offgrid-ai 0.3.20 → 0.3.22

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.3.20",
+  "version": "0.3.22",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/backends.mjs

@@ -72,16 +72,18 @@ async function scanOllamaModels() {
     if (!response.ok) return [];
     const body = await response.json();
     if (!Array.isArray(body?.models)) return [];
-    return body.models.map((model) => ({
-      id: model.name,
-      label: ollamaLabel(model.name),
-      aliasSuggestion: model.name,
-      sizeBytes: model.size ?? 0,
-      quant: model.details?.quantization_level,
-      family: model.details?.family,
-      backend: "ollama",
-      source: "ollama",
-    })).sort((a, b) => a.label.localeCompare(b.label));
+    return body.models
+      .filter((model) => isLocalOllamaModel(model))
+      .map((model) => ({
+        id: model.name,
+        label: ollamaLabel(model.name),
+        aliasSuggestion: model.name,
+        sizeBytes: model.size ?? 0,
+        quant: model.details?.quantization_level,
+        family: model.details?.family,
+        backend: "ollama",
+        source: "ollama",
+      })).sort((a, b) => a.label.localeCompare(b.label));
   } catch {
     return [];
   }
@@ -112,6 +114,13 @@ async function scanOmlxModels() {
 
 // ── Labels ──────────────────────────────────────────────────────────────
 
+function isLocalOllamaModel(model) {
+  const name = String(model?.name ?? "");
+  if (/:cloud(?:$|\b)/i.test(name)) return false;
+  if (!Number.isFinite(model?.size) || model.size <= 0) return false;
+  return true;
+}
+
 function ollamaLabel(name) {
   return name.replace(/[-_]/g, " ").replace(/^gemma\b/i, "Gemma").replace(/^qwen/i, "Qwen");
 }

package/src/cli.mjs

@@ -280,7 +280,7 @@ async function printModelCatalog({ profiles, newModels, managedItems }, items =
     const backendItems = managedItems.filter((item) => item.backendId === backendId);
     if (backendItems.length === 0) continue;
     const be = BACKENDS[backendId];
-    console.log("\n" + pc.bold(`Available from ${be.label}`));
+    console.log("\n" + pc.bold(`Local models via ${be.label}`));
     for (const { model } of backendItems.slice(0, 10)) {
       const num = itemNumber((item) => item.type === "managed" && item.backendId === backendId && item.model.id === model.id);
       console.log(managedModelCard(num, model, be));
@@ -307,13 +307,13 @@ function downloadedModelCard(num, model, caps) {
     ["Good for", humanCapabilitySummary(caps)],
     ["Size", formatBytes(model.sizeBytes)],
     ["When selected", "offgrid-ai will recommend safe local settings"],
-  ]), { formatBorder: caps.mtp ? pc.blue : pc.yellow });
+  ]), { formatBorder: pc.yellow });
 }
 
 function managedModelCard(num, model, backend) {
   return renderCard(`${num}. ${model.label}`, renderRows([
-    ["Status", statusText("info", "Available from another app")],
-    ["App", backend.label],
+    ["Status", statusText("info", `Local model via ${backend.label}`)],
+    ["Runs with", backend.label],
     ["Model ID", pc.cyan(model.id)],
     ...(model.quant ? [["Size/type", model.quant]] : []),
   ]), { formatBorder: pc.magenta });
@@ -438,7 +438,7 @@ function printGgufModelDetails(model) {
 function printManagedModelDetails(model, backend) {
   console.log("\n" + renderSection(`${backend.label} model`, renderRows([
     ["Name", pc.bold(model.label)],
-    ["Status", pc.green("Available from another app")],
+    ["Status", pc.green(`Local model via ${backend.label}`)],
     ["Model ID", pc.cyan(model.id)],
     ["Quant", model.quant ?? "unknown"],
     ["Family", model.family ?? "unknown"],

package/src/profile-setup.mjs

@@ -1,11 +1,11 @@
 import { estimateMemory } from "./estimate.mjs";
-import { pc, formatBytes, renderRows, renderSection, humanCapabilitySummary } from "./ui.mjs";
+import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 
 const CACHE_CHOICES = [
-  { value: "bf16", label: "Balanced", hint: "recommended: stable, good quality" },
-  { value: "f16", label: "Compatible", hint: "stable fallback, similar memory use" },
-  { value: "q8_0", label: "Lower memory", hint: "usually safe, uses less memory" },
-  { value: "q4_0", label: "Smallest memory", hint: "maximum savings, quality/speed tradeoff" },
+  { value: "bf16", label: "bf16", hint: "default: stable, good quality" },
+  { value: "f16", label: "f16", hint: "stable fallback, similar memory to bf16" },
+  { value: "q8_0", label: "q8_0", hint: "lower memory, usually safe" },
+  { value: "q4_0", label: "q4_0", hint: "lowest memory, quality/speed tradeoff" },
 ];
 
 const GENERAL_DEFAULTS = {
@@ -26,49 +26,64 @@ export async function configureLocalProfile(prompt, profile) {
   const caps = profile.capabilities ?? {};
 
   console.log("");
-  console.log(renderSection("Let's set up this model", renderRows([
+  console.log(renderSection("Model setup", renderRows([
     ["Model", pc.bold(profile.label)],
-    ["Good for", humanCapabilitySummary(caps)],
-    ["Conversation memory", `${profile.flags.ctxSize.toLocaleString()} tokens`],
-    ["Memory mode", `${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV}`],
-    ["Response style", samplingSummary(profile.flags)],
+    ["Detected", detectionSummary(caps)],
+    ["Context", `${profile.flags.ctxSize.toLocaleString()} tokens`],
+    ["KV cache", `${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV}`],
+    ["Sampling", samplingSummary(profile.flags)],
   ])));
-  console.log(pc.dim("You can accept the recommended settings. Bigger conversation memory uses more RAM.\n"));
+  console.log(pc.dim("Larger context windows use more memory. KV cache precision controls memory used by attention history."));
+  console.log(pc.dim("Sampling defaults are shown for transparency; you can edit command.json later if needed.\n"));
 
   if (caps.mtp) {
-    console.log(renderSection("Speed boost available", "This model supports fast draft mode. It can make responses feel faster while keeping everything local.\n\nAdvanced: uses llama.cpp MTP on port 8081."));
-    const useMtp = await prompt.yesNo("Use fast draft mode for this model?", true);
+    console.log(renderSection("MTP detected", renderRows([
+      ["Backend", "llama.cpp MTP"],
+      ["Port", "8081"],
+      ["Flags", "--spec-type draft-mtp --spec-draft-n-max 2"],
+    ])));
+    const useMtp = await prompt.yesNo("Use MTP speculative decoding flags?", true);
     configured = useMtp ? applyMtpDefaults(configured) : removeMtpDefaults(configured);
   }
 
   if (caps.qat) {
     console.log("");
-    console.log(renderSection("Optimized model", "This model was trained to work well after compression. No extra runtime settings are needed."));
+    console.log(renderSection("QAT detected", renderRows([
+      ["Meaning", "quantization-aware trained"],
+      ["Runtime flags", "none required"],
+    ])));
   }
 
   if (caps.thinking) {
     console.log("");
-    console.log(renderSection("Reasoning mode", "This model can reason step by step. offgrid-ai can use safer defaults that reduce repetitive loops."));
-    const useThinking = await prompt.yesNo("Use reasoning-friendly defaults?", true);
+    console.log(renderSection("Thinking model detected", renderRows([
+      ["Defaults", "thinking / loop-safe"],
+      ["Flags", "--top-k 64 --presence-penalty 0 --repeat-penalty 1.1"],
+      ["Template", "--chat-template-kwargs { enable_thinking: true }"],
+    ])));
+    const useThinking = await prompt.yesNo("Use these thinking/loop-safe defaults?", true);
     configured = useThinking ? applyThinkingDefaults(configured) : removeThinkingDefaults(configured);
   }
 
-  const ctxSize = await prompt.number("Conversation memory tokens", configured.flags.ctxSize, 1024, 1048576);
-  const cacheTypeK = await prompt.choice("Memory mode, part 1", CACHE_CHOICES, configured.flags.cacheTypeK);
-  const cacheTypeV = await prompt.choice("Memory mode, part 2", CACHE_CHOICES, configured.flags.cacheTypeV);
+  const ctxSize = await prompt.number("Context window tokens", configured.flags.ctxSize, 1024, 1048576);
+  const cacheTypeK = await prompt.choice("K cache precision", CACHE_CHOICES, configured.flags.cacheTypeK);
+  const cacheTypeV = await prompt.choice("V cache precision", CACHE_CHOICES, configured.flags.cacheTypeV);
   configured = applyRuntimeFlagOverrides(configured, { ctxSize, cacheTypeK, cacheTypeV });
 
   console.log("");
-  console.log(renderSection("Final setup", renderRows([
-    ["Runs with", configured.backend],
-    ["Local address", configured.baseUrl],
-    ["Creativity", configured.flags.temperature],
-    ["Focus", configured.flags.topP],
-    ["Reasoning breadth", configured.flags.topK],
+  console.log(renderSection("Defaults", renderRows([
+    ["Backend", configured.backend],
+    ["Endpoint", configured.baseUrl],
+    ["Temperature", configured.flags.temperature],
+    ["Top-p", configured.flags.topP],
+    ["Top-k", configured.flags.topK],
+    ["Min-p", configured.flags.minP],
+    ["Presence penalty", configured.flags.presencePenalty],
+    ["Repeat penalty", configured.flags.repeatPenalty],
   ])));
 
   console.log("\n" + renderMemoryEstimate(configured));
-  if (!(await prompt.yesNo("Save this model setup?", true))) return null;
+  if (!(await prompt.yesNo("Save profile with these settings?", true))) return null;
   return configured;
 }
 
@@ -155,8 +170,8 @@ function renderMemoryEstimate(profile) {
     const est = estimateMemory(profile.modelPath, profile.mmprojPath, null, profile.flags);
     return renderSection("Memory estimate", renderRows([
       ["Estimated total", pc.bold(`~${formatBytes(est.totalBytes)}`)],
-      ["Model file", formatBytes(est.modelBytes)],
-      ["Conversation memory", est.kvBytes ? `~${formatBytes(est.kvBytes)} (${profile.flags.ctxSize.toLocaleString()} tokens, ${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV})` : "unknown"],
+      ["Model", formatBytes(est.modelBytes)],
+      ["KV cache", est.kvBytes ? `~${formatBytes(est.kvBytes)} (${profile.flags.ctxSize.toLocaleString()} ctx, ${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV})` : "unknown"],
       ...(est.note ? [["Note", pc.yellow(est.note)]] : []),
     ]));
   } catch {
@@ -164,6 +179,18 @@ function renderMemoryEstimate(profile) {
   }
 }
 
+function detectionSummary(caps) {
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.quant) parts.push(caps.quant);
+  if (caps.mtp) parts.push("MTP");
+  if (caps.qat) parts.push("QAT");
+
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  return parts.length > 0 ? parts.join(" · ") : "standard GGUF";
+}
+
 function samplingSummary(flags) {
   return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
 }

package/src/ui.mjs

@@ -83,8 +83,8 @@ export function humanCapabilitySummary(caps = {}) {
   const parts = [];
   if (caps.thinking) parts.push(pc.magenta("Reasoning"));
   if (caps.vision) parts.push(pc.cyan("Vision"));
-  if (caps.mtp) parts.push(pc.blue("Fast draft mode"));
-  if (caps.qat) parts.push(pc.green("Optimized quantization"));
+  if (caps.mtp) parts.push(pc.blue("MTP"));
+  if (caps.qat) parts.push(pc.green("QAT"));
   return parts.length > 0 ? parts.join(" · ") : "General chat";
 }