pleumrouter 0.5.4 → 0.5.6

package/README.md

@@ -33,7 +33,7 @@ cat error.log | npx pleumrouter run gpt-5.4-mini "원인은?"   # 파이프
 | `pleum login [--device]` | 브라우저로 로그인 → CLI가 API 키를 자동 발급·저장(`~/.pleum/config.json`). `--device`는 SSH·CI |
 | `pleum logout` | 저장된 API 키만 삭제(언어·기본 모델 등 나머지 설정은 유지) |
 | `pleum launch [tool] [--model id] [-- args]` | 에이전트 실행. `--` 뒤 인자는 도구로 그대로 전달. 도구 생략 시 지원 목록·설치 여부 표시 |
-| `pleum run [model] ["프롬프트"]` | 채팅 — 대화형 REPL, 1회 실행, stdin 파이프. 모델 생략 시 화살표로 선택. `--skip-cache`(이번 요청 응답 캐시 스킵), `--tag a,b`(리포팅 태그, ≤10) |
+| `pleum run [model] ["프롬프트"] [--attach image]` | 채팅 — 대화형 REPL, 1회 실행, stdin 파이프. 모델 생략 시 화살표로 선택. 이미지 업로드는 명시적으로 `--attach ./shot.png`(1회 실행) 또는 REPL의 `@./shot.png`를 사용한다. `--skip-cache`(이번 요청 응답 캐시 스킵), `--tag a,b`(리포팅 태그, ≤10) |
 | `pleum models [query] [--json]` | 모델 목록(공개) — provider·컨텍스트·가격(₩/1M) 컬럼, `--json`은 스크립트용 |
 | `pleum models <id> --info` | 모델 상세 — 가격·컨텍스트·provider별 지연/처리량/업타임·주간 사용량·형제 모델 |
 | `pleum top [--period daily\|weekly\|monthly]` | 라우터 전체 모델 리더보드 + 제공자 점유율(공개 통계) |
@@ -64,7 +64,7 @@ cat error.log | npx pleumrouter run gpt-5.4-mini "원인은?"   # 파이프
 - 응답 후 비용이 현지 통화(한국 ₩)로 표시(스트리밍 마지막 청크 기준)
 - **Ctrl+C**: 생성 중엔 1회로 중단(대화 유지), 대기 중엔 2초 내 2회로 종료
 - 여러 줄 입력: `"""` 한 줄 → 본문 → `"""` 한 줄
-- 이미지: 메시지에 이미지 파일 경로를 그대로 포함 (예: `이 그림 설명해줘 ./shot.png`)
+- 이미지: 명시적으로 첨부 — REPL에서는 `이 그림 설명해줘 @./shot.png`, 1회 실행에서는 `--attach ./shot.png`. 일반 프롬프트의 `./shot.png` 텍스트는 업로드되지 않는다. 첨부는 일반 파일·이미지 magic bytes·5 MiB/장(총 10 MiB)을 확인하고 심볼릭 링크는 거부한다.
 - `↑` 로 지난 입력 재호출(히스토리는 `~/.pleum/history`에 영속)
 - 색상 끄기: `NO_COLOR=1` (비TTY·파이프에선 자동 무색)
 
@@ -78,7 +78,7 @@ env(또는 격리된 `CODEX_HOME`/임시 설정 문자열)만 세팅하고 바
 
 ## 환경변수
 
-- `PLEUM_BASE_URL` — 게이트웨이 루트 덮어쓰기(기본 `https://router.pleum.ai`). 셀프호스트/스테이징용.
+- `PLEUM_BASE_URL` — 게이트웨이 루트 덮어쓰기(기본 `https://apirouter.pleum.ai`). 셀프호스트/스테이징용.
 
 ## 개발
 

package/bin/pleum.mjs

@@ -15,6 +15,7 @@ import { mainMenu } from "../src/menu.mjs";
 import { lang } from "../src/lang.mjs";
 import { RECIPES } from "../src/recipes.mjs";
 import { t } from "../src/i18n.mjs";
+import { sanitizeTerminalText } from "../src/ansi.mjs";
 
 // 명령 구문(왼쪽)은 로케일과 무관 — 설명(오른쪽)만 번역, 폭은 가장 긴 구문 기준으로 계산.
 function buildHelp() {
@@ -26,7 +27,7 @@ function buildHelp() {
     ["pleum launch [tool] [--model id]", t("Run a coding agent against PleumRouter")],
     ["", `(${Object.keys(RECIPES).join(" · ")})`],
     ["", t("omit the tool to see the supported list and install status")],
-    ['pleum run [model] ["prompt"]', t("Terminal chat (REPL) / one-shot / pipe")],
+    ['pleum run [model] ["prompt"] [--attach image]', t("Terminal chat (REPL) / one-shot / pipe")],
     ["", t("omit the model for an arrow-key picker; /help lists commands")],
     ["pleum models [query] [--json]", t("Model list (price, context included)")],
     ["pleum models <id> --info", t("Model detail (price, perf, per-provider uptime)")],
@@ -146,6 +147,6 @@ try {
       process.exit(1);
   }
 } catch (e) {
-  console.error(`✗ ${e.message}`);
+  console.error(`✗ ${sanitizeTerminalText(e.message)}`);
   process.exit(1);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pleumrouter",
-  "version": "0.5.4",
+  "version": "0.5.6",
   "description": "Official PleumRouter terminal CLI — launch any coding agent against PleumRouter.",
   "type": "module",
   "bin": {

package/src/ansi.mjs

@@ -55,6 +55,18 @@ export function stripAnsi(s) {
     .replace(/\x1b\[[0-9;<]*[A-Za-z]/g, "");
 }
 
+// 원격 응답/메타데이터는 신뢰하지 않는다. 터미널 상태를 바꾸는 ESC·C0/C1 제어문자는
+// 제거하되, 텍스트 스트리밍에 필요한 LF와 TAB은 그대로 둔다. CSI/OSC 등의 본문까지
+// 먼저 제거해 ESC만 떼었을 때 남는 `[2J` 같은 노이즈도 남기지 않는다.
+export function sanitizeTerminalText(value) {
+  return String(value ?? "")
+    .replace(/\x1b\][\s\S]*?(?:\x07|\x1b\\)/g, "") // OSC (title, hyperlink, clipboard…)
+    .replace(/\x1b[PX^_][\s\S]*?\x1b\\/g, "") // DCS, SOS, PM, APC
+    .replace(/(?:\x1b\[|\x9b)[0-?]*[ -/]*[@-~]/g, "") // CSI + C1 CSI
+    .replace(/\x1b(?:[()][0-2A-Z0-9]|[=>][0-?]*|.)/g, "") // remaining ESC sequences
+    .replace(/[\u0000-\u0008\u000b-\u001f\u007f-\u009f]/g, "");
+}
+
 // OSC 8 클릭 가능한 하이퍼링크. 무색(파이프·NO_COLOR·dumb)이면 텍스트만.
 export function hyperlink(url, text = url, level = style.level) {
   return level ? `\x1b]8;;${url}\x1b\\${text}\x1b]8;;\x1b\\` : text;

package/src/config.mjs

@@ -3,8 +3,9 @@ import { join } from "node:path";
 import { readFileSync, writeFileSync, mkdirSync } from "node:fs";
 import { t } from "./i18n.mjs";
 
-// 공개 게이트웨이 루트. 셀프호스트/스테이징은 PLEUM_BASE_URL로 덮어쓴다.
-export const ROOT = (process.env.PLEUM_BASE_URL ?? "https://router.pleum.ai")
+// 공개 게이트웨이 루트(canonical API). 셀프호스트/스테이징은 PLEUM_BASE_URL로 덮어쓴다.
+// Compat: https://router.pleum.ai 도 180일+ 유지(프론트 /v1 rewrite). 신규 기본은 apirouter.
+export const ROOT = (process.env.PLEUM_BASE_URL ?? "https://apirouter.pleum.ai")
   .replace(/\/+$/, "")
   .replace(/\/v1$/, ""); // 사용자가 실수로 /v1까지 붙여도 V1이 /v1/v1 이중 경로가 되지 않게 정리(L12)
 // OpenAI 호환 엔드포인트는 /v1. (Claude Code만 루트를 쓴다 — recipes 참고)

package/src/launch.mjs

@@ -1,19 +1,69 @@
 import { spawn } from "node:child_process";
-import { mkdirSync, writeFileSync, existsSync } from "node:fs";
+import { mkdirSync, writeFileSync, existsSync, readFileSync } from "node:fs";
 import { homedir } from "node:os";
-import { join, delimiter } from "node:path";
+import { join, delimiter, dirname, resolve } from "node:path";
 import { requireKey } from "./config.mjs";
 import { RECIPES, MANUAL, codexToml } from "./recipes.mjs";
 import { t } from "./i18n.mjs";
+import { sanitizeTerminalText } from "./ansi.mjs";
 
 // PATH에서 실행 파일을 찾는다(설치 여부 표시용). menu.mjs도 사용.
 export function onPath(bin) {
-  const exts = process.platform === "win32" ? [".cmd", ".exe", ".bat", ""] : [""];
+  const exts = process.platform === "win32" ? [".exe", ".cmd", ".bat", ""] : [""];
   return (process.env.PATH || "")
     .split(delimiter)
     .some((d) => d && exts.some((e) => existsSync(join(d, bin + e))));
 }
 
+// Windows는 .cmd/.bat 파일을 셸 없이 직접 실행할 수 없다. npm이 만드는 표준 .cmd
+// shim은 실제 Node 엔트리포인트를 가리키므로 그 JS 파일을 현재 Node로 직접 실행한다.
+// 임의의 batch script는 cmd.exe로 넘기지 않고 명시적으로 거부한다. 그래야 --model과
+// `--` 패스스루 인자가 shell 문법으로 재해석될 수 없다.
+function windowsCommandCandidates(bin, path, pathDelimiter) {
+  const hasPath = /[\\/]/.test(bin);
+  const hasExtension = /\.[^\\/]+$/.test(bin);
+  const names = hasExtension ? [bin] : [`${bin}.exe`, `${bin}.cmd`, `${bin}.bat`, bin];
+  const dirs = hasPath ? [""] : path.split(pathDelimiter).filter(Boolean);
+  return dirs.flatMap((dir) => names.map((name) => (dir ? join(dir, name) : name)));
+}
+
+function nodeEntryFromWindowsShim(script, { read = readFileSync, exists = existsSync } = {}) {
+  let source;
+  try {
+    source = String(read(script, "utf8"));
+  } catch {
+    return null;
+  }
+  // npm cmd-shim uses either %dp0% or %~dp0 followed by the package's JS entry.
+  const match = /"(?:%~dp0|%dp0%)[\\/]+([^"\r\n]+?\.(?:[cm]?js))"/i.exec(source);
+  if (!match) return null;
+  const entry = resolve(dirname(script), match[1].replace(/[\\/]+/g, "/"));
+  return exists(entry) ? entry : null;
+}
+
+// 플랫폼·filesystem 주입으로 Windows 경로 정책을 Linux/macOS에서도 회귀 테스트한다.
+export function resolveLaunchCommand(bin, args, {
+  platform = process.platform,
+  path = process.env.PATH || "",
+  delimiter: pathDelimiter = delimiter,
+  exists = existsSync,
+  read = readFileSync,
+  node = process.execPath,
+} = {}) {
+  if (platform !== "win32") return { command: bin, args };
+  const found = windowsCommandCandidates(bin, path, pathDelimiter).find((candidate) => exists(candidate));
+  if (!found) return { command: bin, args };
+  if (!/\.(?:cmd|bat)$/i.test(found)) return { command: found, args };
+
+  const entry = nodeEntryFromWindowsShim(found, { read, exists });
+  if (!entry) {
+    throw new Error(
+      `Windows command script '${bin}' cannot be launched safely. Install an .exe or a standard npm Node shim.`,
+    );
+  }
+  return { command: node, args: [entry, ...args] };
+}
+
 // `pleum launch`(도구 미지정) — 지원 도구 + 설치 여부를 안내.
 export function launchHelp() {
   console.log(t("Usage: pleum launch <tool> [--model id] [-- tool args]\n"));
@@ -60,13 +110,22 @@ export async function launch(tool, model, passthrough, { exit = true } = {}) {
 
   const args = [...r.args(model), ...passthrough];
   if (r.note) console.log(`ℹ ${t(r.note)}`);
-  console.log(`▶ ${t("Running {label}: {cmd}", { label: r.label, cmd: `${r.bin} ${args.join(" ")}` })}\n`);
+  console.log(`▶ ${t("Running {label}: {cmd}", { label: r.label, cmd: sanitizeTerminalText(`${r.bin} ${args.join(" ")}`) })}\n`);
+
+  let command;
+  try {
+    command = resolveLaunchCommand(r.bin, args);
+  } catch (e) {
+    console.error(`✗ ${sanitizeTerminalText(e.message)}`);
+    if (exit) process.exit(1);
+    return 1;
+  }
 
   return await new Promise((resolve) => {
-    const child = spawn(r.bin, args, {
+    const child = spawn(command.command, command.args, {
       env,
       stdio: "inherit",
-      shell: process.platform === "win32",
+      shell: false,
     });
     child.on("error", (e) => {
       if (e.code === "ENOENT") {
@@ -74,7 +133,7 @@ export async function launch(tool, model, passthrough, { exit = true } = {}) {
         if (exit) process.exit(127);
         return resolve(127);
       }
-      console.error(`✗ ${e.message}`);
+      console.error(`✗ ${sanitizeTerminalText(e.message)}`);
       if (exit) process.exit(1);
       resolve(1);
     });

package/src/locales/es.mjs

@@ -12,7 +12,7 @@ export const ES = {
   "Account": "Cuenta",
   "All models — price and context included": "Todos los modelos — precio y contexto incluidos",
   "Anthropic's coding agent — supports subagents": "Agente de codificación de Anthropic — admite subagentes",
-  "Attaching and uploading {n} local image(s) from the prompt.": "Adjuntando y subiendo {n} imagen(es) local(es) del prompt.",
+  "Attaching and uploading {n} explicit local image(s).": "Adjuntando y subiendo {n} imagen(es) local(es) seleccionada(s) explícitamente.",
   "Auto-run:": "Ejecución automática:",
   "Autonomous dev agent": "Agente de desarrollo autónomo",
   "Multi-provider terminal coding agent": "Agente de codificación de terminal multi-proveedor",
@@ -32,7 +32,7 @@ export const ES = {
   "Choose a number, or search (Enter to cancel): ": "Elige un número, o busca (Enter para cancelar): ",
   "Choose a shell": "Elegir una shell",
   "Code": "Código",
-  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: include an image file path right in the message (e.g. describe this ./shot.png)": "Comandos:\n  /help, /?           esta ayuda\n  /clear              reinicia el contexto de la conversación (conserva el system prompt)\n  /system [text]      define el system prompt (sin argumento muestra el valor actual)\n  /model [id]         cambia de modelo durante la sesión (sin argumento = selector con flechas)\n  /models [query]     explora la lista de modelos\n  /save [name]        guarda la sesión actual (por defecto: last)\n  /load [name]        carga una sesión guardada\n  /verbose            activa/desactiva estadísticas de tokens/velocidad\n  /exit, /quit        salir\n  Entrada multilínea: una línea con \"\"\" → cuerpo → una línea con \"\"\"\n  Imágenes: incluye la ruta de un archivo de imagen directamente en el mensaje (p. ej. describe esto ./shot.png)",
+  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: attach explicitly with @./shot.png (or --attach ./shot.png for one-shot commands)": "Comandos:\n  /help, /?           esta ayuda\n  /clear              reinicia el contexto de la conversación (conserva el system prompt)\n  /system [text]      define el system prompt (sin argumento muestra el valor actual)\n  /model [id]         cambia de modelo durante la sesión (sin argumento = selector con flechas)\n  /models [query]     explora la lista de modelos\n  /save [name]        guarda la sesión actual (por defecto: last)\n  /load [name]        carga una sesión guardada\n  /verbose            activa/desactiva estadísticas de tokens/velocidad\n  /exit, /quit        salir\n  Entrada multilínea: una línea con \"\"\" → cuerpo → una línea con \"\"\"\n  Imágenes: adjunta explícitamente con @./shot.png (o --attach ./shot.png para comandos de una sola ejecución)",
   "Context cleared": "Contexto borrado",
   "Ctrl+C again to quit (or /exit)": "Ctrl+C de nuevo para salir (o /exit)",
   "Current language: {name}": "Idioma actual: {name}",

package/src/locales/ja.mjs

@@ -12,7 +12,7 @@ export const JA = {
   "Account": "アカウント",
   "All models — price and context included": "全モデル — 価格・コンテキスト込み",
   "Anthropic's coding agent — supports subagents": "Anthropic のコーディングエージェント — サブエージェント対応",
-  "Attaching and uploading {n} local image(s) from the prompt.": "プロンプトからローカル画像 {n} 件を添付・アップロード中。",
+  "Attaching and uploading {n} explicit local image(s).": "明示的に選択したローカル画像 {n} 件を添付・アップロード中。",
   "Auto-run:": "自動実行:",
   "Autonomous dev agent": "自律開発エージェント",
   "Multi-provider terminal coding agent": "マルチプロバイダー対応のターミナルコーディングエージェント",
@@ -32,7 +32,7 @@ export const JA = {
   "Choose a number, or search (Enter to cancel): ": "番号を選択、または検索してください (Enter でキャンセル): ",
   "Choose a shell": "シェルを選択",
   "Code": "コード",
-  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: include an image file path right in the message (e.g. describe this ./shot.png)": "コマンド:\n  /help, /?           このヘルプ\n  /clear              会話コンテキストをリセット (システムプロンプトは保持)\n  /system [text]      システムプロンプトを設定 (引数なしで現在値を表示)\n  /model [id]         セッション中にモデルを切り替え (引数なし = 矢印キーで選択)\n  /models [query]     モデル一覧を閲覧\n  /save [name]        現在のセッションを保存 (デフォルト: last)\n  /load [name]        保存したセッションを読み込み\n  /verbose            トークン/速度統計の表示切替\n  /exit, /quit        終了\n  複数行入力: \"\"\" の行 → 本文 → \"\"\" の行\n  画像: メッセージ内に画像ファイルパスを直接含める (例: describe this ./shot.png)",
+  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: attach explicitly with @./shot.png (or --attach ./shot.png for one-shot commands)": "コマンド:\n  /help, /?           このヘルプ\n  /clear              会話コンテキストをリセット (システムプロンプトは保持)\n  /system [text]      システムプロンプトを設定 (引数なしで現在値を表示)\n  /model [id]         セッション中にモデルを切り替え (引数なし = 矢印キーで選択)\n  /models [query]     モデル一覧を閲覧\n  /save [name]        現在のセッションを保存 (デフォルト: last)\n  /load [name]        保存したセッションを読み込み\n  /verbose            トークン/速度統計の表示切替\n  /exit, /quit        終了\n  複数行入力: \"\"\" の行 → 本文 → \"\"\" の行\n  画像: @./shot.png で明示的に添付 (1回実行では --attach ./shot.png)",
   "Context cleared": "コンテキストをクリアしました",
   "Ctrl+C again to quit (or /exit)": "もう一度 Ctrl+C で終了 (または /exit)",
   "Current language: {name}": "現在の言語: {name}",

package/src/locales/ko.mjs

@@ -176,7 +176,7 @@ export const KO = {
   [`Specify a model: pleum run <model> ["prompt"]  (model selection isn't available on non-interactive stdin)`]:
     `모델을 지정하세요: pleum run <model> ["프롬프트"]  (비대화형 stdin에선 모델 선택 불가)`,
   'Usage: pleum run <model> ["prompt"]\nModel list: pleum models': '사용법: pleum run <model> ["프롬프트"]\n모델 목록: pleum models',
-  "Attaching and uploading {n} local image(s) from the prompt.": "프롬프트의 로컬 이미지 {n}개를 첨부해 업로드합니다.",
+  "Attaching and uploading {n} explicit local image(s).": "명시적으로 선택한 로컬 이미지 {n}개를 첨부해 업로드합니다.",
   "Goodbye.": "안녕히 가세요.",
   "Ctrl+C again to quit (or /exit)": "한 번 더 Ctrl+C 하면 종료 (/exit 도 가능)",
   '/help for commands · /exit to quit · wrap multi-line input in """ · Ctrl+C to stop generation':
@@ -192,7 +192,7 @@ export const KO = {
   /verbose            toggle token/speed stats
   /exit, /quit        quit
   Multi-line input: a line with """ → body → a line with """
-  Images: include an image file path right in the message (e.g. describe this ./shot.png)`]: `명령:
+  Images: attach explicitly with @./shot.png (or --attach ./shot.png for one-shot commands)`]: `명령:
   /help, /?           이 도움말
   /clear              대화 컨텍스트 초기화(시스템 프롬프트는 유지)
   /system [텍스트]    시스템 프롬프트 설정(인자 없으면 현재값 표시)
@@ -203,7 +203,7 @@ export const KO = {
   /verbose            토큰·속도 통계 표시 토글
   /exit, /quit        종료
   여러 줄 입력: """ 한 줄 → 본문 → """ 한 줄
-  이미지: 메시지에 이미지 파일 경로를 그대로 포함(예: 이 그림 설명해줘 ./shot.png)`,
+  이미지: @./shot.png로 명시 첨부(1회 실행은 --attach ./shot.png)`,
   "Context cleared": "컨텍스트 초기화",
   "System prompt set": "시스템 프롬프트 설정",
   "Current system prompt:": "현재 시스템 프롬프트:",

package/src/locales/zh.mjs

@@ -12,7 +12,7 @@ export const ZH = {
   "Account": "账户",
   "All models — price and context included": "所有模型 — 含价格与上下文",
   "Anthropic's coding agent — supports subagents": "Anthropic 的编程代理 — 支持子代理",
-  "Attaching and uploading {n} local image(s) from the prompt.": "正在附加并上传提示中的 {n} 张本地图片。",
+  "Attaching and uploading {n} explicit local image(s).": "正在附加并上传明确选择的 {n} 张本地图片。",
   "Auto-run:": "自动运行：",
   "Autonomous dev agent": "自主开发代理",
   "Multi-provider terminal coding agent": "多提供商终端编程代理",
@@ -32,7 +32,7 @@ export const ZH = {
   "Choose a number, or search (Enter to cancel): ": "选择编号，或搜索（Enter 取消）： ",
   "Choose a shell": "选择 shell",
   "Code": "代码",
-  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: include an image file path right in the message (e.g. describe this ./shot.png)": "命令：\n  /help, /?           本帮助\n  /clear              重置对话上下文（保留系统提示词）\n  /system [text]      设置系统提示词（不带参数则显示当前值）\n  /model [id]         会话中途切换模型（不带参数 = 方向键选择器）\n  /models [query]     浏览模型列表\n  /save [name]        保存当前会话（默认：last）\n  /load [name]        加载已保存的会话\n  /verbose            切换 token/速度统计\n  /exit, /quit        退出\n  多行输入：一行 \"\"\" → 正文 → 一行 \"\"\"\n  图片：在消息中直接包含图片文件路径（例如 describe this ./shot.png）",
+  "Commands:\n  /help, /?           this help\n  /clear              reset the conversation context (keeps the system prompt)\n  /system [text]      set the system prompt (no arg shows the current value)\n  /model [id]         switch models mid-session (no arg = arrow-key picker)\n  /models [query]     browse the model list\n  /save [name]        save the current session (default: last)\n  /load [name]        load a saved session\n  /verbose            toggle token/speed stats\n  /exit, /quit        quit\n  Multi-line input: a line with \"\"\" → body → a line with \"\"\"\n  Images: attach explicitly with @./shot.png (or --attach ./shot.png for one-shot commands)": "命令：\n  /help, /?           本帮助\n  /clear              重置对话上下文（保留系统提示词）\n  /system [text]      设置系统提示词（不带参数则显示当前值）\n  /model [id]         会话中途切换模型（不带参数 = 方向键选择器）\n  /models [query]     浏览模型列表\n  /save [name]        保存当前会话（默认：last）\n  /load [name]        加载已保存的会话\n  /verbose            切换 token/速度统计\n  /exit, /quit        退出\n  多行输入：一行 \"\"\" → 正文 → 一行 \"\"\"\n  图片：使用 @./shot.png 显式附加（单次命令使用 --attach ./shot.png）",
   "Context cleared": "上下文已清除",
   "Ctrl+C again to quit (or /exit)": "再次按 Ctrl+C 退出（或 /exit）",
   "Current language: {name}": "当前语言：{name}",

package/src/login.mjs

@@ -3,14 +3,24 @@ import { randomUUID } from "node:crypto";
 import { spawn } from "node:child_process";
 import { ROOT, V1, readConfig, writeConfig, mergeLoginConfig } from "./config.mjs";
 import { t } from "./i18n.mjs";
+import { sanitizeTerminalText } from "./ansi.mjs";
+
+// `start`는 Windows cmd.exe의 built-in이라 shell:true가 필요하다. rundll32는 같은
+// 기본 브라우저 핸드오프를 제공하는 실제 실행 파일이므로 URL을 argv로 안전하게 넘긴다.
+export function browserCommand(url, platform = process.platform) {
+  if (platform === "darwin") return { command: "open", args: [url] };
+  if (platform === "win32") {
+    return { command: "rundll32.exe", args: ["url.dll,FileProtocolHandler", url] };
+  }
+  return { command: "xdg-open", args: [url] };
+}
 
 function openBrowser(url) {
-  const cmd =
-    process.platform === "darwin" ? "open" : process.platform === "win32" ? "start" : "xdg-open";
-  spawn(cmd, [url], {
+  const { command, args } = browserCommand(url);
+  spawn(command, args, {
     stdio: "ignore",
     detached: true,
-    shell: process.platform === "win32",
+    shell: false,
   }).unref();
 }
 
@@ -56,7 +66,7 @@ export async function login() {
     server.listen(0, "127.0.0.1", () => {
       const { port } = server.address();
       const authUrl = `${ROOT}/cli-auth?port=${port}&state=${state}`;
-      console.log(`${t("Log in from your browser. If it doesn't open automatically, open this URL:")}\n  ${authUrl}\n`);
+      console.log(`${t("Log in from your browser. If it doesn't open automatically, open this URL:")}\n  ${sanitizeTerminalText(authUrl)}\n`);
       openBrowser(authUrl);
     });
 
@@ -81,8 +91,8 @@ export async function loginDevice() {
   const { device_code, user_code, verification_url, interval, expires_in } = await res.json();
 
   console.log(
-    `\n  ${t("Code")}: \x1b[1;32m${user_code}\x1b[0m\n` +
-    `  URL: ${verification_url}\n\n` +
+    `\n  ${t("Code")}: \x1b[1;32m${sanitizeTerminalText(user_code)}\x1b[0m\n` +
+    `  URL: ${sanitizeTerminalText(verification_url)}\n\n` +
     `  ${t("Log in at the URL above, then enter the code (within {min} min).", { min: expires_in / 60 })}\n`
   );
 
@@ -98,7 +108,7 @@ export async function loginDevice() {
     if (!poll.ok) {
       const e = await poll.json().catch(() => ({}));
       if (e.detail === "expired_token") { console.error(t("The code has expired.")); process.exit(1); }
-      console.error(t("Error:"), e.detail ?? poll.status);
+      console.error(t("Error:"), sanitizeTerminalText(e.detail ?? poll.status));
       process.exit(1);
     }
     const { key } = await poll.json();

package/src/markdown.mjs

@@ -2,7 +2,7 @@
 // 전략: \n이 도착해 완결된 줄은 스타일 입혀 확정 출력(재렌더 없음), 아직 안 끝난
 // 마지막 줄만 매 push마다 지우고 다시 그린다. 비TTY(파이프)는 완전 패스스루.
 
-import { style, displayWidth, terminalWidth } from "./ansi.mjs";
+import { style, displayWidth, terminalWidth, sanitizeTerminalText } from "./ansi.mjs";
 
 // ```lang 펜스 감지(순수 함수). 열기/닫기 구분은 호출자의 inFence 상태가 한다.
 export function detectFence(line) {
@@ -13,7 +13,7 @@ export function detectFence(line) {
 // 인라인 마크다운: `code` → cyan, **bold**, *italic*. 코드 조각 먼저 치환해
 // 코드 안의 별표가 bold로 오인되는 흔한 케이스를 줄인다.
 export function renderInline(text, styler = style) {
-  return text
+  return sanitizeTerminalText(text)
     .replace(/`([^`]+)`/g, (_, c) => styler.cyan(c))
     .replace(/\*\*([^*]+)\*\*/g, (_, c) => styler.bold(c))
     .replace(/(^|[^*])\*([^*\s][^*]*)\*(?!\*)/g, (_, pre, c) => pre + styler.italic(c));
@@ -21,6 +21,7 @@ export function renderInline(text, styler = style) {
 
 // 완결된 한 줄 렌더(순수 함수). state.inFence는 이 줄 "이전"의 펜스 상태.
 export function renderLine(line, state = { inFence: false }, styler = style) {
+  line = sanitizeTerminalText(line);
   const fence = detectFence(line);
   if (fence) {
     return state.inFence ? styler.gray("└─") : styler.gray("┌─ ") + styler.dim(fence.lang || "code");
@@ -67,6 +68,8 @@ export function createMarkdownStream({
 
   return {
     push(delta) {
+      delta = sanitizeTerminalText(delta);
+      if (!delta) return;
       if (!interactive) return void write(delta);
       buf += delta;
       const lines = buf.split("\n");

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 } from "./ansi.mjs";
+import { style, displayWidth, terminalWidth, hyperlink, sanitizeTerminalText } from "./ansi.mjs";
 import { keyFeeder, select } from "./interactive.mjs";
 import { readConfig, writeConfig } from "./config.mjs";
 import { RECIPES, MANUAL } from "./recipes.mjs";
 import { onPath, launch } from "./launch.mjs";
-import { fetchModels, matchModels, listModels } from "./models.mjs";
+import { fetchModels, matchModels, listModels, formatModelLabel } from "./models.mjs";
 import { run } from "./run.mjs";
 import { login } from "./login.mjs";
 import { top } from "./top.mjs";
@@ -115,12 +115,12 @@ async function pickModelFor(item) {
   try {
     models = await fetchModels();
   } catch (e) {
-    console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: e.message })}`);
+    console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: sanitizeTerminalText(e.message) })}`);
     return undefined;
   }
   const m = await select(models, {
     message: `${item.title} — ${t("select a model")}`,
-    formatItem: (x) => `${x.id}${x.display_name ? `  ${style.dim(x.display_name)}` : ""}`,
+    formatItem: (x) => formatModelLabel(x),
     matches: matchModels,
     limit: 12,
   });

package/src/models.mjs

@@ -1,5 +1,5 @@
 import { V1 } from "./config.mjs";
-import { style, padDisplay, divider } from "./ansi.mjs";
+import { style, padDisplay, divider, sanitizeTerminalText } from "./ansi.mjs";
 import { t } from "./i18n.mjs";
 
 // GET /v1/models 는 공개(인증 불필요). 슬러그는 provider/model 형식.
@@ -15,16 +15,28 @@ export function matchModels(models, query) {
   const q = (query || "").toLowerCase();
   if (!q) return models;
   return models.filter(
-    (m) => m.id.toLowerCase().includes(q) || (m.display_name || "").toLowerCase().includes(q),
+    (m) => String(m.id ?? "").toLowerCase().includes(q) || String(m.display_name ?? "").toLowerCase().includes(q),
   );
 }
 
+// Model registry data is remote input. Keep this boundary before any style/render operation so
+// formatting codes emitted by the CLI itself are the only ANSI sequences written to a terminal.
+export function safeModelText(value, fallback = "") {
+  return sanitizeTerminalText(value ?? fallback);
+}
+
+export function formatModelLabel(model, styler = style) {
+  const id = safeModelText(model?.id);
+  const displayName = safeModelText(model?.display_name);
+  return `${id}${displayName ? `  ${styler.dim(displayName)}` : ""}`;
+}
+
 // 컨텍스트 길이를 사람이 읽기 좋게: 128000→128K, 1000000→1M.
 export function fmtCtx(n) {
   if (!n) return "—";
   if (n >= 1_000_000) return `${+(n / 1_000_000).toFixed(1)}M`;
   if (n >= 1000) return `${Math.round(n / 1000)}K`;
-  return String(n);
+  return safeModelText(n);
 }
 
 // pricing(_krw)는 토큰당 단가 문자열. 1M 토큰당으로 환산해 표시(KRW 우선, 없으면 USD).
@@ -49,9 +61,9 @@ export function formatTable(models, styler = style) {
   const rows = models.map((m) => {
     const [pin, pout] = fmtPrice(m);
     return [
-      m.id,
-      m.provider || "",
-      m.modality || "text",
+      safeModelText(m.id),
+      safeModelText(m.provider || ""),
+      safeModelText(m.modality || "text"),
       fmtCtx(m.context_length || m.context_window),
       pin,
       pout,
@@ -119,16 +131,31 @@ const fmtNum = (n) => (n == null ? "—" : Number(n).toLocaleString("en-US"));
 export function renderModelInfo(m, styler = style) {
   const [pin, pout] = fmtPrice(m);
   const lines = [];
-  lines.push(`${styler.bold(styler.plum(m.id))}${m.display_name ? `  ${styler.dim(m.display_name)}` : ""}`);
+  const id = safeModelText(m.id);
+  const displayName = safeModelText(m.display_name);
+  lines.push(`${styler.bold(styler.plum(id))}${displayName ? `  ${styler.dim(displayName)}` : ""}`);
   lines.push(divider(undefined, styler));
   const facts = [
-    ["maker", m.maker || "—"],
-    ["modality", m.modality || "text"],
+    ["maker", safeModelText(m.maker || "—")],
+    ["modality", safeModelText(m.modality || "text")],
     ["context", fmtCtx(m.context_length || m.context_window)],
     ["max output", fmtCtx(m.max_output_tokens || m.max_output)],
-    [t("price/1M"), `${t("in")} ${pin} · ${t("out")} ${pout} ${styler.dim(t("(incl. {pct}% markup)", { pct: m.effective_markup_percent ?? 0 }))}`],
+    [t("price/1M"), `${t("in")} ${pin} · ${t("out")} ${pout} ${styler.dim(t("(incl. {pct}% markup)", { pct: safeModelText(m.effective_markup_percent, 0) }))}`],
   ];
-  if (m.supports_thinking) facts.push(["thinking", m.thinking_forced ? t("forced (can't be turned off)") : t("supported")]);
+  if (m.supports_thinking) {
+    facts.push([
+      "thinking",
+      m.thinking_forced
+        ? t("forced (can't be turned off)")
+        : t("supported (reasoning_effort: minimal|none|low|medium|high|xhigh|max)"),
+    ]);
+  }
+  if (Array.isArray(m.supported_parameters) && m.supported_parameters.includes("service_tier")) {
+    facts.push(["service_tier", "auto|default|flex|background|priority"]);
+  }
+  if (Array.isArray(m.supported_parameters) && m.supported_parameters.includes("reasoning_mode")) {
+    facts.push(["reasoning_mode", "standard|pro"]);
+  }
   for (const [k, v] of facts) lines.push(`${styler.dim(padDisplay(k, 10))} ${v}`);
 
   const u = m.usage;
@@ -147,19 +174,19 @@ export function renderModelInfo(m, styler = style) {
       const [i1m, o1m] = fmtPrice(p);
       lines.push(
         [
-          styler.plum(padDisplay(p.provider, 12)),
+          styler.plum(padDisplay(safeModelText(p.provider), 12)),
           padDisplay(i1m, 10),
           padDisplay(o1m, 10),
           padDisplay(fmtCtx(p.context_window), 5),
           padDisplay(p.latency_ms != null ? `${Math.round(p.latency_ms)}ms` : "—", 8),
           padDisplay(p.throughput_tps != null ? `${Math.round(p.throughput_tps)}/s` : "—", 6),
-          p.uptime_percent != null ? `${p.uptime_percent}%` : "—",
+          p.uptime_percent != null ? `${safeModelText(p.uptime_percent)}%` : "—",
         ].join("  "),
       );
     }
   }
 
-  const sibs = (m.siblings || []).map((s) => s.id);
+  const sibs = (m.siblings || []).map((s) => safeModelText(s.id));
   if (sibs.length) {
     lines.push("");
     lines.push(`${styler.dim(t("Same provider:"))} ${sibs.join(styler.dim(" · "))}`);

package/src/recipes.mjs

@@ -169,15 +169,33 @@ export const RECIPES = {
 
 // Codex는 Chat Completions가 제거돼 Responses API(wire_api="responses")만 지원.
 // 사용자의 ~/.codex/config.toml을 건드리지 않도록 격리 CODEX_HOME에 이 파일을 쓴다.
+// TOML basic string encoder. 모델명과 base_url은 CLI/env에서 오므로, 따옴표·개행이
+// config의 새 키/테이블로 해석되지 않도록 모든 제어문자를 이스케이프한다.
+export function tomlBasicString(value) {
+  const escapes = {
+    "\b": "\\b",
+    "\t": "\\t",
+    "\n": "\\n",
+    "\f": "\\f",
+    "\r": "\\r",
+    '"': '\\"',
+    "\\": "\\\\",
+  };
+  const escaped = String(value).replace(/[\u0000-\u001f\u007f"\\\\]/g, (ch) =>
+    escapes[ch] ?? `\\u${ch.codePointAt(0).toString(16).padStart(4, "0")}`,
+  );
+  return `"${escaped}"`;
+}
+
 export function codexToml(model) {
   const m = model || "gpt-5.4-mini";
   return `# Generated by pleum CLI. Isolated CODEX_HOME — your ~/.codex is untouched.
-model = "${m}"
+model = ${tomlBasicString(m)}
 model_provider = "pleum"
 
 [model_providers.pleum]
 name = "PleumRouter"
-base_url = "${V1}"
+base_url = ${tomlBasicString(V1)}
 env_key = "PLEUM_API_KEY"
 wire_api = "responses"
 `;

package/src/run.mjs

@@ -1,21 +1,57 @@
 import { createInterface } from "node:readline";
-import { existsSync, readFileSync, writeFileSync, mkdirSync } from "node:fs";
+import { readFileSync, writeFileSync, mkdirSync, lstatSync } from "node:fs";
 import { join } from "node:path";
 import { V1, requireKey, HISTORY_FILE, SESSIONS_DIR } from "./config.mjs";
-import { fetchModels, formatTable, matchModels } from "./models.mjs";
-import { style, divider } from "./ansi.mjs";
+import { fetchModels, formatTable, matchModels, formatModelLabel, safeModelText } from "./models.mjs";
+import { style, divider, sanitizeTerminalText } from "./ansi.mjs";
 import { createMarkdownStream } from "./markdown.mjs";
 import { spinner, select } from "./interactive.mjs";
 import { t } from "./i18n.mjs";
 
-const IMG_EXT = /\.(png|jpe?g|gif|webp|bmp)$/i;
+const PNG_MAGIC = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+const JPEG_MAGIC = Buffer.from([0xff, 0xd8, 0xff]);
+const GIF87A_MAGIC = Buffer.from("GIF87a");
+const GIF89A_MAGIC = Buffer.from("GIF89a");
+const RIFF_MAGIC = Buffer.from("RIFF");
+const WEBP_MAGIC = Buffer.from("WEBP");
+const BMP_MAGIC = Buffer.from("BM");
+const startsWith = (bytes, magic, offset = 0) =>
+  bytes.length >= offset + magic.length && bytes.subarray(offset, offset + magic.length).equals(magic);
 
-// `pleum run [model] ["프롬프트"...] [--verbose] [--skip-cache] [--tag a,b]`
+const IMAGE_FORMATS = {
+  png: { mime: "image/png", matches: (bytes) => startsWith(bytes, PNG_MAGIC) },
+  jpg: { mime: "image/jpeg", matches: (bytes) => startsWith(bytes, JPEG_MAGIC) },
+  jpeg: { mime: "image/jpeg", matches: (bytes) => startsWith(bytes, JPEG_MAGIC) },
+  gif: {
+    mime: "image/gif",
+    matches: (bytes) => startsWith(bytes, GIF87A_MAGIC) || startsWith(bytes, GIF89A_MAGIC),
+  },
+  webp: {
+    mime: "image/webp",
+    matches: (bytes) => startsWith(bytes, RIFF_MAGIC) && startsWith(bytes, WEBP_MAGIC, 8),
+  },
+  bmp: { mime: "image/bmp", matches: (bytes) => startsWith(bytes, BMP_MAGIC) },
+};
+
+export const MAX_IMAGE_ATTACHMENT_BYTES = 5 * 1024 * 1024;
+const MAX_IMAGE_ATTACHMENTS = 4;
+const MAX_TOTAL_IMAGE_ATTACHMENT_BYTES = 10 * 1024 * 1024;
+
+// `pleum run [model] ["프롬프트"...] [--verbose] [--skip-cache] [--tag a,b]
+//   [--attach image] [--reasoning-effort LEVEL] [--service-tier TIER] [--reasoning-mode MODE]`
 // 인자 파싱(순수 함수 — 테스트 대상).
+const REASONING_EFFORTS = new Set(["minimal", "none", "low", "medium", "high", "xhigh", "max"]);
+const SERVICE_TIERS = new Set(["auto", "default", "flex", "background", "priority"]);
+const REASONING_MODES = new Set(["standard", "pro"]);
+
 export function parseRunArgs(rest = []) {
   let verbose = false;
   let skipCache = false;
   let tags = [];
+  let reasoningEffort = null;
+  let serviceTier = null;
+  let reasoningMode = null;
+  const attachments = [];
   const pos = [];
   for (let i = 0; i < rest.length; i++) {
     const a = rest[i];
@@ -27,21 +63,101 @@ export function parseRunArgs(rest = []) {
         tags = v.split(",").map((s) => s.trim()).filter(Boolean).slice(0, 10);
         i++;
       }
+    } else if (a === "--reasoning-effort") {
+      const v = rest[i + 1];
+      if (typeof v === "string" && REASONING_EFFORTS.has(v)) {
+        reasoningEffort = v;
+        i++;
+      }
+    } else if (a === "--service-tier") {
+      const v = rest[i + 1];
+      if (typeof v === "string" && SERVICE_TIERS.has(v)) {
+        serviceTier = v;
+        i++;
+      }
+    } else if (a === "--reasoning-mode") {
+      const v = rest[i + 1];
+      if (typeof v === "string" && REASONING_MODES.has(v)) {
+        reasoningMode = v;
+        i++;
+      }
+    } else if (a === "--attach" || a === "-a") {
+      const path = rest[i + 1];
+      if (typeof path !== "string" || !path) throw new Error("--attach requires an image path");
+      attachments.push(path);
+      i++;
     } else pos.push(a);
   }
-  return { model: pos[0], prompt: pos.slice(1).join(" "), verbose, skipCache, tags };
+  return {
+    model: pos[0],
+    prompt: pos.slice(1).join(" "),
+    verbose,
+    skipCache,
+    tags,
+    reasoningEffort,
+    serviceTier,
+    reasoningMode,
+    attachments,
+  };
+}
+
+function imageFormat(path) {
+  const ext = /\.([a-z0-9]+)$/i.exec(path)?.[1]?.toLowerCase();
+  return ext ? IMAGE_FORMATS[ext] : null;
 }
 
-// 입력에서 존재하는 이미지 파일 경로를 찾아 멀티모달 content로 만든다(없으면 문자열 그대로).
-// exists/read 주입 가능 — 테스트용.
-export function buildContent(text, exists = existsSync, read = readFileSync) {
-  const imgs = text.split(/\s+/).filter((t) => IMG_EXT.test(t) && exists(t));
-  if (!imgs.length) return text;
+function inlineAttachmentPaths(text) {
+  // REPL에서도 안전하게 쓸 수 있도록 토큰 맨 앞 @만 명시 첨부 문법으로 해석한다.
+  // 일반적인 ./photo.png 문구와 @mention은 그대로 텍스트로 남는다.
+  return text
+    .split(/\s+/)
+    .filter((token) => token.startsWith("@") && imageFormat(token.slice(1)))
+    .map((token) => token.slice(1));
+}
+
+function imagePart(path, { lstat, read }) {
+  const format = imageFormat(path);
+  if (!format) throw new Error(`Unsupported image attachment: ${path}`);
+
+  let info;
+  try {
+    info = lstat(path);
+  } catch {
+    throw new Error(`Image attachment not found: ${path}`);
+  }
+  if (info.isSymbolicLink()) throw new Error(`Refusing symlink image attachment: ${path}`);
+  if (!info.isFile()) throw new Error(`Image attachment is not a regular file: ${path}`);
+  if (!Number.isSafeInteger(info.size) || info.size < 1) throw new Error(`Image attachment has an invalid size: ${path}`);
+  if (info.size > MAX_IMAGE_ATTACHMENT_BYTES) {
+    throw new Error(`Image attachment is too large (max ${MAX_IMAGE_ATTACHMENT_BYTES} bytes): ${path}`);
+  }
+
+  const bytes = Buffer.from(read(path));
+  if (bytes.length !== info.size) throw new Error(`Image attachment changed while reading: ${path}`);
+  if (bytes.length > MAX_IMAGE_ATTACHMENT_BYTES) {
+    throw new Error(`Image attachment is too large (max ${MAX_IMAGE_ATTACHMENT_BYTES} bytes): ${path}`);
+  }
+  if (!format.matches(bytes)) throw new Error(`Image attachment does not match its file extension: ${path}`);
+  return { type: "image_url", image_url: { url: `data:${format.mime};base64,${bytes.toString("base64")}` }, size: bytes.length };
+}
+
+// Local files are uploaded only when explicitly selected with --attach or an @path REPL token.
+// lstat/read injection keeps the filesystem checks focused and regression-testable.
+export function buildContent(text, { attachments = [], lstat = lstatSync, read = readFileSync } = {}) {
+  const paths = [...new Set([...attachments, ...inlineAttachmentPaths(text)])];
+  if (!paths.length) return text;
+  if (paths.length > MAX_IMAGE_ATTACHMENTS) {
+    throw new Error(`Too many image attachments (max ${MAX_IMAGE_ATTACHMENTS})`);
+  }
   const parts = [{ type: "text", text }];
-  for (const p of imgs) {
-    const ext = p.match(IMG_EXT)[1].toLowerCase().replace("jpg", "jpeg");
-    const b64 = Buffer.from(read(p)).toString("base64");
-    parts.push({ type: "image_url", image_url: { url: `data:image/${ext};base64,${b64}` } });
+  let totalSize = 0;
+  for (const path of paths) {
+    const { size, ...part } = imagePart(path, { lstat, read });
+    totalSize += size;
+    if (totalSize > MAX_TOTAL_IMAGE_ATTACHMENT_BYTES) {
+      throw new Error(`Total image attachments are too large (max ${MAX_TOTAL_IMAGE_ATTACHMENT_BYTES} bytes)`);
+    }
+    parts.push(part);
   }
   return parts;
 }
@@ -51,17 +167,17 @@ export function formatStats({ cost, usage, ms, tokPerSec, verbose }) {
   const bits = [];
   if (cost != null && cost !== "") bits.push(`₩${fmtCost(cost)}`);
   if (verbose && usage) {
-    bits.push(`↑${usage.prompt_tokens ?? "?"} ↓${usage.completion_tokens ?? "?"} tok`);
+    bits.push(`↑${sanitizeTerminalText(usage.prompt_tokens ?? "?")} ↓${sanitizeTerminalText(usage.completion_tokens ?? "?")} tok`);
     if (tokPerSec) bits.push(`${tokPerSec.toFixed(1)} tok/s`);
   }
-  if (verbose) bits.push(`${ms}ms`);
+  if (verbose) bits.push(`${sanitizeTerminalText(ms)}ms`);
   return bits.length ? `  (${bits.join(" · ")})` : "";
 }
 
 // 비용 표시: 숫자면 소수 2자리까지(후행 0 제거), 문자열은 그대로.
 function fmtCost(v) {
   const n = Number(v);
-  return Number.isFinite(n) ? String(+n.toFixed(2)) : String(v);
+  return Number.isFinite(n) ? String(+n.toFixed(2)) : sanitizeTerminalText(v);
 }
 
 function isAbort(e) {
@@ -70,11 +186,22 @@ function isAbort(e) {
 
 export async function run(rest) {
   const args = Array.isArray(rest) ? rest : rest == null ? [] : [rest];
-  const { verbose, prompt: inlinePrompt, model: argModel, skipCache, tags } = parseRunArgs(args);
+  const {
+    verbose,
+    prompt: inlinePrompt,
+    model: argModel,
+    skipCache,
+    tags,
+    reasoningEffort,
+    serviceTier,
+    reasoningMode,
+    attachments,
+  } = parseRunArgs(args);
   let model = argModel;
   const piped = await readStdin();
-  const oneShot = Boolean(inlinePrompt || piped);
+  const oneShot = Boolean(inlinePrompt || piped || attachments.length);
   const key = requireKey();
+  const callOpts = { tags, skipCache, reasoningEffort, serviceTier, reasoningMode };
 
   if (!model && !oneShot) {
     if (!process.stdin.isTTY) {
@@ -92,26 +219,31 @@ export async function run(rest) {
 
   if (oneShot) {
     const prompt = [inlinePrompt, piped].filter(Boolean).join("\n\n");
-    const content = buildContent(prompt);
-    if (Array.isArray(content)) {
-      const n = content.filter((p) => p.type === "image_url").length;
-      // 프롬프트 내 로컬 이미지 경로를 자동 첨부·업송하므로 조용히 올리지 않고 명시 고지한다(M8).
-      console.error(`ℹ ${t("Attaching and uploading {n} local image(s) from the prompt.", { n })}`);
-    }
-    const messages = [{ role: "user", content }];
     try {
-      await streamChat(key, model, messages, { verbose, quiet: true, tags, skipCache });
+      const content = buildContent(prompt, { attachments });
+      if (Array.isArray(content)) {
+        const n = content.filter((p) => p.type === "image_url").length;
+        console.error(`ℹ ${t("Attaching and uploading {n} explicit local image(s).", { n })}`);
+      }
+      const messages = [{ role: "user", content }];
+      await streamChat(key, model, messages, { verbose, quiet: true, ...callOpts });
     } catch (e) {
-      console.error(`${style.red("✗")} ${e.message}`);
+      console.error(`${style.red("✗")} ${sanitizeTerminalText(e.message)}`);
       process.exit(1);
     }
     process.exit(0);
   }
 
-  await repl(key, model, verbose, { tags, skipCache });
+  await repl(key, model, verbose, callOpts);
 }
 
-async function repl(key, initialModel, initialVerbose, { tags = [], skipCache = false } = {}) {
+async function repl(key, initialModel, initialVerbose, {
+  tags = [],
+  skipCache = false,
+  reasoningEffort = null,
+  serviceTier = null,
+  reasoningMode = null,
+} = {}) {
   let model = initialModel;
   let system = "";
   let verbose = initialVerbose;
@@ -123,17 +255,28 @@ async function repl(key, initialModel, initialVerbose, { tags = [], skipCache =
   let rl;
 
   const send = async (text) => {
-    messages.push({ role: "user", content: buildContent(text) });
-    const req = system ? [{ role: "system", content: system }, ...messages] : messages;
-    const ac = new AbortController();
-    currentAbort = ac;
+    let added = false;
     try {
-      const reply = await streamChat(key, model, req, { verbose, signal: ac.signal, tags, skipCache });
+      const content = buildContent(text);
+      messages.push({ role: "user", content });
+      added = true;
+      const req = system ? [{ role: "system", content: system }, ...messages] : messages;
+      const ac = new AbortController();
+      currentAbort = ac;
+      const reply = await streamChat(key, model, req, {
+        verbose,
+        signal: ac.signal,
+        tags,
+        skipCache,
+        reasoningEffort,
+        serviceTier,
+        reasoningMode,
+      });
       if (reply) messages.push({ role: "assistant", content: reply });
       else messages.pop(); // 응답 없음(즉시 중단 등) — user 턴 되돌림
     } catch (e) {
-      console.error(`\n${style.red("✗")} ${e.message}`);
-      messages.pop(); // 실패한 user 턴 제거
+      console.error(`\n${style.red("✗")} ${sanitizeTerminalText(e.message)}`);
+      if (added) messages.pop(); // 실패한 user 턴 제거
     } finally {
       currentAbort = null;
     }
@@ -251,7 +394,7 @@ async function repl(key, initialModel, initialVerbose, { tags = [], skipCache =
   };
 
   makeRl(loadHistory());
-  console.log(`${style.bold(style.plum("PleumRouter"))} ${style.dim("·")} ${model}`);
+  console.log(`${style.bold(style.plum("PleumRouter"))} ${style.dim("·")} ${safeModelText(model)}`);
   console.log(style.dim(t('/help for commands · /exit to quit · wrap multi-line input in """ · Ctrl+C to stop generation')));
   console.log(divider());
   rl.prompt();
@@ -270,7 +413,7 @@ const SLASH_HELP = `Commands:
   /verbose            toggle token/speed stats
   /exit, /quit        quit
   Multi-line input: a line with """ → body → a line with """
-  Images: include an image file path right in the message (e.g. describe this ./shot.png)`;
+  Images: attach explicitly with @./shot.png (or --attach ./shot.png for one-shot commands)`;
 
 async function handleSlash(text, ctx) {
   const [cmd] = text.slice(1).split(/\s+/);
@@ -295,17 +438,17 @@ async function handleSlash(text, ctx) {
     case "model":
       if (arg) {
         ctx.model = arg;
-        console.log(`${style.green("✓")} ${t("Switched model →")} ${style.plum(arg)}`);
+        console.log(`${style.green("✓")} ${t("Switched model →")} ${style.plum(safeModelText(arg))}`);
       } else if (ctx.pickModel) {
         const picked = await ctx.pickModel();
         if (picked) {
           ctx.model = picked;
-          console.log(`${style.green("✓")} ${t("Switched model →")} ${style.plum(picked)}`);
+          console.log(`${style.green("✓")} ${t("Switched model →")} ${style.plum(safeModelText(picked))}`);
         } else {
-          console.log(`${t("Current model:")} ${style.plum(ctx.model)}`);
+          console.log(`${t("Current model:")} ${style.plum(safeModelText(ctx.model))}`);
         }
       } else {
-        console.log(`${t("Current model:")} ${ctx.model}\n${t("switch: /model <id>  ·  list: /models")}`);
+        console.log(`${t("Current model:")} ${safeModelText(ctx.model)}\n${t("switch: /model <id>  ·  list: /models")}`);
       }
       break;
     case "models":
@@ -314,7 +457,7 @@ async function handleSlash(text, ctx) {
         const rows = matchModels(models, arg);
         console.log(rows.length ? formatTable(rows) : t("No matching models"));
       } catch (e) {
-        console.error(`${style.red("✗")} ${e.message}`);
+        console.error(`${style.red("✗")} ${sanitizeTerminalText(e.message)}`);
       }
       break;
     case "save":
@@ -322,13 +465,13 @@ async function handleSlash(text, ctx) {
         const name = saveSession(arg || "last", ctx.snapshot());
         console.log(`${style.green("✓")} ${t("Session saved:")} ${name}`);
       } catch (e) {
-        console.error(`${style.red("✗")} ${t("Save failed: {error}", { error: e.message })}`);
+        console.error(`${style.red("✗")} ${t("Save failed: {error}", { error: sanitizeTerminalText(e.message) })}`);
       }
       break;
     case "load":
       try {
         ctx.restore(loadSession(arg || "last"));
-        console.log(`${style.green("✓")} ${t("Session loaded (model: {model})", { model: ctx.model })}`);
+        console.log(`${style.green("✓")} ${t("Session loaded (model: {model})", { model: safeModelText(ctx.model) })}`);
       } catch {
         console.error(`${style.red("✗")} ${t("No such session: {name}", { name: arg || "last" })}`);
       }
@@ -344,9 +487,19 @@ async function handleSlash(text, ctx) {
 
 // SSE 스트리밍 채팅. fetchImpl 주입 가능(테스트용 — 네트워크 없이 SSE 파싱 검증).
 // 비용은 마지막 SSE 청크의 cost.krw — 스트리밍 응답엔 X-Cost-Krw 헤더가 없다(백엔드 계약).
-export async function streamChat(key, model, messages, { verbose = false, quiet = false, signal, tags = [], skipCache = false, fetchImpl = fetch } = {}) {
+export async function streamChat(key, model, messages, {
+  verbose = false,
+  quiet = false,
+  signal,
+  tags = [],
+  skipCache = false,
+  reasoningEffort = null,
+  serviceTier = null,
+  reasoningMode = null,
+  fetchImpl = fetch,
+} = {}) {
   const t0 = Date.now();
-  const sp = spinner(t("{model} thinking…", { model }));
+  const sp = spinner(t("{model} thinking…", { model: sanitizeTerminalText(model) }));
   let res;
   try {
     // 게이트웨이 패리티(0095): X-Pleum-Skip-Cache 헤더 + bounded tags 본문.
@@ -354,6 +507,9 @@ export async function streamChat(key, model, messages, { verbose = false, quiet
     if (skipCache) headers["X-Pleum-Skip-Cache"] = "1";
     const bodyObj = { model, messages, stream: true, stream_options: { include_usage: true } };
     if (tags.length > 0) bodyObj.tags = tags.slice(0, 10);
+    if (reasoningEffort) bodyObj.reasoning_effort = reasoningEffort;
+    if (serviceTier) bodyObj.service_tier = serviceTier;
+    if (reasoningMode) bodyObj.reasoning_mode = reasoningMode;
     res = await fetchImpl(`${V1}/chat/completions`, {
       method: "POST",
       headers,
@@ -371,7 +527,7 @@ export async function streamChat(key, model, messages, { verbose = false, quiet
   if (!res.ok) {
     sp.stop();
     const body = await res.text().catch(() => "");
-    throw new Error(`${res.status} ${body.slice(0, 300)}`);
+    throw new Error(`${res.status} ${sanitizeTerminalText(body.slice(0, 300))}`);
   }
 
   const md = createMarkdownStream({ prefix: quiet ? "" : `${style.plum("bot")}${style.dim(">")} ` });
@@ -398,12 +554,14 @@ export async function streamChat(key, model, messages, { verbose = false, quiet
           if (j.cost) cost = j.cost; // 마지막 청크 — {usd, krw, fx_rate, markup_rate}
           const delta = j.choices?.[0]?.delta?.content;
           if (delta) {
+            const safeDelta = sanitizeTerminalText(delta);
+            if (!safeDelta) continue;
             if (!firstAt) {
               firstAt = Date.now();
               sp.stop();
             }
-            md.push(delta);
-            full += delta;
+            md.push(safeDelta);
+            full += safeDelta;
           }
         } catch {
           // 부분 청크 — 무시
@@ -440,14 +598,14 @@ async function pickModel() {
   try {
     models = await fetchModels();
   } catch (e) {
-    console.error(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: e.message })}`);
+    console.error(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: sanitizeTerminalText(e.message) })}`);
     return null;
   }
   if (!models.length) return null;
   if (process.stdin.isTTY && process.stdout.isTTY) {
     return select(models, {
       message: t("Select a model"),
-      formatItem: (m) => `${m.id}${m.display_name ? `  ${style.dim(m.display_name)}` : ""}`,
+      formatItem: (m) => formatModelLabel(m),
       matches: matchModels,
       limit: 12,
     }).then((m) => m?.id ?? null);
@@ -464,7 +622,7 @@ async function pickModelNumbered(models) {
     for (;;) {
       const shown = pool.slice(0, 30);
       shown.forEach((m, i) =>
-        console.log(`  ${String(i + 1).padStart(2)}. ${m.id}${m.display_name ? `  (${m.display_name})` : ""}`),
+        console.log(`  ${String(i + 1).padStart(2)}. ${formatModelLabel(m, { dim: (s) => `(${s})` })}`),
       );
       if (pool.length > shown.length) console.log(`  ${t("… {n} more — narrow with a search term", { n: pool.length - shown.length })}`);
       const ans = (await ask(`\n${t("Choose a number, or search (Enter to cancel): ")}`)).trim();

package/src/settings.mjs

@@ -6,12 +6,12 @@
 // 즉시 참조하면 순환의 평가 순서에 따라 깨질 수 있다).
 
 import { readFileSync } from "node:fs";
-import { style } from "./ansi.mjs";
+import { style, sanitizeTerminalText } from "./ansi.mjs";
 import { select, confirm } from "./interactive.mjs";
 import { readConfig, writeConfig, clearCredentials } from "./config.mjs";
 import { getLocale, LANGS, t } from "./i18n.mjs";
 import { RECIPES } from "./recipes.mjs";
-import { fetchModels, matchModels } from "./models.mjs";
+import { fetchModels, matchModels, formatModelLabel } from "./models.mjs";
 import { login } from "./login.mjs";
 import { upgrade } from "./upgrade.mjs";
 import { completion } from "./completion.mjs";
@@ -82,12 +82,12 @@ async function pickDefaultModelFor(item) {
   try {
     models = await fetchModels();
   } catch (e) {
-    console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: e.message })}`);
+    console.log(`${style.red("✗")} ${t("Failed to fetch model list: {error}", { error: sanitizeTerminalText(e.message) })}`);
     return undefined;
   }
   const picked = await select([CLEAR_OPTION, ...models], {
     message: `${item.title} — ${t("select a model")}`,
-    formatItem: (x) => (x.isClear ? style.dim(t("Use default (clear)")) : `${x.id}${x.display_name ? `  ${style.dim(x.display_name)}` : ""}`),
+    formatItem: (x) => (x.isClear ? style.dim(t("Use default (clear)")) : formatModelLabel(x)),
     matches: matchModels, // 쿼리 입력 시 CLEAR_OPTION은 자연히 필터링됨(id/display_name 불일치) — 별도 처리 불요
     limit: 12,
   });