specmem-hardwicksoftware 3.7.39 → 3.7.41

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "specmem-hardwicksoftware",
-  "version": "3.7.39",
+  "version": "3.7.41",
   "type": "module",
   "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",

package/scripts/specmem-init.cjs

@@ -9140,15 +9140,15 @@ function registerProjectInRegistry(projectPath, globalDir) {
 /**
  * Stage 1: MODEL DOWNLOAD
  * Checks for all required ML models. If any are missing, auto-downloads
- * the combined models tarball from GitHub release — no prompting.
+ * the combined models tarball from a permanent GitHub release (tag: "models").
+ * Version-independent — all npm versions pull from the same release.
  *
- * Models bundled in npm (always present after install):
- *   - all-MiniLM-L6-v2 (embedding)
- *   - minisbd (sentence boundary)
- *
- * Models NOT in npm (LFS-only, downloaded via release tarball):
- *   - pythia-410m-onnx-quant (mini-COT)
- *   - argos-translate (translation en↔zh/zt)
+ * All models (~662MB uncompressed):
+ *   - all-MiniLM-L6-v2 (23MB) — embedding model (ONNX)
+ *   - pythia-410m-onnx-quant (394MB) — mini-COT (ONNX int8)
+ *   - argos-translate (246MB) — translation (en↔zh/zt)
+ *   - minisbd (192KB) — sentence boundary detection
+ *   - argos-translate-packages (4KB) — package registry
  */
 async function ensureModels(ui) {
   const specmemRoot = path.resolve(__dirname, '..');
@@ -9156,8 +9156,10 @@ async function ensureModels(ui) {
   const version = SPECMEM_VERSION;
 
   const sentinels = [
+    { name: 'Embedding (MiniLM)', path: path.join(modelsDir, 'all-MiniLM-L6-v2', 'onnx', 'model_quint8_avx2.onnx') },
     { name: 'Mini-COT (Pythia)', path: path.join(modelsDir, 'pythia-410m-onnx-quant', 'model_quantized.onnx') },
     { name: 'Translation (Argos)', path: path.join(modelsDir, 'argos-translate', 'translate-en_zh-1_9', 'model', 'model.bin') },
+    { name: 'SBD (minisbd)', path: path.join(modelsDir, 'minisbd', 'en.onnx') },
   ];
 
   const missing = sentinels.filter(s => !fs.existsSync(s.path));
@@ -9188,16 +9190,49 @@ async function ensureModels(ui) {
     return;
   }
 
-  const releaseUrl = `https://github.com/jonhardwick-spec/specmem/releases/download/v${version}/specmem-models-${version}.tar.gz`;
-  const tmpTarball = path.join(os.tmpdir(), `specmem-models-${version}.tar.gz`);
+  // Permanent models release — version-independent, all npm versions pull from here
+  const releaseUrl = 'https://github.com/jonhardwick-spec/specmem/releases/download/models/specmem-models.tar.gz';
+  const tmpTarball = path.join(os.tmpdir(), 'specmem-models.tar.gz');
+  const expectedSizeMB = 562; // approximate tarball size
 
   initLog(`[MODELS] Downloading: ${releaseUrl}`);
 
+  // Use spawn (async) instead of execSync so blessed TUI can render during download
   try {
-    const dlCmd = hasCurl
-      ? `curl -fSL --progress-bar -o "${tmpTarball}" "${releaseUrl}"`
-      : `wget --progress=bar:force -O "${tmpTarball}" "${releaseUrl}"`;
-    execSync(dlCmd, { stdio: 'pipe', timeout: 900000 }); // 15 min timeout for ~570MB
+    await new Promise((resolve, reject) => {
+      const args = hasCurl
+        ? ['curl', ['-fSL', '-o', tmpTarball, releaseUrl]]
+        : ['wget', ['-O', tmpTarball, releaseUrl]];
+      const dl = spawn(args[0], args[1], { stdio: 'ignore' });
+
+      // Poll file size for progress updates
+      const progressInterval = setInterval(() => {
+        try {
+          const stat = fs.statSync(tmpTarball);
+          const mb = (stat.size / (1024 * 1024)).toFixed(0);
+          const pct = Math.min(99, Math.round((stat.size / (expectedSizeMB * 1024 * 1024)) * 100));
+          if (ui) ui.setSubStatus(`↓ Downloading ML models... ${mb}MB / ~${expectedSizeMB}MB (${pct}%)`);
+        } catch {}
+      }, 2000);
+
+      const timeout = setTimeout(() => {
+        clearInterval(progressInterval);
+        dl.kill('SIGTERM');
+        reject(new Error('Download timed out after 15 minutes'));
+      }, 900000);
+
+      dl.on('close', (code) => {
+        clearInterval(progressInterval);
+        clearTimeout(timeout);
+        if (code === 0) resolve();
+        else reject(new Error(`Download exited with code ${code}`));
+      });
+      dl.on('error', (err) => {
+        clearInterval(progressInterval);
+        clearTimeout(timeout);
+        reject(err);
+      });
+    });
   } catch (e) {
     initLog(`[MODELS] Download failed: ${e.message}`);
     if (ui) ui.setSubStatus(`⚠ Model download failed — mini-COT/translation disabled`);
@@ -9210,14 +9245,15 @@ async function ensureModels(ui) {
 
   try {
     const mkdirCmd = `mkdir -p "${modelsDir}"`;
-    const extractCmd = `tar xzf "${tmpTarball}" -C "${modelsDir}"`;
+    // Tarball has models/ prefix — strip it so contents land directly in modelsDir
+    const extractCmd = `tar xzf "${tmpTarball}" -C "${modelsDir}" --strip-components=1`;
     if (needsSudo) {
       execSync(`sudo ${mkdirCmd}`, { stdio: 'pipe' });
-      execSync(`sudo ${extractCmd}`, { stdio: 'pipe', timeout: 180000 });
+      execSync(`sudo ${extractCmd}`, { stdio: 'pipe', timeout: 300000 });
       execSync(`sudo chmod -R a+rX "${modelsDir}"`, { stdio: 'pipe' });
     } else {
       execSync(mkdirCmd, { stdio: 'pipe' });
-      execSync(extractCmd, { stdio: 'pipe', timeout: 180000 });
+      execSync(extractCmd, { stdio: 'pipe', timeout: 300000 });
     }
   } catch (e) {
     initLog(`[MODELS] Extraction failed: ${e.message}`);