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.
- package/INDEX.md +23 -23
- package/README.zh.md +6 -6
- package/dist/agents/types.d.ts +2 -2
- package/dist/i18n/zh-CN.d.ts +3 -3
- package/dist/report/aggregate.d.ts +32 -26
- package/dist/report/aggregate.js +157 -76
- package/dist/report/built-in/index.d.ts +2 -0
- package/dist/report/built-in/index.js +8 -0
- package/dist/report/components.d.ts +91 -164
- package/dist/report/components.js +377 -114
- package/dist/report/compute.d.ts +86 -73
- package/dist/report/compute.js +592 -432
- package/dist/report/flag.d.ts +28 -17
- package/dist/report/flag.js +86 -16
- package/dist/report/format.d.ts +11 -11
- package/dist/report/format.js +17 -15
- package/dist/report/index.d.ts +16 -17
- package/dist/report/index.js +20 -22
- package/dist/report/load.js +3 -2
- package/dist/report/locale.d.ts +49 -34
- package/dist/report/locale.js +106 -58
- package/dist/report/metrics.d.ts +10 -3
- package/dist/report/metrics.js +46 -12
- package/dist/report/primitives.d.ts +42 -15
- package/dist/report/primitives.js +135 -26
- package/dist/report/react/AttemptList.d.ts +10 -8
- package/dist/report/react/AttemptList.js +18 -10
- package/dist/report/react/DeltaTable.js +19 -18
- package/dist/report/react/EvalList.d.ts +3 -3
- package/dist/report/react/EvalList.js +0 -0
- package/dist/report/react/ExperimentComparison.d.ts +4 -2
- package/dist/report/react/ExperimentComparison.js +5 -4
- package/dist/report/react/ExperimentList.d.ts +3 -3
- package/dist/report/react/ExperimentList.js +16 -15
- package/dist/report/react/MetricBars.js +5 -4
- package/dist/report/react/MetricLine.js +12 -5
- package/dist/report/react/MetricMatrix.js +1 -1
- package/dist/report/react/MetricScatter.js +54 -17
- package/dist/report/react/MetricTable.js +2 -12
- package/dist/report/react/ScopeSummary.d.ts +10 -0
- package/dist/report/react/ScopeSummary.js +17 -0
- package/dist/report/react/Scoreboard.js +6 -6
- package/dist/report/react/cell.js +2 -2
- package/dist/report/react/fixtures.d.ts +5 -9
- package/dist/report/react/fixtures.js +105 -149
- package/dist/report/react/index.d.ts +15 -5
- package/dist/report/react/index.js +18 -7
- package/dist/report/report.d.ts +137 -20
- package/dist/report/report.js +261 -34
- package/dist/report/text/faces.d.ts +17 -19
- package/dist/report/text/faces.js +225 -157
- package/dist/report/text/plot.js +1 -1
- package/dist/report/text/table.js +2 -2
- package/dist/report/tree.d.ts +90 -40
- package/dist/report/tree.js +252 -94
- package/dist/report/types.d.ts +245 -300
- package/dist/report/types.js +4 -3
- package/dist/report/web.d.ts +21 -5
- package/dist/report/web.js +42 -16
- package/dist/results/select.d.ts +38 -16
- package/dist/results/select.js +73 -25
- package/dist/results/types.d.ts +38 -14
- package/dist/shared/aggregate.d.ts +3 -2
- package/dist/shared/aggregate.js +5 -4
- package/docs-site/zh/README.md +44 -0
- package/docs-site/zh/examples/ai-agent-application.mdx +63 -0
- package/docs-site/zh/examples/coding-agent-extensions.mdx +57 -0
- package/docs-site/zh/examples/index.mdx +50 -0
- package/docs-site/zh/{concepts → explanation}/adapter.mdx +11 -11
- package/docs-site/zh/{concepts → explanation}/assert.mdx +7 -7
- package/docs-site/zh/{concepts → explanation}/drive.mdx +8 -8
- package/docs-site/zh/{concepts → explanation}/evals.mdx +4 -4
- package/docs-site/zh/{concepts → explanation}/experiment.mdx +8 -8
- package/docs-site/zh/{concepts → explanation}/hitl.mdx +8 -8
- package/docs-site/zh/{concepts → explanation}/judge.mdx +5 -5
- package/docs-site/zh/{concepts → explanation}/overview.mdx +5 -5
- package/docs-site/zh/{guides → explanation}/runner.mdx +1 -1
- package/docs-site/zh/{concepts → explanation}/tier.mdx +6 -6
- package/docs-site/zh/{guides → how-to}/agent-feedback-loop.mdx +7 -7
- package/docs-site/zh/{guides → how-to}/authoring.mdx +2 -2
- package/docs-site/zh/{guides → how-to}/connect-otel.mdx +6 -6
- package/docs-site/zh/{guides → how-to}/connect-your-agent.mdx +18 -18
- package/docs-site/zh/{guides → how-to}/custom-reports.mdx +6 -6
- package/docs-site/zh/{guides → how-to}/experiments.mdx +3 -3
- package/docs-site/zh/{guides → how-to}/publish-report.mdx +2 -2
- package/docs-site/zh/{guides → how-to}/sandbox-agent.mdx +2 -2
- package/docs-site/zh/{guides → how-to}/sandbox-providers.mdx +1 -1
- package/docs-site/zh/{guides → how-to}/viewing-results.mdx +6 -6
- package/docs-site/zh/{guides → how-to}/write-experiment.mdx +3 -3
- package/docs-site/zh/{guides → how-to}/write-send.mdx +13 -13
- package/docs-site/zh/index.mdx +23 -25
- package/docs-site/zh/introduction.mdx +8 -8
- package/docs-site/zh/reference/builtin-agents.mdx +5 -5
- package/docs-site/zh/reference/capabilities.mdx +6 -6
- package/docs-site/zh/reference/cli.mdx +9 -7
- package/docs-site/zh/reference/define-agent.mdx +1 -1
- package/docs-site/zh/reference/events.mdx +3 -3
- package/docs-site/zh/{guides → reference}/official-adapters.mdx +7 -7
- package/docs-site/zh/{guides → reference}/report-components.mdx +5 -5
- package/docs-site/zh/{guides → reference}/results-data.mdx +5 -5
- package/docs-site/zh/{guides → troubleshooting}/debug-sandbox.mdx +2 -2
- package/docs-site/zh/{guides → troubleshooting}/debugging.mdx +4 -2
- package/docs-site/zh/{quickstart.mdx → tutorials/quickstart.mdx} +5 -17
- package/package.json +6 -2
- package/src/agents/index.ts +2 -2
- package/src/agents/openai-compat.ts +1 -1
- package/src/agents/streaming.ts +2 -2
- package/src/agents/types.ts +3 -3
- package/src/cli.ts +42 -23
- package/src/context/context.ts +1 -1
- package/src/context/session.test.ts +1 -1
- package/src/context/session.ts +1 -1
- package/src/i18n/en.ts +18 -16
- package/src/i18n/zh-CN.ts +16 -15
- package/src/report/aggregate.ts +175 -87
- package/src/report/built-in/index.tsx +9 -0
- package/src/report/components.tsx +625 -285
- package/src/report/compute.ts +717 -515
- package/src/report/dual-render.test.tsx +738 -1148
- package/src/report/flag.ts +97 -33
- package/src/report/format.ts +18 -22
- package/src/report/index.ts +113 -58
- package/src/report/load.ts +3 -2
- package/src/report/locale.ts +120 -69
- package/src/report/metrics.ts +42 -12
- package/src/report/primitives.tsx +190 -45
- package/src/report/react/AttemptList.tsx +32 -20
- package/src/report/react/DeltaTable.tsx +63 -45
- package/src/report/react/EvalList.tsx +0 -0
- package/src/report/react/ExperimentComparison.tsx +12 -7
- package/src/report/react/ExperimentList.tsx +38 -26
- package/src/report/react/MetricBars.tsx +5 -4
- package/src/report/react/MetricLine.tsx +13 -8
- package/src/report/react/MetricMatrix.tsx +2 -2
- package/src/report/react/MetricScatter.tsx +74 -20
- package/src/report/react/MetricTable.tsx +4 -76
- package/src/report/react/ScopeSummary.tsx +86 -0
- package/src/report/react/Scoreboard.tsx +28 -10
- package/src/report/react/cell.tsx +2 -2
- package/src/report/react/enhance.js +57 -5
- package/src/report/react/fixtures.ts +109 -156
- package/src/report/react/index.tsx +24 -39
- package/src/report/react/render.test.tsx +139 -104
- package/src/report/react/styles.css +181 -91
- package/src/report/report.test.ts +761 -1031
- package/src/report/report.ts +425 -47
- package/src/report/text/faces.ts +257 -164
- package/src/report/text/plot.ts +1 -1
- package/src/report/text/table.ts +2 -2
- package/src/report/tree.ts +362 -104
- package/src/report/types.ts +257 -287
- package/src/report/web.ts +63 -20
- package/src/results/attempt-evidence.test.ts +4 -4
- package/src/results/attempt-evidence.ts +5 -5
- package/src/results/copy.ts +6 -6
- package/src/results/host-equivalence.test.ts +26 -14
- package/src/results/index.ts +10 -4
- package/src/results/open.ts +8 -4
- package/src/results/results.test.ts +4 -3
- package/src/results/select.ts +104 -34
- package/src/results/types.ts +36 -14
- package/src/runner/feedback/human.test.ts +1 -1
- package/src/runner/run.ts +1 -1
- package/src/sandbox/cli-commands.ts +2 -2
- package/src/scoring/judge.test.ts +1 -1
- package/src/shared/aggregate.ts +5 -4
- package/src/show/compose.ts +50 -67
- package/src/show/index.ts +107 -56
- package/src/show/render.ts +43 -27
- package/src/show/report-host.test.ts +188 -0
- package/src/show/report-host.ts +375 -0
- package/src/show/show.test.ts +86 -36
- package/src/view/app/App.test.tsx +69 -0
- package/src/view/app/App.tsx +144 -48
- package/src/view/app/components/AttemptModal.tsx +324 -63
- package/src/view/app/components/CodeView.tsx +10 -4
- package/src/view/app/i18n.ts +31 -17
- package/src/view/app/lib/artifact-url.ts +14 -3
- package/src/view/app/main.tsx +13 -8
- package/src/view/app/pages/{RunsPage.tsx → AttemptsPage.tsx} +6 -6
- package/src/view/app/types.ts +4 -1
- package/src/view/artifact-serving.test.ts +21 -1
- package/src/view/client-dist/app.css +1 -1
- package/src/view/client-dist/app.js +6 -6
- package/src/view/data.test.ts +9 -3
- package/src/view/data.ts +145 -49
- package/src/view/index.ts +48 -44
- package/src/view/server.ts +35 -15
- package/src/view/shared/types.ts +34 -5
- package/src/view/styles.css +224 -0
- package/src/view/view-report.test.ts +161 -57
- package/dist/report/built-ins/experiment-comparison.d.ts +0 -39
- package/dist/report/built-ins/experiment-comparison.js +0 -119
- package/dist/report/built-ins/index.d.ts +0 -2
- package/dist/report/built-ins/index.js +0 -2
- package/dist/report/react/GroupSummary.d.ts +0 -8
- package/dist/report/react/GroupSummary.js +0 -8
- package/dist/report/react/RunOverview.d.ts +0 -8
- package/dist/report/react/RunOverview.js +0 -12
- package/docs-site/zh/example/ai-agent-application.mdx +0 -152
- package/docs-site/zh/example/claude-code-codex-plugin.mdx +0 -167
- package/docs-site/zh/example/claude-code-codex-skill.mdx +0 -152
- package/docs-site/zh/example/showcase.mdx +0 -39
- package/src/report/built-in-user-parity.test.tsx +0 -597
- package/src/report/built-ins/experiment-comparison.tsx +0 -179
- package/src/report/built-ins/index.ts +0 -7
- package/src/report/react/GroupSummary.tsx +0 -66
- package/src/report/react/RunOverview.tsx +0 -109
- /package/docs-site/zh/{example/tier1-ai-sdk-v7.mdx → examples/integrations/ai-sdk-v7.mdx} +0 -0
- /package/docs-site/zh/{example/tier1-claude-sdk.mdx → examples/integrations/claude-sdk.mdx} +0 -0
- /package/docs-site/zh/{example/tier1-codex-sdk.mdx → examples/integrations/codex-sdk.mdx} +0 -0
- /package/docs-site/zh/{example/tier1-langgraph.mdx → examples/integrations/langgraph.mdx} +0 -0
- /package/docs-site/zh/{example/tier1-pi-sdk.mdx → examples/integrations/pi-sdk.mdx} +0 -0
- /package/docs-site/zh/{guides → how-to}/ci-integration.mdx +0 -0
- /package/docs-site/zh/{guides → how-to}/dataset-fanout.mdx +0 -0
- /package/docs-site/zh/{guides → how-to}/fixtures.mdx +0 -0
- /package/docs-site/zh/{guides → how-to}/reporters.mdx +0 -0
- /package/docs-site/zh/{guides → how-to}/scoring-guide.mdx +0 -0
|
@@ -1,17 +1,17 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import { defineComponent } from "./tree.js";
|
|
3
|
-
import { localeText } from "./locale.js";
|
|
4
|
-
import { indentBlock, joinColumns, wrapDisplay } from "./text/layout.js";
|
|
2
|
+
import { COMPONENT_RAW_CHILDREN, COMPONENT_ROLE, defineComponent } from "./tree.js";
|
|
3
|
+
import { localeText, resolveLocalizedText } from "./locale.js";
|
|
4
|
+
import { indentBlock, joinColumns, stringWidth, wrapDisplay } from "./text/layout.js";
|
|
5
5
|
import { renderTableText } from "./text/table.js";
|
|
6
6
|
function childArray(children) {
|
|
7
7
|
if (children === null || children === undefined || typeof children === "boolean")
|
|
8
8
|
return [];
|
|
9
|
-
return Array.isArray(children) ? children : [children];
|
|
9
|
+
return Array.isArray(children) ? [...children] : [children];
|
|
10
10
|
}
|
|
11
11
|
function cx(...parts) {
|
|
12
12
|
return parts.filter(Boolean).join(" ");
|
|
13
13
|
}
|
|
14
|
-
/** 纵向依次排列:网页是块级堆叠,终端是逐块输出(块间空一行)
|
|
14
|
+
/** 纵向依次排列:网页是块级堆叠,终端是逐块输出(块间空一行)。两面都按声明序。 */
|
|
15
15
|
export const Col = defineComponent({
|
|
16
16
|
web({ children, className }) {
|
|
17
17
|
return _jsx("div", { className: cx("nre", "nre-col", className), children: children });
|
|
@@ -24,10 +24,15 @@ export const Col = defineComponent({
|
|
|
24
24
|
},
|
|
25
25
|
});
|
|
26
26
|
Col.displayName = "Col";
|
|
27
|
-
// Row 的每栏至少留这个宽度,不硬挤;不够就降级纵向
|
|
28
|
-
const MIN_COLUMN_WIDTH = 24;
|
|
29
27
|
const COLUMN_SEPARATOR = " │ ";
|
|
30
|
-
/**
|
|
28
|
+
/** 一段已渲染文本的自然显示宽度(最长一行)。 */
|
|
29
|
+
function blockWidth(block) {
|
|
30
|
+
return Math.max(...block.split("\n").map((line) => stringWidth(line)), 0);
|
|
31
|
+
}
|
|
32
|
+
/**
|
|
33
|
+
* 并排:web 面横排;text 面在可用宽度装得下全部子块时按显示宽度并排(与 `columns` 工具
|
|
34
|
+
* 同一把尺),装不下时整块退化为纵向堆叠——不截断、不隐藏任何子块。
|
|
35
|
+
*/
|
|
31
36
|
export const Row = defineComponent({
|
|
32
37
|
web({ children, className }) {
|
|
33
38
|
return _jsx("div", { className: cx("nre", "nre-row", className), children: children });
|
|
@@ -38,44 +43,49 @@ export const Row = defineComponent({
|
|
|
38
43
|
return "";
|
|
39
44
|
if (blocks.length === 1)
|
|
40
45
|
return ctx.render(blocks[0]);
|
|
41
|
-
const
|
|
42
|
-
if (
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
46
|
+
const rendered = blocks.map((child) => ctx.render(child)).filter((block) => block.length > 0);
|
|
47
|
+
if (rendered.length === 0)
|
|
48
|
+
return "";
|
|
49
|
+
const widths = rendered.map(blockWidth);
|
|
50
|
+
const total = widths.reduce((sum, w) => sum + w, 0) + COLUMN_SEPARATOR.length * (rendered.length - 1);
|
|
51
|
+
if (total > ctx.width) {
|
|
52
|
+
// 装不下:整块退化为纵向堆叠,与 Col 同一形态
|
|
53
|
+
return rendered.join("\n\n");
|
|
48
54
|
}
|
|
49
|
-
|
|
50
|
-
return joinColumns(rendered, rendered.map(() => columnWidth), COLUMN_SEPARATOR);
|
|
55
|
+
return joinColumns(rendered, widths, COLUMN_SEPARATOR);
|
|
51
56
|
},
|
|
52
57
|
});
|
|
53
58
|
Row.displayName = "Row";
|
|
54
59
|
/** 带标题的块:网页是标题层级,终端是标题行加缩进。 */
|
|
55
60
|
export const Section = defineComponent({
|
|
56
|
-
web({ title, children, className }) {
|
|
57
|
-
return (_jsxs("section", { className: cx("nre", "nre-section", className), children: [_jsx("h2", { className: "nre-section-title", children: title }), children] }));
|
|
61
|
+
web({ title, children, className }, ctx) {
|
|
62
|
+
return (_jsxs("section", { className: cx("nre", "nre-section", className), children: [_jsx("h2", { className: "nre-section-title", children: resolveLocalizedText(title, ctx.locale) }), children] }));
|
|
58
63
|
},
|
|
59
64
|
text({ title, children }, ctx) {
|
|
65
|
+
const heading = resolveLocalizedText(title, ctx.locale);
|
|
60
66
|
const body = childArray(children)
|
|
61
67
|
.map((child) => ctx.render(child, ctx.width - 2))
|
|
62
68
|
.filter((block) => block.length > 0)
|
|
63
69
|
.join("\n\n");
|
|
64
|
-
return body.length > 0 ? `${
|
|
70
|
+
return body.length > 0 ? `${heading}\n${indentBlock(body, " ")}` : heading;
|
|
65
71
|
},
|
|
66
72
|
});
|
|
67
73
|
Section.displayName = "Section";
|
|
68
|
-
/**
|
|
74
|
+
/** 自由文本的显式载体:web 面负责转义,text 面按显示宽度折行。 */
|
|
69
75
|
export const Text = defineComponent({
|
|
70
76
|
web({ children, className }) {
|
|
71
77
|
return _jsx("p", { className: cx("nre", "nre-text", className), children: children });
|
|
72
78
|
},
|
|
73
79
|
text({ children }, ctx) {
|
|
74
|
-
return wrapDisplay(
|
|
80
|
+
return wrapDisplay(String(children), ctx.width).join("\n");
|
|
75
81
|
},
|
|
76
82
|
});
|
|
77
83
|
Text.displayName = "Text";
|
|
78
|
-
|
|
84
|
+
Text[COMPONENT_RAW_CHILDREN] = true;
|
|
85
|
+
/**
|
|
86
|
+
* 注入页级全局 CSS:树位置只决定声明顺序,不限定作用域;text 面零输出。
|
|
87
|
+
* 配置对象形态的报告要全站样式优先用外壳 styles,两条通道注入同一增强层。
|
|
88
|
+
*/
|
|
79
89
|
export const Style = defineComponent({
|
|
80
90
|
web({ children }) {
|
|
81
91
|
return _jsx("style", { children: children });
|
|
@@ -85,7 +95,101 @@ export const Style = defineComponent({
|
|
|
85
95
|
},
|
|
86
96
|
});
|
|
87
97
|
Style.displayName = "Style";
|
|
98
|
+
Style[COMPONENT_RAW_CHILDREN] = true;
|
|
99
|
+
/** Tabs 的直接 Tab 子项(数组 / Fragment 已展平;结构合法性由树校验保证)。 */
|
|
100
|
+
function tabEntries(children) {
|
|
101
|
+
const out = [];
|
|
102
|
+
const visit = (node) => {
|
|
103
|
+
if (node === null || node === undefined || typeof node === "boolean")
|
|
104
|
+
return;
|
|
105
|
+
if (Array.isArray(node)) {
|
|
106
|
+
for (const child of node)
|
|
107
|
+
visit(child);
|
|
108
|
+
return;
|
|
109
|
+
}
|
|
110
|
+
if (typeof node === "object" && node !== null && "props" in node) {
|
|
111
|
+
const element = node;
|
|
112
|
+
if (element.type === Symbol.for("react.fragment")) {
|
|
113
|
+
visit(element.props.children);
|
|
114
|
+
return;
|
|
115
|
+
}
|
|
116
|
+
out.push({
|
|
117
|
+
title: element.props.title ?? "",
|
|
118
|
+
children: element.props.children,
|
|
119
|
+
className: element.props.className,
|
|
120
|
+
});
|
|
121
|
+
}
|
|
122
|
+
};
|
|
123
|
+
visit(children);
|
|
124
|
+
return out;
|
|
125
|
+
}
|
|
126
|
+
/**
|
|
127
|
+
* 页内并列视图的可切换块。tab 是页内浏览状态,不是数据边界,也不是宿主寻址单位——
|
|
128
|
+
* 两个渲染面都输出全部 tab 的完整内容:web 静态 HTML 每 tab 一个 <details>(首个默认展开,
|
|
129
|
+
* 渐进增强升级成单选 tab 条,切换不改变数据);text 面按声明序输出带标题分节,不折成索引
|
|
130
|
+
* 也不省略(tab 没有选择器,索引只能是死路)。
|
|
131
|
+
*/
|
|
132
|
+
export const Tabs = defineComponent({
|
|
133
|
+
web({ children, className }, ctx) {
|
|
134
|
+
const tabs = tabEntries(children);
|
|
135
|
+
return (_jsx("div", { className: cx("nre", "nre-tabs", className), "data-nre-tabs": true, children: tabs.map((tab, i) => (_jsxs("details", { className: cx("nre-tab", tab.className), open: i === 0, children: [_jsx("summary", { className: "nre-tab-title", children: resolveLocalizedText(tab.title, ctx.locale) }), _jsx("div", { className: "nre-tab-body", children: tab.children })] }, i))) }));
|
|
136
|
+
},
|
|
137
|
+
text({ children }, ctx) {
|
|
138
|
+
const tabs = tabEntries(children);
|
|
139
|
+
return tabs
|
|
140
|
+
.map((tab) => {
|
|
141
|
+
const heading = resolveLocalizedText(tab.title, ctx.locale);
|
|
142
|
+
const body = childArray(tab.children)
|
|
143
|
+
.map((child) => ctx.render(child, ctx.width - 2))
|
|
144
|
+
.filter((block) => block.length > 0)
|
|
145
|
+
.join("\n\n");
|
|
146
|
+
return body.length > 0 ? `${heading}\n${indentBlock(body, " ")}` : heading;
|
|
147
|
+
})
|
|
148
|
+
.join("\n\n");
|
|
149
|
+
},
|
|
150
|
+
});
|
|
151
|
+
Tabs.displayName = "Tabs";
|
|
152
|
+
Tabs[COMPONENT_ROLE] = "tabs";
|
|
153
|
+
/** 只能直接放在 <Tabs> 下;除通用 children / className 外只有 title。不参与路由,没有 id。 */
|
|
154
|
+
export const Tab = defineComponent({
|
|
155
|
+
web({ children, className }) {
|
|
156
|
+
return _jsx("div", { className: cx("nre", "nre-tab-body", className), children: children });
|
|
157
|
+
},
|
|
158
|
+
text({ children }, ctx) {
|
|
159
|
+
return childArray(children)
|
|
160
|
+
.map((child) => ctx.render(child))
|
|
161
|
+
.filter((block) => block.length > 0)
|
|
162
|
+
.join("\n\n");
|
|
163
|
+
},
|
|
164
|
+
});
|
|
165
|
+
Tab.displayName = "Tab";
|
|
166
|
+
Tab[COMPONENT_ROLE] = "tab";
|
|
88
167
|
const MISSING_MARK = "—";
|
|
168
|
+
/** 列 key 唯一、行 cells 不携带未声明 key、空列拒绝——无类型 JS 输入在渲染前同样校验。 */
|
|
169
|
+
function validateTableProps(props) {
|
|
170
|
+
if (!Array.isArray(props.columns) || props.columns.length === 0) {
|
|
171
|
+
throw new Error("Table needs at least one column: pass columns: [{ key, header }] — an empty column list renders nothing readable.");
|
|
172
|
+
}
|
|
173
|
+
const keys = new Set();
|
|
174
|
+
for (const column of props.columns) {
|
|
175
|
+
if (keys.has(column.key)) {
|
|
176
|
+
throw new Error(`Table column key "${column.key}" is declared twice — column keys address row.cells and must be unique. Rename one column.`);
|
|
177
|
+
}
|
|
178
|
+
keys.add(column.key);
|
|
179
|
+
}
|
|
180
|
+
const rowKeys = new Set();
|
|
181
|
+
for (const row of props.rows) {
|
|
182
|
+
if (rowKeys.has(row.key)) {
|
|
183
|
+
throw new Error(`Table row key "${row.key}" is declared twice — row keys are the row identity and must be unique.`);
|
|
184
|
+
}
|
|
185
|
+
rowKeys.add(row.key);
|
|
186
|
+
for (const cellKey of Object.keys(row.cells)) {
|
|
187
|
+
if (!keys.has(cellKey)) {
|
|
188
|
+
throw new Error(`Table row "${row.key}" has a cell "${cellKey}" that no column declares. Declare the column in columns, or drop the stray cell.`);
|
|
189
|
+
}
|
|
190
|
+
}
|
|
191
|
+
}
|
|
192
|
+
}
|
|
89
193
|
/**
|
|
90
194
|
* 自定义表的标准件:列由报告作者定,格子是算好的显示值,两个面各自排整齐。
|
|
91
195
|
*
|
|
@@ -98,16 +202,21 @@ const MISSING_MARK = "—";
|
|
|
98
202
|
* 这个组件上:自定义表和官方表用同一把尺子。
|
|
99
203
|
*/
|
|
100
204
|
export const Table = defineComponent({
|
|
101
|
-
web(
|
|
205
|
+
web(props, ctx) {
|
|
206
|
+
validateTableProps(props);
|
|
207
|
+
const { columns, rows, locale, className } = props;
|
|
102
208
|
const chrome = locale ?? ctx.locale;
|
|
103
209
|
const hasLocator = rows.some((row) => row.locator !== undefined);
|
|
104
210
|
const alignClass = (align) => (align === "right" ? "nre-align-right" : undefined);
|
|
105
|
-
return (_jsxs("table", { className: cx("nre", "nre-table", className), children: [_jsx("thead", { children: _jsxs("tr", { children: [columns.map((column) => (_jsx("th", { scope: "col", className: alignClass(column.align), children: column.header }, column.key))), hasLocator ? _jsx("th", { scope: "col", children: localeText(chrome, "table.attempt") }) : null] }) }), _jsx("tbody", { children: rows.map((row) => (_jsxs("tr", { children: [columns.map((column) => {
|
|
211
|
+
return (_jsxs("table", { className: cx("nre", "nre-table", className), children: [_jsx("thead", { children: _jsxs("tr", { children: [columns.map((column) => (_jsx("th", { scope: "col", className: alignClass(column.align), children: resolveLocalizedText(column.header, chrome) }, column.key))), hasLocator ? _jsx("th", { scope: "col", children: localeText(chrome, "table.attempt") }) : null] }) }), _jsx("tbody", { children: rows.map((row) => (_jsxs("tr", { children: [columns.map((column) => {
|
|
106
212
|
const value = row.cells[column.key];
|
|
107
213
|
const missing = value === null || value === undefined;
|
|
108
214
|
return (_jsx("td", { className: alignClass(column.align), children: missing ? _jsx("span", { className: "nre-missing", children: MISSING_MARK }) : value }, column.key));
|
|
109
215
|
}), hasLocator ? (_jsx("td", { children: row.locator ? (_jsx("a", { className: "nre-locator", href: ctx.attemptHref(row.locator), children: row.locator })) : (_jsx("span", { className: "nre-missing", children: MISSING_MARK })) })) : null] }, row.key))) })] }));
|
|
110
216
|
},
|
|
111
|
-
text
|
|
217
|
+
text(props, ctx) {
|
|
218
|
+
validateTableProps(props);
|
|
219
|
+
return renderTableText(props, ctx);
|
|
220
|
+
},
|
|
112
221
|
});
|
|
113
222
|
Table.displayName = "Table";
|
|
@@ -5,19 +5,21 @@ import { type ReportLocale } from "../locale.ts";
|
|
|
5
5
|
/** locator + 判定符的普通 <a>,AttemptList/EvalList/ExperimentList 共用。 */
|
|
6
6
|
export declare function AttemptLocatorBadge({ item, attemptHref, }: {
|
|
7
7
|
item: Pick<AttemptListItem, "locator" | "verdict">;
|
|
8
|
-
attemptHref
|
|
8
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
9
9
|
}): ReactElement;
|
|
10
|
-
/**
|
|
11
|
-
export declare function
|
|
10
|
+
/** failureSummary + moreFailures 的展示形态:摘要原样,+N 计数由 moreFailures 驱动。 */
|
|
11
|
+
export declare function failureSummaryText(item: Pick<AttemptListItem, "failureSummary" | "moreFailures">, locale: ReportLocale): string | null;
|
|
12
|
+
/** 一条 Attempt 的比较卡片;完整 assertions 通过 locator 下钻,不在列表内展开。 */
|
|
13
|
+
export declare function AttemptRow({ item, attemptHref, locale, }: {
|
|
12
14
|
item: AttemptListItem;
|
|
13
|
-
attemptHref
|
|
15
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
14
16
|
locale?: ReportLocale;
|
|
15
17
|
}): ReactElement;
|
|
16
|
-
export declare function AttemptList({
|
|
17
|
-
|
|
18
|
-
/**
|
|
18
|
+
export declare function AttemptList({ data, total, attemptHref, className, locale, }: {
|
|
19
|
+
data: readonly AttemptListItem[];
|
|
20
|
+
/** data 被 slice 时的原始数量;如实显示还剩多少条没展示。 */
|
|
19
21
|
total?: number;
|
|
20
|
-
attemptHref
|
|
22
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
21
23
|
className?: string;
|
|
22
24
|
locale?: ReportLocale;
|
|
23
25
|
}): ReactElement;
|
|
@@ -1,18 +1,26 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import { DEFAULT_REPORT_LOCALE, localeText } from "../locale.js";
|
|
3
|
-
import { attemptItemReason } from "../format.js";
|
|
2
|
+
import { DEFAULT_REPORT_LOCALE, countText, localeText } from "../locale.js";
|
|
4
3
|
import { colorClassForKey } from "./colors.js";
|
|
5
4
|
import { cx, formatDurationMs, formatUSD, verdictMark } from "./format.js";
|
|
5
|
+
const DEFAULT_ATTEMPT_HREF = (locator) => `#/attempt/${locator}`;
|
|
6
6
|
/** locator + 判定符的普通 <a>,AttemptList/EvalList/ExperimentList 共用。 */
|
|
7
|
-
export function AttemptLocatorBadge({ item, attemptHref, }) {
|
|
7
|
+
export function AttemptLocatorBadge({ item, attemptHref = DEFAULT_ATTEMPT_HREF, }) {
|
|
8
8
|
return (_jsxs("a", { className: cx("nre-locator", `nre-verdict-${item.verdict}`), href: attemptHref(item.locator), children: [item.locator, _jsx("span", { className: "nre-locator-mark", children: verdictMark(item.verdict) })] }));
|
|
9
9
|
}
|
|
10
|
-
/**
|
|
11
|
-
export function
|
|
12
|
-
|
|
13
|
-
|
|
10
|
+
/** failureSummary + moreFailures 的展示形态:摘要原样,+N 计数由 moreFailures 驱动。 */
|
|
11
|
+
export function failureSummaryText(item, locale) {
|
|
12
|
+
if (item.failureSummary === null)
|
|
13
|
+
return null;
|
|
14
|
+
return item.moreFailures > 0
|
|
15
|
+
? `${item.failureSummary} · ${countText(locale, "entityList.moreFailures", item.moreFailures)}`
|
|
16
|
+
: item.failureSummary;
|
|
14
17
|
}
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
+
/** 一条 Attempt 的比较卡片;完整 assertions 通过 locator 下钻,不在列表内展开。 */
|
|
19
|
+
export function AttemptRow({ item, attemptHref = DEFAULT_ATTEMPT_HREF, locale = DEFAULT_REPORT_LOCALE, }) {
|
|
20
|
+
const reason = failureSummaryText(item, locale);
|
|
21
|
+
return (_jsxs("li", { className: cx("nre-attempt", `nre-attempt-${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 })] }));
|
|
22
|
+
}
|
|
23
|
+
export function AttemptList({ data, total, attemptHref = DEFAULT_ATTEMPT_HREF, className, locale = DEFAULT_REPORT_LOCALE, }) {
|
|
24
|
+
const remaining = (total ?? data.length) - data.length;
|
|
25
|
+
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 }) }))] }));
|
|
18
26
|
}
|
|
@@ -1,24 +1,25 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import { DEFAULT_REPORT_LOCALE, localeText, resolveMetricLabel } from "../locale.js";
|
|
2
|
+
import { DEFAULT_REPORT_LOCALE, localeText, resolveLocalizedText, resolveMetricLabel } from "../locale.js";
|
|
3
3
|
import { MetricCellView } from "./cell.js";
|
|
4
4
|
import { colorClassForKey } from "./colors.js";
|
|
5
5
|
import { cx } from "./format.js";
|
|
6
|
-
|
|
7
|
-
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
if (!better)
|
|
13
|
-
return "nre-delta-neutral"; // 指标没表态方向,只显示不评判
|
|
14
|
-
const improved = better === "higher" ? delta > 0 : delta < 0;
|
|
15
|
-
return improved ? "nre-delta-good" : "nre-delta-bad";
|
|
16
|
-
}
|
|
6
|
+
const OUTCOME_CLASS = {
|
|
7
|
+
improved: "nre-delta-good",
|
|
8
|
+
regressed: "nre-delta-bad",
|
|
9
|
+
unchanged: "nre-delta-flat",
|
|
10
|
+
unavailable: "nre-delta-missing",
|
|
11
|
+
};
|
|
17
12
|
export function DeltaTable({ data, className, locale = DEFAULT_REPORT_LOCALE, }) {
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
13
|
+
if (data.rows.length === 0) {
|
|
14
|
+
return (_jsx("div", { className: cx("nre", "nre-delta-table-empty", className), children: _jsx("p", { className: "nre-missing", children: localeText(locale, "delta.empty", { experiments: data.experiments ?? 0 }) }) }));
|
|
15
|
+
}
|
|
16
|
+
return (_jsxs("table", { className: cx("nre", "nre-delta-table", className), children: [_jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", children: localeText(locale, "delta.pairHeader") }), data.columns.map((col) => (_jsxs("th", { scope: "col", className: "nre-metric-col", children: [resolveMetricLabel(col.label, locale, col.key), col.unit && _jsxs("span", { className: "nre-unit", children: ["(", col.unit, ")"] })] }, col.key)))] }) }), _jsx("tbody", { children: data.rows.map((row) => {
|
|
17
|
+
const label = resolveLocalizedText(row.label, locale);
|
|
18
|
+
return (_jsxs("tr", { children: [_jsxs("th", { scope: "row", className: "nre-pair", children: [_jsx("span", { className: cx("nre-row-key", "nre-key", colorClassForKey(label)), children: label }), _jsxs("span", { className: "nre-pair-ids", children: [row.a.key, " \u2192 ", row.b.key] })] }), data.columns.map((col) => {
|
|
19
|
+
const cell = row.cells[col.key];
|
|
20
|
+
if (!cell)
|
|
21
|
+
return _jsx("td", { className: "nre-td-empty" }, col.key);
|
|
22
|
+
return (_jsxs("td", { className: "nre-delta-cell", children: [_jsxs("span", { className: "nre-delta-side nre-delta-a", children: [_jsx("span", { className: "nre-delta-tag", children: "A" }), _jsx(MetricCellView, { cell: cell.a, locale: locale })] }), _jsxs("span", { className: "nre-delta-side nre-delta-b", children: [_jsx("span", { className: "nre-delta-tag", children: "B" }), _jsx(MetricCellView, { cell: cell.b, locale: locale })] }), _jsxs("span", { className: cx("nre-delta-side", "nre-delta-d", OUTCOME_CLASS[cell.outcome]), "data-outcome": cell.outcome, children: [_jsx("span", { className: "nre-delta-tag", children: "\u0394" }), cell.delta === null ? (_jsx("span", { className: "nre-missing", children: localeText(locale, "cell.missing") })) : (resolveLocalizedText(cell.display, locale))] })] }, col.key));
|
|
23
|
+
})] }, row.key));
|
|
24
|
+
}) })] }));
|
|
24
25
|
}
|
|
@@ -2,9 +2,9 @@ import type { ReactElement } from "react";
|
|
|
2
2
|
import type { AttemptLocator } from "../../results/locator.ts";
|
|
3
3
|
import type { EvalListItem } from "../types.ts";
|
|
4
4
|
import { type ReportLocale } from "../locale.ts";
|
|
5
|
-
export declare function EvalList({
|
|
6
|
-
|
|
7
|
-
attemptHref
|
|
5
|
+
export declare function EvalList({ data, attemptHref, className, locale, }: {
|
|
6
|
+
data: readonly EvalListItem[];
|
|
7
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
8
8
|
className?: string;
|
|
9
9
|
locale?: ReportLocale;
|
|
10
10
|
}): ReactElement;
|
|
Binary file
|
|
@@ -1,8 +1,10 @@
|
|
|
1
1
|
import type { ReactElement } from "react";
|
|
2
|
-
import type { ExperimentComparisonData } from "../
|
|
2
|
+
import type { ExperimentComparisonData } from "../types.ts";
|
|
3
3
|
import { type ReportLocale } from "../locale.ts";
|
|
4
|
-
|
|
4
|
+
import type { AttemptLocator } from "../../results/locator.ts";
|
|
5
|
+
export declare function ExperimentComparisonView({ data, className, locale, attemptHref, }: {
|
|
5
6
|
data: ExperimentComparisonData;
|
|
6
7
|
className?: string;
|
|
7
8
|
locale?: ReportLocale;
|
|
9
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
8
10
|
}): ReactElement;
|
|
@@ -1,11 +1,12 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
2
|
import { DEFAULT_REPORT_LOCALE, localeText } from "../locale.js";
|
|
3
|
-
import {
|
|
4
|
-
import {
|
|
3
|
+
import { ScopeSummary } from "./ScopeSummary.js";
|
|
4
|
+
import { MetricScatter } from "./MetricScatter.js";
|
|
5
|
+
import { ExperimentList } from "./ExperimentList.js";
|
|
5
6
|
import { cx } from "./format.js";
|
|
6
|
-
export function ExperimentComparisonView({ data, className, locale = DEFAULT_REPORT_LOCALE, }) {
|
|
7
|
+
export function ExperimentComparisonView({ data, className, locale = DEFAULT_REPORT_LOCALE, attemptHref, }) {
|
|
7
8
|
if (data.groups.length === 0) {
|
|
8
9
|
return (_jsx("div", { className: cx("nre", "nre-experiment-comparison", className), children: _jsx("p", { className: "nre-experiment-groups-empty", children: localeText(locale, "experimentComparison.empty") }) }));
|
|
9
10
|
}
|
|
10
|
-
return (_jsxs("div", { className: cx("nre", "nre-experiment-comparison", className), "data-nre-experiment-groups": true, children: [_jsx("div", { className: "nre-experiment-group-tabs", role: "tablist", "aria-label": localeText(locale, "experimentComparison.groups"), children: data.groups.map((group, index) => (_jsxs("div", { className: "nre-experiment-group-tab", role: "tab", tabIndex: index === 0 ? 0 : -1, "aria-selected": index === 0 ? "true" : "false", "aria-label": localeText(locale, "experimentComparison.selectGroup", { group: group.key }), "data-nre-experiment-group-select": index, children: [_jsx("strong", { className: "nre-experiment-group-name", children: group.key }), _jsx(
|
|
11
|
+
return (_jsxs("div", { className: cx("nre", "nre-experiment-comparison", className), "data-nre-experiment-groups": true, children: [_jsx("div", { className: "nre-experiment-group-tabs", role: "tablist", "aria-label": localeText(locale, "experimentComparison.groups"), children: data.groups.map((group, index) => (_jsxs("div", { className: "nre-experiment-group-tab", role: "tab", tabIndex: index === 0 ? 0 : -1, "aria-selected": index === 0 ? "true" : "false", "aria-label": localeText(locale, "experimentComparison.selectGroup", { group: group.key }), "data-nre-experiment-group-select": index, children: [_jsx("strong", { className: "nre-experiment-group-name", children: group.key }), _jsx(ScopeSummary, { data: group.summary, locale: locale })] }, group.key))) }), _jsx("div", { className: "nre-experiment-group-panels", children: data.groups.map((group, index) => (_jsxs("details", { className: "nre-experiment-group-panel", "data-nre-experiment-group-panel": index, role: "tabpanel", open: index === 0, children: [_jsx("summary", { children: group.key }), _jsx(MetricScatter, { data: group.scatter, locale: locale }), _jsx(ExperimentList, { data: group.experiments, filter: true, locale: locale, relativeTo: group.key, attemptHref: attemptHref })] }, group.key))) })] }));
|
|
11
12
|
}
|
|
@@ -2,9 +2,9 @@ import type { ReactElement } from "react";
|
|
|
2
2
|
import type { AttemptLocator } from "../../results/locator.ts";
|
|
3
3
|
import type { ExperimentListItem } from "../types.ts";
|
|
4
4
|
import { type ReportLocale } from "../locale.ts";
|
|
5
|
-
export declare function ExperimentList({
|
|
6
|
-
|
|
7
|
-
attemptHref
|
|
5
|
+
export declare function ExperimentList({ data, attemptHref, filter, className, locale, relativeTo, }: {
|
|
6
|
+
data: readonly ExperimentListItem[];
|
|
7
|
+
attemptHref?: (locator: AttemptLocator) => string;
|
|
8
8
|
filter?: boolean;
|
|
9
9
|
className?: string;
|
|
10
10
|
locale?: ReportLocale;
|
|
@@ -1,10 +1,11 @@
|
|
|
1
1
|
import { Fragment as _Fragment, jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import {
|
|
2
|
+
import { experimentDisplayName } from "../format.js";
|
|
3
3
|
import { DEFAULT_REPORT_LOCALE, localeText } from "../locale.js";
|
|
4
|
-
import { AttemptLocatorBadge } from "./AttemptList.js";
|
|
4
|
+
import { AttemptLocatorBadge, failureSummaryText } from "./AttemptList.js";
|
|
5
5
|
import { MetricCellView } from "./cell.js";
|
|
6
6
|
import { colorClassForKey } from "./colors.js";
|
|
7
7
|
import { cx, formatDurationMs, formatUSD, verdictMark } from "./format.js";
|
|
8
|
+
const DEFAULT_ATTEMPT_HREF = (locator) => `#/attempt/${locator}`;
|
|
8
9
|
const verdictOrder = ["passed", "failed", "errored", "skipped"];
|
|
9
10
|
function passRateTone(value) {
|
|
10
11
|
if (value === null)
|
|
@@ -28,28 +29,28 @@ function formatDate(value, locale) {
|
|
|
28
29
|
}
|
|
29
30
|
function VerdictSummary({ item, locale }) {
|
|
30
31
|
const parts = verdictOrder
|
|
31
|
-
.filter((verdict) => item.
|
|
32
|
-
.map((verdict) => `${item.
|
|
32
|
+
.filter((verdict) => item.evalVerdicts[verdict] > 0)
|
|
33
|
+
.map((verdict) => `${item.evalVerdicts[verdict]} ${localeText(locale, `verdict.${verdict}`)}`);
|
|
33
34
|
return _jsx("span", { className: "nre-experiment-pill", children: parts.join(" / ") || "—" });
|
|
34
35
|
}
|
|
35
|
-
function ExperimentAttemptRow({ attempt, last, attemptHref, }) {
|
|
36
|
-
const reason =
|
|
37
|
-
return (_jsxs("li", { className: cx("nre-experiment-attempt-row", `nre-eval-${attempt.verdict}`), children: [_jsx("span", { className: "nre-attempt-branch", "aria-hidden": "true", children: last ? "└─" : "├─" }), _jsx("span", { className: "nre-eval-attempt-badges", children: _jsx(AttemptLocatorBadge, { item: attempt, attemptHref: attemptHref }) }), _jsxs("span", { className: "nre-eval-attempt-metrics", children: [formatDurationMs(attempt.durationMs), attempt.costUSD !==
|
|
36
|
+
function ExperimentAttemptRow({ attempt, last, attemptHref, locale, }) {
|
|
37
|
+
const reason = failureSummaryText(attempt, locale);
|
|
38
|
+
return (_jsxs("li", { className: cx("nre-experiment-attempt-row", `nre-eval-${attempt.verdict}`), children: [_jsx("span", { className: "nre-attempt-branch", "aria-hidden": "true", children: last ? "└─" : "├─" }), _jsx("span", { className: "nre-eval-attempt-badges", children: _jsx(AttemptLocatorBadge, { item: attempt, attemptHref: attemptHref }) }), _jsxs("span", { className: "nre-eval-attempt-metrics", children: [formatDurationMs(attempt.durationMs), attempt.costUSD !== null && _jsxs(_Fragment, { children: [" \u00B7 ", formatUSD(attempt.costUSD)] })] }), _jsx("span", { className: "nre-eval-reason", children: reason ?? "—" })] }));
|
|
38
39
|
}
|
|
39
40
|
function EvalAttempts({ row, attemptHref, locale, }) {
|
|
40
|
-
const duration = row.
|
|
41
|
-
const cost = row.
|
|
42
|
-
return (_jsxs("li", { className: "nre-experiment-eval", children: [_jsxs("div", { className: cx("nre-experiment-eval-header", `nre-eval-${row.verdict}`), children: [_jsx("span", { className: cx("nre-eval-verdict", `nre-verdict-${row.verdict}`), children: verdictMark(row.verdict) }), _jsx("span", { className: "nre-eval-id", children: row.evalId }), _jsx("span", { className: "nre-eval-attempt-count", children: localeText(locale, "overview.attemptsCount", { n: row.attempts.length }) }), _jsxs("span", { className: "nre-eval-rollup", children: [localeText(locale, "entityList.average", { value: duration }), " · ", localeText(locale, "entityList.average", { value: cost })] })] }), _jsx("ul", { className: "nre-experiment-attempts", children: row.attempts.map((attempt, index) => (_jsx(ExperimentAttemptRow, { attempt: attempt, last: index === row.attempts.length - 1, attemptHref: attemptHref }, attempt.locator))) })] }));
|
|
41
|
+
const duration = row.durationMs.value === null ? localeText(locale, "cell.missing") : formatDurationMs(row.durationMs.value);
|
|
42
|
+
const cost = row.costUSD.value === null ? localeText(locale, "cell.missing") : formatUSD(row.costUSD.value);
|
|
43
|
+
return (_jsxs("li", { className: "nre-experiment-eval", children: [_jsxs("div", { className: cx("nre-experiment-eval-header", `nre-eval-${row.verdict}`), children: [_jsx("span", { className: cx("nre-eval-verdict", `nre-verdict-${row.verdict}`), children: verdictMark(row.verdict) }), _jsx("span", { className: "nre-eval-id", children: row.evalId }), _jsx("span", { className: "nre-eval-attempt-count", children: localeText(locale, "overview.attemptsCount", { n: row.attempts.length }) }), _jsxs("span", { className: "nre-eval-rollup", children: [localeText(locale, "entityList.average", { value: duration }), " · ", localeText(locale, "entityList.average", { value: cost })] })] }), _jsx("ul", { className: "nre-experiment-attempts", children: row.attempts.map((attempt, index) => (_jsx(ExperimentAttemptRow, { attempt: attempt, last: index === row.attempts.length - 1, attemptHref: attemptHref, locale: locale }, attempt.locator))) })] }));
|
|
43
44
|
}
|
|
44
45
|
function Flags({ flags, locale }) {
|
|
45
46
|
if (!flags || Object.keys(flags).length === 0)
|
|
46
47
|
return null;
|
|
47
|
-
return (_jsxs("div", { className: "nre-experiment-flags", children: [_jsx("span", { children: localeText(locale, "experimentList.flags") }), Object.entries(flags).map(([key, value]) => (_jsxs("b", { children: [key, "=",
|
|
48
|
+
return (_jsxs("div", { className: "nre-experiment-flags", children: [_jsx("span", { children: localeText(locale, "experimentList.flags") }), Object.entries(flags).map(([key, value]) => (_jsxs("b", { children: [key, "=", typeof value === "string" ? value : JSON.stringify(value)] }, key)))] }));
|
|
48
49
|
}
|
|
49
50
|
function ExperimentRow({ item, attemptHref, locale, relativeTo, }) {
|
|
50
|
-
return (_jsxs("details", { className: "nre-experiment-entry", children: [_jsxs("summary", { className: "nre-experiment-summary", children: [_jsxs("span", { className: "nre-experiment-name", "data-sort-value": item.experimentId, children: [_jsx("b", { className: cx("nre-experiment-id", "nre-key", colorClassForKey(item.experimentId)), children: experimentDisplayName(item.experimentId, relativeTo) }), _jsxs("small", { children: [localeText(locale, "overview.evalsCount", { n: item.evals }), item.attempts > item.evals ? ` · ${localeText(locale, "overview.attemptsCount", { n: item.attempts })}` : "", ` · ${formatDate(item.lastRunAt, locale)}`] })] }), _jsx("span", { "data-sort-value": item.model ?? "", children: item.model ?? localeText(locale, "experimentList.defaultModel") }), _jsx("span", { "data-sort-value": item.agent, children: item.agent }), _jsx("span", { className: "nre-num", "data-sort-value": item.
|
|
51
|
+
return (_jsxs("details", { className: "nre-experiment-entry", children: [_jsxs("summary", { className: "nre-experiment-summary", children: [_jsxs("span", { className: "nre-experiment-name", "data-sort-value": item.experimentId, children: [_jsx("b", { className: cx("nre-experiment-id", "nre-key", colorClassForKey(item.experimentId)), children: experimentDisplayName(item.experimentId, relativeTo) }), _jsxs("small", { children: [localeText(locale, "overview.evalsCount", { n: item.evals }), item.attempts > item.evals ? ` · ${localeText(locale, "overview.attemptsCount", { n: item.attempts })}` : "", ` · ${formatDate(item.lastRunAt, locale)}`] })] }), _jsx("span", { "data-sort-value": item.model ?? "", children: item.model ?? localeText(locale, "experimentList.defaultModel") }), _jsx("span", { "data-sort-value": item.agent, children: item.agent }), _jsx("span", { className: "nre-num", "data-sort-value": item.durationMs.value ?? "", children: _jsx(MetricCellView, { cell: item.durationMs, locale: locale }) }), _jsx("span", { className: cx("nre-num", passRateTone(item.endToEndPassRate.value)), "data-sort-value": item.endToEndPassRate.value ?? "", children: _jsx(MetricCellView, { cell: item.endToEndPassRate, locale: locale }) }), _jsx("span", { className: "nre-num", "data-sort-value": item.tokens.value ?? "", children: _jsx(MetricCellView, { cell: item.tokens, locale: locale }) }), _jsx("span", { className: "nre-num", "data-sort-value": item.costUSD.value ?? "", children: _jsx(MetricCellView, { cell: item.costUSD, locale: locale }) }), _jsx("span", { "data-sort-value": item.evalVerdicts.passed, children: _jsx(VerdictSummary, { item: item, locale: locale }) })] }), _jsxs("div", { className: "nre-experiment-detail", children: [_jsx(Flags, { flags: item.flags, locale: locale }), _jsx("ul", { className: "nre-experiment-evals", children: item.evalRows.map((row) => (_jsx(EvalAttempts, { row: row, attemptHref: attemptHref, locale: locale }, row.evalId))) })] })] }));
|
|
51
52
|
}
|
|
52
|
-
export function ExperimentList({
|
|
53
|
+
export function ExperimentList({ data, attemptHref = DEFAULT_ATTEMPT_HREF, filter = false, className, locale = DEFAULT_REPORT_LOCALE, relativeTo, }) {
|
|
53
54
|
const labels = [
|
|
54
55
|
localeText(locale, "experimentList.experiment"),
|
|
55
56
|
localeText(locale, "table.model"),
|
|
@@ -57,9 +58,9 @@ export function ExperimentList({ items, attemptHref, filter = false, className,
|
|
|
57
58
|
localeText(locale, "experimentList.avgDuration"),
|
|
58
59
|
localeText(locale, "experimentList.passRate"),
|
|
59
60
|
localeText(locale, "experimentList.tokens"),
|
|
60
|
-
localeText(locale, "experimentList.
|
|
61
|
+
localeText(locale, "experimentList.cost"),
|
|
61
62
|
localeText(locale, "experimentList.result"),
|
|
62
63
|
];
|
|
63
|
-
const board = (_jsxs("div", { className: "nre-experiment-table", children: [_jsx("div", { className: "nre-experiment-head", children: labels.map((label, index) => (_jsx("button", { type: "button", "data-nre-experiment-sort": index, className: index === 4 ? "nre-sort-desc" : undefined, children: label }, label))) }),
|
|
64
|
+
const board = (_jsxs("div", { className: "nre-experiment-table", children: [_jsx("div", { className: "nre-experiment-head", children: labels.map((label, index) => (_jsx("button", { type: "button", "data-nre-experiment-sort": index, className: index === 4 ? "nre-sort-desc" : undefined, children: label }, label))) }), data.length === 0 && _jsx("p", { className: "nre-experiment-list-empty", children: localeText(locale, "attemptList.empty") }), data.map((item) => (_jsx(ExperimentRow, { item: item, attemptHref: attemptHref, locale: locale, relativeTo: relativeTo }, item.experimentId)))] }));
|
|
64
65
|
return (_jsxs("div", { className: cx("nre", "nre-experiment-list", filter && "nre-experiment-table-wrap", className), children: [filter && (_jsx("input", { className: "nre-filter", "data-nre-experiment-filter": "", type: "search", placeholder: localeText(locale, "experimentList.filterPlaceholder") })), board] }));
|
|
65
66
|
}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import { DEFAULT_REPORT_LOCALE, resolveMetricLabel } from "../locale.js";
|
|
2
|
+
import { DEFAULT_REPORT_LOCALE, resolveLocalizedText, resolveMetricLabel } from "../locale.js";
|
|
3
3
|
import { colorClassForKey, seriesClassForKey } from "./colors.js";
|
|
4
4
|
import { cx } from "./format.js";
|
|
5
5
|
const WIDTH = 640;
|
|
@@ -26,7 +26,7 @@ export function MetricBars({ data, attemptHref, className, locale = DEFAULT_REPO
|
|
|
26
26
|
const groupWidth = (PLOT.right - PLOT.left) / Math.max(1, groupKeys.length);
|
|
27
27
|
const barWidth = Math.min(36, (groupWidth - 16) / Math.max(1, seriesKeys.length));
|
|
28
28
|
const plotHeight = PLOT.bottom - PLOT.top;
|
|
29
|
-
return (_jsxs("figure", { className: cx("nre", "nre-metric-bars", className), children: [_jsxs("svg", { className: "nre-bars-svg", viewBox: `0 0 ${WIDTH} ${HEIGHT}`, role: "img", "aria-label": `${metricLabel} by ${data.
|
|
29
|
+
return (_jsxs("figure", { className: cx("nre", "nre-metric-bars", className), children: [_jsxs("svg", { className: "nre-bars-svg", viewBox: `0 0 ${WIDTH} ${HEIGHT}`, role: "img", "aria-label": `${metricLabel} by ${data.rowDimension} × ${data.columnDimension}`, children: [groupKeys.map((group, gi) => {
|
|
30
30
|
const x0 = PLOT.left + gi * groupWidth;
|
|
31
31
|
// 组内按值排序,方向随 better(缺数据的系列不画柱)
|
|
32
32
|
const entries = seriesKeys
|
|
@@ -39,10 +39,11 @@ export function MetricBars({ data, attemptHref, className, locale = DEFAULT_REPO
|
|
|
39
39
|
const h = Math.max(1, ratioOf(cell.value) * plotHeight);
|
|
40
40
|
const x = startX + si * barWidth;
|
|
41
41
|
const y = PLOT.bottom - h;
|
|
42
|
-
const
|
|
42
|
+
const display = resolveLocalizedText(cell.display, locale);
|
|
43
|
+
const title = `${group} · ${series}: ${display}(${cell.samples}/${cell.total})`;
|
|
43
44
|
const rect = (_jsx("rect", { className: cx("nre-bar", seriesClassForKey(series)), x: x + 2, y: y, width: barWidth - 4, height: h, children: _jsx("title", { children: title }) }));
|
|
44
45
|
const locator = cell.refs[0];
|
|
45
|
-
return (_jsxs("g", { children: [attemptHref && locator ? _jsx("a", { href: attemptHref(locator), children: rect }) : rect, _jsx("text", { className: "nre-bar-value", x: x + barWidth / 2, y: y - 4, textAnchor: "middle", children: cell.samples < cell.total ? `${
|
|
46
|
+
return (_jsxs("g", { children: [attemptHref && locator ? _jsx("a", { href: attemptHref(locator), children: rect }) : rect, _jsx("text", { className: "nre-bar-value", x: x + barWidth / 2, y: y - 4, textAnchor: "middle", children: cell.samples < cell.total ? `${display} ${cell.samples}/${cell.total}` : display })] }, series));
|
|
46
47
|
}), _jsx("text", { className: "nre-bars-group-label", x: x0 + groupWidth / 2, y: HEIGHT - 20, textAnchor: "middle", children: group })] }, group));
|
|
47
48
|
}), _jsx("line", { className: "nre-bars-baseline", x1: PLOT.left, y1: PLOT.bottom, x2: PLOT.right, y2: PLOT.bottom })] }), _jsxs("figcaption", { className: "nre-bars-legend", children: [_jsxs("span", { className: "nre-bars-metric", children: [metricLabel, data.metric.unit && _jsxs("span", { className: "nre-unit", children: ["(", data.metric.unit, ")"] })] }), seriesKeys.map((series) => (_jsx("span", { className: cx("nre-legend-key", "nre-key", colorClassForKey(series)), children: series }, series)))] })] }));
|
|
48
49
|
}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
|
|
2
|
-
import { DEFAULT_REPORT_LOCALE, countText, localeText, resolveMetricLabel } from "../locale.js";
|
|
2
|
+
import { DEFAULT_REPORT_LOCALE, countText, localeText, resolveLocalizedText, resolveMetricLabel } from "../locale.js";
|
|
3
3
|
import { seriesClassForKey } from "./colors.js";
|
|
4
4
|
import { cx } from "./format.js";
|
|
5
5
|
const WIDTH = 640;
|
|
@@ -22,6 +22,7 @@ export function MetricLine({ data, pointHref, className, locale = DEFAULT_REPORT
|
|
|
22
22
|
const missing = data.rows.filter((r) => r.x === null || r.y.value === null);
|
|
23
23
|
const drawableRows = data.rows.filter((r) => r.x !== null && r.y.value !== null);
|
|
24
24
|
const yLabel = resolveMetricLabel(data.y.label, locale, data.y.key);
|
|
25
|
+
const xLabel = resolveMetricLabel(data.x.label, locale, data.x.key);
|
|
25
26
|
const missingNote = missing.length > 0 ? (_jsx("p", { className: "nre-line-missing", title: missing.map((r) => r.key).join(", "), children: countText(locale, "pointsMissing", missing.length) })) : null;
|
|
26
27
|
if (drawableRows.length === 0) {
|
|
27
28
|
return (_jsxs("figure", { className: cx("nre", "nre-metric-line", className), children: [_jsx("p", { className: "nre-missing", children: localeText(locale, "cell.missing") }), missingNote] }));
|
|
@@ -34,7 +35,7 @@ export function MetricLine({ data, pointHref, className, locale = DEFAULT_REPORT
|
|
|
34
35
|
series: r.series,
|
|
35
36
|
xValue: r.x,
|
|
36
37
|
yValue: r.y.value,
|
|
37
|
-
title: `${r.key}\n${
|
|
38
|
+
title: `${r.key}\n${xLabel}: ${resolveLocalizedText(r.xDisplay, locale)}\n${yLabel}: ${resolveLocalizedText(r.y.display, locale)}(${r.y.samples}/${r.y.total})`,
|
|
38
39
|
px: xScale.scale(r.x),
|
|
39
40
|
py: yScale.scale(r.y.value),
|
|
40
41
|
}));
|
|
@@ -53,9 +54,15 @@ export function MetricLine({ data, pointHref, className, locale = DEFAULT_REPORT
|
|
|
53
54
|
list.sort((a, b) => a.xValue - b.xValue);
|
|
54
55
|
const xTicks = xScale.lo === xScale.hi ? [xScale.lo] : [xScale.lo, xScale.hi];
|
|
55
56
|
const yTicks = yScale.lo === yScale.hi ? [yScale.lo] : [yScale.lo, yScale.hi];
|
|
56
|
-
const xDisplayFor = (value) =>
|
|
57
|
-
|
|
58
|
-
|
|
57
|
+
const xDisplayFor = (value) => {
|
|
58
|
+
const row = drawableRows.find((r) => r.x === value);
|
|
59
|
+
return row === undefined ? String(value) : resolveLocalizedText(row.xDisplay, locale);
|
|
60
|
+
};
|
|
61
|
+
const yDisplayFor = (value) => {
|
|
62
|
+
const row = drawableRows.find((r) => r.y.value === value);
|
|
63
|
+
return row === undefined ? String(value) : resolveLocalizedText(row.y.display, locale);
|
|
64
|
+
};
|
|
65
|
+
return (_jsxs("figure", { className: cx("nre", "nre-metric-line", className), children: [_jsxs("svg", { className: "nre-line-svg", viewBox: `0 0 ${WIDTH} ${HEIGHT}`, role: "img", "aria-label": `${yLabel} by ${xLabel}`, children: [_jsx("rect", { className: "nre-line-plot", x: PLOT.left, y: PLOT.top, width: PLOT.right - PLOT.left, height: PLOT.bottom - PLOT.top, fill: "none" }), _jsxs("text", { className: "nre-line-xlabel", x: (PLOT.left + PLOT.right) / 2, y: HEIGHT - 8, textAnchor: "middle", children: [xLabel, data.x.unit ? `(${data.x.unit})` : ""] }), _jsxs("text", { className: "nre-line-ylabel", x: 16, y: (PLOT.top + PLOT.bottom) / 2, textAnchor: "middle", transform: `rotate(-90 16 ${(PLOT.top + PLOT.bottom) / 2})`, children: [yLabel, data.y.unit ? `(${data.y.unit})` : ""] }), xTicks.map((v) => (_jsx("text", { className: "nre-line-tick", x: xScale.scale(v), y: PLOT.bottom + 16, textAnchor: "middle", children: xDisplayFor(v) }, `x${v}`))), yTicks.map((v) => (_jsx("text", { className: "nre-line-tick", x: PLOT.left - 6, y: yScale.scale(v) + 4, textAnchor: "end", children: yDisplayFor(v) }, `y${v}`))), seriesOrder.map((series) => {
|
|
59
66
|
const list = bySeries.get(series);
|
|
60
67
|
const seriesClass = series === "" ? "nre-series-none" : seriesClassForKey(series);
|
|
61
68
|
const labelAt = list.reduce((a, b) => (b.px > a.px ? b : a));
|
|
@@ -16,7 +16,7 @@ export function MetricMatrix({ data, attemptHref, className, locale = DEFAULT_RE
|
|
|
16
16
|
columnKeys.push(entry.column);
|
|
17
17
|
byPosition.set(`${entry.row}\u0000${entry.column}`, entry.cell);
|
|
18
18
|
}
|
|
19
|
-
return (_jsxs("table", { className: cx("nre", "nre-metric-matrix", className), children: [_jsxs("caption", { className: "nre-matrix-caption", children: [resolveMetricLabel(data.metric.label, locale, data.metric.key), data.metric.unit && _jsxs("span", { className: "nre-unit", children: ["(", data.metric.unit, ")"] }), _jsxs("span", { className: "nre-matrix-axes", children: [data.
|
|
19
|
+
return (_jsxs("table", { className: cx("nre", "nre-metric-matrix", className), children: [_jsxs("caption", { className: "nre-matrix-caption", children: [resolveMetricLabel(data.metric.label, locale, data.metric.key), data.metric.unit && _jsxs("span", { className: "nre-unit", children: ["(", data.metric.unit, ")"] }), _jsxs("span", { className: "nre-matrix-axes", children: [data.rowDimension, " \u00D7 ", data.columnDimension] })] }), _jsx("thead", { children: _jsxs("tr", { children: [_jsx("th", { scope: "col", className: "nre-dimension", children: data.rowDimension }), columnKeys.map((column) => (_jsx("th", { scope: "col", className: cx("nre-col-key", "nre-key", colorClassForKey(column)), children: column }, column)))] }) }), _jsx("tbody", { children: rowKeys.map((row) => (_jsxs("tr", { children: [_jsx("th", { scope: "row", className: "nre-row-key", children: row }), columnKeys.map((column) => {
|
|
20
20
|
const cell = byPosition.get(`${row}\u0000${column}`);
|
|
21
21
|
return (_jsx("td", { className: cx("nre-td", !cell && "nre-td-empty"), children: cell ? _jsx(MetricCellView, { cell: cell, attemptHref: attemptHref, locale: locale }) : null }, column));
|
|
22
22
|
})] }, row))) })] }));
|