@tiens.nguyen/gu-cli 1.0.772 → 1.0.774

package/gu-repl.mjs

@@ -11622,10 +11622,18 @@ async function main() {
             }
           }
         }
-        const chatUrl = String(p.agentBaseURL || "").trim() || `http://127.0.0.1:${md.DEFAULT_CHAT_PORT}`;
-        // A CLIENT hosts no models, so "the chat model server isn't responding" must not lead
-        // to a local model setup here — its model is remote by design.
-        if (!clientModeBanner && !(await md.probeModelServer(chatUrl))) {
+        const chatUrl = String(p.agentBaseURL || "").trim();
+        // ONLY an Apple-Silicon Mac in Host mode whose agent model is its own local MLX server
+        // is offered a local model setup (task #311): a CLIENT or a CLIENT EXPERT keeps its models
+        // remote, Linux/Windows/Intel cannot run MLX, and no configured agent URL is not a reason
+        // to probe 127.0.0.1:8082 — the bar already says "agent model is not set".
+        const offerLocal = md.shouldOfferLocalModelSetup({
+          canHostMlx: md.canHostMlxModels(),
+          clientMode: Boolean(clientModeBanner),
+          clientExpert: isClientExpert(),
+          agentUrl: chatUrl,
+        });
+        if (offerLocal && !(await md.probeModelServer(chatUrl))) {
           console.log(
             dim(`\ngu: the chat model server (${md.modelRoot(chatUrl)}) isn't responding.`)
           );

package/model-doctor.mjs

@@ -1172,6 +1172,30 @@ export function canHostMlxModels() {
   return platform() === "darwin" && arch() === "arm64";
 }
 
+/**
+ * Should gu offer to download and start the LOCAL chat model when it is not answering? (task #311)
+ *
+ * That offer is for one machine: an Apple-Silicon Mac running its own MLX chat model on 127.0.0.1.
+ * It used to be asked whenever the hosted-Client flag was off — so a Client Expert box (models
+ * REMOTE by definition) on Ubuntu was offered an MLX download, and with no agent model configured
+ * the check fell back to 127.0.0.1:8082, which can never answer there.
+ *   · canHostMlx — canHostMlxModels() (darwin + arm64)
+ *   · neither hosted Client nor Client Expert
+ *   · an agent model URL IS configured and points at THIS machine
+ * Pure: every input is passed in.
+ */
+export function shouldOfferLocalModelSetup({ canHostMlx, clientMode, clientExpert, agentUrl }) {
+  if (!canHostMlx || clientMode || clientExpert) return false;
+  const url = String(agentUrl ?? "").trim();
+  if (!url) return false;
+  try {
+    const h = new URL(url).hostname.replace(/^\[|\]$/g, "");
+    return h === "localhost" || h === "::1" || h.startsWith("127.");
+  } catch {
+    return false;
+  }
+}
+
 /**
  * True when a local OLLAMA is serving models here. Matters because canHostMlxModels() is
  * false on an Intel Mac or a Linux box, yet such a machine may well host models via Ollama —

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.772",
+  "version": "1.0.774",
   "description": "Gu CLI — the gu terminal plus the local worker that runs agent / OCR / PDF / embedding jobs on your Mac (Ollama / MLX / OpenAI-compatible).",
   "type": "module",
   "license": "MIT",