pleumrouter 0.5.30 → 0.5.32

package/src/menu.mjs

@@ -11,12 +11,12 @@
 import { readFileSync } from "node:fs";
 import { spawn } from "node:child_process";
 import { createInterface } from "node:readline";
-import { style, displayWidth, terminalWidth, hyperlink, sanitizeTerminalText } from "./ansi.mjs";
+import { style, displayWidth, terminalWidth, terminalHeight, hyperlink, sanitizeTerminalText } from "./ansi.mjs";
 import { keyFeeder, select } from "./interactive.mjs";
-import { readConfig, writeConfig, isLoggedIn, touchSession, getActiveContext } from "./config.mjs";
+import { readConfig, writeConfig, isLoggedIn, touchSession, getActiveContext, isAutoUpdateEnabled } from "./config.mjs";
 import { RECIPES, MANUAL, DEFAULT_MODEL } from "./recipes.mjs";
 import { onPath, launch } from "./launch.mjs";
-import { fetchModels, matchModels, listModels, formatModelLabel } from "./models.mjs";
+import { fetchModels, matchModels, listModels, formatModelLabel, fetchCoreCodingModels, CORE_MODEL_PROVIDERS } from "./models.mjs";
 import { run } from "./run.mjs";
 import { login } from "./login.mjs";
 import { top } from "./top.mjs";
@@ -24,6 +24,7 @@ import { usage } from "./usage.mjs";
 import { balance } from "./balance.mjs";
 import { health } from "./health.mjs";
 import { runSettings } from "./settings.mjs";
+import { checkLatestVersion, performUpdate } from "./update.mjs";
 import { t } from "./i18n.mjs";
 
 // 값은 t() 키(영어 원문)만 담는다 — 여기서 t()를 호출하면 모듈 로드 시점(import 1회)에
@@ -123,24 +124,195 @@ function saveLastSelection(id) {
   writeConfig(cfg);
 }
 
+export async function pickProviderForModel(modelId, { input = process.stdin, output = process.stdout } = {}) {
+  if (!modelId || modelId === "__view_all__") return modelId;
+  const providers = CORE_MODEL_PROVIDERS[modelId] || [];
+  if (!providers.length) return modelId;
+
+  const options = [
+    {
+      id: null,
+      title: `⚡ ${t("Auto (best route)")}`,
+      desc: t("PleumRouter auto-routing — lowest price, lowest latency, automatic failover"),
+    },
+    ...providers.map((p) => ({
+      id: p.id,
+      title: p.name,
+      desc: t("Pin to {provider} (@{provider})", { provider: p.id }),
+    })),
+  ];
+
+  const chosen = await select(options, {
+    message: `${modelId} — ${t("Select routing provider")}`,
+    formatItem: (x) => `${x.title}  ${style.dim(x.desc || "")}`,
+    limit: 8,
+    input,
+    output,
+  });
+
+  if (!chosen) return null;
+  if (!chosen.id) return modelId;
+  return `${modelId}@${chosen.id}`;
+}
+
 export async function pickModelFor(item, defaultModel) {
-  let models;
-  try {
-    models = await fetchModels();
-  } catch (e) {
-    console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: sanitizeTerminalText(e.message) })}`);
-    return undefined;
+  for (;;) {
+    const core = fetchCoreCodingModels();
+    const items = [
+      ...core,
+      {
+        id: "__view_all__",
+        display_name: t("... view all 200+ models"),
+      },
+    ];
+    const targetId = defaultModel || DEFAULT_MODEL;
+    const initialIndex = items.findIndex((x) => x.id === targetId || targetId.startsWith(`${x.id}@`));
+    const m = await select(items, {
+      message: `${item.title} — ${t("select a model")}`,
+      formatItem: (x) => (x.id === "__view_all__" ? style.dim(x.display_name) : formatModelLabel(x)),
+      matches: matchModels,
+      limit: 11,
+      initialCursor: initialIndex >= 0 ? initialIndex : 0,
+    });
+    if (!m) return undefined;
+
+    let pickedModelId = m.id;
+    if (pickedModelId === "__view_all__") {
+      let allModels;
+      try {
+        allModels = await fetchModels();
+      } catch (e) {
+        console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: sanitizeTerminalText(e.message) })}`);
+        return undefined;
+      }
+      const allPicked = await select(allModels, {
+        message: `${item.title} — ${t("select a model")}`,
+        formatItem: (x) => formatModelLabel(x),
+        matches: matchModels,
+        limit: 10,
+      });
+      if (!allPicked) continue;
+      pickedModelId = allPicked.id;
+    }
+
+    const withProvider = await pickProviderForModel(pickedModelId);
+    if (!withProvider) continue;
+    return withProvider;
+  }
+}
+
+export async function pickClaudeCodeTiers(agentItem) {
+  const cfg = readConfig() ?? {};
+  const savedTiers = cfg.claude_tiers;
+  const defaultTiers = {
+    sonnet: "claude-sonnet-5",
+    opus: "claude-opus-5-5",
+    haiku: DEFAULT_MODEL,
+  };
+
+  const options = [
+    {
+      id: "preset:recommended",
+      title: `⚡ ${t("Recommended Preset")}`,
+      desc: `Main: ${defaultTiers.sonnet} · Thinker: ${defaultTiers.opus} · Subagent: ${defaultTiers.haiku}`,
+      tiers: defaultTiers,
+    },
+  ];
+
+  if (
+    savedTiers &&
+    (savedTiers.sonnet !== defaultTiers.sonnet ||
+      savedTiers.opus !== defaultTiers.opus ||
+      savedTiers.haiku !== defaultTiers.haiku)
+  ) {
+    options.push({
+      id: "preset:previous",
+      title: `↻ ${t("Use Previous Configuration")}`,
+      desc: `Main: ${savedTiers.sonnet} · Thinker: ${savedTiers.opus} · Subagent: ${savedTiers.haiku}`,
+      tiers: savedTiers,
+    });
+  }
+
+  options.push({
+    id: "preset:custom",
+    title: `⚙️ ${t("Configure Tiers Manually")}`,
+    desc: t("Choose specific models for Sonnet (main), Opus (thinker), and Haiku (subagent)"),
+  });
+
+  const picked = await select(options, {
+    message: `${agentItem.title} — ${t("Select model tier mapping")}`,
+    formatItem: (x) => `${x.title}\n    ${style.dim(x.desc)}`,
+    limit: 5,
+  });
+
+  if (!picked) return null;
+
+  if (picked.tiers) {
+    cfg.claude_tiers = picked.tiers;
+    writeConfig(cfg);
+    return picked.tiers;
   }
-  const targetId = defaultModel || DEFAULT_MODEL;
-  const initialIndex = models.findIndex((x) => x.id === targetId);
-  const m = await select(models, {
-    message: `${item.title} — ${t("select a model")}`,
-    formatItem: (x) => formatModelLabel(x),
-    matches: matchModels,
-    limit: 10,
+
+  const baseTiers = savedTiers || defaultTiers;
+
+  // Step 1: Sonnet (Main)
+  const sonnet = await pickModelFor(
+    { title: `${agentItem.title} — ${t("Main model (Sonnet tier)")}` },
+    baseTiers.sonnet,
+  );
+  if (!sonnet) return null;
+
+  // Step 2: Opus (Thinker)
+  const opus = await pickModelFor(
+    { title: `${agentItem.title} — ${t("Thinker model (Opus tier)")}` },
+    baseTiers.opus,
+  );
+  if (!opus) return null;
+
+  // Step 3: Haiku (Subagent)
+  const haiku = await pickModelFor(
+    { title: `${agentItem.title} — ${t("Subagent model (Haiku tier)")}` },
+    baseTiers.haiku,
+  );
+  if (!haiku) return null;
+
+  const resultTiers = { sonnet, opus, haiku };
+  cfg.claude_tiers = resultTiers;
+  writeConfig(cfg);
+  return resultTiers;
+}
+
+export async function pickCodexEffort(defaultEffort = "medium") {
+  const options = [
+    { id: "medium", title: `⚡ ${t("Medium (Recommended)")}`, desc: t("Balanced reasoning speed and depth") },
+    { id: "high", title: `🔥 ${t("High (Deep reasoning)")}`, desc: t("Maximum reasoning depth for complex tasks") },
+    { id: "low", title: `⚡ ${t("Low (Fast response)")}`, desc: t("Lightweight reasoning for quick answers") },
+  ];
+  const initialIndex = options.findIndex((x) => x.id === defaultEffort);
+  const picked = await select(options, {
+    message: `Codex — ${t("Select reasoning effort")}`,
+    formatItem: (x) => `${x.title}  ${style.dim(x.desc)}`,
+    limit: 5,
+    initialCursor: initialIndex >= 0 ? initialIndex : 0,
+  });
+  return picked ? picked.id : null;
+}
+
+export async function pickAgyEffort(defaultEffort = "high") {
+  const options = [
+    { id: "high", title: `⚡ ${t("High (Recommended)")}`, desc: t("High reasoning effort for standard tasks") },
+    { id: "max", title: `🚀 ${t("Max (Maximum reasoning)")}`, desc: t("Full reasoning effort for difficult problems") },
+    { id: "medium", title: `⚖️ ${t("Medium (Standard)")}`, desc: t("Balanced reasoning effort") },
+    { id: "low", title: `⚡ ${t("Low (Fast response)")}`, desc: t("Low reasoning effort for rapid iteration") },
+  ];
+  const initialIndex = options.findIndex((x) => x.id === defaultEffort);
+  const picked = await select(options, {
+    message: `Antigravity CLI (agy) — ${t("Select reasoning effort")}`,
+    formatItem: (x) => `${x.title}  ${style.dim(x.desc)}`,
+    limit: 6,
     initialCursor: initialIndex >= 0 ? initialIndex : 0,
   });
-  return m?.id;
+  return picked ? picked.id : null;
 }
 
 // 미설치 도구: ollama처럼 확인 후 그 자리에서 설치(stdio 상속). 성공 시 true.
@@ -184,6 +356,7 @@ export async function runList(items, { header, footer, initialCursor = 0 } = {})
   if (!input.isTTY || !output.isTTY || typeof input.setRawMode !== "function") return undefined;
 
   let cursor = Math.max(0, Math.min(initialCursor, items.length - 1));
+  let windowStart = 0;
   let lastRows = 0;
 
   const rowsFor = (s) => Math.max(1, Math.ceil(displayWidth(s) / Math.max(terminalWidth(), 1)));
@@ -192,28 +365,68 @@ export async function runList(items, { header, footer, initialCursor = 0 } = {})
     output.write("\r" + (lastRows > 1 ? `\x1b[${lastRows - 1}A` : "") + "\x1b[0J");
     lastRows = 0;
   };
+
+  const getMaxVisible = () => {
+    const h = terminalHeight();
+    return Math.max(3, Math.min(9, Math.floor((h - 6) / 2)));
+  };
+
   const render = () => {
+    const maxVisible = getMaxVisible();
+    if (cursor < windowStart) {
+      windowStart = cursor;
+    } else if (cursor >= windowStart + maxVisible) {
+      windowStart = cursor - maxVisible + 1;
+    }
+    windowStart = Math.max(0, Math.min(windowStart, Math.max(0, items.length - maxVisible)));
+
+    const visibleItems = items.slice(windowStart, windowStart + maxVisible);
     const lines = [
       header,
       "",
-      ...items.flatMap((it, i) =>
-        i === cursor
-          ? [`${style.plum("▸")} ${style.bold(it.title)}`, `    ${style.dim(it.desc)}`]
-          : [`  ${it.title}`, `    ${style.dim(it.desc)}`],
-      ),
-      "",
-      style.dim(footer),
     ];
+
+    if (windowStart > 0) {
+      lines.push(style.dim(`  ${t("↑ {n} more above", { n: windowStart })}`));
+    }
+
+    for (let idx = 0; idx < visibleItems.length; idx++) {
+      const i = windowStart + idx;
+      const it = visibleItems[idx];
+      if (i === cursor) {
+        lines.push(`${style.plum("▸")} ${style.bold(it.title)}`);
+        lines.push(`    ${style.dim(it.desc)}`);
+      } else {
+        lines.push(`  ${it.title}`);
+        lines.push(`    ${style.dim(it.desc)}`);
+      }
+    }
+
+    const remainingBelow = items.length - (windowStart + visibleItems.length);
+    if (remainingBelow > 0) {
+      lines.push(style.dim(`  ${t("↓ {n} more below", { n: remainingBelow })}`));
+    }
+
+    lines.push("");
+    lines.push(style.dim(footer));
+
     erase();
     output.write(lines.join("\n"));
     lastRows = lines.reduce((n, l) => n + rowsFor(l), 0);
   };
+
   const move = (dir) => {
     cursor = Math.max(0, Math.min(items.length - 1, cursor + dir));
   };
 
   return new Promise((resolve) => {
+    const onResize = () => {
+      render();
+    };
+    output.on("resize", onResize);
+
     const suspend = () => {
+      output.off("resize", onResize);
       input.off("data", onData);
       input.setRawMode(false);
       output.write("\x1b[?25h");
@@ -223,10 +436,12 @@ export async function runList(items, { header, footer, initialCursor = 0 } = {})
       input.setRawMode(true);
       input.resume();
       output.write("\x1b[?25l");
+      output.on("resize", onResize);
       input.on("data", onData);
       render();
     };
     const finish = (value) => {
+      output.off("resize", onResize);
       suspend();
       input.pause();
       resolve(value);
@@ -275,7 +490,7 @@ export async function runList(items, { header, footer, initialCursor = 0 } = {})
   });
 }
 
-async function menuSelect() {
+async function menuSelect(updateInfo = null) {
   const version = JSON.parse(readFileSync(new URL("../package.json", import.meta.url), "utf8")).version;
   const cfg = readConfig();
   const items = buildItems({
@@ -284,8 +499,19 @@ async function menuSelect() {
   });
   const last = cfg?.last_selection;
   const initialCursor = last ? Math.max(0, items.findIndex((it) => it.id === last)) : 0;
+  const headerLines = [`${style.bold(style.plum("PleumRouter"))} ${style.dim(`pleum ${version}`)}`];
+  if (updateInfo?.isOutdated) {
+    headerLines.push(
+      style.yellow(
+        `💡 ${t("Update available: v{current} → v{latest}. Run 'pleum update' to update.", {
+          current: updateInfo.currentVersion,
+          latest: updateInfo.latestVersion,
+        })}`,
+      ),
+    );
+  }
   return runList(items, {
-    header: `${style.bold(style.plum("PleumRouter"))} ${style.dim(`pleum ${version}`)}`,
+    header: headerLines.join("\n"),
     footer: t("↑/↓ move • enter run • → set model • esc quit"),
     initialCursor,
   });
@@ -303,7 +529,7 @@ async function agentSelect() {
 }
 
 // 에이전트 실행 메뉴 — bare pleum의 "에이전트 실행"과 pleum launch (인자 없음)가 공유.
-// 에이전트를 고르면 곧바로 핵심 10개 모델 피커를 띄워 기본 모델(deepseek-v4.1-flash) 또는
+// 에이전트를 고르면 곧바로 모델/티어/추론 피커를 띄워 기본 모델 또는
 // 선택한 우리 모델로 코딩 에이전트를 실행한다.
 export async function runAgentMenu({ loop = false } = {}) {
   const io = process.stdin;
@@ -330,15 +556,53 @@ export async function runAgentMenu({ loop = false } = {}) {
     }
 
     const agentSaved = readConfig()?.launch_models ?? {};
-    let chosenModel = agentPicked.model;
-    if (!chosenModel) {
-      const defaultForTool = agentSaved[agentItem.tool] || DEFAULT_MODEL;
-      chosenModel = await pickModelFor(agentItem, defaultForTool);
-      if (!chosenModel) continue; // Esc 취소 시 에이전트 목록으로 복귀
+    let launchModelPayload = null;
+
+    if (agentItem.tool === "claude") {
+      const tiers = await pickClaudeCodeTiers(agentItem);
+      if (!tiers) continue;
+      launchModelPayload = tiers;
+      saveModel("claude", tiers.sonnet);
+    } else if (agentItem.tool === "codex") {
+      let chosenModel = agentPicked.model;
+      if (!chosenModel) {
+        const defaultForTool = agentSaved[agentItem.tool] || DEFAULT_MODEL;
+        chosenModel = await pickModelFor(agentItem, defaultForTool);
+        if (!chosenModel) continue;
+      }
+      saveModel("codex", chosenModel);
+      const cfg = readConfig() ?? {};
+      const effort = await pickCodexEffort(cfg.codex_effort || "medium");
+      if (!effort) continue;
+      cfg.codex_effort = effort;
+      writeConfig(cfg);
+      launchModelPayload = { model: chosenModel, effort };
+    } else if (agentItem.tool === "agy") {
+      let chosenModel = agentPicked.model;
+      if (!chosenModel) {
+        const defaultForTool = agentSaved[agentItem.tool] || DEFAULT_MODEL;
+        chosenModel = await pickModelFor(agentItem, defaultForTool);
+        if (!chosenModel) continue;
+      }
+      saveModel("agy", chosenModel);
+      const cfg = readConfig() ?? {};
+      const effort = await pickAgyEffort(cfg.agy_effort || "high");
+      if (!effort) continue;
+      cfg.agy_effort = effort;
+      writeConfig(cfg);
+      launchModelPayload = { model: chosenModel, effort };
+    } else {
+      let chosenModel = agentPicked.model;
+      if (!chosenModel) {
+        const defaultForTool = agentSaved[agentItem.tool] || DEFAULT_MODEL;
+        chosenModel = await pickModelFor(agentItem, defaultForTool);
+        if (!chosenModel) continue;
+      }
+      saveModel(agentItem.tool, chosenModel);
+      launchModelPayload = chosenModel;
     }
 
-    saveModel(agentItem.tool, chosenModel);
-    await launch(agentItem.tool, chosenModel, [], { exit: false });
+    await launch(agentItem.tool, launchModelPayload, [], { exit: false });
     console.log();
     if (!loop) return true;
   }
@@ -357,8 +621,24 @@ export async function mainMenu() {
   if (!isLoggedIn()) await login();
   else await touchSession().catch(() => {});
 
+  let updateInfo = null;
+  try {
+    updateInfo = await checkLatestVersion();
+    if (updateInfo?.isOutdated) {
+      if (isAutoUpdateEnabled()) {
+        console.log(`💡 ${t("Update available: v{current} → v{latest}", { current: updateInfo.currentVersion, latest: updateInfo.latestVersion })}`);
+        console.log(`▶ ${t("Auto-updating PleumRouter CLI...")}\n`);
+        const code = await performUpdate();
+        if (code === 0) {
+          console.log(`\n✓ ${t("Update completed! Please run `pleum` again.")}`);
+          return true;
+        }
+      }
+    }
+  } catch {}
+
   for (;;) {
-    const picked = await menuSelect();
+    const picked = await menuSelect(updateInfo);
     if (picked === undefined) return false; // 비TTY
     if (picked === null || picked === "quit") return true; // esc/q/ctrl+c — ollama처럼 조용히 종료
 

package/src/models.mjs

@@ -1,13 +1,199 @@
+import { homedir } from "node:os";
+import { join } from "node:path";
+import { readFileSync, writeFileSync, mkdirSync, statSync } from "node:fs";
 import { V1 } from "./config.mjs";
 import { style, padDisplay, divider, sanitizeTerminalText } from "./ansi.mjs";
 import { t } from "./i18n.mjs";
 
+export const CORE_MODEL_PROVIDERS = {
+  "deepseek-ai/DeepSeek-V4.1-Flash": [
+    { id: "deepseek", name: "DeepSeek (Official)" },
+    { id: "together", name: "Together AI" },
+    { id: "fireworks", name: "Fireworks AI" },
+    { id: "siliconflow", name: "SiliconFlow" },
+  ],
+  "claude-sonnet-5": [
+    { id: "anthropic", name: "Anthropic (Official)" },
+    { id: "aws_bedrock", name: "AWS Bedrock" },
+    { id: "gcp_vertex", name: "Google Cloud Vertex AI" },
+  ],
+  "claude-opus-5-5": [
+    { id: "anthropic", name: "Anthropic (Official)" },
+    { id: "aws_bedrock", name: "AWS Bedrock" },
+    { id: "gcp_vertex", name: "Google Cloud Vertex AI" },
+  ],
+  "gpt-6-astra": [
+    { id: "openai", name: "OpenAI (Official)" },
+    { id: "azure", name: "Microsoft Azure OpenAI" },
+  ],
+  "gpt-6-sol": [
+    { id: "openai", name: "OpenAI (Official)" },
+    { id: "azure", name: "Microsoft Azure OpenAI" },
+  ],
+  "gpt-6-luna": [
+    { id: "openai", name: "OpenAI (Official)" },
+    { id: "azure", name: "Microsoft Azure OpenAI" },
+  ],
+  "gemini-3.8-flash": [
+    { id: "google", name: "Google AI Studio (Official)" },
+    { id: "vertex", name: "Google Cloud Vertex AI" },
+  ],
+  "qwen3.8-max": [
+    { id: "alibaba", name: "Alibaba DashScope (Official)" },
+    { id: "together", name: "Together AI" },
+    { id: "siliconflow", name: "SiliconFlow" },
+  ],
+  "glm-5.3-flash": [
+    { id: "zhipu", name: "Zhipu AI BigModel (Official)" },
+    { id: "modelark", name: "ByteDance ModelArk" },
+  ],
+  "kimi-k3": [
+    { id: "moonshot", name: "Moonshot AI (Official)" },
+  ],
+};
+
+export const CORE_MODEL_DEFAULTS = [
+  {
+    id: "deepseek-ai/DeepSeek-V4.1-Flash",
+    display_name: "DeepSeek V4.1 Flash (Ultra-fast coding)",
+    maker: "DeepSeek",
+    context_length: 131072,
+    pricing: { prompt: "0.00000014", completion: "0.00000028" },
+  },
+  {
+    id: "claude-sonnet-5",
+    display_name: "Claude Sonnet 5 (Anthropic flagship coding)",
+    maker: "Anthropic",
+    context_length: 200000,
+    pricing: { prompt: "0.000003", completion: "0.000015" },
+  },
+  {
+    id: "claude-opus-5-5",
+    display_name: "Claude Opus 5.5 (Deep reasoning & architecture)",
+    maker: "Anthropic",
+    context_length: 200000,
+    pricing: { prompt: "0.000005", completion: "0.000025" },
+  },
+  {
+    id: "gpt-6-astra",
+    display_name: "GPT-6 Astra (OpenAI flagship reasoning)",
+    maker: "OpenAI",
+    context_length: 128000,
+    pricing: { prompt: "0.0000025", completion: "0.00001" },
+  },
+  {
+    id: "gpt-6-sol",
+    display_name: "GPT-6 Sol (Fast coding & logic)",
+    maker: "OpenAI",
+    context_length: 128000,
+    pricing: { prompt: "0.000001", completion: "0.000004" },
+  },
+  {
+    id: "gpt-6-luna",
+    display_name: "GPT-6 Luna (Lightweight coding)",
+    maker: "OpenAI",
+    context_length: 128000,
+    pricing: { prompt: "0.0000005", completion: "0.000002" },
+  },
+  {
+    id: "gemini-3.8-flash",
+    display_name: "Gemini 3.8 Flash (1M context fast multimodal)",
+    maker: "Google",
+    context_length: 1048576,
+    pricing: { prompt: "0.00000035", completion: "0.0000014" },
+  },
+  {
+    id: "qwen3.8-max",
+    display_name: "Qwen 3.8 Max (Alibaba flagship coding)",
+    maker: "Alibaba",
+    context_length: 131072,
+    pricing: { prompt: "0.0000016", completion: "0.0000064" },
+  },
+  {
+    id: "glm-5.3-flash",
+    display_name: "GLM 5.3 Flash (Zhipu AI high-speed coding)",
+    maker: "Zhipu AI",
+    context_length: 128000,
+    pricing: { prompt: "0.0000001", completion: "0.0000002" },
+  },
+  {
+    id: "kimi-k3",
+    display_name: "Kimi K3 (Moonshot 256K long-context coding)",
+    maker: "Moonshot",
+    context_length: 262144,
+    pricing: { prompt: "0.0000006", completion: "0.0000024" },
+  },
+];
+
+let memoryModelsCache = null;
+let memoryCacheExpiry = 0;
+const CACHE_TTL_MS = 60 * 60 * 1000; // 1 hour
+
+function getCacheFilePath() {
+  return join(homedir(), ".pleum", "models-cache.json");
+}
+
+export function fetchCoreCodingModels() {
+  const cacheFile = getCacheFilePath();
+  let liveList = memoryModelsCache;
+  if (!liveList) {
+    try {
+      liveList = JSON.parse(readFileSync(cacheFile, "utf8"));
+    } catch {}
+  }
+  const byId = new Map((liveList || []).map((m) => [m.id, m]));
+  return CORE_MODEL_DEFAULTS.map((def) => {
+    const live = byId.get(def.id);
+    return live ? { ...def, ...live } : { ...def };
+  });
+}
+
 // GET /v1/models 는 공개(인증 불필요). 슬러그는 provider/model 형식.
-export async function fetchModels() {
-  const res = await fetch(`${V1}/models`);
-  if (!res.ok) throw new Error(t("Failed to fetch model list ({status})", { status: res.status }));
-  const data = await res.json();
-  return data.data ?? [];
+export async function fetchModels({ forceRefresh = false } = {}) {
+  const now = Date.now();
+  if (!forceRefresh && memoryModelsCache && now < memoryCacheExpiry) {
+    return memoryModelsCache;
+  }
+  const cacheFile = getCacheFilePath();
+  if (!forceRefresh && !memoryModelsCache) {
+    try {
+      const stats = statSync(cacheFile);
+      if (now - stats.mtimeMs < CACHE_TTL_MS) {
+        const cached = JSON.parse(readFileSync(cacheFile, "utf8"));
+        if (Array.isArray(cached) && cached.length > 0) {
+          memoryModelsCache = cached;
+          memoryCacheExpiry = stats.mtimeMs + CACHE_TTL_MS;
+          return cached;
+        }
+      }
+    } catch {}
+  }
+
+  try {
+    const res = await fetch(`${V1}/models`);
+    if (!res.ok) throw new Error(t("Failed to fetch model list ({status})", { status: res.status }));
+    const data = await res.json();
+    const list = data.data ?? [];
+    if (Array.isArray(list) && list.length > 0) {
+      memoryModelsCache = list;
+      memoryCacheExpiry = now + CACHE_TTL_MS;
+      try {
+        mkdirSync(join(homedir(), ".pleum"), { recursive: true, mode: 0o700 });
+        writeFileSync(cacheFile, JSON.stringify(list), { mode: 0o600 });
+      } catch {}
+    }
+    return list;
+  } catch (err) {
+    try {
+      const cached = JSON.parse(readFileSync(cacheFile, "utf8"));
+      if (Array.isArray(cached) && cached.length > 0) {
+        memoryModelsCache = cached;
+        memoryCacheExpiry = now + 60 * 1000;
+        return cached;
+      }
+    } catch {}
+    throw err;
+  }
 }
 
 // id/표시명 부분일치 필터 — run 피커·REPL /models·목록이 공유하는 단일 구현.