Fix ONNX shutdown crash: let event loop drain instead of process.exit()

Calling process.exit(0) triggers C++ atexit handlers including the ONNX
global thread pool destructor, which crashes on macOS with:
  libc++abi: terminating due to uncaught exception of type
  std::__1::system_error: mutex lock failed: Invalid argument

Root cause: when process.exit() is called explicitly, the V8 GC has not
run, so ONNX InferenceSession objects are still live when the global thread
pool destructor fires, causing invalid mutex state.

Fix: remove process.exit(0) from all shutdown handlers (serve, mcp workspace,
mcp single-project). After cleanup, the event loop drains naturally — V8 GC
disposes sessions before native teardown, so the thread pool destructor runs
cleanly. A ref'd 5s force timer guards against hangs.

Also revert disposeModels() and session_options workarounds (ineffective).

Author: prih

src/cli/index.ts

@@ -222,9 +222,9 @@ program
         shuttingDown = true;
         process.stderr.write('[mcp] Shutting down...\n');
         const forceTimer = setTimeout(() => { process.stderr.write('[mcp] Shutdown timeout, force exit\n'); process.exit(1); }, 5000);
-        forceTimer.unref();
         try { await manager.shutdown(); } catch { /* ignore */ }
-        process.exit(0);
+        clearTimeout(forceTimer);
+        // Let event loop drain naturally — avoids ONNX global thread pool destructor crash on macOS
       }
 
       process.on('SIGINT',  () => { void shutdown(); });
@@ -304,7 +304,6 @@ program
         process.stderr.write('[mcp] Shutdown timeout, force exit\n');
         process.exit(1);
       }, 5000);
-      forceTimer.unref();
       try {
         if (watcher) await watcher.close();
         if (indexer) await indexer.drain();
@@ -314,7 +313,8 @@ program
         saveFileIndexGraph(fileIndexGraph, project.graphMemory, embeddingFingerprint(ge.files));
         saveTaskGraph(taskGraph, project.graphMemory, embeddingFingerprint(ge.tasks));
       } catch { /* ignore */ }
-      process.exit(0);
+      clearTimeout(forceTimer);
+      // Let event loop drain naturally — avoids ONNX global thread pool destructor crash on macOS
     }
 
     process.on('SIGINT',  () => { void shutdown(); });
@@ -479,7 +479,6 @@ program
         process.stderr.write('[serve] Shutdown timeout, force exit\n');
         process.exit(1);
       }, 5000);
-      forceTimer.unref();
       try {
         httpServer.close();
         // Destroy all open connections (including WebSocket) so the server can close
@@ -490,7 +489,8 @@ program
         await configWatcher.close();
         await manager.shutdown();
       } catch { /* ignore */ }
-      process.exit(0);
+      clearTimeout(forceTimer);
+      // Let event loop drain naturally — avoids ONNX global thread pool destructor crash on macOS
     }
 
     process.on('SIGINT',  () => { void shutdown(); });

src/lib/embedder.ts

@@ -40,10 +40,6 @@ export async function loadModel(
 
   const pipeOpts: Record<string, unknown> = {};
   if (config.dtype) pipeOpts.dtype = config.dtype;
-  // Use per-session threads (intraOpNumThreads: 1) instead of the global ONNX thread pool.
-  // The global thread pool destructor crashes on macOS (libc++ mutex Invalid argument) during
-  // process exit. With intraOpNumThreads: 1 the global pool is never created.
-  pipeOpts.session_options = { intraOpNumThreads: 1, interOpNumThreads: 1 };
 
   const pipe = await pipeline('feature-extraction', config.model, pipeOpts);
   _pipeCache.set(cacheKey, pipe);
@@ -100,15 +96,6 @@ export function resetEmbedder(): void {
   _pipeCache.clear();
 }
 
-/** Dispose all loaded ONNX pipelines. Call before process.exit() to avoid macOS mutex crash. */
-export async function disposeModels(): Promise<void> {
-  for (const pipe of _pipeCache.values()) {
-    try { await pipe.dispose(); } catch { /* ignore */ }
-  }
-  _models.clear();
-  _pipeCache.clear();
-}
-
 // Vectors are L2-normalized → dot product = cosine similarity
 export function cosineSimilarity(a: number[], b: number[]): number {
   if (a.length !== b.length) return 0;

src/lib/project-manager.ts

@@ -1,5 +1,5 @@
 import { EventEmitter } from 'events';
-import { loadModel, embed, embedQuery, disposeModels } from '@/lib/embedder';
+import { loadModel, embed, embedQuery } from '@/lib/embedder';
 import { loadGraph, saveGraph, type DocGraph, DocGraphManager } from '@/graphs/docs';
 import { loadCodeGraph, saveCodeGraph, type CodeGraph, CodeGraphManager } from '@/graphs/code';
 import { loadKnowledgeGraph, saveKnowledgeGraph, KnowledgeGraphManager } from '@/graphs/knowledge';
@@ -463,7 +463,6 @@ export class ProjectManager extends EventEmitter {
     }
     this.projects.clear();
     this.workspaces.clear();
-    await disposeModels();
     process.stderr.write('[project-manager] Shutdown complete\n');
   }