niceeval 0.8.1 → 0.9.0

package/dist/report/compute.js

@@ -774,3 +774,126 @@ function deltaOutcome(metric, delta) {
     const better = metric.better ?? "higher";
     return (delta > 0) === (better === "higher") ? "improved" : "regressed";
 }
+// ───────────────────────── 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) {
+    const { snapshots } = resolveInput(input);
+    let latest = 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) {
+    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) {
+    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) {
+    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) {
+    const { snapshots } = resolveInput(input);
+    const items = collectItems(snapshots);
+    return Promise.all(items.map(async (item) => {
+        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) => ({
+            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/dist/report/index.d.ts

@@ -10,10 +10,10 @@ export { stringWidth, padDisplay as padEnd, padStartDisplay as padStart, wrapDis
 export type { ColumnAlign } from "./text/layout.ts";
 export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel } from "./locale.ts";
 export type { LocalizedText, ReportLocale } from "./locale.ts";
-export { AttemptList, DeltaTable, EvalList, ExperimentComparison, ExperimentList, FailureList, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, Scoreboard, ScopeSummary, } from "./components.tsx";
-export type { AttemptListProps, DataProps, DeltaTableProps, EvalListProps, ExperimentComparisonProps, ExperimentListProps, FailureListProps, MetricBarsProps, MetricLineProps, MetricMatrixProps, MetricScatterProps, MetricTableProps, ScoreboardProps, ScopeSummaryProps, } from "./components.tsx";
-export { attemptListData, deltaTableData, evalListData, experimentComparisonData, experimentListData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scopeSummaryData, scoreboardData, } from "./compute.ts";
+export { AttemptList, CopyFixPrompt, DeltaTable, EvalList, ExperimentComparison, ExperimentList, FailureList, Hero, HeroCard, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, PoweredBy, Scoreboard, ScopeSummary, ScopeWarnings, TraceWaterfall, } from "./components.tsx";
+export type { AttemptListProps, CopyFixPromptProps, DataProps, DeltaTableProps, EvalListProps, ExperimentComparisonProps, ExperimentListProps, FailureListProps, HeroCardProps, HeroProps, MetricBarsProps, MetricLineProps, MetricMatrixProps, MetricScatterProps, MetricTableProps, ScoreboardProps, ScopeSummaryProps, ScopeWarningsProps, TraceWaterfallProps, } from "./components.tsx";
+export { attemptListData, copyFixPromptData, deltaTableData, evalListData, experimentComparisonData, experimentListData, heroData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scopeSummaryData, scopeWarningsData, scoreboardData, traceWaterfallData, } from "./compute.ts";
 export type { DeltaTableOptions, MetricLineOptions, MetricMatrixOptions, MetricScatterOptions, MetricTableOptions, ScoreboardOptions, } from "./compute.ts";
-export type { Aggregator, AttemptListItem, AttemptLocator, BuiltInDimension, CustomDimension, DeltaData, DeltaPair, DimensionInput, DimensionOptions, DimensionRef, EvalListItem, ExperimentComparisonData, ExperimentComparisonGroupData, ExperimentListEvalRow, ExperimentListItem, FlagPairs, LineData, MatrixData, Metric, MetricAggregate, MetricCell, MetricColumn, NumericAxis, NumericAxisOptions, NumericRunConfigAxisOptions, ReportInput, RunConfigKey, ScatterData, ScopeSummaryData, ScopeWarning, ScoreboardData, TableData, VerdictTally, } from "./types.ts";
+export type { Aggregator, AttemptListItem, AttemptLocator, BuiltInDimension, CopyFixPromptData, CustomDimension, DeltaData, DeltaPair, DimensionInput, DimensionOptions, DimensionRef, EvalListItem, ExperimentComparisonData, ExperimentComparisonGroupData, ExperimentListEvalRow, ExperimentListItem, FlagPairs, HeroData, LineData, MatrixData, Metric, MetricAggregate, MetricCell, MetricColumn, NumericAxis, NumericAxisOptions, NumericRunConfigAxisOptions, ReportInput, RunConfigKey, ScatterData, ScopeSummaryData, ScopeWarning, ScoreboardData, TableData, TraceSpanSummary, TraceWaterfallRow, VerdictTally, } from "./types.ts";
 export type { AttemptHandle, Results, Scope, Snapshot } from "../results/types.ts";
 export { experimentGroupOf } from "../shared/aggregate.ts";

package/dist/report/index.js

@@ -20,10 +20,11 @@ export { stringWidth, padDisplay as padEnd, padStartDisplay as padStart, wrapDis
 // locale:官方组件 chrome 文案的语言(内置词典覆盖 en / zh-CN,其它 locale 走回退)
 export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel } from "./locale.js";
 // 官方双面组件(spec / data 双形态;配套 *Data 计算函数在下面成对导出)
-export { AttemptList, DeltaTable, EvalList, ExperimentComparison, ExperimentList, FailureList, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, Scoreboard, ScopeSummary, } from "./components.js";
+// 与站点组件(Hero / HeroCard / PoweredBy / ScopeWarnings / CopyFixPrompt / TraceWaterfall)
+export { AttemptList, CopyFixPrompt, DeltaTable, EvalList, ExperimentComparison, ExperimentList, FailureList, Hero, HeroCard, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, PoweredBy, Scoreboard, ScopeSummary, ScopeWarnings, TraceWaterfall, } from "./components.js";
 // 计算函数(组件解析面的具名形式,与组件成对;spec 形态下由管线代调,data 形态与
 // 嵌入场景下由作者手工调)
-export { attemptListData, deltaTableData, evalListData, experimentComparisonData, experimentListData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scopeSummaryData, scoreboardData, } from "./compute.js";
+export { attemptListData, copyFixPromptData, deltaTableData, evalListData, experimentComparisonData, experimentListData, heroData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scopeSummaryData, scopeWarningsData, scoreboardData, traceWaterfallData, } from "./compute.js";
 // experiment id 的组键推导(id 的目录前缀,如 `compare/bub-low` 的 `compare`)。
 // 重新导出,让自定义报告能按同一份口径把 experiment 分组,不必自己重写这两行逻辑。
 export { experimentGroupOf } from "../shared/aggregate.js";

package/dist/report/locale.d.ts

@@ -104,13 +104,51 @@ declare const en: {
     readonly "scoreboard.subjectTitle": "{questions} evals, weighted {earned} of {possible}";
     readonly "delta.pairHeader": "pair (A → B)";
     readonly "delta.empty": "{experiments} experiments, 0 comparable pairs";
+    /** ScopeWarnings 聚合层的 chrome:汇总行、kind 徽标、组头与明细折叠标签;message 本体不经字典。 */
+    readonly "warnings.summary.experiments.one": "{n} experiment flagged";
+    readonly "warnings.summary.experiments.other": "{n} experiments flagged";
+    readonly "warnings.group.unreadableSnapshot.one": "{n} snapshot skipped";
+    readonly "warnings.group.unreadableSnapshot.other": "{n} snapshots skipped";
+    readonly "warnings.details.one": "{n} warning";
+    readonly "warnings.details.other": "{n} warnings";
+    readonly "warnings.badge.partialCoverage": "coverage {covered}/{total}";
+    readonly "warnings.badge.staleSnapshot": "{gap} behind";
+    readonly "warnings.badge.unfinishedSnapshot": "unfinished";
+    readonly "warnings.gap.second.one": "{n} second";
+    readonly "warnings.gap.second.other": "{n} seconds";
+    readonly "warnings.gap.minute.one": "{n} minute";
+    readonly "warnings.gap.minute.other": "{n} minutes";
+    readonly "warnings.gap.hour.one": "{n} hour";
+    readonly "warnings.gap.hour.other": "{n} hours";
+    readonly "warnings.gap.day.one": "{n} day";
+    readonly "warnings.gap.day.other": "{n} days";
+    /** Hero / HeroCard 的运行 meta(hero.noRuns 是 latestStartedAt 为 null 时的内置文案)。 */
+    readonly "hero.lastRun": "Last run {time}";
+    readonly "hero.noRuns": "No runs yet";
+    /** web 面的合成来源标注(仅 snapshots > 1 时显示)。 */
+    readonly "hero.composedRuns": "composed from {n} runs";
+    /** text 面的合成来源标注(show 页首 meta 行,仅 snapshots > 1 时显示)。 */
+    readonly "hero.composedSnapshots": "composed from {n} snapshots";
+    /** CopyFixPrompt 的 web 面 chrome(prompt 本身面向 agent、固定英文,不经词典)。 */
+    readonly "copyFixPrompt.summary.one": "Fix prompt · {n} failure";
+    readonly "copyFixPrompt.summary.other": "Fix prompt · {n} failures";
+    readonly "copyFixPrompt.copy": "Copy fix prompt";
+    /** TraceWaterfall 的 chrome。 */
+    readonly "traceWaterfall.empty": "No attempts";
+    readonly "traceWaterfall.noTrace": "no trace";
+    readonly "traceWaterfall.spans.one": "{n} span";
+    readonly "traceWaterfall.spans.other": "{n} spans";
+    readonly "traceWaterfall.failedSpans.one": "{n} failed";
+    readonly "traceWaterfall.failedSpans.other": "{n} failed";
+    /** AttemptList 的 web 面过滤框占位符(filter 渐进增强)。 */
+    readonly "attemptList.filterPlaceholder": "Filter attempts…";
     readonly "tabs.tab": "Tab";
 };
 export type ReportMessageKey = keyof typeof en;
 /** 查字典 + 简单插值({name} 占位符)。内置词典未覆盖的 locale 回退 en。 */
 export declare function localeText(locale: ReportLocale, key: ReportMessageKey, vars?: Record<string, string | number>): string;
 /** 带单复数的计数文案:n === 1 用 `<base>.one`,其余 `<base>.other`(zh-CN 两键同值)。 */
-export declare function countText(locale: ReportLocale, base: "overview.experiments" | "pointsMissing" | "scoreboard.notRun" | "scoreboard.unscorable" | "scoreboard.ignored" | "entityList.moreFailures" | "table.columnsHidden", n: number): string;
+export declare function countText(locale: ReportLocale, base: "overview.experiments" | "pointsMissing" | "scoreboard.notRun" | "scoreboard.unscorable" | "scoreboard.ignored" | "entityList.moreFailures" | "table.columnsHidden" | "copyFixPrompt.summary" | "traceWaterfall.spans" | "traceWaterfall.failedSpans", n: number): string;
 /**
  * 按 locale 解析指标 / 列 label:字符串原样;字典按 LocalizedText 回退规则取值;
  * undefined 回退 fallback(= metric.name)。渲染面(web / text)共用。

package/dist/report/locale.js

@@ -129,6 +129,44 @@ const en = {
     "scoreboard.subjectTitle": "{questions} evals, weighted {earned} of {possible}",
     "delta.pairHeader": "pair (A → B)",
     "delta.empty": "{experiments} experiments, 0 comparable pairs",
+    /** ScopeWarnings 聚合层的 chrome:汇总行、kind 徽标、组头与明细折叠标签;message 本体不经字典。 */
+    "warnings.summary.experiments.one": "{n} experiment flagged",
+    "warnings.summary.experiments.other": "{n} experiments flagged",
+    "warnings.group.unreadableSnapshot.one": "{n} snapshot skipped",
+    "warnings.group.unreadableSnapshot.other": "{n} snapshots skipped",
+    "warnings.details.one": "{n} warning",
+    "warnings.details.other": "{n} warnings",
+    "warnings.badge.partialCoverage": "coverage {covered}/{total}",
+    "warnings.badge.staleSnapshot": "{gap} behind",
+    "warnings.badge.unfinishedSnapshot": "unfinished",
+    "warnings.gap.second.one": "{n} second",
+    "warnings.gap.second.other": "{n} seconds",
+    "warnings.gap.minute.one": "{n} minute",
+    "warnings.gap.minute.other": "{n} minutes",
+    "warnings.gap.hour.one": "{n} hour",
+    "warnings.gap.hour.other": "{n} hours",
+    "warnings.gap.day.one": "{n} day",
+    "warnings.gap.day.other": "{n} days",
+    /** Hero / HeroCard 的运行 meta(hero.noRuns 是 latestStartedAt 为 null 时的内置文案)。 */
+    "hero.lastRun": "Last run {time}",
+    "hero.noRuns": "No runs yet",
+    /** web 面的合成来源标注(仅 snapshots > 1 时显示)。 */
+    "hero.composedRuns": "composed from {n} runs",
+    /** text 面的合成来源标注(show 页首 meta 行,仅 snapshots > 1 时显示)。 */
+    "hero.composedSnapshots": "composed from {n} snapshots",
+    /** CopyFixPrompt 的 web 面 chrome(prompt 本身面向 agent、固定英文,不经词典)。 */
+    "copyFixPrompt.summary.one": "Fix prompt · {n} failure",
+    "copyFixPrompt.summary.other": "Fix prompt · {n} failures",
+    "copyFixPrompt.copy": "Copy fix prompt",
+    /** TraceWaterfall 的 chrome。 */
+    "traceWaterfall.empty": "No attempts",
+    "traceWaterfall.noTrace": "no trace",
+    "traceWaterfall.spans.one": "{n} span",
+    "traceWaterfall.spans.other": "{n} spans",
+    "traceWaterfall.failedSpans.one": "{n} failed",
+    "traceWaterfall.failedSpans.other": "{n} failed",
+    /** AttemptList 的 web 面过滤框占位符(filter 渐进增强)。 */
+    "attemptList.filterPlaceholder": "Filter attempts…",
     "tabs.tab": "Tab",
 };
 const zhCN = {
@@ -216,6 +254,37 @@ const zhCN = {
     "scoreboard.subjectTitle": "{questions} 道题,加权得分 {earned}/{possible}",
     "delta.pairHeader": "对比 (A → B)",
     "delta.empty": "{experiments} 个实验、0 个可配对",
+    "warnings.summary.experiments.one": "{n} 个实验的数字带警告",
+    "warnings.summary.experiments.other": "{n} 个实验的数字带警告",
+    "warnings.group.unreadableSnapshot.one": "{n} 个快照被跳过",
+    "warnings.group.unreadableSnapshot.other": "{n} 个快照被跳过",
+    "warnings.details.one": "{n} 条原始警告",
+    "warnings.details.other": "{n} 条原始警告",
+    "warnings.badge.partialCoverage": "覆盖 {covered}/{total}",
+    "warnings.badge.staleSnapshot": "落后 {gap}",
+    "warnings.badge.unfinishedSnapshot": "未收尾",
+    "warnings.gap.second.one": "{n} 秒",
+    "warnings.gap.second.other": "{n} 秒",
+    "warnings.gap.minute.one": "{n} 分钟",
+    "warnings.gap.minute.other": "{n} 分钟",
+    "warnings.gap.hour.one": "{n} 小时",
+    "warnings.gap.hour.other": "{n} 小时",
+    "warnings.gap.day.one": "{n} 天",
+    "warnings.gap.day.other": "{n} 天",
+    "hero.lastRun": "最后运行 {time}",
+    "hero.noRuns": "暂无运行",
+    "hero.composedRuns": "由 {n} 次运行合成",
+    "hero.composedSnapshots": "由 {n} 份快照合成",
+    "copyFixPrompt.summary.one": "修复 prompt · {n} 个失败",
+    "copyFixPrompt.summary.other": "修复 prompt · {n} 个失败",
+    "copyFixPrompt.copy": "复制修复 prompt",
+    "traceWaterfall.empty": "没有 attempt",
+    "traceWaterfall.noTrace": "无 trace",
+    "traceWaterfall.spans.one": "{n} 个 span",
+    "traceWaterfall.spans.other": "{n} 个 span",
+    "traceWaterfall.failedSpans.one": "{n} 个失败",
+    "traceWaterfall.failedSpans.other": "{n} 个失败",
+    "attemptList.filterPlaceholder": "筛选 attempt…",
     "tabs.tab": "Tab",
 };
 const dictionaries = {

package/dist/report/react/AttemptList.d.ts

@@ -15,10 +15,12 @@ export declare function AttemptRow({ item, attemptHref, locale, }: {
     attemptHref?: (locator: AttemptLocator) => string;
     locale?: ReportLocale;
 }): ReactElement;
-export declare function AttemptList({ data, total, attemptHref, className, locale, }: {
+export declare function AttemptList({ data, total, filter, attemptHref, className, locale, }: {
     data: readonly AttemptListItem[];
     /** data 被 slice 时的原始数量;如实显示还剩多少条没展示。 */
     total?: number;
+    /** web 面加过滤输入框(按 experiment、eval、agent、verdict 或摘要文本收窄行);渐进增强,不改变数据。 */
+    filter?: boolean;
     attemptHref?: (locator: AttemptLocator) => string;
     className?: string;
     locale?: ReportLocale;

package/dist/report/react/AttemptList.js

@@ -18,9 +18,9 @@ export function failureSummaryText(item, locale) {
 /** 一条 Attempt 的比较卡片;完整 assertions 通过 locator 下钻,不在列表内展开。 */
 export function AttemptRow({ item, attemptHref = DEFAULT_ATTEMPT_HREF, locale = DEFAULT_REPORT_LOCALE, }) {
     const reason = failureSummaryText(item, locale);
-    return (_jsxs("li", { className: cx("nre-attempt", `nre-attempt-${item.verdict}`), children: [_jsxs("div", { className: "nre-attempt-head", children: [_jsx(AttemptLocatorBadge, { item: item, attemptHref: attemptHref }), _jsx("span", { className: "nre-attempt-eval", children: item.evalId }), _jsx("span", { className: "nre-attempt-experiment", children: item.experimentId }), _jsx("span", { className: cx("nre-attempt-agent", "nre-key", colorClassForKey(item.agent)), children: item.agent }), _jsx("span", { className: "nre-attempt-duration", children: formatDurationMs(item.durationMs) }), item.costUSD !== null && _jsx("span", { className: "nre-attempt-cost", children: formatUSD(item.costUSD) })] }), reason && _jsx("p", { className: "nre-attempt-result", children: reason })] }));
+    return (_jsxs("li", { className: cx("nre-attempt", `nre-attempt-${item.verdict}`), "data-nre-verdict": item.verdict, children: [_jsxs("div", { className: "nre-attempt-head", children: [_jsx(AttemptLocatorBadge, { item: item, attemptHref: attemptHref }), _jsx("span", { className: "nre-attempt-eval", children: item.evalId }), _jsx("span", { className: "nre-attempt-experiment", children: item.experimentId }), _jsx("span", { className: cx("nre-attempt-agent", "nre-key", colorClassForKey(item.agent)), children: item.agent }), _jsx("span", { className: "nre-attempt-duration", children: formatDurationMs(item.durationMs) }), item.costUSD !== null && _jsx("span", { className: "nre-attempt-cost", children: formatUSD(item.costUSD) })] }), reason && _jsx("p", { className: "nre-attempt-result", children: reason })] }));
 }
-export function AttemptList({ data, total, attemptHref = DEFAULT_ATTEMPT_HREF, className, locale = DEFAULT_REPORT_LOCALE, }) {
+export function AttemptList({ data, total, filter = false, attemptHref = DEFAULT_ATTEMPT_HREF, className, locale = DEFAULT_REPORT_LOCALE, }) {
     const remaining = (total ?? data.length) - data.length;
-    return (_jsxs("section", { className: cx("nre", "nre-attempt-list", className), children: [data.length === 0 && _jsx("p", { className: "nre-attempt-list-empty", children: localeText(locale, "attemptList.empty") }), _jsx("ul", { className: "nre-attempts", children: data.map((item) => (_jsx(AttemptRow, { item: item, attemptHref: attemptHref, locale: locale }, item.locator))) }), remaining > 0 && (_jsx("p", { className: "nre-truncated", children: localeText(locale, "attemptList.truncated", { n: remaining }) }))] }));
+    return (_jsxs("section", { className: cx("nre", "nre-attempt-list", className), children: [filter && (_jsx("input", { className: "nre-filter", "data-nre-attempt-filter": "", type: "search", placeholder: localeText(locale, "attemptList.filterPlaceholder") })), data.length === 0 && _jsx("p", { className: "nre-attempt-list-empty", children: localeText(locale, "attemptList.empty") }), _jsx("ul", { className: "nre-attempts", children: data.map((item) => (_jsx(AttemptRow, { item: item, attemptHref: attemptHref, locale: locale }, item.locator))) }), remaining > 0 && (_jsx("p", { className: "nre-truncated", children: localeText(locale, "attemptList.truncated", { n: remaining }) }))] }));
 }

package/dist/report/react/CopyFixPrompt.d.ts

@@ -0,0 +1,12 @@
+import type { ReactElement } from "react";
+import type { CopyFixPromptData } from "../types.ts";
+import { type ReportLocale } from "../locale.ts";
+/**
+ * 批量修复 prompt(纯 web 渲染面):折叠块内完整 prompt 文本 + 复制按钮(增强层)。
+ * 嵌入自有 React 页面时配合 `copyFixPromptData()` 使用;failures 为 0 返回 null。
+ */
+export declare function CopyFixPrompt({ data, className, locale, }: {
+    data: CopyFixPromptData;
+    className?: string;
+    locale?: ReportLocale;
+}): ReactElement | null;

package/dist/report/react/CopyFixPrompt.js

@@ -0,0 +1,12 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import { DEFAULT_REPORT_LOCALE, countText, localeText } from "../locale.js";
+import { cx } from "./format.js";
+/**
+ * 批量修复 prompt(纯 web 渲染面):折叠块内完整 prompt 文本 + 复制按钮(增强层)。
+ * 嵌入自有 React 页面时配合 `copyFixPromptData()` 使用;failures 为 0 返回 null。
+ */
+export function CopyFixPrompt({ data, className, locale = DEFAULT_REPORT_LOCALE, }) {
+    if (data.failures === 0)
+        return null;
+    return (_jsxs("details", { className: cx("nre", "nre-copy-fix-prompt", className), children: [_jsx("summary", { className: "nre-copy-fix-prompt-summary", children: countText(locale, "copyFixPrompt.summary", data.failures) }), _jsx("button", { type: "button", className: "nre-copy-fix-prompt-copy", "data-nre-copy": data.prompt, children: localeText(locale, "copyFixPrompt.copy") }), _jsx("pre", { className: "nre-copy-fix-prompt-text", children: data.prompt })] }));
+}

package/dist/report/react/HeroCard.d.ts

@@ -0,0 +1,13 @@
+import type { ReactElement } from "react";
+import type { HeroData } from "../types.ts";
+import { type LocalizedText, type ReportLocale } from "../locale.ts";
+/**
+ * 站点标题区(纯 web 渲染面):`<h1>` 标题 + meta 行 + 品牌行。
+ * 嵌入自有 React 页面时配合 `heroData()` 使用。
+ */
+export declare function HeroCard({ title, data, className, locale, }: {
+    title: LocalizedText;
+    data: HeroData;
+    className?: string;
+    locale?: ReportLocale;
+}): ReactElement;

package/dist/report/react/HeroCard.js

@@ -0,0 +1,35 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import { DEFAULT_REPORT_LOCALE, localeText, resolveLocalizedText } from "../locale.js";
+import { cx } from "./format.js";
+import { PoweredBy } from "./PoweredBy.js";
+/** ISO 时间 → 按 locale 的「最后运行」显示(年月日 时:分);不可解析原样返回。 */
+function formatLastRun(iso, locale) {
+    const date = new Date(iso);
+    if (Number.isNaN(date.valueOf()))
+        return iso;
+    try {
+        return new Intl.DateTimeFormat(locale, {
+            year: "numeric",
+            month: "short",
+            day: "numeric",
+            hour: "2-digit",
+            minute: "2-digit",
+        }).format(date);
+    }
+    catch {
+        return iso;
+    }
+}
+/**
+ * 站点标题区(纯 web 渲染面):`<h1>` 标题 + meta 行 + 品牌行。
+ * 嵌入自有 React 页面时配合 `heroData()` 使用。
+ */
+export function HeroCard({ title, data, className, locale = DEFAULT_REPORT_LOCALE, }) {
+    const meta = data.latestStartedAt === null
+        ? localeText(locale, "hero.noRuns")
+        : [
+            localeText(locale, "hero.lastRun", { time: formatLastRun(data.latestStartedAt, locale) }),
+            ...(data.snapshots > 1 ? [localeText(locale, "hero.composedRuns", { n: data.snapshots })] : []),
+        ].join(" · ");
+    return (_jsxs("header", { className: cx("nre", "nre-hero", className), children: [_jsx("h1", { className: "nre-hero-title", children: resolveLocalizedText(title, locale) }), _jsx("p", { className: "nre-hero-meta", children: meta }), _jsx(PoweredBy, {})] }));
+}

package/dist/report/react/PoweredBy.d.ts

@@ -0,0 +1,5 @@
+import type { ReactElement } from "react";
+/** 品牌行的固定去处:niceeval 官网,utm 标记「来自报告的品牌行」。 */
+export declare const POWERED_BY_HREF = "https://niceeval.com/?utm_source=report&utm_medium=powered-by";
+/** 一行品牌色小字 `Powered by NiceEval`,链接官网;HeroCard 的品牌行与它同一渲染。 */
+export declare function PoweredBy(): ReactElement;

package/dist/report/react/PoweredBy.js

@@ -0,0 +1,7 @@
+import { jsx as _jsx } from "react/jsx-runtime";
+/** 品牌行的固定去处:niceeval 官网,utm 标记「来自报告的品牌行」。 */
+export const POWERED_BY_HREF = "https://niceeval.com/?utm_source=report&utm_medium=powered-by";
+/** 一行品牌色小字 `Powered by NiceEval`,链接官网;HeroCard 的品牌行与它同一渲染。 */
+export function PoweredBy() {
+    return (_jsx("p", { className: "nre nre-powered-by", children: _jsx("a", { href: POWERED_BY_HREF, target: "_blank", rel: "noopener", children: "Powered by NiceEval" }) }));
+}

package/dist/report/react/ScopeWarnings.d.ts

@@ -0,0 +1,12 @@
+import type { ReactElement } from "react";
+import type { ScopeWarning } from "../types.ts";
+import { type ReportLocale } from "../locale.ts";
+/**
+ * 选择警告区(纯 web 渲染面):按动作聚合的警告组。嵌入自有 React 页面时传
+ * `data={scope.warnings}`;空集返回 null,不渲染空容器。
+ */
+export declare function ScopeWarnings({ data, className, locale, }: {
+    data: readonly ScopeWarning[];
+    className?: string;
+    locale?: ReportLocale;
+}): ReactElement | null;

package/dist/report/react/ScopeWarnings.js

@@ -0,0 +1,18 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import { DEFAULT_REPORT_LOCALE } from "../locale.js";
+import { groupScopeWarnings, warningDetailsLabel } from "../scope-warnings.js";
+import { cx } from "./format.js";
+/** 命令的可复制块:无 JS 时点击全选(user-select: all),enhance.js 在场时点击复制。 */
+function CommandBlock({ command }) {
+    return (_jsx("code", { className: "nre-warning-command", "data-nre-copy": command, children: command }));
+}
+/**
+ * 选择警告区(纯 web 渲染面):按动作聚合的警告组。嵌入自有 React 页面时传
+ * `data={scope.warnings}`;空集返回 null,不渲染空容器。
+ */
+export function ScopeWarnings({ data, className, locale = DEFAULT_REPORT_LOCALE, }) {
+    if (data.length === 0)
+        return null;
+    const { summary, groups, detailsOpen } = groupScopeWarnings(data, locale);
+    return (_jsx("div", { className: cx("nre", "nre-scope-warnings", className), children: _jsxs("details", { className: "nre-warnings", children: [_jsx("summary", { className: "nre-warnings-summary", children: summary }), _jsx("ul", { className: "nre-warning-groups", children: groups.map((group, i) => (_jsxs("li", { className: "nre-warning-group", "data-category": group.category, children: [_jsxs("div", { className: "nre-warning-head", children: [_jsx("span", { className: "nre-warning-title", children: group.title }), group.badges.map((badge, j) => (_jsx("span", { className: "nre-warning-badge", "data-kind": badge.kind, children: badge.text }, j))), group.headCommand !== null && _jsx(CommandBlock, { command: group.headCommand })] }), _jsxs("details", { className: "nre-warning-details", open: detailsOpen || undefined, children: [_jsx("summary", { children: warningDetailsLabel(locale, group.warnings.length) }), _jsx("ul", { children: group.warnings.map((w, j) => (_jsxs("li", { className: "nre-warning", "data-kind": w.kind, children: [w.message, group.headCommand === null && "command" in w && w.command !== undefined && (_jsx(CommandBlock, { command: w.command }))] }, j))) })] })] }, i))) })] }) }));
+}

package/dist/report/react/TraceWaterfall.d.ts

@@ -0,0 +1,14 @@
+import type { ReactElement } from "react";
+import type { TraceWaterfallRow } from "../types.ts";
+import type { AttemptLocator } from "../../results/locator.ts";
+import { type ReportLocale } from "../locale.ts";
+/**
+ * 执行时间瀑布(纯 web 渲染面):嵌入自有 React 页面时配合 `traceWaterfallData()` 使用。
+ * 只画被测 agent 的原始 span;runner 生命周期节点不在 data 里,组合视图归 attempt 详情。
+ */
+export declare function TraceWaterfall({ data, attemptHref, className, locale, }: {
+    data: readonly TraceWaterfallRow[];
+    attemptHref?: (locator: AttemptLocator) => string;
+    className?: string;
+    locale?: ReportLocale;
+}): ReactElement;

package/dist/report/react/TraceWaterfall.js

@@ -0,0 +1,22 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+import { DEFAULT_REPORT_LOCALE, countText, localeText } from "../locale.js";
+import { cx, formatDurationMs } from "./format.js";
+const DEFAULT_ATTEMPT_HREF = (locator) => `#/attempt/${locator}`;
+function pct(part, total) {
+    if (total <= 0)
+        return "0%";
+    return `${Math.min(100, Math.max(0, (part / total) * 100)).toFixed(2)}%`;
+}
+/**
+ * 执行时间瀑布(纯 web 渲染面):嵌入自有 React 页面时配合 `traceWaterfallData()` 使用。
+ * 只画被测 agent 的原始 span;runner 生命周期节点不在 data 里,组合视图归 attempt 详情。
+ */
+export function TraceWaterfall({ data, attemptHref = DEFAULT_ATTEMPT_HREF, className, locale = DEFAULT_REPORT_LOCALE, }) {
+    return (_jsxs("section", { className: cx("nre", "nre-trace-waterfall", className), children: [data.length === 0 && _jsx("p", { className: "nre-waterfall-empty", children: localeText(locale, "traceWaterfall.empty") }), _jsx("ul", { className: "nre-waterfall", children: data.map((row) => {
+                    const failedSpans = row.spans.filter((span) => span.failed).length;
+                    return (_jsxs("li", { className: "nre-waterfall-row", children: [_jsxs("div", { className: "nre-waterfall-head", children: [_jsx("a", { className: "nre-locator", href: attemptHref(row.locator), children: row.locator }), _jsx("span", { className: "nre-waterfall-eval", children: row.evalId }), _jsx("span", { className: "nre-waterfall-experiment", children: row.experimentId }), _jsx("span", { className: "nre-waterfall-duration", children: row.durationMs === null ? localeText(locale, "traceWaterfall.noTrace") : formatDurationMs(row.durationMs) }), _jsx("span", { className: "nre-waterfall-count", children: countText(locale, "traceWaterfall.spans", row.spans.length) }), failedSpans > 0 && (_jsxs("span", { className: "nre-waterfall-failed", children: ["\u2717 ", countText(locale, "traceWaterfall.failedSpans", failedSpans)] }))] }), row.durationMs !== null && row.spans.length > 0 && (_jsx("div", { className: "nre-waterfall-track", children: row.spans.map((span, i) => (_jsx("span", { className: cx("nre-waterfall-span", `nre-span-${span.kind}`, span.failed && "nre-span-failed"), style: {
+                                        left: pct(span.startOffsetMs, row.durationMs),
+                                        width: `max(${pct(span.durationMs, row.durationMs)}, 0.5%)`,
+                                    }, title: `${span.name} · ${formatDurationMs(span.durationMs)}${span.failed ? " · ✗" : ""}` }, i))) }))] }, row.locator));
+                }) })] }));
+}

package/dist/report/react/index.d.ts

@@ -10,7 +10,12 @@ export { MetricScatter } from "./MetricScatter.tsx";
 export { MetricLine } from "./MetricLine.tsx";
 export { DeltaTable } from "./DeltaTable.tsx";
 export { Scoreboard } from "./Scoreboard.tsx";
-export type { AttemptListItem, AttemptLocator, DeltaData, EvalListItem, ExperimentComparisonData, ExperimentComparisonGroupData, ExperimentListEvalRow, ExperimentListItem, LineData, MatrixData, MetricCell, MetricColumn, ScatterData, ScopeSummaryData, ScopeWarning, ScoreboardData, TableData, VerdictTally, } from "../types.ts";
+export { HeroCard } from "./HeroCard.tsx";
+export { PoweredBy } from "./PoweredBy.tsx";
+export { ScopeWarnings } from "./ScopeWarnings.tsx";
+export { CopyFixPrompt } from "./CopyFixPrompt.tsx";
+export { TraceWaterfall } from "./TraceWaterfall.tsx";
+export type { AttemptListItem, AttemptLocator, CopyFixPromptData, DeltaData, EvalListItem, ExperimentComparisonData, ExperimentComparisonGroupData, ExperimentListEvalRow, ExperimentListItem, HeroData, LineData, MatrixData, MetricCell, MetricColumn, ScatterData, ScopeSummaryData, ScopeWarning, ScoreboardData, TableData, TraceSpanSummary, TraceWaterfallRow, VerdictTally, } from "../types.ts";
 export { DEFAULT_REPORT_LOCALE, resolveLocalizedText, resolveMetricLabel } from "../locale.ts";
 export type { LocalizedText, ReportLocale } from "../locale.ts";
 export { NRE_PALETTE, colorClassForKey, colorHexForKey, colorIndexForKey, seriesClassForKey } from "./colors.ts";

package/dist/report/react/index.js

@@ -19,6 +19,12 @@ export { MetricScatter } from "./MetricScatter.js";
 export { MetricLine } from "./MetricLine.js";
 export { DeltaTable } from "./DeltaTable.js";
 export { Scoreboard } from "./Scoreboard.js";
+// 站点组件的纯 web 面(data 形态;Hero 是组合组件,只住 niceeval/report)
+export { HeroCard } from "./HeroCard.js";
+export { PoweredBy } from "./PoweredBy.js";
+export { ScopeWarnings } from "./ScopeWarnings.js";
+export { CopyFixPrompt } from "./CopyFixPrompt.js";
+export { TraceWaterfall } from "./TraceWaterfall.js";
 // locale(官方组件 chrome 文案;LocalizedText 的按 locale 解析也用它)
 export { DEFAULT_REPORT_LOCALE, resolveLocalizedText, resolveMetricLabel } from "../locale.js";
 // 稳定配色(自定义组件想与官方组件同键同色时用;seriesClassForKey 配 CSS 的 --nre-series)

package/dist/report/report.d.ts

@@ -37,11 +37,11 @@ export type HeadTag = {
     children?: string;
 };
 export interface ReportShell {
-    /** 标题:首页 hero 与浏览器标题。页头左端是恒定的 NiceEval 品牌字标,不由 title 覆盖;回退链 def.title → 唯一快照 name → 内置文案「Eval 运行结果 / Eval Results」。 */
+    /** 站点标题:浏览器标题、show 页索引标题行与 `ctx.report.title` 的取值源;`Hero` 组件缺省消费它。回退链 def.title → 唯一快照 name → 内置文案「Eval 运行结果 / Eval Results」。 */
     title?: LocalizedText;
     /** 页头右侧的外部链接,如 GitHub、文档、CI。 */
     links?: ReportLink[];
-    /** 每页页脚的一段文字;省略时不渲染页脚(品牌行恒在 hero 下方,不占页脚)。 */
+    /** 每页页脚的一段文字;省略时不渲染页脚(品牌行归 PoweredBy 组件,不占页脚)。 */
     footer?: LocalizedText;
     /**
      * 注入每页 `<head>` 的结构化标签,在官方与外壳样式之后按声明顺序渲染。
@@ -63,13 +63,27 @@ export interface ReportPage {
     /** 这一页的报告树;ReportDefinition 不是 ReportNode,页装不进外壳。 */
     content: ReportNode;
 }
-/** content / pages 互斥由类型表达,不把非法状态留到运行期。 */
+/** content / pages / extends 三选一由类型表达,不把非法状态留到运行期。 */
 export type ReportDef = ReportShell & ({
+    /** 单页缩写,等价于只含 id `report` 的页列表。 */
     content: ReportNode;
     pages?: never;
+    extends?: never;
 } | {
+    /** 非空页列表;导航按数组顺序显示。 */
     pages: NonEmptyArray<ReportPage>;
     content?: never;
+    extends?: never;
+} | {
+    /**
+     * 在另一份报告上叠外壳:页列表取 base 的页列表;本对象声明的外壳字段整字段覆盖
+     * base 的同名字段,未声明的沿用 base——没有数组拼接、没有深合并。base 是任何
+     * `defineReport` 产物(内建视图或自己别的报告文件的具名导出);合并在
+     * `defineReport` 调用时折叠完成,产物仍是普通 ReportDefinition,可以再被 extends。
+     */
+    extends: ReportDefinition;
+    content?: never;
+    pages?: never;
 });
 /**
  * defineReport 的唯一产物:只作 --report 文件的默认导出,交给宿主装载。
@@ -136,7 +150,9 @@ export interface RenderReportTextOptions extends TextRenderOptions {
 }
 /**
  * text 宿主的装载语义:选页 → resolve(组合展开 + spec 取数,唯一的 await 边界)→ 树校验 →
- * 遍历渲染 text 面;Scope 有挑选警告时在报告顶部前置一块 "! <message>"。不需要 react-dom。
+ * 遍历渲染 text 面。不需要 react-dom。宿主不在报告树外另设警告通道——挑选警告的呈现件是
+ * `ScopeWarnings` 组件,内建报告每页都放它,自定义报告放不放是作者义务
+ * (docs/feature/reports/architecture.md「Scope 是计算入口」)。
  */
 export declare function renderReportToText(definition: ReportDefinition, ctx: ReportHostContext, options?: RenderReportTextOptions): Promise<string>;
 /** 页索引标题行(show 多页索引 / view 导航共用的解析结果):按 locale 解析的标题字符串。 */
@@ -161,7 +177,7 @@ export interface RenderTreeTextOptions extends TextRenderOptions {
 }
 /**
  * 渲染一页报告树的 text 面(宿主逐页调用;页选择归宿主):
- * resolve(组合展开 + spec 取数)→ validate → render。Scope 有挑选警告时在页顶前置
- * "! <message>" 块——宿主是 warning 的唯一呈现者,组件数据不复制 warning。
+ * resolve(组合展开 + spec 取数)→ validate → render。宿主不在报告树外另设警告通道,
+ * 挑选警告由页内的 `ScopeWarnings` 组件呈现(内建报告每页都放它)。
  */
 export declare function renderReportTreeToText(tree: ReportNode, ctx: ReportTreeHostContext, options?: RenderTreeTextOptions): Promise<string>;