umadev 1.0.9 → 1.0.10

package/bin/cli.js

@@ -182,7 +182,7 @@ function downloadTo(url, dest, withBar, label) {
             const now = Date.now();
             const pct = Math.floor((got / total) * 100);
             // Redraw on each new percent OR every ~250ms — keeps the live speed
-            // ticking even while a single percent of a 224MB file streams in.
+            // ticking even while a single percent of a large file streams in.
             if (pct !== lastPct || now - lastDraw > 250) {
               lastPct = pct;
               lastDraw = now;
@@ -213,11 +213,11 @@ function downloadTo(url, dest, withBar, label) {
     req.setTimeout(120000, () => req.destroy(new Error('timeout')));
   });
 }
-// Ordered list of base URLs to try for the release assets. An explicit override
-// (UMADEV_MODEL_BASE_URL) wins; otherwise zh-CN / China-timezone users get GitHub
-// PROXY MIRRORS first (github.com's release CDN is frequently slow or blocked in
-// mainland China), and everyone else gets github.com first — with the others as
-// fallback either way, so a blocked github.com or a down mirror still recovers.
+// Ordered list of base URLs to try for the model files. An explicit override
+// (UMADEV_MODEL_BASE_URL) wins; otherwise EVERYONE pulls the SAME upstream f32 from
+// HuggingFace (international) / hf-mirror.com (China), region-ordered, with the
+// GitHub fp16 release as a last-resort fallback — so the download is consistent in
+// size everywhere and the model is always reachable.
 function releaseBases(version) {
   if (process.env.UMADEV_MODEL_BASE_URL) {
     return [process.env.UMADEV_MODEL_BASE_URL.replace(/\/+$/, '')];
@@ -242,9 +242,13 @@ function releaseBases(version) {
   } catch (_) {
     /* default to international order */
   }
-  // China: hf-mirror first (fast + reliable in CN), then GitHub proxies + direct.
-  // International: GitHub Release first (smaller fp16), then HuggingFace + mirror.
-  return cn ? [hfMirror, ...ghProxies, gh, hf] : [gh, hf, hfMirror, ...ghProxies];
+  // Unified on the upstream f32 model from HuggingFace (international) + its China
+  // mirror hf-mirror.com, so BOTH regions download the SAME ~448MB f32 — consistent
+  // everywhere (no more "some get 200MB, some 400MB"). China: hf-mirror first (fast
+  // in CN); international: huggingface.co first. The GitHub Release fp16 (~224MB) +
+  // proxies stay only as a LAST-RESORT fallback if both HF endpoints are down (the
+  // candle loader casts either precision to f32, so a fallback still loads).
+  return cn ? [hfMirror, hf, gh, ...ghProxies] : [hf, hfMirror, gh, ...ghProxies];
 }
 // Try each base for `name` in order; resolve on first success, throw the last
 // error if all fail. A China mirror can cover a blocked github.com (or vice
@@ -300,6 +304,22 @@ async function ensureModel() {
   }
 }
 
+// Subcommands that do NOT drive the agent runtime — they never retrieve
+// knowledge, so they must not trigger the one-time vector-model download.
+// Without this, `umadev update` (and `--version` / `--help` / `doctor` / …) would
+// appear to hang on a machine that doesn't have the model yet, while it streams
+// in — which reads as "the update command broke". The model is fetched lazily on
+// the first command that actually needs it (the TUI / run / quick / …).
+const NO_MODEL_COMMANDS = new Set([
+  'update', 'install', 'uninstall', 'init',
+  '--version', '-V', 'version',
+  '--help', '-h', 'help',
+  'doctor', 'mcp', 'hook', 'ci',
+  'usage', 'lessons', 'history',
+  'examples', 'guide',
+  'mcp-manage', 'skill', 'knowledge-manage', 'pr',
+]);
+
 async function main() {
   const binary = findBinary();
   // npm artifact round-trips (upload/download-artifact in CI) can strip the
@@ -310,10 +330,15 @@ async function main() {
     // read-only install dir or already +x — spawnSync below reports real errors
   }
   const extraEnv = {};
+  // A utility command (update/--version/--help/doctor/…) skips the model fetch so
+  // it returns instantly even before the model is installed; the agent runtime
+  // commands still fetch it lazily on first use.
+  const firstArg = process.argv[2] || '';
+  const needsModel = !NO_MODEL_COMMANDS.has(firstArg);
   // Prefer a bundled npm model package (dev / sibling layout); otherwise fetch
   // it on demand into ~/.umadev/embed-model (the binary's model_dir() fallback).
   let modelDir = findModelDir();
-  if (!modelDir) modelDir = await ensureModel();
+  if (needsModel && !modelDir) modelDir = await ensureModel();
   if (modelDir && !process.env.UMADEV_EMBED_MODEL_DIR) {
     extraEnv.UMADEV_EMBED_MODEL_DIR = modelDir;
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "umadev",
-  "version": "1.0.9",
+  "version": "1.0.10",
   "description": "A project-director Agent for AI coding hosts — drives your logged-in Claude Code / Codex through a 9-phase commercial delivery pipeline with governance. No API key needed.",
   "keywords": [
     "ai",
@@ -37,11 +37,11 @@
     "node": ">=18"
   },
   "optionalDependencies": {
-    "@umacloud/cli-darwin-arm64": "1.0.9",
-    "@umacloud/cli-darwin-x64": "1.0.9",
-    "@umacloud/cli-linux-x64": "1.0.9",
-    "@umacloud/cli-linux-arm64": "1.0.9",
-    "@umacloud/cli-win32-x64": "1.0.9",
-    "@umacloud/knowledge": "1.0.9"
+    "@umacloud/cli-darwin-arm64": "1.0.10",
+    "@umacloud/cli-darwin-x64": "1.0.10",
+    "@umacloud/cli-linux-x64": "1.0.10",
+    "@umacloud/cli-linux-arm64": "1.0.10",
+    "@umacloud/cli-win32-x64": "1.0.10",
+    "@umacloud/knowledge": "1.0.10"
   }
 }