specmem-hardwicksoftware 3.7.19 → 3.7.20

package/dist/codebase/codebaseIndexer.js

@@ -900,7 +900,7 @@ export class CodebaseIndexer {
         const analyzableLanguages = [
             'typescript', 'typescript-react', 'javascript', 'javascript-react',
             'python', 'go', 'rust', 'java', 'kotlin', 'scala',
-            'ruby', 'php', 'c', 'cpp', 'swift'
+            'ruby', 'php', 'c', 'cpp', 'swift', 'html'
         ];
         return analyzableLanguages.includes(language);
     }

package/dist/tools/goofy/findWhatISaid.js

@@ -821,7 +821,7 @@ export class FindWhatISaid {
                 });
             }
             catch (embeddingError) {
-                clearTimeout(embeddingTimeoutId); // Prevent dangling timer on error path
+                // embeddingTimeoutId is scoped inside withEmbeddingRetry — already cleared there
                 const embeddingDuration = Date.now() - embeddingStartTime;
                 const err = embeddingError;
                 // ============================================================================

package/embedding-sandbox/frankenstein-embeddings.py

@@ -293,32 +293,48 @@ def _detect_best_onnx_file():
     """
     Detect CPU features and return the best ONNX model file name.
     Priority: avx512_vnni > avx512 > avx2 > default
+    Falls back to whatever .onnx file exists if the optimal one isn't found.
     """
+    # Ordered by preference (best first)
+    candidates = []
+
     try:
         with open('/proc/cpuinfo', 'r') as f:
             cpuinfo = f.read().lower()
 
-        # Check for AVX512 VNNI (best for INT8)
         if 'avx512_vnni' in cpuinfo or 'avx512vnni' in cpuinfo:
-            print("🚀 CPU supports AVX512-VNNI - using optimized INT8 model", file=sys.stderr)
-            return "onnx/model_qint8_avx512_vnni.onnx"
-
-        # Check for AVX512 (good INT8 support)
+            candidates.append(("onnx/model_qint8_avx512_vnni.onnx", "AVX512-VNNI"))
         if 'avx512f' in cpuinfo or 'avx512' in cpuinfo:
-            print("🚀 CPU supports AVX512 - using INT8 quantized model", file=sys.stderr)
-            return "onnx/model_qint8_avx512.onnx"
-
-        # Check for AVX2 (common, decent performance)
+            candidates.append(("onnx/model_qint8_avx512.onnx", "AVX512"))
         if 'avx2' in cpuinfo:
-            print("🚀 CPU supports AVX2 - using UINT8 quantized model", file=sys.stderr)
-            return "onnx/model_quint8_avx2.onnx"
-
-        # Fallback to unoptimized
-        print("ℹ️ Using default ONNX model (no AVX optimization)", file=sys.stderr)
-        return "onnx/model.onnx"
+            candidates.append(("onnx/model_quint8_avx2.onnx", "AVX2"))
     except Exception as e:
         print(f"⚠️ Could not detect CPU features: {e}", file=sys.stderr)
-        return "onnx/model.onnx"
+
+    # Always add standard fallbacks
+    candidates.append(("onnx/model_quantized.onnx", "quantized"))
+    candidates.append(("onnx/model.onnx", "default"))
+
+    # Check which files actually exist in the bundled model dir
+    bundled_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'models', 'all-MiniLM-L6-v2')
+    for onnx_file, label in candidates:
+        full_path = os.path.join(bundled_dir, onnx_file)
+        if os.path.isfile(full_path):
+            print(f"🚀 Using {label} ONNX model: {onnx_file}", file=sys.stderr)
+            return onnx_file
+
+    # Last resort: find ANY .onnx file in the bundled dir
+    onnx_dir = os.path.join(bundled_dir, 'onnx')
+    if os.path.isdir(onnx_dir):
+        for f in os.listdir(onnx_dir):
+            if f.endswith('.onnx'):
+                result = f"onnx/{f}"
+                print(f"🔍 Auto-detected ONNX model: {result}", file=sys.stderr)
+                return result
+
+    # Nothing found - return default and let SentenceTransformer handle it
+    print("ℹ️ No bundled ONNX model found - using default", file=sys.stderr)
+    return "onnx/model.onnx"
 
 _BEST_ONNX_FILE = _detect_best_onnx_file()
 

package/embedding-sandbox/server.mjs

@@ -57,6 +57,13 @@ const getMachineSocketPath = () => {
 const SOCKET_PATH = process.env.SOCKET_PATH || getMachineSocketPath();
 const MODEL_NAME = 'Xenova/all-MiniLM-L6-v2';
 
+// Bundled model: shipped with npm package, used as fallback when HF cache unavailable
+import { fileURLToPath } from 'url';
+import { dirname } from 'path';
+const __filename_esm = fileURLToPath(import.meta.url);
+const __dirname_esm = dirname(__filename_esm);
+const BUNDLED_MODEL_DIR = join(__dirname_esm, 'models', 'all-MiniLM-L6-v2');
+
 // Dynamic dimensions - detected from model and database
 let NATIVE_DIM = null;
 let TARGET_DIM = null;
@@ -78,13 +85,39 @@ async function loadModel() {
   try {
     console.log('[Sandbox] Loading model from local cache...');
 
-    // Force local-only mode - will fail if model not pre-downloaded
-    extractor = await pipeline('feature-extraction', MODEL_NAME, {
-      // Use local cache only - no downloads allowed
-      local_files_only: true,
-      // Use CPU only (safer, no GPU driver access)
-      device: 'cpu'
-    });
+    // Try HF cache first, fall back to bundled model
+    let modelSource = MODEL_NAME;
+    try {
+      extractor = await pipeline('feature-extraction', MODEL_NAME, {
+        local_files_only: true,
+        device: 'cpu'
+      });
+    } catch (hfErr) {
+      // HF cache miss — try bundled model shipped with npm package
+      if (existsSync(BUNDLED_MODEL_DIR)) {
+        console.log(`[Sandbox] HF cache miss, loading bundled model: ${BUNDLED_MODEL_DIR}`);
+        // Ensure model.onnx exists (bundled may only have model_quint8_avx2.onnx)
+        const onnxDir = join(BUNDLED_MODEL_DIR, 'onnx');
+        const modelOnnx = join(onnxDir, 'model.onnx');
+        if (!existsSync(modelOnnx) && existsSync(onnxDir)) {
+          // Find any .onnx file and symlink as model.onnx
+          const { readdirSync, symlinkSync } = await import('fs');
+          const onnxFiles = readdirSync(onnxDir).filter(f => f.endsWith('.onnx'));
+          if (onnxFiles.length > 0) {
+            try { symlinkSync(onnxFiles[0], modelOnnx); } catch {}
+          }
+        }
+        extractor = await pipeline('feature-extraction', BUNDLED_MODEL_DIR, {
+          local_files_only: true,
+          device: 'cpu'
+        });
+        modelSource = BUNDLED_MODEL_DIR;
+      } else {
+        throw hfErr;
+      }
+    }
+
+    // Skip the duplicate pipeline call below — extractor is already loaded
 
     modelReady = true;
 

package/mcp-proxy.cjs

@@ -261,7 +261,15 @@ function spawnServer() {
     stopHeartbeat();
 
     if (!shuttingDown) {
-      scheduleRestart();
+      // If bootstrap was intentionally killed (SIGTERM/SIGKILL from init or system),
+      // don't restart — exit the proxy too. Init will start a fresh bootstrap.
+      // Only restart on crashes (non-zero exit code without signal).
+      if (signal === 'SIGTERM' || signal === 'SIGKILL') {
+        log(`Bootstrap was intentionally killed (${signal}) — proxy exiting`);
+        shutdown();
+      } else {
+        scheduleRestart();
+      }
     }
   });
 
@@ -386,6 +394,15 @@ function shutdown() {
 process.on('SIGTERM', shutdown);
 process.on('SIGINT', shutdown);
 
+// Orphan detection: if parent (Claude) dies, proxy gets reparented to PID 1
+// Check every 10s and exit if orphaned
+setInterval(() => {
+  if (process.ppid === 1) {
+    log('Parent died (PPID=1), proxy shutting down');
+    shutdown();
+  }
+}, 10000);
+
 // ============================================================================
 // Start
 // ============================================================================

package/package.json

@@ -1,8 +1,8 @@
 {
   "name": "specmem-hardwicksoftware",
-  "version": "3.7.19",
+  "version": "3.7.20",
   "type": "module",
-  "description": "Persistent memory system for coding sessions - semantic search with pgvector, token compression, team coordination, file watching. Needs root: installs system-wide hooks, manages docker/PostgreSQL, writes global configs, handles screen sessions. justcalljon.pro",
+  "description": "Your Claude Code sessions don't have to start from scratch anymore — SpecMem gives your AI real memory. It won't forget your conversations, your code, or your architecture decisions between sessions. That's the whole point. Semantic code indexing that actually works: TypeScript, JavaScript, Python, Go, Rust, Java, Kotlin, C, C++, HTML and more. It doesn't just track functions — it gets classes, methods, fields, constants, enums, macros, imports, structs, the whole codebase graph. There's chat memory too, powered by pgvector embeddings. You've also got token compression, team coordination, multi-agent comms, and file watching built in. 74+ MCP tools. Runs on PostgreSQL + Docker. It's kind of a big deal. justcalljon.pro",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
   "bin": {
@@ -44,7 +44,18 @@
     "anthropic",
     "embeddings",
     "context",
-    "hooks"
+    "hooks",
+    "code-indexing",
+    "typescript",
+    "javascript",
+    "python",
+    "java",
+    "cpp",
+    "rust",
+    "html",
+    "codebase-analysis",
+    "multi-language",
+    "team-coordination"
   ],
   "author": {
     "name": "Jonathan Hardwick",