@morojs/moro 1.8.9 → 1.8.11

package/dist/moro.js

@@ -9,11 +9,16 @@ import { UnifiedRouter, } from './core/routing/unified-router.js';
 import { PathMatcher } from './core/routing/path-matcher.js';
 import { AppDocumentationManager } from './core/docs/index.js';
 import { EventEmitter } from 'events';
-import cluster from 'cluster';
-import os from 'os';
+import { requireBuiltin } from './core/utilities/builtin.js';
+// node:cluster (it drags in child_process, dgram and net) is loaded only when
+// process clustering actually runs; worker_threads only on the clustering
+// path. See core/utilities/builtin.ts.
+const nodeCluster = () => requireBuiltin('cluster');
+const wt = () => requireBuiltin('worker_threads');
+import { ThreadClusterPrimary, ThreadClusterWorker, computeWorkerCount, isUserTunedYoungGen, resolveClusterTransport, threadWorkerInfo, YOUNG_GEN_MB_FOR_SEMI_SPACE_4, } from './core/cluster/thread-cluster.js';
 import { normalizeValidationError } from './core/validation/schema-interface.js';
 import { buildModuleBasePath } from './core/utilities/module-path.js';
-import { filePathToImportURL } from './core/utilities/package-utils.js';
+import { filePathToImportURL, loadNativeEngine } from './core/utilities/package-utils.js';
 // Configuration System Integration
 import { initializeConfig, initializeConfigAsync } from './core/config/index.js';
 // Runtime System Integration
@@ -339,7 +344,7 @@ export class Moro extends EventEmitter {
         };
     }
     get(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             // Direct route registration
             return this.addRoute('GET', path, handler, options);
         }
@@ -347,7 +352,7 @@ export class Moro extends EventEmitter {
         return this.unifiedRouter.get(path);
     }
     post(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             // Direct route registration
             return this.addRoute('POST', path, handler, options);
         }
@@ -355,7 +360,7 @@ export class Moro extends EventEmitter {
         return this.unifiedRouter.post(path);
     }
     put(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             // Direct route registration
             return this.addRoute('PUT', path, handler, options);
         }
@@ -363,7 +368,7 @@ export class Moro extends EventEmitter {
         return this.unifiedRouter.put(path);
     }
     delete(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             // Direct route registration
             return this.addRoute('DELETE', path, handler, options);
         }
@@ -371,7 +376,7 @@ export class Moro extends EventEmitter {
         return this.unifiedRouter.delete(path);
     }
     patch(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             // Direct route registration
             return this.addRoute('PATCH', path, handler, options);
         }
@@ -379,13 +384,13 @@ export class Moro extends EventEmitter {
         return this.unifiedRouter.patch(path);
     }
     head(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             return this.addRoute('HEAD', path, handler, options);
         }
         return this.unifiedRouter.head(path);
     }
     options(path, handler, options) {
-        if (handler) {
+        if (handler !== undefined) {
             return this.addRoute('OPTIONS', path, handler, options);
         }
         return this.unifiedRouter.options(path);
@@ -778,7 +783,8 @@ export class Moro extends EventEmitter {
         }
         // Check if clustering is enabled for massive performance gains
         if (this.config.performance?.clustering?.enabled) {
-            this.logger.info('Clustering enabled - using Node.js cluster', 'Cluster');
+            // Worker threads (engine backend, POSIX, engine >= 1.1.6) or node:cluster
+            // processes - startWithClustering picks and logs the transport.
             this.startWithClustering(port, host, callback);
             return;
         }
@@ -1231,6 +1237,23 @@ export class Moro extends EventEmitter {
     }
     // Private methods
     addRoute(method, path, handler, options = {}) {
+        // A literal body - app.get('/', '') - is the same route the builder's
+        // .handler('') registers: the unified router turns it into a handler and,
+        // with nothing else configured, a reply the native engine answers itself.
+        // It never enters the legacy direct-route table, whose wrapper would
+        // call it as a function.
+        if (typeof handler !== 'function') {
+            this.unifiedRouter.registerRoute({
+                method: method,
+                path,
+                handler: handler,
+                validation: options.validation,
+                rateLimit: options.rateLimit,
+                cache: options.cache,
+                middleware: options.middleware,
+            });
+            return this;
+        }
         // Register with unified router (primary routing system)
         this.unifiedRouter.addRoute(method, path, handler, options.middleware || []);
         // Also store in legacy routes array for backward compatibility
@@ -1276,6 +1299,19 @@ export class Moro extends EventEmitter {
         if (httpServer && typeof httpServer.setRouterHandler === 'function') {
             const router = this.unifiedRouter;
             httpServer.setRouterHandler((req, res) => router.handleRequest(req, res));
+            // A server that answers fixed replies itself (MoroEngineServer on an
+            // engine with capabilities.staticRoutes) gets every route whose handler
+            // is a literal body, so those requests never enter JS. The route's
+            // handler stays registered: it answers HEAD, and everything on servers
+            // without the capability.
+            if (typeof httpServer.setStaticRoute === 'function') {
+                for (const schema of router.getAllRoutes()) {
+                    const fixed = schema.static;
+                    if (fixed) {
+                        httpServer.setStaticRoute(schema.method, schema.path, fixed.status, fixed.headers, fixed.body);
+                    }
+                }
+            }
             // Servers with a native router (uWS) additionally get fast-path routes
             // registered directly on it - evaluated at listen so late-loaded routes
             // are included; error handling shares the unified router's semantics
@@ -1652,150 +1688,275 @@ export class Moro extends EventEmitter {
      * - BetterStack Node.js Clustering: https://betterstack.com/community/guides/scaling-nodejs/node-clustering/
      */
     clusterWorkers = new Map();
+    threadCluster = undefined;
+    // Entry point for performance.clustering: decides whether THIS instance is
+    // a worker (thread or process) or the primary, and which transport the
+    // primary uses. Config keys are unchanged; the transport is automatic.
     startWithClustering(port, host, callback) {
-        // Worker count calculation - respect user choice
-        let workerCount = this.config.performance?.clustering?.workers || os.cpus().length;
-        // Only auto-optimize if user hasn't specified a number or set it to 'auto'
-        if (workerCount === 'auto') {
-            const cpuCount = os.cpus().length;
-            const totalMemoryGB = os.totalmem() / (1024 * 1024 * 1024);
-            // Get memory per worker from config - if not set by user, calculate dynamically
-            let memoryPerWorkerGB = this.config.performance?.clustering?.memoryPerWorkerGB;
-            if (!memoryPerWorkerGB) {
-                // Dynamic calculation: (Total RAM - 4GB headroom) / CPU cores
-                const headroomGB = 4;
-                memoryPerWorkerGB = Math.max(0.5, Math.floor((totalMemoryGB - headroomGB) / cpuCount));
-            }
-            // Conservative formula based on general guidelines:
-            // - Don't exceed CPU cores
-            // - Respect user's memory allocation preference
-            // - Let the system resources determine the limit
-            workerCount = Math.min(cpuCount, // Don't exceed CPU cores
-            Math.floor(totalMemoryGB / memoryPerWorkerGB) // User-configurable memory per worker
-            );
-            this.logger.info(`Auto-calculated worker count: ${workerCount} (CPU: ${cpuCount}, RAM: ${totalMemoryGB.toFixed(1)}GB, ${memoryPerWorkerGB}GB per worker)`, 'Cluster');
-        }
-        else if (typeof workerCount === 'number') {
-            // User specified a number - respect their choice
-            this.logger.info(`Using user-specified worker count: ${workerCount}`, 'Cluster');
-        }
-        if (cluster.isPrimary) {
-            this.logger.info(`Starting ${workerCount} workers`, 'Cluster');
-            // Optimize cluster scheduling for high concurrency
-            // Round-robin is the default on all platforms except Windows (Node.js docs)
-            // Provides better load distribution than shared socket approach
-            cluster.schedulingPolicy = cluster.SCHED_RR;
-            // Workers inherit the primary's execArgv; additionally bound each
-            // worker's V8 young generation unless the user tuned it themselves.
-            // On big machines V8 sizes the nursery from total RAM and grows it
-            // under sustained load (~10-34 MB extra RSS per process, measured on a
-            // 64 GB box) - and in cluster mode that cost multiplies by the worker
-            // count. A/B-benched 2026-08-01: identical throughput at 4 MB semi-space
-            // in both the realistic profile (~92k vs ~90k req/s) and the pipelined
-            // x10 microbenchmark (~560k both ways), while sustained-load RSS drops
-            // ~30 MB per worker (115 -> 81 MB on the 25-route POST soak).
-            const userTunedGC = process.execArgv.some(a => a.includes('--max-semi-space-size')) ||
-                (process.env.NODE_OPTIONS || '').includes('--max-semi-space-size');
-            // Set cluster settings for better performance
-            cluster.setupMaster({
-                exec: process.argv[1] || process.execPath,
-                args: process.argv.slice(2),
-                silent: false,
-                ...(userTunedGC ? {} : { execArgv: [...process.execArgv, '--max-semi-space-size=4'] }),
+        const { count: workerCount, detail } = computeWorkerCount(this.config.performance?.clustering);
+        const tinfo = threadWorkerInfo();
+        if (tinfo) {
+            this.startClusterWorker({ kind: 'thread', info: tinfo }, port, host);
+            return;
+        }
+        if (nodeCluster().isWorker) {
+            this.startClusterWorker({ kind: 'process' }, port, host);
+            return;
+        }
+        this.logger.info(detail, 'Cluster');
+        const engineKind = this.engine;
+        const transport = resolveClusterTransport({
+            platform: process.platform,
+            isMainThread: wt().isMainThread,
+            argv1: process.argv[1],
+            engineServer: engineKind.server,
+            enginePackage: engineKind.enginePackage,
+            workerThreadsCapable: loadNativeEngine()?.capabilities?.workerThreads === true,
+        });
+        if (transport.kind === 'threads') {
+            this.startThreadPrimary(transport.entry, workerCount, port, host, callback);
+            return;
+        }
+        this.logger.info(`Clustering: ${workerCount} worker processes (node:cluster; ${transport.reason})`, 'Cluster');
+        this.startProcessPrimary(workerCount, port, host, callback);
+    }
+    // Services that must run exactly ONCE per cluster, in the primary: the job
+    // scheduler (leader election already treats the main thread of a
+    // non-cluster-worker process as a participant) and the gRPC server (its own
+    // port). GraphQL registers HTTP/WS routes, so it initialises in every
+    // worker instead. Before 1.9.0 none of these started under clustering.
+    startPrimaryServices() {
+        this.startJobScheduler().catch(err => {
+            this.logger.error(`Failed to start job scheduler: ${String(err)}`);
+        });
+        if (this.grpcManager && !this.grpcStarted) {
+            this.startGrpc().catch(error => {
+                this.logger.error(`Failed to start gRPC server: ${error}`, 'GRPC');
             });
-            // IPC Optimization: Reduce communication overhead between master and workers
-            // Research shows excessive IPC can create bottlenecks in clustered applications
-            // (Source: BetterStack - Node.js Clustering Guide)
-            process.env.NODE_CLUSTER_SCHED_POLICY = 'rr'; // Ensure round-robin
-            process.env.NODE_DISABLE_COLORS = '1'; // Reduce IPC message size by disabling color codes
-            // Graceful shutdown handler
-            const gracefulShutdown = () => {
-                this.logger.info('Gracefully shutting down cluster...', 'Cluster');
-                // Clean up all workers
-                // eslint-disable-next-line @typescript-eslint/no-unused-vars
-                for (const [pid, worker] of this.clusterWorkers) {
-                    worker.removeAllListeners();
-                    worker.kill('SIGTERM');
+        }
+    }
+    startThreadPrimary(entry, workerCount, port, host, callback) {
+        // The libuv threadpool is process-wide and sized once at first use;
+        // the process path set this per worker process. Only if the operator
+        // did not choose a size.
+        if (!process.env.UV_THREADPOOL_SIZE)
+            process.env.UV_THREADPOOL_SIZE = '64';
+        // Same young-generation bound the process path applies through execArgv
+        // (see YOUNG_GEN_MB_FOR_SEMI_SPACE_4), unless the operator tuned GC.
+        const resourceLimits = isUserTunedYoungGen()
+            ? undefined
+            : { maxYoungGenerationSizeMb: YOUNG_GEN_MB_FOR_SEMI_SPACE_4 };
+        const engineKind = this.engine;
+        this.logger.info(`Clustering: ${workerCount} worker threads (worker_threads + SO_REUSEPORT, ` +
+            `${engineKind.enginePackage}${engineKind.engineVersion ? ` v${engineKind.engineVersion}` : ''})`, 'Cluster');
+        let fellBack = false;
+        const fallBackToProcesses = (why) => {
+            if (fellBack)
+                return;
+            fellBack = true;
+            this.logger.warn(`Worker threads unavailable (${why}); falling back to worker processes`, 'Cluster');
+            const primary = this.threadCluster;
+            this.threadCluster = undefined;
+            void primary?.shutdown(1000).finally(() => {
+                this.startProcessPrimary(workerCount, port, host, callback);
+            });
+        };
+        const primary = new ThreadClusterPrimary({
+            entry,
+            argv: process.argv.slice(2),
+            workerCount,
+            logger: this.logger,
+            ...(resourceLimits ? { resourceLimits } : {}),
+            onReady: () => {
+                this.startPrimaryServices();
+                if (callback)
+                    callback();
+            },
+            onFatal: (error, kind) => {
+                if (kind === 'bind') {
+                    // Processes would fail identically: surface it like the single-process
+                    // path does (Node's server convention; rethrows without a listener).
+                    this.emit('error', error);
+                    return;
                 }
-                // Clean up cluster listeners
-                cluster.removeAllListeners();
+                fallBackToProcesses(error.message);
+            },
+        });
+        this.threadCluster = primary;
+        try {
+            primary.start();
+        }
+        catch (error) {
+            fallBackToProcesses(error instanceof Error ? error.message : String(error));
+            return;
+        }
+        // Graceful shutdown: drain every worker thread (each closes its server),
+        // then the primary's own services; a second signal exits at once.
+        let shuttingDown = false;
+        const gracefulShutdown = (code) => {
+            if (shuttingDown) {
+                process.exit(code);
+                return;
+            }
+            shuttingDown = true;
+            this.logger.info('Gracefully shutting down cluster...', 'Cluster');
+            this.close()
+                .then(() => process.exit(code))
+                .catch(() => process.exit(1));
+        };
+        process.on('SIGINT', () => gracefulShutdown(0));
+        process.on('SIGTERM', () => gracefulShutdown(0));
+    }
+    startProcessPrimary(workerCount, port, host, callback) {
+        void port;
+        void host;
+        this.logger.info(`Starting ${workerCount} workers`, 'Cluster');
+        // Optimize cluster scheduling for high concurrency
+        // Round-robin is the default on all platforms except Windows (Node.js docs)
+        // Provides better load distribution than shared socket approach
+        nodeCluster().schedulingPolicy = nodeCluster().SCHED_RR;
+        // Workers inherit the primary's execArgv; additionally bound each
+        // worker's V8 young generation unless the user tuned it themselves.
+        // On big machines V8 sizes the nursery from total RAM and grows it
+        // under sustained load (~10-34 MB extra RSS per process, measured on a
+        // 64 GB box) - and in cluster mode that cost multiplies by the worker
+        // count. A/B-benched 2026-08-01: identical throughput at 4 MB semi-space
+        // in both the realistic profile (~92k vs ~90k req/s) and the pipelined
+        // x10 microbenchmark (~560k both ways), while sustained-load RSS drops
+        // ~30 MB per worker (115 -> 81 MB on the 25-route POST soak).
+        const userTunedGC = isUserTunedYoungGen();
+        // Set cluster settings for better performance
+        nodeCluster().setupMaster({
+            exec: process.argv[1] || process.execPath,
+            args: process.argv.slice(2),
+            silent: false,
+            ...(userTunedGC ? {} : { execArgv: [...process.execArgv, '--max-semi-space-size=4'] }),
+        });
+        // IPC Optimization: Reduce communication overhead between master and workers
+        // Research shows excessive IPC can create bottlenecks in clustered applications
+        // (Source: BetterStack - Node.js Clustering Guide)
+        process.env.NODE_CLUSTER_SCHED_POLICY = 'rr'; // Ensure round-robin
+        process.env.NODE_DISABLE_COLORS = '1'; // Reduce IPC message size by disabling color codes
+        // Graceful shutdown: SIGTERM every worker (each stops accepting, drains
+        // its in-flight responses, then exits), wait for their exits up to a
+        // bounded deadline, and only then exit. Exiting right after the kill
+        // used to close the cluster IPC channel, and a node:cluster worker whose
+        // primary disappears exits immediately - cutting the very responses the
+        // worker was draining. A second signal while draining exits at once.
+        let shuttingDown = false;
+        const gracefulShutdown = () => {
+            if (shuttingDown)
+                process.exit(0);
+            shuttingDown = true;
+            this.logger.info('Gracefully shutting down cluster...', 'Cluster');
+            if (this.clusterWorkers.size === 0) {
+                nodeCluster().removeAllListeners();
+                process.exit(0);
+            }
+            const pending = new Set(this.clusterWorkers.keys());
+            const timer = setTimeout(() => {
+                this.logger.warn(`Shutdown deadline reached; terminating ${pending.size} worker process(es)`, 'Cluster');
+                for (const pid of pending)
+                    this.clusterWorkers.get(pid)?.kill('SIGKILL');
+                finish();
+            }, 3000);
+            const finish = () => {
+                clearTimeout(timer);
+                nodeCluster().removeAllListeners();
                 process.exit(0);
             };
-            // Handle process signals for graceful shutdown
-            process.on('SIGINT', gracefulShutdown);
-            process.on('SIGTERM', gracefulShutdown);
-            // Fork workers with basic tracking
-            for (let i = 0; i < workerCount; i++) {
-                const worker = cluster.fork();
+            for (const [pid, worker] of this.clusterWorkers) {
+                worker.removeAllListeners(); // no restart on these exits
+                worker.once('exit', () => {
+                    pending.delete(pid);
+                    if (pending.size === 0)
+                        finish();
+                });
+                worker.kill('SIGTERM');
+            }
+        };
+        // Handle process signals for graceful shutdown
+        process.on('SIGINT', gracefulShutdown);
+        process.on('SIGTERM', gracefulShutdown);
+        // Fork workers with basic tracking
+        for (let i = 0; i < workerCount; i++) {
+            const worker = nodeCluster().fork();
+            // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+            this.clusterWorkers.set(worker.process.pid, worker);
+            this.logger.info(`Worker ${worker.process.pid} started`, 'Cluster');
+            // Handle individual worker messages
+            worker.on('message', this.handleWorkerMessage.bind(this));
+        }
+        // Primary-only services start once the first worker is serving.
+        let servicesStarted = false;
+        nodeCluster().on('listening', () => {
+            if (servicesStarted)
+                return;
+            servicesStarted = true;
+            this.startPrimaryServices();
+        });
+        // Simple worker exit handling
+        nodeCluster().on('exit', (worker, code, signal) => {
+            const pid = worker.process.pid;
+            this.clusterWorkers.delete(pid);
+            if (code !== 0 && !worker.exitedAfterDisconnect) {
+                this.logger.warn(`Worker ${pid} died unexpectedly (${signal || code}). Restarting...`, 'Cluster');
+                // Simple restart
+                const newWorker = nodeCluster().fork();
                 // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
-                this.clusterWorkers.set(worker.process.pid, worker);
-                this.logger.info(`Worker ${worker.process.pid} started`, 'Cluster');
-                // Handle individual worker messages
-                worker.on('message', this.handleWorkerMessage.bind(this));
-            }
-            // Simple worker exit handling
-            cluster.on('exit', (worker, code, signal) => {
-                const pid = worker.process.pid;
-                this.clusterWorkers.delete(pid);
-                if (code !== 0 && !worker.exitedAfterDisconnect) {
-                    this.logger.warn(`Worker ${pid} died unexpectedly (${signal || code}). Restarting...`, 'Cluster');
-                    // Simple restart
-                    const newWorker = cluster.fork();
-                    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
-                    this.clusterWorkers.set(newWorker.process.pid, newWorker);
-                    this.logger.info(`Worker ${newWorker.process.pid} restarted`, 'Cluster');
-                }
+                this.clusterWorkers.set(newWorker.process.pid, newWorker);
+                this.logger.info(`Worker ${newWorker.process.pid} restarted`, 'Cluster');
+            }
+        });
+        // Master process callback
+        if (callback)
+            callback();
+    }
+    // One worker, either a worker_threads thread or a node:cluster process:
+    // boot the app exactly like the single-process path (auto-discovery,
+    // file-based routes, WebSocket registrations, GraphQL) and listen with
+    // SO_REUSEPORT. Signals/exit are the differences: a thread gets 'shutdown'
+    // from the primary over parentPort and never calls process.exit().
+    startClusterWorker(mode, port, host) {
+        const label = mode.kind === 'thread' ? `Worker thread ${wt().threadId}` : `Worker ${process.pid}`;
+        const workerRef = mode.kind === 'thread' ? wt().threadId : process.pid;
+        this.logger.info(`${label} initializing`, 'Worker');
+        if (mode.kind === 'process') {
+            // Worker-specific optimizations for high concurrency (a thread shares
+            // the primary's pool, sized before it was spawned)
+            process.env.UV_THREADPOOL_SIZE = '64';
+        }
+        // Reduce logging contention in workers (major bottleneck)
+        // Multiple workers writing to same log files creates I/O contention
+        // ONLY reduce log level if user didn't explicitly set one
+        if (!this.userSetLogger) {
+            // Workers log less frequently to reduce I/O contention (only if not explicitly configured)
+            applyLoggingConfiguration(undefined, { level: 'warn' }); // Only warnings and errors
+        }
+        // Optimize garbage collection for workers
+        // eslint-disable-next-line no-undef
+        if (global.gc) {
+            setInterval(() => {
+                // eslint-disable-next-line no-undef
+                if (global.gc)
+                    global.gc();
+            }, 60000); // GC every 60 seconds (less frequent)
+        }
+        let tcw = null;
+        if (mode.kind === 'thread') {
+            tcw = new ThreadClusterWorker(this.logger);
+            // 'shutdown' from the primary: drain and close; the thread then exits
+            // on its own once the engine's handles are gone.
+            tcw.onShutdown(async () => {
+                this.eventBus.removeAllListeners();
+                this.removeAllListeners();
+                await this.close();
             });
-            // Master process callback
-            if (callback)
-                callback();
         }
         else {
-            // Worker process - start the actual server with proper cleanup
-            this.logger.info(`Worker ${process.pid} initializing`, 'Worker');
-            // Worker-specific optimizations for high concurrency
-            process.env.UV_THREADPOOL_SIZE = '64';
-            // Reduce logging contention in workers (major bottleneck)
-            // Multiple workers writing to same log files creates I/O contention
-            // ONLY reduce log level if user didn't explicitly set one
-            if (!this.userSetLogger) {
-                // Workers log less frequently to reduce I/O contention (only if not explicitly configured)
-                applyLoggingConfiguration(undefined, { level: 'warn' }); // Only warnings and errors
-            }
-            // Research-based memory optimization for workers
-            const totalMemoryGB = os.totalmem() / (1024 * 1024 * 1024);
-            const workerCount = Object.keys(cluster.workers || {}).length || 1;
-            // Conservative memory allocation
-            const heapSizePerWorkerMB = Math.min(Math.floor(((totalMemoryGB * 1024) / workerCount) * 0.8), // 80% of available memory
-            1536 // Cap at 1.5GB (GC efficiency threshold from research)
-            );
-            process.env.NODE_OPTIONS = `--max-old-space-size=${heapSizePerWorkerMB}`;
-            this.logger.debug(`Worker memory allocated: ${heapSizePerWorkerMB}MB heap (${workerCount} workers, ${totalMemoryGB.toFixed(1)}GB total)`, 'Worker');
-            // Optimize V8 flags for better performance
-            if (process.env.NODE_ENV === 'production') {
-                // Aggressive V8 optimizations for maximum performance
-                const v8Flags = [
-                    '--optimize-for-size', // Trade memory for speed
-                    '--always-opt', // Always optimize functions
-                    '--turbo-fast-api-calls', // Optimize API calls
-                    '--turbo-escape-analysis', // Escape analysis optimization
-                    '--turbo-inline-api-calls', // Inline API calls
-                    `--max-old-space-size=${heapSizePerWorkerMB}`, // Limit memory to prevent GC pressure
-                ];
-                process.env.NODE_OPTIONS = (process.env.NODE_OPTIONS || '') + ' ' + v8Flags.join(' ');
-            }
-            // Optimize garbage collection for workers
-            // eslint-disable-next-line no-undef
-            if (global.gc) {
-                setInterval(() => {
-                    // eslint-disable-next-line no-undef
-                    if (global.gc)
-                        global.gc();
-                }, 60000); // GC every 60 seconds (less frequent)
-            }
             // Graceful shutdown for worker
             const workerShutdown = () => {
-                this.logger.info(`Worker ${process.pid} shutting down gracefully...`, 'Worker');
+                this.logger.info(`${label} shutting down gracefully...`, 'Worker');
                 // Clean up event listeners
                 this.eventBus.removeAllListeners();
                 this.removeAllListeners();
@@ -1812,79 +1973,96 @@ export class Moro extends EventEmitter {
             // Handle worker shutdown signals
             process.on('SIGTERM', workerShutdown);
             process.on('SIGINT', workerShutdown);
-            // Continue with normal server startup for this worker
-            void this.eventBus.emit('server:starting', {
+        }
+        const fatal = (code, message) => {
+            this.logger.error(`${label} ${message}`, 'Worker');
+            if (tcw) {
+                tcw.fatal(code, message);
+                void this.close().finally(() => tcw?.close());
+            }
+            else {
+                process.exit(1);
+            }
+        };
+        // Continue with normal server startup for this worker
+        void this.eventBus.emit('server:starting', {
+            port,
+            runtime: this.runtimeType,
+            worker: workerRef,
+        });
+        // Add documentation middleware first (if enabled)
+        try {
+            const docsMiddleware = this.documentation.getDocsMiddleware();
+            this.coreFramework.addMiddleware(docsMiddleware);
+        }
+        catch {
+            // Documentation not enabled, that's fine
+        }
+        // Attach the unified router (handles both chainable and direct routes)
+        this.attachUnifiedRouter();
+        // Register legacy direct routes with the HTTP server (for backward compatibility)
+        if (this.routes.length > 0) {
+            this.registerDirectRoutes();
+        }
+        const workerCallback = () => {
+            const displayHost = host || 'localhost';
+            this.logger.info(`${label} ready on ${displayHost}:${port}`, 'Worker');
+            void this.eventBus.emit('server:started', {
                 port,
                 runtime: this.runtimeType,
-                worker: process.pid,
+                worker: workerRef,
+                engine: this.engine,
             });
-            // Add documentation middleware first (if enabled)
-            try {
-                const docsMiddleware = this.documentation.getDocsMiddleware();
-                this.coreFramework.addMiddleware(docsMiddleware);
+            // GraphQL registers HTTP/WS routes on this app: every worker.
+            this.ensureGraphQLInitialized().catch(error => {
+                this.logger.error(`Failed to initialize GraphQL: ${error}`, 'GraphQL');
+            });
+            tcw?.ready(port);
+        };
+        const startWorkerListening = () => {
+            if (host) {
+                this.coreFramework.listen(port, host, workerCallback);
             }
-            catch {
-                // Documentation not enabled, that's fine
+            else {
+                this.coreFramework.listen(port, workerCallback);
             }
-            // Attach the unified router (handles both chainable and direct routes)
-            this.attachUnifiedRouter();
-            // Register legacy direct routes with the HTTP server (for backward compatibility)
-            if (this.routes.length > 0) {
-                this.registerDirectRoutes();
+        };
+        // Same init as the single-process path: auto-discovery, file-based
+        // routes, queued WebSocket registrations - before 1.9.0 workers skipped
+        // discovery/route loading entirely.
+        Promise.all([
+            this.ensureAutoDiscoveryComplete().then(() => this.ensureRoutesLoaded()),
+            this.processQueuedWebSocketRegistrations(),
+        ])
+            .then(startWorkerListening)
+            .catch(error => {
+            const message = error instanceof Error ? error.message : String(error);
+            const code = error?.code;
+            // listen() itself failed (EADDRINUSE etc.): retrying is guaranteed to
+            // fail again and would surface as an unhandled rejection crash-loop.
+            // Report fatally so the primary sees a clean death.
+            if (code === 'EADDRINUSE' ||
+                code === 'EACCES' ||
+                /EADDRINUSE|failed to bind/i.test(message)) {
+                fatal(code === 'EACCES' ? 'EACCES' : 'EADDRINUSE', `could not bind port ${port}: ${message}`);
+                return;
             }
-            const workerCallback = () => {
-                const displayHost = host || 'localhost';
-                this.logger.info(`Worker ${process.pid} ready on ${displayHost}:${port}`, 'Worker');
-                void this.eventBus.emit('server:started', {
-                    port,
-                    runtime: this.runtimeType,
-                    worker: process.pid,
-                    engine: this.engine,
-                });
-            };
-            // Ensure WebSocket setup is complete before starting worker
-            const startWorkerListening = () => {
-                if (host) {
-                    this.coreFramework.listen(port, host, workerCallback);
-                }
-                else {
-                    this.coreFramework.listen(port, workerCallback);
-                }
-            };
-            this.processQueuedWebSocketRegistrations()
-                .then(startWorkerListening)
-                .catch(error => {
-                const message = error instanceof Error ? error.message : String(error);
-                const code = error?.code;
-                // listen() itself failed (EADDRINUSE etc.): retrying is guaranteed to
-                // fail again and would surface as an unhandled rejection crash-loop.
-                // Log fatally and exit so the cluster primary sees a clean death.
-                // Prefer the code the native/Node servers set; fall back to message.
-                if (code === 'EADDRINUSE' ||
-                    code === 'EACCES' ||
-                    /EADDRINUSE|failed to bind/i.test(message)) {
-                    this.logger.error(`Worker ${process.pid} could not bind port ${port}: ${message}`, 'Worker');
-                    process.exit(1);
-                }
-                this.logger.error('WebSocket initialization failed in worker', 'Worker', {
-                    error: message,
-                });
-                // For WebSocket failures with queued registrations, error will propagate
-                if (error instanceof Error &&
-                    error.message.includes('WebSocket features require a WebSocket adapter')) {
-                    throw error;
-                }
-                // WebSocket setup failed but the HTTP server can still serve: start
-                // anyway, and surface (not swallow) a listen failure from this path.
-                try {
-                    startWorkerListening();
-                }
-                catch (listenError) {
-                    this.logger.error(`Worker ${process.pid} failed to start: ${listenError instanceof Error ? listenError.message : String(listenError)}`, 'Worker');
-                    process.exit(1);
-                }
-            });
-        }
+            this.logger.error('Initialization failed in worker', 'Worker', { error: message });
+            // For WebSocket failures with queued registrations, error will propagate
+            if (error instanceof Error &&
+                error.message.includes('WebSocket features require a WebSocket adapter')) {
+                throw error;
+            }
+            // Init failed but the HTTP server can still serve: start anyway, and
+            // surface (not swallow) a listen failure from this path.
+            try {
+                startWorkerListening();
+            }
+            catch (listenError) {
+                const lm = listenError instanceof Error ? listenError.message : String(listenError);
+                fatal(listenError?.code === 'EACCES' ? 'EACCES' : 'EADDRINUSE', `failed to start: ${lm}`);
+            }
+        });
     }
     // Simple worker message handler
     handleWorkerMessage(message) {
@@ -1999,6 +2177,18 @@ export class Moro extends EventEmitter {
                 }
             }
         }
+        // Cluster worker threads first: each drains its server via its own
+        // close(); the primary's services below stop after the traffic did.
+        if (this.threadCluster) {
+            const primary = this.threadCluster;
+            this.threadCluster = undefined;
+            try {
+                await primary.shutdown();
+            }
+            catch (err) {
+                this.logger.error(`Error shutting down cluster worker threads: ${String(err)}`);
+            }
+        }
         // Shutdown worker threads first (fastest)
         try {
             this.logger.info('Shutting down worker threads...');