niceeval 0.6.1 → 0.6.2

package/dist/report/react/chart-math.js

@@ -1,5 +1,4 @@
-// 通用图表数值工具:round-number 刻度生成与标签防撞布局,零依赖纯函数。
-// 从 src/view/app/lib/chart.ts 迁来(view 侧从这里 re-export),报告组件与 view 共用同一份。
+// 通用图表数值工具:round-number 刻度生成与点标签布局,零依赖纯函数。
 function niceNumber(range, round) {
     if (range <= 0)
         return 1;
@@ -43,26 +42,79 @@ export function niceTicks(min, max, count = 5) {
         ticks.push(Number(v.toFixed(10)));
     return ticks;
 }
+/** 标签框高度(基线上方 10px、下方 2px,对应 11px 字号)。 */
+const LABEL_ASCENT = 10;
+const LABEL_DESCENT = 2;
+/** 数据点视为边长 2×DOT_R 的方形障碍(点半径 4.5 + 命中余量)。 */
+const DOT_R = 6;
+/** 紧邻位与点的水平间距;候选环按它逐环外扩。 */
+const NEAR = 10;
+const RINGS = [NEAR, NEAR + 12, NEAR + 26];
+function labelBoxAt(x, y, anchor, width) {
+    const x0 = anchor === "start" ? x : anchor === "end" ? x - width : x - width / 2;
+    return { x0, y0: y - LABEL_ASCENT, x1: x0 + width, y1: y + LABEL_DESCENT };
+}
+function overlapArea(a, b) {
+    const w = Math.min(a.x1, b.x1) - Math.max(a.x0, b.x0);
+    const h = Math.min(a.y1, b.y1) - Math.max(a.y0, b.y0);
+    return w > 0 && h > 0 ? w * h : 0;
+}
+function outOfBoundsArea(box, bounds) {
+    const total = (box.x1 - box.x0) * (box.y1 - box.y0);
+    const w = Math.min(box.x1, bounds.x1) - Math.max(box.x0, bounds.x0);
+    const h = Math.min(box.y1, bounds.y1) - Math.max(box.y0, bounds.y0);
+    const inside = w > 0 && h > 0 ? w * h : 0;
+    return total - inside;
+}
+/** 一个点在距离 d 的候选环:左右紧邻、四个斜角、正上正下。顺序即同代价时的偏好序。 */
+function ringCandidates(p, d) {
+    const diag = d * 0.75;
+    return [
+        { x: p.cx + d, y: p.cy + 4, anchor: "start" },
+        { x: p.cx - d, y: p.cy + 4, anchor: "end" },
+        { x: p.cx + diag, y: p.cy - diag - 2, anchor: "start" },
+        { x: p.cx - diag, y: p.cy - diag - 2, anchor: "end" },
+        { x: p.cx, y: p.cy - d - LABEL_DESCENT - 2, anchor: "middle" },
+        { x: p.cx + diag, y: p.cy + diag + LABEL_ASCENT, anchor: "start" },
+        { x: p.cx - diag, y: p.cy + diag + LABEL_ASCENT, anchor: "end" },
+        { x: p.cx, y: p.cy + d + LABEL_ASCENT + 2, anchor: "middle" },
+    ];
+}
 /**
- * 贪心地把互相重叠的直接标签往下推开(按 y 排序、逐个查重叠再下移),点本身位置不变。
- * 散点常见同分不同价的聚簇,这一步避免标签叠成一团糊字。
+ * 散点直接标签的候选位择优布局(docs/feature/reports/library.md「MetricScatter」):
+ * 每个标签在点四周由近及远的候选环上打分——与已放置标签的重叠、与任何数据点的重叠、
+ * 越出画布的面积、离点距离逐项累加,取代价最小的候选。存在无冲突候选时标签必不遮点、
+ * 不叠标签、不越界;全候选冲突时取重叠最小者,绝不丢标签。重合点簇因此向不同方向散开,
+ * 而不是单向堆叠压到下方的数据点上。
  */
-export function layoutLabelOffsets(items, lineHeight = 14, padX = 4) {
-    const offsets = new Array(items.length).fill(0);
-    const order = [...items.keys()].sort((a, b) => items[a].cy - items[b].cy || items[a].cx - items[b].cx);
-    const placed = [];
+export function placePointLabels(points, bounds) {
+    const dots = points.map((p) => ({ x0: p.cx - DOT_R, y0: p.cy - DOT_R, x1: p.cx + DOT_R, y1: p.cy + DOT_R }));
+    const placedBoxes = [];
+    const result = new Array(points.length);
+    // 由上到下、由左到右逐个放置,布局确定可复现
+    const order = [...points.keys()].sort((a, b) => points[a].cy - points[b].cy || points[a].cx - points[b].cx);
     for (const i of order) {
-        const it = items[i];
-        const x0 = it.anchorLeft ? it.cx - it.width : it.cx;
-        const x1 = it.anchorLeft ? it.cx : it.cx + it.width;
-        let y = it.cy;
-        let guard = 0;
-        while (guard < 40 && placed.some((p) => x0 < p.x1 + padX && x1 > p.x0 - padX && Math.abs(y - p.y) < lineHeight)) {
-            y += lineHeight;
-            guard++;
+        const p = points[i];
+        let best = null;
+        for (let ring = 0; ring < RINGS.length; ring++) {
+            const candidates = ringCandidates(p, RINGS[ring]);
+            for (let c = 0; c < candidates.length; c++) {
+                const candidate = candidates[c];
+                const box = labelBoxAt(candidate.x, candidate.y, candidate.anchor, p.width);
+                let cost = 0;
+                for (const placed of placedBoxes)
+                    cost += overlapArea(box, placed) * 4;
+                for (const dot of dots)
+                    cost += overlapArea(box, dot) * 3;
+                cost += outOfBoundsArea(box, bounds) * 6;
+                cost += ring * 8 + c * 0.5; // 近环、偏好序靠前者优先,只作同冲突程度下的排序
+                if (best === null || cost < best.cost) {
+                    best = { candidate, box, cost, leader: !(ring === 0 && c <= 1) };
+                }
+            }
         }
-        placed.push({ x0, x1, y });
-        offsets[i] = y - it.cy;
+        placedBoxes.push(best.box);
+        result[i] = { ...best.candidate, leader: best.leader };
     }
-    return offsets;
+    return result;
 }

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

@@ -6,14 +6,14 @@ export declare const overviewData: OverviewData;
 export declare const overviewWithCost: OverviewData;
 /** GroupSummary.data 的产物形态:eval 级折叠计票 + 旧 GroupSelector 口径的通过率(包成 MetricCell)。 */
 export declare const groupSummaryData: GroupSummaryData;
-export declare const tableData: TableData<"pass-rate" | "code-lines">;
+export declare const tableData: TableData<"task-pass-rate" | "code-lines">;
 /** rows: "experiment" 的榜单形态:行携带 agent/model 元信息与 eval 级折叠计票。 */
-export declare const tableDataWithMeta: TableData<"pass-rate">;
+export declare const tableDataWithMeta: TableData<"task-pass-rate">;
 export declare const matrixData: MatrixData;
 export declare const scoreboardData: ScoreboardData;
 export declare const scatterData: ScatterData;
 export declare const lineData: LineData;
-export declare const deltaData: DeltaData<"pass-rate" | "cost">;
+export declare const deltaData: DeltaData<"task-pass-rate" | "cost">;
 export declare const attemptListItems: AttemptListItem[];
 export declare const evalListItems: EvalListItem[];
 export declare const experimentListItems: ExperimentListItem[];

package/dist/report/react/fixtures.js

@@ -3,7 +3,7 @@
 // 数字刻意覆盖诚实细节:coverage 角标(samples<total)、全 null 格子、
 // 稀疏矩阵、缺数据的散点、delta 的 null 不硬算、truncated 计数。
 const locator = (s) => s;
-export const passRateColumn = { key: "pass-rate", label: "pass rate", unit: "%", better: "higher" };
+export const passRateColumn = { key: "task-pass-rate", label: "pass rate", unit: "%", better: "higher" };
 export const codeLinesColumn = { key: "code-lines", label: "code lines", unit: "lines", better: "lower" };
 export const costColumn = { key: "cost", label: "cost", unit: "$", better: "lower" };
 export const overviewData = {
@@ -59,7 +59,7 @@ export const tableData = {
         {
             key: "codex",
             cells: {
-                "pass-rate": { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
+                "task-pass-rate": { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
                 // 全 null:一个有效样本都没有 → 缺数据文案,绝不画 0
                 "code-lines": { value: null, display: "—", samples: 0, total: 6, refs: [] },
             },
@@ -67,7 +67,7 @@ export const tableData = {
         {
             key: "bub",
             cells: {
-                "pass-rate": {
+                "task-pass-rate": {
                     value: 0.87,
                     display: "87%",
                     samples: 6,
@@ -87,7 +87,7 @@ export const tableDataWithMeta = {
     rows: [
         {
             key: "compare/bub",
-            cells: { "pass-rate": { value: 0.5, display: "50%", samples: 2, total: 2, refs: [] } },
+            cells: { "task-pass-rate": { value: 0.5, display: "50%", samples: 2, total: 2, refs: [] } },
             meta: {
                 agent: "bub",
                 model: "gpt-5.4",
@@ -96,7 +96,7 @@ export const tableDataWithMeta = {
         },
         {
             key: "compare/codex",
-            cells: { "pass-rate": { value: 1, display: "100%", samples: 2, total: 2, refs: [] } },
+            cells: { "task-pass-rate": { value: 1, display: "100%", samples: 2, total: 2, refs: [] } },
             meta: {
                 agent: "codex",
                 verdicts: { passed: 2, failed: 0, errored: 0, skipped: 0 },
@@ -242,7 +242,7 @@ export const deltaData = {
             b: { experimentId: "compare/bub--agents-md" },
             cells: {
                 // 通过率 +12pp:better higher → 好(绿)
-                "pass-rate": {
+                "task-pass-rate": {
                     a: { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
                     b: { value: 0.62, display: "62%", samples: 6, total: 6, refs: [] },
                     delta: 0.12,
@@ -262,7 +262,7 @@ export const deltaData = {
             a: { experimentId: "compare/codex" },
             b: { experimentId: "compare/codex--agents-md" },
             cells: {
-                "pass-rate": {
+                "task-pass-rate": {
                     a: { value: 0.4, display: "40%", samples: 6, total: 6, refs: [] },
                     b: { value: 0.4, display: "40%", samples: 6, total: 6, refs: [] },
                     delta: 0,
@@ -292,8 +292,9 @@ const failedAttempt = {
             name: "roots-correct",
             severity: "gate",
             score: 0,
-            passed: false,
-            detail: "expected x=2, got x=3",
+            outcome: "failed",
+            expected: "x=2",
+            received: "x=3",
             evidence: "judge: sign flipped when substituting into the quadratic formula",
         },
     ],
@@ -307,7 +308,17 @@ const erroredAttempt = {
     attempt: 0,
     agent: "codex",
     verdict: "errored",
-    error: "TypeError: cannot read properties of undefined (reading 'foo')",
+    // 结构化 error:列表只渲染 message 一层摘要;cause/stack 与 diagnostics 是下钻详情,随数据携带
+    error: {
+        code: "unexpected-error",
+        message: "TypeError: cannot read properties of undefined (reading 'foo')",
+        phase: "eval.run",
+        stack: "TypeError: cannot read properties of undefined (reading 'foo')\n    at run (adapter.ts:42:7)",
+        cause: { name: "TypeError", message: "cannot read properties of undefined (reading 'foo')" },
+    },
+    diagnostics: [
+        { code: "sandbox-teardown-failed", level: "warning", message: "sandbox teardown timed out", phase: "sandbox.teardown" },
+    ],
     assertions: [],
     durationMs: 4_500,
     locator: locator("@1c1c1c1"),
@@ -324,7 +335,6 @@ export const evalListItems = [
         evalId: "algebra/quadratic",
         experimentId: "compare/bub",
         verdict: "failed",
-        reason: "roots-correct: expected x=2, got x=3",
         score: { value: 0, display: "0%", samples: 1, total: 1, refs: [failedAttempt.locator] },
         duration: { value: 32_000, display: "32.0s", samples: 1, total: 1, refs: [failedAttempt.locator] },
         cost: { value: 0.12, display: "$0.12", samples: 1, total: 1, refs: [failedAttempt.locator] },
@@ -334,7 +344,6 @@ export const evalListItems = [
         evalId: "geometry/angles",
         experimentId: "compare/codex",
         verdict: "errored",
-        reason: erroredAttempt.error,
         score: { value: 0, display: "0%", samples: 1, total: 1, refs: [erroredAttempt.locator] },
         duration: { value: 4_500, display: "4.5s", samples: 1, total: 1, refs: [erroredAttempt.locator] },
         cost: { value: null, display: "—", samples: 0, total: 1, refs: [] },
@@ -359,7 +368,6 @@ export const experimentListItems = [
             {
                 evalId: "algebra/quadratic",
                 verdict: "failed",
-                reason: "roots-correct: expected x=2, got x=3",
                 duration: { value: 32_000, display: "32.0s", samples: 1, total: 1, refs: [failedAttempt.locator] },
                 cost: { value: 0.12, display: "$0.12", samples: 1, total: 1, refs: [failedAttempt.locator] },
                 attempts: [failedAttempt, failedRetryAttempt],
@@ -388,7 +396,6 @@ export const experimentListItems = [
             {
                 evalId: "geometry/angles",
                 verdict: "errored",
-                reason: erroredAttempt.error,
                 duration: { value: 4_500, display: "4.5s", samples: 1, total: 1, refs: [erroredAttempt.locator] },
                 cost: { value: null, display: "—", samples: 0, total: 1, refs: [] },
                 attempts: [erroredAttempt],

package/dist/report/report.d.ts

@@ -13,7 +13,11 @@ export interface ReportDefinition {
     [REPORT_DEFINITION]: true;
 }
 export declare function defineReport(build: (ctx: ReportContext) => ReportNode | Promise<ReportNode>): ReportDefinition;
-/** 宿主装载报告文件时用:默认导出是不是 defineReport 的产物。 */
+/**
+ * 宿主装载报告文件时用:默认导出是不是 ReportDefinition。defineReport 的产物是普通对象;
+ * 「组件兼报告」(把 defineReport 产物 Object.assign 到双面组件上)是挂了 build 面的
+ * 可调用函数,同样算数——判据只看 build 面与标记,不看宿主形态。
+ */
 export declare function isReportDefinition(value: unknown): value is ReportDefinition;
 /**
  * text 宿主的装载语义:build → 渲染前解析数据组件(唯一的 await 边界)→ 树校验 → 遍历渲染

package/dist/report/report.js

@@ -14,9 +14,13 @@ export function defineReport(build) {
     }
     return { build, [REPORT_DEFINITION]: true };
 }
-/** 宿主装载报告文件时用:默认导出是不是 defineReport 的产物。 */
+/**
+ * 宿主装载报告文件时用:默认导出是不是 ReportDefinition。defineReport 的产物是普通对象;
+ * 「组件兼报告」(把 defineReport 产物 Object.assign 到双面组件上)是挂了 build 面的
+ * 可调用函数,同样算数——判据只看 build 面与标记,不看宿主形态。
+ */
 export function isReportDefinition(value) {
-    return (typeof value === "object" &&
+    return ((typeof value === "object" || typeof value === "function") &&
         value !== null &&
         value[REPORT_DEFINITION] === true &&
         typeof value.build === "function");

package/dist/report/text/faces.d.ts

@@ -24,6 +24,6 @@ export declare function scoreboardText(data: ScoreboardData, ctx: TextContext):
 export declare function scatterText(data: ScatterData, ctx: TextContext): string;
 export declare function lineText(data: LineData, ctx: TextContext): string;
 export declare function deltaText(data: DeltaData, ctx: TextContext): string;
-export declare function experimentListText(items: ExperimentListItem[], ctx: TextContext): string;
+export declare function experimentListText(items: ExperimentListItem[], ctx: TextContext, relativeTo?: string): string;
 export declare function evalListText(items: EvalListItem[], ctx: TextContext): string;
 export declare function attemptListText(items: AttemptListItem[], total: number | undefined, ctx: TextContext): string;

package/dist/report/text/faces.js

@@ -4,9 +4,9 @@
 // 零 react、零 IO、纯同步 —— 这是 text 宿主不需要 react-dom 的那一半。
 // chrome 文案(注脚、verdict 词、截断提示)经 ctx.locale 查 locale 字典,
 // 默认 en 与历史输出逐字一致;数据(display、键、warnings message)不本地化。
-import { attemptItemReason, formatDurationMs, formatMetricValue, formatPlainNumber, formatUSD, verdictMark, } from "../format.js";
+import { attemptItemReason, experimentDisplayName, formatDurationMs, formatMetricValue, formatPlainNumber, formatUSD, verdictMark, } from "../format.js";
 import { countText, localeText, resolveMetricLabel } from "../locale.js";
-import { indentBlock, padDisplay, textBar, wrapDisplay } from "./layout.js";
+import { padDisplay, stringWidth, textBar, wrapDisplay } from "./layout.js";
 import { renderTableText } from "./table.js";
 import { renderCharPlot, renderCoordinateTable } from "./plot.js";
 const MISSING_MARK = "—";
@@ -399,12 +399,12 @@ export function deltaText(data, ctx) {
 // 三面共用的紧凑标记:`locator✓`(判定符紧跟 locator,中间不留空格)。
 // ExperimentList / EvalList 逐 attempt 只列这一个标记 + 各自的原因/耗时摘要,不重复整段
 // niceeval show 命令;要看某个 attempt 的完整证据,agent 自己拼 `niceeval show <locator>`——
-// 命令模板只在 AttemptList(叶子层)展示完整断言明细时才值得,不在中间层重复。
+// 比较列表都不内联完整断言或命令模板。
 function locatorBadge(item) {
     return `${item.locator}${verdictMark(item.verdict)}`;
 }
 // ── ExperimentList ──
-function experimentSummaryTable(items, ctx) {
+function experimentSummaryTable(items, ctx, relativeTo) {
     const locale = ctx.locale;
     const compact = ctx.width < 100;
     const columns = [
@@ -420,7 +420,7 @@ function experimentSummaryTable(items, ctx) {
     const rows = items.map((item) => ({
         key: item.experimentId,
         cells: {
-            experiment: item.experimentId,
+            experiment: experimentDisplayName(item.experimentId, relativeTo),
             model: item.model ?? localeText(locale, "experimentList.defaultModel"),
             agent: item.agent,
             duration: cellText(item.duration),
@@ -430,18 +430,25 @@ function experimentSummaryTable(items, ctx) {
             cost: cellText(item.cost),
         },
     }));
-    const metadata = items.flatMap((item) => wrapDisplay(`${item.experimentId}: ${localeText(locale, "overview.evalsCount", { n: item.evals })} · ${localeText(locale, "overview.attemptsCount", { n: item.attempts })} · ${item.lastRunAt}`, Math.max(8, ctx.width - 2)).map((line) => `  ${line}`));
+    const metadata = items.flatMap((item) => wrapDisplay(`${experimentDisplayName(item.experimentId, relativeTo)}: ${localeText(locale, "overview.evalsCount", { n: item.evals })} · ${localeText(locale, "overview.attemptsCount", { n: item.attempts })} · ${item.lastRunAt}`, Math.max(8, ctx.width - 2)).map((line) => `  ${line}`));
     return [renderTableText({ columns, rows, locale }, ctx), metadata.join("\n")].join("\n");
 }
-function experimentDetailTable(item, ctx) {
+function experimentDetailTable(item, ctx, relativeTo) {
     const locale = ctx.locale;
     const columns = [
         { key: "status", header: localeText(locale, "experimentList.status") },
         { key: "entity", header: localeText(locale, "experimentList.evalAttempt") },
-        { key: "result", header: localeText(locale, "experimentList.result") },
+        // Result 是可扫读的失败预览,不是证据面:两行放不下的以 … 收口,完整值走 locator 下钻。
+        { key: "result", header: localeText(locale, "experimentList.result"), maxLines: 2 },
         { key: "duration", header: localeText(locale, "experimentList.duration"), align: "right" },
         { key: "cost", header: localeText(locale, "experimentList.cost"), align: "right" },
     ];
+    // Result 的字符预算 ≈ 两行 × 它能分到的列宽(总宽减其它列的自然宽与列距)。这里只做
+    // 优先级让位(标题先截、received 最后截)的粗预算;精确的按宽度收口由列的 maxLines 兜底。
+    const statusWidth = Math.max(stringWidth(localeText(locale, "experimentList.status")), ...item.evalRows.map((row) => stringWidth(`${verdictMark(row.verdict)} ${localeText(locale, `verdict.${row.verdict}`)}`)));
+    const entityWidth = Math.max(stringWidth(localeText(locale, "experimentList.evalAttempt")), ...item.evalRows.flatMap((row) => [stringWidth(row.evalId), ...row.attempts.map((a) => stringWidth(a.locator) + 3)]));
+    const fixedWidth = statusWidth + entityWidth + 8 /* duration */ + 6 /* cost */ + 3 * 4; /* 4 段列距 */
+    const resultBudget = Math.max(24, (ctx.width - fixedWidth) * 2);
     const rows = item.evalRows.flatMap((row) => {
         const parent = {
             key: row.evalId,
@@ -449,8 +456,8 @@ function experimentDetailTable(item, ctx) {
                 status: `${verdictMark(row.verdict)} ${localeText(locale, `verdict.${row.verdict}`)}`,
                 entity: row.evalId,
                 result: "",
-                duration: "",
-                cost: "",
+                duration: localeText(locale, "entityList.average", { value: cellText(row.duration) }),
+                cost: localeText(locale, "entityList.average", { value: cellText(row.cost) }),
             },
         };
         const attempts = row.attempts.map((attempt, index) => ({
@@ -458,7 +465,7 @@ function experimentDetailTable(item, ctx) {
             cells: {
                 status: `  ${verdictMark(attempt.verdict)}`,
                 entity: `${index === row.attempts.length - 1 ? "└─" : "├─"} ${attempt.locator}`,
-                result: attemptItemReason(attempt) ?? MISSING_MARK,
+                result: attemptItemReason(attempt, resultBudget) ?? MISSING_MARK,
                 duration: attempt.verdict === "skipped" && attempt.durationMs === 0 ? null : formatDurationMs(attempt.durationMs),
                 cost: attempt.costUSD === undefined ? null : formatUSD(attempt.costUSD),
             },
@@ -468,16 +475,22 @@ function experimentDetailTable(item, ctx) {
     const flags = item.flags && Object.keys(item.flags).length > 0
         ? `${localeText(locale, "experimentList.flags")} ${Object.entries(item.flags).map(([key, value]) => `${key}=${String(value)}`).join(" · ")}`
         : undefined;
-    return [item.experimentId, flags, renderTableText({ columns, rows, locale }, ctx)].filter(Boolean).join("\n");
+    return [experimentDisplayName(item.experimentId, relativeTo), flags, renderTableText({ columns, rows, locale }, ctx)]
+        .filter(Boolean)
+        .join("\n");
 }
-export function experimentListText(items, ctx) {
+export function experimentListText(items, ctx, relativeTo) {
     if (items.length === 0)
         return localeText(ctx.locale, "attemptList.empty");
-    return [experimentSummaryTable(items, ctx), ...items.map((item) => experimentDetailTable(item, ctx))].join("\n\n");
+    return [
+        experimentSummaryTable(items, ctx, relativeTo),
+        ...items.map((item) => experimentDetailTable(item, ctx, relativeTo)),
+    ].join("\n\n");
 }
 // ── EvalList ──
-function evalListAttemptLine(item) {
-    const reason = attemptItemReason(item);
+function evalListAttemptLine(item, ctx) {
+    // 行式列表同守「Result 最多两行」:预算 = 两行终端宽,超出由 fit 投影按优先级让位。
+    const reason = attemptItemReason(item, Math.max(24, ctx.width * 2 - stringWidth(locatorBadge(item)) - 6));
     return `  ${locatorBadge(item)}${reason ? ` · ${reason}` : ""}`;
 }
 export function evalListText(items, ctx) {
@@ -489,18 +502,21 @@ export function evalListText(items, ctx) {
         const summary = [
             localeText(locale, "attemptList.score", { score: cellText(item.score) }),
             localeText(locale, "overview.attemptsCount", { n: item.attempts.length }),
-            `${formatDurationMs(item.duration.value ?? 0)} avg`,
-            item.cost.value === null ? `${missingText(locale)} avg` : `${formatUSD(item.cost.value)} avg`,
+            localeText(locale, "entityList.average", {
+                value: item.duration.value === null ? missingText(locale) : formatDurationMs(item.duration.value),
+            }),
+            localeText(locale, "entityList.average", {
+                value: item.cost.value === null ? missingText(locale) : formatUSD(item.cost.value),
+            }),
         ].join(" · ");
-        const attemptLines = item.attempts.map(evalListAttemptLine);
+        const attemptLines = item.attempts.map((attempt) => evalListAttemptLine(attempt, ctx));
         return [identity, `  ${summary}`, ...attemptLines].join("\n");
     });
     return blocks.join("\n\n");
 }
 // ── AttemptList ──
-/** 一个 AttemptListItem 的完整 text 卡片:判定符 + locator + 身份 + 耗时/成本,
- * 然后逐条断言(gate 与 soft 都列,与 web 面的 AttemptRow 同一份材料)。 */
-function attemptListItemText(item, ctx, locale) {
+/** Attempt 比较卡片：只显示一条主失败摘要（至多两行终端宽）；完整 assertions 走 locator 下钻。 */
+function attemptListItemText(item, ctx) {
     const head = [
         `${verdictMark(item.verdict)} ${item.locator}`,
         item.evalId,
@@ -509,31 +525,16 @@ function attemptListItemText(item, ctx, locale) {
         ...(item.costUSD !== undefined ? [formatUSD(item.costUSD)] : []),
     ].join(" · ");
     const lines = [head];
-    if (item.error) {
-        lines.push(indentBlock(wrapDisplay(item.error, ctx.width - 4).join("\n"), "    "));
-    }
-    for (const assertion of item.assertions) {
-        const scoreText = assertion.threshold !== undefined
-            ? `${formatPlainNumber(assertion.score)}/${formatPlainNumber(assertion.threshold)}`
-            : formatPlainNumber(assertion.score);
-        lines.push(`  ${assertion.severity} ${assertion.name} · ${localeText(locale, `verdict.${assertion.passed ? "passed" : "failed"}`)}${assertion.severity === "soft" ? ` ${scoreText}` : ""}`);
-        if (assertion.detail)
-            lines.push(indentBlock(wrapDisplay(assertion.detail, ctx.width - 4).join("\n"), "    "));
-        if (assertion.evidence) {
-            const limit = Math.max(240, ctx.width * 5);
-            const evidence = assertion.evidence.length <= limit
-                ? assertion.evidence
-                : `${assertion.evidence.slice(0, limit)}… (${assertion.evidence.length - limit} more chars; open ${item.locator} for full evidence)`;
-            lines.push(indentBlock(wrapDisplay(evidence, ctx.width - 6).join("\n"), "      "));
-        }
-    }
+    const reason = attemptItemReason(item, Math.max(24, ctx.width * 2 - 4));
+    if (reason)
+        lines.push(`  ${reason}`);
     return lines.join("\n");
 }
 export function attemptListText(items, total, ctx) {
     const locale = ctx.locale;
     if (items.length === 0)
         return localeText(locale, "attemptList.empty");
-    const blocks = items.map((item) => attemptListItemText(item, ctx, locale));
+    const blocks = items.map((item) => attemptListItemText(item, ctx));
     const remaining = (total ?? items.length) - items.length;
     if (remaining > 0)
         blocks.push(localeText(locale, "attemptList.truncatedText", { n: remaining }));

package/dist/report/text/table.js

@@ -6,7 +6,7 @@
 // 列宽按显示宽度算(CJK 记 2 列);null 渲染 —,不补 0;超宽先折最宽的左对齐列,
 // 压到下限仍放不下就从右侧丢列并如实报数(「截断报剩余」是既有契约,不在这里破例)。
 import { countText, localeText } from "../locale.js";
-import { renderAlignedRows, stringWidth, wrapDisplay } from "./layout.js";
+import { charDisplayWidth, renderAlignedRows, stringWidth, wrapDisplay } from "./layout.js";
 const MISSING_MARK = "—";
 /** 列间距,与 renderAlignedRows 的 join("   ") 一致。 */
 const COLUMN_GAP = 3;
@@ -47,12 +47,36 @@ function fitWidths(natural, align, available) {
     }
     return { widths, hidden };
 }
-/** 逻辑行 → 物理行:每格折到自己的列宽,行高取最高的那格,矮格补空串。 */
-function toPhysicalRows(cells, widths) {
+/** 按显示宽度截到 width(不够放 `…` 时原样返回)。 */
+function truncateDisplay(text, width) {
+    if (stringWidth(text) <= width)
+        return text;
+    let out = "";
+    let used = 0;
+    for (const ch of text) {
+        const w = charDisplayWidth(ch.codePointAt(0));
+        if (used + w > width)
+            break;
+        out += ch;
+        used += w;
+    }
+    return out;
+}
+/** 列的 maxLines 收口:超出的行丢弃,最后一行以 `…` 如实标注被收口。 */
+function clampCellLines(lines, maxLines, width) {
+    if (maxLines === undefined || lines.length <= maxLines)
+        return lines;
+    const kept = lines.slice(0, Math.max(1, maxLines));
+    const last = kept[kept.length - 1];
+    kept[kept.length - 1] = `${truncateDisplay(last, Math.max(1, width - 1))}…`;
+    return kept;
+}
+/** 逻辑行 → 物理行:每格折到自己的列宽,列带 maxLines 就收口,行高取最高的那格,矮格补空串。 */
+function toPhysicalRows(cells, widths, maxLines) {
     const wrapped = cells.map((cell, c) => 
     // 放得下就原样保留:wrapDisplay 会把连续空格折成一个,而 DeltaTable 的
     // "50% → 62%   +12pp" 这类格子内含固定间距 —— 不折行时一个字节都不能动。
-    !cell.includes("\n") && stringWidth(cell) <= widths[c] ? [cell] : wrapDisplay(cell, widths[c]));
+    clampCellLines(!cell.includes("\n") && stringWidth(cell) <= widths[c] ? [cell] : wrapDisplay(cell, widths[c]), maxLines[c], widths[c]));
     const height = Math.max(...wrapped.map((lines) => lines.length), 1);
     const out = [];
     for (let i = 0; i < height; i++)
@@ -80,7 +104,14 @@ export function renderTableText(props, ctx) {
     const matrix = [header, ...body];
     const natural = header.map((_, c) => Math.max(...matrix.map((row) => stringWidth(row[c] ?? ""))));
     const { widths, hidden } = fitWidths(natural, align, ctx.width);
-    const physical = matrix.flatMap((row) => toPhysicalRows(row.slice(0, widths.length), widths));
+    const maxLines = props.columns.map((column) => column.maxLines);
+    if (hasLocator)
+        maxLines.push(undefined);
+    // 表头不参与 maxLines 收口:表头是自己写的短词,收口只服务数据格。
+    const physical = [
+        ...toPhysicalRows(header.slice(0, widths.length), widths, widths.map(() => undefined)),
+        ...body.flatMap((row) => toPhysicalRows(row.slice(0, widths.length), widths, maxLines)),
+    ];
     const table = renderAlignedRows(physical, align);
     return hidden > 0 ? `${table}\n${countText(locale, "table.columnsHidden", hidden)}` : table;
 }