llm-api-gateway-cli 1.0.4 → 1.0.6

package/cli-agent.js

@@ -27,7 +27,7 @@ import { spawn } from 'node:child_process';
 import { existsSync, readFileSync, writeFileSync } from 'node:fs';
 
 import { loadDotEnv, parseArgs, DEFAULT_BASE_URL, maskKey } from './lib/common.js';
-import { resolveConfig, numArg, fail, KEY_HINT } from './lib/config.js';
+import { resolveConfig, numArg, fail, keyHint } from './lib/config.js';
 import { checkWorkDir, setBashEnabled, availableTools } from './lib/tools.js';
 import { createSession, restoreSession, runTurn, resumePendingTurn, countTurns } from './lib/runner.js';
 import { createSessionStore, resolveSessionDir, SESSION_TTL_MS } from './lib/sessionstore.js';
@@ -40,6 +40,8 @@ import { secretSourceText } from './lib/secrets.js';
 import { AGENT_LIMITS } from './lib/agent.js';
 import { MEMORY_FILES, readMemory } from './lib/memory.js';
 import { refreshMcpTools, mcpStatusText } from './lib/mcp.js';
+// CLI 侧语言（只覆盖命令行输出；Web 界面与手册页仍是中文）。默认 zh，见 lib/i18n.js
+import { pick, setLang, resolveLang } from './lib/i18n.js';
 
 const SCRIPT_DIR = path.dirname(fileURLToPath(import.meta.url));
 loadDotEnv(SCRIPT_DIR);
@@ -47,7 +49,7 @@ loadDotEnv(SCRIPT_DIR);
 const DEFAULT_MODEL = process.env.GATEWAY_MODEL || 'qwen3:8b';
 const HISTORY_LIMIT = 200;
 
-const HELP = `用法：gateway-agent                    启动器：先问「对话 / 任务」，再问「命令行 / 网页」
+const HELP_ZH = `用法：gateway-agent                    启动器：先问「对话 / 任务」，再问「命令行 / 网页」
        gateway-agent [选项] [提示词]
        gateway-agent setup [--base-url <地址>] [--key sk-xxx]      首次配置引导
        gateway-agent config <list|get|set|unset|path> [键] [值]
@@ -62,6 +64,7 @@ const HELP = `用法：gateway-agent                    启动器：先问「对
   http://127.0.0.1:3100/        聊天页
   http://127.0.0.1:3100/task    任务页（「设置」里可改网关地址 / 模型 / 密钥）
   http://127.0.0.1:3100/manual  操作手册页（安装、首次配置、指令一览的完整版）
+  http://127.0.0.1:3100/sessions 命令行会话的只读存档（就是这里聊过的对话，--resume 接着聊）
 
 子命令：
   setup                    首次配置引导：问一次「网关地址」与「网关密钥」并写进本机配置
@@ -93,11 +96,73 @@ const HELP = `用法：gateway-agent                    启动器：先问「对
       --yes                自动批准所有**文件写入**（危险：模型可直接改文件，慎用）。MCP 调用不需要它
       --verbose            打印工具结果正文（默认只显示一行摘要）
       --no-color           关闭彩色输出（也遵循 NO_COLOR 环境变量）
+      --lang <zh|en>       输出语言（默认 zh，也可用环境变量 LLM_GATEWAY_LANG）
   -h, --help               显示帮助
 
 单个任务最多 ${AGENT_LIMITS.MAX_STEPS} 轮工具调用（--max-steps 可调）。写入被拒后模型会改为向你说明思路。
 `;
 
+const HELP_EN = `Usage: gateway-agent                    launcher: asks "chat / task", then "terminal / web"
+       gateway-agent [options] [prompt]
+       gateway-agent setup [--base-url <url>] [--key sk-xxx]      first-time setup
+       gateway-agent config <list|get|set|unset|path> [key] [value]
+       (alias of the same entry point: llm-api-gateway-cli; from the sources: node cli-agent.js …)
+
+Mode 6 · native agent: brings its own tool loop, reads and writes files inside the working
+directory and asks before every write. No Claude Code required (no @anthropic-ai/claude-code).
+
+Related commands (local Web UI — this command never opens a browser for you):
+  gateway-task [--port 3100]   start the local Web UI and task page (gateway-web is an alias);
+                               it starts without an API key; set the gateway URL and key on the
+                               task page under "Settings"
+  http://127.0.0.1:3100/        chat
+  http://127.0.0.1:3100/task    task mode (gateway URL / model / key are editable in "Settings")
+  http://127.0.0.1:3100/manual  operation manual (install, first-time setup, full command list)
+  http://127.0.0.1:3100/sessions read-only archive of CLI sessions (conversations started here)
+
+Subcommands:
+  setup                    first-time setup: asks once for the gateway URL and the key, then writes
+                           them to your machine (URL -> config.json, key -> credentials.json).
+                           With --base-url / --key it writes without asking; a fresh machine enters
+                           this step automatically on the first run.
+  config list|get|set|unset|path
+                           view or change persisted settings (stored on disk, survives cache
+                           clearing). config list marks where each effective value comes from:
+                           command line, environment, config file or built-in default — that column
+                           tells you what is overriding the value you just set.
+
+Options:
+  -p, --prompt <text>      one-shot task (falls back to the positional argument / stdin)
+  -i, --interactive        multi-turn interactive mode (REPL)
+  -C, --cwd <dir>          working directory; tools are sandboxed inside it (default: current dir)
+  -m, --model <model>      model name (default ${DEFAULT_MODEL})
+  -k, --key <sk-key>       gateway API key (or env SK/GATEWAY_KEY/OPENAI_API_KEY, or .env)
+      --base-url <url>     gateway base URL (default ${DEFAULT_BASE_URL})
+  -s, --system <text>      extra system prompt
+      --temperature <v>    sampling temperature
+      --max-tokens <v>     max generated tokens (a thinking model given too few may only emit reasoning)
+      --max-steps <n>      max model turns per task (default ${AGENT_LIMITS.MAX_STEPS})
+      --continue           continue the most recent session (context + pending approval)
+      --resume <session-id> resume the given session (UUID)
+      --session-dir <dir>  where sessions are written (default: .llm-api-gateway-cli in your home dir)
+      --no-session         do not persist the session (memory only, lost on exit)
+      --no-memory          do not read AGENTS.md / CLAUDE.md project memory
+      --allow-bash         enable the bash tool (off by default; each command still needs approval)
+      --output-format <v>  output format: text (default) or stream-json (one JSON event per line)
+      --yes                auto-approve all **file writes** (dangerous: the model edits files directly). MCP calls never need it
+      --verbose            print tool output in full (default: a one-line summary)
+      --no-color           disable colors (also honors the NO_COLOR env var)
+      --lang <zh|en>       output language (default zh; env LLM_GATEWAY_LANG)
+  -h, --help               show this help
+
+One task runs at most ${AGENT_LIMITS.MAX_STEPS} model turns (--max-steps). After a rejected write the model explains its plan instead.
+`;
+
+/** 大段帮助取词必须在**调用时**（模块顶层调 pick 会早于 setLang，永远拿到中文） */
+function mainHelp() {
+  return pick(HELP_ZH, HELP_EN);
+}
+
 const VALUE_FLAGS = {
   '--prompt': 'prompt',
   '-p': 'prompt',
@@ -116,6 +181,7 @@ const VALUE_FLAGS = {
   '--resume': 'resume',
   '--session-dir': 'session-dir',
   '--output-format': 'output-format',
+  '--lang': 'lang',
 };
 const BOOL_FLAGS = {
   '--interactive': 'interactive',
@@ -131,6 +197,14 @@ const BOOL_FLAGS = {
   '-h': 'help',
 };
 
+/** 从原始 argv 里取 --lang：要在 parseArgs 之前用（setup / config 子命令也会输出） */
+function langFromArgv(argv) {
+  const i = argv.indexOf('--lang');
+  if (i >= 0 && argv[i + 1] && !String(argv[i + 1]).startsWith('-')) return argv[i + 1];
+  const eq = argv.find((a) => String(a).startsWith('--lang='));
+  return eq ? String(eq).slice('--lang='.length) : '';
+}
+
 function readStdin() {
   return new Promise((resolve) => {
     if (process.stdin.isTTY) return resolve('');
@@ -233,17 +307,27 @@ function createRenderer({ color, verbose, json }) {
 // ---------- 写入批准 ----------
 
 function renderApproval(preview, color) {
-  const top = '┌─ 待批准写入 ─────────────────────────────';
+  const top = pick('┌─ 待批准写入 ─────────────────────────────', '┌─ pending write ──────────────────────────');
   const mid = '├──────────────────────────────────────────';
   const bot = '└──────────────────────────────────────────';
-  const kind = preview.shell ? '执行命令' : preview.patch ? '增量修改' : preview.exists ? '覆盖已有文件' : '新建文件';
-  const out = [top, `│ ${preview.path}  (${kind})`, `│ ${preview.lines} 行 / ${preview.bytes} 字节`];
+  const kind = preview.shell
+    ? pick('执行命令', 'run command')
+    : preview.patch
+      ? pick('增量修改', 'patch')
+      : preview.exists
+        ? pick('覆盖已有文件', 'overwrite existing file')
+        : pick('新建文件', 'create file');
+  const out = [
+    top,
+    `│ ${preview.path}  (${kind})`,
+    pick(`│ ${preview.lines} 行 / ${preview.bytes} 字节`, `│ ${preview.lines} lines / ${preview.bytes} bytes`),
+  ];
 
   // 覆盖已有文件时给出行级 diff：只看新内容根本判断不出改了什么
   const canDiff = !preview.shell && preview.exists && typeof preview.old === 'string' && preview.old !== preview.content;
   if (canDiff) {
     const { lines, removed, added } = diffLines(preview.old, preview.content);
-    out.push(mid, `│ 改动：-${removed} 行 / +${added} 行`);
+    out.push(mid, pick(`│ 改动：-${removed} 行 / +${added} 行`, `│ changes: -${removed} lines / +${added} lines`));
     for (const line of lines) {
       const painted = line.startsWith('-') ? color.red(line) : line.startsWith('+') ? color.green(line) : line;
       out.push(`│ ${painted}`);
@@ -253,7 +337,9 @@ function renderApproval(preview, color) {
     const contentLines = String(preview.content ?? '').split('\n');
     const shown = contentLines.slice(0, 40);
     for (const line of shown) out.push(`│ ${line}`);
-    if (contentLines.length > shown.length) out.push(`│ …（预览截断，共 ${contentLines.length} 行）`);
+    if (contentLines.length > shown.length) {
+      out.push(pick(`│ …（预览截断，共 ${contentLines.length} 行）`, `│ ... (preview truncated, ${contentLines.length} lines in total)`));
+    }
   }
   if (preview.note) out.push(`│ ${color.yellow(preview.note)}`);
   out.push(bot);
@@ -269,7 +355,7 @@ function question(rl, prompt) {
  * 不引 TUI 库：readline 就够，多行只是多问几次。
  */
 async function readInput(rl, color) {
-  const first = await question(rl, color.cyan('你> '));
+  const first = await question(rl, color.cyan(pick('你> ', 'you> ')));
   if (!first.trimEnd().endsWith('\\')) return first;
   const parts = [first.trimEnd().slice(0, -1)];
   for (;;) {
@@ -292,14 +378,22 @@ function makeApproval({ autoApprove, rl, color }) {
   return async (pending, preview) => {
     renderApproval(preview, color);
     if (autoApprove) {
-      process.stderr.write(color.yellow('  [--yes] 已自动批准这次写入\n'));
+      process.stderr.write(color.yellow(pick('  [--yes] 已自动批准这次写入\n', '  [--yes] write auto-approved\n')));
       return true;
     }
     if (!rl || !process.stdin.isTTY) {
-      process.stderr.write(color.red('  非交互环境，默认拒绝文件写入（确需自动批准请加 --yes；MCP 调用不受此限）\n'));
+      process.stderr.write(color.red(pick('  非交互环境，默认拒绝文件写入（确需自动批准请加 --yes；MCP 调用不受此限）\n', '  non-interactive environment: file writes are denied by default (pass --yes to auto-approve; MCP calls are not affected)\n')));
       return false;
     }
-    const ans = await question(rl, color.cyan(`批准 ${preview.shell ? '执行该命令' : `写入 ${preview.path}`} ？(y/N) `));
+    const ans = await question(
+      rl,
+      color.cyan(
+        pick(
+          `批准 ${preview.shell ? '执行该命令' : `写入 ${preview.path}`} ？(y/N) `,
+          `Approve ${preview.shell ? 'running this command' : `writing ${preview.path}`}? (y/N) `,
+        ),
+      ),
+    );
     return /^(y|yes)$/i.test(ans.trim());
   };
 }
@@ -385,6 +479,9 @@ async function main() {
   // 「配置还没弄好」恰恰是用户最可能先跑它的场景；交给 parseArgs 还会把
   // `set system --foo` 这类取值当成选项吃掉。
   const raw = process.argv.slice(2);
+  // 语言要在**任何输出之前**定下来：setup / config 子命令走的是 parseArgs 之前的路径，
+  // 所以这里从原始 argv 里取一次 --lang（不重复 parseArgs，免得未知参数被报两遍）。
+  setLang(resolveLang({ lang: langFromArgv(raw) }, process.env));
   // `gateway-agent setup`：首次配置引导。和 `config` 一样放在 parseArgs 之前 ——
   // 「还没配」正是最可能先跑它的场景，而它自己收 --base-url / --key（给了就不问）。
   if (raw[0] === 'setup') {
@@ -413,7 +510,7 @@ async function main() {
   const args = parseArgs(process.argv.slice(2), VALUE_FLAGS, BOOL_FLAGS);
 
   if (args.help) {
-    console.log(HELP);
+    console.log(mainHelp());
     return;
   }
 
@@ -440,15 +537,27 @@ async function main() {
   const maxSteps = numArg(args['max-steps']);
   if (!key) {
     process.stderr.write(
-      `[gateway] 还没配密钥 · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n` +
+      pick(
+        `[gateway] 还没配密钥 · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n`,
+        `[gateway] no API key configured · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n`,
+      ) +
         (srcInfo.allDefault && !srcInfo.configExists
-          ? '          这台机器还没有任何配置（没找到 config.json / credentials.json / .env / 相关环境变量）：\n' +
-            '          上面那两个值是**内置默认**，不是「已经配好了」——配一次之后就一直用它。\n'
+          ? pick(
+              '          这台机器还没有任何配置（没找到 config.json / credentials.json / .env / 相关环境变量）：\n' +
+                '          上面那两个值是**内置默认**，不是「已经配好了」——配一次之后就一直用它。\n',
+              '          This machine has no configuration yet (no config.json / credentials.json / .env / env vars):\n' +
+                '          the two values above are **built-in defaults**, not something you configured — set them once and they stick.\n',
+            )
           : ''),
     );
     // 没密钥时最实用的出口是 Web UI：它没有密钥也能起（KEY_HINT 第①条没说那个页面怎么起起来）
-    process.stderr.write('          Web UI（没密钥也能起）：gateway-task  →  http://127.0.0.1:3100/task\n');
-    process.stderr.write(`[错误] ${KEY_HINT}\n`);
+    process.stderr.write(
+      pick(
+        '          Web UI（没密钥也能起）：gateway-task  →  http://127.0.0.1:3100/task\n',
+        '          Web UI (starts without a key): gateway-task  ->  http://127.0.0.1:3100/task\n',
+      ),
+    );
+    process.stderr.write(`${pick('[错误]', '[error]')} ${keyHint()}\n`);
     process.exit(1);
   }
 
@@ -474,7 +583,7 @@ async function main() {
       },
     });
     if (!r.ok) {
-      process.stderr.write(`[错误] ${r.error}\n`);
+      process.stderr.write(pick(`[错误] ${r.error}\n`, `[error] ${r.error}\n`));
       process.exitCode = r.code || 1;
       return;
     }
@@ -495,7 +604,7 @@ async function main() {
       sessionNotice = resolved.notice;
       store = createSessionStore(sessionDir);
     } catch (e) {
-      sessionWarning = `会话落盘不可用（${e.message}），本次会话不会保存`;
+      sessionWarning = pick(`会话落盘不可用（${e.message}），本次会话不会保存`, `Session persistence unavailable (${e.message}); this session will not be saved`);
     }
   }
 
@@ -504,23 +613,23 @@ async function main() {
   if (args.resume !== undefined) {
     const id = typeof args.resume === 'string' ? args.resume.trim() : '';
     if (!store) {
-      process.stderr.write('[错误] --resume 与 --no-session 不能同时用\n');
+      process.stderr.write(pick('[错误] --resume 与 --no-session 不能同时用\n', '[error] --resume and --no-session cannot be used together\n'));
       process.exit(1);
     }
     if (!id) {
-      process.stderr.write(`[错误] --resume 需要一个会话 id（UUID）。会话目录：${sessionDir}\n`);
+      process.stderr.write(pick(`[错误] --resume 需要一个会话 id（UUID）。会话目录：${sessionDir}\n`, `[error] --resume needs a session id (UUID). Session directory: ${sessionDir}\n`));
       process.exit(1);
     }
     restored = store.read(id);
     if (!restored) {
-      process.stderr.write(`[错误] 找不到会话 ${id}（id 需为 UUID，且未被 TTL 清理）。会话目录：${sessionDir}\n`);
+      process.stderr.write(pick(`[错误] 找不到会话 ${id}（id 需为 UUID，且未被 TTL 清理）。会话目录：${sessionDir}\n`, `[error] session ${id} not found (id must be a UUID and not expired by TTL). Session directory: ${sessionDir}\n`));
       process.exit(1);
     }
   } else if (args.continue) {
     restored = store ? store.latest(SESSION_TTL_MS) : null;
     if (!restored) {
       const plain = makeColor(false);
-      process.stderr.write(`${plain.yellow('!')} --continue 没有找到可恢复的会话，将新建一个\n`);
+      process.stderr.write(pick(`${plain.yellow('!')} --continue 没有找到可恢复的会话，将新建一个\n`, `${plain.yellow('!')} --continue found no session to resume; starting a new one\n`));
     }
   }
 
@@ -529,7 +638,7 @@ async function main() {
     const want = typeof args.cwd === 'string' ? args.cwd : restored?.workingDir || process.cwd();
     workDir = checkWorkDir(want);
   } catch (e) {
-    process.stderr.write(`[错误] 工作目录无效：${e.message}\n`);
+    process.stderr.write(pick(`[错误] 工作目录无效：${e.message}\n`, `[error] invalid working directory: ${e.message}\n`));
     process.exit(1);
   }
 
@@ -549,7 +658,7 @@ async function main() {
   const color = makeColor(!args['no-color'] && !process.env.NO_COLOR && Boolean(process.stdout.isTTY));
   const jsonOut = args['output-format'] === 'stream-json';
   if (args['output-format'] && !jsonOut && args['output-format'] !== 'text') {
-    process.stderr.write('[错误] --output-format 只支持 text / stream-json\n');
+    process.stderr.write(pick('[错误] --output-format 只支持 text / stream-json\n', '[error] --output-format only supports text / stream-json\n'));
     process.exit(1);
   }
   const renderer = createRenderer({ color, verbose: Boolean(args.verbose), json: jsonOut });
@@ -559,7 +668,7 @@ async function main() {
     try {
       store.write(session);
     } catch (e) {
-      process.stderr.write(color.yellow(`! 会话落盘失败：${e.message}\n`));
+      process.stderr.write(color.yellow(pick(`! 会话落盘失败：${e.message}\n`, `! could not save the session: ${e.message}\n`)));
     }
   };
 
@@ -573,7 +682,7 @@ async function main() {
   }
 
   if (!interactive && !prompt) {
-    console.log(HELP);
+    console.log(mainHelp());
     return;
   }
 
@@ -594,7 +703,7 @@ async function main() {
   let controller = null;
   const onSigint = () => {
     if (controller) {
-      process.stderr.write(`\n${color.yellow('!')} 已中断当前任务\n`);
+      process.stderr.write(pick(`\n${color.yellow('!')} 已中断当前任务\n`, `\n${color.yellow('!')} current task interrupted\n`));
       controller.abort();
       controller = null;
     } else if (rl) {
@@ -622,25 +731,46 @@ async function main() {
   // 注意后半句的措辞：`GATEWAY_MCP=off` 时工具是网关自己注入的，说「没有 MCP」会误导，
   // 只能保证「CLI 不携带」。
   const mcpLine = mcpInfo?.error
-    ? `${mcpStatusText()} · ${mcpInfo.attemptedAt
-        ? `已请求 GET ${baseUrl}/v1/mcp/tools`
-        : '按上述原因未向网关发出请求'}（CLI 本次不携带 MCP 工具）`
+    ? `${mcpStatusText()} · ${
+        mcpInfo.attemptedAt
+          ? pick(`已请求 GET ${baseUrl}/v1/mcp/tools`, `requested GET ${baseUrl}/v1/mcp/tools`)
+          : pick('按上述原因未向网关发出请求', 'no request sent to the gateway (see above)')
+      }${pick('（CLI 本次不携带 MCP 工具）', ' (this CLI run carries no MCP tools)')}`
     : mcpStatusText();
   process.stderr.write(
-    `[gateway] 原生 Agent · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n` +
+    pick(
+      `[gateway] 原生 Agent · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n`,
+      `[gateway] native agent · base=${baseUrl}${srcOf('baseUrl')} model=${model}${srcOf('model')}\n`,
+    ) +
       // 密钥这一行上轮只做在 Web 横幅里，CLI 的横幅反倒不说密钥从哪来 —— 而「我没配过密钥，
       // 它怎么跑起来了」正是最需要一句解释的情形（多半是当前目录的 .env 给的）
-      `          密钥 ${maskKey(key)}（${keySrcOf}）\n` +
-      `          工作目录 ${workDir}\n` +
-      `          工具 ${toolNames}${args['allow-bash'] ? '  [bash 已开启]' : ''}\n` +
-      `          Web UI gateway-task → http://127.0.0.1:3100/task（本命令不会替你打开浏览器）\n` +
+      pick(`          密钥 ${maskKey(key)}（${keySrcOf}）\n`, `          key ${maskKey(key)} (${keySrcOf})\n`) +
+      pick(`          工作目录 ${workDir}\n`, `          working dir ${workDir}\n`) +
+      pick(
+        `          工具 ${toolNames}${args['allow-bash'] ? '  [bash 已开启]' : ''}\n`,
+        `          tools ${toolNames}${args['allow-bash'] ? '  [bash enabled]' : ''}\n`,
+      ) +
+      pick(
+        '          Web UI gateway-task → http://127.0.0.1:3100/task（本命令不会替你打开浏览器）\n',
+        '          Web UI: gateway-task -> http://127.0.0.1:3100/task (this command never opens a browser)\n',
+      ) +
       `          ${mcpLine}\n` +
-      `          轮次上限 ${session.maxSteps || AGENT_LIMITS.MAX_STEPS} 轮\n` +
-      (resumedFrom ? `          已恢复会话 ${resumedFrom}（${countTurns(session)} 条消息）\n` : '') +
-      (store ? `          会话目录 ${sessionDir}\n` : '') +
+      pick(
+        `          轮次上限 ${session.maxSteps || AGENT_LIMITS.MAX_STEPS} 轮\n`,
+        `          max steps ${session.maxSteps || AGENT_LIMITS.MAX_STEPS}\n`,
+      ) +
+      (resumedFrom
+        ? pick(`          已恢复会话 ${resumedFrom}（${countTurns(session)} 条消息）\n`, `          resumed session ${resumedFrom} (${countTurns(session)} messages)\n`)
+        : '') +
+      (store ? pick(`          会话目录 ${sessionDir}\n`, `          session dir ${sessionDir}\n`) : '') +
       (sessionNotice ? `${color.yellow('!')} ${sessionNotice}\n` : '') +
       (sessionWarning ? `${color.yellow('!')} ${sessionWarning}\n` : '') +
-      (memory ? `          已注入项目记忆（${MEMORY_FILES.join(' / ')}）\n` : ''),
+      (memory
+        ? pick(
+            `          已注入项目记忆（${MEMORY_FILES.join(' / ')}）\n`,
+            `          project memory injected (${MEMORY_FILES.join(' / ')})\n`,
+          )
+        : ''),
   );
 
   const runOnce = async (text) => {
@@ -653,17 +783,17 @@ async function main() {
     }
     persist();
     const u = session.usage;
-    process.stderr.write(color.dim(`[usage] 输入 ${u.prompt} / 输出 ${u.completion} / 合计 ${u.prompt + u.completion} tokens\n`));
+    process.stderr.write(color.dim(pick(`[usage] 输入 ${u.prompt} / 输出 ${u.completion} / 合计 ${u.prompt + u.completion} tokens\n`, `[usage] prompt ${u.prompt} / completion ${u.completion} / total ${u.prompt + u.completion} tokens\n`)));
     if (jsonOut) process.stdout.write(`${JSON.stringify({ type: 'done', sessionId: session.id, usage: { ...u } })}\n`);
   };
 
   // 会话里有待批准的写入（进程退出前留下的）→ 先问批不批再接着跑
   if (session.pending) {
-    process.stderr.write(color.yellow(`! 恢复了一个待批准的写入：${session.pending.name}\n`));
+    process.stderr.write(color.yellow(pick(`! 恢复了一个待批准的写入：${session.pending.name}\n`, `! recovered a pending write: ${session.pending.name}\n`)));
     controller = new AbortController();
     try {
       const outcome = await resumePendingTurn({ cfg, session, onEvent: renderer.handle, askApproval, persist, signal: controller.signal });
-      if (outcome) process.stderr.write(color.dim(`[恢复完成] 状态 ${outcome.status}\n`));
+      if (outcome) process.stderr.write(color.dim(pick(`[恢复完成] 状态 ${outcome.status}\n`, `[resume done] status ${outcome.status}\n`)));
     } catch (e) {
       fail(e, { exit: false });
     } finally {
@@ -673,7 +803,7 @@ async function main() {
   }
 
   if (interactive) {
-    process.stderr.write(color.dim('交互模式：直接输入任务；/help 看命令，/exit 退出。\n'));
+    process.stderr.write(color.dim(pick('交互模式：直接输入任务；/help 看命令，/exit 退出。\n', 'Interactive mode: type a task; /help lists commands, /exit quits.\n')));
     const ctx = {
       get session() {
         return session;
@@ -701,6 +831,10 @@ async function main() {
         })),
       // /config 用：与 `gateway-agent config` 完全同一份实现与同一个文件
       configOpts: () => ({ env: process.env, file: settingsFilePath() }),
+      // 需要用户当场回答一个问题的命令（现在只有 /mcp add 的追问与 /mcp rm 的确认）。
+      // **非 TTY 返回 null**：那种环境里没人能回答，命令层据此改成「要求写成参数」，
+      // 而不是把 null 当默认值用——删服务器这种事猜错一次就没了。
+      ask: (prompt) => (rl && process.stdin.isTTY ? question(rl, prompt) : Promise.resolve(null)),
     };
 
     for (;;) {
@@ -717,7 +851,7 @@ async function main() {
       const cmd = parseSlash(line);
       if (cmd) {
         if (cmd.error) {
-          process.stderr.write(color.red(`  ${cmd.error}\n  用法：${cmd.cmd.usage}\n`));
+          process.stderr.write(color.red(pick(`  ${cmd.error}\n  用法：${cmd.cmd.usage}\n`, `  ${cmd.error}\n  usage: ${cmd.cmd.usage}\n`)));
           continue;
         }
         // await 是必须的：命令可以是异步的（如 /mcp 要出网重拉清单）。
@@ -745,9 +879,12 @@ async function main() {
         const near = suggest(head.slice(1));
         if (near.length) {
           process.stderr.write(color.dim(
-            `  命令名不完整：${head}\n  候选：${near.map((c) => c.usage.split(' ')[0]).join('　')}\n`));
+            pick(
+              `  命令名不完整：${head}\n  候选：${near.map((c) => c.usage.split(' ')[0]).join('　')}\n`,
+              `  incomplete command: ${head}\n  candidates: ${near.map((c) => c.usage.split(' ')[0]).join('  ')}\n`,
+            )));
         } else {
-          process.stderr.write(color.dim(`未知命令 ${head}（/help 看可用命令）。\n`));
+          process.stderr.write(color.dim(pick(`未知命令 ${head}（/help 看可用命令）。\n`, `unknown command ${head} (see /help for the list).\n`)));
         }
         continue;
       }
@@ -762,7 +899,7 @@ async function main() {
       rl.close();
       saveHistory(sessionDir, rl);
     }
-    process.stderr.write('再见。\n');
+    process.stderr.write(pick('再见。\n', 'Bye.\n'));
     return;
   }
 
@@ -778,6 +915,6 @@ async function main() {
 }
 
 main().catch((e) => {
-  process.stderr.write(`[错误] ${e?.message || e}\n`);
+  process.stderr.write(pick(`[错误] ${e?.message || e}\n`, `[error] ${e?.message || e}\n`));
   process.exit(1);
 });

package/cli-anthropic.js

@@ -27,7 +27,9 @@ import readline from 'node:readline';
 import { randomUUID } from 'node:crypto';
 
 import { loadDotEnv, parseArgs, sessionBaggage, DEFAULT_BASE_URL } from './lib/common.js';
-import { resolveConfig, fail, KEY_HINT } from './lib/config.js';
+import { resolveConfig, fail, keyHint } from './lib/config.js';
+// CLI 侧语言（只覆盖命令行输出；Web 界面与手册页仍是中文）。默认 zh，见 lib/i18n.js
+import { pick, setLang, resolveLang, LANG_FLAG } from './lib/i18n.js';
 
 const SCRIPT_DIR = path.dirname(fileURLToPath(import.meta.url));
 loadDotEnv(SCRIPT_DIR); // 先加载 .env，让 GATEWAY_MODEL 等环境变量参与默认值计算
@@ -35,7 +37,7 @@ loadDotEnv(SCRIPT_DIR); // 先加载 .env，让 GATEWAY_MODEL 等环境变量参
 const DEFAULT_MODEL = process.env.GATEWAY_MODEL || 'qwen3:8b';
 const DEFAULT_MAX_TOKENS = 4096;
 
-const HELP = `用法：node cli-anthropic.js [选项] [提示词]
+const HELP_ZH = `用法：node cli-anthropic.js [选项] [提示词]
 
 选项：
   -p, --prompt <文本>      单轮提问内容（省略则读位置参数 / 标准输入）
@@ -48,9 +50,32 @@ const HELP = `用法：node cli-anthropic.js [选项] [提示词]
       --no-stream          关闭流式输出（默认开启）
       --json               输出完整原始响应（非流式）
   -i, --interactive        进入多轮交互（REPL）
+      --lang <zh|en>       输出语言（默认 zh，也可用环境变量 LLM_GATEWAY_LANG）
   -h, --help               显示帮助
 `;
 
+const HELP_EN = `Usage: node cli-anthropic.js [options] [prompt]
+
+Options:
+  -p, --prompt <text>      one-shot prompt (falls back to the positional argument / stdin)
+  -s, --system <text>      system prompt
+  -m, --model <model>      model name (default ${DEFAULT_MODEL})
+  -k, --key <sk-key>       gateway API key (or env SK/GATEWAY_KEY/ANTHROPIC_API_KEY, or .env)
+      --base-url <url>     gateway base URL (default ${DEFAULT_BASE_URL})
+      --max-tokens <v>     max generated tokens (required by Anthropic, default ${DEFAULT_MAX_TOKENS})
+      --temperature <v>    sampling temperature
+      --no-stream          disable streaming output (on by default)
+      --json               print the full raw response (non-streaming)
+  -i, --interactive        multi-turn interactive mode (REPL)
+      --lang <zh|en>       output language (default zh; env LLM_GATEWAY_LANG)
+  -h, --help               show this help
+`;
+
+/** 大段帮助取词必须在**调用时**（模块顶层调 pick 会早于 setLang，永远拿到中文） */
+function help() {
+  return pick(HELP_ZH, HELP_EN);
+}
+
 // 值型选项：会消费下一个参数作为值；布尔型选项：不消费后续参数
 const VALUE_FLAGS = {
   '--prompt': 'prompt', '-p': 'prompt',
@@ -60,6 +85,7 @@ const VALUE_FLAGS = {
   '--base-url': 'base-url',
   '--temperature': 'temperature',
   '--max-tokens': 'max-tokens',
+  ...LANG_FLAG, // --lang <zh|en>：与另外三个入口同一份定义（lib/i18n.js）
 };
 const BOOL_FLAGS = {
   '--no-stream': 'no-stream',
@@ -79,13 +105,23 @@ function readStdin() {
   });
 }
 
+/** usage 行：Anthropic 的字段名与 OpenAI 不同，这里统一成一句按当前语言取词 */
+function usageLine(input, output) {
+  return pick(
+    `[usage] 输入 ${input ?? '?'} / 输出 ${output ?? '?'} tokens`,
+    `[usage] input ${input ?? '?'} / output ${output ?? '?'} tokens`,
+  );
+}
+
 // ---------- 主流程 ----------
 
 async function main() {
   const args = parseArgs(process.argv.slice(2), VALUE_FLAGS, BOOL_FLAGS);
+  // 语言要在**任何输出之前**定下来：--help / 报错 / 横幅都按它取词
+  setLang(resolveLang(args, process.env));
 
   if (args.help) {
-    console.log(HELP);
+    console.log(help());
     return;
   }
 
@@ -95,14 +131,19 @@ async function main() {
   });
 
   if (!key) {
-    console.error(`[错误] ${KEY_HINT}`);
+    console.error(`${pick('[错误]', '[error]')} ${keyHint()}`);
     process.exit(1);
   }
 
   // 一次进程运行 = 一次对话。用 W3C baggage 告诉网关，好让它把同一会话的请求聚成一条。
   const sessionId = randomUUID();
   const baggage = sessionBaggage(sessionId);
-  if (baggage) process.stderr.write(`[gateway] 会话 ${sessionId}（W3C baggage: session.id）\n`);
+  if (baggage) {
+    process.stderr.write(pick(
+      `[gateway] 会话 ${sessionId}（W3C baggage: session.id）\n`,
+      `[gateway] session ${sessionId} (W3C baggage: session.id)\n`,
+    ));
+  }
 
   const client = new Anthropic({
     baseURL: baseUrl,
@@ -120,7 +161,7 @@ async function main() {
   if (!prompt && args._.length) prompt = args._.join(' ');
   if (!prompt) prompt = await readStdin();
   if (!prompt) {
-    console.log(HELP);
+    console.log(help());
     return;
   }
 
@@ -166,7 +207,7 @@ async function chat(client, model, prompt, { system, stream, maxTokens, temperat
       process.stdout.write('\n');
       // 跨方言（Anthropic 入口 → OpenAI 上游）流式时网关不逐块回传 usage，仅在确有输出 token 时才展示
       if (outputTokens !== null) {
-        process.stderr.write(`[usage] 输入 ${inputTokens ?? '?'} / 输出 ${outputTokens} tokens\n`);
+        process.stderr.write(`${usageLine(inputTokens, outputTokens)}\n`);
       }
     } else {
       const resp = await client.messages.create(req);
@@ -174,12 +215,12 @@ async function chat(client, model, prompt, { system, stream, maxTokens, temperat
         console.log(JSON.stringify(resp, null, 2));
       } else {
         const text = (resp.content || [])
-          .map((b) => (b.type === 'text' ? b.text : b.type === 'tool_use' ? `[工具调用 ${b.name}]` : ''))
+          .map((b) => (b.type === 'text' ? b.text : b.type === 'tool_use' ? pick(`[工具调用 ${b.name}]`, `[tool call ${b.name}]`) : ''))
           .join('');
         console.log(text);
       }
       const u = resp.usage;
-      if (u) process.stderr.write(`[usage] 输入 ${u.input_tokens ?? '?'} / 输出 ${u.output_tokens ?? '?'} tokens\n`);
+      if (u) process.stderr.write(`${usageLine(u.input_tokens, u.output_tokens)}\n`);
     }
   } catch (e) {
     fail(e, { exit: exitOnError });
@@ -187,10 +228,13 @@ async function chat(client, model, prompt, { system, stream, maxTokens, temperat
 }
 
 async function repl(client, model, opts) {
-  console.error(`[gateway] 交互模式 model=${model}，输入文本提问，\`exit\` 退出`);
+  console.error(pick(
+    `[gateway] 交互模式 model=${model}，输入文本提问，\`exit\` 退出`,
+    `[gateway] interactive mode model=${model}; type a prompt, \`exit\` to quit`,
+  ));
   const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
   const messages = [];
-  const ask = () => new Promise((resolve) => rl.question('你> ', resolve));
+  const ask = () => new Promise((resolve) => rl.question(pick('你> ', 'you> '), resolve));
   for (;;) {
     let line;
     try {
@@ -203,7 +247,7 @@ async function repl(client, model, opts) {
     messages.push({ role: 'user', content: line });
 
     // 复用底层流式，逐字打印并把助手回复累积进上下文
-    process.stdout.write('助手> ');
+    process.stdout.write(pick('助手> ', 'assistant> '));
     const req = { model, max_tokens: opts.maxTokens, messages, stream: true, temperature: opts.temperature };
     try {
       let out = '';
@@ -221,7 +265,7 @@ async function repl(client, model, opts) {
       }
       process.stdout.write('\n');
       if (outputTokens !== null) {
-        process.stderr.write(`[usage] 输入 ${inputTokens ?? '?'} / 输出 ${outputTokens} tokens\n`);
+        process.stderr.write(`${usageLine(inputTokens, outputTokens)}\n`);
       }
       messages.push({ role: 'assistant', content: out });
     } catch (e) {