@aisa-one/cli 0.3.0 → 0.5.0

package/dist/commands/flow.js

@@ -0,0 +1,697 @@
+/**
+ * What `aisa connect` asks, in the order it asks it — as data.
+ *
+ * The page and the terminal are two renderers of this file, not two programs
+ * that happen to agree. Before it existed they were the latter: every sentence
+ * on the T2 page lived inside a template string (82% of that file is markup),
+ * so the terminal carried a hand-written paraphrase of the same choices. Two
+ * copies of the same wording drift, and nothing catches it — there is no test
+ * that can compare a sentence in HTML against a sentence in a log line.
+ *
+ * So copy lives here, beside the structure it belongs to. A step that the page
+ * renders as a card and the terminal renders as a numbered menu is still one
+ * step, with one question and one set of options.
+ *
+ * ── Language ──────────────────────────────────────────────────────────────
+ * Every user-facing string is a `Text`: `{ en, zh }`. This is a product
+ * surface, not the machine-readable one — `aisa manifest`, MCP tool
+ * descriptions and OpenAPI summaries stay English-only, per the workspace
+ * convention, because those are read by agents everywhere. What a person sees
+ * while choosing is a different audience.
+ *
+ * The chosen language is stored by the CLI (see config.ts), not in the
+ * browser: the page is served by this process, and a language kept only in
+ * localStorage would leave the terminal in English while the page is in
+ * Chinese — the exact split this file exists to prevent.
+ *
+ * ── Scope ─────────────────────────────────────────────────────────────────
+ * Step 2 (Your agent) is here now. The remaining steps still render from
+ * connect-t2.ts and are being moved one at a time, each behind the byte-exact
+ * page snapshots in tests/connect-snapshot.test.ts.
+ */
+export const LANGS = ["en", "zh"];
+export function t(text, lang) {
+    return text[lang] ?? text.en;
+}
+/**
+ * Resolve the language for this run: explicit flag, then what was chosen and
+ * remembered, then the OS, then English.
+ *
+ * The OS check is deliberately loose — `zh_CN.UTF-8`, `zh-Hans`, `zh_TW` all
+ * mean a reader who would rather not be handed English.
+ */
+export function resolveLang(flag, stored, env = process.env) {
+    const known = (v) => LANGS.includes(v) ? v : undefined;
+    if (known(flag))
+        return known(flag);
+    if (known(stored))
+        return known(stored);
+    const locale = env.LC_ALL || env.LC_MESSAGES || env.LANG || "";
+    if (/^zh/i.test(locale))
+        return "zh";
+    return "en";
+}
+// ── step 2: your agent ──────────────────────────────────────────────────────
+export const STEP_AGENT = {
+    id: "agent",
+    n: 2,
+    title: { en: "Your agent", zh: "你的 agent" },
+    sub: { en: "Pick the tool to connect", zh: "选择要连接的工具" },
+    question: {
+        en: "Which agent should AIsa plug into?",
+        zh: "AIsa 要接入哪个 agent?",
+    },
+    lede: {
+        en: "One agent per run, so a problem is always easy to place. Detected tools are ready to connect; a missing one can be installed right here through its official installer.",
+        zh: "一次只连一个 agent,出问题时永远知道该看哪里。已检测到的可以直接连接;没装的可以在这里用它官方的安装方式装上。",
+    },
+};
+/**
+ * Fixed card order, whatever is or is not installed: the terminal agents
+ * first, then the editors, then the web targets. Order is part of the flow,
+ * not of one renderer — a menu that renumbers itself between the page and the
+ * terminal would make "pick 2" mean two different things.
+ */
+export const AGENT_ORDER = [
+    "claude-code",
+    "codex",
+    "opencode",
+    "vscode",
+    "cursor",
+    "claude-desktop",
+    "claude-ai",
+];
+export function agentRank(id) {
+    const i = AGENT_ORDER.indexOf(id);
+    return i === -1 ? 99 : i;
+}
+/** Badges on an agent card. The terminal renders the same three states. */
+export const AGENT_BADGE = {
+    detected: { en: "✓ detected", zh: "✓ 已检测到" },
+    absent: { en: "not installed", zh: "未安装" },
+    web: { en: "web · no install", zh: "网页版 · 无需安装" },
+    notFound: { en: "not found", zh: "未找到" },
+    soon: { en: "soon", zh: "即将支持" },
+};
+export const AGENT_SIDE_TITLES = {
+    how: { en: "How it connects", zh: "它是怎么连上的" },
+    have: { en: "Already have it set up?", zh: "已经配置过了怎么办?" },
+};
+/** Offer to install a missing agent, with the exact command that will run. */
+/** Fill {name} placeholders. Keeps punctuation inside the translated string. */
+export function fill(text, lang, vars) {
+    return Object.entries(vars).reduce((s, [k, v]) => s.split(`{${k}}`).join(String(v)), t(text, lang));
+}
+export function installOffer(command, lang) {
+    return lang === "zh"
+        ? `安装<b>并</b>连接 — <code>${command}</code>`
+        : `Install <b>and</b> connect it — <code>${command}</code>`;
+}
+/**
+ * What connecting each agent actually does, in its own terms.
+ *
+ * These name the real mechanism — `claude mcp add`, a provider block in
+ * config.toml — because the question a reader has at this step is "what are
+ * you about to do to my machine", and the honest answer is the command.
+ */
+export const AGENT_NOTES = {
+    "claude-code": {
+        en: "Servers are added with <code>claude mcp add</code> in user scope — the official mechanism, reversible with <code>claude mcp remove</code>. Models, if you choose to, are set through Claude Code's own settings file.",
+        zh: "通过 <code>claude mcp add</code> 写入 user scope —— 这是官方机制,用 <code>claude mcp remove</code> 就能撤销。模型(如果你选择切换)走 Claude Code 自己的设置文件。",
+    },
+    codex: {
+        en: "Servers are added with <code>codex mcp add</code>, Codex's own command. Models go in an <code>aisa</code> provider inside <code>~/.codex/config.toml</code> — only keys we wrote are ever touched.",
+        zh: "通过 Codex 自己的 <code>codex mcp add</code> 写入。模型写成 <code>~/.codex/config.toml</code> 里的一个 <code>aisa</code> provider —— 只动我们自己写的那些键。",
+    },
+    opencode: {
+        en: "Servers are added with <code>opencode mcp add</code>. Models become an extra <code>aisa</code> provider in <code>opencode.json</code>; pick <code>aisa/…</code> from its model list.",
+        zh: "通过 <code>opencode mcp add</code> 写入。模型会成为 <code>opencode.json</code> 里额外的一个 <code>aisa</code> provider,在它的模型列表里选 <code>aisa/…</code> 即可。",
+    },
+    vscode: {
+        en: "Nothing is done by hand. The AIsa servers go into VS Code's <code>mcp.json</code> (with your key, or VS Code signs in to them itself), and a small <b>AIsa extension</b> is installed that asks VS Code to store your key and add the models as a <b>Custom Endpoint</b> group named AIsa — beside Copilot's own, which stay. Inline completions keep using Copilot.",
+        zh: "全程无需手动操作。AIsa 的 server 写进 VS Code 的 <code>mcp.json</code>(带上你的 key,或者由 VS Code 自己去登录),同时装一个很小的 <b>AIsa 扩展</b>,由它请 VS Code 保管你的 key 并把模型添加为名为 AIsa 的 <b>Custom Endpoint</b> 分组 —— 与 Copilot 自己的并存,不动它。行内补全仍然用 Copilot。",
+    },
+    cursor: {
+        en: "One click per server: each becomes an <b>Add to Cursor</b> link on the last step. Cursor opens, shows you the exact entry, and writes it into its own MCP settings when you confirm. Models stay as they are — Cursor picks its own.",
+        zh: "每个 server 一次点击:最后一步会给出 <b>Add to Cursor</b> 链接。Cursor 会打开、把确切的配置内容给你看,你确认后由它自己写进 MCP 设置。模型保持原样 —— Cursor 用它自己的。",
+    },
+    "claude-desktop": {
+        en: "Claude Desktop's config file only takes MCP servers over stdio, so each AIsa server is written in as a small <code>mcp-remote</code> bridge carrying your key (plus the free <code>aisa-docs</code> server). The bridges live and die with the app — no background service, no open ports. Models cannot be changed: Claude Desktop runs Anthropic's own. Restart the app and the servers appear under its tools menu.",
+        zh: "Claude Desktop 的配置文件只接受 stdio 形式的 MCP server,所以每个 AIsa server 都写成一个带你 key 的 <code>mcp-remote</code> 小桥接(外加免费的 <code>aisa-docs</code>)。桥接随 app 启停 —— 没有后台服务,不开端口。模型无法更改:Claude Desktop 跑的是 Anthropic 自家模型。重启 app 后就能在工具菜单里看到这些 server。",
+    },
+    "claude-ai": {
+        en: "Nothing is installed. claude.ai takes remote MCP servers as <b>Connectors</b>: on the last step you get one URL per server with a Copy button, and add each under <b>Settings → Connectors → Add custom connector</b>, then press Connect — claude.ai runs the AIsa sign-in itself. Models cannot be changed: claude.ai runs Anthropic's own.",
+        zh: "不安装任何东西。claude.ai 以 <b>Connectors</b> 的形式接受远程 MCP server:最后一步会给出每个 server 的 URL 和复制按钮,在 <b>Settings → Connectors → Add custom connector</b> 里逐个添加,然后点 Connect —— claude.ai 会自己跑 AIsa 的登录流程。模型无法更改:claude.ai 跑的是 Anthropic 自家模型。",
+    },
+};
+/**
+ * The answer to "will this break what I already have", per agent.
+ *
+ * Asked on this step rather than later because it is what decides whether
+ * someone continues at all.
+ */
+export const AGENT_HAVE_NOTES = {
+    "claude-code": {
+        en: "Nothing of yours is replaced. Your <code>claude</code> keeps its login, models and settings; the next step lets you add AIsa <b>beside</b> it as a separate <b><code>claude-aisa</code></b> command, or switch — your call.",
+        zh: "你原有的东西一样都不会被替换。你的 <code>claude</code> 保留登录、模型和设置;下一步你可以选择把 AIsa 装在<b>旁边</b>(一个独立的 <b><code>claude-aisa</code></b> 命令),也可以直接切换 —— 你说了算。",
+    },
+    codex: {
+        en: "Nothing of yours is replaced. Your <code>codex</code> keeps its login and config; the next step lets you add AIsa <b>beside</b> it as an <code>aisa</code> profile and a <b><code>codex-aisa</code></b> command, or switch — your call.",
+        zh: "你原有的东西一样都不会被替换。你的 <code>codex</code> 保留登录和配置;下一步你可以选择把 AIsa 装在<b>旁边</b>(一个 <code>aisa</code> profile 加一个 <b><code>codex-aisa</code></b> 命令),也可以直接切换 —— 你说了算。",
+    },
+    opencode: {
+        en: "Nothing of yours is replaced. Your default model stays; the next step can add AIsa as an extra <code>aisa/…</code> provider you pick from the model list, or switch — your call.",
+        zh: "你原有的东西一样都不会被替换。默认模型保持不变;下一步可以把 AIsa 添加为额外的 <code>aisa/…</code> provider(在模型列表里选),也可以直接切换 —— 你说了算。",
+    },
+    "claude-desktop": {
+        en: "Nothing of yours is replaced. Anything already in <code>claude_desktop_config.json</code> stays; AIsa's entries are added beside it under <code>aisa-*</code> names and removing them is deleting those entries.",
+        zh: "你原有的东西一样都不会被替换。<code>claude_desktop_config.json</code> 里已有的内容保持原样;AIsa 的条目以 <code>aisa-*</code> 命名加在旁边,要移除就是删掉这些条目。",
+    },
+    "claude-ai": {
+        en: "Nothing of yours is touched — the connectors live in your claude.ai account, beside any you already have, and can be removed there in one click.",
+        zh: "完全不动你原有的东西 —— connector 存在你的 claude.ai 账户里,与你已有的并存,在那里一键就能移除。",
+    },
+    vscode: {
+        en: "Nothing of yours is replaced. Other MCP servers and model groups in those two files stay; only the <code>aisa-*</code> entries and the AIsa group are ours, and removing them is deleting those entries.",
+        zh: "你原有的东西一样都不会被替换。那两个文件里其他的 MCP server 和模型分组保持原样;只有 <code>aisa-*</code> 条目和 AIsa 那个分组是我们写的,要移除就是删掉它们。",
+    },
+    cursor: {
+        en: "Nothing of yours is replaced. Cursor adds each AIsa server beside your existing MCP entries under <code>aisa-*</code> names, and you approve every one inside Cursor first.",
+        zh: "你原有的东西一样都不会被替换。Cursor 会把每个 AIsa server 以 <code>aisa-*</code> 命名加在你已有的 MCP 条目旁边,而且每一个都要你在 Cursor 里先确认。",
+    },
+};
+// ── the rail: every step's name ─────────────────────────────────────────────
+/**
+ * The six steps, and one glyph each.
+ *
+ * `icon` is for the terminal only. The page has a numbered rail down its left
+ * edge — position, colour and a tick already say which step you are on, and a
+ * sixth signal would be decoration. A terminal has none of that: the headers
+ * scroll past in one column of text, and after two of them "3/6" and "4/6"
+ * are the same shape. The glyph is what the eye finds when scrolling back.
+ *
+ * One per step and never repeated, so it identifies rather than decorates.
+ * Deliberately not the marks the journal already uses for outcomes
+ * (✅ ⚠️ ❌ ℹ️) — a step heading is not a result.
+ */
+export const STEP_TITLES = [
+    { n: 1, icon: "👋", title: { en: "Welcome", zh: "开始" },
+        sub: { en: "What you are about to get", zh: "你将得到什么" } },
+    { n: 2, icon: "🤖", title: STEP_AGENT.title, sub: STEP_AGENT.sub },
+    { n: 3, icon: "🧠", title: { en: "Models", zh: "模型" },
+        sub: { en: "What it runs on", zh: "它用什么模型跑" } },
+    { n: 4, icon: "⚡", title: { en: "Capabilities", zh: "能力" },
+        sub: { en: "Live data for your agent", zh: "给 agent 的实时数据" } },
+    { n: 5, icon: "📦", title: { en: "Install", zh: "安装" },
+        sub: { en: "Sign in and wire it up", zh: "登录并接线" } },
+    { n: 6, icon: "🚀", title: { en: "Done", zh: "完成" },
+        sub: { en: "Try it now", zh: "现在就试试" } },
+];
+// ── step 1: welcome ─────────────────────────────────────────────────────────
+export const STEP_WELCOME = {
+    h1: {
+        en: "One connection. Every major model, live data and skills — inside the agent you already use.",
+        zh: "一次连接,把主流大模型、实时数据和技能,装进你已经在用的 agent 里。",
+    },
+    lede: {
+        en: "AIsa is the capability layer for AI agents: one account, one key, and your coding agent can reach the best models and the real world. Setting it up takes about a minute and nothing here is irreversible.",
+        zh: "AIsa 是 AI agent 的能力层:一个账户、一把 key,让你的编程 agent 既能用上最好的模型,也能接触真实世界的数据。整个过程大约一分钟,而且没有一步是不可撤销的。",
+    },
+    tiles: [
+        {
+            h3: { en: "All the well-known LLM models, one key, lower prices",
+                zh: "所有知名大模型,一把 key,更低的价格" },
+            p: { en: "Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok — all behind one endpoint, well below going direct. Switch between them with a single setting, no re-configuring.",
+                zh: "Claude、GPT、Gemini、DeepSeek、Kimi、GLM、Qwen、Grok —— 全部在同一个接入点后面,价格明显低于直接购买。改一个设置就能切换,不用重新配置。" },
+        },
+        {
+            h3: { en: "Rich, advanced tools your agent can act on",
+                zh: "丰富、先进,而且 agent 真能调用的工具" },
+            p: { en: "Market analysis and expansion research, finance data, social media signals, B2B prospecting — a wide network of officially licensed commercial APIs, integrated and refined by AIsa into MCP tools your agent can use out of the box.",
+                zh: "市场分析与拓展调研、金融数据、社交媒体信号、B2B 获客 —— 背后是一大批正式授权的商业 API,由 AIsa 集成打磨成开箱即用的 MCP 工具。" },
+        },
+        {
+            h3: { en: "Agent skills, ready to run", zh: "开箱即用的 agent 技能" },
+            p: { en: "A public catalogue of skills that teach your agent how to use each capability well — install with <code>aisa skills</code>, no prompt engineering needed.",
+                zh: "一个公开的技能目录,教你的 agent 怎么把每项能力用好 —— 用 <code>aisa skills</code> 安装,不需要自己调提示词。" },
+        },
+        {
+            h3: { en: "MCP the friendly way", zh: "让 MCP 变简单" },
+            p: { en: "One sign-in in your browser, zero keys to paste, every entry written through your agent's own official command. Remove it all just as easily.",
+                zh: "浏览器里登录一次,不用粘贴任何 key,每一条配置都通过 agent 自己的官方命令写入。要移除同样简单。" },
+        },
+    ],
+};
+// ── step 3: models ──────────────────────────────────────────────────────────
+export const STEP_MODELS = {
+    h1: { en: "Every major model, one account — and switching is a one-liner",
+        zh: "所有主流模型,一个账户 —— 切换只要改一行" },
+    lede: {
+        en: "Whatever you decide below, this is what sits behind your AIsa key. No separate accounts, no separate billing, no re-wiring when a better model ships next month.",
+        zh: "无论你在下面怎么选,这些都在你的 AIsa key 后面。不用开多个账户、不用分开付费,下个月出了更好的模型也不用重新接线。",
+    },
+    callout: {
+        en: "<b>Switching is the whole point.</b> One key, one endpoint: change the model name and you are on a different lab's best model — Claude today, DeepSeek for the cheap batch job tonight, GPT-5.5 tomorrow. No new sign-ups, no new config, no juggling keys. ",
+        zh: "<b>能随时切换才是重点。</b>一把 key、一个接入点:改个模型名字,你就用上了另一家的旗舰模型 —— 今天 Claude,晚上跑批量任务换便宜的 DeepSeek,明天 GPT-5.5。不用重新注册、不用改配置、不用管理一堆 key。 ",
+    },
+    rerun: {
+        en: "Change the default any time — just run <code>aisa connect</code> again.",
+        zh: "默认模型随时可改 —— 再跑一次 <b><code>aisa connect</code></b> 就行。",
+    },
+    h2Prefix: { en: "How should ", zh: "要让 " },
+    h2Suffix: { en: " use them?", zh: " 怎么用这些模型?" },
+    yourAgent: { en: "your agent", zh: "你的 agent" },
+    recommended: { en: "recommended", zh: "推荐" },
+    freshSwitch: {
+        name: { en: "Run it on AIsa models", zh: "就用 AIsa 的模型跑" },
+        brief: { en: "A fresh install has no model backend yet. This writes the agent's own provider settings so it starts on {model} through AIsa — reversible.",
+            zh: "全新安装还没有模型后端。这会写入 agent 自己的 provider 设置,让它一上来就通过 AIsa 使用 {model} —— 可撤销。" },
+    },
+    notNow: {
+        name: { en: "Not now", zh: "暂时不用" },
+        briefFresh: { en: "Install it without a model. It will not answer a prompt until you configure a provider by hand.",
+            zh: "只安装,不配模型。在你手动配好 provider 之前,它无法回答任何问题。" },
+        briefDetected: { en: "Leave models exactly as they are; only the MCP tools are added.",
+            zh: "模型保持原样,只添加 MCP 工具。" },
+    },
+    backup: {
+        name: { en: "Add AIsa beside it", zh: "把 AIsa 装在旁边" },
+    },
+    switchIt: {
+        name: { en: "Switch it to AIsa", zh: "切换到 AIsa" },
+        brief: { en: "Points this agent's model traffic at AIsa ({model}). Writes the agent's own provider settings and nothing else — reversible.",
+            zh: "把这个 agent 的模型流量指向 AIsa({model})。只写 agent 自己的 provider 设置,别的都不动 —— 可撤销。" },
+    },
+    warn: {
+        head: { en: "⚠︎ Installing without a model backend", zh: "⚠︎ 安装后将没有模型后端" },
+        body: { en: "A fresh install <b>cannot answer a single prompt</b> until you configure a provider by hand. Turn on AIsa models and it leaves here ready to work.",
+            zh: "全新安装在你手动配好 provider 之前<b>一个问题都答不了</b>。打开 AIsa 模型,它离开这里时就能直接干活。" },
+        fix: { en: "Use AIsa models →", zh: "使用 AIsa 模型 →" },
+    },
+};
+/** Why the Models step has nothing to offer a file-configured client. */
+export const FILE_MODEL_NOTE = {
+    "claude-desktop": {
+        en: "<b>Claude Desktop runs on Anthropic's own models</b>, and AIsa cannot stand in for them — so nothing changes here. You still get every model above through the coding agents on the same AIsa account; Claude Desktop gets the MCP tools.",
+        zh: "<b>Claude Desktop 跑的是 Anthropic 自家的模型</b>,AIsa 无法替代它们 —— 所以这一步不做任何改动。上面那些模型你仍然可以在同一个 AIsa 账户下,通过编程 agent 使用;Claude Desktop 这边拿到的是 MCP 工具。",
+    },
+    "claude-ai": {
+        en: "<b>claude.ai runs on Anthropic's own models</b>, and AIsa cannot stand in for them — so nothing changes here. You still get every model above through the coding agents (Claude Code, Codex, opencode) on the same AIsa account; claude.ai gets the MCP connectors.",
+        zh: "<b>claude.ai 跑的是 Anthropic 自家的模型</b>,AIsa 无法替代它们 —— 所以这一步不做任何改动。上面那些模型你仍然可以在同一个 AIsa 账户下,通过编程 agent(Claude Code、Codex、opencode)使用;claude.ai 这边拿到的是 MCP connector。",
+    },
+    cursor: {
+        en: "<b>Cursor picks its models inside the app</b>, so nothing changes here — only the MCP servers are added.",
+        zh: "<b>Cursor 在它自己的应用里选模型</b>,所以这一步不做改动 —— 只添加 MCP server。",
+    },
+};
+/** Backup-mode consent copy, per client — the same contract T1 shows. */
+export const BACKUP_COPY = {
+    "claude-code": {
+        en: "Installs one small command, <b><code>claude-aisa</code></b>, next to your other tools. Your original <b><code>claude</code></b> keeps its login, models and settings <b>exactly as they are</b>.<br>You can use the newly added <b><code>claude-aisa</code></b> whenever you want AIsa's models at lower prices. Delete that one file to remove it.",
+        zh: "在你其他工具旁边装一个很小的命令 <b><code>claude-aisa</code></b>。你原来的 <b><code>claude</code></b> 的登录、模型和设置<b>原封不动</b>。<br>想用 AIsa 的低价模型时,就用新加的 <b><code>claude-aisa</code></b>。删掉那一个文件就等于移除。",
+    },
+    codex: {
+        en: "Adds an <b>aisa profile</b> inside Codex's own config and a <b><code>codex-aisa</code></b> command. Your default Codex is <b>untouched</b>.<br>You can use the newly added <b><code>codex-aisa</code></b> (or <code>codex --profile aisa</code>) whenever you want AIsa's models; it applies to that session only.",
+        zh: "在 Codex 自己的配置里加一个 <b>aisa profile</b>,再加一个 <b><code>codex-aisa</code></b> 命令。你默认的 Codex <b>完全不动</b>。<br>想用 AIsa 的模型时,就用新加的 <b><code>codex-aisa</code></b>(或 <code>codex --profile aisa</code>);它只对那一次会话生效。",
+    },
+    vscode: {
+        en: "Adds a <b>Custom Endpoint</b> group named <b>AIsa</b> to VS Code's chat models — Claude, GPT, DeepSeek, Kimi, GLM, Qwen — beside Copilot's own. Your Copilot setup is <b>untouched</b>.<br>A small <b>AIsa extension</b> is installed so VS Code stores your key itself — nothing to paste. Then pick any AIsa model from the chat model picker; inline completions stay on Copilot.",
+        zh: "在 VS Code 的聊天模型里加一个名为 <b>AIsa</b> 的 <b>Custom Endpoint</b> 分组 —— Claude、GPT、DeepSeek、Kimi、GLM、Qwen —— 与 Copilot 自己的并存。你的 Copilot 配置<b>完全不动</b>。<br>会装一个很小的 <b>AIsa 扩展</b>,由 VS Code 自己保管你的 key —— 不用粘贴任何东西。之后在聊天模型选择器里挑任意 AIsa 模型即可;行内补全仍然走 Copilot。",
+    },
+    opencode: {
+        en: "Adds AIsa as an <b>extra provider</b> in opencode's config. Your default model is <b>untouched</b>.<br>You can pick the newly added <code>aisa/…</code> models from opencode's model list whenever you want them.",
+        zh: "在 opencode 的配置里把 AIsa 添加为<b>额外的 provider</b>。你的默认模型<b>完全不动</b>。<br>想用的时候,在 opencode 的模型列表里选新加的 <code>aisa/…</code> 就行。",
+    },
+};
+// ── step 4: capabilities ────────────────────────────────────────────────────
+export const STEP_CAPS = {
+    h1: { en: "What should your agent be able to reach?", zh: "你的 agent 需要够得着什么?" },
+    ledeTail: {
+        en: "Open an area to pick the servers inside it. Everything here is live production data with licensed sources; you can add more later with one more <code>aisa connect</code>.",
+        zh: "点开一个领域来挑里面的 server。这里全都是有授权来源的生产环境实时数据;之后再跑一次 <code>aisa connect</code> 就能加更多。",
+    },
+    pickArea: { en: "Pick an area above", zh: "先在上面选一个领域" },
+    selectAll: { en: "Select all in this area", zh: "全选这个领域" },
+    /**
+     * The counts sentence as one string, because punctuation is not shared
+     * between languages: joining "N", "areas", ", " in code produced
+     * "2 个领域, 2 个 server, 48 个工具." — Chinese text with ASCII commas and a
+     * full stop. Whatever varies goes in as a placeholder instead.
+     */
+    counts: {
+        en: "{areas} areas, {servers} servers, {tools} tools.",
+        zh: "{areas} 个领域、{servers} 个 server、{tools} 个工具。",
+    },
+    /** Category tile meta: "3 servers · 47 tools". */
+    tileMeta: {
+        en: "{n} {noun} · {tools} tools",
+        zh: "{n} 个 {noun} · {tools} 个工具",
+    },
+    /**
+     * A placeholder both renderers can fill. The page used to carry
+     * <b id="mModel"></b> here and fill it with script — which left the
+     * terminal printing "AIsa()" with nothing inside, because stripping tags is
+     * all it can do with markup meant for the other surface.
+     */
+    serversWord: { en: "servers", zh: "server" },
+    serverWord: { en: "server", zh: "server" },
+    toolsWord: { en: "tools", zh: "个工具" },
+};
+/** Shown for a file-configured client that has no note of its own. */
+export const FILE_MODEL_FALLBACK = {
+    en: "This client picks its own models inside the app, so nothing is changed here — only the MCP servers are added.",
+    zh: "这个客户端在它自己的应用里选模型,所以这一步不做改动 —— 只添加 MCP server。",
+};
+export const CATEGORY_BLURB = {
+    "Search & Research": {
+        en: "Ranked web results with page text already extracted, plus YouTube.",
+        zh: "带排序的网页搜索结果,正文已经抽好,另有 YouTube。",
+    },
+    Finance: {
+        en: "US equities, crypto, prediction markets and what X is saying about a ticker.",
+        zh: "美股、加密货币、预测市场,以及 X 上关于某只票的讨论。",
+    },
+    Social: {
+        en: "Public X/Twitter, Reddit, Instagram and Pinterest — profiles, posts, engagement.",
+        zh: "公开的 X/Twitter、Reddit、Instagram 和 Pinterest —— 资料、帖子、互动数据。",
+    },
+    Sales: {
+        en: "Apollo B2B data — enrich people and companies, find prospects.",
+        zh: "Apollo 的 B2B 数据 —— 补全人和公司信息、找潜在客户。",
+    },
+    "SEO & Search Data": {
+        en: "Keywords, rankings, backlinks and on-page checks — and how a brand shows up in AI answers.",
+        zh: "关键词、排名、外链、页面检查 —— 还有一个品牌在 AI 回答里的露出情况。",
+    },
+    Communication: {
+        en: "An inbox an agent can actually use — its own address, sending and receiving.",
+        zh: "agent 能真正用起来的邮箱 —— 自己的地址,能收能发。",
+    },
+    Creators: {
+        en: "Find creators by audience, topic and reach, across platforms.",
+        zh: "按受众、话题和影响力找创作者,跨平台。",
+    },
+    Search: {
+        en: "YouTube search on its own.",
+        zh: "单独的 YouTube 搜索。",
+    },
+    "Market Intelligence": {
+        en: "Similarweb data — a site's traffic, engagement and where its visitors come from.",
+        zh: "Similarweb 数据 —— 网站流量、互动情况,以及访客从哪来。",
+    },
+};
+// ── step 5: install ─────────────────────────────────────────────────────────
+export const STEP_INSTALL = {
+    h1: { en: "Ready to connect", zh: "准备接线" },
+    lede: {
+        en: "Here is everything that happens, in order. It starts on its own; each step reports as it finishes, and your results open on the last step.",
+        zh: "下面是将要发生的全部事情,按顺序列出。它会自动开始;每一步完成后都会汇报,结果会在最后一步打开。",
+    },
+    authKeyed: {
+        en: "Your configured AIsa API key is written into each entry — <b>no sign-in needed</b>.",
+        zh: "你已配置的 AIsa API key 会写进每一条配置 —— <b>无需登录</b>。",
+    },
+    authFresh: {
+        en: "<b>One sign-in, nothing to paste.</b> Your browser opens the AIsa approval once; it issues a long-lived key for this machine, and every server and model is configured with it.",
+        zh: "<b>登录一次,不用粘贴任何东西。</b>浏览器会打开一次 AIsa 授权页;它会为这台机器签发一把长期有效的 key,所有 server 和模型都用它来配置。",
+    },
+};
+// ── step 6: done ────────────────────────────────────────────────────────────
+export const STEP_DONE = {
+    eyebrow: { en: "Step 6 of 6", zh: "第 6 步 / 共 6 步" },
+    h1: { en: "Almost there…", zh: "就快好了…" },
+    lede: { en: "The results appear here as soon as the run finishes.", zh: "运行一结束,结果就会显示在这里。" },
+};
+// ── strings the page script needs at runtime ────────────────────────────────
+// Resolved to the current language and injected as a single object, because
+// the browser has no access to this file. Same source, one hop later.
+export const RUNTIME_COPY = {
+    nothingSelected: {
+        en: "Nothing selected yet — pick at least one server.",
+        zh: "还没有选择 —— 至少选一个 server。",
+    },
+    planBackup: {
+        "claude-code": { en: "Install the claude-aisa command", zh: "安装 claude-aisa 命令" },
+        codex: { en: "Add the aisa profile and codex-aisa", zh: "添加 aisa profile 和 codex-aisa" },
+        vscode: { en: "Add AIsa models to VS Code chat", zh: "把 AIsa 模型加进 VS Code 聊天" },
+        other: { en: "Add AIsa as a backup provider", zh: "把 AIsa 添加为备用 provider" },
+    },
+    planBackupNote: {
+        en: "your current setup stays untouched",
+        zh: "你现有的配置原封不动",
+    },
+    ledeFailed: {
+        en: "Some steps did not complete — the details are in the list; your results page explains how to retry.",
+        zh: "有些步骤没有完成 —— 明细在上面的列表里,结果页会说明怎么重试。",
+    },
+    ledeAllRan: {
+        en: "Everything ran. Your results, balance and try-it-now prompts are on the last step.",
+        zh: "全部执行完毕。结果、余额和可以直接试的提示语都在最后一步。",
+    },
+    backupIntact: { en: "Your usual setup is untouched.", zh: "你平时用的配置原封不动。" },
+    backupOpencode: {
+        en: "Pick <code>aisa/…</code> from opencode's model list whenever you want AIsa.",
+        zh: "想用 AIsa 的时候,在 opencode 的模型列表里选 <code>aisa/…</code> 即可。",
+    },
+    // Bold, like every other command a reader is meant to type later. The two
+    // sit in the same paragraph on the results page and were marked up
+    // differently — one command looking more like a command than the other,
+    // for no reason a reader could infer.
+    backupBin: {
+        // Plain <code>, like every other command in this section's prose. The
+        // extra weight was added to make two neighbours match; matching the whole
+        // page is the better answer — one page with one way of writing a command
+        // beats one paragraph with its own.
+        en: "Run <code>{bin}</code> whenever you want AIsa's models; delete that one file to remove it.",
+        zh: "想用 AIsa 的模型时就跑 <code>{bin}</code>;删掉那一个文件就等于移除。",
+    },
+    noKeyStored: {
+        en: "No key stored here — run <code>aisa login</code>, then copy it from <code>~/.aisa/key</code>.",
+        zh: "本机没有存 key —— 先跑 <code>aisa login</code>,再从 <code>~/.aisa/key</code> 复制。",
+    },
+    cannotStartApp: {
+        en: "Could not start {name} — open it from your Applications folder.",
+        zh: "无法启动 {name} —— 请从应用程序文件夹里手动打开。",
+    },
+    // Louder than the prose around it, because when this appears it is the only
+    // instruction left: the button that would have done it is gone, and the
+    // command is the whole of what the reader has to act on.
+    cannotOpenTerminal: {
+        en: "Could not open a terminal — run <b><code>{bin}</code></b> in any terminal.",
+        zh: "无法打开终端 —— 在任意终端里跑 <b><code>{bin}</code></b>。",
+    },
+};
+/** Flatten RUNTIME_COPY to plain strings for the page script. */
+export function runtimeCopy(lang) {
+    const walk = (v) => {
+        if (v && typeof v === "object" && "en" in v && "zh" in v)
+            return t(v, lang);
+        if (v && typeof v === "object") {
+            return Object.fromEntries(Object.entries(v).map(([k, x]) => [k, walk(x)]));
+        }
+        return v;
+    };
+    return walk(RUNTIME_COPY);
+}
+/** What each language calls itself — never translated. */
+export const LANG_LABEL = { en: "EN", zh: "中文" };
+export const NAV = {
+    back: { en: "← Back", zh: "← 返回" },
+};
+/** The forward button, worded for the step it leads out of. */
+export const NEXT_WORD = {
+    1: { en: "Let's get started ", zh: "开始吧 " },
+    2: { en: "Continue to models ", zh: "下一步:模型 " },
+    3: { en: "Continue to capabilities ", zh: "下一步:能力 " },
+    4: { en: "Review and connect ", zh: "确认并连接 " },
+    5: { en: "", zh: "" },
+    6: { en: "", zh: "" },
+};
+/** The last gate before anything is written. */
+export const CONFIRM = {
+    heading: { en: "Ready to connect", zh: "准备就绪" },
+    agent: { en: "Agent", zh: "Agent" },
+    models: { en: "Models", zh: "模型" },
+    capabilities: { en: "Capabilities", zh: "能力" },
+    /**
+     * A word, not a keystroke. Everything up to here was reversible browsing;
+     * this is the point where files on the machine change, and a prompt you can
+     * clear with a stray return is the wrong shape for it.
+     */
+    /**
+     * A picker rather than a typed word.
+     *
+     * Typing was the first shape, and it fought the arrow-key steps beside it:
+     * keys left in the buffer by the picker arrived at readline as empty lines,
+     * which re-asked, which read again — the prompt repeated down the screen
+     * and tangled with the one after it. One input method for the whole flow.
+     *
+     * The deliberateness is kept by where the cursor starts: on "go back", so
+     * a stray return changes nothing.
+     */
+    apply: { en: "Apply this — write it to my machine", zh: "开始配置 —— 写入这台机器" },
+    goBack: { en: "Go back and change something", zh: "返回修改" },
+    ask: { en: "Confirm", zh: "确认" },
+    cancelled: { en: "Nothing was written.", zh: "什么都没有写入。" },
+    /** n means change something, not give up — the run is still yours. */
+    backToEdit: {
+        en: "Back to the start — your choices are kept.",
+        zh: "回到开头重新选 —— 刚才选的都还在。",
+    },
+};
+/** The last question: run something now, or hand the terminal back. */
+export const LAUNCH = {
+    question: { en: "What would you like to do now?", zh: "现在想做什么?" },
+    exit: { en: "Exit", zh: "退出" },
+    /** Says what it does: the page goes with it. Claiming the page survives
+     *  was true only while the process lingered, which is what made Exit feel
+     *  broken — it did not exit. */
+    exitNote: {
+        en: "back to your prompt — closes the results page",
+        zh: "回到命令提示符 —— 同时关闭结果页",
+    },
+};
+/** Shown on a page a newer run has taken over. */
+/**
+ * How a run introduces the two surfaces it is about to use.
+ *
+ * The console case used to lead with what was missing — "no browser on this
+ * machine" — which reads like a fault report for something that is not a
+ * fault at all. A server without a desktop is a perfectly ordinary place to
+ * run this, and the terminal is a first-class way through the flow, not a
+ * consolation. So it is introduced as a mode, by name, and the sentence
+ * after it gets on with the work.
+ */
+export const SURFACE = {
+    pageTitle: { en: "Open this page to choose", zh: "打开这个页面来选择" },
+    opening: {
+        en: "opening your browser… (or answer here — both stay in step)",
+        zh: "正在打开浏览器…(也可以直接在这里回答 —— 两边始终同步)",
+    },
+    urlOnly: {
+        en: "open the URL above, or answer here — both stay in step",
+        zh: "打开上面的地址,或者直接在这里回答 —— 两边始终同步",
+    },
+    consoleTitle: { en: "Console mode", zh: "控制台模式" },
+    consoleBody: {
+        en: "The browser view is not open here, so we will do this together in the terminal — the same questions, and it is very easy. Off we go.",
+        zh: "这里没有开浏览器界面,我们就用命令行一起把它选完 —— 问题都一样,非常简单。这就开始。",
+    },
+};
+/**
+ * Signing in from a machine whose browser is somewhere else.
+ *
+ * Clerk offers authorization_code and refresh_token and nothing else — no
+ * device grant — so there is no "type this short code" flow to reach for.
+ * What there is works, but only if the person is told plainly that the last
+ * page is *supposed* to fail: they are being asked to copy an address out of
+ * a browser that is showing an error, which nobody does unless told to.
+ */
+export const SIGNIN = {
+    start: { en: "Signing in to AIsa…", zh: "正在登录 AIsa…" },
+    willOpen: { en: "Your browser will open — approve the sign-in there.", zh: "浏览器就要打开了 —— 在那里确认登录。" },
+    ifNot: { en: "If it does not open, visit:", zh: "如果没有自动打开,访问:" },
+    here: {
+        en: "This machine has no browser, so sign in on the one in front of you.",
+        zh: "这台机器上没有浏览器,请在你面前那台电脑上登录。",
+    },
+    step1: { en: "1. Open this address there:", zh: "1. 在那台电脑上打开这个地址:" },
+    step2: {
+        en: "2. Sign in. A page then shows you a short block of text — press Copy.",
+        zh: "2. 登录。之后会出现一个页面,上面有一小段文字 —— 点「复制」。",
+    },
+    step3: { en: "3. Paste it below.", zh: "3. 粘贴到下面。" },
+    prompt: { en: "Paste it here: ", zh: "粘贴到这里: " },
+    badPaste: {
+        en: "that does not look like the redirect — it should contain code=",
+        zh: "这看起来不是那个跳转地址 —— 里面应该有 code=",
+    },
+};
+/** What the browser is told when a sign-in lands back on this machine. */
+export const SIGNIN_PAGE = {
+    ok: {
+        title: { en: "Signed in", zh: "登录成功" },
+        body: {
+            en: "Your key has been created on the machine you started from. You can close this tab.",
+            zh: "key 已经在你发起登录的那台机器上生成好了。可以关掉这个标签页了。",
+        },
+    },
+    failed: {
+        title: { en: "Sign-in was not completed", zh: "登录没有完成" },
+        body: {
+            en: "Nothing was changed. Return to your terminal and try again.",
+            zh: "什么都没有改动。回到终端重新试一次。",
+        },
+    },
+};
+export const SUPERSEDED = {
+    title: { en: "A newer setup is running", zh: "已有新的配置在运行" },
+    /** The other way a page ends: nobody replaced it, its time ran out. */
+    expiringTitle: { en: "This page is about to close", zh: "本页面即将关闭" },
+    expiringBody: {
+        en: "The run that opened it has finished. Nothing here needs you any more; the page closes in",
+        zh: "打开它的那次运行已经结束,这里没有需要你做的事了。本窗口将在",
+    },
+    body: {
+        en: "You started <b>aisa connect</b> again, and this page describes the run before it. It closes in",
+        zh: "你又运行了一次 <b>aisa connect</b>,这个页面显示的是上一次的结果。本窗口将在",
+    },
+    seconds: { en: "seconds", zh: "秒后关闭" },
+    closed: {
+        en: "This page has closed. The newer run has its own.",
+        zh: "本页面已关闭。新的运行有它自己的页面。",
+    },
+    /**
+     * The two lines every ending carries, whatever ended.
+     *
+     * A surface that is closing has the reader's attention and nothing left to
+     * spend it on — which is the one moment these are useful rather than in the
+     * way. Both are addresses, not buttons: a button is clicked once and
+     * forgotten, while `console.aisa.one` and `aisa connect` are things someone
+     * can still type next week. That is the whole point of putting them here.
+     */
+    console: {
+        en: "Usage, spending and top-ups live at",
+        zh: "用量、消费和充值都在",
+    },
+    consoleDomain: "console.aisa.one",
+    againBefore: { en: "Run", zh: "随时再跑" },
+    againCmd: "aisa connect",
+    againAfter: {
+        en: "any time — add servers, switch models, or connect another agent.",
+        zh: "—— 加 server、换模型、接另一个 agent。",
+    },
+};
+/**
+ * The second tab, once the sign-in is done.
+ *
+ * A sign-in started from inside `connect` opens a new tab, and that tab has
+ * to end somewhere. Ending it on "you can close this tab" is right for
+ * standalone `aisa login`, where closing it is the whole of what is left —
+ * but wrong here, where a setup is still running one tab over and the reader
+ * has just been sent away from it. This page goes back instead.
+ */
+export const RETURN_PAGE = {
+    ok: {
+        title: { en: "Signed in", zh: "登录成功" },
+        body: {
+            en: "Taking you back to your setup — it carries on from where it was.",
+            zh: "正在带你回到刚才的配置流程 —— 会从原来的位置继续。",
+        },
+    },
+    failed: {
+        title: { en: "Sign-in was not completed", zh: "登录没有完成" },
+        body: {
+            en: "Nothing was changed. Go back to your setup and try the sign-in again.",
+            zh: "什么都没有改动。回到配置流程再登录一次。",
+        },
+    },
+    stale: {
+        title: { en: "Nothing is waiting for this", zh: "这次回调已经没人在等了" },
+        body: {
+            en: "This sign-in has already been answered. Back to your setup.",
+            zh: "这次登录已经处理过了。回到你的配置流程。",
+        },
+    },
+    link: { en: "Back to your setup", zh: "回到配置流程" },
+};