fix(server): prevent data races during concurrent config access and reloads

- Access configuration safely via getters across server routes, middleware, and service routines.
- Clone configuration snapshots for runtime and synchronize access in management handlers.
- Guard configuration YAML snapshots with mutex locks during reloads.

Closes: #6592

internal/api/runtime_reload_regression_test.go

@@ -0,0 +1,44 @@
+package api
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"sync"
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v8/internal/config"
+)
+
+func TestHandlerRuntimeReloadConcurrentWithModelLists(t *testing.T) {
+	server := newTestServer(t)
+	host := &mockModelListInterceptorHost{}
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(2)
+	go func() {
+		defer wg.Done()
+		<-start
+		for i := 0; i < 500; i++ {
+			server.handlers.SetPluginHost(host)
+			server.handlers.UpdateClients(&config.SDKConfig{RequestLog: i%2 == 0})
+			server.handlers.SetPluginHost(nil)
+		}
+	}()
+	go func() {
+		defer wg.Done()
+		<-start
+		for i := 0; i < 500; i++ {
+			req := httptest.NewRequest(http.MethodGet, "/v1/models", nil)
+			req.Header.Set("Authorization", "Bearer test-key")
+			req.Header.Set("User-Agent", "codex_cli_rs/0.100.0")
+			rr := httptest.NewRecorder()
+			server.engine.ServeHTTP(rr, req)
+			if rr.Code != http.StatusOK {
+				t.Errorf("model list status = %d, want 200", rr.Code)
+				return
+			}
+		}
+	}()
+	close(start)
+	wg.Wait()
+}

internal/api/server_routes.go

@@ -236,8 +236,10 @@ func (s *Server) codexAlphaSearchModelRouterHost() handlers.PluginModelRouterHos
 	if s.pluginHost != nil {
 		return s.pluginHost
 	}
-	if s.handlers != nil && s.handlers.ModelRouterHost != nil {
-		return s.handlers.ModelRouterHost
+	if s.handlers != nil {
+		if host := s.handlers.ModelRouterHostSnapshot(); host != nil {
+			return host
+		}
 	}
 	return nil
 }

sdk/api/handlers/claude/code_handlers.go

@@ -157,7 +157,8 @@ func rewriteClaudeDDModelInBody(rawJSON []byte) []byte {
 // Parameters:
 //   - c: The Gin context for the request.
 func (h *ClaudeCodeAPIHandler) ClaudeModels(c *gin.Context) {
-	disableCloaking := h.Cfg != nil && h.Cfg.ClaudeCode.DisableCloakingModelList
+	cfg := h.CurrentConfig()
+	disableCloaking := cfg != nil && cfg.ClaudeCode.DisableCloakingModelList
 	h.WriteModelListResponse(c, h.HandlerType(), claudemodels.BuildResponse(h.Models(), disableCloaking))
 }
 
@@ -261,7 +262,7 @@ func (h *ClaudeCodeAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON [
 		flusher.Flush()
 		return true
 	}
-	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	defer stopKeepAlive()
 
 	dataChan, resolvedHeaders, errChan, okBootstrap := h.ExecuteStreamWithPreFTKeepAlive(
@@ -451,7 +452,7 @@ func (h *ClaudeCodeAPIHandler) WriteErrorResponse(c *gin.Context, msg *interface
 			c.Writer.Header().Add("Retry-After", value)
 		}
 	}
-	if msg != nil && msg.Addon != nil && handlers.PassthroughHeadersEnabled(h.Cfg) {
+	if msg != nil && msg.Addon != nil && handlers.PassthroughHeadersEnabled(h.CurrentConfig()) {
 		for key, values := range msg.Addon {
 			if len(values) == 0 || handlers.IsCPAReservedResponseHeader(key) {
 				continue

sdk/api/handlers/gemini/gemini_handlers.go

@@ -225,7 +225,7 @@ func (h *GeminiAPIHandler) handleStreamGenerateContent(c *gin.Context, modelName
 	var keepAliveC <-chan time.Time
 	stopKeepAlive := func() {}
 	if alt == "" {
-		keepAliveC, stopKeepAlive = handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+		keepAliveC, stopKeepAlive = handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	}
 	defer stopKeepAlive()
 

sdk/api/handlers/gemini/interactions_handlers.go

@@ -151,7 +151,7 @@ func (h *GeminiAPIHandler) handleInteractionsStream(c *gin.Context, cliCtx conte
 		handlers.WriteUpstreamHeaders(c.Writer.Header(), upstreamHeaders)
 		committed = true
 	}
-	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	defer stopKeepAlive()
 
 	var stream handlers.ModelExecutionStream

sdk/api/handlers/handlers.go

@@ -408,6 +408,8 @@ func setGenerateMetadata(meta map[string]any, rawJSON []byte) {
 // It holds a pool of clients to interact with the backend service and manages
 // load balancing, client selection, and configuration.
 type BaseAPIHandler struct {
+	runtimeMu sync.RWMutex
+
 	// AuthManager manages auth lifecycle and execution in the new architecture.
 	AuthManager *coreauth.Manager
 
@@ -444,7 +446,25 @@ func NewBaseAPIHandlers(cfg *config.SDKConfig, authManager *coreauth.Manager) *B
 // Parameters:
 //   - clients: The new slice of AI service clients
 //   - cfg: The new application configuration
-func (h *BaseAPIHandler) UpdateClients(cfg *config.SDKConfig) { h.Cfg = cfg }
+func (h *BaseAPIHandler) UpdateClients(cfg *config.SDKConfig) {
+	if h == nil {
+		return
+	}
+	h.runtimeMu.Lock()
+	h.Cfg = cfg
+	h.runtimeMu.Unlock()
+}
+
+// CurrentConfig returns the current application configuration snapshot.
+func (h *BaseAPIHandler) CurrentConfig() *config.SDKConfig {
+	if h == nil {
+		return nil
+	}
+	h.runtimeMu.RLock()
+	cfg := h.Cfg
+	h.runtimeMu.RUnlock()
+	return cfg
+}
 
 // SetPluginHost configures the optional plugin interceptor host.
 func (h *BaseAPIHandler) SetPluginHost(host PluginInterceptorHost) {
@@ -452,10 +472,11 @@ func (h *BaseAPIHandler) SetPluginHost(host PluginInterceptorHost) {
 		return
 	}
 	if isNilPluginInterceptorHost(host) {
-		h.PluginHost = nil
-		return
+		host = nil
 	}
+	h.runtimeMu.Lock()
 	h.PluginHost = host
+	h.runtimeMu.Unlock()
 }
 
 // SetModelRouterHost configures the optional plugin model router host.
@@ -464,10 +485,25 @@ func (h *BaseAPIHandler) SetModelRouterHost(host PluginModelRouterHost) {
 		return
 	}
 	if isNilPluginModelRouterHost(host) {
-		h.ModelRouterHost = nil
-		return
+		host = nil
 	}
+	h.runtimeMu.Lock()
 	h.ModelRouterHost = host
+	h.runtimeMu.Unlock()
+}
+
+// ModelRouterHostSnapshot returns the configured plugin model router host snapshot.
+func (h *BaseAPIHandler) ModelRouterHostSnapshot() PluginModelRouterHost {
+	if h == nil {
+		return nil
+	}
+	h.runtimeMu.RLock()
+	host := h.ModelRouterHost
+	h.runtimeMu.RUnlock()
+	if isNilPluginModelRouterHost(host) {
+		return nil
+	}
+	return host
 }
 
 func isNilPluginInterceptorHost(host PluginInterceptorHost) bool {
@@ -596,7 +632,7 @@ func (h *BaseAPIHandler) GetContextWithCancel(handler interfaces.APIHandler, c *
 		if c != nil {
 			logging.SetResponseStatus(cancelCtx, c.Writer.Status())
 		}
-		if h.Cfg.RequestLog && len(params) == 1 {
+		if cfg := h.CurrentConfig(); cfg != nil && cfg.RequestLog && len(params) == 1 {
 			if captured, exists := c.Get(logging.APIResponseCapturedContextKey); exists {
 				if capturedBool, ok := captured.(bool); ok && capturedBool {
 					cancel()
@@ -648,7 +684,7 @@ func (h *BaseAPIHandler) StartNonStreamingKeepAlive(c *gin.Context, ctx context.
 	if h == nil || c == nil {
 		return func() {}
 	}
-	interval := NonStreamingKeepAliveInterval(h.Cfg)
+	interval := NonStreamingKeepAliveInterval(h.CurrentConfig())
 	if interval <= 0 {
 		return func() {}
 	}

sdk/api/handlers/handlers_errors.go

@@ -128,7 +128,7 @@ func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.Erro
 			c.Writer.Header().Add("Retry-After", value)
 		}
 	}
-	if msg != nil && msg.Addon != nil && PassthroughHeadersEnabled(h.Cfg) {
+	if msg != nil && msg.Addon != nil && PassthroughHeadersEnabled(h.CurrentConfig()) {
 		for key, values := range msg.Addon {
 			if len(values) == 0 || IsCPAReservedResponseHeader(key) {
 				continue
@@ -196,7 +196,7 @@ func writeDirectErrorResponse(c *gin.Context, status int, msg *interfaces.ErrorM
 }
 
 func (h *BaseAPIHandler) LoggingAPIResponseError(ctx context.Context, err *interfaces.ErrorMessage) {
-	if h.Cfg.RequestLog {
+	if cfg := h.CurrentConfig(); cfg != nil && cfg.RequestLog {
 		if ginContext, ok := ctx.Value("gin").(*gin.Context); ok {
 			if apiResponseErrors, isExist := ginContext.Get("API_RESPONSE_ERROR"); isExist {
 				if slicesAPIResponseError, isOk := apiResponseErrors.([]*interfaces.ErrorMessage); isOk {

sdk/api/handlers/handlers_execution.go

@@ -112,7 +112,7 @@ func (h *BaseAPIHandler) executeWithAuthManagerFormats(ctx context.Context, entr
 	executedReq, executedOpts := afterAuthCapture.apply(req, opts)
 	ctx = enrichContextWithSessionHierarchy(ctx, executedOpts.Headers, executedReq.Payload, executedOpts.Metadata)
 	rawResponseHeaders := cloneHeader(resp.Headers)
-	passthroughHeaders := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeaders := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	responseHeaders := downstreamHeadersFromExecutor(rawResponseHeaders, passthroughHeaders)
 	body, responseHeaders := h.applyResponseInterceptors(ctx, lifecycle.requestID(), responseProtocol, normalizedModel, originalRequestedModel, executedOpts, rawResponseHeaders, responseHeaders, executedOpts.OriginalRequest, executedReq.Payload, resp.Payload, http.StatusOK, execOptions.SkipInterceptorPluginID, passthroughHeaders)
 	lifecycle.complete(pluginapi.RequestCompletionSucceeded, http.StatusOK, nil)
@@ -182,7 +182,7 @@ func (h *BaseAPIHandler) executeCountWithAuthManager(ctx context.Context, handle
 	executedReq, executedOpts := afterAuthCapture.apply(req, opts)
 	ctx = enrichContextWithSessionHierarchy(ctx, executedOpts.Headers, executedReq.Payload, executedOpts.Metadata)
 	rawResponseHeaders := cloneHeader(resp.Headers)
-	passthroughHeaders := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeaders := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	responseHeaders := downstreamHeadersFromExecutor(rawResponseHeaders, passthroughHeaders)
 	body, responseHeaders := h.applyResponseInterceptors(ctx, lifecycle.requestID(), handlerType, normalizedModel, originalRequestedModel, executedOpts, rawResponseHeaders, responseHeaders, executedOpts.OriginalRequest, executedReq.Payload, resp.Payload, http.StatusOK, execOptions.SkipInterceptorPluginID, passthroughHeaders)
 	lifecycle.complete(pluginapi.RequestCompletionSucceeded, http.StatusOK, nil)
@@ -232,7 +232,7 @@ func (h *BaseAPIHandler) executeWithPluginExecutor(ctx context.Context, entryPro
 		reporter.EnsurePublished(execCtx)
 	}
 	rawResponseHeaders := cloneHeader(resp.Headers)
-	passthroughHeaders := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeaders := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	responseHeaders := downstreamHeadersFromExecutor(rawResponseHeaders, passthroughHeaders)
 	body, responseHeaders := h.applyResponseInterceptors(execCtx, lifecycle.requestID(), responseProtocol, modelName, originalRequestedModel, opts, rawResponseHeaders, responseHeaders, opts.OriginalRequest, req.Payload, resp.Payload, http.StatusOK, execOptions.SkipInterceptorPluginID, passthroughHeaders)
 	lifecycle.complete(pluginapi.RequestCompletionSucceeded, http.StatusOK, nil)
@@ -268,7 +268,7 @@ func (h *BaseAPIHandler) countWithPluginExecutor(ctx context.Context, handlerTyp
 		return nil, nil, errMsg
 	}
 	rawResponseHeaders := cloneHeader(resp.Headers)
-	passthroughHeaders := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeaders := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	responseHeaders := downstreamHeadersFromExecutor(rawResponseHeaders, passthroughHeaders)
 	body, responseHeaders := h.applyResponseInterceptors(ctx, lifecycle.requestID(), handlerType, modelName, originalRequestedModel, opts, rawResponseHeaders, responseHeaders, opts.OriginalRequest, req.Payload, resp.Payload, http.StatusOK, execOptions.SkipInterceptorPluginID, passthroughHeaders)
 	lifecycle.complete(pluginapi.RequestCompletionSucceeded, http.StatusOK, nil)
@@ -400,11 +400,19 @@ func (h *BaseAPIHandler) pluginExecutorHost() PluginExecutorHost {
 	if h == nil {
 		return nil
 	}
-	if executorHost, ok := h.ModelRouterHost.(PluginExecutorHost); ok && executorHost != nil {
-		return executorHost
+	h.runtimeMu.RLock()
+	routerHost := h.ModelRouterHost
+	pluginHost := h.PluginHost
+	h.runtimeMu.RUnlock()
+	if !isNilPluginModelRouterHost(routerHost) {
+		if executorHost, ok := routerHost.(PluginExecutorHost); ok && executorHost != nil {
+			return executorHost
+		}
 	}
-	if executorHost, ok := h.PluginHost.(PluginExecutorHost); ok && executorHost != nil {
-		return executorHost
+	if !isNilPluginInterceptorHost(pluginHost) {
+		if executorHost, ok := pluginHost.(PluginExecutorHost); ok && executorHost != nil {
+			return executorHost
+		}
 	}
 	return nil
 }

sdk/api/handlers/handlers_interceptors.go

@@ -340,7 +340,13 @@ func (h *BaseAPIHandler) interceptorHost() PluginInterceptorHost {
 	if h == nil {
 		return nil
 	}
-	return h.PluginHost
+	h.runtimeMu.RLock()
+	host := h.PluginHost
+	h.runtimeMu.RUnlock()
+	if isNilPluginInterceptorHost(host) {
+		return nil
+	}
+	return host
 }
 
 func streamInterceptorsEnabled(host PluginInterceptorHost) bool {

sdk/api/handlers/handlers_routing.go

@@ -301,14 +301,17 @@ func (h *BaseAPIHandler) modelRouterHost() PluginModelRouterHost {
 	if h == nil {
 		return nil
 	}
-	if !isNilPluginModelRouterHost(h.ModelRouterHost) {
-		return h.ModelRouterHost
+	h.runtimeMu.RLock()
+	routerHost := h.ModelRouterHost
+	pluginHost := h.PluginHost
+	h.runtimeMu.RUnlock()
+	if !isNilPluginModelRouterHost(routerHost) {
+		return routerHost
 	}
-	host := h.interceptorHost()
-	if host == nil {
+	if isNilPluginInterceptorHost(pluginHost) {
 		return nil
 	}
-	router, ok := host.(PluginModelRouterHost)
+	router, ok := pluginHost.(PluginModelRouterHost)
 	if !ok {
 		return nil
 	}

sdk/api/handlers/handlers_stream.go

@@ -92,7 +92,7 @@ func (h *BaseAPIHandler) streamWithPluginExecutor(ctx context.Context, entryProt
 		return nil, nil, errChan
 	}
 
-	passthroughHeadersEnabled := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeadersEnabled := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	interceptorHost := h.interceptorHost()
 	streamInterceptorsActive := streamInterceptorsEnabled(interceptorHost)
 	rawStreamHeaders := cloneHeader(streamResult.Headers)
@@ -375,7 +375,7 @@ func (h *BaseAPIHandler) executeStreamWithAuthManagerFormats(ctx context.Context
 	if executedReq, executedOpts := executedRequest(); len(executedOpts.Headers) > 0 || len(executedReq.Payload) > 0 || len(executedOpts.Metadata) > 0 {
 		ctx = enrichContextWithSessionHierarchy(ctx, executedOpts.Headers, executedReq.Payload, executedOpts.Metadata)
 	}
-	passthroughHeadersEnabled := executionPassthroughHeaders(h.Cfg, execOptions.InternalSource)
+	passthroughHeadersEnabled := executionPassthroughHeaders(h.CurrentConfig(), execOptions.InternalSource)
 	interceptorHost := h.interceptorHost()
 	streamInterceptorsActive := streamInterceptorsEnabled(interceptorHost)
 	// Resolve bootstrap retries and header initialization before returning so the
@@ -539,7 +539,7 @@ func (h *BaseAPIHandler) executeStreamWithAuthManagerFormats(ctx context.Context
 		}
 	}
 
-	maxBootstrapRetries := StreamingBootstrapRetries(h.Cfg)
+	maxBootstrapRetries := StreamingBootstrapRetries(h.CurrentConfig())
 	if h.AuthManager.HomeEnabled() {
 		maxBootstrapRetries = 0
 	}

sdk/api/handlers/openai/codex_client_models.go

@@ -11,10 +11,15 @@ func (h *OpenAIAPIHandler) codexClientModelsResponse(clientVersion ...string) ma
 	if len(clientVersion) > 0 {
 		version = clientVersion[0]
 	}
-	optimizeMultiAgentV2 := h != nil && h.Cfg != nil && h.Cfg.Client.Codex.OptimizeMultiAgentV2
+	optimizeMultiAgentV2 := false
 	var applyPatchCapabilityForModel codexmodels.ApplyPatchCapabilityForModelFunc
-	if h != nil && h.Cfg != nil && h.Cfg.Client.Codex.EnableApplyPatch {
-		applyPatchCapabilityForModel = h.SupportsApplyPatchModel
+	if h != nil {
+		if cfg := h.CurrentConfig(); cfg != nil {
+			optimizeMultiAgentV2 = cfg.Client.Codex.OptimizeMultiAgentV2
+			if cfg.Client.Codex.EnableApplyPatch {
+				applyPatchCapabilityForModel = h.SupportsApplyPatchModel
+			}
+		}
 	}
 	var toolSearchCapabilityForModel codexmodels.ToolSearchCapabilityForModelFunc
 	if h != nil && responsesclaude.ClientToolSearchEnabled() {

sdk/api/handlers/openai/openai_handlers.go

@@ -507,7 +507,7 @@ func (h *OpenAIAPIHandler) handleStreamingResponse(c *gin.Context, rawJSON []byt
 		flusher.Flush()
 		return true
 	}
-	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	defer stopKeepAlive()
 
 	dataChan, resolvedHeaders, errChan, okBootstrap := h.ExecuteStreamWithPreFTKeepAlive(
@@ -666,7 +666,7 @@ func (h *OpenAIAPIHandler) handleCompletionsStreamingResponse(c *gin.Context, ra
 		flusher.Flush()
 		return true
 	}
-	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	defer stopKeepAlive()
 
 	dataChan, resolvedHeaders, errChan, okBootstrap := h.ExecuteStreamWithPreFTKeepAlive(

sdk/api/handlers/openai/openai_images_handlers.go

@@ -79,7 +79,7 @@ func (h *OpenAIAPIHandler) newImagesStreamKeepAliveTicker() (func(), <-chan time
 	if h == nil || h.BaseAPIHandler == nil {
 		return func() {}, nil
 	}
-	keepAliveC, stop := handlers.NewStreamingKeepAliveTicker(handlers.StreamingKeepAliveInterval(h.Cfg))
+	keepAliveC, stop := handlers.NewStreamingKeepAliveTicker(handlers.StreamingKeepAliveInterval(h.CurrentConfig()))
 	return stop, keepAliveC
 }
 
@@ -87,7 +87,7 @@ func (h *OpenAIAPIHandler) newImagesPreFirstTokenKeepAliveTicker() (func(), <-ch
 	if h == nil || h.BaseAPIHandler == nil {
 		return func() {}, nil
 	}
-	keepAliveC, stop := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stop := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	return stop, keepAliveC
 }
 
@@ -624,9 +624,11 @@ func parseBoolField(raw string, fallback bool) bool {
 }
 
 func (h *OpenAIAPIHandler) ImagesGenerations(c *gin.Context) {
-	if h != nil && h.BaseAPIHandler != nil && h.BaseAPIHandler.Cfg != nil && h.BaseAPIHandler.Cfg.DisableImageGeneration == internalconfig.DisableImageGenerationAll {
-		c.AbortWithStatus(http.StatusNotFound)
-		return
+	if h != nil {
+		if cfg := h.CurrentConfig(); cfg != nil && cfg.DisableImageGeneration == internalconfig.DisableImageGenerationAll {
+			c.AbortWithStatus(http.StatusNotFound)
+			return
+		}
 	}
 
 	rawJSON, err := handlers.ReadRequestBody(c)
@@ -728,9 +730,11 @@ func (h *OpenAIAPIHandler) ImagesGenerations(c *gin.Context) {
 }
 
 func (h *OpenAIAPIHandler) ImagesEdits(c *gin.Context) {
-	if h != nil && h.BaseAPIHandler != nil && h.BaseAPIHandler.Cfg != nil && h.BaseAPIHandler.Cfg.DisableImageGeneration == internalconfig.DisableImageGenerationAll {
-		c.AbortWithStatus(http.StatusNotFound)
-		return
+	if h != nil {
+		if cfg := h.CurrentConfig(); cfg != nil && cfg.DisableImageGeneration == internalconfig.DisableImageGenerationAll {
+			c.AbortWithStatus(http.StatusNotFound)
+			return
+		}
 	}
 
 	contentType := strings.ToLower(strings.TrimSpace(c.GetHeader("Content-Type")))

sdk/api/handlers/openai/openai_responses_handlers.go

@@ -551,7 +551,11 @@ func (h *OpenAIResponsesAPIHandler) OpenAIResponsesModels(c *gin.Context) {
 }
 
 func (h *OpenAIResponsesAPIHandler) prepareCodexMultiAgentV2Tools(c *gin.Context, payload []byte) []byte {
-	if h == nil || h.Cfg == nil {
+	if h == nil {
+		return payload
+	}
+	cfg := h.CurrentConfig()
+	if cfg == nil {
 		return payload
 	}
 
@@ -570,7 +574,7 @@ func (h *OpenAIResponsesAPIHandler) prepareCodexMultiAgentV2Tools(c *gin.Context
 		requestCtx,
 		requestHeaders,
 		payload,
-		h.Cfg.Client.Codex.OptimizeMultiAgentV2,
+		cfg.Client.Codex.OptimizeMultiAgentV2,
 		homeEnabled,
 	)
 	if prepared && c != nil {
@@ -580,7 +584,11 @@ func (h *OpenAIResponsesAPIHandler) prepareCodexMultiAgentV2Tools(c *gin.Context
 }
 
 func (h *OpenAIResponsesAPIHandler) prepareCodexOrphanDelegation(c *gin.Context, payload []byte) []byte {
-	if h == nil || h.Cfg == nil || !h.Cfg.CodexOrphanDelegationCompatibility {
+	if h == nil {
+		return payload
+	}
+	cfg := h.CurrentConfig()
+	if cfg == nil || !cfg.CodexOrphanDelegationCompatibility {
 		return payload
 	}
 	requestCtx := context.Background()
@@ -754,7 +762,7 @@ func (h *OpenAIResponsesAPIHandler) handleStreamingResponse(c *gin.Context, rawJ
 		}
 		return true
 	}
-	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.Cfg)
+	keepAliveC, stopKeepAlive := handlers.NewPreFirstTokenKeepAliveTicker(h.CurrentConfig())
 	defer stopKeepAlive()
 
 	dataChan, resolvedHeaders, errChan, okBootstrap := h.ExecuteStreamWithPreFTKeepAlive(

sdk/api/handlers/openai/openai_responses_websocket.go

@@ -278,14 +278,21 @@ func (h *OpenAIResponsesAPIHandler) ResponsesWebsocket(c *gin.Context) {
 	c.Request = c.Request.WithContext(socketCtx)
 	// The reader owns the client socket so an in-flight interrupt is not stuck
 	// behind the response currently being forwarded.
-	requestLogEnabled := h != nil && h.Cfg != nil && h.Cfg.RequestLog
+	requestLogEnabled := false
+	codexResponseSteering := false
+	if h != nil {
+		if cfg := h.CurrentConfig(); cfg != nil {
+			requestLogEnabled = cfg.RequestLog
+			codexResponseSteering = cfg.CodexResponseSteering
+		}
+	}
 	wsTimelineLog := newWebsocketTimelineLog(requestLogEnabled, websocketTimelineSourceFromContext(c))
 	localInterrupt := newResponsesLocalInterrupt()
 	input, inputDone := readResponsesWebsocketInput(socketCtx, cancelSocket, conn, func(payload []byte) error {
 		return h.forwardResponsesWebsocketInterrupt(socketCtx, passthroughSessionID, payload)
 	}, localInterrupt, writer, wsTimelineLog)
 	var duplexInput <-chan cliproxyexecutor.WebsocketInput
-	if h != nil && h.Cfg != nil && h.Cfg.CodexResponseSteering {
+	if codexResponseSteering {
 		duplexInput = input
 	}
 	downstreamSessionKey := websocketDownstreamSessionKey(c.Request)

sdk/api/handlers/openai/openai_responses_websocket_forward.go

@@ -61,7 +61,7 @@ func (h *OpenAIResponsesAPIHandler) forwardResponsesWebsocket(
 	var keepAliveC <-chan time.Time
 	keepAliveInterval := time.Duration(0)
 	if h != nil {
-		keepAliveInterval = handlers.StreamingKeepAliveInterval(h.Cfg)
+		keepAliveInterval = handlers.StreamingKeepAliveInterval(h.CurrentConfig())
 	}
 	if opts.keepAliveInterval != nil {
 		keepAliveInterval = *opts.keepAliveInterval

sdk/api/handlers/openai/openai_videos_handlers.go

@@ -296,11 +296,13 @@ func firstPostForm(c *gin.Context, keys ...string) string {
 }
 
 func (h *OpenAIAPIHandler) videoAuthBindingTTL() time.Duration {
-	if h != nil && h.BaseAPIHandler != nil && h.Cfg != nil {
-		raw := strings.TrimSpace(h.Cfg.VideoResultAuthCacheTTL)
-		if raw != "" {
-			if ttl, err := time.ParseDuration(raw); err == nil && ttl > 0 {
-				return ttl
+	if h != nil {
+		if cfg := h.CurrentConfig(); cfg != nil {
+			raw := strings.TrimSpace(cfg.VideoResultAuthCacheTTL)
+			if raw != "" {
+				if ttl, err := time.ParseDuration(raw); err == nil && ttl > 0 {
+					return ttl
+				}
 			}
 		}
 	}
@@ -947,8 +949,10 @@ func (h *OpenAIAPIHandler) videoContentHTTPClient(c *gin.Context) *http.Client {
 		ctx = c.Request.Context()
 	}
 	var cfg *config.Config
-	if h != nil && h.BaseAPIHandler != nil && h.Cfg != nil {
-		cfg = &config.Config{SDKConfig: *h.Cfg}
+	if h != nil {
+		if sdkCfg := h.CurrentConfig(); sdkCfg != nil {
+			cfg = &config.Config{SDKConfig: *sdkCfg}
+		}
 	}
 	return helps.NewProxyAwareHTTPClient(ctx, cfg, h.videoContentDownloadAuth(c), 0)
 }

sdk/api/handlers/stream_forwarder.go

@@ -148,7 +148,7 @@ func (h *BaseAPIHandler) ForwardStream(c *gin.Context, flusher http.Flusher, can
 		}
 	}
 
-	keepAliveInterval := StreamingKeepAliveInterval(h.Cfg)
+	keepAliveInterval := StreamingKeepAliveInterval(h.CurrentConfig())
 	if opts.KeepAliveInterval != nil {
 		keepAliveInterval = *opts.KeepAliveInterval
 	}