niceeval 0.6.2 → 0.7.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (218) hide show
  1. package/INDEX.md +23 -23
  2. package/README.zh.md +6 -6
  3. package/dist/agents/types.d.ts +2 -2
  4. package/dist/i18n/zh-CN.d.ts +3 -3
  5. package/dist/report/aggregate.d.ts +32 -26
  6. package/dist/report/aggregate.js +157 -76
  7. package/dist/report/built-in/index.d.ts +2 -0
  8. package/dist/report/built-in/index.js +8 -0
  9. package/dist/report/components.d.ts +91 -164
  10. package/dist/report/components.js +377 -114
  11. package/dist/report/compute.d.ts +86 -73
  12. package/dist/report/compute.js +592 -432
  13. package/dist/report/flag.d.ts +28 -17
  14. package/dist/report/flag.js +86 -16
  15. package/dist/report/format.d.ts +11 -11
  16. package/dist/report/format.js +17 -15
  17. package/dist/report/index.d.ts +16 -17
  18. package/dist/report/index.js +20 -22
  19. package/dist/report/load.js +3 -2
  20. package/dist/report/locale.d.ts +49 -34
  21. package/dist/report/locale.js +106 -58
  22. package/dist/report/metrics.d.ts +10 -3
  23. package/dist/report/metrics.js +46 -12
  24. package/dist/report/primitives.d.ts +42 -15
  25. package/dist/report/primitives.js +135 -26
  26. package/dist/report/react/AttemptList.d.ts +10 -8
  27. package/dist/report/react/AttemptList.js +18 -10
  28. package/dist/report/react/DeltaTable.js +19 -18
  29. package/dist/report/react/EvalList.d.ts +3 -3
  30. package/dist/report/react/EvalList.js +0 -0
  31. package/dist/report/react/ExperimentComparison.d.ts +4 -2
  32. package/dist/report/react/ExperimentComparison.js +5 -4
  33. package/dist/report/react/ExperimentList.d.ts +3 -3
  34. package/dist/report/react/ExperimentList.js +16 -15
  35. package/dist/report/react/MetricBars.js +5 -4
  36. package/dist/report/react/MetricLine.js +12 -5
  37. package/dist/report/react/MetricMatrix.js +1 -1
  38. package/dist/report/react/MetricScatter.js +54 -17
  39. package/dist/report/react/MetricTable.js +2 -12
  40. package/dist/report/react/ScopeSummary.d.ts +10 -0
  41. package/dist/report/react/ScopeSummary.js +17 -0
  42. package/dist/report/react/Scoreboard.js +6 -6
  43. package/dist/report/react/cell.js +2 -2
  44. package/dist/report/react/fixtures.d.ts +5 -9
  45. package/dist/report/react/fixtures.js +105 -149
  46. package/dist/report/react/index.d.ts +15 -5
  47. package/dist/report/react/index.js +18 -7
  48. package/dist/report/report.d.ts +137 -20
  49. package/dist/report/report.js +261 -34
  50. package/dist/report/text/faces.d.ts +17 -19
  51. package/dist/report/text/faces.js +225 -157
  52. package/dist/report/text/plot.js +1 -1
  53. package/dist/report/text/table.js +2 -2
  54. package/dist/report/tree.d.ts +90 -40
  55. package/dist/report/tree.js +252 -94
  56. package/dist/report/types.d.ts +245 -300
  57. package/dist/report/types.js +4 -3
  58. package/dist/report/web.d.ts +21 -5
  59. package/dist/report/web.js +42 -16
  60. package/dist/results/select.d.ts +38 -16
  61. package/dist/results/select.js +73 -25
  62. package/dist/results/types.d.ts +38 -14
  63. package/dist/shared/aggregate.d.ts +3 -2
  64. package/dist/shared/aggregate.js +5 -4
  65. package/docs-site/zh/README.md +44 -0
  66. package/docs-site/zh/examples/ai-agent-application.mdx +63 -0
  67. package/docs-site/zh/examples/coding-agent-extensions.mdx +57 -0
  68. package/docs-site/zh/examples/index.mdx +50 -0
  69. package/docs-site/zh/{concepts → explanation}/adapter.mdx +11 -11
  70. package/docs-site/zh/{concepts → explanation}/assert.mdx +7 -7
  71. package/docs-site/zh/{concepts → explanation}/drive.mdx +8 -8
  72. package/docs-site/zh/{concepts → explanation}/evals.mdx +4 -4
  73. package/docs-site/zh/{concepts → explanation}/experiment.mdx +8 -8
  74. package/docs-site/zh/{concepts → explanation}/hitl.mdx +8 -8
  75. package/docs-site/zh/{concepts → explanation}/judge.mdx +5 -5
  76. package/docs-site/zh/{concepts → explanation}/overview.mdx +5 -5
  77. package/docs-site/zh/{guides → explanation}/runner.mdx +1 -1
  78. package/docs-site/zh/{concepts → explanation}/tier.mdx +6 -6
  79. package/docs-site/zh/{guides → how-to}/agent-feedback-loop.mdx +7 -7
  80. package/docs-site/zh/{guides → how-to}/authoring.mdx +2 -2
  81. package/docs-site/zh/{guides → how-to}/connect-otel.mdx +6 -6
  82. package/docs-site/zh/{guides → how-to}/connect-your-agent.mdx +18 -18
  83. package/docs-site/zh/{guides → how-to}/custom-reports.mdx +6 -6
  84. package/docs-site/zh/{guides → how-to}/experiments.mdx +3 -3
  85. package/docs-site/zh/{guides → how-to}/publish-report.mdx +2 -2
  86. package/docs-site/zh/{guides → how-to}/sandbox-agent.mdx +2 -2
  87. package/docs-site/zh/{guides → how-to}/sandbox-providers.mdx +1 -1
  88. package/docs-site/zh/{guides → how-to}/viewing-results.mdx +6 -6
  89. package/docs-site/zh/{guides → how-to}/write-experiment.mdx +3 -3
  90. package/docs-site/zh/{guides → how-to}/write-send.mdx +13 -13
  91. package/docs-site/zh/index.mdx +23 -25
  92. package/docs-site/zh/introduction.mdx +8 -8
  93. package/docs-site/zh/reference/builtin-agents.mdx +5 -5
  94. package/docs-site/zh/reference/capabilities.mdx +6 -6
  95. package/docs-site/zh/reference/cli.mdx +9 -7
  96. package/docs-site/zh/reference/define-agent.mdx +1 -1
  97. package/docs-site/zh/reference/events.mdx +3 -3
  98. package/docs-site/zh/{guides → reference}/official-adapters.mdx +7 -7
  99. package/docs-site/zh/{guides → reference}/report-components.mdx +5 -5
  100. package/docs-site/zh/{guides → reference}/results-data.mdx +5 -5
  101. package/docs-site/zh/{guides → troubleshooting}/debug-sandbox.mdx +2 -2
  102. package/docs-site/zh/{guides → troubleshooting}/debugging.mdx +4 -2
  103. package/docs-site/zh/{quickstart.mdx → tutorials/quickstart.mdx} +5 -17
  104. package/package.json +6 -2
  105. package/src/agents/index.ts +2 -2
  106. package/src/agents/openai-compat.ts +1 -1
  107. package/src/agents/streaming.ts +2 -2
  108. package/src/agents/types.ts +3 -3
  109. package/src/cli.ts +42 -23
  110. package/src/context/context.ts +1 -1
  111. package/src/context/session.test.ts +1 -1
  112. package/src/context/session.ts +1 -1
  113. package/src/i18n/en.ts +18 -16
  114. package/src/i18n/zh-CN.ts +16 -15
  115. package/src/report/aggregate.ts +175 -87
  116. package/src/report/built-in/index.tsx +9 -0
  117. package/src/report/components.tsx +625 -285
  118. package/src/report/compute.ts +717 -515
  119. package/src/report/dual-render.test.tsx +738 -1148
  120. package/src/report/flag.ts +97 -33
  121. package/src/report/format.ts +18 -22
  122. package/src/report/index.ts +113 -58
  123. package/src/report/load.ts +3 -2
  124. package/src/report/locale.ts +120 -69
  125. package/src/report/metrics.ts +42 -12
  126. package/src/report/primitives.tsx +190 -45
  127. package/src/report/react/AttemptList.tsx +32 -20
  128. package/src/report/react/DeltaTable.tsx +63 -45
  129. package/src/report/react/EvalList.tsx +0 -0
  130. package/src/report/react/ExperimentComparison.tsx +12 -7
  131. package/src/report/react/ExperimentList.tsx +38 -26
  132. package/src/report/react/MetricBars.tsx +5 -4
  133. package/src/report/react/MetricLine.tsx +13 -8
  134. package/src/report/react/MetricMatrix.tsx +2 -2
  135. package/src/report/react/MetricScatter.tsx +74 -20
  136. package/src/report/react/MetricTable.tsx +4 -76
  137. package/src/report/react/ScopeSummary.tsx +86 -0
  138. package/src/report/react/Scoreboard.tsx +28 -10
  139. package/src/report/react/cell.tsx +2 -2
  140. package/src/report/react/enhance.js +57 -5
  141. package/src/report/react/fixtures.ts +109 -156
  142. package/src/report/react/index.tsx +24 -39
  143. package/src/report/react/render.test.tsx +139 -104
  144. package/src/report/react/styles.css +181 -91
  145. package/src/report/report.test.ts +761 -1031
  146. package/src/report/report.ts +425 -47
  147. package/src/report/text/faces.ts +257 -164
  148. package/src/report/text/plot.ts +1 -1
  149. package/src/report/text/table.ts +2 -2
  150. package/src/report/tree.ts +362 -104
  151. package/src/report/types.ts +257 -287
  152. package/src/report/web.ts +63 -20
  153. package/src/results/attempt-evidence.test.ts +4 -4
  154. package/src/results/attempt-evidence.ts +5 -5
  155. package/src/results/copy.ts +6 -6
  156. package/src/results/host-equivalence.test.ts +26 -14
  157. package/src/results/index.ts +10 -4
  158. package/src/results/open.ts +8 -4
  159. package/src/results/results.test.ts +4 -3
  160. package/src/results/select.ts +104 -34
  161. package/src/results/types.ts +36 -14
  162. package/src/runner/feedback/human.test.ts +1 -1
  163. package/src/runner/run.ts +1 -1
  164. package/src/sandbox/cli-commands.ts +2 -2
  165. package/src/scoring/judge.test.ts +1 -1
  166. package/src/shared/aggregate.ts +5 -4
  167. package/src/show/compose.ts +50 -67
  168. package/src/show/index.ts +107 -56
  169. package/src/show/render.ts +43 -27
  170. package/src/show/report-host.test.ts +188 -0
  171. package/src/show/report-host.ts +375 -0
  172. package/src/show/show.test.ts +86 -36
  173. package/src/view/app/App.test.tsx +69 -0
  174. package/src/view/app/App.tsx +144 -48
  175. package/src/view/app/components/AttemptModal.tsx +324 -63
  176. package/src/view/app/components/CodeView.tsx +10 -4
  177. package/src/view/app/i18n.ts +31 -17
  178. package/src/view/app/lib/artifact-url.ts +14 -3
  179. package/src/view/app/main.tsx +13 -8
  180. package/src/view/app/pages/{RunsPage.tsx → AttemptsPage.tsx} +6 -6
  181. package/src/view/app/types.ts +4 -1
  182. package/src/view/artifact-serving.test.ts +21 -1
  183. package/src/view/client-dist/app.css +1 -1
  184. package/src/view/client-dist/app.js +6 -6
  185. package/src/view/data.test.ts +9 -3
  186. package/src/view/data.ts +145 -49
  187. package/src/view/index.ts +48 -44
  188. package/src/view/server.ts +35 -15
  189. package/src/view/shared/types.ts +34 -5
  190. package/src/view/styles.css +224 -0
  191. package/src/view/view-report.test.ts +161 -57
  192. package/dist/report/built-ins/experiment-comparison.d.ts +0 -39
  193. package/dist/report/built-ins/experiment-comparison.js +0 -119
  194. package/dist/report/built-ins/index.d.ts +0 -2
  195. package/dist/report/built-ins/index.js +0 -2
  196. package/dist/report/react/GroupSummary.d.ts +0 -8
  197. package/dist/report/react/GroupSummary.js +0 -8
  198. package/dist/report/react/RunOverview.d.ts +0 -8
  199. package/dist/report/react/RunOverview.js +0 -12
  200. package/docs-site/zh/example/ai-agent-application.mdx +0 -152
  201. package/docs-site/zh/example/claude-code-codex-plugin.mdx +0 -167
  202. package/docs-site/zh/example/claude-code-codex-skill.mdx +0 -152
  203. package/docs-site/zh/example/showcase.mdx +0 -39
  204. package/src/report/built-in-user-parity.test.tsx +0 -597
  205. package/src/report/built-ins/experiment-comparison.tsx +0 -179
  206. package/src/report/built-ins/index.ts +0 -7
  207. package/src/report/react/GroupSummary.tsx +0 -66
  208. package/src/report/react/RunOverview.tsx +0 -109
  209. /package/docs-site/zh/{example/tier1-ai-sdk-v7.mdx → examples/integrations/ai-sdk-v7.mdx} +0 -0
  210. /package/docs-site/zh/{example/tier1-claude-sdk.mdx → examples/integrations/claude-sdk.mdx} +0 -0
  211. /package/docs-site/zh/{example/tier1-codex-sdk.mdx → examples/integrations/codex-sdk.mdx} +0 -0
  212. /package/docs-site/zh/{example/tier1-langgraph.mdx → examples/integrations/langgraph.mdx} +0 -0
  213. /package/docs-site/zh/{example/tier1-pi-sdk.mdx → examples/integrations/pi-sdk.mdx} +0 -0
  214. /package/docs-site/zh/{guides → how-to}/ci-integration.mdx +0 -0
  215. /package/docs-site/zh/{guides → how-to}/dataset-fanout.mdx +0 -0
  216. /package/docs-site/zh/{guides → how-to}/fixtures.mdx +0 -0
  217. /package/docs-site/zh/{guides → how-to}/reporters.mdx +0 -0
  218. /package/docs-site/zh/{guides → how-to}/scoring-guide.mdx +0 -0
@@ -1,6 +1,6 @@
1
1
  import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
2
2
  import { formatMetricValue } from "../format.js";
3
- import { DEFAULT_REPORT_LOCALE, countText, localeText, resolveMetricLabel } from "../locale.js";
3
+ import { DEFAULT_REPORT_LOCALE, countText, localeText, resolveLocalizedText, resolveMetricLabel } from "../locale.js";
4
4
  import { niceTicks, placePointLabels } from "./chart-math.js";
5
5
  import { colorClassForKey, seriesClassForKey } from "./colors.js";
6
6
  import { cx } from "./format.js";
@@ -9,22 +9,55 @@ const HEIGHT = 360;
9
9
  const MARGIN = { top: 26, right: 24, bottom: 46, left: 62 };
10
10
  const PLOT_W = WIDTH - MARGIN.left - MARGIN.right;
11
11
  const PLOT_H = HEIGHT - MARGIN.top - MARGIN.bottom;
12
- /** 点的直接标签:experiment id 的末段(完整 id 在 <title> 里)。 */
13
- function pointLabel(key) {
14
- return key.split("/").filter(Boolean).at(-1) ?? key;
12
+ /**
13
+ * 点的直接标签:末段在当前 data 中唯一才缩成末段;重名时逐步加长为能区分它们的最短
14
+ * 路径后缀(完整 id 与两轴值仍进 <title>)。
15
+ */
16
+ function pointLabels(keys) {
17
+ const segsOf = (key) => key.split("/").filter(Boolean);
18
+ const depth = new Map(keys.map((key) => [key, 1]));
19
+ for (;;) {
20
+ const byLabel = new Map();
21
+ for (const key of keys) {
22
+ const segs = segsOf(key);
23
+ const label = segs.slice(-Math.min(depth.get(key), segs.length)).join("/") || key;
24
+ byLabel.set(label, [...(byLabel.get(label) ?? []), key]);
25
+ }
26
+ let grew = false;
27
+ for (const group of byLabel.values()) {
28
+ if (group.length < 2)
29
+ continue;
30
+ for (const key of group) {
31
+ const segs = segsOf(key);
32
+ if (depth.get(key) < segs.length) {
33
+ depth.set(key, depth.get(key) + 1);
34
+ grew = true;
35
+ }
36
+ }
37
+ }
38
+ if (!grew) {
39
+ const out = new Map();
40
+ for (const key of keys) {
41
+ const segs = segsOf(key);
42
+ out.set(key, segs.slice(-Math.min(depth.get(key), segs.length)).join("/") || key);
43
+ }
44
+ return out;
45
+ }
46
+ }
15
47
  }
16
48
  /**
17
- * 一根轴:niceTicks 撑出整齐的值域,值 → 像素做线性映射;
18
- * better: "lower" 时反向,好的一端固定在右 / 上。
49
+ * 一根轴:niceTicks 撑出整齐的值域,值 → 像素做线性映射。轴方向跟随指标的 `better`:
50
+ * `better: "lower"` 的轴反向渲染(值大的一端在左 / 下),「更好」因此恒指向右与上;
51
+ * 刻度标签始终显示真实值,反向只改方向不改数字(docs/feature/reports/library/metric-views.md)。
19
52
  */
20
- function axisScale(values, better, pixelLo, pixelHi) {
53
+ function axisScale(values, pixelLo, pixelHi, invert) {
21
54
  const ticks = niceTicks(Math.min(...values), Math.max(...values), 5);
22
55
  const lo = ticks[0];
23
56
  const hi = ticks[ticks.length - 1];
24
57
  const scale = (v) => {
25
58
  let t = (v - lo) / (hi - lo || 1);
26
- if (better === "lower")
27
- t = 1 - t; // 反向:值越低越靠「好」的一端
59
+ if (invert)
60
+ t = 1 - t;
28
61
  return pixelLo + t * (pixelHi - pixelLo);
29
62
  };
30
63
  return { ticks, scale };
@@ -41,21 +74,25 @@ export function MetricScatter({ data, pointHref, className, locale = DEFAULT_REP
41
74
  return (_jsxs("figure", { className: cx("nre", "nre-metric-scatter", className), children: [_jsx("p", { className: "nre-scatter-empty nre-missing", children: localeText(locale, "scatter.noData", { x: xLabel, y: yLabel }) }), missingNote] }));
42
75
  }
43
76
  const axisLabel = (label, col) => `${label}${col.unit ? `(${col.unit})` : ""}`;
44
- const xScale = axisScale(drawableRows.map((r) => r.x.value), data.x.better, MARGIN.left, MARGIN.left + PLOT_W);
45
- // y 像素轴向下增长:better:"higher" 高值在上 映射到 [bottom, top];"lower" 由 axisScale 反向后同样落到上方
46
- const yScale = axisScale(drawableRows.map((r) => r.y.value), data.y.better, MARGIN.top + PLOT_H, MARGIN.top);
77
+ // 轴方向跟随 better:lower 反向(值大在左 / ),higher 与未声明正向。
78
+ const xScale = axisScale(drawableRows.map((r) => r.x.value), MARGIN.left, MARGIN.left + PLOT_W, data.x.better === "lower");
79
+ // y 像素轴向下增长:正向 = 高值在上 映射到 [bottom, top];lower 反向 = 高值在下。
80
+ const yScale = axisScale(drawableRows.map((r) => r.y.value), MARGIN.top + PLOT_H, MARGIN.top, data.y.better === "lower");
81
+ const labelByKey = pointLabels(drawableRows.map((r) => r.key));
82
+ // 方向提示恒为「越靠右上越好」,仅当两轴都声明 better 时显示——任一轴未声明,
83
+ // 组件不猜「更好」朝哪边,整图无提示。
84
+ const showBetterHint = data.x.better !== undefined && data.y.better !== undefined;
47
85
  const points = drawableRows.map((r) => {
48
86
  const xValue = r.x.value;
49
87
  const yValue = r.y.value;
50
88
  return {
51
89
  key: r.key,
52
90
  series: r.series,
53
- label: pointLabel(r.key),
91
+ label: labelByKey.get(r.key) ?? r.key,
54
92
  xValue,
55
93
  yValue,
56
- // hover 内容:experiment(点键)+ 系列(series,ExperimentComparison 传的是 agent 维度,
57
- // 有则加一行;无系列的散点没有这行)+ 两轴 display samples/total(docs/feature/reports/library.md「MetricScatter」行为清单)
58
- title: `${r.key}${r.series !== undefined ? `\n${r.series}` : ""}\n${xLabel}: ${r.x.display}(${r.x.samples}/${r.x.total})\n${yLabel}: ${r.y.display}(${r.y.samples}/${r.y.total})`,
94
+ // hover 内容:experiment(点键)+ 系列(有则加一行)+ 两轴 display 与 samples/total
95
+ title: `${r.key}${r.series !== undefined ? `\n${r.series}` : ""}\n${xLabel}: ${resolveLocalizedText(r.x.display, locale)}(${r.x.samples}/${r.x.total})\n${yLabel}: ${resolveLocalizedText(r.y.display, locale)}(${r.y.samples}/${r.y.total})`,
59
96
  px: xScale.scale(xValue),
60
97
  py: yScale.scale(yValue),
61
98
  };
@@ -76,7 +113,7 @@ export function MetricScatter({ data, pointHref, className, locale = DEFAULT_REP
76
113
  list.sort((a, b) => a.xValue - b.xValue);
77
114
  // 直接标签的候选位择优布局:锚向、避让方向都由布局按空间决定,画布边界含边距
78
115
  const labels = placePointLabels(points.map((p) => ({ cx: p.px, cy: p.py, width: p.label.length * 6.4 + 10 })), { x0: 2, y0: 2, x1: WIDTH - 2, y1: HEIGHT - 2 });
79
- return (_jsxs("figure", { className: cx("nre", "nre-metric-scatter", className), children: [_jsxs("svg", { className: "nre-scatter-svg", viewBox: `0 0 ${WIDTH} ${HEIGHT}`, role: "img", "aria-label": `${axisLabel(xLabel, data.x)} × ${axisLabel(yLabel, data.y)}`, children: [_jsxs("g", { className: "nre-scatter-grid", children: [yScale.ticks.map((tick) => (_jsx("line", { x1: MARGIN.left, x2: MARGIN.left + PLOT_W, y1: yScale.scale(tick), y2: yScale.scale(tick) }, `gy${tick}`))), xScale.ticks.map((tick) => (_jsx("line", { y1: MARGIN.top, y2: MARGIN.top + PLOT_H, x1: xScale.scale(tick), x2: xScale.scale(tick) }, `gx${tick}`)))] }), _jsx("text", { className: "nre-scatter-better-hint", x: MARGIN.left + PLOT_W - 6, y: MARGIN.top + 14, textAnchor: "end", children: localeText(locale, "scatter.betterHint") }), _jsx("g", { className: "nre-scatter-axis nre-scatter-axis-y", children: yScale.ticks.map((tick) => (_jsx("text", { className: "nre-scatter-tick", x: MARGIN.left - 8, y: yScale.scale(tick) + 3, textAnchor: "end", children: formatMetricValue(tick, data.y.unit) }, `ay${tick}`))) }), _jsx("g", { className: "nre-scatter-axis nre-scatter-axis-x", children: xScale.ticks.map((tick) => (_jsx("text", { className: "nre-scatter-tick", x: xScale.scale(tick), y: MARGIN.top + PLOT_H + 16, textAnchor: "middle", children: formatMetricValue(tick, data.x.unit) }, `ax${tick}`))) }), _jsx("text", { className: "nre-scatter-xlabel", x: MARGIN.left + PLOT_W / 2, y: HEIGHT - 8, textAnchor: "middle", children: axisLabel(xLabel, data.x) }), _jsx("text", { className: "nre-scatter-ylabel", x: 14, y: MARGIN.top + PLOT_H / 2, textAnchor: "middle", transform: `rotate(-90 14 ${MARGIN.top + PLOT_H / 2})`, children: axisLabel(yLabel, data.y) }), seriesOrder.map((series) => {
116
+ return (_jsxs("figure", { className: cx("nre", "nre-metric-scatter", className), children: [_jsxs("svg", { className: "nre-scatter-svg", viewBox: `0 0 ${WIDTH} ${HEIGHT}`, role: "img", "aria-label": `${axisLabel(xLabel, data.x)} × ${axisLabel(yLabel, data.y)}`, children: [_jsxs("g", { className: "nre-scatter-grid", children: [yScale.ticks.map((tick) => (_jsx("line", { x1: MARGIN.left, x2: MARGIN.left + PLOT_W, y1: yScale.scale(tick), y2: yScale.scale(tick) }, `gy${tick}`))), xScale.ticks.map((tick) => (_jsx("line", { y1: MARGIN.top, y2: MARGIN.top + PLOT_H, x1: xScale.scale(tick), x2: xScale.scale(tick) }, `gx${tick}`)))] }), showBetterHint && (_jsx("text", { className: "nre-scatter-better-hint", x: MARGIN.left + PLOT_W - 6, y: MARGIN.top + 14, textAnchor: "end", children: localeText(locale, "scatter.betterUpperRight") })), _jsx("g", { className: "nre-scatter-axis nre-scatter-axis-y", children: yScale.ticks.map((tick) => (_jsx("text", { className: "nre-scatter-tick", x: MARGIN.left - 8, y: yScale.scale(tick) + 3, textAnchor: "end", children: formatMetricValue(tick, data.y.unit) }, `ay${tick}`))) }), _jsx("g", { className: "nre-scatter-axis nre-scatter-axis-x", children: xScale.ticks.map((tick) => (_jsx("text", { className: "nre-scatter-tick", x: xScale.scale(tick), y: MARGIN.top + PLOT_H + 16, textAnchor: "middle", children: formatMetricValue(tick, data.x.unit) }, `ax${tick}`))) }), _jsx("text", { className: "nre-scatter-xlabel", x: MARGIN.left + PLOT_W / 2, y: HEIGHT - 8, textAnchor: "middle", children: axisLabel(xLabel, data.x) }), _jsx("text", { className: "nre-scatter-ylabel", x: 14, y: MARGIN.top + PLOT_H / 2, textAnchor: "middle", transform: `rotate(-90 14 ${MARGIN.top + PLOT_H / 2})`, children: axisLabel(yLabel, data.y) }), seriesOrder.map((series) => {
80
117
  const list = bySeries.get(series);
81
118
  if (list.length < 2)
82
119
  return null;
@@ -3,21 +3,11 @@ import { DEFAULT_REPORT_LOCALE, localeText, resolveMetricLabel } from "../locale
3
3
  import { MetricCellView } from "./cell.js";
4
4
  import { colorClassForKey } from "./colors.js";
5
5
  import { cx } from "./format.js";
6
- /** verdict 计票 pill(「3 passed / 1 failed」):非零判定各一枚,全零如实空。 */
7
- function VerdictTally({ verdicts, locale, }) {
8
- const kinds = ["passed", "failed", "errored", "skipped"].filter((k) => verdicts[k] > 0);
9
- return (_jsxs("span", { className: "nre-verdict-tally", children: [kinds.map((kind) => (_jsxs("span", { className: cx("nre-verdict-pill", `nre-verdict-${kind}`), children: [verdicts[kind], " ", localeText(locale, `verdict.${kind}`)] }, kind))), kinds.length === 0 && _jsx("span", { className: "nre-missing", children: "\u2014" })] }));
10
- }
11
6
  export function MetricTable({ data, attemptHref, filter, className, locale = DEFAULT_REPORT_LOCALE, }) {
12
- const hasMeta = data.rows.some((row) => row.meta !== undefined);
13
- const hasModel = data.rows.some((row) => row.meta?.model !== undefined);
14
- const hasVerdicts = data.rows.some((row) => row.meta?.verdicts !== undefined);
15
- const table = (_jsxs("table", { className: cx("nre", "nre-metric-table", !filter && className), children: [_jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", "data-nre-sort": "", children: data.dimension }), hasMeta && hasModel && (_jsx("th", { scope: "col", className: "nre-meta-col", "data-nre-sort": "", children: localeText(locale, "table.model") })), hasMeta && (_jsx("th", { scope: "col", className: "nre-meta-col", "data-nre-sort": "", children: localeText(locale, "table.agent") })), data.columns.map((col) => (_jsxs("th", { scope: "col", className: "nre-metric-col", "data-nre-sort": "", children: [resolveMetricLabel(col.label, locale, col.key), col.unit && _jsxs("span", { className: "nre-unit", children: ["(", col.unit, ")"] }), col.better && (_jsx("span", { className: "nre-better", title: localeText(locale, col.better === "higher" ? "table.higherBetter" : "table.lowerBetter"), children: col.better === "higher" ? "↑" : "↓" }))] }, col.key))), hasVerdicts && (_jsx("th", { scope: "col", className: "nre-verdicts-col", "data-nre-sort": "", children: localeText(locale, "table.verdicts") }))] }) }), _jsx("tbody", { children: data.rows.map((row) => (_jsxs("tr", { children: [_jsxs("th", { scope: "row", className: cx("nre-row-key", "nre-key", colorClassForKey(row.key)), "data-sort-value": row.key, children: [row.key, row.meta?.evals !== undefined && (_jsxs("div", { className: "nre-row-meta-sub", children: [localeText(locale, "overview.evalsCount", { n: row.meta.evals }), row.meta.attempts !== undefined && row.meta.attempts > row.meta.evals
16
- ? ` · ${localeText(locale, "overview.attemptsCount", { n: row.meta.attempts })}`
17
- : "", row.meta.lastRunAt ? ` · ${localeText(locale, "latestRun", { run: row.meta.lastRunAt })}` : ""] }))] }), hasMeta && hasModel && (_jsx("td", { className: "nre-td nre-meta-cell", "data-sort-value": row.meta?.model ?? "", children: row.meta?.model ?? _jsx("span", { className: "nre-missing", children: "\u2014" }) })), hasMeta && (_jsx("td", { className: "nre-td nre-meta-cell", "data-sort-value": row.meta?.agent ?? "", children: row.meta?.agent ?? _jsx("span", { className: "nre-missing", children: "\u2014" }) })), data.columns.map((col) => {
7
+ const table = (_jsxs("table", { className: cx("nre", "nre-metric-table", !filter && className), children: [_jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", "data-nre-sort": "", children: data.rowDimension }), data.columns.map((col) => (_jsxs("th", { scope: "col", className: "nre-metric-col", "data-nre-sort": "", children: [resolveMetricLabel(col.label, locale, col.key), col.unit && _jsxs("span", { className: "nre-unit", children: ["(", col.unit, ")"] }), col.better && (_jsx("span", { className: "nre-better", title: localeText(locale, col.better === "higher" ? "table.higherBetter" : "table.lowerBetter"), children: col.better === "higher" ? "↑" : "↓" }))] }, col.key)))] }) }), _jsx("tbody", { children: data.rows.map((row) => (_jsxs("tr", { children: [_jsx("th", { scope: "row", className: cx("nre-row-key", "nre-key", colorClassForKey(row.key)), "data-sort-value": row.key, children: row.key }), data.columns.map((col) => {
18
8
  const cell = row.cells[col.key];
19
9
  return (_jsx("td", { className: "nre-td", "data-sort-value": cell?.value ?? "", children: cell ? (_jsx(MetricCellView, { cell: cell, attemptHref: attemptHref, locale: locale })) : (_jsx("span", { className: "nre-empty" })) }, col.key));
20
- }), hasVerdicts && (_jsx("td", { className: "nre-td nre-verdicts-cell", "data-sort-value": row.meta?.verdicts?.passed ?? "", children: row.meta?.verdicts ? (_jsx(VerdictTally, { verdicts: row.meta.verdicts, locale: locale })) : (_jsx("span", { className: "nre-missing", children: "\u2014" })) }))] }, row.key))) })] }));
10
+ })] }, row.key))) })] }));
21
11
  if (!filter)
22
12
  return table;
23
13
  // 过滤输入框渲染在表格前(同一个 wrap 里),enhance.js 经 data-nre-filter 接管;
@@ -0,0 +1,10 @@
1
+ import type { ReactElement } from "react";
2
+ import type { ScopeSummaryData } from "../types.ts";
3
+ import { type ReportLocale } from "../locale.ts";
4
+ export declare function ScopeSummary({ data, votes, className, locale, }: {
5
+ data: ScopeSummaryData;
6
+ /** 显示哪一级计票;默认 "eval"。data 恒携带两级,votes 只选择呈现。 */
7
+ votes?: "eval" | "attempt";
8
+ className?: string;
9
+ locale?: ReportLocale;
10
+ }): ReactElement;
@@ -0,0 +1,17 @@
1
+ import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
2
+ import { DEFAULT_REPORT_LOCALE, localeText } from "../locale.js";
3
+ import { MetricCellView } from "./cell.js";
4
+ import { cx } from "./format.js";
5
+ function VerdictTallyView({ tally, locale }) {
6
+ const kinds = ["passed", "failed", "errored", "skipped"].filter((k) => tally[k] > 0);
7
+ return (_jsxs("span", { className: "nre-verdict-tally", children: [kinds.map((kind) => (_jsxs("span", { className: cx("nre-verdict-pill", `nre-verdict-${kind}`), children: [tally[kind], " ", localeText(locale, `verdict.${kind}`)] }, kind))), kinds.length === 0 && _jsx("span", { className: "nre-missing", children: "\u2014" })] }));
8
+ }
9
+ export function ScopeSummary({ data, votes = "eval", className, locale = DEFAULT_REPORT_LOCALE, }) {
10
+ const tally = votes === "attempt" ? data.attemptVerdicts : data.evalVerdicts;
11
+ return (_jsxs("div", { className: cx("nre", "nre-scope-summary", className), "data-votes": votes, children: [_jsxs("dl", { className: "nre-scope-kpis", children: [_jsxs("div", { className: "nre-scope-kpi nre-scope-kpi-rate", children: [_jsx("dt", { children: localeText(locale, "scopeSummary.passRate") }), _jsx("dd", { children: _jsx(MetricCellView, { cell: data.endToEndPassRate, locale: locale }) })] }), _jsxs("div", { className: "nre-scope-kpi", children: [_jsx("dt", { children: localeText(locale, "scopeSummary.experiments") }), _jsx("dd", { children: data.experiments })] }), _jsxs("div", { className: "nre-scope-kpi", children: [_jsx("dt", { children: localeText(locale, "scopeSummary.evals") }), _jsx("dd", { children: data.evals })] }), _jsxs("div", { className: "nre-scope-kpi", children: [_jsx("dt", { children: localeText(locale, "scopeSummary.attempts") }), _jsx("dd", { children: data.attempts })] }), _jsxs("div", { className: "nre-scope-kpi nre-scope-kpi-verdicts", children: [_jsx("dt", { children: localeText(locale, votes === "attempt" ? "scopeSummary.votesAttempt" : "scopeSummary.votesEval") }), _jsx("dd", { children: _jsx(VerdictTallyView, { tally: tally, locale: locale }) })] }), _jsxs("div", { className: "nre-scope-kpi", children: [_jsx("dt", { children: localeText(locale, "scopeSummary.totalCost") }), _jsx("dd", { children: _jsx(MetricCellView, { cell: data.totalCostUSD, locale: locale }) })] })] }), data.range.latestStartedAt !== null && (_jsx("p", { className: "nre-scope-summary-range", children: data.range.earliestStartedAt !== null && data.range.earliestStartedAt !== data.range.latestStartedAt
12
+ ? localeText(locale, "scopeSummary.range", {
13
+ from: data.range.earliestStartedAt,
14
+ to: data.range.latestStartedAt,
15
+ })
16
+ : data.range.latestStartedAt }))] }));
17
+ }
@@ -1,5 +1,5 @@
1
1
  import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
2
- import { DEFAULT_REPORT_LOCALE, countText, localeText } from "../locale.js";
2
+ import { DEFAULT_REPORT_LOCALE, countText, localeText, resolveLocalizedText } from "../locale.js";
3
3
  import { colorClassForKey } from "./colors.js";
4
4
  import { cx } from "./format.js";
5
5
  export function Scoreboard({ data, className, locale = DEFAULT_REPORT_LOCALE, }) {
@@ -11,16 +11,16 @@ export function Scoreboard({ data, className, locale = DEFAULT_REPORT_LOCALE, })
11
11
  subjectKeys.push(subject.key);
12
12
  }
13
13
  }
14
- return (_jsxs("section", { className: cx("nre", "nre-scoreboard", className), children: [_jsxs("table", { className: "nre-scoreboard-table", children: [_jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", children: data.dimension }), _jsxs("th", { scope: "col", className: "nre-total-col", children: [localeText(locale, "scoreboard.total"), _jsxs("span", { className: "nre-full-marks", children: ["/ ", data.fullMarks] })] }), subjectKeys.map((key) => (_jsx("th", { scope: "col", className: "nre-subject-col", children: key }, key)))] }) }), _jsx("tbody", { children: data.rows.map((row) => (_jsxs("tr", { children: [_jsx("th", { scope: "row", className: cx("nre-row-key", "nre-key", colorClassForKey(row.key)), children: row.key }), _jsx("td", { className: "nre-total", children: row.total.display }), subjectKeys.map((key) => {
14
+ return (_jsxs("section", { className: cx("nre", "nre-scoreboard", className), children: [_jsxs("table", { className: "nre-scoreboard-table", children: [_jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", children: data.rowDimension }), _jsxs("th", { scope: "col", className: "nre-total-col", children: [localeText(locale, "scoreboard.total"), _jsxs("span", { className: "nre-full-marks", children: ["/ ", data.fullMarks] })] }), subjectKeys.map((key) => (_jsx("th", { scope: "col", className: "nre-subject-col", children: key }, key)))] }) }), _jsx("tbody", { children: data.rows.map((row) => (_jsxs("tr", { children: [_jsx("th", { scope: "row", className: cx("nre-row-key", "nre-key", colorClassForKey(row.key)), children: row.key }), _jsxs("td", { className: "nre-total", children: [resolveLocalizedText(row.total.display, locale), row.total.notRun > 0 && (_jsx("span", { className: "nre-total-not-run", children: countText(locale, "scoreboard.notRun", row.total.notRun) })), row.total.unscorable > 0 && (_jsx("span", { className: "nre-total-unscorable", children: countText(locale, "scoreboard.unscorable", row.total.unscorable) }))] }), subjectKeys.map((key) => {
15
15
  const subject = row.subjects.find((s) => s.key === key);
16
16
  if (!subject)
17
17
  return _jsx("td", { className: "nre-td-empty" }, key);
18
- return (_jsxs("td", { className: "nre-subject", children: [_jsxs("span", { className: "nre-subject-score", title: localeText(locale, "scoreboard.subjectTitle", {
19
- evals: subject.evals,
18
+ return (_jsxs("td", { className: "nre-subject", children: [_jsx("span", { className: "nre-subject-score", title: localeText(locale, "scoreboard.subjectTitle", {
19
+ questions: subject.questions,
20
20
  earned: subject.earned,
21
21
  possible: subject.possible,
22
- }), children: [subject.earned, "/", subject.possible] }), subject.missing > 0 && (_jsx("span", { className: "nre-subject-missing", children: countText(locale, "scoreboard.missing", subject.missing) }))] }, key));
22
+ }), children: resolveLocalizedText(subject.display, locale) }), subject.notRun > 0 && (_jsx("span", { className: "nre-subject-not-run", children: localeText(locale, "scoreboard.notRunText", { n: subject.notRun }) })), subject.unscorable > 0 && (_jsx("span", { className: "nre-subject-unscorable", children: localeText(locale, "scoreboard.unscorableText", { n: subject.unscorable }) }))] }, key));
23
23
  })] }, row.key))) })] }), _jsxs("p", { className: "nre-weights", children: [localeText(locale, "scoreboard.weights"), " ", data.weights.length === 0
24
24
  ? localeText(locale, "scoreboard.allWeights")
25
- : data.weights.map((w) => (_jsxs("span", { className: "nre-weight", children: [w.prefix, " \u00D7", w.weight] }, w.prefix))), data.weights.length > 0 && _jsx("span", { className: "nre-weight-rest", children: localeText(locale, "scoreboard.othersWeight") })] })] }));
25
+ : data.weights.map((w) => (_jsxs("span", { className: "nre-weight", children: [w.prefix, " \u00D7", w.weight] }, w.prefix))), data.weights.length > 0 && _jsx("span", { className: "nre-weight-rest", children: localeText(locale, "scoreboard.othersWeight") })] }), data.ignoredEvals > 0 && (_jsx("p", { className: "nre-scoreboard-ignored", children: countText(locale, "scoreboard.ignored", data.ignoredEvals) }))] }));
26
26
  }
@@ -1,9 +1,9 @@
1
1
  import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
2
- import { DEFAULT_REPORT_LOCALE, localeText } from "../locale.js";
2
+ import { DEFAULT_REPORT_LOCALE, localeText, resolveLocalizedText } from "../locale.js";
3
3
  export function MetricCellView({ cell, attemptHref, locale = DEFAULT_REPORT_LOCALE, }) {
4
4
  // 全 null(没有任何有效样本)→ 缺数据文案,绝不画 0;total 仍如实入 title
5
5
  if (cell.value === null) {
6
6
  return (_jsx("span", { className: "nre-cell nre-cell-missing", children: _jsx("span", { className: "nre-missing", title: localeText(locale, "cell.noneMeasurableTitle", { total: cell.total }), children: localeText(locale, "cell.missing") }) }));
7
7
  }
8
- return (_jsxs("span", { className: "nre-cell", children: [_jsx("span", { className: "nre-value", title: localeText(locale, "cell.measuredTitle", { samples: cell.samples, total: cell.total }), children: cell.display }), cell.samples < cell.total && (_jsxs("sup", { className: "nre-coverage", title: localeText(locale, "cell.coverageTitle", { samples: cell.samples, total: cell.total }), children: [cell.samples, "/", cell.total] })), attemptHref && cell.refs && cell.refs.length > 0 && (_jsx("span", { className: "nre-refs", children: cell.refs.map((locator, i) => (_jsxs("a", { className: "nre-ref", href: attemptHref(locator), children: ["#", i + 1] }, locator))) }))] }));
8
+ return (_jsxs("span", { className: "nre-cell", children: [_jsx("span", { className: "nre-value", title: localeText(locale, "cell.measuredTitle", { samples: cell.samples, total: cell.total }), children: resolveLocalizedText(cell.display, locale) }), cell.samples < cell.total && (_jsxs("sup", { className: "nre-coverage", title: localeText(locale, "cell.coverageTitle", { samples: cell.samples, total: cell.total }), children: [cell.samples, "/", cell.total] })), attemptHref && cell.refs && cell.refs.length > 0 && (_jsx("span", { className: "nre-refs", children: cell.refs.map((locator, i) => (_jsxs("a", { className: "nre-ref", href: attemptHref(locator), children: ["#", i + 1] }, locator))) }))] }));
9
9
  }
@@ -1,19 +1,15 @@
1
- import type { AttemptListItem, DeltaData, EvalListItem, ExperimentListItem, GroupSummaryData, LineData, MatrixData, MetricColumn, OverviewData, ScatterData, ScoreboardData, TableData } from "../types.ts";
1
+ import type { AttemptListItem, DeltaData, EvalListItem, ExperimentListItem, LineData, MatrixData, MetricColumn, ScatterData, ScopeSummaryData, ScoreboardData, TableData } from "../types.ts";
2
2
  export declare const passRateColumn: MetricColumn;
3
3
  export declare const codeLinesColumn: MetricColumn;
4
4
  export declare const costColumn: MetricColumn;
5
- export declare const overviewData: OverviewData;
6
- export declare const overviewWithCost: OverviewData;
7
- /** GroupSummary.data 的产物形态:eval 级折叠计票 + 旧 GroupSelector 口径的通过率(包成 MetricCell)。 */
8
- export declare const groupSummaryData: GroupSummaryData;
9
- export declare const tableData: TableData<"task-pass-rate" | "code-lines">;
10
- /** rows: "experiment" 的榜单形态:行携带 agent/model 元信息与 eval 级折叠计票。 */
11
- export declare const tableDataWithMeta: TableData<"task-pass-rate">;
5
+ /** scopeSummaryData 的产物形态:两级计票恒随行,成功率与总成本是官方 MetricCell。 */
6
+ export declare const scopeSummaryData: ScopeSummaryData;
7
+ export declare const tableData: TableData;
12
8
  export declare const matrixData: MatrixData;
13
9
  export declare const scoreboardData: ScoreboardData;
14
10
  export declare const scatterData: ScatterData;
15
11
  export declare const lineData: LineData;
16
- export declare const deltaData: DeltaData<"task-pass-rate" | "cost">;
12
+ export declare const deltaData: DeltaData;
17
13
  export declare const attemptListItems: AttemptListItem[];
18
14
  export declare const evalListItems: EvalListItem[];
19
15
  export declare const experimentListItems: ExperimentListItem[];