niceeval 0.5.3 → 0.6.0

package/src/view/view-report.test.ts

@@ -1,24 +1,32 @@
-// niceeval view 的 --report 报告槽与宿主组合语义(docs/reports.md「宿主输入的组合语义」
+// niceeval view 的报告槽与宿主组合语义(docs/feature/reports/architecture.md「Selection 是计算入口」
 // 与裁决记录 6;公开行为准绳 docs-site/zh/guides/viewing-results.mdx / custom-reports.mdx)。
 // 覆盖:
-// - 组合语义与 show 对齐:位置前缀收窄报告槽选集、--experiment 过滤、匹配不到直说;
+// - 组合语义与 show 对齐:位置前缀收窄报告槽 Selection、--experiment 过滤、匹配不到直说;
 // - 单文件模式共存:存在的文件路径 → 单文件模式,目录报错直说走 --run,其余按 eval 前缀;
-// - 报告槽整槽替换:报告 HTML(含 <Style> 产物与证据室深链)烘进静态块,证据室数据原样保留;
-// - --out 静态导出:index.html 含报告块与官方样式,报告 HTML 零 <script>;
+// - 报告槽恒在:裸跑填充 CostPassRateComparison,--report 整槽替换;en / zh-CN 双语各渲染一遍;
+//   裸跑 ≡ --report <re-export CostPassRateComparison 的文件>(等价性);
+// - --out 静态导出:index.html 含两个语言的报告块、官方样式与增强 runtime,报告块零 <script>;
 // - dev server 装载语义:报告文件变更 → 下次装载整页重算(mtime cache-busting)。
+//
+// fixture 直接写新布局(<expDir>/<snapDir>/snapshot.json + <evalId>/a<n>/result.json),
+// 依据是 docs/feature/results/architecture.md 的稳定磁盘契约,不经 writer 运行时 API。
 
 import { mkdir, mkdtemp, readFile, rm, utimes, writeFile } from "node:fs/promises";
 import { existsSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { join, resolve } from "node:path";
 import { afterEach, describe, expect, it } from "vitest";
-import { ReportLoadError } from "../report/load.ts";
+// dist-sourced: this must be the exact class loadViewScan()/data.ts's loadReportFile() throws
+// (see src/view/data.ts's comment) — a raw-src import would be a structurally-identical but
+// `instanceof`-incompatible class.
+import { ReportLoadError } from "../../dist/report/load.js";
 import { ViewInputError, loadViewScan } from "./data.ts";
 import { buildView, resolveViewInput } from "./index.ts";
 import { runShow } from "../show/index.ts";
-import { RESULTS_FORMAT, RESULTS_SCHEMA_VERSION, type EvalResult, type RunSummary } from "../types.ts";
+import { RESULTS_FORMAT, RESULTS_SCHEMA_VERSION, type EvalResult, type Verdict } from "../types.ts";
 
 const EXAM_REPORT = resolve(__dirname, "../../test/fixtures/report/exam-report.tsx");
+const DEFAULT_REPORT_REEXPORT = resolve(__dirname, "../../test/fixtures/report/default-report-reexport.tsx");
 
 const roots: string[] = [];
 async function makeRoot(): Promise<string> {
@@ -30,66 +38,58 @@ afterEach(async () => {
   await Promise.all(roots.splice(0).map((r) => rm(r, { recursive: true, force: true })));
 });
 
-function res(over: Partial<EvalResult> & Pick<EvalResult, "id">): EvalResult {
-  return { agent: "bub", outcome: "passed", attempt: 0, durationMs: 1000, assertions: [], ...over };
+type AttemptFixture = Pick<EvalResult, "id" | "verdict"> & Partial<Pick<EvalResult, "attempt" | "durationMs" | "assertions">>;
+
+function res(id: string, verdict: Verdict, extra: Partial<AttemptFixture> = {}): AttemptFixture {
+  return { id, verdict, attempt: 0, durationMs: 1000, assertions: [], ...extra };
+}
+
+interface SnapshotOpts {
+  experimentId: string;
+  agent?: string;
+  startedAt: string;
 }
 
-function summaryOf(results: EvalResult[], over: Partial<RunSummary> = {}): RunSummary {
-  const count = (o: EvalResult["outcome"]) => results.filter((r) => r.outcome === o).length;
-  return {
+/** 写一份新布局快照:snapshot.json + 各 attempt 的 result.json。返回快照目录绝对路径。 */
+async function writeSnapshot(
+  root: string,
+  expDirName: string,
+  snapDirName: string,
+  opts: SnapshotOpts,
+  results: AttemptFixture[],
+): Promise<string> {
+  const dir = join(root, expDirName, snapDirName);
+  await mkdir(dir, { recursive: true });
+  const meta = {
     format: RESULTS_FORMAT,
     schemaVersion: RESULTS_SCHEMA_VERSION,
     producer: { name: "niceeval", version: "0.4.6" },
-    agent: results[0]?.agent ?? "bub",
-    startedAt: "2026-07-08T10:00:00.000Z",
-    completedAt: "2026-07-08T10:10:00.000Z",
-    passed: count("passed"),
-    failed: count("failed"),
-    skipped: count("skipped"),
-    errored: count("errored"),
-    durationMs: 60_000,
-    results,
-    ...over,
+    experimentId: opts.experimentId,
+    agent: opts.agent ?? "bub",
+    startedAt: opts.startedAt,
+    completedAt: opts.startedAt,
   };
-}
-
-async function writeRun(root: string, dirName: string, summary: RunSummary): Promise<string> {
-  const dir = join(root, dirName);
-  await mkdir(dir, { recursive: true });
-  await writeFile(join(dir, "summary.json"), JSON.stringify(summary, null, 2), "utf-8");
+  await writeFile(join(dir, "snapshot.json"), JSON.stringify(meta, null, 2), "utf-8");
+  for (const r of results) {
+    const attemptDir = join(dir, r.id, `a${r.attempt ?? 0}`);
+    await mkdir(attemptDir, { recursive: true });
+    await writeFile(join(attemptDir, "result.json"), JSON.stringify(r, null, 2), "utf-8");
+  }
   return dir;
 }
 
 /** 两个实验、weather 通过 + button 失败:报告槽/证据室/深链断言都要用到失败案例。 */
 async function seedRoot(): Promise<string> {
   const root = await makeRoot();
-  await writeRun(
-    root,
-    "2026-07-08T10-00-00-000Z",
-    summaryOf(
-      [
-        res({ id: "weather/brooklyn", experimentId: "compare/bub", startedAt: "2026-07-08T10:00:01.000Z" }),
-        res({
-          id: "fixtures/button",
-          experimentId: "compare/bub",
-          outcome: "failed",
-          startedAt: "2026-07-08T10:00:02.000Z",
-          assertions: [
-            { name: 'fileChanged("Button.tsx")', severity: "gate", score: 0, passed: false },
-          ],
-        }),
-      ],
-      { startedAt: "2026-07-08T10:00:00.000Z" },
-    ),
-  );
-  await writeRun(
-    root,
-    "2026-07-09T10-00-00-000Z",
-    summaryOf(
-      [res({ id: "weather/brooklyn", experimentId: "compare/codex", agent: "codex", startedAt: "2026-07-09T10:00:01.000Z" })],
-      { agent: "codex", startedAt: "2026-07-09T10:00:00.000Z" },
-    ),
-  );
+  await writeSnapshot(root, "compare_bub", "2026-07-08T10-00-00-000Z", { experimentId: "compare/bub", agent: "bub", startedAt: "2026-07-08T10:00:00.000Z" }, [
+    res("weather/brooklyn", "passed"),
+    res("fixtures/button", "failed", {
+      assertions: [{ name: 'fileChanged("Button.tsx")', severity: "gate", score: 0, passed: false }],
+    }),
+  ]);
+  await writeSnapshot(root, "compare_codex", "2026-07-09T10-00-00-000Z", { experimentId: "compare/codex", agent: "codex", startedAt: "2026-07-09T10:00:00.000Z" }, [
+    res("weather/brooklyn", "passed"),
+  ]);
   return root;
 }
 
@@ -98,7 +98,7 @@ async function seedRoot(): Promise<string> {
 describe("resolveViewInput · 位置参数语义", () => {
   it("存在的文件路径 → 单文件模式;其余位置参数按 eval id 前缀", async () => {
     const root = await seedRoot();
-    const file = join(root, "2026-07-08T10-00-00-000Z", "summary.json");
+    const file = join(root, "compare_bub", "2026-07-08T10-00-00-000Z", "snapshot.json");
     expect(resolveViewInput(root, [file])).toEqual({ input: file, patterns: [] });
     expect(resolveViewInput(root, ["weather", "fixtures/button"])).toEqual({
       patterns: ["weather", "fixtures/button"],
@@ -108,13 +108,13 @@ describe("resolveViewInput · 位置参数语义", () => {
 
   it("目录位置参数:报错直说走 --run(位置参数留给 eval id 前缀)", async () => {
     const root = await seedRoot();
-    expect(() => resolveViewInput(root, ["2026-07-08T10-00-00-000Z"])).toThrow(ViewInputError);
-    expect(() => resolveViewInput(root, ["2026-07-08T10-00-00-000Z"])).toThrow(/--run/);
+    expect(() => resolveViewInput(root, ["compare_bub"])).toThrow(ViewInputError);
+    expect(() => resolveViewInput(root, ["compare_bub"])).toThrow(/--run/);
   });
 
   it("单文件模式不与其它位置参数或 --run 混用,歧义报错直说", async () => {
     const root = await seedRoot();
-    const file = join(root, "2026-07-08T10-00-00-000Z", "summary.json");
+    const file = join(root, "compare_bub", "2026-07-08T10-00-00-000Z", "snapshot.json");
     expect(() => resolveViewInput(root, [file, "weather"])).toThrow(/exactly one path/);
     expect(() => resolveViewInput(root, [file], root)).toThrow(/--run/);
   });
@@ -133,24 +133,23 @@ describe("resolveViewInput · 位置参数语义", () => {
 // ───────────────────────── 组合语义(与 show 对齐) ─────────────────────────
 
 describe("loadViewScan · 组合语义", () => {
-  it("位置前缀收窄报告槽选集(榜单/overview);证据室快照不收窄,深链恒可达", async () => {
+  it("位置前缀收窄报告槽 Selection;证据室快照不收窄,深链恒可达", async () => {
     const root = await seedRoot();
-    const { viewData } = await loadViewScan(root, { patterns: ["weather"] });
-    // 报告槽:两实验都只剩 weather/brooklyn 一题,全过。
-    expect(viewData.overview.totals.evals).toBe(1);
-    expect(viewData.overview.totals.attempts).toBe(2); // 2 个实验各 1 attempt
-    for (const row of viewData.table.rows) {
-      expect(row.cells["pass-rate"]!.value).toBe(1);
-    }
+    const { viewData, reportHtml } = await loadViewScan(root, { patterns: ["weather"] });
+    // 报告槽:两实验都只剩 weather/brooklyn 一题,范围外的失败不再出现。
+    expect(reportHtml.en).toContain("compare/bub");
+    expect(reportHtml.en).toContain("compare/codex");
+    expect(reportHtml.en).not.toContain("fixtures/button");
     // 证据室:快照明细仍含 fixtures/button(attempt 深链在收窄下也能解析)。
     const allIds = viewData.snapshots.flatMap((s) => s.results.map((r) => r.id));
     expect(allIds).toContain("fixtures/button");
   });
 
-  it("--experiment 过滤:选集只留该实验", async () => {
+  it("--experiment 过滤:报告槽 Selection 只留该实验", async () => {
     const root = await seedRoot();
-    const { viewData } = await loadViewScan(root, { experiment: "compare/codex" });
-    expect(viewData.table.rows.map((r) => r.key)).toEqual(["compare/codex"]);
+    const { reportHtml } = await loadViewScan(root, { experiment: "compare/codex" });
+    expect(reportHtml.en).toContain("compare/codex");
+    expect(reportHtml.en).not.toContain("compare/bub");
   });
 
   it("前缀/实验匹配不到:直说,不渲染空页面", async () => {
@@ -160,11 +159,51 @@ describe("loadViewScan · 组合语义", () => {
     await expect(loadViewScan(root, { experiment: "nosuch" })).rejects.toBeInstanceOf(ViewInputError);
   });
 
-  it("全部缺省:默认行为不变(results.latest() 口径,不走合成)", async () => {
+  it("全部缺省:报告槽使用现刻水位口径(与 show 相同的合成规则)", async () => {
+    const root = await seedRoot();
+    const { reportHtml } = await loadViewScan(root);
+    // 默认报告的榜单含两个实验;官方组件的稳定类名在场。
+    expect(reportHtml.en).toContain("compare/bub");
+    expect(reportHtml.en).toContain("compare/codex");
+    expect(reportHtml.en).toContain("nre-");
+  });
+});
+
+// ─────────────────── 报告槽恒在:裸跑 ≡ --report <CostPassRateComparison> ───────────────────
+
+describe("loadViewScan · 默认报告槽(裸跑)", () => {
+  it("裸跑产出的报告槽 HTML 与 --report <re-export CostPassRateComparison 的文件> 完全一致(双语)", async () => {
+    const root = await seedRoot();
+    const bare = await loadViewScan(root);
+    const viaReport = await loadViewScan(root, {
+      report: { path: DEFAULT_REPORT_REEXPORT, cwd: root },
+    });
+    expect(bare.reportHtml.en).toBe(viaReport.reportHtml.en);
+    expect(bare.reportHtml["zh-CN"]).toBe(viaReport.reportHtml["zh-CN"]);
+  });
+
+  it("报告槽双语渲染:同一棵树按 locale 渲染两遍,chrome 文案分语言、数据不分语言", async () => {
+    const root = await seedRoot();
+    const { reportHtml } = await loadViewScan(root);
+    expect(reportHtml.en).toContain("Pass rate"); // ExperimentList 主行(en)
+    expect(reportHtml["zh-CN"]).toContain("成功率"); // ExperimentList 主行(zh-CN,passRate 的 zh label)
+    for (const html of [reportHtml.en, reportHtml["zh-CN"]]) {
+      expect(html).toContain("compare/bub");
+      // 失败案例深链进证据室:不透明 AttemptLocator 单段路由 `#/attempt/@<locator>`,
+      // 不再是旧的两段式 `#/attempt/<snapshot>/<attempt>`。
+      expect(html).toMatch(/href="#\/attempt\/@[0-9a-z]+"/);
+      expect(html).not.toContain("<script"); // 报告槽产物零客户端 JS,不 hydrate
+    }
+  });
+
+  it("失败清单与警告住在报告槽里:ExperimentList 列出失败,壳的 viewData 不携带统计产物", async () => {
     const root = await seedRoot();
     const { viewData, reportHtml } = await loadViewScan(root);
-    expect(reportHtml).toBeUndefined();
-    expect(viewData.table.rows.map((r) => r.key).sort()).toEqual(["compare/bub", "compare/codex"]);
+    expect(reportHtml.en).toContain("fixtures/button");
+    // viewData 只有证据室数据:快照 + skipped + 壳元信息。
+    expect(viewData).not.toHaveProperty("overview");
+    expect(viewData).not.toHaveProperty("table");
+    expect(viewData).not.toHaveProperty("overall");
   });
 });
 
@@ -174,31 +213,33 @@ describe("loadViewScan · --report 报告槽", () => {
   it("报告树渲染为静态 HTML:官方水位 + 自定义摆法 + <Style> 产物 + 证据室深链,零 <script>", async () => {
     const root = await seedRoot();
     const scan = await loadViewScan(root, { report: { path: EXAM_REPORT, cwd: root } });
-    const html = scan.reportHtml!;
+    const html = scan.reportHtml.en;
     expect(html).toContain("考试成绩单"); // 自定义 Section
     expect(html).toContain("nre-"); // 官方组件的稳定类名
     expect(html).toContain("<style>.exam-note { color: #4a7; }</style>"); // <Style> 随树带走
-    expect(html).toContain("#/attempt/2026-07-08T10-00-00-000Z/"); // 失败案例深链进证据室
+    expect(html).toMatch(/href="#\/attempt\/@[0-9a-z]+"/); // 失败案例深链进证据室(单段 locator 路由)
     expect(html).not.toContain("<script"); // 报告槽产物零客户端 JS,不 hydrate
-    // 证据室数据契约(__NICEEVAL_VIEW_DATA__)原样保留:快照、attemptRef、skipped 不动。
+    // 用户报告同样双语渲染两遍(壳按界面语言摆放)。
+    expect(scan.reportHtml["zh-CN"]).toContain("考试成绩单");
+    // 证据室数据契约(__NICEEVAL_VIEW_DATA__)原样保留:快照、locator、skipped 不动。
     expect(scan.viewData.snapshots.length).toBeGreaterThan(0);
-    expect(scan.viewData.snapshots.flatMap((s) => s.results).every((r) => r.attemptRef)).toBe(true);
+    expect(scan.viewData.snapshots.flatMap((s) => s.results).every((r) => r.locator)).toBe(true);
     expect(JSON.stringify(scan.viewData)).not.toContain("考试成绩单"); // 报告块不进 viewData
   });
 
-  it("位置前缀对 --report 生效:收窄注入选集,报告只见范围内的 eval", async () => {
+  it("位置前缀对 --report 生效:收窄注入 Selection,报告只见范围内的 eval", async () => {
     const root = await seedRoot();
     const scan = await loadViewScan(root, {
       patterns: ["weather"],
       report: { path: EXAM_REPORT, cwd: root },
     });
     // 范围外的失败(fixtures/button)不再出现在报告里;范围内的实验行都在。
-    expect(scan.reportHtml).not.toContain("fixtures/button");
-    expect(scan.reportHtml).toContain("compare/bub");
-    expect(scan.reportHtml).toContain("compare/codex");
+    expect(scan.reportHtml.en).not.toContain("fixtures/button");
+    expect(scan.reportHtml.en).toContain("compare/bub");
+    expect(scan.reportHtml.en).toContain("compare/codex");
   });
 
-  it("show --report 与 view --report 吃同一个报告文件,判决口径一致", async () => {
+  it("show --report 与 view --report 吃同一个报告文件,判定口径一致", async () => {
     const root = await seedRoot();
     let text = "";
     const code = await runShow(root, [], { run: root, report: EXAM_REPORT }, {
@@ -211,7 +252,7 @@ describe("loadViewScan · --report 报告槽", () => {
     // 同一棵树的两个面:同一份失败清单与同一个自定义 Section。
     for (const needle of ["考试成绩单", "compare/bub", "compare/codex", "fixtures/button"]) {
       expect(text).toContain(needle);
-      expect(scan.reportHtml).toContain(needle);
+      expect(scan.reportHtml.en).toContain(needle);
     }
   });
 
@@ -258,51 +299,58 @@ describe("loadViewScan · 报告文件变更整页重算", () => {
     const path = join(root, "report.mjs");
     await writeFile(path, reportSource("FIRST_RENDER"), "utf-8");
     const first = await loadViewScan(root, { report: { path, cwd: root } });
-    expect(first.reportHtml).toContain("FIRST_RENDER");
+    expect(first.reportHtml.en).toContain("FIRST_RENDER");
 
     await writeFile(path, reportSource("SECOND_RENDER"), "utf-8");
     // mtime 精度兜底:显式把 mtime 拨到未来,确保与首次装载可区分。
     const future = new Date(Date.now() + 5000);
     await utimes(path, future, future);
     const second = await loadViewScan(root, { report: { path, cwd: root } });
-    expect(second.reportHtml).toContain("SECOND_RENDER");
-    expect(second.reportHtml).not.toContain("FIRST_RENDER");
+    expect(second.reportHtml.en).toContain("SECOND_RENDER");
+    expect(second.reportHtml.en).not.toContain("FIRST_RENDER");
   });
 });
 
 // ───────────────────────── --out 静态导出 ─────────────────────────
 
 describe("buildView · --out 与 --report", () => {
-  it("报告页为首页报告槽,证据室同站:index.html 含报告块与官方样式,工件照常复制", async () => {
+  it("报告页为首页报告槽,证据室同站:index.html 含报告块与官方样式,artifact 照常复制", async () => {
     const root = await seedRoot();
-    // 给 weather attempt 一份 events 工件,验证证据室工件照常进导出。
-    const artifactDir = join(root, "2026-07-09T10-00-00-000Z", "artifacts", "weather");
+    // 给 weather attempt 一份 events artifact,验证证据室 artifact 照常进导出。
+    // 本快照跑出的条目:artifact 目录 = <snapshot.dir>/<evalId>/a<n>,不需要声明字段。
+    const artifactDir = join(root, "compare_codex", "2026-07-09T10-00-00-000Z", "weather", "brooklyn", "a0");
     await mkdir(artifactDir, { recursive: true });
     await writeFile(join(artifactDir, "events.json"), "[]", "utf-8");
-    const summaryPath = join(root, "2026-07-09T10-00-00-000Z", "summary.json");
-    const summary = JSON.parse(await readFile(summaryPath, "utf-8")) as RunSummary;
-    summary.results[0]!.artifactsDir = "artifacts/weather";
-    await writeFile(summaryPath, JSON.stringify(summary), "utf-8");
 
     const out = join(root, "site");
     await buildView({ input: root, out, scan: { report: { path: EXAM_REPORT, cwd: root } } });
 
     const html = await readFile(join(out, "index.html"), "utf-8");
-    expect(html).toContain('<template id="niceeval-report">');
+    // 双语两个 <template> 静态块都在,壳按界面语言摆放。
+    expect(html).toContain('<template id="niceeval-report-en">');
+    expect(html).toContain('<template id="niceeval-report-zh-CN">');
     expect(html).toContain("考试成绩单");
     expect(html).toContain("nre-"); // 官方组件样式(report/react/styles.css)随页注入
     // 报告块本体零 <script>:静态块起于 <template>,内部只有标记与 <style>。
-    const block = html.split('<template id="niceeval-report">')[1]!.split("</template>")[0]!;
+    const block = html.split('<template id="niceeval-report-en">')[1]!.split("</template>")[0]!;
     expect(block).not.toContain("<script");
-    // 证据室同站:工件按 /artifact/<base>/ 布局复制。
-    expect(existsSync(join(out, "artifact", "2026-07-09T10-00-00-000Z/artifacts/weather", "events.json"))).toBe(true);
+    // 证据室同站: artifact 按 /artifact/<base>/ 布局复制。
+    expect(
+      existsSync(join(out, "artifact", "compare_codex/2026-07-09T10-00-00-000Z/weather/brooklyn/a0", "events.json")),
+    ).toBe(true);
   });
 
-  it("默认导出(无 --report)不含报告块,行为不变", async () => {
+  it("默认导出(无 --report):报告槽填充 CostPassRateComparison,双语块与增强 runtime 恒内联", async () => {
     const root = await seedRoot();
     const out = join(root, "site");
     await buildView({ input: root, out });
     const html = await readFile(join(out, "index.html"), "utf-8");
-    expect(html).not.toContain('<template id="niceeval-report">');
+    expect(html).toContain('<template id="niceeval-report-en">');
+    expect(html).toContain('<template id="niceeval-report-zh-CN">');
+    expect(html).toContain("nre-"); // 官方组件的稳定类名(KPI / 榜单 / 散点来自 nre 组件)
+    // 渐进增强 runtime(排序 / 过滤 / tooltip)恒内联;报告块本体仍零 <script>。
+    expect(html).toContain("__nreEnhanced");
+    const block = html.split('<template id="niceeval-report-en">')[1]!.split("</template>")[0]!;
+    expect(block).not.toContain("<script");
   });
 });

package/docs-site/zh/guides/sandbox-backends.mdx

@@ -1,75 +0,0 @@
----
-title: "Sandbox 后端：Docker、Vercel 与第三方"
-sidebarTitle: "沙箱后端"
-description: "NiceEval 在 Docker 或 Vercel sandbox 中运行 coding agents。了解如何选择后端、配置权限，并通过 warm pools 改善性能。"
----
-
-Sandbox backend 是创建和管理隔离运行环境的基础设施。[NiceEval](https://niceeval.com/) 把它们包装成同一个 `Sandbox` 接口，所以 adapter 不需要知道当前用的是本地 Docker、Vercel micro-VM 还是第三方云服务。
-
-## `Sandbox` 接口
-
-Adapter 常用操作包括：
-
-| 方法 | 用途 |
-|---|---|
-| `runCommand(cmd, args)` | 运行命令 |
-| `runShell(script)` | 运行 shell 脚本 |
-| `readFile(path)` | 读取文件 |
-| `writeFiles(files)` | 写入一组文件 |
-| `uploadFiles(files)` | 上传 workspace 或测试文件 |
-| `workdir` | 后端真实工作目录；省略的 `cwd` / `targetDir` 都解析到这里 |
-| `runCommand(..., { cwd })` | 单条命令临时切换工作目录；相对路径按 `workdir` 解析 |
-| `stop()` | 销毁环境 |
-
-## 选择后端
-
-选择后端只能写在代码里——没有对应的 CLI flag，[NiceEval](https://niceeval.com/) 也不会自动探测后端。在 experiment 里设置 `sandbox` 字段（或者在 `niceeval.config.ts` 里设置作为项目级兜底），三选一：
-
-```ts
-// experiments/local.ts
-import { defineExperiment } from "niceeval";
-import { dockerSandbox } from "niceeval/sandbox";
-
-export default defineExperiment({
-  agent: myCodingAgent,
-  model: "claude-sonnet-4-6",
-  sandbox: dockerSandbox(), // 或 vercelSandbox() / e2bSandbox()
-});
-```
-
-如果 experiment 和 `niceeval.config.ts` 都没有设置 `sandbox`，[NiceEval](https://niceeval.com/) 在创建 sandbox 时会直接报错，而不会猜一个后端。
-
-三个内置后端的 SDK 不随 NiceEval 一起安装——用哪个就装哪个，避免把用不到的依赖（以及它们的原生构建脚本）带进你的项目：
-
-| 后端 | 安装命令 |
-|---|---|
-| `dockerSandbox()` | `pnpm add dockerode @types/dockerode` |
-| `vercelSandbox()` | `pnpm add @vercel/sandbox` |
-| `e2bSandbox()` | `pnpm add e2b` |
-
-漏装时不会静默失败：NiceEval 在创建 sandbox 的那一刻报错并直接给出上面的安装命令，例如 `Docker sandbox requires 'dockerode'. Install it with: pnpm add dockerode @types/dockerode`。
-
-## Docker
-
-Docker 适合本地开发和标准 CI。优点是简单、可控、无云端依赖；缺点是机器资源有限，冷启动和安装依赖可能较慢。
-
-## Vercel Sandbox
-
-Vercel Sandbox 适合需要云端隔离、更多资源或更稳定环境的任务。需要相应 token 或 OIDC 配置。
-
-## 第三方后端
-
-只要实现 `Sandbox` 接口，就可以接入其他 sandbox provider。adapter 仍然只依赖接口，不依赖 provider 私有 API。
-
-## 权限和 root
-
-不同后端对 root 权限、网络、文件系统和进程生命周期的约束不同。编写 fixture 时尽量避免依赖宿主机环境，把依赖写进 `package.json` 或 fixture setup。
-
-## 性能建议
-
-- 减少 fixture 依赖体积。
-- 使用缓存或预热机制。
-- 控制 `sandboxConcurrency`，避免本地 Docker 资源耗尽。
-- 把慢测试拆成必要的 gate 和可选的 soft 检查。
-
-Warm pools 和复用属于 runner / scheduler 层面的能力，详见 [Runner](/zh/guides/runner)。