedModelMetadataKey] = normalizedModel
@@ -572,7 +585,7 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		SourceFormat:    sdktranslator.FromString(handlerType),
 	}
 	opts.Metadata = reqMeta
-	chunks, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+	streamResult, err := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 	if err != nil {
 		errChan := make(chan *interfaces.ErrorMessage, 1)
 		status := http.StatusInternalServerError
@@ -589,8 +602,19 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 		}
 		errChan <- &interfaces.ErrorMessage{StatusCode: status, Error: err, Addon: addon}
 		close(errChan)
-		return nil, errChan
+		return nil, nil, errChan
 	}
+	passthroughHeadersEnabled := PassthroughHeadersEnabled(h.Cfg)
+	// Capture upstream headers from the initial connection synchronously before the goroutine starts.
+	// Keep a mutable map so bootstrap retries can replace it before first payload is sent.
+	var upstreamHeaders http.Header
+	if passthroughHeadersEnabled {
+		upstreamHeaders = cloneHeader(FilterUpstreamHeaders(streamResult.Headers))
+		if upstreamHeaders == nil {
+			upstreamHeaders = make(http.Header)
+		}
+	}
+	chunks := streamResult.Chunks
 	dataChan := make(chan []byte)
 	errChan := make(chan *interfaces.ErrorMessage, 1)
 	go func() {
@@ -664,9 +688,12 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 					if !sentPayload {
 						if bootstrapRetries < maxBootstrapRetries && bootstrapEligible(streamErr) {
 							bootstrapRetries++
-							retryChunks, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
+							retryResult, retryErr := h.AuthManager.ExecuteStream(ctx, providers, req, opts)
 							if retryErr == nil {
-								chunks = retryChunks
+								if passthroughHeadersEnabled {
+									replaceHeader(upstreamHeaders, FilterUpstreamHeaders(retryResult.Headers))
+								}
+								chunks = retryResult.Chunks
 								continue outer
 							}
 							streamErr = retryErr
@@ -689,6 +716,12 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 					return
 				}
 				if len(chunk.Payload) > 0 {
+					if handlerType == "openai-response" {
+						if err := validateSSEDataJSON(chunk.Payload); err != nil {
+							_ = sendErr(&interfaces.ErrorMessage{StatusCode: http.StatusBadGateway, Error: err})
+							return
+						}
+					}
 					sentPayload = true
 					if okSendData := sendData(cloneBytes(chunk.Payload)); !okSendData {
 						return
@@ -697,7 +730,36 @@ func (h *BaseAPIHandler) ExecuteStreamWithAuthManager(ctx context.Context, handl
 			}
 		}
 	}()
-	return dataChan, errChan
+	return dataChan, upstreamHeaders, errChan
+}
+
+func validateSSEDataJSON(chunk []byte) error {
+	for _, line := range bytes.Split(chunk, []byte("\n")) {
+		line = bytes.TrimSpace(line)
+		if len(line) == 0 {
+			continue
+		}
+		if !bytes.HasPrefix(line, []byte("data:")) {
+			continue
+		}
+		data := bytes.TrimSpace(line[5:])
+		if len(data) == 0 {
+			continue
+		}
+		if bytes.Equal(data, []byte("[DONE]")) {
+			continue
+		}
+		if json.Valid(data) {
+			continue
+		}
+		const max = 512
+		preview := data
+		if len(preview) > max {
+			preview = preview[:max]
+		}
+		return fmt.Errorf("invalid SSE data JSON (len=%d): %q", len(data), preview)
+	}
+	return nil
 }
 
 func statusFromError(err error) int {
@@ -757,13 +819,33 @@ func cloneBytes(src []byte) []byte {
 	return dst
 }
 
+func cloneHeader(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	dst := make(http.Header, len(src))
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+	return dst
+}
+
+func replaceHeader(dst http.Header, src http.Header) {
+	for key := range dst {
+		delete(dst, key)
+	}
+	for key, values := range src {
+		dst[key] = append([]string(nil), values...)
+	}
+}
+
 // 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
 	if msg != nil && msg.StatusCode > 0 {
 		status = msg.StatusCode
 	}
-	if msg != nil && msg.Addon != nil {
+	if msg != nil && msg.Addon != nil && PassthroughHeadersEnabled(h.Cfg) {
 		for key, values := range msg.Addon {
 			if len(values) == 0 {
 				continue

sdk/api/handlers/handlers_error_response_test.go

@@ -0,0 +1,68 @@
+package handlers
+
+import (
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestWriteErrorResponse_AddonHeadersDisabledByDefault(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+
+	handler := NewBaseAPIHandlers(nil, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"req-1"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "" {
+		t.Fatalf("Retry-After should be empty when passthrough is disabled, got %q", got)
+	}
+	if got := recorder.Header().Get("X-Request-Id"); got != "" {
+		t.Fatalf("X-Request-Id should be empty when passthrough is disabled, got %q", got)
+	}
+}
+
+func TestWriteErrorResponse_AddonHeadersEnabled(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodGet, "/", nil)
+	c.Writer.Header().Set("X-Request-Id", "old-value")
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{PassthroughHeaders: true}, nil)
+	handler.WriteErrorResponse(c, &interfaces.ErrorMessage{
+		StatusCode: http.StatusTooManyRequests,
+		Error:      errors.New("rate limit"),
+		Addon: http.Header{
+			"Retry-After":  {"30"},
+			"X-Request-Id": {"new-1", "new-2"},
+		},
+	})
+
+	if recorder.Code != http.StatusTooManyRequests {
+		t.Fatalf("status = %d, want %d", recorder.Code, http.StatusTooManyRequests)
+	}
+	if got := recorder.Header().Get("Retry-After"); got != "30" {
+		t.Fatalf("Retry-After = %q, want %q", got, "30")
+	}
+	if got := recorder.Header().Values("X-Request-Id"); !reflect.DeepEqual(got, []string{"new-1", "new-2"}) {
+		t.Fatalf("X-Request-Id = %#v, want %#v", got, []string{"new-1", "new-2"})
+	}
+}

sdk/api/handlers/handlers_stream_bootstrap_test.go

@@ -23,7 +23,7 @@ func (e *failOnceStreamExecutor) Execute(context.Context, *coreauth.Auth, coreex
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	call := e.calls
@@ -40,12 +40,18 @@ func (e *failOnceStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth,
 			},
 		}
 		close(ch)
-		return ch, nil
+		return &coreexecutor.StreamResult{
+			Headers: http.Header{"X-Upstream-Attempt": {"1"}},
+			Chunks:  ch,
+		}, nil
 	}
 
 	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{
+		Headers: http.Header{"X-Upstream-Attempt": {"2"}},
+		Chunks:  ch,
+	}, nil
 }
 
 func (e *failOnceStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -81,7 +87,7 @@ func (e *payloadThenErrorStreamExecutor) Execute(context.Context, *coreauth.Auth
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	e.mu.Lock()
 	e.calls++
 	e.mu.Unlock()
@@ -97,7 +103,7 @@ func (e *payloadThenErrorStreamExecutor) ExecuteStream(context.Context, *coreaut
 		},
 	}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *payloadThenErrorStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -128,13 +134,44 @@ type authAwareStreamExecutor struct {
 	authIDs []string
 }
 
+type invalidJSONStreamExecutor struct{}
+
+func (e *invalidJSONStreamExecutor) Identifier() string { return "codex" }
+
+func (e *invalidJSONStreamExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
+}
+
+func (e *invalidJSONStreamExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
+	ch := make(chan coreexecutor.StreamChunk, 1)
+	ch <- coreexecutor.StreamChunk{Payload: []byte("event: response.completed\ndata: {\"type\"")}
+	close(ch)
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
+}
+
+func (e *invalidJSONStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *invalidJSONStreamExecutor) CountTokens(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
+	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "CountTokens not implemented"}
+}
+
+func (e *invalidJSONStreamExecutor) HttpRequest(ctx context.Context, auth *coreauth.Auth, req *http.Request) (*http.Response, error) {
+	return nil, &coreauth.Error{
+		Code:       "not_implemented",
+		Message:    "HttpRequest not implemented",
+		HTTPStatus: http.StatusNotImplemented,
+	}
+}
+
 func (e *authAwareStreamExecutor) Identifier() string { return "codex" }
 
 func (e *authAwareStreamExecutor) Execute(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (coreexecutor.Response, error) {
 	return coreexecutor.Response{}, &coreauth.Error{Code: "not_implemented", Message: "Execute not implemented"}
 }
 
-func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *coreauth.Auth, req coreexecutor.Request, opts coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	_ = ctx
 	_ = req
 	_ = opts
@@ -160,12 +197,12 @@ func (e *authAwareStreamExecutor) ExecuteStream(ctx context.Context, auth *corea
 			},
 		}
 		close(ch)
-		return ch, nil
+		return &coreexecutor.StreamResult{Chunks: ch}, nil
 	}
 
 	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
 	close(ch)
-	return ch, nil
+	return &coreexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *authAwareStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
@@ -231,11 +268,12 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	})
 
 	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		PassthroughHeaders: true,
 		Streaming: sdkconfig.StreamingConfig{
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -257,6 +295,70 @@ func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	if executor.Calls() != 2 {
 		t.Fatalf("expected 2 stream attempts, got %d", executor.Calls())
 	}
+	upstreamAttemptHeader := upstreamHeaders.Get("X-Upstream-Attempt")
+	if upstreamAttemptHeader != "2" {
+		t.Fatalf("expected upstream header from retry attempt, got %q", upstreamAttemptHeader)
+	}
+}
+
+func TestExecuteStreamWithAuthManager_HeaderPassthroughDisabledByDefault(t *testing.T) {
+	executor := &failOnceStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	auth2 := &coreauth.Auth{
+		ID:       "auth2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test2@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth2); err != nil {
+		t.Fatalf("manager.Register(auth2): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	registry.GetGlobalRegistry().RegisterClient(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 1,
+		},
+	}, manager)
+	dataChan, upstreamHeaders, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+	for msg := range errChan {
+		if msg != nil {
+			t.Fatalf("unexpected error: %+v", msg)
+		}
+	}
+
+	if string(got) != "ok" {
+		t.Fatalf("expected payload ok, got %q", string(got))
+	}
+	if upstreamHeaders != nil {
+		t.Fatalf("expected nil upstream headers when passthrough is disabled, got %#v", upstreamHeaders)
+	}
 }
 
 func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
@@ -296,7 +398,7 @@ func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
 			BootstrapRetries: 1,
 		},
 	}, manager)
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -367,7 +469,7 @@ func TestExecuteStreamWithAuthManager_PinnedAuthKeepsSameUpstream(t *testing.T)
 		},
 	}, manager)
 	ctx := WithPinnedAuthID(context.Background(), "auth1")
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -431,7 +533,7 @@ func TestExecuteStreamWithAuthManager_SelectedAuthCallbackReceivesAuthID(t *test
 	ctx := WithSelectedAuthIDCallback(context.Background(), func(authID string) {
 		selectedAuthID = authID
 	})
-	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(ctx, "openai", "test-model", []byte(`{"model":"test-model"}`), "")
 	if dataChan == nil || errChan == nil {
 		t.Fatalf("expected non-nil channels")
 	}
@@ -453,3 +555,55 @@ func TestExecuteStreamWithAuthManager_SelectedAuthCallbackReceivesAuthID(t *test
 		t.Fatalf("selectedAuthID = %q, want %q", selectedAuthID, "auth2")
 	}
 }
+
+func TestExecuteStreamWithAuthManager_ValidatesOpenAIResponsesStreamDataJSON(t *testing.T) {
+	executor := &invalidJSONStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	auth1 := &coreauth.Auth{
+		ID:       "auth1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+		Metadata: map[string]any{"email": "test1@example.com"},
+	}
+	if _, err := manager.Register(context.Background(), auth1); err != nil {
+		t.Fatalf("manager.Register(auth1): %v", err)
+	}
+
+	registry.GetGlobalRegistry().RegisterClient(auth1.ID, auth1.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth1.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, manager)
+	dataChan, _, errChan := handler.ExecuteStreamWithAuthManager(context.Background(), "openai-response", "test-model", []byte(`{"model":"test-model"}`), "")
+	if dataChan == nil || errChan == nil {
+		t.Fatalf("expected non-nil channels")
+	}
+
+	var got []byte
+	for chunk := range dataChan {
+		got = append(got, chunk...)
+	}
+	if len(got) != 0 {
+		t.Fatalf("expected empty payload, got %q", string(got))
+	}
+
+	gotErr := false
+	for msg := range errChan {
+		if msg == nil {
+			continue
+		}
+		if msg.StatusCode != http.StatusBadGateway {
+			t.Fatalf("expected status %d, got %d", http.StatusBadGateway, msg.StatusCode)
+		}
+		if msg.Error == nil {
+			t.Fatalf("expected error")
+		}
+		gotErr = true
+	}
+	if !gotErr {
+		t.Fatalf("expected terminal error")
+	}
+}

sdk/api/handlers/header_filter.go

@@ -0,0 +1,80 @@
+package handlers
+
+import (
+	"net/http"
+	"strings"
+)
+
+// hopByHopHeaders lists RFC 7230 Section 6.1 hop-by-hop headers that MUST NOT
+// be forwarded by proxies, plus security-sensitive headers that should not leak.
+var hopByHopHeaders = map[string]struct{}{
+	// RFC 7230 hop-by-hop
+	"Connection":          {},
+	"Keep-Alive":          {},
+	"Proxy-Authenticate":  {},
+	"Proxy-Authorization": {},
+	"Te":                  {},
+	"Trailer":             {},
+	"Transfer-Encoding":   {},
+	"Upgrade":             {},
+	// Security-sensitive
+	"Set-Cookie": {},
+	// CPA-managed (set by handlers, not upstream)
+	"Content-Length":   {},
+	"Content-Encoding": {},
+}
+
+// FilterUpstreamHeaders returns a copy of src with hop-by-hop and security-sensitive
+// headers removed. Returns nil if src is nil or empty after filtering.
+func FilterUpstreamHeaders(src http.Header) http.Header {
+	if src == nil {
+		return nil
+	}
+	connectionScoped := connectionScopedHeaders(src)
+	dst := make(http.Header)
+	for key, values := range src {
+		canonicalKey := http.CanonicalHeaderKey(key)
+		if _, blocked := hopByHopHeaders[canonicalKey]; blocked {
+			continue
+		}
+		if _, scoped := connectionScoped[canonicalKey]; scoped {
+			continue
+		}
+		dst[key] = values
+	}
+	if len(dst) == 0 {
+		return nil
+	}
+	return dst
+}
+
+func connectionScopedHeaders(src http.Header) map[string]struct{} {
+	scoped := make(map[string]struct{})
+	for _, rawValue := range src.Values("Connection") {
+		for _, token := range strings.Split(rawValue, ",") {
+			headerName := strings.TrimSpace(token)
+			if headerName == "" {
+				continue
+			}
+			scoped[http.CanonicalHeaderKey(headerName)] = struct{}{}
+		}
+	}
+	return scoped
+}
+
+// WriteUpstreamHeaders writes filtered upstream headers to the gin response writer.
+// Headers already set by CPA (e.g., Content-Type) are NOT overwritten.
+func WriteUpstreamHeaders(dst http.Header, src http.Header) {
+	if src == nil {
+		return
+	}
+	for key, values := range src {
+		// Don't overwrite headers already set by CPA handlers
+		if dst.Get(key) != "" {
+			continue
+		}
+		for _, v := range values {
+			dst.Add(key, v)
+		}
+	}
+}

sdk/api/handlers/header_filter_test.go

@@ -0,0 +1,55 @@
+package handlers
+
+import (
+	"net/http"
+	"testing"
+)
+
+func TestFilterUpstreamHeaders_RemovesConnectionScopedHeaders(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "keep-alive, x-hop-a, x-hop-b")
+	src.Add("Connection", "x-hop-c")
+	src.Set("Keep-Alive", "timeout=5")
+	src.Set("X-Hop-A", "a")
+	src.Set("X-Hop-B", "b")
+	src.Set("X-Hop-C", "c")
+	src.Set("X-Request-Id", "req-1")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered == nil {
+		t.Fatalf("expected filtered headers, got nil")
+	}
+
+	requestID := filtered.Get("X-Request-Id")
+	if requestID != "req-1" {
+		t.Fatalf("expected X-Request-Id to be preserved, got %q", requestID)
+	}
+
+	blockedHeaderKeys := []string{
+		"Connection",
+		"Keep-Alive",
+		"X-Hop-A",
+		"X-Hop-B",
+		"X-Hop-C",
+		"Set-Cookie",
+	}
+	for _, key := range blockedHeaderKeys {
+		value := filtered.Get(key)
+		if value != "" {
+			t.Fatalf("expected %s to be removed, got %q", key, value)
+		}
+	}
+}
+
+func TestFilterUpstreamHeaders_ReturnsNilWhenAllHeadersBlocked(t *testing.T) {
+	src := http.Header{}
+	src.Add("Connection", "x-hop-a")
+	src.Set("X-Hop-A", "a")
+	src.Set("Set-Cookie", "session=secret")
+
+	filtered := FilterUpstreamHeaders(src)
+	if filtered != nil {
+		t.Fatalf("expected nil when all headers are filtered, got %#v", filtered)
+	}
+}

sdk/api/handlers/openai/openai_handlers.go

@@ -431,12 +431,13 @@ func (h *OpenAIAPIHandler) handleNonStreamingResponse(c *gin.Context, rawJSON []
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -463,7 +464,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, h.GetAlt(c))
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -496,6 +497,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 			if !ok {
 				// Stream closed without data? Send DONE or just headers.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -504,6 +506,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 
 			// Success! Commit to streaming headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			_, _ = fmt.Fprintf(c.Writer, "data: %s\n\n", string(chunk))
 			flusher.Flush()
@@ -531,13 +534,14 @@ func (h *OpenAIAPIHandler) handleCompletionsNonStreamingResponse(c *gin.Context,
 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	completionsResp := convertChatCompletionsResponseToCompletions(resp)
 	_, _ = c.Writer.Write(completionsResp)
 	cliCancel()
@@ -568,7 +572,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 
 	modelName := gjson.GetBytes(chatCompletionsJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, chatCompletionsJSON, "")
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -599,6 +603,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 		case chunk, ok := <-dataChan:
 			if !ok {
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = fmt.Fprintf(c.Writer, "data: [DONE]\n\n")
 				flusher.Flush()
 				cliCancel(nil)
@@ -607,6 +612,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 
 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write the first chunk
 			converted := convertChatCompletionsStreamChunkToCompletions(chunk)

sdk/api/handlers/openai/openai_responses_compact_test.go

@@ -31,7 +31,7 @@ func (e *compactCaptureExecutor) Execute(ctx context.Context, auth *coreauth.Aut
 	return coreexecutor.Response{Payload: []byte(`{"ok":true}`)}, nil
 }
 
-func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (<-chan coreexecutor.StreamChunk, error) {
+func (e *compactCaptureExecutor) ExecuteStream(context.Context, *coreauth.Auth, coreexecutor.Request, coreexecutor.Options) (*coreexecutor.StreamResult, error) {
 	return nil, errors.New("not implemented")
 }
 

sdk/api/handlers/openai/openai_responses_handlers.go

@@ -124,13 +124,14 @@ func (h *OpenAIResponsesAPIHandler) Compact(c *gin.Context) {
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "responses/compact")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -149,13 +150,14 @@ func (h *OpenAIResponsesAPIHandler) handleNonStreamingResponse(c *gin.Context, r
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
 	stopKeepAlive := h.StartNonStreamingKeepAlive(c, cliCtx)
 
-	resp, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	resp, upstreamHeaders, errMsg := h.ExecuteWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 	stopKeepAlive()
 	if errMsg != nil {
 		h.WriteErrorResponse(c, errMsg)
 		cliCancel(errMsg.Error)
 		return
 	}
+	handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 	_, _ = c.Writer.Write(resp)
 	cliCancel()
 }
@@ -183,7 +185,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 	// New core execution path
 	modelName := gjson.GetBytes(rawJSON, "model").String()
 	cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
-	dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
+	dataChan, upstreamHeaders, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, rawJSON, "")
 
 	setSSEHeaders := func() {
 		c.Header("Content-Type", "text/event-stream")
@@ -216,6 +218,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 			if !ok {
 				// Stream closed without data? Send headers and done.
 				setSSEHeaders()
+				handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 				_, _ = c.Writer.Write([]byte("\n"))
 				flusher.Flush()
 				cliCancel(nil)
@@ -224,6 +227,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 
 			// Success! Set headers.
 			setSSEHeaders()
+			handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 
 			// Write first chunk logic (matching forwardResponsesStream)
 			if bytes.HasPrefix(chunk, []byte("event:")) {
@@ -261,8 +265,8 @@ func (h *OpenAIResponsesAPIHandler) forwardResponsesStream(c *gin.Context, flush
 			if errMsg.Error != nil && errMsg.Error.Error() != "" {
 				errText = errMsg.Error.Error()
 			}
-			body := handlers.BuildErrorResponseBody(status, errText)
-			_, _ = fmt.Fprintf(c.Writer, "\nevent: error\ndata: %s\n\n", string(body))
+			chunk := handlers.BuildOpenAIResponsesStreamErrorChunk(status, errText, 0)
+			_, _ = fmt.Fprintf(c.Writer, "\nevent: error\ndata: %s\n\n", string(chunk))
 		},
 		WriteDone: func() {
 			_, _ = c.Writer.Write([]byte("\n"))

sdk/api/handlers/openai/openai_responses_handlers_stream_error_test.go

@@ -0,0 +1,43 @@
+package openai
+
+import (
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	sdkconfig "github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestForwardResponsesStreamTerminalErrorUsesResponsesErrorChunk(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	base := handlers.NewBaseAPIHandlers(&sdkconfig.SDKConfig{}, nil)
+	h := NewOpenAIResponsesAPIHandler(base)
+
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+	c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+
+	flusher, ok := c.Writer.(http.Flusher)
+	if !ok {
+		t.Fatalf("expected gin writer to implement http.Flusher")
+	}
+
+	data := make(chan []byte)
+	errs := make(chan *interfaces.ErrorMessage, 1)
+	errs <- &interfaces.ErrorMessage{StatusCode: http.StatusInternalServerError, Error: errors.New("unexpected EOF")}
+	close(errs)
+
+	h.forwardResponsesStream(c, flusher, func(error) {}, data, errs)
+	body := recorder.Body.String()
+	if !strings.Contains(body, `"type":"error"`) {
+		t.Fatalf("expected responses error chunk, got: %q", body)
+	}
+	if strings.Contains(body, `"error":{`) {
+		t.Fatalf("expected streaming error chunk (top-level type), got HTTP error body: %q", body)
+	}
+}

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -153,7 +153,7 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 				pinnedAuthID = strings.TrimSpace(authID)
 			})
 		}
-		dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
+		dataChan, _, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
 
 		completedOutput, errForward := h.forwardResponsesWebsocket(c, conn, cliCancel, dataChan, errChan, &wsBodyLog, passthroughSessionID)
 		if errForward != nil {

sdk/api/handlers/openai_responses_stream_error.go

@@ -0,0 +1,119 @@
+package handlers
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+type openAIResponsesStreamErrorChunk struct {
+	Type           string `json:"type"`
+	Code           string `json:"code"`
+	Message        string `json:"message"`
+	SequenceNumber int    `json:"sequence_number"`
+}
+
+func openAIResponsesStreamErrorCode(status int) string {
+	switch status {
+	case http.StatusUnauthorized:
+		return "invalid_api_key"
+	case http.StatusForbidden:
+		return "insufficient_quota"
+	case http.StatusTooManyRequests:
+		return "rate_limit_exceeded"
+	case http.StatusNotFound:
+		return "model_not_found"
+	case http.StatusRequestTimeout:
+		return "request_timeout"
+	default:
+		if status >= http.StatusInternalServerError {
+			return "internal_server_error"
+		}
+		if status >= http.StatusBadRequest {
+			return "invalid_request_error"
+		}
+		return "unknown_error"
+	}
+}
+
+// BuildOpenAIResponsesStreamErrorChunk builds an OpenAI Responses streaming error chunk.
+//
+// Important: OpenAI's HTTP error bodies are shaped like {"error":{...}}; those are valid for
+// non-streaming responses, but streaming clients validate SSE `data:` payloads against a union
+// of chunks that requires a top-level `type` field.
+func BuildOpenAIResponsesStreamErrorChunk(status int, errText string, sequenceNumber int) []byte {
+	if status <= 0 {
+		status = http.StatusInternalServerError
+	}
+	if sequenceNumber < 0 {
+		sequenceNumber = 0
+	}
+
+	message := strings.TrimSpace(errText)
+	if message == "" {
+		message = http.StatusText(status)
+	}
+
+	code := openAIResponsesStreamErrorCode(status)
+
+	trimmed := strings.TrimSpace(errText)
+	if trimmed != "" && json.Valid([]byte(trimmed)) {
+		var payload map[string]any
+		if err := json.Unmarshal([]byte(trimmed), &payload); err == nil {
+			if t, ok := payload["type"].(string); ok && strings.TrimSpace(t) == "error" {
+				if m, ok := payload["message"].(string); ok && strings.TrimSpace(m) != "" {
+					message = strings.TrimSpace(m)
+				}
+				if v, ok := payload["code"]; ok && v != nil {
+					if c, ok := v.(string); ok && strings.TrimSpace(c) != "" {
+						code = strings.TrimSpace(c)
+					} else {
+						code = strings.TrimSpace(fmt.Sprint(v))
+					}
+				}
+				if v, ok := payload["sequence_number"].(float64); ok && sequenceNumber == 0 {
+					sequenceNumber = int(v)
+				}
+			}
+			if e, ok := payload["error"].(map[string]any); ok {
+				if m, ok := e["message"].(string); ok && strings.TrimSpace(m) != "" {
+					message = strings.TrimSpace(m)
+				}
+				if v, ok := e["code"]; ok && v != nil {
+					if c, ok := v.(string); ok && strings.TrimSpace(c) != "" {
+						code = strings.TrimSpace(c)
+					} else {
+						code = strings.TrimSpace(fmt.Sprint(v))
+					}
+				}
+			}
+		}
+	}
+
+	if strings.TrimSpace(code) == "" {
+		code = "unknown_error"
+	}
+
+	data, err := json.Marshal(openAIResponsesStreamErrorChunk{
+		Type:           "error",
+		Code:           code,
+		Message:        message,
+		SequenceNumber: sequenceNumber,
+	})
+	if err == nil {
+		return data
+	}
+
+	// Extremely defensive fallback.
+	data, _ = json.Marshal(openAIResponsesStreamErrorChunk{
+		Type:           "error",
+		Code:           "internal_server_error",
+		Message:        message,
+		SequenceNumber: sequenceNumber,
+	})
+	if len(data) > 0 {
+		return data
+	}
+	return []byte(`{"type":"error","code":"internal_server_error","message":"internal error","sequence_number":0}`)
+}

sdk/api/handlers/openai_responses_stream_error_test.go

@@ -0,0 +1,48 @@
+package handlers
+
+import (
+	"encoding/json"
+	"net/http"
+	"testing"
+)
+
+func TestBuildOpenAIResponsesStreamErrorChunk(t *testing.T) {
+	chunk := BuildOpenAIResponsesStreamErrorChunk(http.StatusInternalServerError, "unexpected EOF", 0)
+	var payload map[string]any
+	if err := json.Unmarshal(chunk, &payload); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if payload["type"] != "error" {
+		t.Fatalf("type = %v, want %q", payload["type"], "error")
+	}
+	if payload["code"] != "internal_server_error" {
+		t.Fatalf("code = %v, want %q", payload["code"], "internal_server_error")
+	}
+	if payload["message"] != "unexpected EOF" {
+		t.Fatalf("message = %v, want %q", payload["message"], "unexpected EOF")
+	}
+	if payload["sequence_number"] != float64(0) {
+		t.Fatalf("sequence_number = %v, want %v", payload["sequence_number"], 0)
+	}
+}
+
+func TestBuildOpenAIResponsesStreamErrorChunkExtractsHTTPErrorBody(t *testing.T) {
+	chunk := BuildOpenAIResponsesStreamErrorChunk(
+		http.StatusInternalServerError,
+		`{"error":{"message":"oops","type":"server_error","code":"internal_server_error"}}`,
+		0,
+	)
+	var payload map[string]any
+	if err := json.Unmarshal(chunk, &payload); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if payload["type"] != "error" {
+		t.Fatalf("type = %v, want %q", payload["type"], "error")
+	}
+	if payload["code"] != "internal_server_error" {
+		t.Fatalf("code = %v, want %q", payload["code"], "internal_server_error")
+	}
+	if payload["message"] != "oops" {
+		t.Fatalf("message = %v, want %q", payload["message"], "oops")
+	}
+}

sdk/auth/codex.go

@@ -2,8 +2,6 @@ package auth
 
 import (
 	"context"
-	"crypto/sha256"
-	"encoding/hex"
 	"fmt"
 	"net/http"
 	"strings"
@@ -48,6 +46,10 @@ func (a *CodexAuthenticator) Login(ctx context.Context, cfg *config.Config, opts
 		opts = &LoginOptions{}
 	}
 
+	if shouldUseCodexDeviceFlow(opts) {
+		return a.loginWithDeviceFlow(ctx, cfg, opts)
+	}
+
 	callbackPort := a.CallbackPort
 	if opts.CallbackPort > 0 {
 		callbackPort = opts.CallbackPort
@@ -186,39 +188,5 @@ waitForCallback:
 		return nil, codex.NewAuthenticationError(codex.ErrCodeExchangeFailed, err)
 	}
 
-	tokenStorage := authSvc.CreateTokenStorage(authBundle)
-
-	if tokenStorage == nil || tokenStorage.Email == "" {
-		return nil, fmt.Errorf("codex token storage missing account information")
-	}
-
-	planType := ""
-	hashAccountID := ""
-	if tokenStorage.IDToken != "" {
-		if claims, errParse := codex.ParseJWTToken(tokenStorage.IDToken); errParse == nil && claims != nil {
-			planType = strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType)
-			accountID := strings.TrimSpace(claims.CodexAuthInfo.ChatgptAccountID)
-			if accountID != "" {
-				digest := sha256.Sum256([]byte(accountID))
-				hashAccountID = hex.EncodeToString(digest[:])[:8]
-			}
-		}
-	}
-	fileName := codex.CredentialFileName(tokenStorage.Email, planType, hashAccountID, true)
-	metadata := map[string]any{
-		"email": tokenStorage.Email,
-	}
-
-	fmt.Println("Codex authentication successful")
-	if authBundle.APIKey != "" {
-		fmt.Println("Codex API key obtained and stored")
-	}
-
-	return &coreauth.Auth{
-		ID:       fileName,
-		Provider: a.Provider(),
-		FileName: fileName,
-		Storage:  tokenStorage,
-		Metadata: metadata,
-	}, nil
+	return a.buildAuthRecord(authSvc, authBundle)
 }

sdk/auth/codex_device.go

@@ -0,0 +1,291 @@
+package auth
+
+import (
+	"bytes"
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/codex"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/browser"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	codexLoginModeMetadataKey             = "codex_login_mode"
+	codexLoginModeDevice                  = "device"
+	codexDeviceUserCodeURL                = "https://auth.openai.com/api/accounts/deviceauth/usercode"
+	codexDeviceTokenURL                   = "https://auth.openai.com/api/accounts/deviceauth/token"
+	codexDeviceVerificationURL            = "https://auth.openai.com/codex/device"
+	codexDeviceTokenExchangeRedirectURI   = "https://auth.openai.com/deviceauth/callback"
+	codexDeviceTimeout                    = 15 * time.Minute
+	codexDeviceDefaultPollIntervalSeconds = 5
+)
+
+type codexDeviceUserCodeRequest struct {
+	ClientID string `json:"client_id"`
+}
+
+type codexDeviceUserCodeResponse struct {
+	DeviceAuthID string          `json:"device_auth_id"`
+	UserCode     string          `json:"user_code"`
+	UserCodeAlt  string          `json:"usercode"`
+	Interval     json.RawMessage `json:"interval"`
+}
+
+type codexDeviceTokenRequest struct {
+	DeviceAuthID string `json:"device_auth_id"`
+	UserCode     string `json:"user_code"`
+}
+
+type codexDeviceTokenResponse struct {
+	AuthorizationCode string `json:"authorization_code"`
+	CodeVerifier      string `json:"code_verifier"`
+	CodeChallenge     string `json:"code_challenge"`
+}
+
+func shouldUseCodexDeviceFlow(opts *LoginOptions) bool {
+	if opts == nil || opts.Metadata == nil {
+		return false
+	}
+	return strings.EqualFold(strings.TrimSpace(opts.Metadata[codexLoginModeMetadataKey]), codexLoginModeDevice)
+}
+
+func (a *CodexAuthenticator) loginWithDeviceFlow(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+
+	httpClient := util.SetProxy(&cfg.SDKConfig, &http.Client{})
+
+	userCodeResp, err := requestCodexDeviceUserCode(ctx, httpClient)
+	if err != nil {
+		return nil, err
+	}
+
+	deviceCode := strings.TrimSpace(userCodeResp.UserCode)
+	if deviceCode == "" {
+		deviceCode = strings.TrimSpace(userCodeResp.UserCodeAlt)
+	}
+	deviceAuthID := strings.TrimSpace(userCodeResp.DeviceAuthID)
+	if deviceCode == "" || deviceAuthID == "" {
+		return nil, fmt.Errorf("codex device flow did not return required fields")
+	}
+
+	pollInterval := parseCodexDevicePollInterval(userCodeResp.Interval)
+
+	fmt.Println("Starting Codex device authentication...")
+	fmt.Printf("Codex device URL: %s\n", codexDeviceVerificationURL)
+	fmt.Printf("Codex device code: %s\n", deviceCode)
+
+	if !opts.NoBrowser {
+		if !browser.IsAvailable() {
+			log.Warn("No browser available; please open the device URL manually")
+		} else if errOpen := browser.OpenURL(codexDeviceVerificationURL); errOpen != nil {
+			log.Warnf("Failed to open browser automatically: %v", errOpen)
+		}
+	}
+
+	tokenResp, err := pollCodexDeviceToken(ctx, httpClient, deviceAuthID, deviceCode, pollInterval)
+	if err != nil {
+		return nil, err
+	}
+
+	authCode := strings.TrimSpace(tokenResp.AuthorizationCode)
+	codeVerifier := strings.TrimSpace(tokenResp.CodeVerifier)
+	codeChallenge := strings.TrimSpace(tokenResp.CodeChallenge)
+	if authCode == "" || codeVerifier == "" || codeChallenge == "" {
+		return nil, fmt.Errorf("codex device flow token response missing required fields")
+	}
+
+	authSvc := codex.NewCodexAuth(cfg)
+	authBundle, err := authSvc.ExchangeCodeForTokensWithRedirect(
+		ctx,
+		authCode,
+		codexDeviceTokenExchangeRedirectURI,
+		&codex.PKCECodes{
+			CodeVerifier:  codeVerifier,
+			CodeChallenge: codeChallenge,
+		},
+	)
+	if err != nil {
+		return nil, codex.NewAuthenticationError(codex.ErrCodeExchangeFailed, err)
+	}
+
+	return a.buildAuthRecord(authSvc, authBundle)
+}
+
+func requestCodexDeviceUserCode(ctx context.Context, client *http.Client) (*codexDeviceUserCodeResponse, error) {
+	body, err := json.Marshal(codexDeviceUserCodeRequest{ClientID: codex.ClientID})
+	if err != nil {
+		return nil, fmt.Errorf("failed to encode codex device request: %w", err)
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, codexDeviceUserCodeURL, bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create codex device request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "application/json")
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to request codex device code: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read codex device code response: %w", err)
+	}
+
+	if !codexDeviceIsSuccessStatus(resp.StatusCode) {
+		trimmed := strings.TrimSpace(string(respBody))
+		if resp.StatusCode == http.StatusNotFound {
+			return nil, fmt.Errorf("codex device endpoint is unavailable (status %d)", resp.StatusCode)
+		}
+		if trimmed == "" {
+			trimmed = "empty response body"
+		}
+		return nil, fmt.Errorf("codex device code request failed with status %d: %s", resp.StatusCode, trimmed)
+	}
+
+	var parsed codexDeviceUserCodeResponse
+	if err := json.Unmarshal(respBody, &parsed); err != nil {
+		return nil, fmt.Errorf("failed to decode codex device code response: %w", err)
+	}
+
+	return &parsed, nil
+}
+
+func pollCodexDeviceToken(ctx context.Context, client *http.Client, deviceAuthID, userCode string, interval time.Duration) (*codexDeviceTokenResponse, error) {
+	deadline := time.Now().Add(codexDeviceTimeout)
+
+	for {
+		if time.Now().After(deadline) {
+			return nil, fmt.Errorf("codex device authentication timed out after 15 minutes")
+		}
+
+		body, err := json.Marshal(codexDeviceTokenRequest{
+			DeviceAuthID: deviceAuthID,
+			UserCode:     userCode,
+		})
+		if err != nil {
+			return nil, fmt.Errorf("failed to encode codex device poll request: %w", err)
+		}
+
+		req, err := http.NewRequestWithContext(ctx, http.MethodPost, codexDeviceTokenURL, bytes.NewReader(body))
+		if err != nil {
+			return nil, fmt.Errorf("failed to create codex device poll request: %w", err)
+		}
+		req.Header.Set("Content-Type", "application/json")
+		req.Header.Set("Accept", "application/json")
+
+		resp, err := client.Do(req)
+		if err != nil {
+			return nil, fmt.Errorf("failed to poll codex device token: %w", err)
+		}
+
+		respBody, readErr := io.ReadAll(resp.Body)
+		_ = resp.Body.Close()
+		if readErr != nil {
+			return nil, fmt.Errorf("failed to read codex device poll response: %w", readErr)
+		}
+
+		switch {
+		case codexDeviceIsSuccessStatus(resp.StatusCode):
+			var parsed codexDeviceTokenResponse
+			if err := json.Unmarshal(respBody, &parsed); err != nil {
+				return nil, fmt.Errorf("failed to decode codex device token response: %w", err)
+			}
+			return &parsed, nil
+		case resp.StatusCode == http.StatusForbidden || resp.StatusCode == http.StatusNotFound:
+			select {
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			case <-time.After(interval):
+				continue
+			}
+		default:
+			trimmed := strings.TrimSpace(string(respBody))
+			if trimmed == "" {
+				trimmed = "empty response body"
+			}
+			return nil, fmt.Errorf("codex device token polling failed with status %d: %s", resp.StatusCode, trimmed)
+		}
+	}
+}
+
+func parseCodexDevicePollInterval(raw json.RawMessage) time.Duration {
+	defaultInterval := time.Duration(codexDeviceDefaultPollIntervalSeconds) * time.Second
+	if len(raw) == 0 {
+		return defaultInterval
+	}
+
+	var asString string
+	if err := json.Unmarshal(raw, &asString); err == nil {
+		if seconds, convErr := strconv.Atoi(strings.TrimSpace(asString)); convErr == nil && seconds > 0 {
+			return time.Duration(seconds) * time.Second
+		}
+	}
+
+	var asInt int
+	if err := json.Unmarshal(raw, &asInt); err == nil && asInt > 0 {
+		return time.Duration(asInt) * time.Second
+	}
+
+	return defaultInterval
+}
+
+func codexDeviceIsSuccessStatus(code int) bool {
+	return code >= 200 && code < 300
+}
+
+func (a *CodexAuthenticator) buildAuthRecord(authSvc *codex.CodexAuth, authBundle *codex.CodexAuthBundle) (*coreauth.Auth, error) {
+	tokenStorage := authSvc.CreateTokenStorage(authBundle)
+
+	if tokenStorage == nil || tokenStorage.Email == "" {
+		return nil, fmt.Errorf("codex token storage missing account information")
+	}
+
+	planType := ""
+	hashAccountID := ""
+	if tokenStorage.IDToken != "" {
+		if claims, errParse := codex.ParseJWTToken(tokenStorage.IDToken); errParse == nil && claims != nil {
+			planType = strings.TrimSpace(claims.CodexAuthInfo.ChatgptPlanType)
+			accountID := strings.TrimSpace(claims.CodexAuthInfo.ChatgptAccountID)
+			if accountID != "" {
+				digest := sha256.Sum256([]byte(accountID))
+				hashAccountID = hex.EncodeToString(digest[:])[:8]
+			}
+		}
+	}
+
+	fileName := codex.CredentialFileName(tokenStorage.Email, planType, hashAccountID, true)
+	metadata := map[string]any{
+		"email": tokenStorage.Email,
+	}
+
+	fmt.Println("Codex authentication successful")
+	if authBundle.APIKey != "" {
+		fmt.Println("Codex API key obtained and stored")
+	}
+
+	return &coreauth.Auth{
+		ID:       fileName,
+		Provider: a.Provider(),
+		FileName: fileName,
+		Storage:  tokenStorage,
+		Metadata: metadata,
+	}, nil
+}

sdk/auth/filestore.go

@@ -64,8 +64,16 @@ func (s *FileTokenStore) Save(ctx context.Context, auth *cliproxyauth.Auth) (str
 		return "", fmt.Errorf("auth filestore: create dir failed: %w", err)
 	}
 
+	// metadataSetter is a private interface for TokenStorage implementations that support metadata injection.
+	type metadataSetter interface {
+		SetMetadata(map[string]any)
+	}
+
 	switch {
 	case auth.Storage != nil:
+		if setter, ok := auth.Storage.(metadataSetter); ok {
+			setter.SetMetadata(auth.Metadata)
+		}
 		if err = auth.Storage.SaveTokenToFile(path); err != nil {
 			return "", err
 		}

sdk/cliproxy/auth/conductor.go

@@ -30,8 +30,9 @@ type ProviderExecutor interface {
 	Identifier() string
 	// Execute handles non-streaming execution and returns the provider response payload.
 	Execute(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error)
-	// ExecuteStream handles streaming execution and returns a channel of provider chunks.
-	ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error)
+	// ExecuteStream handles streaming execution and returns a StreamResult containing
+	// upstream headers and a channel of provider chunks.
+	ExecuteStream(ctx context.Context, auth *Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error)
 	// Refresh attempts to refresh provider credentials and returns the updated auth state.
 	Refresh(ctx context.Context, auth *Auth) (*Auth, error)
 	// CountTokens returns the token count for the given request.
@@ -558,7 +559,7 @@ func (m *Manager) ExecuteCount(ctx context.Context, providers []string, req clip
 
 // ExecuteStream performs a streaming execution using the configured selector and executor.
 // It supports multiple providers for the same model and round-robins the starting provider per model.
-func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	normalized := m.normalizeProviders(providers)
 	if len(normalized) == 0 {
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
@@ -568,9 +569,9 @@ func (m *Manager) ExecuteStream(ctx context.Context, providers []string, req cli
 
 	var lastErr error
 	for attempt := 0; ; attempt++ {
-		chunks, errStream := m.executeStreamMixedOnce(ctx, normalized, req, opts)
+		result, errStream := m.executeStreamMixedOnce(ctx, normalized, req, opts)
 		if errStream == nil {
-			return chunks, nil
+			return result, nil
 		}
 		lastErr = errStream
 		wait, shouldRetry := m.shouldRetryAfterError(errStream, attempt, normalized, req.Model, maxWait)
@@ -699,7 +700,7 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 	}
 }
 
-func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	if len(providers) == 0 {
 		return nil, &Error{Code: "provider_not_found", Message: "no provider supplied"}
 	}
@@ -730,7 +731,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 		execReq.Model = rewriteModelForAuth(routeModel, auth)
 		execReq.Model = m.applyOAuthModelAlias(auth, execReq.Model)
 		execReq.Model = m.applyAPIKeyModelAlias(auth, execReq.Model)
-		chunks, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
+		streamResult, errStream := executor.ExecuteStream(execCtx, auth, execReq, opts)
 		if errStream != nil {
 			if errCtx := execCtx.Err(); errCtx != nil {
 				return nil, errCtx
@@ -778,8 +779,11 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 			if !failed {
 				m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: true})
 			}
-		}(execCtx, auth.Clone(), provider, chunks)
-		return out, nil
+		}(execCtx, auth.Clone(), provider, streamResult.Chunks)
+		return &cliproxyexecutor.StreamResult{
+			Headers: streamResult.Headers,
+			Chunks:  out,
+		}, nil
 	}
 }
 
@@ -1824,9 +1828,7 @@ func (m *Manager) persist(ctx context.Context, auth *Auth) error {
 // every few seconds and triggers refresh operations when required.
 // Only one loop is kept alive; starting a new one cancels the previous run.
 func (m *Manager) StartAutoRefresh(parent context.Context, interval time.Duration) {
-	if interval <= 0 || interval > refreshCheckInterval {
-		interval = refreshCheckInterval
-	} else {
+	if interval <= 0 {
 		interval = refreshCheckInterval
 	}
 	if m.refreshCancel != nil {

sdk/cliproxy/auth/conductor_executor_replace_test.go

@@ -24,10 +24,10 @@ func (e *replaceAwareExecutor) Execute(context.Context, *Auth, cliproxyexecutor.
 	return cliproxyexecutor.Response{}, nil
 }
 
-func (e *replaceAwareExecutor) ExecuteStream(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+func (e *replaceAwareExecutor) ExecuteStream(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (*cliproxyexecutor.StreamResult, error) {
 	ch := make(chan cliproxyexecutor.StreamChunk)
 	close(ch)
-	return ch, nil
+	return &cliproxyexecutor.StreamResult{Chunks: ch}, nil
 }
 
 func (e *replaceAwareExecutor) Refresh(_ context.Context, auth *Auth) (*Auth, error) {
@@ -89,7 +89,11 @@ func TestManagerExecutorReturnsRegisteredExecutor(t *testing.T) {
 	if !okResolved {
 		t.Fatal("expected registered executor to be found")
 	}
-	if resolved != current {
+	resolvedExecutor, okResolvedExecutor := resolved.(*replaceAwareExecutor)
+	if !okResolvedExecutor {
+		t.Fatalf("expected resolved executor type %T, got %T", current, resolved)
+	}
+	if resolvedExecutor != current {
 		t.Fatal("expected resolved executor to match registered executor")
 	}
 

sdk/cliproxy/auth/selector.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"math"
+	"math/rand/v2"
 	"net/http"
 	"sort"
 	"strconv"
@@ -248,6 +249,9 @@ func getAvailableAuths(auths []*Auth, provider, model string, now time.Time) ([]
 }
 
 // Pick selects the next available auth for the provider in a round-robin manner.
+// For gemini-cli virtual auths (identified by the gemini_virtual_parent attribute),
+// a two-level round-robin is used: first cycling across credential groups (parent
+// accounts), then cycling within each group's project auths.
 func (s *RoundRobinSelector) Pick(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, auths []*Auth) (*Auth, error) {
 	_ = opts
 	now := time.Now()
@@ -265,21 +269,87 @@ func (s *RoundRobinSelector) Pick(ctx context.Context, provider, model string, o
 	if limit <= 0 {
 		limit = 4096
 	}
-	if _, ok := s.cursors[key]; !ok && len(s.cursors) >= limit {
-		s.cursors = make(map[string]int)
-	}
-	index := s.cursors[key]
 
+	// Check if any available auth has gemini_virtual_parent attribute,
+	// indicating gemini-cli virtual auths that should use credential-level polling.
+	groups, parentOrder := groupByVirtualParent(available)
+	if len(parentOrder) > 1 {
+		// Two-level round-robin: first select a credential group, then pick within it.
+		groupKey := key + "::group"
+		s.ensureCursorKey(groupKey, limit)
+		if _, exists := s.cursors[groupKey]; !exists {
+			// Seed with a random initial offset so the starting credential is randomized.
+			s.cursors[groupKey] = rand.IntN(len(parentOrder))
+		}
+		groupIndex := s.cursors[groupKey]
+		if groupIndex >= 2_147_483_640 {
+			groupIndex = 0
+		}
+		s.cursors[groupKey] = groupIndex + 1
+
+		selectedParent := parentOrder[groupIndex%len(parentOrder)]
+		group := groups[selectedParent]
+
+		// Second level: round-robin within the selected credential group.
+		innerKey := key + "::cred:" + selectedParent
+		s.ensureCursorKey(innerKey, limit)
+		innerIndex := s.cursors[innerKey]
+		if innerIndex >= 2_147_483_640 {
+			innerIndex = 0
+		}
+		s.cursors[innerKey] = innerIndex + 1
+		s.mu.Unlock()
+		return group[innerIndex%len(group)], nil
+	}
+
+	// Flat round-robin for non-grouped auths (original behavior).
+	s.ensureCursorKey(key, limit)
+	index := s.cursors[key]
 	if index >= 2_147_483_640 {
 		index = 0
 	}
-
 	s.cursors[key] = index + 1
 	s.mu.Unlock()
-	// log.Debugf("available: %d, index: %d, key: %d", len(available), index, index%len(available))
 	return available[index%len(available)], nil
 }
 
+// ensureCursorKey ensures the cursor map has capacity for the given key.
+// Must be called with s.mu held.
+func (s *RoundRobinSelector) ensureCursorKey(key string, limit int) {
+	if _, ok := s.cursors[key]; !ok && len(s.cursors) >= limit {
+		s.cursors = make(map[string]int)
+	}
+}
+
+// groupByVirtualParent groups auths by their gemini_virtual_parent attribute.
+// Returns a map of parentID -> auths and a sorted slice of parent IDs for stable iteration.
+// Only auths with a non-empty gemini_virtual_parent are grouped; if any auth lacks
+// this attribute, nil/nil is returned so the caller falls back to flat round-robin.
+func groupByVirtualParent(auths []*Auth) (map[string][]*Auth, []string) {
+	if len(auths) == 0 {
+		return nil, nil
+	}
+	groups := make(map[string][]*Auth)
+	for _, a := range auths {
+		parent := ""
+		if a.Attributes != nil {
+			parent = strings.TrimSpace(a.Attributes["gemini_virtual_parent"])
+		}
+		if parent == "" {
+			// Non-virtual auth present; fall back to flat round-robin.
+			return nil, nil
+		}
+		groups[parent] = append(groups[parent], a)
+	}
+	// Collect parent IDs in sorted order for stable cursor indexing.
+	parentOrder := make([]string, 0, len(groups))
+	for p := range groups {
+		parentOrder = append(parentOrder, p)
+	}
+	sort.Strings(parentOrder)
+	return groups, parentOrder
+}
+
 // Pick selects the first available auth for the provider in a deterministic manner.
 func (s *FillFirstSelector) Pick(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, auths []*Auth) (*Auth, error) {
 	_ = opts

sdk/cliproxy/auth/selector_test.go

@@ -402,3 +402,128 @@ func TestRoundRobinSelectorPick_CursorKeyCap(t *testing.T) {
 		t.Fatalf("selector.cursors missing key %q", "gemini:m3")
 	}
 }
+
+func TestRoundRobinSelectorPick_GeminiCLICredentialGrouping(t *testing.T) {
+	t.Parallel()
+
+	selector := &RoundRobinSelector{}
+
+	// Simulate two gemini-cli credentials, each with multiple projects:
+	// Credential A (parent = "cred-a.json") has 3 projects
+	// Credential B (parent = "cred-b.json") has 2 projects
+	auths := []*Auth{
+		{ID: "cred-a.json::proj-a1", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-a.json::proj-a2", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-a.json::proj-a3", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-b.json::proj-b1", Attributes: map[string]string{"gemini_virtual_parent": "cred-b.json"}},
+		{ID: "cred-b.json::proj-b2", Attributes: map[string]string{"gemini_virtual_parent": "cred-b.json"}},
+	}
+
+	// Two-level round-robin: consecutive picks must alternate between credentials.
+	// Credential group order is randomized, but within each call the group cursor
+	// advances by 1, so consecutive picks should cycle through different parents.
+	picks := make([]string, 6)
+	parents := make([]string, 6)
+	for i := 0; i < 6; i++ {
+		got, err := selector.Pick(context.Background(), "gemini-cli", "gemini-2.5-pro", cliproxyexecutor.Options{}, auths)
+		if err != nil {
+			t.Fatalf("Pick() #%d error = %v", i, err)
+		}
+		if got == nil {
+			t.Fatalf("Pick() #%d auth = nil", i)
+		}
+		picks[i] = got.ID
+		parents[i] = got.Attributes["gemini_virtual_parent"]
+	}
+
+	// Verify property: consecutive picks must alternate between credential groups.
+	for i := 1; i < len(parents); i++ {
+		if parents[i] == parents[i-1] {
+			t.Fatalf("Pick() #%d and #%d both from same parent %q (IDs: %q, %q); expected alternating credentials",
+				i-1, i, parents[i], picks[i-1], picks[i])
+		}
+	}
+
+	// Verify property: each credential's projects are picked in sequence (round-robin within group).
+	credPicks := map[string][]string{}
+	for i, id := range picks {
+		credPicks[parents[i]] = append(credPicks[parents[i]], id)
+	}
+	for parent, ids := range credPicks {
+		for i := 1; i < len(ids); i++ {
+			if ids[i] == ids[i-1] {
+				t.Fatalf("Credential %q picked same project %q twice in a row", parent, ids[i])
+			}
+		}
+	}
+}
+
+func TestRoundRobinSelectorPick_SingleParentFallsBackToFlat(t *testing.T) {
+	t.Parallel()
+
+	selector := &RoundRobinSelector{}
+
+	// All auths from the same parent - should fall back to flat round-robin
+	// because there's only one credential group (no benefit from two-level).
+	auths := []*Auth{
+		{ID: "cred-a.json::proj-a1", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-a.json::proj-a2", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-a.json::proj-a3", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+	}
+
+	// With single parent group, parentOrder has length 1, so it uses flat round-robin.
+	// Sorted by ID: proj-a1, proj-a2, proj-a3
+	want := []string{
+		"cred-a.json::proj-a1",
+		"cred-a.json::proj-a2",
+		"cred-a.json::proj-a3",
+		"cred-a.json::proj-a1",
+	}
+
+	for i, expectedID := range want {
+		got, err := selector.Pick(context.Background(), "gemini-cli", "gemini-2.5-pro", cliproxyexecutor.Options{}, auths)
+		if err != nil {
+			t.Fatalf("Pick() #%d error = %v", i, err)
+		}
+		if got == nil {
+			t.Fatalf("Pick() #%d auth = nil", i)
+		}
+		if got.ID != expectedID {
+			t.Fatalf("Pick() #%d auth.ID = %q, want %q", i, got.ID, expectedID)
+		}
+	}
+}
+
+func TestRoundRobinSelectorPick_MixedVirtualAndNonVirtualFallsBackToFlat(t *testing.T) {
+	t.Parallel()
+
+	selector := &RoundRobinSelector{}
+
+	// Mix of virtual and non-virtual auths (e.g., a regular gemini-cli auth without projects
+	// alongside virtual ones). Should fall back to flat round-robin.
+	auths := []*Auth{
+		{ID: "cred-a.json::proj-a1", Attributes: map[string]string{"gemini_virtual_parent": "cred-a.json"}},
+		{ID: "cred-regular.json"}, // no gemini_virtual_parent
+	}
+
+	// groupByVirtualParent returns nil when any auth lacks the attribute,
+	// so flat round-robin is used. Sorted by ID: cred-a.json::proj-a1, cred-regular.json
+	want := []string{
+		"cred-a.json::proj-a1",
+		"cred-regular.json",
+		"cred-a.json::proj-a1",
+	}
+
+	for i, expectedID := range want {
+		got, err := selector.Pick(context.Background(), "gemini-cli", "", cliproxyexecutor.Options{}, auths)
+		if err != nil {
+			t.Fatalf("Pick() #%d error = %v", i, err)
+		}
+		if got == nil {
+			t.Fatalf("Pick() #%d auth = nil", i)
+		}
+		if got.ID != expectedID {
+			t.Fatalf("Pick() #%d auth.ID = %q, want %q", i, got.ID, expectedID)
+		}
+	}
+}

sdk/cliproxy/auth/types.go

@@ -1,9 +1,12 @@
 package auth
 
 import (
+	"context"
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
+	"net/http"
+	"net/url"
 	"strconv"
 	"strings"
 	"sync"
@@ -12,6 +15,33 @@ import (
 	baseauth "github.com/router-for-me/CLIProxyAPI/v6/internal/auth"
 )
 
+// PostAuthHook defines a function that is called after an Auth record is created
+// but before it is persisted to storage. This allows for modification of the
+// Auth record (e.g., injecting metadata) based on external context.
+type PostAuthHook func(context.Context, *Auth) error
+
+// RequestInfo holds information extracted from the HTTP request.
+// It is injected into the context passed to PostAuthHook.
+type RequestInfo struct {
+	Query   url.Values
+	Headers http.Header
+}
+
+type requestInfoKey struct{}
+
+// WithRequestInfo returns a new context with the given RequestInfo attached.
+func WithRequestInfo(ctx context.Context, info *RequestInfo) context.Context {
+	return context.WithValue(ctx, requestInfoKey{}, info)
+}
+
+// GetRequestInfo retrieves the RequestInfo from the context, if present.
+func GetRequestInfo(ctx context.Context) *RequestInfo {
+	if val, ok := ctx.Value(requestInfoKey{}).(*RequestInfo); ok {
+		return val
+	}
+	return nil
+}
+
 // Auth encapsulates the runtime state and metadata associated with a single credential.
 type Auth struct {
 	// ID uniquely identifies the auth record across restarts.

sdk/cliproxy/builder.go

@@ -153,6 +153,16 @@ func (b *Builder) WithLocalManagementPassword(password string) *Builder {
 	return b
 }
 
+// WithPostAuthHook registers a hook to be called after an Auth record is created
+// but before it is persisted to storage.
+func (b *Builder) WithPostAuthHook(hook coreauth.PostAuthHook) *Builder {
+	if hook == nil {
+		return b
+	}
+	b.serverOptions = append(b.serverOptions, api.WithPostAuthHook(hook))
+	return b
+}
+
 // Build validates inputs, applies defaults, and returns a ready-to-run service.
 func (b *Builder) Build() (*Service, error) {
 	if b.cfg == nil {

sdk/cliproxy/executor/types.go

@@ -57,6 +57,8 @@ type Response struct {
 	Payload []byte
 	// Metadata exposes optional structured data for translators.
 	Metadata map[string]any
+	// Headers carries upstream HTTP response headers for passthrough to clients.
+	Headers http.Header
 }
 
 // StreamChunk represents a single streaming payload unit emitted by provider executors.
@@ -67,6 +69,15 @@ type StreamChunk struct {
 	Err error
 }
 
+// StreamResult wraps the streaming response, providing both the chunk channel
+// and the upstream HTTP response headers captured before streaming begins.
+type StreamResult struct {
+	// Headers carries upstream HTTP response headers from the initial connection.
+	Headers http.Header
+	// Chunks is the channel of streaming payload units.
+	Chunks <-chan StreamChunk
+}
+
 // StatusError represents an error that carries an HTTP-like status code.
 // Provider executors should implement this when possible to enable
 // better auth state updates on failures (e.g., 401/402/429).