offgrid-ai 0.3.28 → 0.3.30

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.3.28",
+  "version": "0.3.30",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/cli.mjs

@@ -21,7 +21,7 @@ import { offerManagedLlamaRuntimeUpdate } from "./runtime.mjs";
 
 async function offerUpdate(argv) {
   const invocation = detectInvocation();
-  const update = await checkForUpdate({ force: invocation === "npx" });
+  const update = await checkForUpdate({ force: true });
   if (!update) return false;
 
   const plan = updateCommand(invocation, argv);
@@ -1132,7 +1132,7 @@ async function printVersion() {
   const version = currentPackageVersion();
   console.log(`offgrid-ai v${version}`);
   const invocation = detectInvocation();
-  const update = await checkForUpdate({ force: invocation === "npx" });
+  const update = await checkForUpdate({ force: true });
   if (update) {
     const plan = updateCommand(invocation, ["version"]);
     console.log(pc.yellow(`Update available: v${update.latest}. Run: ${plan.display}`));

package/src/profile-setup.mjs

@@ -1,6 +1,11 @@
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
 import { estimateMemory } from "./estimate.mjs";
+import { findLlamaServer } from "./config.mjs";
 import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 
+const execFileAsync = promisify(execFile);
+
 const CACHE_CHOICES = [
   { value: "bf16", label: "bf16", hint: "default: stable, good quality" },
   { value: "f16", label: "f16", hint: "stable fallback, similar memory to bf16" },
@@ -57,13 +62,14 @@ export async function configureLocalProfile(prompt, profile) {
   if (caps.vision && profile.mmprojPath) {
     console.log("");
     const gemma4Unified = isGemma4UnifiedProjector(caps.mmprojProjectorType);
+    const supported = !gemma4Unified || await runtimeSupportsGemma4Unified();
     console.log(renderSection("Vision projector detected", renderRows([
       ["Projector", caps.mmprojProjectorType ?? "unknown"],
       ["Flag", `--mmproj ${profile.mmprojPath}`],
-      ...(gemma4Unified ? [["Note", pc.yellow("Gemma 4 unified projectors need newer llama.cpp than current Homebrew stable.")]] : []),
+      ...(gemma4Unified && !supported ? [["Note", pc.yellow("Gemma 4 unified projectors need llama.cpp b9549+.")]] : []),
     ])));
-    const useVision = await prompt.yesNo("Enable vision with --mmproj?", !gemma4Unified);
-    configured = useVision ? applyVisionDefaults(configured) : removeVisionDefaults(configured, gemma4Unified ? "gemma4-unified-unsupported" : "user-disabled");
+    const useVision = await prompt.yesNo("Enable vision with --mmproj?", supported);
+    configured = useVision ? applyVisionDefaults(configured) : removeVisionDefaults(configured, gemma4Unified && !supported ? "gemma4-unified-unsupported" : "user-disabled");
   }
 
   if (caps.thinking) {
@@ -212,6 +218,19 @@ function isGemma4UnifiedProjector(projectorType) {
   return /gemma4u[va]/i.test(String(projectorType ?? ""));
 }
 
+async function runtimeSupportsGemma4Unified() {
+  try {
+    const binary = await findLlamaServer();
+    if (!binary) return false;
+    const { stdout, stderr } = await execFileAsync(binary, ["--version"]);
+    const output = `${stdout}\n${stderr}`;
+    const version = Number(output.match(/version:\s*(\d+)/i)?.[1]);
+    return Number.isFinite(version) && version >= 9549;
+  } catch {
+    return false;
+  }
+}
+
 function detectionSummary(caps) {
   const parts = [];
   if (caps.architecture) parts.push(caps.architecture);