open-agents-ai 0.104.0 → 0.104.2

package/dist/index.js

@@ -30893,6 +30893,7 @@ ${activitySummary}
 });
 
 // packages/cli/dist/tui/model-picker.js
+import { totalmem as totalmem2 } from "node:os";
 async function fetchOllamaModels(baseUrl) {
   const url = `${normalizeBaseUrl(baseUrl)}/api/tags`;
   const resp = await fetch(url, {
@@ -30932,10 +30933,22 @@ async function fetchOllamaModels(baseUrl) {
       }
     }
     if (show.model_info) {
-      for (const [key, value] of Object.entries(show.model_info)) {
-        if (key.endsWith(".context_length") && typeof value === "number") {
-          result[i].contextLength = value;
-          break;
+      const info = show.model_info;
+      const arch = info["general.architecture"];
+      const paramCount = info["general.parameter_count"];
+      const fileSizeGB = result[i].sizeBytes > 0 ? result[i].sizeBytes / 1024 ** 3 : paramCount ? paramCount * 0.6 / 1024 ** 3 : 4;
+      if (arch) {
+        const archMax = info[`${arch}.context_length`];
+        const nLayers = info[`${arch}.block_count`];
+        const nKVHeads = info[`${arch}.attention.head_count_kv`] ?? info[`${arch}.attention.head_count`];
+        const keyDim = info[`${arch}.attention.key_length`];
+        const valDim = info[`${arch}.attention.value_length`] ?? keyDim;
+        if (archMax && nLayers && nKVHeads && keyDim && valDim) {
+          const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
+          result[i].contextLength = estimateRealisticContext(kvBytesPerToken, archMax, fileSizeGB);
+        } else if (archMax) {
+          const kvEstimate = fileSizeGB <= 5 ? 524288 : fileSizeGB <= 20 ? 1048576 : 1572864;
+          result[i].contextLength = estimateRealisticContext(kvEstimate, archMax, fileSizeGB);
         }
       }
     }
@@ -31231,9 +31244,23 @@ async function queryModelContextSize(baseUrl, modelName) {
         return parseInt(match[1], 10);
     }
     if (data.model_info) {
-      for (const [key, value] of Object.entries(data.model_info)) {
-        if (key.endsWith(".context_length") && typeof value === "number") {
-          return value;
+      const info = data.model_info;
+      const arch = info["general.architecture"];
+      const paramCount = info["general.parameter_count"];
+      const modelSizeGB2 = paramCount ? paramCount * 0.6 / 1024 ** 3 : 4;
+      if (arch) {
+        const archMax = info[`${arch}.context_length`];
+        const nLayers = info[`${arch}.block_count`];
+        const nKVHeads = info[`${arch}.attention.head_count_kv`] ?? info[`${arch}.attention.head_count`];
+        const keyDim = info[`${arch}.attention.key_length`];
+        const valDim = info[`${arch}.attention.value_length`] ?? keyDim;
+        if (archMax && nLayers && nKVHeads && keyDim && valDim) {
+          const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
+          return estimateRealisticContext(kvBytesPerToken, archMax, modelSizeGB2);
+        }
+        if (archMax) {
+          const kvEstimate = modelSizeGB2 <= 5 ? 524288 : modelSizeGB2 <= 20 ? 1048576 : 1572864;
+          return estimateRealisticContext(kvEstimate, archMax, modelSizeGB2);
         }
       }
     }
@@ -31242,10 +31269,25 @@ async function queryModelContextSize(baseUrl, modelName) {
     return null;
   }
 }
+function estimateRealisticContext(kvBytesPerToken, archMax, modelSizeGB2) {
+  const totalMemGB = totalmem2() / 1024 ** 3;
+  const usableBytes = totalMemGB * 0.7 * 1024 ** 3;
+  const maxTokens = Math.floor(usableBytes / kvBytesPerToken);
+  let numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
+  numCtx = Math.min(numCtx, 131072, archMax);
+  if (modelSizeGB2 && modelSizeGB2 > 0) {
+    const maxKVBytes = modelSizeGB2 * 4 * 1024 ** 3;
+    const budgetCap = Math.max(2048, Math.floor(maxKVBytes / kvBytesPerToken / 1024) * 1024);
+    numCtx = Math.min(numCtx, budgetCap);
+  }
+  return numCtx;
+}
 async function queryOpenAIContextSize(baseUrl, modelName, apiKey) {
   try {
     const models = await fetchOpenAIModels(baseUrl, apiKey);
     const model = models.find((m) => m.name === modelName);
+    if (model?.contextLength)
+      return model.contextLength;
     if (model?.size) {
       const match = model.size.match(/(\d+)K ctx/);
       if (match)
@@ -32155,20 +32197,30 @@ function recommendModel(specs) {
   }
   return QWEN_VARIANTS.find((v) => v.tag === "qwen3.5:cloud");
 }
-function calculateContextWindow(specs, modelSizeGB2) {
+function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
   const totalAvail = Math.max(specs.gpuVramGB, specs.totalRamGB);
-  const remaining = totalAvail - modelSizeGB2;
-  if (remaining >= 200)
-    return { numCtx: 131072, label: "128K" };
-  if (remaining >= 100)
-    return { numCtx: 65536, label: "64K" };
-  if (remaining >= 50)
-    return { numCtx: 32768, label: "32K" };
-  if (remaining >= 20)
-    return { numCtx: 16384, label: "16K" };
-  if (remaining >= 8)
-    return { numCtx: 8192, label: "8K" };
-  return { numCtx: 4096, label: "4K" };
+  const remaining = Math.max(0, totalAvail - modelSizeGB2);
+  const usableGB = remaining * 0.85;
+  let numCtx;
+  if (kvBytesPerToken && kvBytesPerToken > 0) {
+    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 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;
+    const maxTokensFromBudget = Math.floor(maxKVBytes / kvBytesPerToken);
+    const budgetCap = Math.max(2048, Math.floor(maxTokensFromBudget / 1024) * 1024);
+    numCtx = Math.min(numCtx, budgetCap);
+  }
+  const label = numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
+  return { numCtx, label };
 }
 function modelSupportsToolCalling(modelName) {
   const lower = modelName.toLowerCase();
@@ -33512,9 +33564,40 @@ function modelSizeGB(models, modelName) {
   const known = QWEN_VARIANTS.find((v) => modelName.includes(v.tag.split(":")[1] ?? ""));
   return known?.sizeGB ?? 4;
 }
-async function createExpandedVariantAsync(baseModel, specs, sizeGB) {
+async function queryModelKVInfo(backendUrl, modelName) {
+  try {
+    const normalized = backendUrl.replace(/\/+$/, "");
+    const res = await fetch(`${normalized}/api/show`, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ name: modelName }),
+      signal: AbortSignal.timeout(5e3)
+    });
+    if (!res.ok)
+      return null;
+    const data = await res.json();
+    if (!data.model_info)
+      return null;
+    const info = data.model_info;
+    const arch = info["general.architecture"];
+    if (!arch)
+      return null;
+    const nLayers = info[`${arch}.block_count`];
+    const nKVHeads = info[`${arch}.attention.head_count_kv`] ?? info[`${arch}.attention.head_count`];
+    const keyDim = info[`${arch}.attention.key_length`];
+    const valDim = info[`${arch}.attention.value_length`] ?? keyDim;
+    const archMax = info[`${arch}.context_length`];
+    if (!nLayers || !nKVHeads || !keyDim || !valDim || !archMax)
+      return null;
+    const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
+    return { kvBytesPerToken, archMax };
+  } catch {
+    return null;
+  }
+}
+async function createExpandedVariantAsync(baseModel, specs, sizeGB, kvBytesPerToken, archMax) {
   const customName = expandedModelName(baseModel);
-  const ctx = calculateContextWindow(specs, sizeGB);
+  const ctx = calculateContextWindow(specs, sizeGB, kvBytesPerToken, archMax);
   try {
     const numPredict = Math.min(16384, Math.max(2048, Math.floor(ctx.numCtx * 0.25)));
     const modelfileContent = [
@@ -33537,37 +33620,51 @@ async function createExpandedVariantAsync(baseModel, specs, sizeGB) {
   }
 }
 async function ensureExpandedContext(modelName, backendUrl) {
-  if (modelName.startsWith("open-agents-")) {
-    const specs2 = await detectSystemSpecsAsync();
-    const ctx2 = calculateContextWindow(specs2, 4);
-    return { model: modelName, created: false, contextLabel: ctx2.label, numCtx: ctx2.numCtx };
-  }
   if (modelName.includes("cloud") || modelName.includes(":cloud")) {
     return { model: modelName, created: false, contextLabel: "remote", numCtx: 0 };
   }
-  const existing = await checkExpandedVariant(modelName, backendUrl);
-  if (existing === null) {
-    return { model: modelName, created: false, contextLabel: "", numCtx: 0 };
-  }
   const specs = await detectSystemSpecsAsync();
-  if (typeof existing === "string") {
-    let sizeGB2 = 4;
-    try {
-      const models = await fetchOllamaModels(backendUrl);
-      sizeGB2 = modelSizeGB(models, modelName);
-    } catch {
-    }
-    const ctx2 = calculateContextWindow(specs, sizeGB2);
-    return { model: existing, created: false, contextLabel: ctx2.label, numCtx: ctx2.numCtx };
-  }
+  const kvInfo = await queryModelKVInfo(backendUrl, modelName);
   let sizeGB = 4;
   try {
     const models = await fetchOllamaModels(backendUrl);
     sizeGB = modelSizeGB(models, modelName);
   } catch {
   }
-  const ctx = calculateContextWindow(specs, sizeGB);
-  const created = await createExpandedVariantAsync(modelName, specs, sizeGB);
+  const ctx = calculateContextWindow(specs, sizeGB, kvInfo?.kvBytesPerToken, kvInfo?.archMax);
+  if (modelName.startsWith("open-agents-")) {
+    try {
+      const normalized = backendUrl.replace(/\/+$/, "");
+      const showRes = await fetch(`${normalized}/api/show`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({ name: modelName }),
+        signal: AbortSignal.timeout(5e3)
+      });
+      if (showRes.ok) {
+        const showData = await showRes.json();
+        const numCtxMatch = showData.parameters?.match(/num_ctx\s+(\d+)/);
+        const currentNumCtx = numCtxMatch ? parseInt(numCtxMatch[1], 10) : 0;
+        if (currentNumCtx !== ctx.numCtx) {
+          const fromMatch = showData.modelfile?.match(/^FROM\s+(.+)$/m);
+          const baseModel = fromMatch?.[1]?.trim();
+          if (baseModel && !baseModel.startsWith("open-agents-")) {
+            await createExpandedVariantAsync(baseModel, specs, sizeGB, kvInfo?.kvBytesPerToken, kvInfo?.archMax);
+          }
+        }
+      }
+    } catch {
+    }
+    return { model: modelName, created: false, contextLabel: ctx.label, numCtx: ctx.numCtx };
+  }
+  const existing = await checkExpandedVariant(modelName, backendUrl);
+  if (existing === null) {
+    return { model: modelName, created: false, contextLabel: "", numCtx: 0 };
+  }
+  if (typeof existing === "string") {
+    return { model: existing, created: false, contextLabel: ctx.label, numCtx: ctx.numCtx };
+  }
+  const created = await createExpandedVariantAsync(modelName, specs, sizeGB, kvInfo?.kvBytesPerToken, kvInfo?.archMax);
   if (created) {
     return { model: created, created: true, contextLabel: ctx.label, numCtx: ctx.numCtx };
   }
@@ -33720,7 +33817,7 @@ function tuiSelect(opts) {
     cursor = first >= 0 ? first : 0;
   }
   const selectChrome = 3;
-  const contentArea = process.stdout.rows ?? 24;
+  const contentArea = opts.availableRows ?? process.stdout.rows ?? 24;
   const maxVisible = opts.maxVisible ?? Math.max(3, contentArea - selectChrome);
   let scrollOffset = 0;
   let lastRenderedLines = 0;
@@ -46337,7 +46434,7 @@ var init_braille_spinner = __esm({
 });
 
 // packages/cli/dist/tui/system-metrics.js
-import { loadavg as loadavg2, cpus as cpus2, totalmem as totalmem2, freemem as freemem2, platform as platform4 } from "node:os";
+import { loadavg as loadavg2, cpus as cpus2, totalmem as totalmem3, freemem as freemem2, platform as platform4 } from "node:os";
 import { exec as exec3 } from "node:child_process";
 import { readFile as readFile16 } from "node:fs/promises";
 function formatRate(bytesPerSec) {
@@ -46471,7 +46568,7 @@ function collectCpuRam() {
   const [l1] = loadavg2();
   const cores = cpus2().length;
   const cpuModel = cpus2()[0]?.model ?? "";
-  const totalMem = totalmem2();
+  const totalMem = totalmem3();
   const usedMem = totalMem - freemem2();
   return {
     cpuUtil: Math.min(100, Math.round(l1 / cores * 100)),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.104.0",
+  "version": "0.104.2",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",