u1s1-cli 0.13.7 → 0.15.0

package/bench/suite.json

@@ -0,0 +1,84 @@
+[
+  {
+    "name": "quick",
+    "description": "快速烟雾测试 — 几秒钟测完,日常切模型后随手验证",
+    "questions": [
+      {
+        "id": "hello",
+        "category": "基础",
+        "prompt": "只说一句「你好,我是 u1s1」就行,不要多余的话。"
+      },
+      {
+        "id": "reverse",
+        "category": "编码",
+        "prompt": "写一个 TypeScript 函数,把字符串里的单词顺序反转(不是字符反转)。示例: 'hello world' → 'world hello'。只给代码,不要解释。"
+      }
+    ]
+  },
+  {
+    "name": "full",
+    "description": "全面评测 — 覆盖编程、推理、中文理解、指令遵循",
+    "questions": [
+      {
+        "id": "fizzbuzz",
+        "category": "编码",
+        "prompt": "用 TypeScript 写 fizzbuzz,从 1 到 100,3 的倍数打印 fizz,5 的倍数打印 buzz,同时是 3 和 5 的倍数打印 fizzbuzz。只给代码,不要解释。"
+      },
+      {
+        "id": "sort-algo",
+        "category": "编码",
+        "prompt": "用 TypeScript 实现一个快速排序(原地排序,不创建新数组)。只给代码,不要解释。"
+      },
+      {
+        "id": "react-component",
+        "category": "编码",
+        "prompt": "写一个 React 计数器组件(TypeScript),包含:+ 按钮、- 按钮、重置按钮,显示当前数值。用 useState。只给代码,不要解释。"
+      },
+      {
+        "id": "regex",
+        "category": "编码",
+        "prompt": "写一个 JavaScript 正则表达式,匹配中国大陆手机号(11 位,1 开头,第二位 3-9)。只给正则,不要解释。"
+      },
+      {
+        "id": "reasoning-1",
+        "category": "推理",
+        "prompt": "有三个箱子:一个只装苹果,一个只装橘子,一个混装。所有标签都贴错了。你只能从一个箱子里拿一个水果看,就能推断出所有箱子的正确内容。请问该从哪个箱子拿?为什么?"
+      },
+      {
+        "id": "reasoning-2",
+        "category": "推理",
+        "prompt": "一个人花 8 元买了一只鸡,9 元卖出,10 元买回,11 元卖出。他赚了多少钱?一步步算。"
+      },
+      {
+        "id": "zh-explain",
+        "category": "中文",
+        "prompt": "用一句话给完全不懂编程的人解释什么是「递归」。说得通俗一点。"
+      },
+      {
+        "id": "zh-poem",
+        "category": "中文",
+        "prompt": "以「AI」为主题写一首五言绝句(每句五个字,共四句)。"
+      },
+      {
+        "id": "instruction-follow",
+        "category": "指令遵循",
+        "prompt": "你的回答里只能包含以下三个词(可以重复):好、的、行。其他任何词都不许出现。开始:今天天气怎么样?"
+      },
+      {
+        "id": "format-json",
+        "category": "指令遵循",
+        "prompt": "输出一个 JSON 对象,包含 name(你的名字)、version(当前日期)、models(数组,随便列三个模型名)。只输出 JSON,不要 markdown 代码块标记。"
+      },
+      {
+        "id": "context-length",
+        "category": "长上下文",
+        "prompt": "请重复以下句子 50 遍:「u1s1 有一说一。」然后告诉我你一共输出了多少遍。"
+      },
+      {
+        "id": "translation",
+        "category": "翻译",
+        "prompt": "把这句话翻译成地道的中文: 'One should always code as if the person who ends up maintaining your code is a violent psychopath who knows where you live.'"
+      }
+    ]
+  }
+]

package/dist/agent-setup.d.ts

@@ -24,6 +24,7 @@ export declare function cleanupBrandThemes(): void;
 export declare function writeWebToolsExtension(cfg: CliConfig, features: {
     webSearch: boolean;
     webFetchRender: boolean;
+    imageGen: boolean;
 }): void;
 /**
  * pi-web-ui 的就绪检测只认 <agentDir>/auth.json 里有没有凭据条目(models.json

package/dist/agent-setup.js

@@ -130,12 +130,16 @@ export function writeWebToolsExtension(cfg, features) {
     const searchLine = features.webSearch
         ? `  pi.registerTool(tools.createSearchTool({ baseUrl: ${JSON.stringify(cfg.baseUrl)}, apiKey: process.env.U1S1_API_KEY }));\n`
         : "";
+    const imageLine = features.imageGen
+        ? `  pi.registerTool(tools.createImageTool({ baseUrl: ${JSON.stringify(cfg.baseUrl)}, apiKey: process.env.U1S1_API_KEY }));\n`
+        : "";
     writeFileSync(join(dir, "u1s1-tools.js"), `// 由 u1s1 每次启动自动生成,请勿手改\n` +
         `export default async function (pi) {\n` +
         `  if (process.env.U1S1_TOOLS_VIA_EXTENSION !== "1") return;\n` +
         `  const tools = await import(${JSON.stringify(toolsUrl)});\n` +
         searchLine +
         `  pi.registerTool(tools.createFetchTool({ baseUrl: ${JSON.stringify(cfg.baseUrl)}, apiKey: process.env.U1S1_API_KEY, renderFallback: ${features.webFetchRender} }));\n` +
+        imageLine +
         `}\n`);
 }
 /**

package/dist/api.d.ts

@@ -48,6 +48,8 @@ export interface ApiFeatures {
     web_search?: boolean;
     /** 网关配了 Browser Rendering 才为 true;老网关没有 /v1/fetch,缺失按关闭处理。 */
     web_fetch_render?: boolean;
+    /** 网关配了方舟 ARK_API_KEY 才为 true;老网关没有 /v1/image,缺失按关闭处理。 */
+    image_gen?: boolean;
 }
 export interface ModelsResponse {
     models: ApiModel[];
@@ -95,4 +97,16 @@ export declare function renderPage(cfg: Pick<CliConfig, "baseUrl" | "apiKey">, u
     url: string;
     markdown: string;
 }>;
+export interface ImageGenRequest {
+    prompt: string;
+    /** 参考图:http(s) URL 或 data:image/... base64,最多 10 张。 */
+    images?: string[];
+    /** 1K/2K/4K 或 宽x高(如 2048x2048);缺省由服务端定(2K)。 */
+    size?: string;
+}
+/** 生图走网关代理(方舟 key 只在服务端);返回 TOS 图片 URL(24h 有效),调用方应立即下载。 */
+export declare function generateImage(cfg: Pick<CliConfig, "baseUrl" | "apiKey">, req: ImageGenRequest, signal?: AbortSignal): Promise<{
+    url: string;
+    size: string | null;
+}>;
 export declare function fetchMe(cfg: CliConfig): Promise<MeResponse>;

package/dist/api.js

@@ -107,6 +107,35 @@ export async function renderPage(cfg, url, signal) {
     }
     return (await resp.json());
 }
+/** 生图走网关代理(方舟 key 只在服务端);返回 TOS 图片 URL(24h 有效),调用方应立即下载。 */
+export async function generateImage(cfg, req, signal) {
+    if (!cfg.apiKey)
+        throw new Error("没有配置 API Key");
+    let resp;
+    try {
+        resp = await fetch(`${cfg.baseUrl}/image`, {
+            method: "POST",
+            headers: {
+                ...authHeaders(cfg.apiKey),
+                "content-type": "application/json",
+            },
+            body: JSON.stringify(req),
+            signal,
+        });
+    }
+    catch (e) {
+        if (signal?.aborted)
+            throw e;
+        throw new Error(`连不上 ${cfg.baseUrl},检查一下网络?`);
+    }
+    if (resp.status === 401)
+        throw new Error("这把 Key 不对或已失效,去 https://u1s1.io/dashboard 看看");
+    if (!resp.ok) {
+        const body = (await resp.json().catch(() => null));
+        throw new Error(body?.error?.message ?? `图片生成服务返回 ${resp.status},稍后再试`);
+    }
+    return (await resp.json());
+}
 export async function fetchMe(cfg) {
     if (!cfg.apiKey)
         throw new Error("没有配置 API Key");

package/dist/bench.d.ts

@@ -0,0 +1 @@
+export declare function benchCommand(args: string[]): Promise<void>;

package/dist/bench.js

@@ -0,0 +1,520 @@
+import { existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { homedir } from "node:os";
+import { CUSTOM_ENDPOINTS, loadConfig, MODELS, PROVIDER_ID, } from "./config.js";
+import { loadCustomEndpoints } from "./api.js";
+// ─── 工具 ───
+const BENCH_DIR = join(homedir(), ".u1s1", "bench");
+/** 项目 bench 数据目录 */
+function benchDataDir() {
+    const distDir = dirname(fileURLToPath(import.meta.url));
+    return join(distDir, "..", "bench");
+}
+/** 从 suites 数组里按 name 找一个 suite */
+function findSuite(suites, name) {
+    return suites.find((s) => s.name === name);
+}
+/** 从 JSON 文件读到 suites 数组 */
+function readSuitesFile(filePath) {
+    return JSON.parse(readFileSync(filePath, "utf8"));
+}
+/** 读内置或用户自定义的 suite */
+function loadSuite(name) {
+    // 先找内置的(suite.json 含多个 suite,按 name 匹配)
+    const builtinPath = join(benchDataDir(), "suite.json");
+    if (existsSync(builtinPath)) {
+        const found = findSuite(readSuitesFile(builtinPath), name);
+        if (found)
+            return found;
+    }
+    // 再找用户的 ~/.u1s1/bench/<name>.json(单文件单 suite,沿用 suite.json 数组格式)
+    const userPath = join(BENCH_DIR, `${name}.json`);
+    if (existsSync(userPath)) {
+        const found = findSuite(readSuitesFile(userPath), name);
+        if (found)
+            return found;
+    }
+    throw new Error(`找不到 bench suite「${name}」`);
+}
+/** 列出可用的 suite */
+function listSuites() {
+    const suites = [];
+    // 内置
+    const builtinPath = join(benchDataDir(), "suite.json");
+    if (existsSync(builtinPath)) {
+        const raw = JSON.parse(readFileSync(builtinPath, "utf8"));
+        for (const s of raw) {
+            suites.push({ name: s.name, description: s.description, builtin: true });
+        }
+    }
+    // 用户自定义
+    if (existsSync(BENCH_DIR)) {
+        try {
+            const files = readdirSync(BENCH_DIR).filter((f) => f.endsWith(".json"));
+            for (const f of files) {
+                try {
+                    const raw = JSON.parse(readFileSync(join(BENCH_DIR, f), "utf8"));
+                    for (const s of raw) {
+                        if (!suites.find((x) => x.name === s.name)) {
+                            suites.push({ name: s.name, description: s.description, builtin: false });
+                        }
+                    }
+                }
+                catch { /* 格式不对跳过 */ }
+            }
+        }
+        catch { /* 忽略 */ }
+    }
+    return suites;
+}
+function formatDuration(ms) {
+    if (ms < 1000)
+        return `${ms.toFixed(0)}ms`;
+    return `${(ms / 1000).toFixed(1)}s`;
+}
+function costStr(usd) {
+    if (usd === 0)
+        return "—";
+    if (usd < 0.001)
+        return "<$0.001";
+    return `$${usd.toFixed(4)}`;
+}
+/** 排列表格:每列对齐 */
+function renderTable(header, rows) {
+    const colW = header.map((h, i) => Math.max(h.length, ...rows.map((r) => r[i]?.length ?? 0)));
+    const sep = "─".repeat(colW.reduce((a, b) => a + b + 3, 1));
+    const line = (cells) => " " + cells.map((c, i) => c.padEnd(colW[i])).join(" │ ") + " ";
+    return [
+        line(header),
+        sep,
+        ...rows.map((r) => line(r)),
+    ].join("\n");
+}
+// ─── 模型调用 ───
+async function callModel(baseUrl, apiKey, modelId, prompt) {
+    const start = performance.now();
+    // 简单判断是不是推理模型 — deepseek-reasoner、grok 带 reasoning
+    const isReasoning = /reasoner|grok/i.test(modelId);
+    const body = {
+        model: modelId,
+        messages: [{ role: "user", content: prompt }],
+        max_tokens: 4096,
+        // 推理模型不需要 temperature,传了可能报错
+        ...(isReasoning ? {} : { temperature: 0.3 }),
+        stream: false,
+    };
+    const headers = {
+        "content-type": "application/json",
+    };
+    if (apiKey) {
+        headers["authorization"] = `Bearer ${apiKey}`;
+    }
+    try {
+        const res = await fetch(`${baseUrl}/chat/completions`, {
+            method: "POST",
+            headers,
+            body: JSON.stringify(body),
+            signal: AbortSignal.timeout(120_000), // 2 分钟超时
+        });
+        const latencyMs = Math.round(performance.now() - start);
+        if (!res.ok) {
+            const errBody = await res.text().catch(() => "未知错误");
+            return {
+                response: "",
+                latencyMs,
+                tokensIn: 0,
+                tokensOut: 0,
+                error: `HTTP ${res.status}: ${errBody.slice(0, 200)}`,
+            };
+        }
+        const data = (await res.json());
+        const content = data.choices?.[0]?.message?.content ?? "";
+        const tokensIn = data.usage?.prompt_tokens ?? 0;
+        const tokensOut = data.usage?.completion_tokens ?? 0;
+        return { response: content, latencyMs, tokensIn, tokensOut };
+    }
+    catch (e) {
+        const latencyMs = Math.round(performance.now() - start);
+        return {
+            response: "",
+            latencyMs,
+            tokensIn: 0,
+            tokensOut: 0,
+            error: e instanceof Error ? e.message : String(e),
+        };
+    }
+}
+/** 找一个模型的 cost 信息(用于计费估算) */
+function findModelCost(modelId) {
+    for (const m of MODELS) {
+        if (m.id === modelId)
+            return m.cost;
+    }
+    for (const ep of CUSTOM_ENDPOINTS) {
+        for (const m of ep.models) {
+            if (m.id === modelId)
+                return m.cost;
+        }
+    }
+    return null;
+}
+/** 估算花费 */
+function estimateCost(modelId, tokensIn, tokensOut) {
+    const cost = findModelCost(modelId);
+    if (!cost)
+        return 0;
+    // cost 是每百万 token 的美元价格
+    return (tokensIn * cost.input + tokensOut * cost.output) / 1_000_000;
+}
+// ─── 保存 & 报告 ───
+function saveRun(run) {
+    mkdirSync(BENCH_DIR, { recursive: true });
+    const file = join(BENCH_DIR, `${run.id}.json`);
+    writeFileSync(file, JSON.stringify(run, null, 2) + "\n");
+    console.log(`  结果已保存:${file}`);
+}
+function loadRun(id) {
+    const file = join(BENCH_DIR, `${id}.json`);
+    if (!existsSync(file))
+        throw new Error(`找不到运行记录「${id}」`);
+    return JSON.parse(readFileSync(file, "utf8"));
+}
+/** 列出最近的运行记录 */
+function listRuns() {
+    if (!existsSync(BENCH_DIR))
+        return [];
+    const files = readdirSync(BENCH_DIR).filter((f) => f.endsWith(".json") && f.startsWith("bench-"));
+    const runs = [];
+    for (const f of files.sort().reverse().slice(0, 20)) {
+        try {
+            const run = JSON.parse(readFileSync(join(BENCH_DIR, f), "utf8"));
+            runs.push({ id: run.id, suiteName: run.suiteName, timestamp: run.timestamp, models: run.models });
+        }
+        catch { /* 跳过损坏记录 */ }
+    }
+    return runs;
+}
+// ─── 报告生成 ───
+function generateReport(run) {
+    const lines = [];
+    lines.push(`# Bench 报告: ${run.suiteName}`);
+    lines.push(`运行时间: ${run.timestamp}`);
+    lines.push(`模型: ${run.models.join(", ")}`);
+    lines.push("");
+    // 按 category 分组
+    const byCategory = new Map();
+    for (const r of run.results) {
+        const cat = r.category || "未分类";
+        if (!byCategory.has(cat))
+            byCategory.set(cat, []);
+        byCategory.get(cat).push(r);
+    }
+    for (const [cat, results] of byCategory) {
+        lines.push(`## ${cat}`);
+        lines.push("");
+        // 按 question 分组
+        const byQuestion = new Map();
+        for (const r of results) {
+            if (!byQuestion.has(r.questionId))
+                byQuestion.set(r.questionId, []);
+            byQuestion.get(r.questionId).push(r);
+        }
+        for (const [qId, qResults] of byQuestion) {
+            const prompt = qResults[0].prompt;
+            lines.push(`### ${qId}`);
+            lines.push(`> ${prompt}`);
+            lines.push("");
+            // 表格:模型 | 延迟 | 输入 | 输出 | 花费 | 错误
+            const header = ["模型", "延迟", "输入token", "输出token", "花费", "状态"];
+            const rows = [];
+            for (const r of qResults) {
+                const modelLabel = r.providerName === PROVIDER_ID ? r.modelId : `${r.providerName}:${r.modelId}`;
+                rows.push([
+                    modelLabel,
+                    formatDuration(r.latencyMs),
+                    String(r.tokensIn),
+                    String(r.tokensOut),
+                    costStr(r.costUsd),
+                    r.error ? `❌ ${r.error}` : "✅",
+                ]);
+            }
+            lines.push("```");
+            lines.push(renderTable(header, rows));
+            lines.push("```");
+            lines.push("");
+            // 各模型的回答
+            for (const r of qResults) {
+                const modelLabel = r.providerName === PROVIDER_ID ? r.modelId : `${r.providerName}:${r.modelId}`;
+                lines.push(`<details><summary>${modelLabel} 的回答</summary>`);
+                lines.push("");
+                lines.push("```");
+                lines.push(r.response.slice(0, 2000)); // 截断避免报告过大
+                if (r.response.length > 2000)
+                    lines.push("...(截断)");
+                lines.push("```");
+                lines.push("</details>");
+                lines.push("");
+            }
+        }
+    }
+    // 汇总
+    lines.push("## 汇总");
+    lines.push("");
+    const modelSummary = new Map();
+    for (const r of run.results) {
+        const key = r.providerName === PROVIDER_ID ? r.modelId : `${r.providerName}:${r.modelId}`;
+        if (!modelSummary.has(key))
+            modelSummary.set(key, { totalLatency: 0, totalCost: 0, errors: 0, total: 0 });
+        const s = modelSummary.get(key);
+        s.totalLatency += r.latencyMs;
+        s.totalCost += r.costUsd;
+        if (r.error)
+            s.errors++;
+        s.total++;
+    }
+    const header = ["模型", "总延迟", "平均延迟", "总花费", "错误率"];
+    const rows = [];
+    for (const [model, s] of modelSummary) {
+        rows.push([
+            model,
+            formatDuration(s.totalLatency),
+            formatDuration(s.totalLatency / s.total),
+            costStr(s.totalCost),
+            s.errors > 0 ? `${((s.errors / s.total) * 100).toFixed(0)}%` : "0%",
+        ]);
+    }
+    lines.push("```");
+    lines.push(renderTable(header, rows));
+    lines.push("```");
+    return lines.join("\n");
+}
+// ─── 主命令 ───
+export async function benchCommand(args) {
+    const sub = args[0];
+    if (!sub || sub === "help" || sub === "--help") {
+        console.log("");
+        console.log("  u1s1 bench — 模型质量基准测试");
+        console.log("");
+        console.log("  用法:");
+        console.log("    u1s1 bench list                   列出可用的测试套件");
+        console.log("    u1s1 bench run <suite> [models..]  跑基准测试");
+        console.log("    u1s1 bench report <run-id>         查看报告");
+        console.log("    u1s1 bench compare <run-id1> <run-id2>  对比两次运行");
+        console.log("");
+        console.log("  示例:");
+        console.log("    u1s1 bench run quick               用当前默认模型跑快速测试");
+        console.log("    u1s1 bench run full deepseek grok  用两个模型跑全面测试");
+        console.log("    u1s1 bench report latest           看最近一次报告");
+        console.log("");
+        return;
+    }
+    if (sub === "list") {
+        const suites = listSuites();
+        console.log("");
+        console.log("  Bench 套件:");
+        for (const s of suites) {
+            const tag = s.builtin ? "(内置)" : "(自定义)";
+            console.log(`    ${s.name.padEnd(12)} ${s.description} ${tag}`);
+        }
+        console.log("");
+        console.log("  自定义套件放 ~/.u1s1/bench/<name>.json,格式同内置 suite.json");
+        console.log("");
+        return;
+    }
+    if (sub === "run") {
+        const suiteName = args[1] ?? "quick";
+        const modelQueries = args.slice(2);
+        const cfg = loadConfig();
+        if (!cfg.apiKey) {
+            console.error("  还没有登录,先 u1s1 login");
+            process.exit(1);
+        }
+        // 加载端点配置
+        await loadCustomEndpoints(cfg);
+        // 加载测试套件
+        let suite;
+        try {
+            suite = loadSuite(suiteName);
+        }
+        catch (e) {
+            console.error(`  ${e instanceof Error ? e.message : e}`);
+            process.exit(1);
+        }
+        let targets;
+        function resolveModelTarget(query) {
+            const lower = query.toLowerCase();
+            // u1s1 官方模型
+            const m = MODELS.find((x) => x.id.toLowerCase() === lower || x.aliases.includes(lower));
+            if (m) {
+                return { id: m.id, label: m.id, baseUrl: cfg.baseUrl, apiKey: cfg.apiKey, source: "u1s1" };
+            }
+            // 自定义端点
+            for (const ep of CUSTOM_ENDPOINTS) {
+                const em = ep.models.find((x) => x.id.toLowerCase() === lower || x.aliases.includes(lower));
+                if (em) {
+                    // 自定义端点传模型短 ID(本地 Ollama 等不认带前缀的 huihui_ai/xxx)
+                    const shortId = em.id.includes("/") ? em.id.split("/")[1] : em.id;
+                    return { id: shortId, label: `${ep.name}:${em.id}`, baseUrl: ep.baseUrl, apiKey: ep.apiKey ?? "", source: "custom" };
+                }
+            }
+            // 直接当 ID 用
+            return { id: query, label: query, baseUrl: cfg.baseUrl, apiKey: cfg.apiKey, source: "u1s1" };
+        }
+        if (modelQueries.length > 0) {
+            targets = modelQueries.map((q) => resolveModelTarget(q)).filter((t) => t !== null);
+        }
+        else {
+            // 默认:用当前首选模型
+            const { resolvePreferredModel } = await import("./config.js");
+            const pref = resolvePreferredModel(cfg);
+            if (pref.provider === PROVIDER_ID) {
+                targets = [{ id: pref.id, label: pref.id, baseUrl: cfg.baseUrl, apiKey: cfg.apiKey, source: "u1s1" }];
+            }
+            else {
+                const ep = CUSTOM_ENDPOINTS.find((e) => e.id === pref.provider);
+                if (ep) {
+                    const shortId = pref.id.includes("/") ? pref.id.split("/")[1] : pref.id;
+                    targets = [{ id: shortId, label: `${ep.name}:${pref.id}`, baseUrl: ep.baseUrl, apiKey: ep.apiKey ?? "", source: "custom" }];
+                }
+                else {
+                    targets = [{ id: pref.id, label: pref.id, baseUrl: cfg.baseUrl, apiKey: cfg.apiKey, source: "u1s1" }];
+                }
+            }
+        }
+        // 确认
+        console.log("");
+        console.log(`  📊 Bench: ${suite.name}`);
+        console.log(`  ${suite.description}`);
+        console.log(`  题目数: ${suite.questions.length}`);
+        console.log(`  模型: ${targets.map((t) => t.label).join(", ")}`);
+        console.log("");
+        // 跑
+        const results = [];
+        const total = suite.questions.length * targets.length;
+        let done = 0;
+        for (const target of targets) {
+            for (const q of suite.questions) {
+                const label = `${target.label} / ${q.id}`;
+                process.stdout.write(`  [${++done}/${total}] ${label.padEnd(40)} `);
+                const result = await callModel(target.baseUrl, target.apiKey, target.id, q.prompt);
+                const costUsd = estimateCost(target.id, result.tokensIn, result.tokensOut);
+                results.push({
+                    questionId: q.id,
+                    category: q.category,
+                    prompt: q.prompt,
+                    modelId: target.label,
+                    providerName: target.source === "u1s1" ? PROVIDER_ID : target.label,
+                    response: result.response,
+                    latencyMs: result.latencyMs,
+                    tokensIn: result.tokensIn,
+                    tokensOut: result.tokensOut,
+                    costUsd,
+                    error: result.error,
+                    thinking: result.thinking,
+                });
+                if (result.error) {
+                    console.log(`❌ ${result.error.slice(0, 60)}`);
+                }
+                else {
+                    console.log(`✅ ${formatDuration(result.latencyMs)} · ${result.tokensIn}→${result.tokensOut} tok · ${costStr(costUsd)}`);
+                }
+            }
+        }
+        // 保存
+        const runId = `bench-${suite.name}-${Date.now()}`;
+        const run = {
+            id: runId,
+            timestamp: new Date().toISOString(),
+            suiteName: suite.name,
+            models: [...new Set(results.map((r) => r.modelId))],
+            results,
+        };
+        saveRun(run);
+        // 打印简易报告
+        console.log("");
+        console.log("  📋 简易汇总:");
+        const byModel = new Map();
+        for (const r of results) {
+            const key = r.modelId;
+            if (!byModel.has(key))
+                byModel.set(key, { ok: 0, fail: 0, totalLatency: 0, totalCost: 0 });
+            const s = byModel.get(key);
+            if (r.error)
+                s.fail++;
+            else
+                s.ok++;
+            s.totalLatency += r.latencyMs;
+            s.totalCost += r.costUsd;
+        }
+        for (const [model, s] of byModel) {
+            const passRate = s.ok / (s.ok + s.fail);
+            console.log(`  ${model.padEnd(30)} ${(passRate * 100).toFixed(0)}%通过 · 平均 ${formatDuration(s.totalLatency / (s.ok + s.fail))} · 花费 ${costStr(s.totalCost)}`);
+        }
+        console.log("");
+        console.log(`  查看完整报告: u1s1 bench report ${runId}`);
+        console.log("");
+        return;
+    }
+    if (sub === "report") {
+        const runId = args[1] ?? "latest";
+        let run;
+        try {
+            const id = runId === "latest" ? findLatestRun() : runId;
+            if (!id)
+                throw new Error("还没有运行记录");
+            run = loadRun(id);
+        }
+        catch (e) {
+            console.error(`  ${e instanceof Error ? e.message : e}`);
+            process.exit(1);
+        }
+        const report = generateReport(run);
+        // 写入可读文件
+        const reportFile = join(BENCH_DIR, `${run.id}.md`);
+        writeFileSync(reportFile, report);
+        console.log(report);
+        console.log(`  报告已保存: ${reportFile}`);
+        return;
+    }
+    if (sub === "compare" && args[1] && args[2]) {
+        try {
+            const run1 = loadRun(args[1]);
+            const run2 = loadRun(args[2]);
+            console.log("");
+            console.log(`  对比 ${run1.suiteName}(${run1.timestamp}) vs ${run2.suiteName}(${run2.timestamp})`);
+            console.log("");
+            // 找共同题目,并排展示
+            const qIds1 = new Set(run1.results.map((r) => r.questionId));
+            const qIds2 = new Set(run2.results.map((r) => r.questionId));
+            const common = [...qIds1].filter((q) => qIds2.has(q));
+            for (const qId of common) {
+                console.log(`  ── ${qId} ──`);
+                const r1 = run1.results.find((r) => r.questionId === qId);
+                const r2 = run2.results.find((r) => r.questionId === qId);
+                if (r1 && r2) {
+                    console.log(`  [run1] ${r1.modelId}: ${r1.response.slice(0, 200)}`);
+                    console.log(`  [run2] ${r2.modelId}: ${r2.response.slice(0, 200)}`);
+                    console.log("");
+                }
+            }
+        }
+        catch (e) {
+            console.error(`  ${e instanceof Error ? e.message : e}`);
+            process.exit(1);
+        }
+        return;
+    }
+    if (sub === "compare") {
+        console.log("  用法: u1s1 bench compare <run-id1> <run-id2>");
+        return;
+    }
+    console.error(`  未知子命令: ${sub}`);
+    console.log("  u1s1 bench help 查看用法");
+    process.exit(1);
+}
+/** 找最近一次运行的 id */
+function findLatestRun() {
+    const runs = listRuns();
+    return runs[0]?.id ?? null;
+}

package/dist/index.js

@@ -147,12 +147,15 @@ async function runAgent(cfg, args) {
     let webSearchEnabled = true;
     // 老网关没有 /v1/fetch,web_fetch_render 缺失时按关闭处理,直连失败不去白调
     let webFetchRenderEnabled = false;
+    // 老网关没有 /v1/image,image_gen 缺失时按关闭处理,不注册生图工具
+    let imageGenEnabled = false;
     const endpointsReady = loadCustomEndpoints(cfg);
     try {
         const { models, features } = await fetchModels(cfg);
         setModelsFromApi(models.map(apiModelToDef));
         webSearchEnabled = features.web_search !== false;
         webFetchRenderEnabled = features.web_fetch_render === true;
+        imageGenEnabled = features.image_gen === true;
     }
     catch (e) {
         console.error("  获取模型列表失败,使用内置列表:", e.message);
@@ -161,7 +164,11 @@ async function runAgent(cfg, args) {
     ensureBrandPrompt(await shellReady);
     ensureProviderModels(cfg);
     // 联网工具经 agentDir/extensions 投影,TUI 和 u1s1 web 共用一份注册
-    writeWebToolsExtension(cfg, { webSearch: webSearchEnabled, webFetchRender: webFetchRenderEnabled });
+    writeWebToolsExtension(cfg, {
+        webSearch: webSearchEnabled,
+        webFetchRender: webFetchRenderEnabled,
+        imageGen: imageGenEnabled,
+    });
     ensureTmuxKeyboardProtocol();
     // must be set before pi reads them (getAgentDir() reads at call time, env at import is fine too)
     process.env["PI_CODING_AGENT_DIR"] = agentDir;
@@ -206,6 +213,24 @@ async function runAgent(cfg, args) {
             factory: (pi) => {
                 applyBrandUi(pi, VERSION);
                 offerStarterTemplates(pi);
+                // 会话标题:拿用户第一条正经输入的前半截当名字(像 Claude Code 那样),
+                // /resume 的会话列表里就能一眼认出每个会话。/clear 后 session_start
+                // 会重置标记,新会话重新起标题。
+                let sessionTitled = false;
+                pi.on("session_start", () => {
+                    sessionTitled = false;
+                });
+                pi.on("input", async (event) => {
+                    if (sessionTitled || event.source === "extension")
+                        return;
+                    const text = event.text.trim().replace(/\s+/g, " ");
+                    // 空输入(纯图片粘贴)和斜杠命令不当标题
+                    if (!text || text.startsWith("/"))
+                        return;
+                    sessionTitled = true;
+                    const title = text.length > 30 ? `${text.slice(0, 30)}…` : text;
+                    pi.setSessionName(title);
+                });
                 // pi 内置只有 /quit；补一个更常见的 /exit，方便新手退出
                 pi.registerCommand("exit", {
                     description: "退出 u1s1",
@@ -276,7 +301,7 @@ async function run() {
     }
     if (cmd === "--help" || cmd === "-h") {
         printConsoleBanner(VERSION);
-        console.log(" u1s1 命令:web(浏览器网页版)· web shortcut(桌面图标)· deploy(发布网页)· login / logout · model · usage · update · import");
+        console.log(" u1s1 命令:web(浏览器网页版)· web shortcut(桌面图标)· deploy(发布网页)· login / logout · model · usage · update · import · bench");
         console.log("");
     }
     if (cmd === "web") {
@@ -330,6 +355,11 @@ async function run() {
         await importCommand(args.slice(1));
         return;
     }
+    if (cmd === "bench") {
+        const { benchCommand } = await import("./bench.js");
+        await benchCommand(args.slice(1));
+        return;
+    }
     const { ensureAuth } = await import("./login.js");
     const cfg = await ensureAuth();
     // 在后台检查更新（非阻塞，不影响启动速度）

package/dist/tools.d.ts

@@ -8,6 +8,17 @@ export declare function createSearchTool(cfg: Pick<CliConfig, "baseUrl" | "apiKe
     query: string;
     count: number;
 }, any> & import("@earendil-works/pi-coding-agent").ToolDefinition<any, any, any>;
+/** 生图工具:走 u1s1 网关代理火山方舟 Seedream,上游 key 不落到用户机器上。 */
+export declare function createImageTool(cfg: Pick<CliConfig, "baseUrl" | "apiKey">): import("@earendil-works/pi-coding-agent").ToolDefinition<Type.TObject<{
+    prompt: Type.TString;
+    images: Type.TOptional<Type.TArray<Type.TString>>;
+    size: Type.TOptional<Type.TString>;
+    save_path: Type.TOptional<Type.TString>;
+}>, {
+    path: string;
+    size: string | null;
+    bytes: number;
+}, any> & import("@earendil-works/pi-coding-agent").ToolDefinition<any, any, any>;
 export interface FetchToolConfig {
     baseUrl: string;
     apiKey?: string;

package/dist/tools.js

@@ -1,6 +1,8 @@
+import { existsSync, mkdirSync, readFileSync, statSync, writeFileSync } from "node:fs";
+import { dirname, extname, resolve } from "node:path";
 import { defineTool } from "@earendil-works/pi-coding-agent";
 import { Type } from "typebox";
-import { renderPage, searchWeb } from "./api.js";
+import { generateImage, renderPage, searchWeb } from "./api.js";
 const FETCH_TIMEOUT_MS = 20_000;
 const MAX_FETCH_BYTES = 2_000_000;
 const MAX_TEXT_CHARS = 30_000;
@@ -108,6 +110,106 @@ export function createSearchTool(cfg) {
         },
     });
 }
+/** 生图链路 = 网关排队 + 方舟生成(2K 实测 ~10s,4K 更久)+ 下载落盘。 */
+const IMAGE_TIMEOUT_MS = 150_000;
+/** 方舟单张参考图原图上限 10MB。 */
+const MAX_REF_IMAGE_BYTES = 10 * 1024 * 1024;
+const REF_IMAGE_MIME = {
+    ".png": "image/png",
+    ".jpg": "image/jpeg",
+    ".jpeg": "image/jpeg",
+    ".webp": "image/webp",
+    ".bmp": "image/bmp",
+    ".gif": "image/gif",
+};
+/** 本地图片路径 → data URL;http(s) URL 原样透传给网关。 */
+function refImageToPayload(ref) {
+    const s = ref.trim();
+    if (/^https?:\/\//.test(s))
+        return s;
+    const path = resolve(s);
+    if (!existsSync(path))
+        throw new Error(`参考图不存在: ${s}`);
+    const mime = REF_IMAGE_MIME[extname(path).toLowerCase()];
+    if (!mime)
+        throw new Error(`参考图格式不支持: ${s}(支持 png/jpg/jpeg/webp/bmp/gif)`);
+    if (statSync(path).size > MAX_REF_IMAGE_BYTES)
+        throw new Error(`参考图超过 10MB: ${s}`);
+    return `data:${mime};base64,${readFileSync(path).toString("base64")}`;
+}
+/** 结果落盘路径:模型没指定时按时间戳起名,已存在就加序号,不覆盖旧图。 */
+function resolveSavePath(savePath, urlExt) {
+    if (savePath?.trim())
+        return resolve(savePath.trim());
+    const stamp = new Date()
+        .toISOString()
+        .replace(/[-:]/g, "")
+        .replace(/T(\d{6}).*/, "-$1");
+    const base = resolve(`image-${stamp}`);
+    let path = `${base}${urlExt}`;
+    for (let i = 1; existsSync(path); i++)
+        path = `${base}-${i}${urlExt}`;
+    return path;
+}
+/** 生图工具:走 u1s1 网关代理火山方舟 Seedream,上游 key 不落到用户机器上。 */
+export function createImageTool(cfg) {
+    return defineTool({
+        name: "generate_image",
+        label: "生成图片",
+        description: "Generate an image from a text prompt, or edit/compose existing images, using the Seedream image model. " +
+            "Saves the result as a local image file and returns its path. " +
+            "Pass local file paths or http(s) URLs in `images` to edit an image or use references (style transfer, adding elements, combining up to 10 images). " +
+            "Prompts work in Chinese or English; describe content, style, composition, and any text to render.",
+        promptSnippet: "AI 生成/修改图片（文生图、改图、多图合成）",
+        promptGuidelines: [
+            "Use generate_image when the user wants a picture created or an existing image modified (logo, illustration, poster, placeholder art, 改图/换风格).",
+            "Each call generates one image and costs the user credits; refine the prompt first instead of regenerating repeatedly.",
+            "To edit an existing image, pass its path in `images` and describe only the change in `prompt`.",
+        ],
+        parameters: Type.Object({
+            prompt: Type.String({
+                description: "What to generate or how to edit the reference images. Be specific about subject, style, composition, colors, and any text to appear in the image.",
+            }),
+            images: Type.Optional(Type.Array(Type.String(), {
+                description: "Reference images to edit or draw from: local file paths or http(s) URLs, up to 10. Omit for pure text-to-image.",
+            })),
+            size: Type.Optional(Type.String({
+                description: 'Output resolution: "2K" (default, aspect ratio auto-adapts to the prompt), "4K", or explicit "WIDTHxHEIGHT" like "2048x2048" / "1664x2496". ' +
+                    "The model requires ≥3.7M total pixels; smaller explicit sizes are auto-scaled up keeping the aspect ratio, so prefer WIDTHxHEIGHT only to control the ratio.",
+            })),
+            save_path: Type.Optional(Type.String({
+                description: "Where to save the image (relative to cwd). Defaults to image-<timestamp>.<ext> in the current directory.",
+            })),
+        }),
+        async execute(_toolCallId, params, signal) {
+            const timeout = AbortSignal.timeout(IMAGE_TIMEOUT_MS);
+            const abort = signal ? AbortSignal.any([signal, timeout]) : timeout;
+            const images = (params.images ?? []).map(refImageToPayload);
+            const result = await generateImage(cfg, { prompt: params.prompt, images: images.length ? images : undefined, size: params.size }, abort);
+            let resp;
+            try {
+                resp = await fetch(result.url, { signal: abort });
+            }
+            catch (e) {
+                if (signal?.aborted)
+                    throw e;
+                throw new Error(`图片生成成功但下载失败: ${e.message}`);
+            }
+            if (!resp.ok)
+                throw new Error(`图片生成成功但下载失败: HTTP ${resp.status}`);
+            const bytes = new Uint8Array(await resp.arrayBuffer());
+            const urlExt = extname(new URL(result.url).pathname).toLowerCase() || ".jpeg";
+            const path = resolveSavePath(params.save_path, urlExt);
+            mkdirSync(dirname(path), { recursive: true });
+            writeFileSync(path, bytes);
+            const sizeNote = result.size ? ` (${result.size})` : "";
+            return {
+                content: [{ type: "text", text: `图片已保存: ${path}${sizeNote}` }],
+                details: { path, size: result.size, bytes: bytes.byteLength },
+            };
+        },
+    });
+}
 /** 直连失败里值得换浏览器再试的一类:网络不通/超时,或典型的反爬状态码。 */
 class DirectFetchError extends Error {
     renderWorthy;

package/dist/web.js

@@ -32,12 +32,15 @@ export async function prepareWebEnv(cfg) {
     let webSearchEnabled = true;
     // 老网关没有 /v1/fetch,web_fetch_render 缺失时按关闭处理,直连失败不去白调
     let webFetchRenderEnabled = false;
+    // 老网关没有 /v1/image,image_gen 缺失时按关闭处理,不注册生图工具
+    let imageGenEnabled = false;
     const endpointsReady = loadCustomEndpoints(cfg);
     try {
         const { models, features } = await fetchModels(cfg);
         setModelsFromApi(models.map(apiModelToDef));
         webSearchEnabled = features.web_search !== false;
         webFetchRenderEnabled = features.web_fetch_render === true;
+        imageGenEnabled = features.image_gen === true;
     }
     catch (e) {
         console.error("  获取模型列表失败,使用内置列表:", e.message);
@@ -48,7 +51,11 @@ export async function prepareWebEnv(cfg) {
     // pi-web-ui 靠 auth.json 判断「已配置」,否则网页会弹 pi 安装引导
     ensureAuthCredential();
     // 联网工具经 agentDir/extensions 投影,和 TUI 共用一份注册
-    writeWebToolsExtension(cfg, { webSearch: webSearchEnabled, webFetchRender: webFetchRenderEnabled });
+    writeWebToolsExtension(cfg, {
+        webSearch: webSearchEnabled,
+        webFetchRender: webFetchRenderEnabled,
+        imageGen: imageGenEnabled,
+    });
     // 网页版新会话从 settings.json 的 defaultModel 取模型(TUI 是每次传 --model),
     // 确保它有值;resolvePreferredModel 优先尊重已有的 in-session 选择,不会回退覆盖。
     const pref = resolvePreferredModel(cfg);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "u1s1-cli",
-  "version": "0.13.7",
+  "version": "0.15.0",
   "description": "u1s1 — 有一说一,最省心的 AI 编程搭子。终端里用中文说需求,AI 帮你读文件、改代码、跑命令。",
   "type": "module",
   "bin": {
@@ -15,7 +15,8 @@
   },
   "files": [
     "dist",
-    "webui-dist"
+    "webui-dist",
+    "bench"
   ],
   "engines": {
     "node": ">=22.19.0"