@tiens.nguyen/gonext-local-worker 1.0.114 → 1.0.117

package/gonext-local-worker.mjs

@@ -1669,6 +1669,40 @@ async function checkOllamaTags(rawBase) {
   }
 }
 
+/**
+ * Discover the mlx_lm.server processes actually running on this machine by
+ * parsing `ps` output. Returns deduped `[{ port, modelPath }]` so the model
+ * dropdown reflects live servers instead of a hand-maintained port list.
+ * mlx_lm.server defaults to port 8080 when `--port` is omitted. macOS only.
+ */
+async function discoverMlxServers() {
+  if (platform() !== "darwin") return [];
+  let stdout = "";
+  try {
+    ({ stdout } = await execFile("ps", ["-Ao", "args="], {
+      maxBuffer: 8 * 1024 * 1024,
+    }));
+  } catch {
+    return [];
+  }
+  const byPort = new Map();
+  for (const raw of stdout.split("\n")) {
+    const line = raw.trim();
+    if (!line) continue;
+    // Match `mlx_lm.server` or `mlx_lm server` (python -m mlx_lm server).
+    if (!/mlx_lm[.\s]server\b/.test(line)) continue;
+    // Ignore our own grep/ps invocations that happen to mention the string.
+    if (/\bgrep\b/.test(line)) continue;
+    const portMatch = line.match(/--port(?:=|\s+)(\d{2,5})/);
+    const port = portMatch ? parseInt(portMatch[1], 10) : 8080;
+    if (!(port > 0)) continue;
+    const modelMatch = line.match(/--model(?:=|\s+)(\S+)/);
+    const modelPath = modelMatch ? modelMatch[1] : "";
+    if (!byPort.has(port)) byPort.set(port, { port, modelPath });
+  }
+  return [...byPort.values()];
+}
+
 async function checkOpenAiModels(base, apiKey, source) {
   const endpoint = `${base}/models`;
   const headers = {};
@@ -1842,13 +1876,6 @@ async function runLocalHealthJob(job) {
         if (!dedup.has(m.value)) dedup.set(m.value, m);
       }
     }
-    // Support both single mlxOpenAiBaseUrl and multi-port mlxOpenAiBaseUrls
-    const mlxRootSingle = normalizeOpenAiV1Root(payload?.mlxOpenAiBaseUrl);
-    const mlxRootsMulti = Array.isArray(payload?.mlxOpenAiBaseUrls)
-      ? payload.mlxOpenAiBaseUrls.map(normalizeOpenAiV1Root).filter(Boolean)
-      : [];
-    const mlxRoots = mlxRootsMulti.length > 0 ? mlxRootsMulti : (mlxRootSingle ? [mlxRootSingle] : []);
-    const mlxRoot = mlxRoots[0] || null;
     const targetWorkerHostId =
       typeof payload?.targetWorkerHostId === "string"
         ? payload.targetWorkerHostId.trim()
@@ -1857,6 +1884,30 @@ async function runLocalHealthJob(job) {
       typeof payload?.targetWorkerHostName === "string"
         ? payload.targetWorkerHostName.trim()
         : "";
+    // MLX servers are discovered dynamically from running mlx_lm.server
+    // processes (macOS), so the dropdown reflects reality with no saved port
+    // list. Only discover when a worker host is targeted so the encoded models
+    // are host-scoped (routable). Each server is probed at 127.0.0.1:<port>/v1.
+    const discoveredMlx = targetWorkerHostId ? await discoverMlxServers() : [];
+    const discoveredByPort = new Map(discoveredMlx.map((s) => [s.port, s]));
+    const mlxRootsDiscovered = discoveredMlx.map(
+      (s) => `http://127.0.0.1:${s.port}/v1`
+    );
+    // Fall back to any explicitly provided base (e.g. a remote host) so
+    // non-local MLX endpoints still work when no local process is found.
+    const mlxRootSingle = normalizeOpenAiV1Root(payload?.mlxOpenAiBaseUrl);
+    const mlxRootsMulti = Array.isArray(payload?.mlxOpenAiBaseUrls)
+      ? payload.mlxOpenAiBaseUrls.map(normalizeOpenAiV1Root).filter(Boolean)
+      : [];
+    const mlxRoots =
+      mlxRootsDiscovered.length > 0
+        ? mlxRootsDiscovered
+        : mlxRootsMulti.length > 0
+          ? mlxRootsMulti
+          : mlxRootSingle
+            ? [mlxRootSingle]
+            : [];
+    const mlxRoot = mlxRoots[0] || null;
     let mlxNative = null;
     const modelPathChecksRaw = Array.isArray(payload?.modelPathChecks)
       ? payload.modelPathChecks
@@ -1892,12 +1943,29 @@ async function runLocalHealthJob(job) {
     // Merge all per-port results
     const mlxModelDedup = new Map();
     let mlxHttpOnline = false;
+    const portsWithModels = new Set();
     for (const r of mlxHttpResults) {
       if (r.online) mlxHttpOnline = true;
       for (const m of r.models) {
+        const pm = String(m.value ?? "").match(/@@port:(\d+)/);
+        if (pm) portsWithModels.add(parseInt(pm[1], 10));
         if (!mlxModelDedup.has(m.value)) mlxModelDedup.set(m.value, m);
       }
     }
+    // A discovered server may be booting (process up, /v1/models not ready).
+    // Synthesize an entry from its `--model` path so it is still visible.
+    for (const { port, modelPath } of discoveredByPort.values()) {
+      if (portsWithModels.has(port) || !modelPath) continue;
+      const hostSuffix = targetWorkerHostId ? `@@host:${targetWorkerHostId}` : "";
+      const value = `mlx:${modelPath}${hostSuffix}@@port:${port}`;
+      if (mlxModelDedup.has(value)) continue;
+      const label = targetWorkerHostName || sourceLabelFromBase(`http://127.0.0.1:${port}`);
+      mlxModelDedup.set(value, {
+        id: targetWorkerHostId ? `${modelPath}@@${targetWorkerHostId}` : modelPath,
+        name: `${modelPath} (${label}:${port}) — starting…`,
+        value,
+      });
+    }
     const mlxHttp = mlxRoots.length > 0 ? {
       online: mlxHttpOnline,
       endpoint: mlxHttpResults[0]?.endpoint,
@@ -1905,7 +1973,7 @@ async function runLocalHealthJob(job) {
     } : null;
 
     const wantNativeFallback =
-      mlxRoot &&
+      (targetWorkerHostId || mlxRoot) &&
       payload?.mlxNativeFallback !== false &&
       platform() === "darwin" &&
       (!mlxHttp?.online || (mlxHttp?.models?.length ?? 0) === 0);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.114",
+  "version": "1.0.117",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",