niceeval 0.9.1 → 0.10.2

package/src/report/components/site-components/compute.ts

@@ -0,0 +1,144 @@
+// 计算函数(*Data):ReportInput → 一份组件数据。站点组件族(Hero / ScopeWarnings /
+// CopyFixPrompt / TraceWaterfall)的 *Data 都住在这里
+// (docs/feature/reports/library/site-components.md)。
+//
+// 共同约定(docs/feature/reports/architecture.md「指标聚合不变量」):
+// - 第一参收 ReportInput = Scope | readonly Snapshot[];warnings 不进组件数据(宿主统一显示);
+// - null ≠ 0:缺数据不编数。
+
+import type { CopyFixPromptData, HeroData, ReportInput, ScopeWarning, TraceSpanSummary, TraceWaterfallRow } from "../../model/types.ts";
+import type { TraceSpan } from "../../../types.ts";
+import { collectItems, evalIdOf, experimentIdOf, locatorOf, resolveInput } from "../../model/aggregate.ts";
+import { attemptListData } from "../entity-lists/compute.ts";
+
+// ───────────────────────── 站点组件的计算函数(hero / warnings / fix prompt / trace)─────────────────────────
+
+/**
+ * `heroData(input)`:站点标题区的运行 meta——`latestStartedAt` 取范围内最新快照的开始时间
+ * (空范围为 null,不编造当前时间),`snapshots` 计贡献当前水位的快照数
+ * (docs/feature/reports/library/site-components.md「HeroCard」)。
+ */
+export async function heroData(input: ReportInput): Promise<HeroData> {
+  const { snapshots } = resolveInput(input);
+  let latest: string | null = null;
+  for (const snapshot of snapshots) {
+    if (latest === null || snapshot.startedAt > latest) latest = snapshot.startedAt;
+  }
+  return { latestStartedAt: latest, snapshots: snapshots.length };
+}
+
+/**
+ * `scopeWarningsData(input)`:Scope 携带的挑选警告原样透出;`input` 是裸 `Snapshot[]` 时
+ * 没有挑选过程、没有警告,返回空数组,也如实
+ * (docs/feature/reports/library/site-components.md「ScopeWarnings」)。
+ */
+export async function scopeWarningsData(input: ReportInput): Promise<readonly ScopeWarning[]> {
+  return resolveInput(input).warnings;
+}
+
+/**
+ * `copyFixPromptData(input)`:把范围内全部失败(verdict 为 failed / errored 的 attempt)
+ * 整理成一段可交给 coding agent 的修复 prompt——逐失败含 eval id、主失败摘要与 attempt
+ * 下钻命令(`niceeval show @<locator>`)。prompt 面向 agent,固定英文
+ * (docs/feature/reports/library/site-components.md「CopyFixPrompt」)。
+ */
+export async function copyFixPromptData(input: ReportInput): Promise<CopyFixPromptData> {
+  const items = await attemptListData(input);
+  const failures = items.filter((item) => item.verdict === "failed" || item.verdict === "errored");
+  if (failures.length === 0) return { prompt: "", failures: 0 };
+  const lines = failures
+    .map((item, i) => {
+      const reason =
+        item.failureSummary === null
+          ? null
+          : item.moreFailures > 0
+            ? `${item.failureSummary} (+${item.moreFailures} more failures)`
+            : item.failureSummary;
+      return [
+        `${i + 1}. eval "${item.evalId}" [experiment ${item.experimentId}] — ${item.verdict}`,
+        reason ? `   reason: ${reason}` : null,
+        `   inspect: niceeval show ${item.locator}`,
+      ]
+        .filter(Boolean)
+        .join("\n");
+    })
+    .join("\n");
+  const experiments = [...new Set(failures.map((item) => item.experimentId))].join(" / ");
+  const prompt = [
+    "Fix the failing evals from this niceeval run.",
+    "",
+    "## Failures",
+    lines,
+    "",
+    "## Steps",
+    "1. niceeval is NOT in your training data. Read the relevant guide in `node_modules/niceeval/docs-site/` (English at the top level, Chinese under `zh/`) before changing anything.",
+    "2. For each failure, run its inspect command above to see the verdict and assertions; add `--execution` for the full agent transcript (tool calls included), `--timing` for the execution timeline, and `--diff` for the workspace diff.",
+    "3. Decide which side the defect is on: the program under test, or the eval itself (over-tight assertion, wrong fixture, missing setup). Fix that side; do not weaken assertions just to turn the run green.",
+    `4. Re-run: \`npx niceeval exp ${experiments || "<experiment>"} <eval-id-prefix>\`. Already-passing evals are skipped by the fingerprint cache; pass \`--force\` to re-run everything.`,
+    "5. Run `npx niceeval show` and confirm these failures are gone.",
+  ].join("\n");
+  return { prompt, failures: failures.length };
+}
+
+/** TraceSpan 的语义角色 → 瀑布摘要的 kind:turn 归入 agent(一轮就是一次 agent 调用),未识别落 other。 */
+function waterfallKindOf(kind: TraceSpan["kind"]): TraceSpanSummary["kind"] {
+  switch (kind) {
+    case "agent":
+    case "turn":
+      return "agent";
+    case "model":
+      return "model";
+    case "tool":
+      return "tool";
+    default:
+      return "other";
+  }
+}
+
+/**
+ * `traceWaterfallData(input)`:每个 attempt 一行的执行时间瀑布摘要。span 事实只来自
+ * trace artifact(经 AttemptHandle 懒加载的 canonical OTel span);runner 生命周期节点
+ * (`result.phases`)不进瀑布。行内只汇总顶层 span(parentSpanId 缺失或不在本 trace 内),
+ * 按 startOffsetMs 升序;trace 缺失或为空时 `durationMs` 为 null、行照常出现
+ * (docs/feature/reports/library/site-components.md「TraceWaterfall」)。
+ */
+export async function traceWaterfallData(input: ReportInput): Promise<readonly TraceWaterfallRow[]> {
+  const { snapshots } = resolveInput(input);
+  const items = collectItems(snapshots);
+  return Promise.all(
+    items.map(async (item): Promise<TraceWaterfallRow> => {
+      const spans = await item.attempt.trace();
+      if (spans === null || spans.length === 0) {
+        return {
+          experimentId: experimentIdOf(item),
+          evalId: evalIdOf(item),
+          locator: locatorOf(item),
+          durationMs: null,
+          spans: [],
+        };
+      }
+      const t0 = Math.min(...spans.map((s) => s.startMs));
+      const t1 = Math.max(...spans.map((s) => s.endMs));
+      const ids = new Set(spans.map((s) => s.spanId));
+      const topLevel = spans.filter((s) => s.parentSpanId === undefined || !ids.has(s.parentSpanId));
+      const summaries = topLevel
+        .map(
+          (s): TraceSpanSummary => ({
+            name: s.name,
+            kind: waterfallKindOf(s.kind),
+            startOffsetMs: s.startMs - t0,
+            durationMs: s.endMs - s.startMs,
+            failed: s.status === "error",
+          }),
+        )
+        .sort((a, b) => a.startOffsetMs - b.startOffsetMs);
+      return {
+        experimentId: experimentIdOf(item),
+        evalId: evalIdOf(item),
+        locator: locatorOf(item),
+        durationMs: Math.max(0, t1 - t0),
+        spans: summaries,
+      };
+    }),
+  );
+}

package/src/report/components/site-components/faces.ts

@@ -0,0 +1,80 @@
+// 站点组件族(HeroCard / ScopeWarnings / TraceWaterfall)的 text 面:同一份算好的数据,
+// 渲染成终端字符(niceeval show 的形态)。零 react、零 IO、纯同步。
+
+import type { HeroData, ScopeWarning, TraceWaterfallRow } from "../../model/types.ts";
+import type { LocalizedText } from "../../model/locale.ts";
+import type { TextContext } from "../../definition/tree.ts";
+import { countText, localeText, resolveLocalizedText } from "../../model/locale.ts";
+import { formatDurationMs } from "../../model/format.ts";
+import { groupScopeWarnings } from "./scope-warnings.ts";
+
+/** ISO 时间 → "YYYY-MM-DD HH:mm"(本地时区);不可解析原样返回。 */
+function formatDateTimeMinute(iso: string): string {
+  const date = new Date(iso);
+  if (Number.isNaN(date.valueOf())) return iso;
+  const pad = (n: number) => String(n).padStart(2, "0");
+  return `${date.getFullYear()}-${pad(date.getMonth() + 1)}-${pad(date.getDate())} ${pad(date.getHours())}:${pad(date.getMinutes())}`;
+}
+
+// ───────────────────────── 站点组件(HeroCard / ScopeWarnings / TraceWaterfall)─────────────────────────
+
+/**
+ * HeroCard 的 text 面:标题行 + meta 行(最后运行时间;空范围为内置「暂无运行」文案;
+ * 多快照时标注合成来源),不含品牌行(品牌行是纯 web 件,text 面零输出)。
+ */
+export function heroCardText(title: LocalizedText, data: HeroData, ctx: TextContext): string {
+  const locale = ctx.locale;
+  const meta =
+    data.latestStartedAt === null
+      ? localeText(locale, "hero.noRuns")
+      : [
+          localeText(locale, "hero.lastRun", { time: formatDateTimeMinute(data.latestStartedAt) }),
+          ...(data.snapshots > 1 ? [localeText(locale, "hero.composedSnapshots", { n: data.snapshots })] : []),
+        ].join(" · ");
+  return `${resolveLocalizedText(title, locale)}\n${meta}`;
+}
+
+/**
+ * ScopeWarnings 的 text 面:按动作聚合(../scope-warnings.ts,与 web 面共用),同构但不折叠——
+ * 多组时首行 "! <分类计数汇总>";每组一行组头 "! <标题> — <徽标> → <组头命令>",其下缩进
+ * 逐条原样打印 message(已以下一步收尾,不截断掉尾段)。空警告集零输出。
+ */
+export function scopeWarningsText(warnings: readonly ScopeWarning[], ctx: TextContext): string {
+  if (warnings.length === 0) return "";
+  const { summary, groups } = groupScopeWarnings(warnings, ctx.locale);
+  const lines: string[] = [];
+  // 汇总行只在多组时打印;单组时组头即汇总,不另起一行(web 面则恒以汇总行作外层 <summary>)。
+  if (groups.length > 1) lines.push(`! ${summary}`);
+  for (const group of groups) {
+    const badges = group.badges.length > 0 ? ` — ${group.badges.map((b) => b.text).join(" · ")}` : "";
+    const command = group.headCommand !== null ? ` → ${group.headCommand}` : "";
+    lines.push(`! ${group.title}${badges}${command}`);
+    for (const w of group.warnings) lines.push(`!   ${w.message}`);
+  }
+  return lines.join("\n");
+}
+
+/**
+ * TraceWaterfall 的 text 面:每 attempt 一行——locator、总耗时(缺 trace 如实显示缺失)、
+ * 顶层 span 计数与失败标记,行尾是可复制的 `--timing` 下钻命令(经宿主注入的 attemptCommand
+ * 通道拼出,携带宿主上下文)。当前报告没有 attempt-input page 时 `ctx.attemptCommand`
+ * 不存在,行退化为纯文本,不生成假命令(architecture.md「Attempt 详情是一张参数化 page」)。
+ */
+export function traceWaterfallText(rows: readonly TraceWaterfallRow[], ctx: TextContext): string {
+  const locale = ctx.locale;
+  if (rows.length === 0) return localeText(locale, "traceWaterfall.empty");
+  return rows
+    .map((row) => {
+      const failedSpans = row.spans.filter((span) => span.failed).length;
+      const parts = [
+        row.locator,
+        row.evalId,
+        row.durationMs === null ? localeText(locale, "traceWaterfall.noTrace") : formatDurationMs(row.durationMs),
+        countText(locale, "traceWaterfall.spans", row.spans.length),
+        ...(failedSpans > 0 ? [`✗ ${countText(locale, "traceWaterfall.failedSpans", failedSpans)}`] : []),
+      ];
+      const line = parts.join(" · ");
+      return ctx.attemptCommand ? `${line}   ${ctx.attemptCommand(row.locator)} --timing` : line;
+    })
+    .join("\n");
+}

package/src/report/components/site-components/index.tsx

@@ -0,0 +1,265 @@
+// 官方双面组件的装配点:web 面(./HeroCard.tsx 等纯 React 组件)+ text 面(./faces.ts)+
+// resolve 解析面(spec 形态由管线代调配套 ./compute.ts)。Hero 是组合组件(装配 HeroCard),
+// 不产生自己的 data;PoweredBy 无 props,没有对应的计算函数或 validate。
+
+import { defineComponent, type ReportComponent } from "../../definition/tree.ts";
+import type { CopyFixPromptData, HeroData, ScopeWarning, TraceWaterfallRow } from "../../model/types.ts";
+import type { LocalizedText } from "../../model/locale.ts";
+import type { AttemptLocator } from "../../../results/locator.ts";
+import {
+  arrayProblem,
+  dataShapeError,
+  isObject,
+  makeDataComponent,
+  hrefOf,
+  type ChromeProps,
+  type DataProps,
+  type Validator,
+} from "../shared.ts";
+import { copyFixPromptData, heroData, scopeWarningsData, traceWaterfallData } from "./compute.ts";
+import { heroCardText, scopeWarningsText, traceWaterfallText } from "./faces.ts";
+import { HeroCard as HeroCardWeb } from "./HeroCard.tsx";
+import { PoweredBy as PoweredByWeb } from "./PoweredBy.tsx";
+import { ScopeWarnings as ScopeWarningsWeb } from "./ScopeWarnings.tsx";
+import { CopyFixPrompt as CopyFixPromptWeb } from "./CopyFixPrompt.tsx";
+import { TraceWaterfall as TraceWaterfallWeb } from "./TraceWaterfall.tsx";
+
+export const validateHeroData: Validator = (data) => {
+  if (!isObject(data)) return "expected an object";
+  if (!(data.latestStartedAt === null || typeof data.latestStartedAt === "string")) {
+    return '"latestStartedAt" must be a string or null';
+  }
+  if (typeof data.snapshots !== "number") return '"snapshots" must be a number';
+  return null;
+};
+
+const UNREADABLE_SNAPSHOT_REASONS = ["incompatible-version", "malformed", "incomplete"];
+
+/**
+ * ScopeWarning(src/results/types.ts「警告 kind 全集」):按 `kind` 判别的联合,已登记 kind
+ * 各自的必填字段单独校验——`unreadable-snapshot` 没有 experimentId(非实验作用域)、
+ * `command` 恒可选。未登记的 `kind`(如未来版本新增的 warning)只要求 `kind` / `message`
+ * 这份两族共用的最小形状,不拒绝——`ScopeWarnings` 的分组渲染对未识别 kind 有专门的
+ * 单独成组回退(message 原样、按 integrity 归位),这条前向兼容路径本身就是契约的一部分,
+ * 结构校验不能比渲染逻辑更严。
+ */
+function scopeWarningProblem(value: unknown, path: string): string | null {
+  if (!isObject(value)) return `"${path}" must be a ScopeWarning object`;
+  if (typeof value.kind !== "string") return `"${path}.kind" must be a string`;
+  if (typeof value.message !== "string") return `"${path}.message" must be a string`;
+  switch (value.kind) {
+    case "partial-coverage":
+      if (typeof value.experimentId !== "string") return `"${path}.experimentId" must be a string`;
+      if (typeof value.covered !== "number") return `"${path}.covered" must be a number`;
+      if (typeof value.total !== "number") return `"${path}.total" must be a number`;
+      if (typeof value.command !== "string") return `"${path}.command" must be a string`;
+      return null;
+    case "stale-snapshot":
+      if (typeof value.experimentId !== "string") return `"${path}.experimentId" must be a string`;
+      if (typeof value.startedAt !== "string") return `"${path}.startedAt" must be a string`;
+      if (typeof value.latestStartedAt !== "string") return `"${path}.latestStartedAt" must be a string`;
+      if (typeof value.command !== "string") return `"${path}.command" must be a string`;
+      return null;
+    case "unfinished-snapshot":
+      if (typeof value.experimentId !== "string") return `"${path}.experimentId" must be a string`;
+      if (typeof value.startedAt !== "string") return `"${path}.startedAt" must be a string`;
+      if (typeof value.dir !== "string") return `"${path}.dir" must be a string`;
+      if (typeof value.command !== "string") return `"${path}.command" must be a string`;
+      return null;
+    case "unreadable-snapshot":
+      if (typeof value.dir !== "string") return `"${path}.dir" must be a string`;
+      if (typeof value.reason !== "string" || !UNREADABLE_SNAPSHOT_REASONS.includes(value.reason)) {
+        return `"${path}.reason" must be one of ${JSON.stringify(UNREADABLE_SNAPSHOT_REASONS)}`;
+      }
+      if (value.command !== undefined && typeof value.command !== "string") {
+        return `"${path}.command" must be a string when present`;
+      }
+      return null;
+    default:
+      return null;
+  }
+}
+
+export const validateScopeWarningsData: Validator = (data) => arrayProblem(data, "data", scopeWarningProblem);
+
+export const validateCopyFixPromptData: Validator = (data) => {
+  if (!isObject(data)) return "expected an object";
+  if (typeof data.prompt !== "string") return '"prompt" must be a string';
+  if (typeof data.failures !== "number") return '"failures" must be a number';
+  return null;
+};
+
+const TRACE_SPAN_KINDS = ["agent", "model", "tool", "other"];
+
+/** TraceSpanSummary(src/report/model/types.ts):瀑布行里的一个顶层 span 摘要。 */
+function traceSpanSummaryProblem(value: unknown, path: string): string | null {
+  if (!isObject(value)) return `"${path}" must be a TraceSpanSummary { name, kind, startOffsetMs, durationMs, failed }`;
+  if (typeof value.name !== "string") return `"${path}.name" must be a string`;
+  if (typeof value.kind !== "string" || !TRACE_SPAN_KINDS.includes(value.kind)) {
+    return `"${path}.kind" must be one of ${JSON.stringify(TRACE_SPAN_KINDS)}`;
+  }
+  if (typeof value.startOffsetMs !== "number") return `"${path}.startOffsetMs" must be a number`;
+  if (typeof value.durationMs !== "number") return `"${path}.durationMs" must be a number`;
+  if (typeof value.failed !== "boolean") return `"${path}.failed" must be a boolean`;
+  return null;
+}
+
+export const validateTraceWaterfallData: Validator = (data) =>
+  arrayProblem(data, "data", (row, path) => {
+    if (!isObject(row)) return `"${path}" must be an object`;
+    if (typeof row.experimentId !== "string") return `"${path}.experimentId" must be a string`;
+    if (typeof row.evalId !== "string") return `"${path}.evalId" must be a string`;
+    if (typeof row.locator !== "string") return `"${path}.locator" must be a string`;
+    if (!(row.durationMs === null || typeof row.durationMs === "number")) {
+      return `"${path}.durationMs" must be a number or null`;
+    }
+    return arrayProblem(row.spans, `${path}.spans`, traceSpanSummaryProblem);
+  });
+
+// ───────────────────────── 站点组件(Hero / PoweredBy / ScopeWarnings / CopyFixPrompt / TraceWaterfall)─────────────────────────
+
+/** `Hero` 的 props:标题缺省取 `ctx.report.title`(回退链后的站点标题)。 */
+export interface HeroProps {
+  /** 覆盖标题;省略时取 ctx.report.title(回退链后的站点标题)。 */
+  title?: LocalizedText;
+  className?: string;
+}
+
+/** HeroCard 的 data 校验入口(它不经 makeDataComponent,数据形态是唯一形态)。 */
+const assertHeroData = (data: unknown): HeroData => {
+  const problem = validateHeroData(data);
+  if (problem !== null) throw dataShapeError("HeroCard", "heroData", "HeroData", problem);
+  return data as HeroData;
+};
+
+/**
+ * `HeroCard`:Hero 的渲染件,双面组件,只收 data 形态——标题输入是站点声明与 Scope 的
+ * 合成物,没有单独的 spec 等价形。web 面渲染 hero 标题(h1)、按渲染 locale 格式化的运行
+ * meta(latestStartedAt 为 null 时内置「暂无运行」文案)与品牌行(等同 PoweredBy,恒含、
+ * 无拆除 prop);text 面输出标题行与 meta 行,不含品牌行
+ * (docs/feature/reports/library/site-components.md「HeroCard」)。
+ */
+export const HeroCard = defineComponent<HeroCardProps>({
+  web: (props, ctx) => {
+    assertHeroData(props.data);
+    return <HeroCardWeb title={props.title} data={props.data} className={props.className} locale={ctx.locale} />;
+  },
+  text: (props, ctx) => {
+    assertHeroData(props.data);
+    return heroCardText(props.title, props.data, ctx);
+  },
+});
+HeroCard.displayName = "HeroCard";
+
+/** `HeroCard` 的 props:标题 + `heroData()` 的产物,只有 data 形态。 */
+export interface HeroCardProps {
+  title: LocalizedText;
+  data: HeroData;
+  className?: string;
+}
+
+/**
+ * `Hero`:页首的站点标题区——标题、最后运行时间、快照合成来源,恒含品牌行。官方组合组件,
+ * 与手写 `<HeroCard title={title ?? ctx.report.title} data={await heroData(ctx.scope)} />`
+ * 严格等价、没有私有能力;读 `ctx.report` 意味着输出跟随站点,要站点无关的标题区直接用
+ * `HeroCard` 显式传值(docs/feature/reports/library/site-components.md「Hero」)。
+ */
+export const Hero = defineComponent<HeroProps>(async ({ title, className }, ctx) => (
+  <HeroCard title={title ?? ctx.report.title} data={await heroData(ctx.scope)} className={className} />
+));
+Hero.displayName = "Hero";
+
+/**
+ * `PoweredBy`:唯一的品牌件,无 props 双面组件。web 面渲染指向 niceeval 官网的一行品牌色
+ * 小字(`utm_source=report&utm_medium=powered-by`,`rel` 仅 `noopener` 以保留 Referer);
+ * text 面零输出。没有任何配置——品牌契约是「提供一个组件,不给开关」:不想要品牌就不用
+ * 这些组件、自己写替代组件(docs/feature/reports/library/site-components.md「PoweredBy」)。
+ */
+export const PoweredBy = defineComponent<Record<never, never>>({
+  web: () => <PoweredByWeb />,
+  text: () => "",
+});
+PoweredBy.displayName = "PoweredBy";
+
+/** `ScopeWarnings` 的 props:spec 形态取宿主 Scope 的 warnings,data 形态收 `ScopeWarning[]`。 */
+export type ScopeWarningsProps = DataProps<readonly ScopeWarning[], Record<never, never>, ChromeProps>;
+
+/**
+ * `ScopeWarnings`:选择警告区,警告的唯一呈现组件。把 Scope 携带的 `ScopeWarning[]`
+ * 按「下一步动作」聚合渲染(带 experimentId 的按实验聚合、非实验作用域按 kind 聚合;
+ * integrity 组在前);web 面组头带去重后的可复制命令、明细收原生 `<details>`(总条数 ≤ 3
+ * 默认展开),text 面同构但不折叠。空警告集与裸 `Snapshot[]` 输入两面零输出
+ * (docs/feature/reports/library/site-components.md「ScopeWarnings」)。
+ */
+export const ScopeWarnings = makeDataComponent<readonly ScopeWarning[], Record<never, never>, ChromeProps>({
+  name: "ScopeWarnings",
+  dataFnName: "scopeWarningsData",
+  shapeName: "ScopeWarning[]",
+  dataFn: (input) => scopeWarningsData(input),
+  specKeys: [],
+  validate: validateScopeWarningsData,
+  web: (props, ctx) =>
+    props.data.length === 0 ? null : (
+      <ScopeWarningsWeb data={props.data} locale={props.locale ?? ctx.locale} className={props.className} />
+    ),
+  text: (props, ctx) => scopeWarningsText(props.data, ctx),
+}) as unknown as ReportComponent<ScopeWarningsProps>;
+
+/** `CopyFixPrompt` 的 props:spec 形态无选项,data 形态收 `copyFixPromptData()` 的产物。 */
+export type CopyFixPromptProps = DataProps<CopyFixPromptData, Record<never, never>, ChromeProps>;
+
+/**
+ * `CopyFixPrompt`:把当前范围的全部失败整理成一段可交给 coding agent 的修复 prompt。
+ * prompt 在 resolve 阶段算好、烘进静态 HTML,无 JS 时在折叠块里完整可读,「复制」是增强层
+ * 行为;`failures` 为 0 时两面零输出;text 面恒零输出——终端里的等价能力是 `show` 的
+ * attempt 下钻命令本身(docs/feature/reports/library/site-components.md「CopyFixPrompt」)。
+ */
+export const CopyFixPrompt = makeDataComponent<CopyFixPromptData, Record<never, never>, ChromeProps>({
+  name: "CopyFixPrompt",
+  dataFnName: "copyFixPromptData",
+  shapeName: "CopyFixPromptData",
+  dataFn: (input) => copyFixPromptData(input),
+  specKeys: [],
+  validate: validateCopyFixPromptData,
+  web: (props, ctx) =>
+    props.data.failures === 0 ? null : (
+      <CopyFixPromptWeb data={props.data} locale={props.locale ?? ctx.locale} className={props.className} />
+    ),
+  text: () => "",
+}) as unknown as ReportComponent<CopyFixPromptProps>;
+
+/** `TraceWaterfall` 的 props:spec 形态无选项,data 形态收 `traceWaterfallData()` 的产物。 */
+export type TraceWaterfallProps = DataProps<
+  readonly TraceWaterfallRow[],
+  Record<never, never>,
+  ChromeProps & { attemptHref?: (locator: AttemptLocator) => string }
+>;
+
+/**
+ * `TraceWaterfall`:每个 attempt 一行的执行时间瀑布,用 canonical OTel 字段显示被测 agent
+ * 的原始 span(agent / model / tool)。web 面静态渲染顶层 span 分解条(失败 span 带失败
+ * 标记),行链接 attempt 详情;text 面每 attempt 一行(locator、总耗时、span 计数与失败
+ * 标记)+ 可复制的 `--timing` 下钻命令。trace 缺失的行照常出现并如实显示缺失;runner
+ * 生命周期节点不进瀑布(docs/feature/reports/library/site-components.md「TraceWaterfall」)。
+ */
+export const TraceWaterfall = makeDataComponent<
+  readonly TraceWaterfallRow[],
+  Record<never, never>,
+  ChromeProps & { attemptHref?: (locator: AttemptLocator) => string }
+>({
+  name: "TraceWaterfall",
+  dataFnName: "traceWaterfallData",
+  shapeName: "TraceWaterfallRow[]",
+  dataFn: (input) => traceWaterfallData(input),
+  specKeys: [],
+  validate: validateTraceWaterfallData,
+  web: (props, ctx) => (
+    <TraceWaterfallWeb
+      data={props.data}
+      attemptHref={hrefOf(props, ctx)}
+      locale={props.locale ?? ctx.locale}
+      className={props.className}
+    />
+  ),
+  text: (props, ctx) => traceWaterfallText(props.data, ctx),
+}) as unknown as ReportComponent<TraceWaterfallProps>;

package/src/report/{scope-warnings.ts → components/site-components/scope-warnings.ts}

@@ -4,9 +4,9 @@
 // (docs/feature/results/library.md「警告 kind 全集」)经 locale 词典渲染,未登记的 kind
 // 回退为单独成组、逐条 message 原样。
 
-import type { ScopeWarning } from "../results/types.ts";
-import { gapParts } from "../results/select.ts";
-import { localeText, type ReportLocale, type ReportMessageKey } from "./locale.ts";
+import type { ScopeWarning } from "../../../results/types.ts";
+import { gapParts } from "../../../results/select.ts";
+import { localeText, type ReportLocale, type ReportMessageKey } from "../../model/locale.ts";
 
 /** kind 表登记的类别:integrity(选中集合的分母可能不对)组排在 freshness(可能过期)之前。 */
 export type WarningCategory = "integrity" | "freshness";

package/src/report/{site-components.test.tsx → components/site-components/site-components.test.tsx}

@@ -7,9 +7,9 @@
 
 import { describe, expect, it } from "vitest";
 
-import type { AssertionResult, EvalResult, TraceSpan, Verdict } from "../types.ts";
-import type { AttemptHandle, Results, Scope, ScopeWarning, Snapshot } from "../results/index.ts";
-import { makeScope } from "../results/select.ts";
+import type { AssertionResult, EvalResult, TraceSpan, Verdict } from "../../../types.ts";
+import type { AttemptHandle, Results, Scope, ScopeWarning, Snapshot } from "../../../results/index.ts";
+import { makeScope } from "../../../results/select.ts";
 import {
   createTextContext,
   defineComponent,
@@ -18,20 +18,15 @@ import {
   validateReportTree,
   ResolveMemo,
   type ReportNode,
-} from "./tree.ts";
-import { buildReportMeta, defineReport, pickReportPage, renderReportToText, reportTitleText } from "./report.ts";
-import { renderReportToStaticHtml } from "./web.ts";
-import {
-  AttemptList,
-  CopyFixPrompt,
-  Hero,
-  HeroCard,
-  PoweredBy,
-  ScopeWarnings,
-  TraceWaterfall,
-} from "./components.tsx";
-import { Text } from "./primitives.tsx";
-import { attemptListData, copyFixPromptData, heroData, traceWaterfallData } from "./compute.ts";
+} from "../../definition/tree.ts";
+import { buildReportMeta, defineReport } from "../../definition/report.ts";
+import { pickReportPage, renderReportToText, reportTitleText } from "../../runtime/text.ts";
+import { renderReportToStaticHtml } from "../../runtime/web.ts";
+import { AttemptList } from "../entity-lists/index.tsx";
+import { CopyFixPrompt, Hero, HeroCard, PoweredBy, ScopeWarnings, TraceWaterfall } from "./index.tsx";
+import { Text } from "../../definition/primitives.tsx";
+import { attemptListData } from "../entity-lists/compute.ts";
+import { copyFixPromptData, heroData, traceWaterfallData } from "./compute.ts";
 
 // ───────────────────────── fake 数据(按 results 读取契约造)─────────────────────────
 
@@ -120,22 +115,29 @@ function hostCtx(scope: Scope) {
   return { scope, results: resultsOf(scope.snapshots) };
 }
 
+/** 站点组件测试固定"宿主已声明 attempt-input page"这条件,验证 attemptCommand/attemptHref 通道本身接得对;
+ * 「没有 attempt page 时退化成纯文本」由 report.test.ts / show.test.ts 的报告级测试单独覆盖。 */
+const DEFAULT_TEST_ATTEMPT_COMMAND = (locator: string) => `niceeval show ${locator}`;
+const DEFAULT_TEST_ATTEMPT_HREF = (locator: string) => `#/attempt/${locator}`;
+
 /** 管线便捷入口:装载 → resolve → validate → text 渲染(与 show 同一条管线,不带宿主前置块)。 */
 async function treeText(node: ReportNode, scope: Scope, width = 120): Promise<string> {
   const definition = defineReport(node);
-  const resolved = await resolveReportTree(pickReportPage(definition).content, {
+  const page = pickReportPage(definition);
+  const resolved = await resolveReportTree(page.content, {
     scope,
     results: resultsOf(scope.snapshots),
-    report: buildReportMeta(definition, scope, "report"),
+    report: buildReportMeta(definition, scope),
+    page: { id: page.id, input: "scope" },
     memo: new ResolveMemo(),
   });
   validateReportTree(resolved);
-  return renderNodeToText(resolved, createTextContext({ width }));
+  return renderNodeToText(resolved, createTextContext({ width, attemptCommand: DEFAULT_TEST_ATTEMPT_COMMAND }));
 }
 
 /** 管线便捷入口:web 面静态 HTML(经 renderReportToStaticHtml;测试用无警告 Scope,免宿主前置块)。 */
 async function treeHtml(node: ReportNode, scope: Scope): Promise<string> {
-  return renderReportToStaticHtml(defineReport(node), hostCtx(scope));
+  return renderReportToStaticHtml(defineReport(node), hostCtx(scope), { attemptHref: DEFAULT_TEST_ATTEMPT_HREF });
 }
 
 // ───────────────────────── 警告 fixture(按 ScopeWarning 联合造)─────────────────────────