Merge branch 'dev'

- Introduced unit tests for request logging middleware to enhance coverage.
- Added WebSocket-based Codex executor to support Responses API upgrade.
- Updated middleware logic to selectively capture request bodies for memory efficiency.
- Enhanced Codex configuration handling with new WebSocket attributes.

.github/workflows/release.yaml

@@ -19,7 +19,7 @@ jobs:
       - run: git fetch --force --tags
       - uses: actions/setup-go@v4
         with:
-          go-version: '>=1.24.0'
+          go-version: '>=1.26.0'
           cache: true
       - name: Generate Build Metadata
         run: |

Dockerfile

@@ -1,4 +1,4 @@
-FROM golang:1.24-alpine AS builder
+FROM golang:1.26-alpine AS builder
 
 WORKDIR /app
 

README.md

@@ -146,6 +146,10 @@ A Windows tray application implemented using PowerShell scripts, without relying
 
 霖君 is a cross-platform desktop application for managing AI programming assistants, supporting macOS, Windows, and Linux systems. Unified management of Claude Code, Gemini CLI, OpenAI Codex, Qwen Code, and other AI coding tools, with local proxy for multi-account quota tracking and one-click configuration.
 
+### [CLIProxyAPI Dashboard](https://github.com/itsmylife44/cliproxyapi-dashboard)
+
+A modern web-based management dashboard for CLIProxyAPI built with Next.js, React, and PostgreSQL. Features real-time log streaming, structured configuration editing, API key management, OAuth provider integration for Claude/Gemini/Codex, usage analytics, container management, and config sync with OpenCode via companion plugin - no manual YAML editing needed.
+
 > [!NOTE]  
 > If you developed a project based on CLIProxyAPI, please open a PR to add it to this list.
 

README_CN.md

@@ -145,6 +145,10 @@ Windows 托盘应用，基于 PowerShell 脚本实现，不依赖任何第三方
 
 霖君是一款用于管理AI编程助手的跨平台桌面应用，支持macOS、Windows、Linux系统。统一管理Claude Code、Gemini CLI、OpenAI Codex、Qwen Code等AI编程工具，本地代理实现多账户配额跟踪和一键配置。
 
+### [CLIProxyAPI Dashboard](https://github.com/itsmylife44/cliproxyapi-dashboard)
+
+一个面向 CLIProxyAPI 的现代化 Web 管理仪表盘，基于 Next.js、React 和 PostgreSQL 构建。支持实时日志流、结构化配置编辑、API Key 管理、Claude/Gemini/Codex 的 OAuth 提供方集成、使用量分析、容器管理，并可通过配套插件与 OpenCode 同步配置，无需手动编辑 YAML。
+
 > [!NOTE]  
 > 如果你开发了基于 CLIProxyAPI 的项目，请提交一个 PR（拉取请求）将其添加到此列表中。
 

go.mod

@@ -1,6 +1,6 @@
 module github.com/router-for-me/CLIProxyAPI/v6
 
-go 1.24.0
+go 1.26.0
 
 require (
 	github.com/andybalholm/brotli v1.0.6

internal/api/handlers/management/auth_files.go

@@ -1188,6 +1188,30 @@ func (h *Handler) RequestGeminiCLIToken(c *gin.Context) {
 			}
 			ts.ProjectID = strings.Join(projects, ",")
 			ts.Checked = true
+		} else if strings.EqualFold(requestedProjectID, "GOOGLE_ONE") {
+			ts.Auto = false
+			if errSetup := performGeminiCLISetup(ctx, gemClient, &ts, ""); errSetup != nil {
+				log.Errorf("Google One auto-discovery failed: %v", errSetup)
+				SetOAuthSessionError(state, "Google One auto-discovery failed")
+				return
+			}
+			if strings.TrimSpace(ts.ProjectID) == "" {
+				log.Error("Google One auto-discovery returned empty project ID")
+				SetOAuthSessionError(state, "Google One auto-discovery returned empty project ID")
+				return
+			}
+			isChecked, errCheck := checkCloudAPIIsEnabled(ctx, gemClient, ts.ProjectID)
+			if errCheck != nil {
+				log.Errorf("Failed to verify Cloud AI API status: %v", errCheck)
+				SetOAuthSessionError(state, "Failed to verify Cloud AI API status")
+				return
+			}
+			ts.Checked = isChecked
+			if !isChecked {
+				log.Error("Cloud AI API is not enabled for the auto-discovered project")
+				SetOAuthSessionError(state, "Cloud AI API not enabled")
+				return
+			}
 		} else {
 			if errEnsure := ensureGeminiProjectAndOnboard(ctx, gemClient, &ts, requestedProjectID); errEnsure != nil {
 				log.Errorf("Failed to complete Gemini CLI onboarding: %v", errEnsure)
@@ -2036,7 +2060,48 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 		}
 	}
 	if projectID == "" {
-		return &projectSelectionRequiredError{}
+		// Auto-discovery: try onboardUser without specifying a project
+		// to let Google auto-provision one (matches Gemini CLI headless behavior
+		// and Antigravity's FetchProjectID pattern).
+		autoOnboardReq := map[string]any{
+			"tierId":   tierID,
+			"metadata": metadata,
+		}
+
+		autoCtx, autoCancel := context.WithTimeout(ctx, 30*time.Second)
+		defer autoCancel()
+		for attempt := 1; ; attempt++ {
+			var onboardResp map[string]any
+			if errOnboard := callGeminiCLI(autoCtx, httpClient, "onboardUser", autoOnboardReq, &onboardResp); errOnboard != nil {
+				return fmt.Errorf("auto-discovery onboardUser: %w", errOnboard)
+			}
+
+			if done, okDone := onboardResp["done"].(bool); okDone && done {
+				if resp, okResp := onboardResp["response"].(map[string]any); okResp {
+					switch v := resp["cloudaicompanionProject"].(type) {
+					case string:
+						projectID = strings.TrimSpace(v)
+					case map[string]any:
+						if id, okID := v["id"].(string); okID {
+							projectID = strings.TrimSpace(id)
+						}
+					}
+				}
+				break
+			}
+
+			log.Debugf("Auto-discovery: onboarding in progress, attempt %d...", attempt)
+			select {
+			case <-autoCtx.Done():
+				return &projectSelectionRequiredError{}
+			case <-time.After(2 * time.Second):
+			}
+		}
+
+		if projectID == "" {
+			return &projectSelectionRequiredError{}
+		}
+		log.Infof("Auto-discovered project ID via onboarding: %s", projectID)
 	}
 
 	onboardReqBody := map[string]any{

internal/api/handlers/management/config_basic.go

@@ -28,8 +28,7 @@ func (h *Handler) GetConfig(c *gin.Context) {
 		c.JSON(200, gin.H{})
 		return
 	}
-	cfgCopy := *h.cfg
-	c.JSON(200, &cfgCopy)
+	c.JSON(200, new(*h.cfg))
 }
 
 type releaseInfo struct {

internal/api/middleware/request_logging.go

@@ -15,10 +15,12 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 )
 
+const maxErrorOnlyCapturedRequestBodyBytes int64 = 1 << 20 // 1 MiB
+
 // RequestLoggingMiddleware creates a Gin middleware that logs HTTP requests and responses.
 // It captures detailed information about the request and response, including headers and body,
-// and uses the provided RequestLogger to record this data. When logging is disabled in the
-// logger, it still captures data so that upstream errors can be persisted.
+// and uses the provided RequestLogger to record this data. When full request logging is disabled,
+// body capture is limited to small known-size payloads to avoid large per-request memory spikes.
 func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 	return func(c *gin.Context) {
 		if logger == nil {
@@ -26,7 +28,7 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 			return
 		}
 
-		if c.Request.Method == http.MethodGet {
+		if shouldSkipMethodForRequestLogging(c.Request) {
 			c.Next()
 			return
 		}
@@ -37,8 +39,10 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 			return
 		}
 
+		loggerEnabled := logger.IsEnabled()
+
 		// Capture request information
-		requestInfo, err := captureRequestInfo(c)
+		requestInfo, err := captureRequestInfo(c, shouldCaptureRequestBody(loggerEnabled, c.Request))
 		if err != nil {
 			// Log error but continue processing
 			// In a real implementation, you might want to use a proper logger here
@@ -48,7 +52,7 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 
 		// Create response writer wrapper
 		wrapper := NewResponseWriterWrapper(c.Writer, logger, requestInfo)
-		if !logger.IsEnabled() {
+		if !loggerEnabled {
 			wrapper.logOnErrorOnly = true
 		}
 		c.Writer = wrapper
@@ -64,10 +68,47 @@ func RequestLoggingMiddleware(logger logging.RequestLogger) gin.HandlerFunc {
 	}
 }
 
+func shouldSkipMethodForRequestLogging(req *http.Request) bool {
+	if req == nil {
+		return true
+	}
+	if req.Method != http.MethodGet {
+		return false
+	}
+	return !isResponsesWebsocketUpgrade(req)
+}
+
+func isResponsesWebsocketUpgrade(req *http.Request) bool {
+	if req == nil || req.URL == nil {
+		return false
+	}
+	if req.URL.Path != "/v1/responses" {
+		return false
+	}
+	return strings.EqualFold(strings.TrimSpace(req.Header.Get("Upgrade")), "websocket")
+}
+
+func shouldCaptureRequestBody(loggerEnabled bool, req *http.Request) bool {
+	if loggerEnabled {
+		return true
+	}
+	if req == nil || req.Body == nil {
+		return false
+	}
+	contentType := strings.ToLower(strings.TrimSpace(req.Header.Get("Content-Type")))
+	if strings.HasPrefix(contentType, "multipart/form-data") {
+		return false
+	}
+	if req.ContentLength <= 0 {
+		return false
+	}
+	return req.ContentLength <= maxErrorOnlyCapturedRequestBodyBytes
+}
+
 // captureRequestInfo extracts relevant information from the incoming HTTP request.
 // It captures the URL, method, headers, and body. The request body is read and then
 // restored so that it can be processed by subsequent handlers.
-func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {
+func captureRequestInfo(c *gin.Context, captureBody bool) (*RequestInfo, error) {
 	// Capture URL with sensitive query parameters masked
 	maskedQuery := util.MaskSensitiveQuery(c.Request.URL.RawQuery)
 	url := c.Request.URL.Path
@@ -86,7 +127,7 @@ func captureRequestInfo(c *gin.Context) (*RequestInfo, error) {
 
 	// Capture request body
 	var body []byte
-	if c.Request.Body != nil {
+	if captureBody && c.Request.Body != nil {
 		// Read the body
 		bodyBytes, err := io.ReadAll(c.Request.Body)
 		if err != nil {

internal/api/middleware/request_logging_test.go

@@ -0,0 +1,138 @@
+package middleware
+
+import (
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"testing"
+)
+
+func TestShouldSkipMethodForRequestLogging(t *testing.T) {
+	tests := []struct {
+		name string
+		req  *http.Request
+		skip bool
+	}{
+		{
+			name: "nil request",
+			req:  nil,
+			skip: true,
+		},
+		{
+			name: "post request should not skip",
+			req: &http.Request{
+				Method: http.MethodPost,
+				URL:    &url.URL{Path: "/v1/responses"},
+			},
+			skip: false,
+		},
+		{
+			name: "plain get should skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/models"},
+				Header: http.Header{},
+			},
+			skip: true,
+		},
+		{
+			name: "responses websocket upgrade should not skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/responses"},
+				Header: http.Header{"Upgrade": []string{"websocket"}},
+			},
+			skip: false,
+		},
+		{
+			name: "responses get without upgrade should skip",
+			req: &http.Request{
+				Method: http.MethodGet,
+				URL:    &url.URL{Path: "/v1/responses"},
+				Header: http.Header{},
+			},
+			skip: true,
+		},
+	}
+
+	for i := range tests {
+		got := shouldSkipMethodForRequestLogging(tests[i].req)
+		if got != tests[i].skip {
+			t.Fatalf("%s: got skip=%t, want %t", tests[i].name, got, tests[i].skip)
+		}
+	}
+}
+
+func TestShouldCaptureRequestBody(t *testing.T) {
+	tests := []struct {
+		name          string
+		loggerEnabled bool
+		req           *http.Request
+		want          bool
+	}{
+		{
+			name:          "logger enabled always captures",
+			loggerEnabled: true,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("{}")),
+				ContentLength: -1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: true,
+		},
+		{
+			name:          "nil request",
+			loggerEnabled: false,
+			req:           nil,
+			want:          false,
+		},
+		{
+			name:          "small known size json in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("{}")),
+				ContentLength: 2,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: true,
+		},
+		{
+			name:          "large known size skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: maxErrorOnlyCapturedRequestBodyBytes + 1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: false,
+		},
+		{
+			name:          "unknown size skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: -1,
+				Header:        http.Header{"Content-Type": []string{"application/json"}},
+			},
+			want: false,
+		},
+		{
+			name:          "multipart skipped in error-only mode",
+			loggerEnabled: false,
+			req: &http.Request{
+				Body:          io.NopCloser(strings.NewReader("x")),
+				ContentLength: 1,
+				Header:        http.Header{"Content-Type": []string{"multipart/form-data; boundary=abc"}},
+			},
+			want: false,
+		},
+	}
+
+	for i := range tests {
+		got := shouldCaptureRequestBody(tests[i].loggerEnabled, tests[i].req)
+		if got != tests[i].want {
+			t.Fatalf("%s: got %t, want %t", tests[i].name, got, tests[i].want)
+		}
+	}
+}

internal/api/middleware/response_writer.go

@@ -14,6 +14,8 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/logging"
 )
 
+const requestBodyOverrideContextKey = "REQUEST_BODY_OVERRIDE"
+
 // RequestInfo holds essential details of an incoming HTTP request for logging purposes.
 type RequestInfo struct {
 	URL       string              // URL is the request URL.
@@ -223,8 +225,8 @@ func (w *ResponseWriterWrapper) detectStreaming(contentType string) bool {
 
 	// Only fall back to request payload hints when Content-Type is not set yet.
 	if w.requestInfo != nil && len(w.requestInfo.Body) > 0 {
-		bodyStr := string(w.requestInfo.Body)
-		return strings.Contains(bodyStr, `"stream": true`) || strings.Contains(bodyStr, `"stream":true`)
+		return bytes.Contains(w.requestInfo.Body, []byte(`"stream": true`)) ||
+			bytes.Contains(w.requestInfo.Body, []byte(`"stream":true`))
 	}
 
 	return false
@@ -310,7 +312,7 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 		return nil
 	}
 
-	return w.logRequest(finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), w.extractAPIResponseTimestamp(c), slicesAPIResponseError, forceLog)
+	return w.logRequest(w.extractRequestBody(c), finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), w.extractAPIResponseTimestamp(c), slicesAPIResponseError, forceLog)
 }
 
 func (w *ResponseWriterWrapper) cloneHeaders() map[string][]string {
@@ -361,16 +363,32 @@ func (w *ResponseWriterWrapper) extractAPIResponseTimestamp(c *gin.Context) time
 	return time.Time{}
 }
 
-func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseTimestamp time.Time, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
+func (w *ResponseWriterWrapper) extractRequestBody(c *gin.Context) []byte {
+	if c != nil {
+		if bodyOverride, isExist := c.Get(requestBodyOverrideContextKey); isExist {
+			switch value := bodyOverride.(type) {
+			case []byte:
+				if len(value) > 0 {
+					return bytes.Clone(value)
+				}
+			case string:
+				if strings.TrimSpace(value) != "" {
+					return []byte(value)
+				}
+			}
+		}
+	}
+	if w.requestInfo != nil && len(w.requestInfo.Body) > 0 {
+		return w.requestInfo.Body
+	}
+	return nil
+}
+
+func (w *ResponseWriterWrapper) logRequest(requestBody []byte, statusCode int, headers map[string][]string, body []byte, apiRequestBody, apiResponseBody []byte, apiResponseTimestamp time.Time, apiResponseErrors []*interfaces.ErrorMessage, forceLog bool) error {
 	if w.requestInfo == nil {
 		return nil
 	}
 
-	var requestBody []byte
-	if len(w.requestInfo.Body) > 0 {
-		requestBody = w.requestInfo.Body
-	}
-
 	if loggerWithOptions, ok := w.logger.(interface {
 		LogRequestWithOptions(string, string, map[string][]string, []byte, int, map[string][]string, []byte, []byte, []byte, []*interfaces.ErrorMessage, bool, string, time.Time, time.Time) error
 	}); ok {

internal/api/middleware/response_writer_test.go

@@ -0,0 +1,43 @@
+package middleware
+
+import (
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+)
+
+func TestExtractRequestBodyPrefersOverride(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	wrapper := &ResponseWriterWrapper{
+		requestInfo: &RequestInfo{Body: []byte("original-body")},
+	}
+
+	body := wrapper.extractRequestBody(c)
+	if string(body) != "original-body" {
+		t.Fatalf("request body = %q, want %q", string(body), "original-body")
+	}
+
+	c.Set(requestBodyOverrideContextKey, []byte("override-body"))
+	body = wrapper.extractRequestBody(c)
+	if string(body) != "override-body" {
+		t.Fatalf("request body = %q, want %q", string(body), "override-body")
+	}
+}
+
+func TestExtractRequestBodySupportsStringOverride(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	wrapper := &ResponseWriterWrapper{}
+	c.Set(requestBodyOverrideContextKey, "override-as-string")
+
+	body := wrapper.extractRequestBody(c)
+	if string(body) != "override-as-string" {
+		t.Fatalf("request body = %q, want %q", string(body), "override-as-string")
+	}
+}

internal/api/modules/amp/amp.go

@@ -127,8 +127,7 @@ func (m *AmpModule) Register(ctx modules.Context) error {
 		m.modelMapper = NewModelMapper(settings.ModelMappings)
 
 		// Store initial config for partial reload comparison
-		settingsCopy := settings
-		m.lastConfig = &settingsCopy
+		m.lastConfig = new(settings)
 
 		// Initialize localhost restriction setting (hot-reloadable)
 		m.setRestrictToLocalhost(settings.RestrictManagementToLocalhost)

internal/api/server.go

@@ -323,6 +323,7 @@ func (s *Server) setupRoutes() {
 		v1.POST("/completions", openaiHandlers.Completions)
 		v1.POST("/messages", claudeCodeHandlers.ClaudeMessages)
 		v1.POST("/messages/count_tokens", claudeCodeHandlers.ClaudeCountTokens)
+		v1.GET("/responses", openaiResponsesHandlers.ResponsesWebsocket)
 		v1.POST("/responses", openaiResponsesHandlers.Responses)
 		v1.POST("/responses/compact", openaiResponsesHandlers.Compact)
 	}

internal/cmd/anthropic_login.go

@@ -40,8 +40,7 @@ func DoClaudeLogin(cfg *config.Config, options *LoginOptions) {
 
 	_, savedPath, err := manager.Login(context.Background(), "claude", cfg, authOpts)
 	if err != nil {
-		var authErr *claude.AuthenticationError
-		if errors.As(err, &authErr) {
+		if authErr, ok := errors.AsType[*claude.AuthenticationError](err); ok {
 			log.Error(claude.GetUserFriendlyMessage(authErr))
 			if authErr.Type == claude.ErrPortInUse.Type {
 				os.Exit(claude.ErrPortInUse.Code)

internal/cmd/iflow_login.go

@@ -32,8 +32,7 @@ func DoIFlowLogin(cfg *config.Config, options *LoginOptions) {
 
 	_, savedPath, err := manager.Login(context.Background(), "iflow", cfg, authOpts)
 	if err != nil {
-		var emailErr *sdkAuth.EmailRequiredError
-		if errors.As(err, &emailErr) {
+		if emailErr, ok := errors.AsType[*sdkAuth.EmailRequiredError](err); ok {
 			log.Error(emailErr.Error())
 			return
 		}

internal/cmd/login.go

@@ -100,49 +100,74 @@ func DoLogin(cfg *config.Config, projectID string, options *LoginOptions) {
 
 	log.Info("Authentication successful.")
 
-	projects, errProjects := fetchGCPProjects(ctx, httpClient)
-	if errProjects != nil {
-		log.Errorf("Failed to get project list: %v", errProjects)
-		return
+	var activatedProjects []string
+
+	useGoogleOne := false
+	if trimmedProjectID == "" && promptFn != nil {
+		fmt.Println("\nSelect login mode:")
+		fmt.Println("  1. Code Assist  (GCP project, manual selection)")
+		fmt.Println("  2. Google One   (personal account, auto-discover project)")
+		choice, errPrompt := promptFn("Enter choice [1/2] (default: 1): ")
+		if errPrompt == nil && strings.TrimSpace(choice) == "2" {
+			useGoogleOne = true
+		}
 	}
 
-	selectedProjectID := promptForProjectSelection(projects, trimmedProjectID, promptFn)
-	projectSelections, errSelection := resolveProjectSelections(selectedProjectID, projects)
-	if errSelection != nil {
-		log.Errorf("Invalid project selection: %v", errSelection)
-		return
-	}
-	if len(projectSelections) == 0 {
-		log.Error("No project selected; aborting login.")
-		return
-	}
-
-	activatedProjects := make([]string, 0, len(projectSelections))
-	seenProjects := make(map[string]bool)
-	for _, candidateID := range projectSelections {
-		log.Infof("Activating project %s", candidateID)
-		if errSetup := performGeminiCLISetup(ctx, httpClient, storage, candidateID); errSetup != nil {
-			var projectErr *projectSelectionRequiredError
-			if errors.As(errSetup, &projectErr) {
-				log.Error("Failed to start user onboarding: A project ID is required.")
-				showProjectSelectionHelp(storage.Email, projects)
-				return
-			}
-			log.Errorf("Failed to complete user setup: %v", errSetup)
+	if useGoogleOne {
+		log.Info("Google One mode: auto-discovering project...")
+		if errSetup := performGeminiCLISetup(ctx, httpClient, storage, ""); errSetup != nil {
+			log.Errorf("Google One auto-discovery failed: %v", errSetup)
 			return
 		}
-		finalID := strings.TrimSpace(storage.ProjectID)
-		if finalID == "" {
-			finalID = candidateID
+		autoProject := strings.TrimSpace(storage.ProjectID)
+		if autoProject == "" {
+			log.Error("Google One auto-discovery returned empty project ID")
+			return
+		}
+		log.Infof("Auto-discovered project: %s", autoProject)
+		activatedProjects = []string{autoProject}
+	} else {
+		projects, errProjects := fetchGCPProjects(ctx, httpClient)
+		if errProjects != nil {
+			log.Errorf("Failed to get project list: %v", errProjects)
+			return
 		}
 
-		// Skip duplicates
-		if seenProjects[finalID] {
-			log.Infof("Project %s already activated, skipping", finalID)
-			continue
+		selectedProjectID := promptForProjectSelection(projects, trimmedProjectID, promptFn)
+		projectSelections, errSelection := resolveProjectSelections(selectedProjectID, projects)
+		if errSelection != nil {
+			log.Errorf("Invalid project selection: %v", errSelection)
+			return
+		}
+		if len(projectSelections) == 0 {
+			log.Error("No project selected; aborting login.")
+			return
+		}
+
+		seenProjects := make(map[string]bool)
+		for _, candidateID := range projectSelections {
+			log.Infof("Activating project %s", candidateID)
+			if errSetup := performGeminiCLISetup(ctx, httpClient, storage, candidateID); errSetup != nil {
+				if _, ok := errors.AsType[*projectSelectionRequiredError](errSetup); ok {
+					log.Error("Failed to start user onboarding: A project ID is required.")
+					showProjectSelectionHelp(storage.Email, projects)
+					return
+				}
+				log.Errorf("Failed to complete user setup: %v", errSetup)
+				return
+			}
+			finalID := strings.TrimSpace(storage.ProjectID)
+			if finalID == "" {
+				finalID = candidateID
+			}
+
+			if seenProjects[finalID] {
+				log.Infof("Project %s already activated, skipping", finalID)
+				continue
+			}
+			seenProjects[finalID] = true
+			activatedProjects = append(activatedProjects, finalID)
 		}
-		seenProjects[finalID] = true
-		activatedProjects = append(activatedProjects, finalID)
 	}
 
 	storage.Auto = false
@@ -235,7 +260,48 @@ func performGeminiCLISetup(ctx context.Context, httpClient *http.Client, storage
 		}
 	}
 	if projectID == "" {
-		return &projectSelectionRequiredError{}
+		// Auto-discovery: try onboardUser without specifying a project
+		// to let Google auto-provision one (matches Gemini CLI headless behavior
+		// and Antigravity's FetchProjectID pattern).
+		autoOnboardReq := map[string]any{
+			"tierId":   tierID,
+			"metadata": metadata,
+		}
+
+		autoCtx, autoCancel := context.WithTimeout(ctx, 30*time.Second)
+		defer autoCancel()
+		for attempt := 1; ; attempt++ {
+			var onboardResp map[string]any
+			if errOnboard := callGeminiCLI(autoCtx, httpClient, "onboardUser", autoOnboardReq, &onboardResp); errOnboard != nil {
+				return fmt.Errorf("auto-discovery onboardUser: %w", errOnboard)
+			}
+
+			if done, okDone := onboardResp["done"].(bool); okDone && done {
+				if resp, okResp := onboardResp["response"].(map[string]any); okResp {
+					switch v := resp["cloudaicompanionProject"].(type) {
+					case string:
+						projectID = strings.TrimSpace(v)
+					case map[string]any:
+						if id, okID := v["id"].(string); okID {
+							projectID = strings.TrimSpace(id)
+						}
+					}
+				}
+				break
+			}
+
+			log.Debugf("Auto-discovery: onboarding in progress, attempt %d...", attempt)
+			select {
+			case <-autoCtx.Done():
+				return &projectSelectionRequiredError{}
+			case <-time.After(2 * time.Second):
+			}
+		}
+
+		if projectID == "" {
+			return &projectSelectionRequiredError{}
+		}
+		log.Infof("Auto-discovered project ID via onboarding: %s", projectID)
 	}
 
 	onboardReqBody := map[string]any{
@@ -617,7 +683,7 @@ func updateAuthRecord(record *cliproxyauth.Auth, storage *gemini.GeminiTokenStor
 		return
 	}
 
-	finalName := gemini.CredentialFileName(storage.Email, storage.ProjectID, false)
+	finalName := gemini.CredentialFileName(storage.Email, storage.ProjectID, true)
 
 	if record.Metadata == nil {
 		record.Metadata = make(map[string]any)

internal/cmd/openai_login.go

@@ -54,8 +54,7 @@ func DoCodexLogin(cfg *config.Config, options *LoginOptions) {
 
 	_, savedPath, err := manager.Login(context.Background(), "codex", cfg, authOpts)
 	if err != nil {
-		var authErr *codex.AuthenticationError
-		if errors.As(err, &authErr) {
+		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)

internal/cmd/qwen_login.go

@@ -44,8 +44,7 @@ func DoQwenLogin(cfg *config.Config, options *LoginOptions) {
 
 	_, savedPath, err := manager.Login(context.Background(), "qwen", cfg, authOpts)
 	if err != nil {
-		var emailErr *sdkAuth.EmailRequiredError
-		if errors.As(err, &emailErr) {
+		if emailErr, ok := errors.AsType[*sdkAuth.EmailRequiredError](err); ok {
 			log.Error(emailErr.Error())
 			return
 		}

internal/config/config.go

@@ -355,6 +355,9 @@ type CodexKey struct {
 	// If empty, the default Codex API URL will be used.
 	BaseURL string `yaml:"base-url" json:"base-url"`
 
+	// Websockets enables the Responses API websocket transport for this credential.
+	Websockets bool `yaml:"websockets,omitempty" json:"websockets,omitempty"`
+
 	// ProxyURL overrides the global proxy setting for this API key if provided.
 	ProxyURL string `yaml:"proxy-url" json:"proxy-url"`
 

internal/registry/model_definitions.go

@@ -19,6 +19,7 @@ import (
 //   - codex
 //   - qwen
 //   - iflow
+//   - kimi
 //   - antigravity (returns static overrides only)
 func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
 	key := strings.ToLower(strings.TrimSpace(channel))
@@ -39,6 +40,8 @@ func GetStaticModelDefinitionsByChannel(channel string) []*ModelInfo {
 		return GetQwenModels()
 	case "iflow":
 		return GetIFlowModels()
+	case "kimi":
+		return GetKimiModels()
 	case "antigravity":
 		cfg := GetAntigravityModelConfig()
 		if len(cfg) == 0 {
@@ -83,6 +86,7 @@ func LookupStaticModelInfo(modelID string) *ModelInfo {
 		GetOpenAIModels(),
 		GetQwenModels(),
 		GetIFlowModels(),
+		GetKimiModels(),
 	}
 	for _, models := range allModels {
 		for _, m := range models {

internal/registry/model_definitions_static_data.go

@@ -28,6 +28,17 @@ func GetClaudeModels() []*ModelInfo {
 			MaxCompletionTokens: 64000,
 			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
 		},
+		{
+			ID:                  "claude-sonnet-4-6",
+			Object:              "model",
+			Created:             1771372800, // 2026-02-17
+			OwnedBy:             "anthropic",
+			Type:                "claude",
+			DisplayName:         "Claude 4.6 Sonnet",
+			ContextLength:       200000,
+			MaxCompletionTokens: 64000,
+			Thinking:            &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: false},
+		},
 		{
 			ID:                  "claude-opus-4-6",
 			Object:              "model",
@@ -742,6 +753,20 @@ func GetOpenAIModels() []*ModelInfo {
 			SupportedParameters: []string{"tools"},
 			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
 		},
+		{
+			ID:                  "gpt-5.3-codex-spark",
+			Object:              "model",
+			Created:             1770912000,
+			OwnedBy:             "openai",
+			Type:                "openai",
+			Version:             "gpt-5.3",
+			DisplayName:         "GPT 5.3 Codex Spark",
+			Description:         "Ultra-fast coding model.",
+			ContextLength:       128000,
+			MaxCompletionTokens: 128000,
+			SupportedParameters: []string{"tools"},
+			Thinking:            &ThinkingSupport{Levels: []string{"low", "medium", "high", "xhigh"}},
+		},
 	}
 }
 
@@ -774,6 +799,19 @@ func GetQwenModels() []*ModelInfo {
 			MaxCompletionTokens: 2048,
 			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
 		},
+		{
+			ID:                  "coder-model",
+			Object:              "model",
+			Created:             1771171200,
+			OwnedBy:             "qwen",
+			Type:                "qwen",
+			Version:             "3.5",
+			DisplayName:         "Qwen 3.5 Plus",
+			Description:         "efficient hybrid model with leading coding performance",
+			ContextLength:       1048576,
+			MaxCompletionTokens: 65536,
+			SupportedParameters: []string{"temperature", "top_p", "max_tokens", "stream", "stop"},
+		},
 		{
 			ID:                  "vision-model",
 			Object:              "model",
@@ -814,6 +852,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "kimi-k2-0905", DisplayName: "Kimi-K2-Instruct-0905", Description: "Moonshot Kimi K2 instruct 0905", Created: 1757030400},
 		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400, Thinking: iFlowThinkingSupport},
 		{ID: "glm-4.7", DisplayName: "GLM-4.7", Description: "Zhipu GLM 4.7 general model", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "glm-5", DisplayName: "GLM-5", Description: "Zhipu GLM 5 general model", Created: 1770768000, Thinking: iFlowThinkingSupport},
 		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
 		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200},
 		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
@@ -828,6 +867,7 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "qwen3-235b", DisplayName: "Qwen3-235B-A22B", Description: "Qwen3 235B A22B", Created: 1753401600},
 		{ID: "minimax-m2", DisplayName: "MiniMax-M2", Description: "MiniMax M2", Created: 1758672000, Thinking: iFlowThinkingSupport},
 		{ID: "minimax-m2.1", DisplayName: "MiniMax-M2.1", Description: "MiniMax M2.1", Created: 1766448000, Thinking: iFlowThinkingSupport},
+		{ID: "minimax-m2.5", DisplayName: "MiniMax-M2.5", Description: "MiniMax M2.5", Created: 1770825600, Thinking: iFlowThinkingSupport},
 		{ID: "iflow-rome-30ba3b", DisplayName: "iFlow-ROME", Description: "iFlow Rome 30BA3B model", Created: 1736899200},
 		{ID: "kimi-k2.5", DisplayName: "Kimi-K2.5", Description: "Moonshot Kimi K2.5", Created: 1769443200, Thinking: iFlowThinkingSupport},
 	}
@@ -866,8 +906,10 @@ func GetAntigravityModelConfig() map[string]*AntigravityModelConfig {
 		"gemini-3-flash":             {Thinking: &ThinkingSupport{Min: 128, Max: 32768, ZeroAllowed: false, DynamicAllowed: true, Levels: []string{"minimal", "low", "medium", "high"}}},
 		"claude-sonnet-4-5-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-opus-4-5-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
-		"claude-opus-4-6-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 128000},
+		"claude-opus-4-6-thinking":   {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"claude-sonnet-4-5":          {MaxCompletionTokens: 64000},
+		"claude-sonnet-4-6":          {MaxCompletionTokens: 64000},
+		"claude-sonnet-4-6-thinking": {Thinking: &ThinkingSupport{Min: 1024, Max: 128000, ZeroAllowed: true, DynamicAllowed: true}, MaxCompletionTokens: 64000},
 		"gpt-oss-120b-medium":        {},
 		"tab_flash_lite_preview":     {},
 	}

internal/registry/model_registry.go

@@ -596,8 +596,7 @@ func (r *ModelRegistry) SetModelQuotaExceeded(clientID, modelID string) {
 	defer r.mutex.Unlock()
 
 	if registration, exists := r.models[modelID]; exists {
-		now := time.Now()
-		registration.QuotaExceededClients[clientID] = &now
+		registration.QuotaExceededClients[clientID] = new(time.Now())
 		log.Debugf("Marked model %s as quota exceeded for client %s", modelID, clientID)
 	}
 }

internal/runtime/executor/antigravity_executor_buildrequest_test.go

@@ -0,0 +1,159 @@
+package executor
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"testing"
+
+	cliproxyauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+)
+
+func TestAntigravityBuildRequest_SanitizesGeminiToolSchema(t *testing.T) {
+	body := buildRequestBodyFromPayload(t, "gemini-2.5-pro")
+
+	decl := extractFirstFunctionDeclaration(t, body)
+	if _, ok := decl["parametersJsonSchema"]; ok {
+		t.Fatalf("parametersJsonSchema should be renamed to parameters")
+	}
+
+	params, ok := decl["parameters"].(map[string]any)
+	if !ok {
+		t.Fatalf("parameters missing or invalid type")
+	}
+	assertSchemaSanitizedAndPropertyPreserved(t, params)
+}
+
+func TestAntigravityBuildRequest_SanitizesAntigravityToolSchema(t *testing.T) {
+	body := buildRequestBodyFromPayload(t, "claude-opus-4-6")
+
+	decl := extractFirstFunctionDeclaration(t, body)
+	params, ok := decl["parameters"].(map[string]any)
+	if !ok {
+		t.Fatalf("parameters missing or invalid type")
+	}
+	assertSchemaSanitizedAndPropertyPreserved(t, params)
+}
+
+func buildRequestBodyFromPayload(t *testing.T, modelName string) map[string]any {
+	t.Helper()
+
+	executor := &AntigravityExecutor{}
+	auth := &cliproxyauth.Auth{}
+	payload := []byte(`{
+		"request": {
+			"tools": [
+				{
+					"function_declarations": [
+						{
+							"name": "tool_1",
+							"parametersJsonSchema": {
+								"$schema": "http://json-schema.org/draft-07/schema#",
+								"$id": "root-schema",
+								"type": "object",
+								"properties": {
+									"$id": {"type": "string"},
+									"arg": {
+										"type": "object",
+										"prefill": "hello",
+										"properties": {
+											"mode": {
+												"type": "string",
+												"enum": ["a", "b"],
+												"enumTitles": ["A", "B"]
+											}
+										}
+									}
+								},
+								"patternProperties": {
+									"^x-": {"type": "string"}
+								}
+							}
+						}
+					]
+				}
+			]
+		}
+	}`)
+
+	req, err := executor.buildRequest(context.Background(), auth, "token", modelName, payload, false, "", "https://example.com")
+	if err != nil {
+		t.Fatalf("buildRequest error: %v", err)
+	}
+
+	raw, err := io.ReadAll(req.Body)
+	if err != nil {
+		t.Fatalf("read request body error: %v", err)
+	}
+
+	var body map[string]any
+	if err := json.Unmarshal(raw, &body); err != nil {
+		t.Fatalf("unmarshal request body error: %v, body=%s", err, string(raw))
+	}
+	return body
+}
+
+func extractFirstFunctionDeclaration(t *testing.T, body map[string]any) map[string]any {
+	t.Helper()
+
+	request, ok := body["request"].(map[string]any)
+	if !ok {
+		t.Fatalf("request missing or invalid type")
+	}
+	tools, ok := request["tools"].([]any)
+	if !ok || len(tools) == 0 {
+		t.Fatalf("tools missing or empty")
+	}
+	tool, ok := tools[0].(map[string]any)
+	if !ok {
+		t.Fatalf("first tool invalid type")
+	}
+	decls, ok := tool["function_declarations"].([]any)
+	if !ok || len(decls) == 0 {
+		t.Fatalf("function_declarations missing or empty")
+	}
+	decl, ok := decls[0].(map[string]any)
+	if !ok {
+		t.Fatalf("first function declaration invalid type")
+	}
+	return decl
+}
+
+func assertSchemaSanitizedAndPropertyPreserved(t *testing.T, params map[string]any) {
+	t.Helper()
+
+	if _, ok := params["$id"]; ok {
+		t.Fatalf("root $id should be removed from schema")
+	}
+	if _, ok := params["patternProperties"]; ok {
+		t.Fatalf("patternProperties should be removed from schema")
+	}
+
+	props, ok := params["properties"].(map[string]any)
+	if !ok {
+		t.Fatalf("properties missing or invalid type")
+	}
+	if _, ok := props["$id"]; !ok {
+		t.Fatalf("property named $id should be preserved")
+	}
+
+	arg, ok := props["arg"].(map[string]any)
+	if !ok {
+		t.Fatalf("arg property missing or invalid type")
+	}
+	if _, ok := arg["prefill"]; ok {
+		t.Fatalf("prefill should be removed from nested schema")
+	}
+
+	argProps, ok := arg["properties"].(map[string]any)
+	if !ok {
+		t.Fatalf("arg.properties missing or invalid type")
+	}
+	mode, ok := argProps["mode"].(map[string]any)
+	if !ok {
+		t.Fatalf("mode property missing or invalid type")
+	}
+	if _, ok := mode["enumTitles"]; ok {
+		t.Fatalf("enumTitles should be removed from nested schema")
+	}
+}

internal/runtime/executor/codex_executor.go

@@ -28,8 +28,8 @@ import (
 )
 
 const (
-	codexClientVersion = "0.98.0"
-	codexUserAgent     = "codex_cli_rs/0.98.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464"
+	codexClientVersion = "0.101.0"
+	codexUserAgent     = "codex_cli_rs/0.101.0 (Mac OS 26.0.1; arm64) Apple_Terminal/464"
 )
 
 var dataTag = []byte("data:")

internal/runtime/executor/gemini_cli_executor.go

@@ -899,8 +899,7 @@ func parseRetryDelay(errorBody []byte) (*time.Duration, error) {
 		if matches := re.FindStringSubmatch(message); len(matches) > 1 {
 			seconds, err := strconv.Atoi(matches[1])
 			if err == nil {
-				duration := time.Duration(seconds) * time.Second
-				return &duration, nil
+				return new(time.Duration(seconds) * time.Second), nil
 			}
 		}
 	}

internal/runtime/executor/qwen_executor.go

@@ -22,9 +22,7 @@ import (
 )
 
 const (
-	qwenUserAgent           = "google-api-nodejs-client/9.15.1"
-	qwenXGoogAPIClient      = "gl-node/22.17.0"
-	qwenClientMetadataValue = "ideType=IDE_UNSPECIFIED,platform=PLATFORM_UNSPECIFIED,pluginType=GEMINI"
+	qwenUserAgent = "QwenCode/0.10.3 (darwin; arm64)"
 )
 
 // QwenExecutor is a stateless executor for Qwen Code using OpenAI-compatible chat completions.
@@ -344,8 +342,18 @@ func applyQwenHeaders(r *http.Request, token string, stream bool) {
 	r.Header.Set("Content-Type", "application/json")
 	r.Header.Set("Authorization", "Bearer "+token)
 	r.Header.Set("User-Agent", qwenUserAgent)
-	r.Header.Set("X-Goog-Api-Client", qwenXGoogAPIClient)
-	r.Header.Set("Client-Metadata", qwenClientMetadataValue)
+	r.Header.Set("X-Dashscope-Useragent", qwenUserAgent)
+	r.Header.Set("X-Stainless-Runtime-Version", "v22.17.0")
+	r.Header.Set("Sec-Fetch-Mode", "cors")
+	r.Header.Set("X-Stainless-Lang", "js")
+	r.Header.Set("X-Stainless-Arch", "arm64")
+	r.Header.Set("X-Stainless-Package-Version", "5.11.0")
+	r.Header.Set("X-Dashscope-Cachecontrol", "enable")
+	r.Header.Set("X-Stainless-Retry-Count", "0")
+	r.Header.Set("X-Stainless-Os", "MacOS")
+	r.Header.Set("X-Dashscope-Authtype", "qwen-oauth")
+	r.Header.Set("X-Stainless-Runtime", "node")
+
 	if stream {
 		r.Header.Set("Accept", "text/event-stream")
 		return

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -344,7 +344,8 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Inject interleaved thinking hint when both tools and thinking are active
 	hasTools := toolDeclCount > 0
 	thinkingResult := gjson.GetBytes(rawJSON, "thinking")
-	hasThinking := thinkingResult.Exists() && thinkingResult.IsObject() && thinkingResult.Get("type").String() == "enabled"
+	thinkingType := thinkingResult.Get("type").String()
+	hasThinking := thinkingResult.Exists() && thinkingResult.IsObject() && (thinkingType == "enabled" || thinkingType == "adaptive")
 	isClaudeThinking := util.IsClaudeThinkingModel(modelName)
 
 	if hasTools && hasThinking && isClaudeThinking {
@@ -377,12 +378,18 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); enableThoughtTranslate && t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

internal/translator/codex/claude/codex_claude_request.go

@@ -222,6 +222,10 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 					reasoningEffort = effort
 				}
 			}
+		case "adaptive":
+			// Claude adaptive means "enable with max capacity"; keep it as highest level
+			// and let ApplyThinking normalize per target model capability.
+			reasoningEffort = string(thinking.LevelXHigh)
 		case "disabled":
 			if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 				reasoningEffort = effort

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

@@ -90,6 +90,9 @@ func ConvertCodexResponseToOpenAI(_ context.Context, modelName string, originalR
 		if inputTokensResult := usageResult.Get("input_tokens"); inputTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokensResult.Int())
 		}
+		if cachedTokensResult := usageResult.Get("input_tokens_details.cached_tokens"); cachedTokensResult.Exists() {
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokensResult.Int())
+		}
 		if reasoningTokensResult := usageResult.Get("output_tokens_details.reasoning_tokens"); reasoningTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", reasoningTokensResult.Int())
 		}
@@ -205,6 +208,9 @@ func ConvertCodexResponseToOpenAINonStream(_ context.Context, _ string, original
 		if inputTokensResult := usageResult.Get("input_tokens"); inputTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.prompt_tokens", inputTokensResult.Int())
 		}
+		if cachedTokensResult := usageResult.Get("input_tokens_details.cached_tokens"); cachedTokensResult.Exists() {
+			template, _ = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokensResult.Int())
+		}
 		if reasoningTokensResult := usageResult.Get("output_tokens_details.reasoning_tokens"); reasoningTokensResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", reasoningTokensResult.Int())
 		}

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

@@ -27,6 +27,9 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
 
+	// Delete the user field as it is not supported by the Codex upstream.
+	rawJSON, _ = sjson.DeleteBytes(rawJSON, "user")
+
 	// Convert role "system" to "developer" in input array to comply with Codex API requirements.
 	rawJSON = convertSystemRoleToDeveloper(rawJSON)
 

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

@@ -263,3 +263,20 @@ func TestConvertSystemRoleToDeveloper_AssistantRole(t *testing.T) {
 		t.Errorf("Expected third role 'assistant', got '%s'", thirdRole.String())
 	}
 }
+
+func TestUserFieldDeletion(t *testing.T) {  
+	inputJSON := []byte(`{  
+		"model": "gpt-5.2",  
+		"user": "test-user",  
+		"input": [{"role": "user", "content": "Hello"}]  
+	}`)  
+	  
+	output := ConvertOpenAIResponsesRequestToCodex("gpt-5.2", inputJSON, false)  
+	outputStr := string(output)  
+	  
+	// Verify user field is deleted  
+	userField := gjson.Get(outputStr, "user")  
+	if userField.Exists() {
+		t.Errorf("user field should be deleted, but it was found with value: %s", userField.Raw)
+	}
+}

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

@@ -173,12 +173,18 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

internal/translator/gemini/claude/gemini_claude_request.go

@@ -154,12 +154,18 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
 	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		if t.Get("type").String() == "enabled" {
+		switch t.Get("type").String() {
+		case "enabled":
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
 				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 			}
+		case "adaptive":
+			// Keep adaptive as a high level sentinel; ApplyThinking resolves it
+			// to model-specific max capability.
+			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingLevel", "high")
+			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
 		}
 	}
 	if v := gjson.GetBytes(rawJSON, "temperature"); v.Exists() && v.Type == gjson.Number {

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

@@ -117,19 +117,29 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 			switch itemType {
 			case "message":
 				if strings.EqualFold(itemRole, "system") {
-					if contentArray := item.Get("content"); contentArray.Exists() && contentArray.IsArray() {
-						var builder strings.Builder
-						contentArray.ForEach(func(_, contentItem gjson.Result) bool {
-							text := contentItem.Get("text").String()
-							if builder.Len() > 0 && text != "" {
-								builder.WriteByte('\n')
-							}
-							builder.WriteString(text)
-							return true
-						})
-						if !gjson.Get(out, "system_instruction").Exists() {
-							systemInstr := `{"parts":[{"text":""}]}`
-							systemInstr, _ = sjson.Set(systemInstr, "parts.0.text", builder.String())
+					if contentArray := item.Get("content"); contentArray.Exists() {
+						systemInstr := ""
+						if systemInstructionResult := gjson.Get(out, "system_instruction"); systemInstructionResult.Exists() {
+							systemInstr = systemInstructionResult.Raw
+						} else {
+							systemInstr = `{"parts":[]}`
+						}
+
+						if contentArray.IsArray() {
+							contentArray.ForEach(func(_, contentItem gjson.Result) bool {
+								part := `{"text":""}`
+								text := contentItem.Get("text").String()
+								part, _ = sjson.Set(part, "text", text)
+								systemInstr, _ = sjson.SetRaw(systemInstr, "parts.-1", part)
+								return true
+							})
+						} else if contentArray.Type == gjson.String {
+							part := `{"text":""}`
+							part, _ = sjson.Set(part, "text", contentArray.String())
+							systemInstr, _ = sjson.SetRaw(systemInstr, "parts.-1", part)
+						}
+
+						if systemInstr != `{"parts":[]}` {
 							out, _ = sjson.SetRaw(out, "system_instruction", systemInstr)
 						}
 					}
@@ -236,8 +246,22 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 					})
 
 					flush()
-				}
+				} else if contentArray.Type == gjson.String {
+					effRole := "user"
+					if itemRole != "" {
+						switch strings.ToLower(itemRole) {
+						case "assistant", "model":
+							effRole = "model"
+						default:
+							effRole = strings.ToLower(itemRole)
+						}
+					}
 
+					one := `{"role":"","parts":[{"text":""}]}`
+					one, _ = sjson.Set(one, "role", effRole)
+					one, _ = sjson.Set(one, "parts.0.text", contentArray.String())
+					out, _ = sjson.SetRaw(out, "contents.-1", one)
+				}
 			case "function_call":
 				// Handle function calls - convert to model message with functionCall
 				name := item.Get("name").String()

internal/translator/openai/claude/openai_claude_request.go

@@ -75,6 +75,10 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 						out, _ = sjson.Set(out, "reasoning_effort", effort)
 					}
 				}
+			case "adaptive":
+				// Claude adaptive means "enable with max capacity"; keep it as highest level
+				// and let ApplyThinking normalize per target model capability.
+				out, _ = sjson.Set(out, "reasoning_effort", string(thinking.LevelXHigh))
 			case "disabled":
 				if effort, ok := thinking.ConvertBudgetToLevel(0); ok && effort != "" {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)

internal/util/gemini_schema.go

@@ -428,8 +428,9 @@ func flattenTypeArrays(jsonStr string) string {
 
 func removeUnsupportedKeywords(jsonStr string) string {
 	keywords := append(unsupportedConstraints,
-		"$schema", "$defs", "definitions", "const", "$ref", "additionalProperties",
-		"propertyNames", // Gemini doesn't support property name validation
+		"$schema", "$defs", "definitions", "const", "$ref", "$id", "additionalProperties",
+		"propertyNames", "patternProperties", // Gemini doesn't support these schema keywords
+		"enumTitles", "prefill", // Claude/OpenCode schema metadata fields unsupported by Gemini
 	)
 
 	deletePaths := make([]string, 0)

internal/util/gemini_schema_test.go

@@ -870,6 +870,57 @@ func TestCleanJSONSchemaForAntigravity_BooleanEnumToString(t *testing.T) {
 	}
 }
 
+func TestCleanJSONSchemaForGemini_RemovesGeminiUnsupportedMetadataFields(t *testing.T) {
+	input := `{
+		"$schema": "http://json-schema.org/draft-07/schema#",
+		"$id": "root-schema",
+		"type": "object",
+		"properties": {
+			"payload": {
+				"type": "object",
+				"prefill": "hello",
+				"properties": {
+					"mode": {
+						"type": "string",
+						"enum": ["a", "b"],
+						"enumTitles": ["A", "B"]
+					}
+				},
+				"patternProperties": {
+					"^x-": {"type": "string"}
+				}
+			},
+			"$id": {
+				"type": "string",
+				"description": "property name should not be removed"
+			}
+		}
+	}`
+
+	expected := `{
+		"type": "object",
+		"properties": {
+			"payload": {
+				"type": "object",
+				"properties": {
+					"mode": {
+						"type": "string",
+						"enum": ["a", "b"],
+						"description": "Allowed: a, b"
+					}
+				}
+			},
+			"$id": {
+				"type": "string",
+				"description": "property name should not be removed"
+			}
+		}
+	}`
+
+	result := CleanJSONSchemaForGemini(input)
+	compareJSON(t, expected, result)
+}
+
 func TestRemoveExtensionFields(t *testing.T) {
 	tests := []struct {
 		name     string

internal/watcher/diff/config_diff.go

@@ -184,6 +184,9 @@ func BuildConfigChangeDetails(oldCfg, newCfg *config.Config) []string {
 			if strings.TrimSpace(o.Prefix) != strings.TrimSpace(n.Prefix) {
 				changes = append(changes, fmt.Sprintf("codex[%d].prefix: %s -> %s", i, strings.TrimSpace(o.Prefix), strings.TrimSpace(n.Prefix)))
 			}
+			if o.Websockets != n.Websockets {
+				changes = append(changes, fmt.Sprintf("codex[%d].websockets: %t -> %t", i, o.Websockets, n.Websockets))
+			}
 			if strings.TrimSpace(o.APIKey) != strings.TrimSpace(n.APIKey) {
 				changes = append(changes, fmt.Sprintf("codex[%d].api-key: updated", i))
 			}

internal/watcher/synthesizer/config.go

@@ -160,6 +160,9 @@ func (s *ConfigSynthesizer) synthesizeCodexKeys(ctx *SynthesisContext) []*coreau
 		if ck.BaseURL != "" {
 			attrs["base_url"] = ck.BaseURL
 		}
+		if ck.Websockets {
+			attrs["websockets"] = "true"
+		}
 		if hash := diff.ComputeCodexModelsHash(ck.Models); hash != "" {
 			attrs["models_hash"] = hash
 		}

internal/watcher/synthesizer/config_test.go

@@ -231,10 +231,11 @@ func TestConfigSynthesizer_CodexKeys(t *testing.T) {
 		Config: &config.Config{
 			CodexKey: []config.CodexKey{
 				{
-					APIKey:   "codex-key-123",
-					Prefix:   "dev",
-					BaseURL:  "https://api.openai.com",
-					ProxyURL: "http://proxy.local",
+					APIKey:     "codex-key-123",
+					Prefix:     "dev",
+					BaseURL:    "https://api.openai.com",
+					ProxyURL:   "http://proxy.local",
+					Websockets: true,
 				},
 			},
 		},
@@ -259,6 +260,9 @@ func TestConfigSynthesizer_CodexKeys(t *testing.T) {
 	if auths[0].ProxyURL != "http://proxy.local" {
 		t.Errorf("expected proxy_url http://proxy.local, got %s", auths[0].ProxyURL)
 	}
+	if auths[0].Attributes["websockets"] != "true" {
+		t.Errorf("expected websockets=true, got %s", auths[0].Attributes["websockets"])
+	}
 }
 
 func TestConfigSynthesizer_CodexKeys_SkipsEmptyAndHeaders(t *testing.T) {

internal/watcher/synthesizer/file.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"time"
 
@@ -92,6 +93,9 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			status = coreauth.StatusDisabled
 		}
 
+		// Read per-account excluded models from the OAuth JSON file
+		perAccountExcluded := extractExcludedModelsFromMetadata(metadata)
+
 		a := &coreauth.Auth{
 			ID:       id,
 			Provider: provider,
@@ -108,11 +112,23 @@ func (s *FileSynthesizer) Synthesize(ctx *SynthesisContext) ([]*coreauth.Auth, e
 			CreatedAt: now,
 			UpdatedAt: now,
 		}
-		ApplyAuthExcludedModelsMeta(a, cfg, nil, "oauth")
+		// Read priority from auth file
+		if rawPriority, ok := metadata["priority"]; ok {
+			switch v := rawPriority.(type) {
+			case float64:
+				a.Attributes["priority"] = strconv.Itoa(int(v))
+			case string:
+				priority := strings.TrimSpace(v)
+				if _, errAtoi := strconv.Atoi(priority); errAtoi == nil {
+					a.Attributes["priority"] = priority
+				}
+			}
+		}
+		ApplyAuthExcludedModelsMeta(a, cfg, perAccountExcluded, "oauth")
 		if provider == "gemini-cli" {
 			if virtuals := SynthesizeGeminiVirtualAuths(a, metadata, now); len(virtuals) > 0 {
 				for _, v := range virtuals {
-					ApplyAuthExcludedModelsMeta(v, cfg, nil, "oauth")
+					ApplyAuthExcludedModelsMeta(v, cfg, perAccountExcluded, "oauth")
 				}
 				out = append(out, a)
 				out = append(out, virtuals...)
@@ -167,6 +183,10 @@ func SynthesizeGeminiVirtualAuths(primary *coreauth.Auth, metadata map[string]an
 		if authPath != "" {
 			attrs["path"] = authPath
 		}
+		// Propagate priority from primary auth to virtual auths
+		if priorityVal, hasPriority := primary.Attributes["priority"]; hasPriority && priorityVal != "" {
+			attrs["priority"] = priorityVal
+		}
 		metadataCopy := map[string]any{
 			"email":             email,
 			"project_id":        projectID,
@@ -239,3 +259,40 @@ func buildGeminiVirtualID(baseID, projectID string) string {
 	replacer := strings.NewReplacer("/", "_", "\\", "_", " ", "_")
 	return fmt.Sprintf("%s::%s", baseID, replacer.Replace(project))
 }
+
+// extractExcludedModelsFromMetadata reads per-account excluded models from the OAuth JSON metadata.
+// Supports both "excluded_models" and "excluded-models" keys, and accepts both []string and []interface{}.
+func extractExcludedModelsFromMetadata(metadata map[string]any) []string {
+	if metadata == nil {
+		return nil
+	}
+	// Try both key formats
+	raw, ok := metadata["excluded_models"]
+	if !ok {
+		raw, ok = metadata["excluded-models"]
+	}
+	if !ok || raw == nil {
+		return nil
+	}
+	var stringSlice []string
+	switch v := raw.(type) {
+	case []string:
+		stringSlice = v
+	case []interface{}:
+		stringSlice = make([]string, 0, len(v))
+		for _, item := range v {
+			if s, ok := item.(string); ok {
+				stringSlice = append(stringSlice, s)
+			}
+		}
+	default:
+		return nil
+	}
+	result := make([]string, 0, len(stringSlice))
+	for _, s := range stringSlice {
+		if trimmed := strings.TrimSpace(s); trimmed != "" {
+			result = append(result, trimmed)
+		}
+	}
+	return result
+}

internal/watcher/synthesizer/file_test.go

@@ -297,6 +297,117 @@ func TestFileSynthesizer_Synthesize_PrefixValidation(t *testing.T) {
 	}
 }
 
+func TestFileSynthesizer_Synthesize_PriorityParsing(t *testing.T) {
+	tests := []struct {
+		name     string
+		priority any
+		want     string
+		hasValue bool
+	}{
+		{
+			name:     "string with spaces",
+			priority: " 10 ",
+			want:     "10",
+			hasValue: true,
+		},
+		{
+			name:     "number",
+			priority: 8,
+			want:     "8",
+			hasValue: true,
+		},
+		{
+			name:     "invalid string",
+			priority: "1x",
+			hasValue: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			tempDir := t.TempDir()
+			authData := map[string]any{
+				"type":     "claude",
+				"priority": tt.priority,
+			}
+			data, _ := json.Marshal(authData)
+			errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+			if errWriteFile != nil {
+				t.Fatalf("failed to write auth file: %v", errWriteFile)
+			}
+
+			synth := NewFileSynthesizer()
+			ctx := &SynthesisContext{
+				Config:      &config.Config{},
+				AuthDir:     tempDir,
+				Now:         time.Now(),
+				IDGenerator: NewStableIDGenerator(),
+			}
+
+			auths, errSynthesize := synth.Synthesize(ctx)
+			if errSynthesize != nil {
+				t.Fatalf("unexpected error: %v", errSynthesize)
+			}
+			if len(auths) != 1 {
+				t.Fatalf("expected 1 auth, got %d", len(auths))
+			}
+
+			value, ok := auths[0].Attributes["priority"]
+			if tt.hasValue {
+				if !ok {
+					t.Fatal("expected priority attribute to be set")
+				}
+				if value != tt.want {
+					t.Fatalf("expected priority %q, got %q", tt.want, value)
+				}
+				return
+			}
+			if ok {
+				t.Fatalf("expected priority attribute to be absent, got %q", value)
+			}
+		})
+	}
+}
+
+func TestFileSynthesizer_Synthesize_OAuthExcludedModelsMerged(t *testing.T) {
+	tempDir := t.TempDir()
+	authData := map[string]any{
+		"type":            "claude",
+		"excluded_models": []string{"custom-model", "MODEL-B"},
+	}
+	data, _ := json.Marshal(authData)
+	errWriteFile := os.WriteFile(filepath.Join(tempDir, "auth.json"), data, 0644)
+	if errWriteFile != nil {
+		t.Fatalf("failed to write auth file: %v", errWriteFile)
+	}
+
+	synth := NewFileSynthesizer()
+	ctx := &SynthesisContext{
+		Config: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"claude": {"shared", "model-b"},
+			},
+		},
+		AuthDir:     tempDir,
+		Now:         time.Now(),
+		IDGenerator: NewStableIDGenerator(),
+	}
+
+	auths, errSynthesize := synth.Synthesize(ctx)
+	if errSynthesize != nil {
+		t.Fatalf("unexpected error: %v", errSynthesize)
+	}
+	if len(auths) != 1 {
+		t.Fatalf("expected 1 auth, got %d", len(auths))
+	}
+
+	got := auths[0].Attributes["excluded_models"]
+	want := "custom-model,model-b,shared"
+	if got != want {
+		t.Fatalf("expected excluded_models %q, got %q", want, got)
+	}
+}
+
 func TestSynthesizeGeminiVirtualAuths_NilInputs(t *testing.T) {
 	now := time.Now()
 
@@ -533,6 +644,7 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		"type":       "gemini",
 		"email":      "multi@example.com",
 		"project_id": "project-a, project-b, project-c",
+		"priority":   " 10 ",
 	}
 	data, _ := json.Marshal(authData)
 	err := os.WriteFile(filepath.Join(tempDir, "gemini-multi.json"), data, 0644)
@@ -565,6 +677,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 	if primary.Status != coreauth.StatusDisabled {
 		t.Errorf("expected primary status disabled, got %s", primary.Status)
 	}
+	if gotPriority := primary.Attributes["priority"]; gotPriority != "10" {
+		t.Errorf("expected primary priority 10, got %q", gotPriority)
+	}
 
 	// Remaining auths should be virtuals
 	for i := 1; i < 4; i++ {
@@ -575,6 +690,9 @@ func TestFileSynthesizer_Synthesize_MultiProjectGemini(t *testing.T) {
 		if v.Attributes["gemini_virtual_parent"] != primary.ID {
 			t.Errorf("expected virtual %d parent to be %s, got %s", i, primary.ID, v.Attributes["gemini_virtual_parent"])
 		}
+		if gotPriority := v.Attributes["priority"]; gotPriority != "10" {
+			t.Errorf("expected virtual %d priority 10, got %q", i, gotPriority)
+		}
 	}
 }
 

internal/watcher/synthesizer/helpers.go

@@ -53,6 +53,8 @@ func (g *StableIDGenerator) Next(kind string, parts ...string) (string, string)
 
 // ApplyAuthExcludedModelsMeta applies excluded models metadata to an auth entry.
 // It computes a hash of excluded models and sets the auth_kind attribute.
+// For OAuth entries, perKey (from the JSON file's excluded-models field) is merged
+// with the global oauth-excluded-models config for the provider.
 func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey []string, authKind string) {
 	if auth == nil || cfg == nil {
 		return
@@ -72,9 +74,13 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	}
 	if authKindKey == "apikey" {
 		add(perKey)
-	} else if cfg.OAuthExcludedModels != nil {
-		providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
-		add(cfg.OAuthExcludedModels[providerKey])
+	} else {
+		// For OAuth: merge per-account excluded models with global provider-level exclusions
+		add(perKey)
+		if cfg.OAuthExcludedModels != nil {
+			providerKey := strings.ToLower(strings.TrimSpace(auth.Provider))
+			add(cfg.OAuthExcludedModels[providerKey])
+		}
 	}
 	combined := make([]string, 0, len(seen))
 	for k := range seen {
@@ -88,6 +94,10 @@ func ApplyAuthExcludedModelsMeta(auth *coreauth.Auth, cfg *config.Config, perKey
 	if hash != "" {
 		auth.Attributes["excluded_models_hash"] = hash
 	}
+	// Store the combined excluded models list so that routing can read it at runtime
+	if len(combined) > 0 {
+		auth.Attributes["excluded_models"] = strings.Join(combined, ",")
+	}
 	if authKind != "" {
 		auth.Attributes["auth_kind"] = authKind
 	}

internal/watcher/synthesizer/helpers_test.go

@@ -6,6 +6,7 @@ import (
 	"testing"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/config"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/watcher/diff"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
 )
 
@@ -200,6 +201,30 @@ func TestApplyAuthExcludedModelsMeta(t *testing.T) {
 	}
 }
 
+func TestApplyAuthExcludedModelsMeta_OAuthMergeWritesCombinedModels(t *testing.T) {
+	auth := &coreauth.Auth{
+		Provider:   "claude",
+		Attributes: make(map[string]string),
+	}
+	cfg := &config.Config{
+		OAuthExcludedModels: map[string][]string{
+			"claude": {"global-a", "shared"},
+		},
+	}
+
+	ApplyAuthExcludedModelsMeta(auth, cfg, []string{"per", "SHARED"}, "oauth")
+
+	const wantCombined = "global-a,per,shared"
+	if gotCombined := auth.Attributes["excluded_models"]; gotCombined != wantCombined {
+		t.Fatalf("expected excluded_models=%q, got %q", wantCombined, gotCombined)
+	}
+
+	expectedHash := diff.ComputeExcludedModelsHash([]string{"global-a", "per", "shared"})
+	if gotHash := auth.Attributes["excluded_models_hash"]; gotHash != expectedHash {
+		t.Fatalf("expected excluded_models_hash=%q, got %q", expectedHash, gotHash)
+	}
+}
+
 func TestAddConfigHeadersToAttrs(t *testing.T) {
 	tests := []struct {
 		name    string

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

@@ -185,8 +185,7 @@ func (h *GeminiCLIAPIHandler) handleInternalGenerateContent(c *gin.Context, rawJ
 func (h *GeminiCLIAPIHandler) forwardCLIStream(c *gin.Context, flusher http.Flusher, alt string, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
 	var keepAliveInterval *time.Duration
 	if alt != "" {
-		disabled := time.Duration(0)
-		keepAliveInterval = &disabled
+		keepAliveInterval = new(time.Duration(0))
 	}
 
 	h.ForwardStream(c, flusher, cancel, data, errs, handlers.StreamForwardOptions{

sdk/api/handlers/gemini/gemini_handlers.go

@@ -300,8 +300,7 @@ func (h *GeminiAPIHandler) handleGenerateContent(c *gin.Context, modelName strin
 func (h *GeminiAPIHandler) forwardGeminiStream(c *gin.Context, flusher http.Flusher, alt string, cancel func(error), data <-chan []byte, errs <-chan *interfaces.ErrorMessage) {
 	var keepAliveInterval *time.Duration
 	if alt != "" {
-		disabled := time.Duration(0)
-		keepAliveInterval = &disabled
+		keepAliveInterval = new(time.Duration(0))
 	}
 
 	h.ForwardStream(c, flusher, cancel, data, errs, handlers.StreamForwardOptions{

sdk/api/handlers/handlers.go

@@ -52,6 +52,45 @@ const (
 	defaultStreamingBootstrapRetries = 0
 )
 
+type pinnedAuthContextKey struct{}
+type selectedAuthCallbackContextKey struct{}
+type executionSessionContextKey struct{}
+
+// WithPinnedAuthID returns a child context that requests execution on a specific auth ID.
+func WithPinnedAuthID(ctx context.Context, authID string) context.Context {
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, pinnedAuthContextKey{}, authID)
+}
+
+// WithSelectedAuthIDCallback returns a child context that receives the selected auth ID.
+func WithSelectedAuthIDCallback(ctx context.Context, callback func(string)) context.Context {
+	if callback == nil {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, selectedAuthCallbackContextKey{}, callback)
+}
+
+// WithExecutionSessionID returns a child context tagged with a long-lived execution session ID.
+func WithExecutionSessionID(ctx context.Context, sessionID string) context.Context {
+	sessionID = strings.TrimSpace(sessionID)
+	if sessionID == "" {
+		return ctx
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, executionSessionContextKey{}, sessionID)
+}
+
 // BuildErrorResponseBody builds an OpenAI-compatible JSON error response body.
 // If errText is already valid JSON, it is returned as-is to preserve upstream error payloads.
 func BuildErrorResponseBody(status int, errText string) []byte {
@@ -152,7 +191,59 @@ func requestExecutionMetadata(ctx context.Context) map[string]any {
 	if key == "" {
 		key = uuid.NewString()
 	}
-	return map[string]any{idempotencyKeyMetadataKey: key}
+
+	meta := map[string]any{idempotencyKeyMetadataKey: key}
+	if pinnedAuthID := pinnedAuthIDFromContext(ctx); pinnedAuthID != "" {
+		meta[coreexecutor.PinnedAuthMetadataKey] = pinnedAuthID
+	}
+	if selectedCallback := selectedAuthIDCallbackFromContext(ctx); selectedCallback != nil {
+		meta[coreexecutor.SelectedAuthCallbackMetadataKey] = selectedCallback
+	}
+	if executionSessionID := executionSessionIDFromContext(ctx); executionSessionID != "" {
+		meta[coreexecutor.ExecutionSessionMetadataKey] = executionSessionID
+	}
+	return meta
+}
+
+func pinnedAuthIDFromContext(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	raw := ctx.Value(pinnedAuthContextKey{})
+	switch v := raw.(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case []byte:
+		return strings.TrimSpace(string(v))
+	default:
+		return ""
+	}
+}
+
+func selectedAuthIDCallbackFromContext(ctx context.Context) func(string) {
+	if ctx == nil {
+		return nil
+	}
+	raw := ctx.Value(selectedAuthCallbackContextKey{})
+	if callback, ok := raw.(func(string)); ok && callback != nil {
+		return callback
+	}
+	return nil
+}
+
+func executionSessionIDFromContext(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	raw := ctx.Value(executionSessionContextKey{})
+	switch v := raw.(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case []byte:
+		return strings.TrimSpace(string(v))
+	default:
+		return ""
+	}
 }
 
 // BaseAPIHandler contains the handlers for API endpoints.

sdk/api/handlers/handlers_stream_bootstrap_test.go

@@ -122,6 +122,82 @@ func (e *payloadThenErrorStreamExecutor) Calls() int {
 	return e.calls
 }
 
+type authAwareStreamExecutor struct {
+	mu      sync.Mutex
+	calls   int
+	authIDs []string
+}
+
+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) {
+	_ = ctx
+	_ = req
+	_ = opts
+	ch := make(chan coreexecutor.StreamChunk, 1)
+
+	authID := ""
+	if auth != nil {
+		authID = auth.ID
+	}
+
+	e.mu.Lock()
+	e.calls++
+	e.authIDs = append(e.authIDs, authID)
+	e.mu.Unlock()
+
+	if authID == "auth1" {
+		ch <- coreexecutor.StreamChunk{
+			Err: &coreauth.Error{
+				Code:       "unauthorized",
+				Message:    "unauthorized",
+				Retryable:  false,
+				HTTPStatus: http.StatusUnauthorized,
+			},
+		}
+		close(ch)
+		return ch, nil
+	}
+
+	ch <- coreexecutor.StreamChunk{Payload: []byte("ok")}
+	close(ch)
+	return ch, nil
+}
+
+func (e *authAwareStreamExecutor) Refresh(ctx context.Context, auth *coreauth.Auth) (*coreauth.Auth, error) {
+	return auth, nil
+}
+
+func (e *authAwareStreamExecutor) 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 *authAwareStreamExecutor) 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) Calls() int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.calls
+}
+
+func (e *authAwareStreamExecutor) AuthIDs() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.authIDs))
+	copy(out, e.authIDs)
+	return out
+}
+
 func TestExecuteStreamWithAuthManager_RetriesBeforeFirstByte(t *testing.T) {
 	executor := &failOnceStreamExecutor{}
 	manager := coreauth.NewManager(nil, nil, nil)
@@ -252,3 +328,128 @@ func TestExecuteStreamWithAuthManager_DoesNotRetryAfterFirstByte(t *testing.T) {
 		t.Fatalf("expected 1 stream attempt, got %d", executor.Calls())
 	}
 }
+
+func TestExecuteStreamWithAuthManager_PinnedAuthKeepsSameUpstream(t *testing.T) {
+	executor := &authAwareStreamExecutor{}
+	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)
+	ctx := WithPinnedAuthID(context.Background(), "auth1")
+	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "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...)
+	}
+
+	var gotErr error
+	for msg := range errChan {
+		if msg != nil && msg.Error != nil {
+			gotErr = msg.Error
+		}
+	}
+
+	if len(got) != 0 {
+		t.Fatalf("expected empty payload, got %q", string(got))
+	}
+	if gotErr == nil {
+		t.Fatalf("expected terminal error, got nil")
+	}
+	authIDs := executor.AuthIDs()
+	if len(authIDs) == 0 {
+		t.Fatalf("expected at least one upstream attempt")
+	}
+	for _, authID := range authIDs {
+		if authID != "auth1" {
+			t.Fatalf("expected all attempts on auth1, got sequence %v", authIDs)
+		}
+	}
+}
+
+func TestExecuteStreamWithAuthManager_SelectedAuthCallbackReceivesAuthID(t *testing.T) {
+	executor := &authAwareStreamExecutor{}
+	manager := coreauth.NewManager(nil, nil, nil)
+	manager.RegisterExecutor(executor)
+
+	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(auth2.ID, auth2.Provider, []*registry.ModelInfo{{ID: "test-model"}})
+	t.Cleanup(func() {
+		registry.GetGlobalRegistry().UnregisterClient(auth2.ID)
+	})
+
+	handler := NewBaseAPIHandlers(&sdkconfig.SDKConfig{
+		Streaming: sdkconfig.StreamingConfig{
+			BootstrapRetries: 0,
+		},
+	}, manager)
+
+	selectedAuthID := ""
+	ctx := WithSelectedAuthIDCallback(context.Background(), func(authID string) {
+		selectedAuthID = authID
+	})
+	dataChan, errChan := handler.ExecuteStreamWithAuthManager(ctx, "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 selectedAuthID != "auth2" {
+		t.Fatalf("selectedAuthID = %q, want %q", selectedAuthID, "auth2")
+	}
+}

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -0,0 +1,662 @@
+package openai
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
+	"github.com/gorilla/websocket"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/api/handlers"
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+	log "github.com/sirupsen/logrus"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+)
+
+const (
+	wsRequestTypeCreate  = "response.create"
+	wsRequestTypeAppend  = "response.append"
+	wsEventTypeError     = "error"
+	wsEventTypeCompleted = "response.completed"
+	wsEventTypeDone      = "response.done"
+	wsDoneMarker         = "[DONE]"
+	wsTurnStateHeader    = "x-codex-turn-state"
+	wsRequestBodyKey     = "REQUEST_BODY_OVERRIDE"
+	wsPayloadLogMaxSize  = 2048
+)
+
+var responsesWebsocketUpgrader = websocket.Upgrader{
+	ReadBufferSize:  4096,
+	WriteBufferSize: 4096,
+	CheckOrigin: func(r *http.Request) bool {
+		return true
+	},
+}
+
+// ResponsesWebsocket handles websocket requests for /v1/responses.
+// It accepts `response.create` and `response.append` requests and streams
+// response events back as JSON websocket text messages.
+func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
+	conn, err := responsesWebsocketUpgrader.Upgrade(c.Writer, c.Request, websocketUpgradeHeaders(c.Request))
+	if err != nil {
+		return
+	}
+	passthroughSessionID := uuid.NewString()
+	clientRemoteAddr := ""
+	if c != nil && c.Request != nil {
+		clientRemoteAddr = strings.TrimSpace(c.Request.RemoteAddr)
+	}
+	log.Infof("responses websocket: client connected id=%s remote=%s", passthroughSessionID, clientRemoteAddr)
+	var wsTerminateErr error
+	var wsBodyLog strings.Builder
+	defer func() {
+		if wsTerminateErr != nil {
+			// log.Infof("responses websocket: session closing id=%s reason=%v", passthroughSessionID, wsTerminateErr)
+		} else {
+			log.Infof("responses websocket: session closing id=%s", passthroughSessionID)
+		}
+		if h != nil && h.AuthManager != nil {
+			h.AuthManager.CloseExecutionSession(passthroughSessionID)
+			log.Infof("responses websocket: upstream execution session closed id=%s", passthroughSessionID)
+		}
+		setWebsocketRequestBody(c, wsBodyLog.String())
+		if errClose := conn.Close(); errClose != nil {
+			log.Warnf("responses websocket: close connection error: %v", errClose)
+		}
+	}()
+
+	var lastRequest []byte
+	lastResponseOutput := []byte("[]")
+	pinnedAuthID := ""
+
+	for {
+		msgType, payload, errReadMessage := conn.ReadMessage()
+		if errReadMessage != nil {
+			wsTerminateErr = errReadMessage
+			appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errReadMessage.Error()))
+			if websocket.IsCloseError(errReadMessage, websocket.CloseNormalClosure, websocket.CloseGoingAway, websocket.CloseNoStatusReceived) {
+				log.Infof("responses websocket: client disconnected id=%s error=%v", passthroughSessionID, errReadMessage)
+			} else {
+				// log.Warnf("responses websocket: read message failed id=%s error=%v", passthroughSessionID, errReadMessage)
+			}
+			return
+		}
+		if msgType != websocket.TextMessage && msgType != websocket.BinaryMessage {
+			continue
+		}
+		// log.Infof(
+		// 	"responses websocket: downstream_in id=%s type=%d event=%s payload=%s",
+		// 	passthroughSessionID,
+		// 	msgType,
+		// 	websocketPayloadEventType(payload),
+		// 	websocketPayloadPreview(payload),
+		// )
+		appendWebsocketEvent(&wsBodyLog, "request", payload)
+
+		allowIncrementalInputWithPreviousResponseID := websocketUpstreamSupportsIncrementalInput(nil, nil)
+		if pinnedAuthID != "" && h != nil && h.AuthManager != nil {
+			if pinnedAuth, ok := h.AuthManager.GetByID(pinnedAuthID); ok && pinnedAuth != nil {
+				allowIncrementalInputWithPreviousResponseID = websocketUpstreamSupportsIncrementalInput(pinnedAuth.Attributes, pinnedAuth.Metadata)
+			}
+		}
+
+		var requestJSON []byte
+		var updatedLastRequest []byte
+		var errMsg *interfaces.ErrorMessage
+		requestJSON, updatedLastRequest, errMsg = normalizeResponsesWebsocketRequestWithMode(
+			payload,
+			lastRequest,
+			lastResponseOutput,
+			allowIncrementalInputWithPreviousResponseID,
+		)
+		if errMsg != nil {
+			h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+			markAPIResponseTimestamp(c)
+			errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+			appendWebsocketEvent(&wsBodyLog, "response", errorPayload)
+			log.Infof(
+				"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+				passthroughSessionID,
+				websocket.TextMessage,
+				websocketPayloadEventType(errorPayload),
+				websocketPayloadPreview(errorPayload),
+			)
+			if errWrite != nil {
+				log.Warnf(
+					"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+					passthroughSessionID,
+					websocketPayloadEventType(errorPayload),
+					errWrite,
+				)
+				return
+			}
+			continue
+		}
+		lastRequest = updatedLastRequest
+
+		modelName := gjson.GetBytes(requestJSON, "model").String()
+		cliCtx, cliCancel := h.GetContextWithCancel(h, c, context.Background())
+		cliCtx = cliproxyexecutor.WithDownstreamWebsocket(cliCtx)
+		cliCtx = handlers.WithExecutionSessionID(cliCtx, passthroughSessionID)
+		if pinnedAuthID != "" {
+			cliCtx = handlers.WithPinnedAuthID(cliCtx, pinnedAuthID)
+		} else {
+			cliCtx = handlers.WithSelectedAuthIDCallback(cliCtx, func(authID string) {
+				pinnedAuthID = strings.TrimSpace(authID)
+			})
+		}
+		dataChan, errChan := h.ExecuteStreamWithAuthManager(cliCtx, h.HandlerType(), modelName, requestJSON, "")
+
+		completedOutput, errForward := h.forwardResponsesWebsocket(c, conn, cliCancel, dataChan, errChan, &wsBodyLog, passthroughSessionID)
+		if errForward != nil {
+			wsTerminateErr = errForward
+			appendWebsocketEvent(&wsBodyLog, "disconnect", []byte(errForward.Error()))
+			log.Warnf("responses websocket: forward failed id=%s error=%v", passthroughSessionID, errForward)
+			return
+		}
+		lastResponseOutput = completedOutput
+	}
+}
+
+func websocketUpgradeHeaders(req *http.Request) http.Header {
+	headers := http.Header{}
+	if req == nil {
+		return headers
+	}
+
+	// Keep the same sticky turn-state across reconnects when provided by the client.
+	turnState := strings.TrimSpace(req.Header.Get(wsTurnStateHeader))
+	if turnState != "" {
+		headers.Set(wsTurnStateHeader, turnState)
+	}
+	return headers
+}
+
+func normalizeResponsesWebsocketRequest(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte) ([]byte, []byte, *interfaces.ErrorMessage) {
+	return normalizeResponsesWebsocketRequestWithMode(rawJSON, lastRequest, lastResponseOutput, true)
+}
+
+func normalizeResponsesWebsocketRequestWithMode(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte, allowIncrementalInputWithPreviousResponseID bool) ([]byte, []byte, *interfaces.ErrorMessage) {
+	requestType := strings.TrimSpace(gjson.GetBytes(rawJSON, "type").String())
+	switch requestType {
+	case wsRequestTypeCreate:
+		// log.Infof("responses websocket: response.create request")
+		if len(lastRequest) == 0 {
+			return normalizeResponseCreateRequest(rawJSON)
+		}
+		return normalizeResponseSubsequentRequest(rawJSON, lastRequest, lastResponseOutput, allowIncrementalInputWithPreviousResponseID)
+	case wsRequestTypeAppend:
+		// log.Infof("responses websocket: response.append request")
+		return normalizeResponseSubsequentRequest(rawJSON, lastRequest, lastResponseOutput, allowIncrementalInputWithPreviousResponseID)
+	default:
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("unsupported websocket request type: %s", requestType),
+		}
+	}
+}
+
+func normalizeResponseCreateRequest(rawJSON []byte) ([]byte, []byte, *interfaces.ErrorMessage) {
+	normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+	if errDelete != nil {
+		normalized = bytes.Clone(rawJSON)
+	}
+	normalized, _ = sjson.SetBytes(normalized, "stream", true)
+	if !gjson.GetBytes(normalized, "input").Exists() {
+		normalized, _ = sjson.SetRawBytes(normalized, "input", []byte("[]"))
+	}
+
+	modelName := strings.TrimSpace(gjson.GetBytes(normalized, "model").String())
+	if modelName == "" {
+		return nil, nil, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("missing model in response.create request"),
+		}
+	}
+	return normalized, bytes.Clone(normalized), nil
+}
+
+func normalizeResponseSubsequentRequest(rawJSON []byte, lastRequest []byte, lastResponseOutput []byte, allowIncrementalInputWithPreviousResponseID bool) ([]byte, []byte, *interfaces.ErrorMessage) {
+	if len(lastRequest) == 0 {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("websocket request received before response.create"),
+		}
+	}
+
+	nextInput := gjson.GetBytes(rawJSON, "input")
+	if !nextInput.Exists() || !nextInput.IsArray() {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("websocket request requires array field: input"),
+		}
+	}
+
+	// Websocket v2 mode uses response.create with previous_response_id + incremental input.
+	// Do not expand it into a full input transcript; upstream expects the incremental payload.
+	if allowIncrementalInputWithPreviousResponseID {
+		if prev := strings.TrimSpace(gjson.GetBytes(rawJSON, "previous_response_id").String()); prev != "" {
+			normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+			if errDelete != nil {
+				normalized = bytes.Clone(rawJSON)
+			}
+			if !gjson.GetBytes(normalized, "model").Exists() {
+				modelName := strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+				if modelName != "" {
+					normalized, _ = sjson.SetBytes(normalized, "model", modelName)
+				}
+			}
+			if !gjson.GetBytes(normalized, "instructions").Exists() {
+				instructions := gjson.GetBytes(lastRequest, "instructions")
+				if instructions.Exists() {
+					normalized, _ = sjson.SetRawBytes(normalized, "instructions", []byte(instructions.Raw))
+				}
+			}
+			normalized, _ = sjson.SetBytes(normalized, "stream", true)
+			return normalized, bytes.Clone(normalized), nil
+		}
+	}
+
+	existingInput := gjson.GetBytes(lastRequest, "input")
+	mergedInput, errMerge := mergeJSONArrayRaw(existingInput.Raw, normalizeJSONArrayRaw(lastResponseOutput))
+	if errMerge != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("invalid previous response output: %w", errMerge),
+		}
+	}
+
+	mergedInput, errMerge = mergeJSONArrayRaw(mergedInput, nextInput.Raw)
+	if errMerge != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("invalid request input: %w", errMerge),
+		}
+	}
+
+	normalized, errDelete := sjson.DeleteBytes(rawJSON, "type")
+	if errDelete != nil {
+		normalized = bytes.Clone(rawJSON)
+	}
+	normalized, _ = sjson.DeleteBytes(normalized, "previous_response_id")
+	var errSet error
+	normalized, errSet = sjson.SetRawBytes(normalized, "input", []byte(mergedInput))
+	if errSet != nil {
+		return nil, lastRequest, &interfaces.ErrorMessage{
+			StatusCode: http.StatusBadRequest,
+			Error:      fmt.Errorf("failed to merge websocket input: %w", errSet),
+		}
+	}
+	if !gjson.GetBytes(normalized, "model").Exists() {
+		modelName := strings.TrimSpace(gjson.GetBytes(lastRequest, "model").String())
+		if modelName != "" {
+			normalized, _ = sjson.SetBytes(normalized, "model", modelName)
+		}
+	}
+	if !gjson.GetBytes(normalized, "instructions").Exists() {
+		instructions := gjson.GetBytes(lastRequest, "instructions")
+		if instructions.Exists() {
+			normalized, _ = sjson.SetRawBytes(normalized, "instructions", []byte(instructions.Raw))
+		}
+	}
+	normalized, _ = sjson.SetBytes(normalized, "stream", true)
+	return normalized, bytes.Clone(normalized), nil
+}
+
+func websocketUpstreamSupportsIncrementalInput(attributes map[string]string, metadata map[string]any) bool {
+	if len(attributes) > 0 {
+		if raw := strings.TrimSpace(attributes["websockets"]); raw != "" {
+			parsed, errParse := strconv.ParseBool(raw)
+			if errParse == nil {
+				return parsed
+			}
+		}
+	}
+	if len(metadata) == 0 {
+		return false
+	}
+	raw, ok := metadata["websockets"]
+	if !ok || raw == nil {
+		return false
+	}
+	switch value := raw.(type) {
+	case bool:
+		return value
+	case string:
+		parsed, errParse := strconv.ParseBool(strings.TrimSpace(value))
+		if errParse == nil {
+			return parsed
+		}
+	default:
+	}
+	return false
+}
+
+func mergeJSONArrayRaw(existingRaw, appendRaw string) (string, error) {
+	existingRaw = strings.TrimSpace(existingRaw)
+	appendRaw = strings.TrimSpace(appendRaw)
+	if existingRaw == "" {
+		existingRaw = "[]"
+	}
+	if appendRaw == "" {
+		appendRaw = "[]"
+	}
+
+	var existing []json.RawMessage
+	if err := json.Unmarshal([]byte(existingRaw), &existing); err != nil {
+		return "", err
+	}
+	var appendItems []json.RawMessage
+	if err := json.Unmarshal([]byte(appendRaw), &appendItems); err != nil {
+		return "", err
+	}
+
+	merged := append(existing, appendItems...)
+	out, err := json.Marshal(merged)
+	if err != nil {
+		return "", err
+	}
+	return string(out), nil
+}
+
+func normalizeJSONArrayRaw(raw []byte) string {
+	trimmed := strings.TrimSpace(string(raw))
+	if trimmed == "" {
+		return "[]"
+	}
+	result := gjson.Parse(trimmed)
+	if result.Type == gjson.JSON && result.IsArray() {
+		return trimmed
+	}
+	return "[]"
+}
+
+func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
+	c *gin.Context,
+	conn *websocket.Conn,
+	cancel handlers.APIHandlerCancelFunc,
+	data <-chan []byte,
+	errs <-chan *interfaces.ErrorMessage,
+	wsBodyLog *strings.Builder,
+	sessionID string,
+) ([]byte, error) {
+	completed := false
+	completedOutput := []byte("[]")
+
+	for {
+		select {
+		case <-c.Request.Context().Done():
+			cancel(c.Request.Context().Err())
+			return completedOutput, c.Request.Context().Err()
+		case errMsg, ok := <-errs:
+			if !ok {
+				errs = nil
+				continue
+			}
+			if errMsg != nil {
+				h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+				markAPIResponseTimestamp(c)
+				errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+				appendWebsocketEvent(wsBodyLog, "response", errorPayload)
+				log.Infof(
+					"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+					sessionID,
+					websocket.TextMessage,
+					websocketPayloadEventType(errorPayload),
+					websocketPayloadPreview(errorPayload),
+				)
+				if errWrite != nil {
+					// log.Warnf(
+					// 	"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+					// 	sessionID,
+					// 	websocketPayloadEventType(errorPayload),
+					// 	errWrite,
+					// )
+					cancel(errMsg.Error)
+					return completedOutput, errWrite
+				}
+			}
+			if errMsg != nil {
+				cancel(errMsg.Error)
+			} else {
+				cancel(nil)
+			}
+			return completedOutput, nil
+		case chunk, ok := <-data:
+			if !ok {
+				if !completed {
+					errMsg := &interfaces.ErrorMessage{
+						StatusCode: http.StatusRequestTimeout,
+						Error:      fmt.Errorf("stream closed before response.completed"),
+					}
+					h.LoggingAPIResponseError(context.WithValue(context.Background(), "gin", c), errMsg)
+					markAPIResponseTimestamp(c)
+					errorPayload, errWrite := writeResponsesWebsocketError(conn, errMsg)
+					appendWebsocketEvent(wsBodyLog, "response", errorPayload)
+					log.Infof(
+						"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+						sessionID,
+						websocket.TextMessage,
+						websocketPayloadEventType(errorPayload),
+						websocketPayloadPreview(errorPayload),
+					)
+					if errWrite != nil {
+						log.Warnf(
+							"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+							sessionID,
+							websocketPayloadEventType(errorPayload),
+							errWrite,
+						)
+						cancel(errMsg.Error)
+						return completedOutput, errWrite
+					}
+					cancel(errMsg.Error)
+					return completedOutput, nil
+				}
+				cancel(nil)
+				return completedOutput, nil
+			}
+
+			payloads := websocketJSONPayloadsFromChunk(chunk)
+			for i := range payloads {
+				eventType := gjson.GetBytes(payloads[i], "type").String()
+				if eventType == wsEventTypeCompleted {
+					// log.Infof("replace %s with %s", wsEventTypeCompleted, wsEventTypeDone)
+					payloads[i], _ = sjson.SetBytes(payloads[i], "type", wsEventTypeDone)
+
+					completed = true
+					completedOutput = responseCompletedOutputFromPayload(payloads[i])
+				}
+				markAPIResponseTimestamp(c)
+				appendWebsocketEvent(wsBodyLog, "response", payloads[i])
+				// log.Infof(
+				// 	"responses websocket: downstream_out id=%s type=%d event=%s payload=%s",
+				// 	sessionID,
+				// 	websocket.TextMessage,
+				// 	websocketPayloadEventType(payloads[i]),
+				// 	websocketPayloadPreview(payloads[i]),
+				// )
+				if errWrite := conn.WriteMessage(websocket.TextMessage, payloads[i]); errWrite != nil {
+					log.Warnf(
+						"responses websocket: downstream_out write failed id=%s event=%s error=%v",
+						sessionID,
+						websocketPayloadEventType(payloads[i]),
+						errWrite,
+					)
+					cancel(errWrite)
+					return completedOutput, errWrite
+				}
+			}
+		}
+	}
+}
+
+func responseCompletedOutputFromPayload(payload []byte) []byte {
+	output := gjson.GetBytes(payload, "response.output")
+	if output.Exists() && output.IsArray() {
+		return bytes.Clone([]byte(output.Raw))
+	}
+	return []byte("[]")
+}
+
+func websocketJSONPayloadsFromChunk(chunk []byte) [][]byte {
+	payloads := make([][]byte, 0, 2)
+	lines := bytes.Split(chunk, []byte("\n"))
+	for i := range lines {
+		line := bytes.TrimSpace(lines[i])
+		if len(line) == 0 || bytes.HasPrefix(line, []byte("event:")) {
+			continue
+		}
+		if bytes.HasPrefix(line, []byte("data:")) {
+			line = bytes.TrimSpace(line[len("data:"):])
+		}
+		if len(line) == 0 || bytes.Equal(line, []byte(wsDoneMarker)) {
+			continue
+		}
+		if json.Valid(line) {
+			payloads = append(payloads, bytes.Clone(line))
+		}
+	}
+
+	if len(payloads) > 0 {
+		return payloads
+	}
+
+	trimmed := bytes.TrimSpace(chunk)
+	if bytes.HasPrefix(trimmed, []byte("data:")) {
+		trimmed = bytes.TrimSpace(trimmed[len("data:"):])
+	}
+	if len(trimmed) > 0 && !bytes.Equal(trimmed, []byte(wsDoneMarker)) && json.Valid(trimmed) {
+		payloads = append(payloads, bytes.Clone(trimmed))
+	}
+	return payloads
+}
+
+func writeResponsesWebsocketError(conn *websocket.Conn, errMsg *interfaces.ErrorMessage) ([]byte, error) {
+	status := http.StatusInternalServerError
+	errText := http.StatusText(status)
+	if errMsg != nil {
+		if errMsg.StatusCode > 0 {
+			status = errMsg.StatusCode
+			errText = http.StatusText(status)
+		}
+		if errMsg.Error != nil && strings.TrimSpace(errMsg.Error.Error()) != "" {
+			errText = errMsg.Error.Error()
+		}
+	}
+
+	body := handlers.BuildErrorResponseBody(status, errText)
+	payload := map[string]any{
+		"type":   wsEventTypeError,
+		"status": status,
+	}
+
+	if errMsg != nil && errMsg.Addon != nil {
+		headers := map[string]any{}
+		for key, values := range errMsg.Addon {
+			if len(values) == 0 {
+				continue
+			}
+			headers[key] = values[0]
+		}
+		if len(headers) > 0 {
+			payload["headers"] = headers
+		}
+	}
+
+	if len(body) > 0 && json.Valid(body) {
+		var decoded map[string]any
+		if errDecode := json.Unmarshal(body, &decoded); errDecode == nil {
+			if inner, ok := decoded["error"]; ok {
+				payload["error"] = inner
+			} else {
+				payload["error"] = decoded
+			}
+		}
+	}
+
+	if _, ok := payload["error"]; !ok {
+		payload["error"] = map[string]any{
+			"type":    "server_error",
+			"message": errText,
+		}
+	}
+
+	data, err := json.Marshal(payload)
+	if err != nil {
+		return nil, err
+	}
+	return data, conn.WriteMessage(websocket.TextMessage, data)
+}
+
+func appendWebsocketEvent(builder *strings.Builder, eventType string, payload []byte) {
+	if builder == nil {
+		return
+	}
+	trimmedPayload := bytes.TrimSpace(payload)
+	if len(trimmedPayload) == 0 {
+		return
+	}
+	if builder.Len() > 0 {
+		builder.WriteString("\n")
+	}
+	builder.WriteString("websocket.")
+	builder.WriteString(eventType)
+	builder.WriteString("\n")
+	builder.Write(trimmedPayload)
+	builder.WriteString("\n")
+}
+
+func websocketPayloadEventType(payload []byte) string {
+	eventType := strings.TrimSpace(gjson.GetBytes(payload, "type").String())
+	if eventType == "" {
+		return "-"
+	}
+	return eventType
+}
+
+func websocketPayloadPreview(payload []byte) string {
+	trimmedPayload := bytes.TrimSpace(payload)
+	if len(trimmedPayload) == 0 {
+		return "<empty>"
+	}
+	preview := trimmedPayload
+	if len(preview) > wsPayloadLogMaxSize {
+		preview = preview[:wsPayloadLogMaxSize]
+	}
+	previewText := strings.ReplaceAll(string(preview), "\n", "\\n")
+	previewText = strings.ReplaceAll(previewText, "\r", "\\r")
+	if len(trimmedPayload) > wsPayloadLogMaxSize {
+		return fmt.Sprintf("%s...(truncated,total=%d)", previewText, len(trimmedPayload))
+	}
+	return previewText
+}
+
+func setWebsocketRequestBody(c *gin.Context, body string) {
+	if c == nil {
+		return
+	}
+	trimmedBody := strings.TrimSpace(body)
+	if trimmedBody == "" {
+		return
+	}
+	c.Set(wsRequestBodyKey, []byte(trimmedBody))
+}
+
+func markAPIResponseTimestamp(c *gin.Context) {
+	if c == nil {
+		return
+	}
+	if _, exists := c.Get("API_RESPONSE_TIMESTAMP"); exists {
+		return
+	}
+	c.Set("API_RESPONSE_TIMESTAMP", time.Now())
+}

sdk/api/handlers/openai/openai_responses_websocket_test.go

@@ -0,0 +1,249 @@
+package openai
+
+import (
+	"bytes"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/tidwall/gjson"
+)
+
+func TestNormalizeResponsesWebsocketRequestCreate(t *testing.T) {
+	raw := []byte(`{"type":"response.create","model":"test-model","stream":false,"input":[{"type":"message","id":"msg-1"}]}`)
+
+	normalized, last, errMsg := normalizeResponsesWebsocketRequest(raw, nil, nil)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized create request must not include type field")
+	}
+	if !gjson.GetBytes(normalized, "stream").Bool() {
+		t.Fatalf("normalized create request must force stream=true")
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+	if !bytes.Equal(last, normalized) {
+		t.Fatalf("last request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestCreateWithHistory(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequest(raw, lastRequest, lastResponseOutput)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized subsequent create request must not include type field")
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 4 {
+		t.Fatalf("merged input len = %d, want 4", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "fc-1" ||
+		input[2].Get("id").String() != "assistant-1" ||
+		input[3].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestWithPreviousResponseIDIncremental(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"instructions":"be helpful","input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","previous_response_id":"resp-1","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequestWithMode(raw, lastRequest, lastResponseOutput, true)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "type").Exists() {
+		t.Fatalf("normalized request must not include type field")
+	}
+	if gjson.GetBytes(normalized, "previous_response_id").String() != "resp-1" {
+		t.Fatalf("previous_response_id must be preserved in incremental mode")
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 1 {
+		t.Fatalf("incremental input len = %d, want 1", len(input))
+	}
+	if input[0].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected incremental input item id: %s", input[0].Get("id").String())
+	}
+	if gjson.GetBytes(normalized, "model").String() != "test-model" {
+		t.Fatalf("unexpected model: %s", gjson.GetBytes(normalized, "model").String())
+	}
+	if gjson.GetBytes(normalized, "instructions").String() != "be helpful" {
+		t.Fatalf("unexpected instructions: %s", gjson.GetBytes(normalized, "instructions").String())
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestWithPreviousResponseIDMergedWhenIncrementalDisabled(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"function_call","id":"fc-1","call_id":"call-1"},
+		{"type":"message","id":"assistant-1"}
+	]`)
+	raw := []byte(`{"type":"response.create","previous_response_id":"resp-1","input":[{"type":"function_call_output","call_id":"call-1","id":"tool-out-1"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequestWithMode(raw, lastRequest, lastResponseOutput, false)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	if gjson.GetBytes(normalized, "previous_response_id").Exists() {
+		t.Fatalf("previous_response_id must be removed when incremental mode is disabled")
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 4 {
+		t.Fatalf("merged input len = %d, want 4", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "fc-1" ||
+		input[2].Get("id").String() != "assistant-1" ||
+		input[3].Get("id").String() != "tool-out-1" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestAppend(t *testing.T) {
+	lastRequest := []byte(`{"model":"test-model","stream":true,"input":[{"type":"message","id":"msg-1"}]}`)
+	lastResponseOutput := []byte(`[
+		{"type":"message","id":"assistant-1"},
+		{"type":"function_call_output","id":"tool-out-1"}
+	]`)
+	raw := []byte(`{"type":"response.append","input":[{"type":"message","id":"msg-2"},{"type":"message","id":"msg-3"}]}`)
+
+	normalized, next, errMsg := normalizeResponsesWebsocketRequest(raw, lastRequest, lastResponseOutput)
+	if errMsg != nil {
+		t.Fatalf("unexpected error: %v", errMsg.Error)
+	}
+	input := gjson.GetBytes(normalized, "input").Array()
+	if len(input) != 5 {
+		t.Fatalf("merged input len = %d, want 5", len(input))
+	}
+	if input[0].Get("id").String() != "msg-1" ||
+		input[1].Get("id").String() != "assistant-1" ||
+		input[2].Get("id").String() != "tool-out-1" ||
+		input[3].Get("id").String() != "msg-2" ||
+		input[4].Get("id").String() != "msg-3" {
+		t.Fatalf("unexpected merged input order")
+	}
+	if !bytes.Equal(next, normalized) {
+		t.Fatalf("next request snapshot should match normalized append request")
+	}
+}
+
+func TestNormalizeResponsesWebsocketRequestAppendWithoutCreate(t *testing.T) {
+	raw := []byte(`{"type":"response.append","input":[]}`)
+
+	_, _, errMsg := normalizeResponsesWebsocketRequest(raw, nil, nil)
+	if errMsg == nil {
+		t.Fatalf("expected error for append without previous request")
+	}
+	if errMsg.StatusCode != http.StatusBadRequest {
+		t.Fatalf("status = %d, want %d", errMsg.StatusCode, http.StatusBadRequest)
+	}
+}
+
+func TestWebsocketJSONPayloadsFromChunk(t *testing.T) {
+	chunk := []byte("event: response.created\n\ndata: {\"type\":\"response.created\",\"response\":{\"id\":\"resp-1\"}}\n\ndata: [DONE]\n")
+
+	payloads := websocketJSONPayloadsFromChunk(chunk)
+	if len(payloads) != 1 {
+		t.Fatalf("payloads len = %d, want 1", len(payloads))
+	}
+	if gjson.GetBytes(payloads[0], "type").String() != "response.created" {
+		t.Fatalf("unexpected payload type: %s", gjson.GetBytes(payloads[0], "type").String())
+	}
+}
+
+func TestWebsocketJSONPayloadsFromPlainJSONChunk(t *testing.T) {
+	chunk := []byte(`{"type":"response.completed","response":{"id":"resp-1"}}`)
+
+	payloads := websocketJSONPayloadsFromChunk(chunk)
+	if len(payloads) != 1 {
+		t.Fatalf("payloads len = %d, want 1", len(payloads))
+	}
+	if gjson.GetBytes(payloads[0], "type").String() != "response.completed" {
+		t.Fatalf("unexpected payload type: %s", gjson.GetBytes(payloads[0], "type").String())
+	}
+}
+
+func TestResponseCompletedOutputFromPayload(t *testing.T) {
+	payload := []byte(`{"type":"response.completed","response":{"id":"resp-1","output":[{"type":"message","id":"out-1"}]}}`)
+
+	output := responseCompletedOutputFromPayload(payload)
+	items := gjson.ParseBytes(output).Array()
+	if len(items) != 1 {
+		t.Fatalf("output len = %d, want 1", len(items))
+	}
+	if items[0].Get("id").String() != "out-1" {
+		t.Fatalf("unexpected output id: %s", items[0].Get("id").String())
+	}
+}
+
+func TestAppendWebsocketEvent(t *testing.T) {
+	var builder strings.Builder
+
+	appendWebsocketEvent(&builder, "request", []byte("  {\"type\":\"response.create\"}\n"))
+	appendWebsocketEvent(&builder, "response", []byte("{\"type\":\"response.created\"}"))
+
+	got := builder.String()
+	if !strings.Contains(got, "websocket.request\n{\"type\":\"response.create\"}\n") {
+		t.Fatalf("request event not found in body: %s", got)
+	}
+	if !strings.Contains(got, "websocket.response\n{\"type\":\"response.created\"}\n") {
+		t.Fatalf("response event not found in body: %s", got)
+	}
+}
+
+func TestSetWebsocketRequestBody(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	recorder := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(recorder)
+
+	setWebsocketRequestBody(c, " \n ")
+	if _, exists := c.Get(wsRequestBodyKey); exists {
+		t.Fatalf("request body key should not be set for empty body")
+	}
+
+	setWebsocketRequestBody(c, "event body")
+	value, exists := c.Get(wsRequestBodyKey)
+	if !exists {
+		t.Fatalf("request body key not set")
+	}
+	bodyBytes, ok := value.([]byte)
+	if !ok {
+		t.Fatalf("request body key type mismatch")
+	}
+	if string(bodyBytes) != "event body" {
+		t.Fatalf("request body = %q, want %q", string(bodyBytes), "event body")
+	}
+}

sdk/auth/antigravity.go

@@ -28,8 +28,7 @@ func (AntigravityAuthenticator) Provider() string { return "antigravity" }
 
 // RefreshLead instructs the manager to refresh five minutes before expiry.
 func (AntigravityAuthenticator) RefreshLead() *time.Duration {
-	lead := 5 * time.Minute
-	return &lead
+	return new(5 * time.Minute)
 }
 
 // Login launches a local OAuth flow to obtain antigravity tokens and persists them.

sdk/auth/claude.go

@@ -32,8 +32,7 @@ func (a *ClaudeAuthenticator) Provider() string {
 }
 
 func (a *ClaudeAuthenticator) RefreshLead() *time.Duration {
-	d := 4 * time.Hour
-	return &d
+	return new(4 * time.Hour)
 }
 
 func (a *ClaudeAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {

sdk/auth/codex.go

@@ -34,8 +34,7 @@ func (a *CodexAuthenticator) Provider() string {
 }
 
 func (a *CodexAuthenticator) RefreshLead() *time.Duration {
-	d := 5 * 24 * time.Hour
-	return &d
+	return new(5 * 24 * time.Hour)
 }
 
 func (a *CodexAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {

sdk/auth/filestore.go

@@ -4,8 +4,10 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"io/fs"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"strings"
@@ -186,15 +188,21 @@ func (s *FileTokenStore) readAuthFile(path, baseDir string) (*cliproxyauth.Auth,
 	if provider == "" {
 		provider = "unknown"
 	}
-	if provider == "antigravity" {
+	if provider == "antigravity" || provider == "gemini" {
 		projectID := ""
 		if pid, ok := metadata["project_id"].(string); ok {
 			projectID = strings.TrimSpace(pid)
 		}
 		if projectID == "" {
-			accessToken := ""
-			if token, ok := metadata["access_token"].(string); ok {
-				accessToken = strings.TrimSpace(token)
+			accessToken := extractAccessToken(metadata)
+			// For gemini type, the stored access_token is likely expired (~1h lifetime).
+			// Refresh it using the long-lived refresh_token before querying.
+			if provider == "gemini" {
+				if tokenMap, ok := metadata["token"].(map[string]any); ok {
+					if refreshed, errRefresh := refreshGeminiAccessToken(tokenMap, http.DefaultClient); errRefresh == nil {
+						accessToken = refreshed
+					}
+				}
 			}
 			if accessToken != "" {
 				fetchedProjectID, errFetch := FetchAntigravityProjectID(context.Background(), accessToken, http.DefaultClient)
@@ -304,6 +312,67 @@ func (s *FileTokenStore) baseDirSnapshot() string {
 	return s.baseDir
 }
 
+func extractAccessToken(metadata map[string]any) string {
+	if at, ok := metadata["access_token"].(string); ok {
+		if v := strings.TrimSpace(at); v != "" {
+			return v
+		}
+	}
+	if tokenMap, ok := metadata["token"].(map[string]any); ok {
+		if at, ok := tokenMap["access_token"].(string); ok {
+			if v := strings.TrimSpace(at); v != "" {
+				return v
+			}
+		}
+	}
+	return ""
+}
+
+func refreshGeminiAccessToken(tokenMap map[string]any, httpClient *http.Client) (string, error) {
+	refreshToken, _ := tokenMap["refresh_token"].(string)
+	clientID, _ := tokenMap["client_id"].(string)
+	clientSecret, _ := tokenMap["client_secret"].(string)
+	tokenURI, _ := tokenMap["token_uri"].(string)
+
+	if refreshToken == "" || clientID == "" || clientSecret == "" {
+		return "", fmt.Errorf("missing refresh credentials")
+	}
+	if tokenURI == "" {
+		tokenURI = "https://oauth2.googleapis.com/token"
+	}
+
+	data := url.Values{
+		"grant_type":    {"refresh_token"},
+		"refresh_token": {refreshToken},
+		"client_id":     {clientID},
+		"client_secret": {clientSecret},
+	}
+
+	resp, err := httpClient.PostForm(tokenURI, data)
+	if err != nil {
+		return "", fmt.Errorf("refresh request: %w", err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	body, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("refresh failed: status %d", resp.StatusCode)
+	}
+
+	var result map[string]any
+	if errUnmarshal := json.Unmarshal(body, &result); errUnmarshal != nil {
+		return "", fmt.Errorf("decode refresh response: %w", errUnmarshal)
+	}
+
+	newAccessToken, _ := result["access_token"].(string)
+	if newAccessToken == "" {
+		return "", fmt.Errorf("no access_token in refresh response")
+	}
+
+	tokenMap["access_token"] = newAccessToken
+	return newAccessToken, nil
+}
+
 // jsonEqual compares two JSON blobs by parsing them into Go objects and deep comparing.
 func jsonEqual(a, b []byte) bool {
 	var objA any

sdk/auth/filestore_test.go

@@ -0,0 +1,80 @@
+package auth
+
+import "testing"
+
+func TestExtractAccessToken(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name     string
+		metadata map[string]any
+		expected string
+	}{
+		{
+			"antigravity top-level access_token",
+			map[string]any{"access_token": "tok-abc"},
+			"tok-abc",
+		},
+		{
+			"gemini nested token.access_token",
+			map[string]any{
+				"token": map[string]any{"access_token": "tok-nested"},
+			},
+			"tok-nested",
+		},
+		{
+			"top-level takes precedence over nested",
+			map[string]any{
+				"access_token": "tok-top",
+				"token":        map[string]any{"access_token": "tok-nested"},
+			},
+			"tok-top",
+		},
+		{
+			"empty metadata",
+			map[string]any{},
+			"",
+		},
+		{
+			"whitespace-only access_token",
+			map[string]any{"access_token": "   "},
+			"",
+		},
+		{
+			"wrong type access_token",
+			map[string]any{"access_token": 12345},
+			"",
+		},
+		{
+			"token is not a map",
+			map[string]any{"token": "not-a-map"},
+			"",
+		},
+		{
+			"nested whitespace-only",
+			map[string]any{
+				"token": map[string]any{"access_token": "  "},
+			},
+			"",
+		},
+		{
+			"fallback to nested when top-level empty",
+			map[string]any{
+				"access_token": "",
+				"token":        map[string]any{"access_token": "tok-fallback"},
+			},
+			"tok-fallback",
+		},
+	}
+
+	for _, tt := range tests {
+		tt := tt
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got := extractAccessToken(tt.metadata)
+			if got != tt.expected {
+				t.Errorf("extractAccessToken() = %q, want %q", got, tt.expected)
+			}
+		})
+	}
+}

sdk/auth/iflow.go

@@ -26,8 +26,7 @@ func (a *IFlowAuthenticator) Provider() string { return "iflow" }
 
 // RefreshLead indicates how soon before expiry a refresh should be attempted.
 func (a *IFlowAuthenticator) RefreshLead() *time.Duration {
-	d := 24 * time.Hour
-	return &d
+	return new(24 * time.Hour)
 }
 
 // Login performs the OAuth code flow using a local callback server.

sdk/auth/qwen.go

@@ -27,8 +27,7 @@ func (a *QwenAuthenticator) Provider() string {
 }
 
 func (a *QwenAuthenticator) RefreshLead() *time.Duration {
-	d := 3 * time.Hour
-	return &d
+	return new(3 * time.Hour)
 }
 
 func (a *QwenAuthenticator) Login(ctx context.Context, cfg *config.Config, opts *LoginOptions) (*coreauth.Auth, error) {

sdk/cliproxy/auth/conductor.go

@@ -41,6 +41,17 @@ type ProviderExecutor interface {
 	HttpRequest(ctx context.Context, auth *Auth, req *http.Request) (*http.Response, error)
 }
 
+// ExecutionSessionCloser allows executors to release per-session runtime resources.
+type ExecutionSessionCloser interface {
+	CloseExecutionSession(sessionID string)
+}
+
+const (
+	// CloseAllExecutionSessionsID asks an executor to release all active execution sessions.
+	// Executors that do not support this marker may ignore it.
+	CloseAllExecutionSessionsID = "__all_execution_sessions__"
+)
+
 // RefreshEvaluator allows runtime state to override refresh decisions.
 type RefreshEvaluator interface {
 	ShouldRefresh(now time.Time, auth *Auth) bool
@@ -389,9 +400,23 @@ func (m *Manager) RegisterExecutor(executor ProviderExecutor) {
 	if executor == nil {
 		return
 	}
+	provider := strings.TrimSpace(executor.Identifier())
+	if provider == "" {
+		return
+	}
+
+	var replaced ProviderExecutor
 	m.mu.Lock()
-	defer m.mu.Unlock()
-	m.executors[executor.Identifier()] = executor
+	replaced = m.executors[provider]
+	m.executors[provider] = executor
+	m.mu.Unlock()
+
+	if replaced == nil || replaced == executor {
+		return
+	}
+	if closer, ok := replaced.(ExecutionSessionCloser); ok && closer != nil {
+		closer.CloseExecutionSession(CloseAllExecutionSessionsID)
+	}
 }
 
 // UnregisterExecutor removes the executor associated with the provider key.
@@ -581,6 +606,7 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 
 		entry := logEntryWithRequestID(ctx)
 		debugLogAuthSelection(entry, auth, provider, req.Model)
+		publishSelectedAuthMetadata(opts.Metadata, auth.ID)
 
 		tried[auth.ID] = struct{}{}
 		execCtx := ctx
@@ -599,8 +625,7 @@ func (m *Manager) executeMixedOnce(ctx context.Context, providers []string, req
 				return cliproxyexecutor.Response{}, errCtx
 			}
 			result.Error = &Error{Message: errExec.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errExec, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
 				result.Error.HTTPStatus = se.StatusCode()
 			}
 			if ra := retryAfterFromError(errExec); ra != nil {
@@ -637,6 +662,7 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 
 		entry := logEntryWithRequestID(ctx)
 		debugLogAuthSelection(entry, auth, provider, req.Model)
+		publishSelectedAuthMetadata(opts.Metadata, auth.ID)
 
 		tried[auth.ID] = struct{}{}
 		execCtx := ctx
@@ -655,8 +681,7 @@ func (m *Manager) executeCountMixedOnce(ctx context.Context, providers []string,
 				return cliproxyexecutor.Response{}, errCtx
 			}
 			result.Error = &Error{Message: errExec.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errExec, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errExec); ok && se != nil {
 				result.Error.HTTPStatus = se.StatusCode()
 			}
 			if ra := retryAfterFromError(errExec); ra != nil {
@@ -693,6 +718,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 
 		entry := logEntryWithRequestID(ctx)
 		debugLogAuthSelection(entry, auth, provider, req.Model)
+		publishSelectedAuthMetadata(opts.Metadata, auth.ID)
 
 		tried[auth.ID] = struct{}{}
 		execCtx := ctx
@@ -710,8 +736,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 				return nil, errCtx
 			}
 			rerr := &Error{Message: errStream.Error()}
-			var se cliproxyexecutor.StatusError
-			if errors.As(errStream, &se) && se != nil {
+			if se, ok := errors.AsType[cliproxyexecutor.StatusError](errStream); ok && se != nil {
 				rerr.HTTPStatus = se.StatusCode()
 			}
 			result := Result{AuthID: auth.ID, Provider: provider, Model: routeModel, Success: false, Error: rerr}
@@ -732,8 +757,7 @@ func (m *Manager) executeStreamMixedOnce(ctx context.Context, providers []string
 				if chunk.Err != nil && !failed {
 					failed = true
 					rerr := &Error{Message: chunk.Err.Error()}
-					var se cliproxyexecutor.StatusError
-					if errors.As(chunk.Err, &se) && se != nil {
+					if se, ok := errors.AsType[cliproxyexecutor.StatusError](chunk.Err); ok && se != nil {
 						rerr.HTTPStatus = se.StatusCode()
 					}
 					m.MarkResult(streamCtx, Result{AuthID: streamAuth.ID, Provider: streamProvider, Model: routeModel, Success: false, Error: rerr})
@@ -798,6 +822,38 @@ func hasRequestedModelMetadata(meta map[string]any) bool {
 	}
 }
 
+func pinnedAuthIDFromMetadata(meta map[string]any) string {
+	if len(meta) == 0 {
+		return ""
+	}
+	raw, ok := meta[cliproxyexecutor.PinnedAuthMetadataKey]
+	if !ok || raw == nil {
+		return ""
+	}
+	switch val := raw.(type) {
+	case string:
+		return strings.TrimSpace(val)
+	case []byte:
+		return strings.TrimSpace(string(val))
+	default:
+		return ""
+	}
+}
+
+func publishSelectedAuthMetadata(meta map[string]any, authID string) {
+	if len(meta) == 0 {
+		return
+	}
+	authID = strings.TrimSpace(authID)
+	if authID == "" {
+		return
+	}
+	meta[cliproxyexecutor.SelectedAuthMetadataKey] = authID
+	if callback, ok := meta[cliproxyexecutor.SelectedAuthCallbackMetadataKey].(func(string)); ok && callback != nil {
+		callback(authID)
+	}
+}
+
 func rewriteModelForAuth(model string, auth *Auth) string {
 	if auth == nil || model == "" {
 		return model
@@ -1431,8 +1487,7 @@ func retryAfterFromError(err error) *time.Duration {
 	if retryAfter == nil {
 		return nil
 	}
-	val := *retryAfter
-	return &val
+	return new(*retryAfter)
 }
 
 func statusCodeFromResult(err *Error) int {
@@ -1555,7 +1610,56 @@ func (m *Manager) GetByID(id string) (*Auth, bool) {
 	return auth.Clone(), true
 }
 
+// Executor returns the registered provider executor for a provider key.
+func (m *Manager) Executor(provider string) (ProviderExecutor, bool) {
+	if m == nil {
+		return nil, false
+	}
+	provider = strings.TrimSpace(provider)
+	if provider == "" {
+		return nil, false
+	}
+
+	m.mu.RLock()
+	executor, okExecutor := m.executors[provider]
+	if !okExecutor {
+		lowerProvider := strings.ToLower(provider)
+		if lowerProvider != provider {
+			executor, okExecutor = m.executors[lowerProvider]
+		}
+	}
+	m.mu.RUnlock()
+
+	if !okExecutor || executor == nil {
+		return nil, false
+	}
+	return executor, true
+}
+
+// CloseExecutionSession asks all registered executors to release the supplied execution session.
+func (m *Manager) CloseExecutionSession(sessionID string) {
+	sessionID = strings.TrimSpace(sessionID)
+	if m == nil || sessionID == "" {
+		return
+	}
+
+	m.mu.RLock()
+	executors := make([]ProviderExecutor, 0, len(m.executors))
+	for _, exec := range m.executors {
+		executors = append(executors, exec)
+	}
+	m.mu.RUnlock()
+
+	for i := range executors {
+		if closer, ok := executors[i].(ExecutionSessionCloser); ok && closer != nil {
+			closer.CloseExecutionSession(sessionID)
+		}
+	}
+}
+
 func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, error) {
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+
 	m.mu.RLock()
 	executor, okExecutor := m.executors[provider]
 	if !okExecutor {
@@ -1576,6 +1680,9 @@ func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cli
 		if candidate.Provider != provider || candidate.Disabled {
 			continue
 		}
+		if pinnedAuthID != "" && candidate.ID != pinnedAuthID {
+			continue
+		}
 		if _, used := tried[candidate.ID]; used {
 			continue
 		}
@@ -1611,6 +1718,8 @@ func (m *Manager) pickNext(ctx context.Context, provider, model string, opts cli
 }
 
 func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model string, opts cliproxyexecutor.Options, tried map[string]struct{}) (*Auth, ProviderExecutor, string, error) {
+	pinnedAuthID := pinnedAuthIDFromMetadata(opts.Metadata)
+
 	providerSet := make(map[string]struct{}, len(providers))
 	for _, provider := range providers {
 		p := strings.TrimSpace(strings.ToLower(provider))
@@ -1638,6 +1747,9 @@ func (m *Manager) pickNextMixed(ctx context.Context, providers []string, model s
 		if candidate == nil || candidate.Disabled {
 			continue
 		}
+		if pinnedAuthID != "" && candidate.ID != pinnedAuthID {
+			continue
+		}
 		providerKey := strings.TrimSpace(strings.ToLower(candidate.Provider))
 		if providerKey == "" {
 			continue

sdk/cliproxy/auth/conductor_executor_replace_test.go

@@ -0,0 +1,100 @@
+package auth
+
+import (
+	"context"
+	"net/http"
+	"sync"
+	"testing"
+
+	cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/executor"
+)
+
+type replaceAwareExecutor struct {
+	id string
+
+	mu               sync.Mutex
+	closedSessionIDs []string
+}
+
+func (e *replaceAwareExecutor) Identifier() string {
+	return e.id
+}
+
+func (e *replaceAwareExecutor) Execute(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e *replaceAwareExecutor) ExecuteStream(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (<-chan cliproxyexecutor.StreamChunk, error) {
+	ch := make(chan cliproxyexecutor.StreamChunk)
+	close(ch)
+	return ch, nil
+}
+
+func (e *replaceAwareExecutor) Refresh(_ context.Context, auth *Auth) (*Auth, error) {
+	return auth, nil
+}
+
+func (e *replaceAwareExecutor) CountTokens(context.Context, *Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	return cliproxyexecutor.Response{}, nil
+}
+
+func (e *replaceAwareExecutor) HttpRequest(context.Context, *Auth, *http.Request) (*http.Response, error) {
+	return nil, nil
+}
+
+func (e *replaceAwareExecutor) CloseExecutionSession(sessionID string) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.closedSessionIDs = append(e.closedSessionIDs, sessionID)
+}
+
+func (e *replaceAwareExecutor) ClosedSessionIDs() []string {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	out := make([]string, len(e.closedSessionIDs))
+	copy(out, e.closedSessionIDs)
+	return out
+}
+
+func TestManagerRegisterExecutorClosesReplacedExecutionSessions(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, nil, nil)
+	replaced := &replaceAwareExecutor{id: "codex"}
+	current := &replaceAwareExecutor{id: "codex"}
+
+	manager.RegisterExecutor(replaced)
+	manager.RegisterExecutor(current)
+
+	closed := replaced.ClosedSessionIDs()
+	if len(closed) != 1 {
+		t.Fatalf("expected replaced executor close calls = 1, got %d", len(closed))
+	}
+	if closed[0] != CloseAllExecutionSessionsID {
+		t.Fatalf("expected close marker %q, got %q", CloseAllExecutionSessionsID, closed[0])
+	}
+	if len(current.ClosedSessionIDs()) != 0 {
+		t.Fatalf("expected current executor to stay open")
+	}
+}
+
+func TestManagerExecutorReturnsRegisteredExecutor(t *testing.T) {
+	t.Parallel()
+
+	manager := NewManager(nil, nil, nil)
+	current := &replaceAwareExecutor{id: "codex"}
+	manager.RegisterExecutor(current)
+
+	resolved, okResolved := manager.Executor("CODEX")
+	if !okResolved {
+		t.Fatal("expected registered executor to be found")
+	}
+	if resolved != current {
+		t.Fatal("expected resolved executor to match registered executor")
+	}
+
+	_, okMissing := manager.Executor("unknown")
+	if okMissing {
+		t.Fatal("expected unknown provider lookup to fail")
+	}
+}

sdk/cliproxy/auth/selector.go

@@ -134,6 +134,62 @@ func canonicalModelKey(model string) string {
 	return modelName
 }
 
+func authWebsocketsEnabled(auth *Auth) bool {
+	if auth == nil {
+		return false
+	}
+	if len(auth.Attributes) > 0 {
+		if raw := strings.TrimSpace(auth.Attributes["websockets"]); raw != "" {
+			parsed, errParse := strconv.ParseBool(raw)
+			if errParse == nil {
+				return parsed
+			}
+		}
+	}
+	if len(auth.Metadata) == 0 {
+		return false
+	}
+	raw, ok := auth.Metadata["websockets"]
+	if !ok || raw == nil {
+		return false
+	}
+	switch v := raw.(type) {
+	case bool:
+		return v
+	case string:
+		parsed, errParse := strconv.ParseBool(strings.TrimSpace(v))
+		if errParse == nil {
+			return parsed
+		}
+	default:
+	}
+	return false
+}
+
+func preferCodexWebsocketAuths(ctx context.Context, provider string, available []*Auth) []*Auth {
+	if len(available) == 0 {
+		return available
+	}
+	if !cliproxyexecutor.DownstreamWebsocket(ctx) {
+		return available
+	}
+	if !strings.EqualFold(strings.TrimSpace(provider), "codex") {
+		return available
+	}
+
+	wsEnabled := make([]*Auth, 0, len(available))
+	for i := 0; i < len(available); i++ {
+		candidate := available[i]
+		if authWebsocketsEnabled(candidate) {
+			wsEnabled = append(wsEnabled, candidate)
+		}
+	}
+	if len(wsEnabled) > 0 {
+		return wsEnabled
+	}
+	return available
+}
+
 func collectAvailableByPriority(auths []*Auth, model string, now time.Time) (available map[int][]*Auth, cooldownCount int, earliest time.Time) {
 	available = make(map[int][]*Auth)
 	for i := 0; i < len(auths); i++ {
@@ -193,13 +249,13 @@ func getAvailableAuths(auths []*Auth, provider, model string, now time.Time) ([]
 
 // Pick selects the next available auth for the provider in a round-robin manner.
 func (s *RoundRobinSelector) Pick(ctx context.Context, provider, model string, opts cliproxyexecutor.Options, auths []*Auth) (*Auth, error) {
-	_ = ctx
 	_ = opts
 	now := time.Now()
 	available, err := getAvailableAuths(auths, provider, model, now)
 	if err != nil {
 		return nil, err
 	}
+	available = preferCodexWebsocketAuths(ctx, provider, available)
 	key := provider + ":" + canonicalModelKey(model)
 	s.mu.Lock()
 	if s.cursors == nil {
@@ -226,13 +282,13 @@ func (s *RoundRobinSelector) Pick(ctx context.Context, provider, model string, o
 
 // 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) {
-	_ = ctx
 	_ = opts
 	now := time.Now()
 	available, err := getAvailableAuths(auths, provider, model, now)
 	if err != nil {
 		return nil, err
 	}
+	available = preferCodexWebsocketAuths(ctx, provider, available)
 	return available[0], nil
 }
 

sdk/cliproxy/executor/context.go

@@ -0,0 +1,23 @@
+package executor
+
+import "context"
+
+type downstreamWebsocketContextKey struct{}
+
+// WithDownstreamWebsocket marks the current request as coming from a downstream websocket connection.
+func WithDownstreamWebsocket(ctx context.Context) context.Context {
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return context.WithValue(ctx, downstreamWebsocketContextKey{}, true)
+}
+
+// DownstreamWebsocket reports whether the current request originates from a downstream websocket connection.
+func DownstreamWebsocket(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	raw := ctx.Value(downstreamWebsocketContextKey{})
+	enabled, ok := raw.(bool)
+	return ok && enabled
+}

sdk/cliproxy/executor/types.go

@@ -10,6 +10,17 @@ import (
 // RequestedModelMetadataKey stores the client-requested model name in Options.Metadata.
 const RequestedModelMetadataKey = "requested_model"
 
+const (
+	// PinnedAuthMetadataKey locks execution to a specific auth ID.
+	PinnedAuthMetadataKey = "pinned_auth_id"
+	// SelectedAuthMetadataKey stores the auth ID selected by the scheduler.
+	SelectedAuthMetadataKey = "selected_auth_id"
+	// SelectedAuthCallbackMetadataKey carries an optional callback invoked with the selected auth ID.
+	SelectedAuthCallbackMetadataKey = "selected_auth_callback"
+	// ExecutionSessionMetadataKey identifies a long-lived downstream execution session.
+	ExecutionSessionMetadataKey = "execution_session_id"
+)
+
 // Request encapsulates the translated payload that will be sent to a provider executor.
 type Request struct {
 	// Model is the upstream model identifier after translation.

sdk/cliproxy/service.go

@@ -325,6 +325,9 @@ func (s *Service) applyCoreAuthRemoval(ctx context.Context, id string) {
 		if _, err := s.coreManager.Update(ctx, existing); err != nil {
 			log.Errorf("failed to disable auth %s: %v", id, err)
 		}
+		if strings.EqualFold(strings.TrimSpace(existing.Provider), "codex") {
+			s.ensureExecutorsForAuth(existing)
+		}
 	}
 }
 
@@ -357,7 +360,24 @@ func openAICompatInfoFromAuth(a *coreauth.Auth) (providerKey string, compatName
 }
 
 func (s *Service) ensureExecutorsForAuth(a *coreauth.Auth) {
-	if s == nil || a == nil {
+	s.ensureExecutorsForAuthWithMode(a, false)
+}
+
+func (s *Service) ensureExecutorsForAuthWithMode(a *coreauth.Auth, forceReplace bool) {
+	if s == nil || s.coreManager == nil || a == nil {
+		return
+	}
+	if strings.EqualFold(strings.TrimSpace(a.Provider), "codex") {
+		if !forceReplace {
+			existingExecutor, hasExecutor := s.coreManager.Executor("codex")
+			if hasExecutor {
+				_, isCodexAutoExecutor := existingExecutor.(*executor.CodexAutoExecutor)
+				if isCodexAutoExecutor {
+					return
+				}
+			}
+		}
+		s.coreManager.RegisterExecutor(executor.NewCodexAutoExecutor(s.cfg))
 		return
 	}
 	// Skip disabled auth entries when (re)binding executors.
@@ -392,8 +412,6 @@ func (s *Service) ensureExecutorsForAuth(a *coreauth.Auth) {
 		s.coreManager.RegisterExecutor(executor.NewAntigravityExecutor(s.cfg))
 	case "claude":
 		s.coreManager.RegisterExecutor(executor.NewClaudeExecutor(s.cfg))
-	case "codex":
-		s.coreManager.RegisterExecutor(executor.NewCodexExecutor(s.cfg))
 	case "qwen":
 		s.coreManager.RegisterExecutor(executor.NewQwenExecutor(s.cfg))
 	case "iflow":
@@ -415,8 +433,15 @@ func (s *Service) rebindExecutors() {
 		return
 	}
 	auths := s.coreManager.List()
+	reboundCodex := false
 	for _, auth := range auths {
-		s.ensureExecutorsForAuth(auth)
+		if auth != nil && strings.EqualFold(strings.TrimSpace(auth.Provider), "codex") {
+			if reboundCodex {
+				continue
+			}
+			reboundCodex = true
+		}
+		s.ensureExecutorsForAuthWithMode(auth, true)
 	}
 }
 
@@ -740,6 +765,13 @@ func (s *Service) registerModelsForAuth(a *coreauth.Auth) {
 		provider = "openai-compatibility"
 	}
 	excluded := s.oauthExcludedModels(provider, authKind)
+	// The synthesizer pre-merges per-account and global exclusions into the "excluded_models" attribute.
+	// If this attribute is present, it represents the complete list of exclusions and overrides the global config.
+	if a.Attributes != nil {
+		if val, ok := a.Attributes["excluded_models"]; ok && strings.TrimSpace(val) != "" {
+			excluded = strings.Split(val, ",")
+		}
+	}
 	var models []*ModelInfo
 	switch provider {
 	case "gemini":

sdk/cliproxy/service_codex_executor_binding_test.go

@@ -0,0 +1,64 @@
+package cliproxy
+
+import (
+	"testing"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestEnsureExecutorsForAuth_CodexDoesNotReplaceInNormalMode(t *testing.T) {
+	service := &Service{
+		cfg:         &config.Config{},
+		coreManager: coreauth.NewManager(nil, nil, nil),
+	}
+	auth := &coreauth.Auth{
+		ID:       "codex-auth-1",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+	}
+
+	service.ensureExecutorsForAuth(auth)
+	firstExecutor, okFirst := service.coreManager.Executor("codex")
+	if !okFirst || firstExecutor == nil {
+		t.Fatal("expected codex executor after first bind")
+	}
+
+	service.ensureExecutorsForAuth(auth)
+	secondExecutor, okSecond := service.coreManager.Executor("codex")
+	if !okSecond || secondExecutor == nil {
+		t.Fatal("expected codex executor after second bind")
+	}
+
+	if firstExecutor != secondExecutor {
+		t.Fatal("expected codex executor to stay unchanged in normal mode")
+	}
+}
+
+func TestEnsureExecutorsForAuthWithMode_CodexForceReplace(t *testing.T) {
+	service := &Service{
+		cfg:         &config.Config{},
+		coreManager: coreauth.NewManager(nil, nil, nil),
+	}
+	auth := &coreauth.Auth{
+		ID:       "codex-auth-2",
+		Provider: "codex",
+		Status:   coreauth.StatusActive,
+	}
+
+	service.ensureExecutorsForAuth(auth)
+	firstExecutor, okFirst := service.coreManager.Executor("codex")
+	if !okFirst || firstExecutor == nil {
+		t.Fatal("expected codex executor after first bind")
+	}
+
+	service.ensureExecutorsForAuthWithMode(auth, true)
+	secondExecutor, okSecond := service.coreManager.Executor("codex")
+	if !okSecond || secondExecutor == nil {
+		t.Fatal("expected codex executor after forced rebind")
+	}
+
+	if firstExecutor == secondExecutor {
+		t.Fatal("expected codex executor replacement in force mode")
+	}
+}

sdk/cliproxy/service_excluded_models_test.go

@@ -0,0 +1,65 @@
+package cliproxy
+
+import (
+	"strings"
+	"testing"
+
+	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestRegisterModelsForAuth_UsesPreMergedExcludedModelsAttribute(t *testing.T) {
+	service := &Service{
+		cfg: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"gemini-cli": {"gemini-2.5-pro"},
+			},
+		},
+	}
+	auth := &coreauth.Auth{
+		ID:       "auth-gemini-cli",
+		Provider: "gemini-cli",
+		Status:   coreauth.StatusActive,
+		Attributes: map[string]string{
+			"auth_kind":       "oauth",
+			"excluded_models": "gemini-2.5-flash",
+		},
+	}
+
+	registry := GlobalModelRegistry()
+	registry.UnregisterClient(auth.ID)
+	t.Cleanup(func() {
+		registry.UnregisterClient(auth.ID)
+	})
+
+	service.registerModelsForAuth(auth)
+
+	models := registry.GetAvailableModelsByProvider("gemini-cli")
+	if len(models) == 0 {
+		t.Fatal("expected gemini-cli models to be registered")
+	}
+
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		modelID := strings.TrimSpace(model.ID)
+		if strings.EqualFold(modelID, "gemini-2.5-flash") {
+			t.Fatalf("expected model %q to be excluded by auth attribute", modelID)
+		}
+	}
+
+	seenGlobalExcluded := false
+	for _, model := range models {
+		if model == nil {
+			continue
+		}
+		if strings.EqualFold(strings.TrimSpace(model.ID), "gemini-2.5-pro") {
+			seenGlobalExcluded = true
+			break
+		}
+	}
+	if !seenGlobalExcluded {
+		t.Fatal("expected global excluded model to be present when attribute override is set")
+	}
+}

test/thinking_conversion_test.go

@@ -2590,6 +2590,135 @@ func TestThinkingE2EMatrix_Body(t *testing.T) {
 	runThinkingTests(t, cases)
 }
 
+// TestThinkingE2EClaudeAdaptive_Body tests Claude thinking.type=adaptive extended body-only cases.
+// These cases validate that adaptive means "thinking enabled without explicit budget", and
+// cross-protocol conversion should resolve to target-model maximum thinking capability.
+func TestThinkingE2EClaudeAdaptive_Body(t *testing.T) {
+	reg := registry.GetGlobalRegistry()
+	uid := fmt.Sprintf("thinking-e2e-claude-adaptive-%d", time.Now().UnixNano())
+
+	reg.RegisterClient(uid, "test", getTestModels())
+	defer reg.UnregisterClient(uid)
+
+	cases := []thinkingTestCase{
+		// A1: Claude adaptive to OpenAI level model -> highest supported level
+		{
+			name:        "A1",
+			from:        "claude",
+			to:          "openai",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning_effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		// A2: Claude adaptive to Gemini level subset model -> highest supported level
+		{
+			name:            "A2",
+			from:            "claude",
+			to:              "gemini",
+			model:           "level-subset-model",
+			inputJSON:       `{"model":"level-subset-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A3: Claude adaptive to Gemini budget model -> max budget
+		{
+			name:            "A3",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-budget-model",
+			inputJSON:       `{"model":"gemini-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A4: Claude adaptive to Gemini mixed model -> highest supported level
+		{
+			name:            "A4",
+			from:            "claude",
+			to:              "gemini",
+			model:           "gemini-mixed-model",
+			inputJSON:       `{"model":"gemini-mixed-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "generationConfig.thinkingConfig.thinkingLevel",
+			expectValue:     "high",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A5: Claude adaptive passthrough for same protocol
+		{
+			name:        "A5",
+			from:        "claude",
+			to:          "claude",
+			model:       "claude-budget-model",
+			inputJSON:   `{"model":"claude-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "thinking.type",
+			expectValue: "adaptive",
+			expectErr:   false,
+		},
+		// A6: Claude adaptive to Antigravity budget model -> max budget
+		{
+			name:            "A6",
+			from:            "claude",
+			to:              "antigravity",
+			model:           "antigravity-budget-model",
+			inputJSON:       `{"model":"antigravity-budget-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField:     "request.generationConfig.thinkingConfig.thinkingBudget",
+			expectValue:     "20000",
+			includeThoughts: "true",
+			expectErr:       false,
+		},
+		// A7: Claude adaptive to iFlow GLM -> enabled boolean
+		{
+			name:        "A7",
+			from:        "claude",
+			to:          "iflow",
+			model:       "glm-test",
+			inputJSON:   `{"model":"glm-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "chat_template_kwargs.enable_thinking",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		// A8: Claude adaptive to iFlow MiniMax -> enabled boolean
+		{
+			name:        "A8",
+			from:        "claude",
+			to:          "iflow",
+			model:       "minimax-test",
+			inputJSON:   `{"model":"minimax-test","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning_split",
+			expectValue: "true",
+			expectErr:   false,
+		},
+		// A9: Claude adaptive to Codex level model -> highest supported level
+		{
+			name:        "A9",
+			from:        "claude",
+			to:          "codex",
+			model:       "level-model",
+			inputJSON:   `{"model":"level-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "reasoning.effort",
+			expectValue: "high",
+			expectErr:   false,
+		},
+		// A10: Claude adaptive on non-thinking model should still be stripped
+		{
+			name:        "A10",
+			from:        "claude",
+			to:          "openai",
+			model:       "no-thinking-model",
+			inputJSON:   `{"model":"no-thinking-model","messages":[{"role":"user","content":"hi"}],"thinking":{"type":"adaptive"}}`,
+			expectField: "",
+			expectErr:   false,
+		},
+	}
+
+	runThinkingTests(t, cases)
+}
+
 // getTestModels returns the shared model definitions for E2E tests.
 func getTestModels() []*registry.ModelInfo {
 	return []*registry.ModelInfo{