niceeval 0.6.0 → 0.6.1

package/src/report/dual-render.test.tsx

@@ -41,7 +41,7 @@ import {
   renderReportToText,
   stringWidth,
 } from "./index.ts";
-import { CostPassRateComparison } from "./built-ins/index.ts";
+import { ExperimentComparison } from "./built-ins/index.ts";
 import { renderReportToStaticHtml } from "./web.ts";
 import { createTextContext, renderNodeToText, validateReportTree } from "./tree.ts";
 import {
@@ -462,7 +462,7 @@ describe("AttemptList 双面", () => {
   const html = renderToStaticMarkup(<AttemptList items={attemptListItems} />);
   const term = text(<AttemptList items={attemptListItems} />);
 
-  it("两面同口径:判定符 + locator + 证据能力标记 + 断言/error 明细 + 下钻命令一致", () => {
+  it("两面同口径:判定符 + locator + 断言/error 明细 + 下钻命令一致", () => {
     for (const piece of ["roots-correct", "expected x=2, got x=3", "TypeError", "@1a4a4a4", "@1c1c1c1"]) {
       expect(html).toContain(piece);
       expect(term).toContain(piece);
@@ -470,9 +470,8 @@ describe("AttemptList 双面", () => {
     // web 面走证据室路由(#/attempt/@<locator>);text 面只列 locator 本身,不重复整条命令
     // (docs/feature/reports/architecture.md「text 输出只在整份报告末尾给一次命令模板」)。
     expect(html).toContain('href="#/attempt/@1a4a4a4"');
-    // 证据能力标记两面一致:failedAttempt 有 eval/execution/timing,erroredAttempt 只有 execution
-    expect(term).toContain("[E,X,⏱]");
-    expect(html).toContain("[E,X,⏱]");
+    expect(term).not.toMatch(/\[[EXD⏱,]+\]/);
+    expect(html).not.toMatch(/\[[EXD⏱,]+\]/);
   });
 
   it("total > items.length 时两面都如实报剩余数量", () => {
@@ -541,10 +540,13 @@ describe("ExperimentList 双面", () => {
     expect(term).toContain("50%");
   });
 
-  it("text 面是逐条卡片,不是 renderTableText 的产物:空行分隔每个 experiment、展开到 eval 的行缩进——表格没有这层嵌套", () => {
+  it("text 面先给比较表,再按 experiment 给 Eval / Attempt 父子表", () => {
     const blocks = term.split("\n\n");
-    expect(blocks).toHaveLength(experimentListItems.length);
-    expect(term.split("\n").some((l) => /^  \S/.test(l))).toBe(true); // 展开到 eval 的行缩进两格
+    expect(blocks).toHaveLength(experimentListItems.length + 1);
+    expect(blocks[0]).toContain("Model");
+    expect(blocks[0]).toContain("Avg");
+    expect(blocks[1]!.match(/algebra\/quadratic/g)).toHaveLength(1);
+    expect(blocks[1]).toMatch(/✗ failed\s+algebra\/quadratic[\s\S]*├─ @1a4a4a4[\s\S]*└─ @1b5b5b5/);
   });
 });
 
@@ -987,11 +989,11 @@ describe("defineReport + 渲染入口", () => {
     const out = await renderReportToText(report, fakeContext(), { width: 100 });
     // RunOverview(data 形态,预计算)
     expect(out).toContain("1 experiment · 2 evals · 2 attempts");
-    // ExperimentList:主行 + eval 级折叠计票 + 失败诊断(locator 徽标带证据能力标记)
-    expect(out).toContain("compare/bub · bub");
+    // ExperimentList:主行 + eval 级折叠计票 + 逐 Attempt locator 与失败诊断
+    expect(out).toMatch(/compare\/bub\s+default\s+bub/);
     expect(out).toContain("1 passed / 1 failed");
     expect(out).toContain("50%");
-    expect(out).toMatch(/✗ algebra\/y\s+@[0-9a-z]+✗/);
+    expect(out).toMatch(/✗ failed\s+algebra\/y[\s\S]*└─ @[0-9a-z]+/);
     // 自己的口径:成绩单
     expect(out).toContain("考试成绩单");
     expect(out).toContain("bub");
@@ -1051,7 +1053,7 @@ describe("defineReport + 渲染入口", () => {
 // ts-expect-error 注释对应一处编译错误。直接标注 Props 类型(而非 JSX)——JSX 的 union
 // 属性检查对「缺必填字段」较宽松,直接的类型赋值才严格钉住互斥不变量:同时传 data 与
 // selection、或两者都不传、或 selection 形态缺必填计算选项,全部编译失败。正向 JSX 用法由
-// CostPassRateComparison / defineReport 报告在同一文件里编译验证。
+// ExperimentComparison / defineReport 报告在同一文件里编译验证。
 function metricScatterPropsTypeChecks(selection: Selection, data: ScatterData): void {
   const ok1: MetricScatterProps = { data }; // 合法:data 形态
   const ok2: MetricScatterProps = { selection, points: "experiment", series: "agent", x: costUSD, y: passRate }; // 合法:selection 形态
@@ -1070,56 +1072,58 @@ function metricScatterPropsTypeChecks(selection: Selection, data: ScatterData):
 
 void metricScatterPropsTypeChecks;
 
-// ───────────────────────── CostPassRateComparison(内置默认报告)─────────────────────────
+// ───────────────────────── ExperimentComparison(内置默认报告)─────────────────────────
 
-describe("CostPassRateComparison", () => {
+describe("ExperimentComparison", () => {
   it("是普通 ReportDefinition;text 面 = 成本×通过率散点 + 实验列表,别无它物", async () => {
-    expect(isReportDefinition(CostPassRateComparison)).toBe(true);
-    const out = await renderReportToText(CostPassRateComparison, fakeContext(), { width: 100 });
+    expect(isReportDefinition(ExperimentComparison)).toBe(true);
+    const out = await renderReportToText(ExperimentComparison, fakeContext(), { width: 100 });
     // 散点:fakeContext 无成本数据 → 0 可画点,显式说明缺哪两个指标(而不是画一张空图)
     expect(out).toContain("No data to plot");
     expect(out).not.toContain("better → upper right");
     // 实验列表主行 + eval 级折叠计票 + 失败诊断(ExperimentList.data 在 build() 里直接 await)
-    expect(out).toContain("compare/bub · bub");
+    expect(out).toMatch(/compare\/bub\s+default\s+bub/);
     expect(out).toContain("1 passed / 1 failed");
     expect(out).toContain("50%");
-    expect(out).toMatch(/✗ algebra\/y\s+@[0-9a-z]+✗/);
+    expect(out).toMatch(/✗ failed\s+algebra\/y[\s\S]*└─ @[0-9a-z]+/);
     // 只有两个直接业务组件:没有 RunOverview / GroupSummary / Section 分组
     expect(out).not.toContain("Current verdicts");
   });
 
   it("web 面:散点空态 + 实验列表 <details> 展开区,无 <script>,无 Section 分组", async () => {
-    const html = await renderReportToStaticHtml(CostPassRateComparison, fakeContext());
+    const html = await renderReportToStaticHtml(ExperimentComparison, fakeContext());
     expect(html).toContain("nre-metric-scatter");
     expect(html).toContain("nre-scatter-empty"); // 0 可画点的空态
-    expect(html).toContain('<li class="nre-experiment-entry">');
+    expect(html).toContain('<details class="nre-experiment-entry">');
+    expect(html).toContain("nre-experiment-head");
+    expect(html).toContain('data-nre-experiment-filter=""');
     expect(html).toContain("nre-experiment-evals");
     expect(html).not.toContain("nre-section");
     expect(html).not.toContain("<script");
   });
 
   it("locale 变体:en / zh-CN 都渲染(chrome 分语言),散点空态两面同一事实", async () => {
-    const zhHtml = await renderReportToStaticHtml(CostPassRateComparison, fakeContext(), { locale: "zh-CN" });
+    const zhHtml = await renderReportToStaticHtml(ExperimentComparison, fakeContext(), { locale: "zh-CN" });
     expect(zhHtml).toContain("成功率"); // passRate 的 zh-CN label(ExperimentList 主行)
-    const enHtml = await renderReportToStaticHtml(CostPassRateComparison, fakeContext(), { locale: "en" });
+    const enHtml = await renderReportToStaticHtml(ExperimentComparison, fakeContext(), { locale: "en" });
     expect(enHtml).toContain("Pass rate");
-    const zhText = await renderReportToText(CostPassRateComparison, fakeContext(), { locale: "zh-CN" });
+    const zhText = await renderReportToText(ExperimentComparison, fakeContext(), { locale: "zh-CN" });
     expect(zhText).toContain("没有可绘制的数据"); // 散点空态 zh
-    const enText = await renderReportToText(CostPassRateComparison, fakeContext(), { locale: "en" });
+    const enText = await renderReportToText(ExperimentComparison, fakeContext(), { locale: "en" });
     expect(enText).toContain("No data to plot");
   });
 
   it("多实验 fixture:每个 experiment 一项,散点如实处理全部无成本", async () => {
-    const out = await renderReportToText(CostPassRateComparison, fakeMultiGroupContext(), { width: 100 });
+    const out = await renderReportToText(ExperimentComparison, fakeMultiGroupContext(), { width: 100 });
     // 三个 experiment 的身份行都出现(experimentId · agent,不再截短成最后一段)
-    expect(out).toContain("compare/bub · bub");
-    expect(out).toContain("other/codex · codex");
-    expect(out).toContain("solo · bub");
+    expect(out).toMatch(/compare\/bub\s+default\s+bub/);
+    expect(out).toMatch(/other\/codex\s+default\s+codex/);
+    expect(out).toMatch(/solo\s+default\s+bub/);
     // 散点空态(三点都无成本)
     expect(out).toContain("No data to plot");
     // 没有组分 Section 标题(不再按目录前缀分组)
     expect(out).not.toMatch(/^compare$/m);
-    const html = await renderReportToStaticHtml(CostPassRateComparison, fakeMultiGroupContext());
+    const html = await renderReportToStaticHtml(ExperimentComparison, fakeMultiGroupContext());
     expect(html).not.toContain("nre-section");
   });
 });

package/src/report/format.ts

@@ -2,7 +2,6 @@
 //   "%" → 87%    "ms" → 1.2s    "$" → $0.31    其余 → 1.2k 缩写(带 unit 后缀)
 // metric.display 可整体覆盖;这里只负责默认。
 
-import type { AttemptEvidenceCapabilities } from "../results/attempt-evidence.ts";
 import type { AssertionResult, Verdict } from "../types.ts";
 
 /** 一位小数、去掉无意义的 ".0" 尾巴。 */
@@ -78,8 +77,7 @@ export function formatPercent(ratio: number): string {
   return `${Math.round(ratio * 100)}%`;
 }
 
-// ── 实体列表(ExperimentList / EvalList / AttemptList)共用的判定符与证据能力位 ──
-// web / text 两面对同一个 AttemptListItem 给出同一个标记(docs-site「终端输出形成反馈闭环」)。
+// ── 实体列表(ExperimentList / EvalList / AttemptList)共用的判定符 ──
 
 /** passed / failed / errored / skipped 的判定符。 */
 export function verdictMark(verdict: Verdict): string {
@@ -95,19 +93,6 @@ export function verdictMark(verdict: Verdict): string {
   }
 }
 
-/**
- * 证据能力标记:`[E,X,⏱]` 式方括号列表,固定顺序 eval → execution → timing → diff,
- * 缺哪个不列哪个;四位全无时返回空串(没有证据可看,不画空 `[]`)。
- */
-export function capabilityBadge(capabilities: AttemptEvidenceCapabilities): string {
-  const marks: string[] = [];
-  if (capabilities.eval) marks.push("E");
-  if (capabilities.execution) marks.push("X");
-  if (capabilities.timing) marks.push("⏱");
-  if (capabilities.diff) marks.push("D");
-  return marks.length > 0 ? `[${marks.join(",")}]` : "";
-}
-
 /**
  * 一个 AttemptListItem 的失败原因摘要,按优先级取第一个在场的:`error` → 未通过的 gate 断言
  * (原始声明顺序,`name`,detail 在场则 `"name: detail"`,多条用「, 」连接)→ 缺席。与

package/src/report/index.ts

@@ -45,7 +45,7 @@ export {
 export type { ColumnAlign } from "./text/layout.ts";
 
 // 内置报告(show / view 裸跑时报告槽的出厂填充,一份普通 ReportDefinition,无 renderer 特权)
-export { CostPassRateComparison } from "./built-ins/index.ts";
+export { ExperimentComparison } from "./built-ins/index.ts";
 
 // locale:官方组件 chrome 文案的语言(en / zh-CN);指标 label 可按 locale 给字典
 export { DEFAULT_REPORT_LOCALE, resolveMetricLabel } from "./locale.ts";
@@ -95,7 +95,6 @@ export type {
 // 数据契约(组件的 data props)
 export type {
   Aggregator,
-  AttemptEvidenceCapabilities,
   AttemptListItem,
   AttemptLocator,
   DeltaData,

package/src/report/locale.ts

@@ -32,6 +32,19 @@ const en = {
   "table.viewBreakdown": "Per-eval breakdown",
   /** <Table> 的 locator 列表头(行带 locator 时自动追加)。 */
   "table.attempt": "attempt",
+  "experimentList.experiment": "Experiment",
+  "experimentList.avgDuration": "Avg duration",
+  "experimentList.passRate": "Pass rate",
+  "experimentList.tokens": "Tokens",
+  "experimentList.estimatedCost": "Est. cost",
+  "experimentList.result": "Result",
+  "experimentList.status": "Status",
+  "experimentList.evalAttempt": "Eval / Attempt",
+  "experimentList.duration": "Duration",
+  "experimentList.cost": "Cost",
+  "experimentList.filterPlaceholder": "Filter experiments…",
+  "experimentList.defaultModel": "default",
+  "experimentList.flags": "Flags",
   /** <Table> 压到下限仍放不下时,从右侧丢列并如实报数。 */
   "table.columnsHidden.one": "({n} more column not shown)",
   "table.columnsHidden.other": "({n} more columns not shown)",
@@ -60,8 +73,6 @@ const en = {
   "scatter.axisReversed": "(axis reversed: right = better)",
   /** 0 个可画点:x/y 指标没有可用数据。 */
   "scatter.noData": "No data to plot {x} × {y}",
-  /** 恰好 1 个可画点:比较至少要两个实验。 */
-  "scatter.needTwo": "At least 2 experiments needed to compare {x} × {y}",
   "pointsMissing.one": "{n} point missing data",
   "pointsMissing.other": "{n} points missing data",
 
@@ -110,6 +121,19 @@ const zhCN: Record<ReportMessageKey, string> = {
   "table.reason": "原因",
   "table.viewBreakdown": "逐题明细",
   "table.attempt": "Attempt",
+  "experimentList.experiment": "实验",
+  "experimentList.avgDuration": "平均耗时",
+  "experimentList.passRate": "成功率",
+  "experimentList.tokens": "Tokens",
+  "experimentList.estimatedCost": "预估成本",
+  "experimentList.result": "结果",
+  "experimentList.status": "状态",
+  "experimentList.evalAttempt": "题目 / Attempt",
+  "experimentList.duration": "耗时",
+  "experimentList.cost": "成本",
+  "experimentList.filterPlaceholder": "筛选实验…",
+  "experimentList.defaultModel": "默认",
+  "experimentList.flags": "Flags",
   "table.columnsHidden.one": "(还有 {n} 列未列出)",
   "table.columnsHidden.other": "(还有 {n} 列未列出)",
 
@@ -135,7 +159,6 @@ const zhCN: Record<ReportMessageKey, string> = {
   "scatter.betterUpperRight": "更好 → 右上",
   "scatter.axisReversed": "(轴反向:右 = 更好)",
   "scatter.noData": "{x} × {y} 没有可绘制的数据",
-  "scatter.needTwo": "至少需要 2 个实验才能比较 {x} × {y}",
   "pointsMissing.one": "{n} 个点缺数据",
   "pointsMissing.other": "{n} 个点缺数据",
 

package/src/report/metrics.ts

@@ -13,7 +13,7 @@
 //
 // 两档指标(docs/feature/reports/library.md「内置指标」):以上除 turns 外全部只读 attempt.result
 // 的瘦身字段——任何 producer、任何 copySnapshots artifacts 选择都算得出,内置报告
-// CostPassRateComparison 只用这一档。turns 读 attempt.o11y()(懒加载 artifact),发布时若
+// ExperimentComparison 只用这一档。turns 读 attempt.o11y()(懒加载 artifact),发布时若
 // o11y 没随行就诚实渲染缺数据「—」,不算 0——报告作者自己摆时心里要有这根弦,内置报告不用它。
 
 import type { EvalResult } from "../types.ts";

package/src/report/react/AttemptList.tsx

@@ -1,5 +1,5 @@
 // AttemptList:实体列表的叶子层——每项一个 Attempt,固定展示判定、断言、error、Judge 评语
-// (assertions 的 detail/evidence)与证据引用(locator + 证据能力标记)。它不预设只看失败,
+// (assertions 的 detail/evidence)与证据引用(locator)。它不预设只看失败,
 // 报告作者过滤 items、用 .slice() 限量,total 让渲染面如实报告剩余数量,不静默截断。
 // ExperimentList / EvalList 的下钻数组是同一个 AttemptListItem[],这里的渲染逻辑因此
 // 也是它们展开区里"逐条 attempt"那一层的唯一实现(通过 AttemptRow 导出复用,不重写一遍)。
@@ -9,17 +9,16 @@ import type { AttemptListItem } from "../types.ts";
 import type { AttemptLocator } from "../../results/locator.ts";
 import { DEFAULT_REPORT_LOCALE, localeText, type ReportLocale } from "../locale.ts";
 import { colorClassForKey } from "./colors.ts";
-import { capabilityBadge, cx, formatDurationMs, formatUSD, verdictMark } from "./format.ts";
+import { cx, formatDurationMs, formatUSD, verdictMark } from "./format.ts";
 
-/** locator + 判定符 + 证据能力标记的普通 <a>,AttemptList/EvalList/ExperimentList 共用。 */
+/** locator + 判定符的普通 <a>,AttemptList/EvalList/ExperimentList 共用。 */
 export function AttemptLocatorBadge({
   item,
   attemptHref,
 }: {
-  item: Pick<AttemptListItem, "locator" | "verdict" | "capabilities">;
+  item: Pick<AttemptListItem, "locator" | "verdict">;
   attemptHref: (locator: AttemptLocator) => string;
 }): ReactElement {
-  const caps = capabilityBadge(item.capabilities);
   return (
     <a
       className={cx("nre-locator", `nre-verdict-${item.verdict}`)}
@@ -27,7 +26,6 @@ export function AttemptLocatorBadge({
     >
       {item.locator}
       <span className="nre-locator-mark">{verdictMark(item.verdict)}</span>
-      {caps && <span className="nre-locator-caps">{caps}</span>}
     </a>
   );
 }

package/src/report/react/ExperimentList.tsx

@@ -1,47 +1,119 @@
-// ExperimentList:实体列表的第一级——每项固定代表一个 experiment,主行是身份(experiment id /
-// agent / model / flags)、Eval 判定构成与官方两级聚合汇总指标(passRate/cost/duration/tokens,
-// 复用 MetricCellView 与其它指标图形同一份渲染)。零 JS 用原生 <details> 展开到这个 experiment
-// 的 Eval:每道题一行,判定符 + 该题全部 Attempt 的 locator 徽标(内联,不用再点开一层)+
-// 通过时的平均耗时/成本、失败/出错时的原因摘要。
+// ExperimentList:实体列表的第一级。web 面是一行一个 experiment 的固定列比较表；
+// 每行用原生 <details> 展开到 Eval 与 Attempt locator。数据仍完全来自
+// ExperimentList.data(),不恢复旧 ExperimentTable 的混合实体计算层。
 
 import type { ReactElement } from "react";
-import type { ExperimentListEvalRow, ExperimentListItem } from "../types.ts";
 import type { AttemptLocator } from "../../results/locator.ts";
+import type { AttemptListItem, ExperimentListEvalRow, ExperimentListItem } from "../types.ts";
+import { attemptItemReason } from "../format.ts";
 import { DEFAULT_REPORT_LOCALE, localeText, type ReportLocale } from "../locale.ts";
-import { MetricCellView } from "./cell.tsx";
 import { AttemptLocatorBadge } from "./AttemptList.tsx";
+import { MetricCellView } from "./cell.tsx";
 import { colorClassForKey } from "./colors.ts";
 import { cx, formatDurationMs, formatUSD, verdictMark } from "./format.ts";
 
-function EvalRow({
-  row,
+const verdictOrder = ["passed", "failed", "errored", "skipped"] as const;
+
+function passRateTone(value: number | null): string | undefined {
+  if (value === null) return undefined;
+  if (value >= 0.8) return "nre-good";
+  if (value >= 0.5) return "nre-warn";
+  return "nre-bad";
+}
+
+function formatDate(value: string, locale: ReportLocale): string {
+  const date = new Date(value);
+  if (Number.isNaN(date.valueOf())) return value;
+  return new Intl.DateTimeFormat(locale === "zh-CN" ? "zh-CN" : "en", {
+    month: "short",
+    day: "numeric",
+    hour: "2-digit",
+    minute: "2-digit",
+  }).format(date);
+}
+
+function VerdictSummary({ item, locale }: { item: ExperimentListItem; locale: ReportLocale }): ReactElement {
+  const parts = verdictOrder
+    .filter((verdict) => item.verdicts[verdict] > 0)
+    .map((verdict) => `${item.verdicts[verdict]} ${localeText(locale, `verdict.${verdict}`)}`);
+  return <span className="nre-experiment-pill">{parts.join(" / ") || "—"}</span>;
+}
+
+function ExperimentAttemptRow({
+  attempt,
+  last,
   attemptHref,
 }: {
-  row: ExperimentListEvalRow;
+  attempt: AttemptListItem;
+  last: boolean;
   attemptHref: (locator: AttemptLocator) => string;
 }): ReactElement {
+  const reason = attemptItemReason(attempt);
   return (
-    <li className={cx("nre-experiment-eval-row", `nre-eval-${row.verdict}`)}>
-      <span className={cx("nre-eval-verdict", `nre-verdict-${row.verdict}`)}>{verdictMark(row.verdict)}</span>
-      <span className="nre-eval-id">{row.evalId}</span>
+    <li className={cx("nre-experiment-attempt-row", `nre-eval-${attempt.verdict}`)}>
+      <span className="nre-attempt-branch" aria-hidden="true">{last ? "└─" : "├─"}</span>
       <span className="nre-eval-attempt-badges">
-        {row.attempts.map((attempt) => (
-          <AttemptLocatorBadge key={attempt.locator} item={attempt} attemptHref={attemptHref} />
-        ))}
+        <AttemptLocatorBadge item={attempt} attemptHref={attemptHref} />
       </span>
-      {row.verdict === "passed" ? (
-        <span className="nre-eval-avg">
-          {formatDurationMs(row.duration.value ?? 0)}
-          {row.cost.value !== null && <> · {formatUSD(row.cost.value)}</>}
+      <span className="nre-eval-attempt-metrics">
+        {formatDurationMs(attempt.durationMs)}
+        {attempt.costUSD !== undefined && <> · {formatUSD(attempt.costUSD)}</>}
+      </span>
+      <span className="nre-eval-reason">{reason ?? "—"}</span>
+    </li>
+  );
+}
+
+function EvalAttempts({
+  row,
+  attemptHref,
+  locale,
+}: {
+  row: ExperimentListEvalRow;
+  attemptHref: (locator: AttemptLocator) => string;
+  locale: ReportLocale;
+}): ReactElement {
+  return (
+    <li className="nre-experiment-eval">
+      <div className={cx("nre-experiment-eval-header", `nre-eval-${row.verdict}`)}>
+        <span className={cx("nre-eval-verdict", `nre-verdict-${row.verdict}`)}>{verdictMark(row.verdict)}</span>
+        <span className="nre-eval-id">{row.evalId}</span>
+        <span className="nre-eval-attempt-count">{localeText(locale, "overview.attemptsCount", { n: row.attempts.length })}</span>
+        <span className="nre-eval-summary">
+          {row.verdict === "passed"
+            ? [formatDurationMs(row.duration.value ?? 0), row.cost.value === null ? undefined : formatUSD(row.cost.value)]
+                .filter((value): value is string => value !== undefined)
+                .join(" · ")
+            : row.reason}
         </span>
-      ) : (
-        <span className="nre-eval-reason">{row.reason}</span>
-      )}
+      </div>
+      <ul className="nre-experiment-attempts">
+        {row.attempts.map((attempt, index) => (
+          <ExperimentAttemptRow
+            key={attempt.locator}
+            attempt={attempt}
+            last={index === row.attempts.length - 1}
+            attemptHref={attemptHref}
+          />
+        ))}
+      </ul>
     </li>
   );
 }
 
-function ExperimentCard({
+function Flags({ flags, locale }: { flags: Record<string, unknown> | undefined; locale: ReportLocale }): ReactElement | null {
+  if (!flags || Object.keys(flags).length === 0) return null;
+  return (
+    <div className="nre-experiment-flags">
+      <span>{localeText(locale, "experimentList.flags")}</span>
+      {Object.entries(flags).map(([key, value]) => (
+        <b key={key}>{key}={String(value)}</b>
+      ))}
+    </div>
+  );
+}
+
+function ExperimentRow({
   item,
   attemptHref,
   locale,
@@ -51,58 +123,98 @@ function ExperimentCard({
   locale: ReportLocale;
 }): ReactElement {
   return (
-    <li className="nre-experiment-entry">
-      <details className="nre-experiment-details">
-        <summary className="nre-experiment-summary">
-          <span className={cx("nre-experiment-id", "nre-key", colorClassForKey(item.experimentId))}>
-            {item.experimentId}
-          </span>
-          <span className="nre-experiment-agent">{item.agent}</span>
-          {item.model && <span className="nre-experiment-model">{item.model}</span>}
-          <span className="nre-experiment-cell">
-            <MetricCellView cell={item.passRate} attemptHref={attemptHref} locale={locale} />
-          </span>
-          <span className="nre-experiment-cell">
-            <MetricCellView cell={item.cost} attemptHref={attemptHref} locale={locale} />
-          </span>
-          <span className="nre-experiment-cell">
-            <MetricCellView cell={item.duration} attemptHref={attemptHref} locale={locale} />
-          </span>
-          <span className="nre-experiment-meta-sub">
+    <details className="nre-experiment-entry">
+      <summary className="nre-experiment-summary">
+        <span className="nre-experiment-name" data-sort-value={item.experimentId}>
+          <b className={cx("nre-experiment-id", "nre-key", colorClassForKey(item.experimentId))}>{item.experimentId}</b>
+          <small>
             {localeText(locale, "overview.evalsCount", { n: item.evals })}
             {item.attempts > item.evals ? ` · ${localeText(locale, "overview.attemptsCount", { n: item.attempts })}` : ""}
-            {` · ${localeText(locale, "latestRun", { run: item.lastRunAt })}`}
-          </span>
-        </summary>
+            {` · ${formatDate(item.lastRunAt, locale)}`}
+          </small>
+        </span>
+        <span data-sort-value={item.model ?? ""}>{item.model ?? localeText(locale, "experimentList.defaultModel")}</span>
+        <span data-sort-value={item.agent}>{item.agent}</span>
+        <span className="nre-num" data-sort-value={item.duration.value ?? ""}>
+          <MetricCellView cell={item.duration} locale={locale} />
+        </span>
+        <span className={cx("nre-num", passRateTone(item.passRate.value))} data-sort-value={item.passRate.value ?? ""}>
+          <MetricCellView cell={item.passRate} locale={locale} />
+        </span>
+        <span className="nre-num" data-sort-value={item.tokens.value ?? ""}>
+          <MetricCellView cell={item.tokens} locale={locale} />
+        </span>
+        <span className="nre-num" data-sort-value={item.cost.value ?? ""}>
+          <MetricCellView cell={item.cost} locale={locale} />
+        </span>
+        <span data-sort-value={item.verdicts.passed}><VerdictSummary item={item} locale={locale} /></span>
+      </summary>
+      <div className="nre-experiment-detail">
+        <Flags flags={item.flags} locale={locale} />
         <ul className="nre-experiment-evals">
           {item.evalRows.map((row) => (
-            <EvalRow key={row.evalId} row={row} attemptHref={attemptHref} />
+            <EvalAttempts key={row.evalId} row={row} attemptHref={attemptHref} locale={locale} />
           ))}
         </ul>
-      </details>
-    </li>
+      </div>
+    </details>
   );
 }
 
 export function ExperimentList({
   items,
   attemptHref,
+  filter = false,
   className,
   locale = DEFAULT_REPORT_LOCALE,
 }: {
   items: ExperimentListItem[];
   attemptHref: (locator: AttemptLocator) => string;
+  filter?: boolean;
   className?: string;
   locale?: ReportLocale;
 }): ReactElement {
-  return (
-    <div className={cx("nre", "nre-experiment-list", className)}>
-      {items.length === 0 && <p className="nre-experiment-list-empty">{localeText(locale, "attemptList.empty")}</p>}
-      <ul className="nre-experiments">
-        {items.map((item) => (
-          <ExperimentCard key={item.experimentId} item={item} attemptHref={attemptHref} locale={locale} />
+  const labels = [
+    localeText(locale, "experimentList.experiment"),
+    localeText(locale, "table.model"),
+    localeText(locale, "table.agent"),
+    localeText(locale, "experimentList.avgDuration"),
+    localeText(locale, "experimentList.passRate"),
+    localeText(locale, "experimentList.tokens"),
+    localeText(locale, "experimentList.estimatedCost"),
+    localeText(locale, "experimentList.result"),
+  ];
+  const board = (
+    <div className="nre-experiment-table">
+      <div className="nre-experiment-head">
+        {labels.map((label, index) => (
+          <button
+            type="button"
+            data-nre-experiment-sort={index}
+            className={index === 4 ? "nre-sort-desc" : undefined}
+            key={label}
+          >
+            {label}
+          </button>
         ))}
-      </ul>
+      </div>
+      {items.length === 0 && <p className="nre-experiment-list-empty">{localeText(locale, "attemptList.empty")}</p>}
+      {items.map((item) => (
+        <ExperimentRow key={item.experimentId} item={item} attemptHref={attemptHref} locale={locale} />
+      ))}
+    </div>
+  );
+  return (
+    <div className={cx("nre", "nre-experiment-list", filter && "nre-experiment-table-wrap", className)}>
+      {filter && (
+        <input
+          className="nre-filter"
+          data-nre-experiment-filter=""
+          type="search"
+          placeholder={localeText(locale, "experimentList.filterPlaceholder")}
+        />
+      )}
+      {board}
     </div>
   );
 }

package/src/report/react/MetricScatter.tsx

@@ -88,16 +88,6 @@ export function MetricScatter({
     );
   }
 
-  // 恰好 1 个可画点:成本 × 通过率的对比至少要两个实验,单点不成图 —— 显式说清而不是画个孤点。
-  if (drawableRows.length === 1) {
-    return (
-      <figure className={cx("nre", "nre-metric-scatter", className)}>
-        <p className="nre-scatter-empty">{localeText(locale, "scatter.needTwo", { x: xLabel, y: yLabel })}</p>
-        {missingNote}
-      </figure>
-    );
-  }
-
   const axisLabel = (label: string, col: MetricColumn) => `${label}${col.unit ? `(${col.unit})` : ""}`;
 
   const xScale = axisScale(drawableRows.map((r) => r.x.value as number), data.x.better, MARGIN.left, MARGIN.left + PLOT_W);
@@ -113,7 +103,7 @@ export function MetricScatter({
       label: pointLabel(r.key),
       xValue,
       yValue,
-      // hover 内容:experiment(点键)+ 系列(series,CostPassRateComparison 传的是 agent 维度,
+      // hover 内容:experiment(点键)+ 系列(series,ExperimentComparison 传的是 agent 维度,
       // 有则加一行;无系列的散点没有这行)+ 两轴 display 与 samples/total(docs/feature/reports/library.md「MetricScatter」行为清单)
       title: `${r.key}${r.series !== undefined ? `\n${r.series}` : ""}\n${xLabel}: ${r.x.display}(${r.x.samples}/${r.x.total})\n${yLabel}: ${r.y.display}(${r.y.samples}/${r.y.total})`,
       px: xScale.scale(xValue),