omnius 1.0.106 → 1.0.108

package/dist/index.js

@@ -578020,6 +578020,23 @@ async function needsTextToolMode(modelName, backendUrl2) {
   const hasTools = await checkToolSupport(modelName, backendUrl2);
   return !hasTools;
 }
+function detectUnifiedMemory() {
+  if (process.platform === "darwin" && process.arch === "arm64") return true;
+  if (process.platform === "linux") {
+    try {
+      if (existsSync91("/etc/nv_tegra_release")) return true;
+    } catch {
+    }
+    try {
+      if (existsSync91("/sys/devices/soc0/family")) {
+        const family = readFileSync74("/sys/devices/soc0/family", "utf8").trim().toLowerCase();
+        if (family.includes("tegra")) return true;
+      }
+    } catch {
+    }
+  }
+  return false;
+}
 function parseNvidiaSmi(stdout) {
   let total = 0;
   let free = 0;
@@ -578035,6 +578052,14 @@ function parseNvidiaSmi(stdout) {
   }
   return { total, free, name: name10 };
 }
+function parseFreeBytes(stdout) {
+  const memLine = stdout.split("\n").find((l2) => /^Mem:/i.test(l2)) ?? "";
+  const nums = memLine.match(/\d+/g);
+  if (!nums || nums.length === 0) return { total: 0, available: 0 };
+  const total = parseInt(nums[0], 10) || 0;
+  const available = nums.length >= 6 ? parseInt(nums[5], 10) || 0 : 0;
+  return { total, available };
+}
 function parseRocmSmi(stdout) {
   let total = 0;
   let used = 0;
@@ -578064,14 +578089,14 @@ function detectSystemSpecs() {
       timeout: 5e3
     });
     if (memInfo.includes("Mem:")) {
-      const match = memInfo.match(/^Mem:\s+(\d+)\s+\d+\s+\d+\s+\d+\s+\d+\s+(\d+)/m);
-      if (match) {
-        totalRamGB = parseInt(match[1], 10) / 1024 ** 3;
-        availableRamGB = parseInt(match[2], 10) / 1024 ** 3;
+      const { total, available } = parseFreeBytes(memInfo);
+      if (total > 0) {
+        totalRamGB = total / 1024 ** 3;
+        availableRamGB = available > 0 ? available / 1024 ** 3 : totalRamGB * 0.75;
       }
     } else {
       const bytes = parseInt(memInfo.trim(), 10);
-      if (!isNaN(bytes)) {
+      if (!isNaN(bytes) && bytes > 0) {
         totalRamGB = bytes / 1024 ** 3;
         availableRamGB = totalRamGB * 0.7;
       }
@@ -578102,12 +578127,18 @@ function detectSystemSpecs() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory();
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
   return {
     totalRamGB: Math.round(totalRamGB * 10) / 10,
     availableRamGB: Math.round(availableRamGB * 10) / 10,
     gpuVramGB: Math.round(gpuVramGB * 10) / 10,
     availableVramGB: Math.round(availableVramGB * 10) / 10,
-    gpuName
+    gpuName,
+    unifiedMemory
   };
 }
 async function detectSystemSpecsAsync() {
@@ -578122,14 +578153,14 @@ async function detectSystemSpecsAsync() {
       { timeout: 5e3 }
     );
     if (memInfo.includes("Mem:")) {
-      const match = memInfo.match(/^Mem:\s+(\d+)\s+\d+\s+\d+\s+\d+\s+\d+\s+(\d+)/m);
-      if (match) {
-        totalRamGB = parseInt(match[1], 10) / 1024 ** 3;
-        availableRamGB = parseInt(match[2], 10) / 1024 ** 3;
+      const { total, available } = parseFreeBytes(memInfo);
+      if (total > 0) {
+        totalRamGB = total / 1024 ** 3;
+        availableRamGB = available > 0 ? available / 1024 ** 3 : totalRamGB * 0.75;
       }
     } else {
       const bytes = parseInt(memInfo.trim(), 10);
-      if (!isNaN(bytes)) {
+      if (!isNaN(bytes) && bytes > 0) {
         totalRamGB = bytes / 1024 ** 3;
         availableRamGB = totalRamGB * 0.7;
       }
@@ -578160,12 +578191,18 @@ async function detectSystemSpecsAsync() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory();
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
   return {
     totalRamGB: Math.round(totalRamGB * 10) / 10,
     availableRamGB: Math.round(availableRamGB * 10) / 10,
     gpuVramGB: Math.round(gpuVramGB * 10) / 10,
     availableVramGB: Math.round(availableVramGB * 10) / 10,
-    gpuName
+    gpuName,
+    unifiedMemory
   };
 }
 function recommendModel(specs) {
@@ -578179,7 +578216,7 @@ function recommendModel(specs) {
   }
   return QWEN_VARIANTS.find((v) => v.tag === "qwen3.5:cloud");
 }
-function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
+function calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, archMax) {
   const ramBudget = specs.availableRamGB > 0 ? specs.availableRamGB : specs.totalRamGB;
   const vramBudget = specs.availableVramGB > 0 ? specs.availableVramGB : specs.gpuVramGB;
   const totalAvail = Math.max(vramBudget, ramBudget);
@@ -578190,11 +578227,19 @@ function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
     const maxTokens = Math.floor(usableGB * 1024 ** 3 / kvBytesPerToken);
     numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
   } else {
-    const kvEstimate = modelSizeGB2 <= 5 ? 524288 : modelSizeGB2 <= 20 ? 1048576 : 1572864;
+    const kvEstimate = modelSizeGB2 <= 5 ? 64 * 1024 : (
+      // 4B-class
+      modelSizeGB2 <= 12 ? 160 * 1024 : (
+        // 9B-class
+        modelSizeGB2 <= 25 ? 256 * 1024 : (
+          // 30B-class
+          384 * 1024
+        )
+      )
+    );
     const maxTokens = Math.floor(usableGB * 1024 ** 3 / kvEstimate);
     numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
   }
-  numCtx = Math.min(numCtx, 131072);
   if (archMax && archMax > 0) numCtx = Math.min(numCtx, archMax);
   if (kvBytesPerToken && kvBytesPerToken > 0 && modelSizeGB2 > 0) {
     const maxKVBytes = modelSizeGB2 * 4 * 1024 ** 3;
@@ -578202,6 +578247,11 @@ function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
     const budgetCap = Math.max(2048, Math.floor(maxTokensFromBudget / 1024) * 1024);
     numCtx = Math.min(numCtx, budgetCap);
   }
+  return numCtx;
+}
+function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
+  let numCtx = calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, archMax);
+  numCtx = Math.min(numCtx, 131072);
   const label = numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
   return { numCtx, label };
 }
@@ -578209,18 +578259,17 @@ function formatContextLabel(numCtx) {
   return numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
 }
 function calculateExpandedVariantContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
-  const memoryBudget = calculateContextWindow(
+  const memoryFit = calculateMemoryBoundedNumCtx(
     specs,
     modelSizeGB2,
     kvBytesPerToken,
     archMax
   );
-  if (archMax && archMax > 0) {
-    const archCtx = Math.max(2048, Math.floor(archMax / 1024) * 1024);
-    const numCtx = Math.min(archCtx, memoryBudget.numCtx);
-    return { numCtx, label: formatContextLabel(numCtx) };
-  }
-  return memoryBudget;
+  const archCtx = archMax && archMax > 0 ? Math.max(2048, Math.floor(archMax / 1024) * 1024) : Number.POSITIVE_INFINITY;
+  const floor = Math.min(EXPANDED_VARIANT_MIN_NUM_CTX, archCtx);
+  const fits = Math.min(memoryFit, archCtx);
+  const numCtx = Math.max(floor, fits);
+  return { numCtx, label: formatContextLabel(numCtx) };
 }
 function ask(rl, question) {
   return new Promise((resolve52) => {
@@ -579988,6 +580037,17 @@ function modelSizeGB(models, modelName) {
   const known = QWEN_VARIANTS.find((v) => modelName.includes(v.tag.split(":")[1] ?? ""));
   return known?.sizeGB ?? 4;
 }
+function defaultLayersForArch(arch3) {
+  const a2 = arch3.toLowerCase();
+  if (a2.includes("qwen")) return 36;
+  if (a2.includes("llama")) return 32;
+  if (a2.includes("mistral")) return 32;
+  if (a2.includes("phi")) return 32;
+  if (a2.includes("gemma")) return 42;
+  if (a2.includes("granite")) return 40;
+  if (a2.includes("command")) return 40;
+  return 32;
+}
 async function queryModelKVInfo(backendUrl2, modelName) {
   try {
     const normalized = backendUrl2.replace(/\/+$/, "");
@@ -580003,13 +580063,17 @@ async function queryModelKVInfo(backendUrl2, modelName) {
     const info = data.model_info;
     const arch3 = info["general.architecture"];
     if (!arch3) return null;
-    const nLayers = info[`${arch3}.block_count`];
-    const nKVHeads = info[`${arch3}.attention.head_count_kv`] ?? info[`${arch3}.attention.head_count`];
-    const keyDim = info[`${arch3}.attention.key_length`];
-    const valDim = info[`${arch3}.attention.value_length`] ?? keyDim;
+    const nLayersRaw = info[`${arch3}.block_count`];
+    const nKVHeadsRaw = info[`${arch3}.attention.head_count_kv`] ?? info[`${arch3}.attention.head_count`];
+    const keyDimRaw = info[`${arch3}.attention.key_length`];
+    const valDimRaw = info[`${arch3}.attention.value_length`] ?? keyDimRaw;
     const archMax = info[`${arch3}.context_length`];
     if (!archMax) return null;
-    if (!nLayers || !nKVHeads || !keyDim || !valDim) return { archMax };
+    const keyDim = keyDimRaw ?? 128;
+    const valDim = valDimRaw ?? 128;
+    const nLayers = nLayersRaw ?? defaultLayersForArch(arch3);
+    const nKVHeads = nKVHeadsRaw ?? 32;
+    if (!nLayers) return { archMax };
     const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
     return { kvBytesPerToken, archMax };
   } catch {
@@ -580412,7 +580476,7 @@ export PATH="${binDir}:$PATH"  # Added by omnius for nvim
   } catch {
   }
 }
-var execAsync2, OMNIUS_FIRST_RUN_BANNER, ANSI_RE, visibleLen2, QWEN_VARIANTS, _toolSupportCache, _cloudflaredInstallPromise;
+var execAsync2, OMNIUS_FIRST_RUN_BANNER, ANSI_RE, visibleLen2, QWEN_VARIANTS, _toolSupportCache, EXPANDED_VARIANT_MIN_NUM_CTX, _cloudflaredInstallPromise;
 var init_setup = __esm({
   "packages/cli/src/tui/setup.ts"() {
     "use strict";
@@ -580445,6 +580509,7 @@ var init_setup = __esm({
       { tag: "qwen3.5:397b-cloud", sizeGB: 0, label: "397B Cloud (Ollama Cloud)", cloud: true }
     ];
     _toolSupportCache = /* @__PURE__ */ new Map();
+    EXPANDED_VARIANT_MIN_NUM_CTX = 32768;
     _cloudflaredInstallPromise = null;
   }
 });

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.106",
+  "version": "1.0.108",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.106",
+      "version": "1.0.108",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.106",
+  "version": "1.0.108",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",