omnius 1.0.107 → 1.0.109

package/dist/index.js

@@ -577955,6 +577955,7 @@ __export(setup_exports, {
   ensurePythonVenv: () => ensurePythonVenv,
   ensureVisionDeps: () => ensureVisionDeps,
   expandedModelName: () => expandedModelName,
+  formatExpandedContextDiagnostic: () => formatExpandedContextDiagnostic,
   getLatestOllamaVersion: () => getLatestOllamaVersion,
   getOllamaVersion: () => getOllamaVersion,
   hasCmd: () => hasCmd,
@@ -578020,6 +578021,33 @@ async function needsTextToolMode(modelName, backendUrl2) {
   const hasTools = await checkToolSupport(modelName, backendUrl2);
   return !hasTools;
 }
+function detectUnifiedMemory(hasDiscreteGpu = false) {
+  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 {
+    }
+    try {
+      if (existsSync91("/proc/device-tree/model")) {
+        const model = readFileSync74("/proc/device-tree/model", "utf8").replace(/\0+$/, "").toLowerCase();
+        if (/jetson|tegra|orin|xavier|nano|raspberry|rockchip|rk\d{4}|mt\d{4}/.test(model)) {
+          return true;
+        }
+      }
+    } catch {
+    }
+    if (process.arch === "arm64" && !hasDiscreteGpu) return true;
+  }
+  return false;
+}
 function parseNvidiaSmi(stdout) {
   let total = 0;
   let free = 0;
@@ -578110,12 +578138,22 @@ function detectSystemSpecs() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory(gpuVramGB > 0);
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
+  if (!unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.75;
+    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() {
@@ -578168,12 +578206,22 @@ async function detectSystemSpecsAsync() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory(gpuVramGB > 0);
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
+  if (!unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.75;
+    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) {
@@ -578198,7 +578246,16 @@ function calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, arch
     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);
   }
@@ -578221,17 +578278,41 @@ function formatContextLabel(numCtx) {
   return numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
 }
 function calculateExpandedVariantContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
-  const memoryFit = calculateMemoryBoundedNumCtx(
-    specs,
-    modelSizeGB2,
-    kvBytesPerToken,
-    archMax
-  );
-  const archCtx = archMax && archMax > 0 ? Math.max(2048, Math.floor(archMax / 1024) * 1024) : Number.POSITIVE_INFINITY;
+  const ramBudget = specs.availableRamGB > 0 ? specs.availableRamGB : specs.totalRamGB;
+  const vramBudget = specs.availableVramGB > 0 ? specs.availableVramGB : specs.gpuVramGB;
+  const totalAvail = Math.max(vramBudget, ramBudget);
+  const remaining = Math.max(0, totalAvail - modelSizeGB2);
+  const usableGB = remaining * 0.85;
+  const memoryFit = calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, archMax);
+  const archCtxValue = archMax && archMax > 0 ? Math.max(2048, Math.floor(archMax / 1024) * 1024) : null;
+  const archCtx = archCtxValue ?? 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) };
+  let limitedBy;
+  if (numCtx === floor && fits < floor) limitedBy = "floor";
+  else if (archCtxValue !== null && numCtx === archCtxValue) limitedBy = "arch";
+  else limitedBy = "memory";
+  const effectiveKvBpt = kvBytesPerToken && kvBytesPerToken > 0 ? kvBytesPerToken : (modelSizeGB2 <= 5 ? 64 : modelSizeGB2 <= 12 ? 160 : modelSizeGB2 <= 25 ? 256 : 384) * 1024;
+  return {
+    numCtx,
+    label: formatContextLabel(numCtx),
+    math: {
+      numCtx,
+      label: formatContextLabel(numCtx),
+      modelSizeGB: modelSizeGB2,
+      kvBytesPerToken: effectiveKvBpt,
+      kvSource: kvBytesPerToken && kvBytesPerToken > 0 ? "model_info" : "fallback",
+      archMax: archMax && archMax > 0 ? archMax : null,
+      ramBudgetGB: ramBudget,
+      vramBudgetGB: vramBudget,
+      usableGB,
+      memoryFit,
+      archCtx: archCtxValue,
+      floor,
+      limitedBy
+    }
+  };
 }
 function ask(rl, question) {
   return new Promise((resolve52) => {
@@ -579999,6 +580080,17 @@ 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(/\/+$/, "");
@@ -580014,13 +580106,17 @@ async function queryModelKVInfo(backendUrl2, modelName) {
     const info = data.model_info;
     const arch3 = info["general.architecture"];
     if (!arch3) return null;
-    const nLayers = info[`${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`];
+    const keyDimRaw = info[`${arch3}.attention.key_length`];
+    const valDimRaw = info[`${arch3}.attention.value_length`] ?? keyDimRaw;
     const archMax = info[`${arch3}.context_length`];
     if (!archMax) return null;
-    if (!nLayers || !nKVHeads || !keyDim || !valDim) return { archMax };
+    const keyDim = keyDimRaw ?? 128;
+    const valDim = valDimRaw ?? 128;
+    const nLayers = nLayersRaw ?? defaultLayersForArch(arch3);
+    const nKVHeads = nKVHeadsRaw ?? 8;
+    if (!nLayers) return { archMax };
     const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
     return { kvBytesPerToken, archMax };
   } catch {
@@ -580151,6 +580247,21 @@ async function createExpandedVariantAsync(baseModel, specs, sizeGB, kvBytesPerTo
     archMax
   );
 }
+function formatExpandedContextDiagnostic(specs, math) {
+  const fmtGB = (n2) => `${n2.toFixed(1)}GB`;
+  const fmtKB = (n2) => `${Math.round(n2 / 1024)}KB`;
+  const fmtK = (n2) => n2 >= 1024 ? `${Math.floor(n2 / 1024)}K` : String(n2);
+  const memBits = [];
+  if (specs.gpuVramGB > 0) {
+    memBits.push(`VRAM ${fmtGB(specs.availableVramGB || specs.gpuVramGB)}/${fmtGB(specs.gpuVramGB)}`);
+  }
+  memBits.push(`RAM ${fmtGB(specs.availableRamGB)}/${fmtGB(specs.totalRamGB)}${specs.unifiedMemory ? " unified" : ""}`);
+  const mem = memBits.join(", ");
+  const kv = `KV ${fmtKB(math.kvBytesPerToken)}/tok (${math.kvSource})`;
+  const fit2 = `fit ${fmtK(math.memoryFit)}, arch ${math.archCtx !== null ? fmtK(math.archCtx) : "n/a"}, floor ${fmtK(math.floor)}`;
+  const limit = `→ ${fmtK(math.numCtx)} (${math.limitedBy === "floor" ? "min floor" : math.limitedBy === "arch" ? "arch-capped" : "memory-fit"})`;
+  return `[${mem} | model ${fmtGB(math.modelSizeGB)} | ${kv} | ${fit2} ${limit}]`;
+}
 async function ensureExpandedContext(modelName, backendUrl2) {
   if (modelName.includes("cloud") || modelName.includes(":cloud")) {
     return { model: modelName, created: false, contextLabel: "remote", numCtx: 0 };
@@ -580177,11 +580288,11 @@ async function ensureExpandedContext(modelName, backendUrl2) {
       kvInfo?.kvBytesPerToken,
       kvInfo?.archMax
     ).catch(() => ({ repaired: false, currentNumCtx: 0, baseModel: null, resolvedModel: modelName }));
-    return { model: repair.resolvedModel, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx };
+    return { model: repair.resolvedModel, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx, specs, math: ctx3.math };
   }
   const existing = await checkExpandedVariant(modelName, backendUrl2);
   if (existing === null) {
-    return { model: modelName, created: false, contextLabel: "", numCtx: 0 };
+    return { model: modelName, created: false, contextLabel: "", numCtx: 0, specs, math: ctx3.math };
   }
   if (typeof existing === "string") {
     const lostTools = await wrapperLacksToolsCapability(backendUrl2, existing).catch(() => false);
@@ -580196,7 +580307,7 @@ async function ensureExpandedContext(modelName, backendUrl2) {
           kvInfo?.archMax
         );
         if (rebuilt) {
-          return { model: rebuilt, created: true, contextLabel: ctx3.label, numCtx: ctx3.numCtx };
+          return { model: rebuilt, created: true, contextLabel: ctx3.label, numCtx: ctx3.numCtx, specs, math: ctx3.math };
         }
       } catch {
       }
@@ -580211,13 +580322,13 @@ async function ensureExpandedContext(modelName, backendUrl2) {
       kvInfo?.kvBytesPerToken,
       kvInfo?.archMax
     ).catch(() => ({ repaired: false, currentNumCtx: 0, baseModel: null, resolvedModel: existing }));
-    return { model: repair.resolvedModel, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx };
+    return { model: repair.resolvedModel, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx, specs, math: ctx3.math };
   }
   const created = await createExpandedVariantAsync(modelName, specs, sizeGB, kvInfo?.kvBytesPerToken, kvInfo?.archMax);
   if (created) {
-    return { model: created, created: true, contextLabel: ctx3.label, numCtx: ctx3.numCtx };
+    return { model: created, created: true, contextLabel: ctx3.label, numCtx: ctx3.numCtx, specs, math: ctx3.math };
   }
-  return { model: modelName, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx };
+  return { model: modelName, created: false, contextLabel: ctx3.label, numCtx: ctx3.numCtx, specs, math: ctx3.math };
 }
 function guessBaseFromVariant(variantName, models) {
   const stripped = stripVariantTag(variantName);
@@ -602188,14 +602299,15 @@ async function switchModel(query, ctx3, local = false) {
         match.name,
         ctx3.config.backendUrl
       );
+      const diag = result.specs && result.math ? "\n  " + c3.dim(formatExpandedContextDiagnostic(result.specs, result.math)) : "";
       if (result.created) {
         renderInfo(
-          `Created expanded context variant: ${c3.bold(result.model)} (${result.contextLabel}, ${result.numCtx} tokens)`
+          `Created expanded context variant: ${c3.bold(result.model)} (${result.contextLabel}, ${result.numCtx} tokens)${diag}`
         );
         finalModel = result.model;
       } else if (result.model !== match.name) {
         renderInfo(
-          `Using expanded context variant: ${c3.bold(result.model)} (${result.contextLabel})`
+          `Using expanded context variant: ${c3.bold(result.model)} (${result.contextLabel})${diag}`
         );
         finalModel = result.model;
       }
@@ -653743,13 +653855,14 @@ This is an independent background session started from /background.`
           currentConfig.model,
           currentConfig.backendUrl
         );
+        const diag = expandResult.specs && expandResult.math ? "\n  " + c3.dim(formatExpandedContextDiagnostic(expandResult.specs, expandResult.math)) : "";
         if (expandResult.created) {
           config = { ...config, model: expandResult.model };
           currentConfig = { ...currentConfig, model: expandResult.model };
           statusBar.setModelName(expandResult.model);
           writeContent(
             () => renderInfo(
-              `Created expanded context model: ${expandResult.model} (${expandResult.contextLabel}, ${expandResult.numCtx} tokens)`
+              `Created expanded context model: ${expandResult.model} (${expandResult.contextLabel}, ${expandResult.numCtx} tokens)${diag}`
             )
           );
         } else if (expandResult.model !== currentConfig.model) {
@@ -653758,7 +653871,7 @@ This is an independent background session started from /background.`
           statusBar.setModelName(expandResult.model);
           writeContent(
             () => renderInfo(
-              `Using expanded context model: ${expandResult.model} (${expandResult.contextLabel})`
+              `Using expanded context model: ${expandResult.model} (${expandResult.contextLabel})${diag}`
             )
           );
         }

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.107",
+  "version": "1.0.109",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.107",
+      "version": "1.0.109",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.107",
+  "version": "1.0.109",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",