niceeval 0.6.0 → 0.6.1

package/docs-site/zh/reference/events.mdx

@@ -1,7 +1,7 @@
 ---
 title: "标准事件流参考"
 sidebarTitle: "事件流"
-description: "StreamEvent 的九种事件：每种的字段、什么时候吐、哪些断言消费它。adapter 的核心工作就是产出这条流。"
+description: "StreamEvent 的十种事件：每种的字段、什么时候吐、哪些断言消费它。adapter 的核心工作就是产出这条流。"
 ---
 
 adapter 的 `send` 返回一个 `Turn`，其中 `events: StreamEvent[]` 是**断言的唯一数据源**：`t.calledTool()`、`t.reply`、`toolOrder`、`noFailedActions`……全部从这条流上读。把你的 agent"这一轮做了什么"翻成这条流，整套断言就都能用。
@@ -79,7 +79,7 @@ token 用量或 OTel span 反推得到)。存在时优先于按价格表(`define
 
 ## StreamEvent 变体一览
 
-`StreamEvent` 的九种变体，逐字段列出（消费它们的断言 / 使用细节见下面的「事件总表」和「逐事件说明」）：
+`StreamEvent` 的十种变体，逐字段列出（消费它们的断言 / 使用细节见下面的「事件总表」和「逐事件说明」）：
 
 {/* GENERATED:BEGIN stream-events */}
 

package/docs-site/zh/reference/expect.mdx

@@ -200,13 +200,18 @@ atLeast(threshold: number): ValueAssertion;
 ## 自定义 matcher
 
 ```ts
-import { makeAssertion } from "niceeval/expect";
+import { makeAssertion, type Assertion } from "niceeval/expect";
 
-const validEmail = makeAssertion("valid email", (value) => {
-  return typeof value === "string" && value.includes("@") ? 1 : 0;
-});
+function validEmail(): Assertion {
+  return makeAssertion({
+    name: "valid email",
+    score(value) {
+      return typeof value === "string" && value.includes("@") ? 1 : 0;
+    },
+  });
+}
 
 t.check(t.reply, validEmail());
 ```
 
-自定义 matcher 应返回 0 到 1 之间的分数，并给出清晰名称。
+`makeAssertion` 接受单个 spec 对象（`name`、可选 `severity`/`threshold`、`score`），直接返回一个 `Assertion`；按惯例用一个同名函数包一层再调用，而不是把 `makeAssertion` 本身当工厂调用。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "niceeval",
-  "version": "0.6.0",
+  "version": "0.6.1",
   "description": "Agent-native eval tool — eval agents, services, functions, and coding-agent fixtures",
   "type": "module",
   "license": "MIT",

package/src/agents/types.ts

@@ -221,7 +221,11 @@ export interface AgentContext {
    * 具体取值)是两个维度——这里只是「跑的是哪个实验」的稳定标识,不携带条件内容。
    */
   readonly experimentId?: string;
-  /** 仅沙箱型 agent 有(运行器按 --sandbox 备好)。 */
+  /**
+   * 所有 agent 都有:沙箱型是运行器按项目/experiment 配置备好的真实沙箱句柄,remote 型是
+   * `createRemoteSandbox()` 产出的 stub(仅含 `workdir`/`sandboxId`/`otlpHost`/`stop` 等
+   * 元信息,其余方法调用即抛错)。
+   */
   readonly sandbox: Sandbox;
   readonly session: AgentSession;
   /**

package/src/i18n/en.ts

@@ -54,8 +54,7 @@ export const en = {
     "  niceeval exp [group|experiment] [eval-id-prefix…]   run experiments\n" +
     "  niceeval show [eval-id-prefix… | @<locator>]        read results in the terminal\n" +
     "      bare: current verdicts per experiment (composed across runs), each row\n" +
-    "        with a compact attempt index (locator + evidence capability letters\n" +
-    "        [E,X,⏱,D] + failure reason)\n" +
+    "        with a compact attempt index (locator + failure reason)\n" +
     "      a single eval id: attempts + assertion details\n" +
     "      @<locator>  exactly one attempt: no flag -> compact overview;\n" +
     "        with a flag -> that evidence slice\n" +

package/src/i18n/zh-CN.ts

@@ -52,7 +52,7 @@ export const zhCN = {
     "  niceeval exp [组|实验] [eval-id 前缀…]   跑实验\n" +
     "  niceeval show [eval-id 前缀… | @<locator>]   终端读结果\n" +
     "      裸跑:每个 experiment 的现刻判定(跨 run 合成),每行带紧凑 attempt 索引\n" +
-    "        (locator + 证据 capability 字母 [E,X,⏱,D] + 失败原因)\n" +
+    "        (locator + 失败原因)\n" +
     "      单个 eval id:attempt 与断言明细\n" +
     "      @<locator>  精确一个 attempt:无 flag → 紧凑全景;带 flag → 对应证据切面\n" +
     "      --eval        该 attempt 运行时保存的 Eval 源码,断言标回源码行\n" +

package/src/report/built-in-user-parity.test.tsx

@@ -1,15 +1,15 @@
-// Phase 4 验收:证明内置报告 CostPassRateComparison 与一份普通用户 --report 文件走的是
+// Phase 4 验收:证明内置报告 ExperimentComparison 与一份普通用户 --report 文件走的是
 // 完全同一套机制 —— 不是长得像,而是同一条 build → resolveReportTree → validate → render 管线。
 //
 // 证据链(plan/built-in-reports-user-parity.md「Phase 4」):
-//   1. test/fixtures/report/cost-pass-rate-comparison-public-copy.tsx 只从公开 barrel import,
+//   1. test/fixtures/report/experiment-comparison-public-copy.tsx 只从公开 barrel import,
 //      与 built-in 逐节点同构(见文件头)。
 //   2. 同一 ReportContext 下分别 build + resolveReportTree,深比较 resolved 树:组件类型 / 顺序 /
 //      resolved data / props keys 全部结构化相等(不靠 definition 引用相等,两份是独立文件)。
 //   3. 两份分别过 renderReportToText / renderReportToStaticHtml,比较事实(散点键 / 表格行 /
 //      attempt 深链),不做 HTML 字面 snapshot。
 //   4. 覆盖 plan 的 9 个必测场景。
-//   5. 架构不变量:renderer 源码零 "CostPassRateComparison" / "built-ins" 分支。
+//   5. 架构不变量:renderer 源码零 "ExperimentComparison" / "built-ins" 分支。
 //   6. show / view 只在「无 --report」分支选内置报告,--report 只换 definition,不碰
 //      Selection / resolve / validate / render。
 
@@ -22,8 +22,8 @@ import type { AttemptHandle, Results, Selection, Snapshot } from "../results/ind
 import type { ExperimentListItem, ScatterData } from "./types.ts";
 import { encodeAttemptLocator } from "../results/locator.ts";
 import { Col, ExperimentList, defineReport } from "./index.ts";
-import { CostPassRateComparison } from "./built-ins/index.ts";
-import publicCopy from "../../test/fixtures/report/cost-pass-rate-comparison-public-copy.tsx";
+import { ExperimentComparison } from "./built-ins/index.ts";
+import publicCopy from "../../test/fixtures/report/experiment-comparison-public-copy.tsx";
 import { renderReportToText } from "./report.ts";
 import { renderReportToStaticHtml } from "./web.ts";
 import {
@@ -278,15 +278,15 @@ function expectedLocatorFor(attempt: AttemptHandle): string {
 
 describe("resolved 树结构化等价(built-in ≡ 包外 public-copy)", () => {
   it("两份是独立 ReportDefinition,不是同一引用", () => {
-    expect(CostPassRateComparison).not.toBe(publicCopy);
-    expect(typeof CostPassRateComparison.build).toBe("function");
+    expect(ExperimentComparison).not.toBe(publicCopy);
+    expect(typeof ExperimentComparison.build).toBe("function");
     expect(typeof publicCopy.build).toBe("function");
   });
 
   for (const { name, ctx } of SCENARIOS) {
     it(`同一 ReportContext 下 resolved 树逐节点相等 — ${name}`, async () => {
       const c = ctx();
-      const builtIn = await resolvedTreeOf(CostPassRateComparison, c);
+      const builtIn = await resolvedTreeOf(ExperimentComparison, c);
       const user = await resolvedTreeOf(publicCopy, ctx()); // 各用一份等价 ctx,证明不依赖共享状态
 
       // 组件类型与顺序:<Col> 包 [MetricScatter, ExperimentList],别无它物。
@@ -308,7 +308,7 @@ describe("resolved 树结构化等价(built-in ≡ 包外 public-copy)", () => {
 
 describe("必测场景:resolved data 事实", () => {
   it("场景 1+2:多 experiment/agent,成本与通过率都有值;缺成本的点如实 x=null,实验列表仍出项", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, richContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, richContext());
     const scatter = scatterOf(tree)!;
     const experiments = experimentListOf(tree)!;
 
@@ -342,7 +342,7 @@ describe("必测场景:resolved data 事实", () => {
   });
 
   it("场景 3:恰好一个可画点(另一实验缺成本)", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, oneDrawableContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, oneDrawableContext());
     const scatter = scatterOf(tree)!;
     const drawable = scatter.rows.filter((r) => r.x.value !== null && r.y.value !== null);
     expect(drawable).toHaveLength(1);
@@ -350,20 +350,20 @@ describe("必测场景:resolved data 事实", () => {
   });
 
   it("场景 4:空 Selection —— 散点无 rows、实验列表为空数组", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, emptyContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, emptyContext());
     expect(scatterOf(tree)!.rows).toEqual([]);
     expect(experimentListOf(tree)).toEqual([]);
   });
 
   it("场景 4 变体:全 skipped —— 每个点 x/y 都不可测,0 可画点", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, allSkippedContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, allSkippedContext());
     const scatter = scatterOf(tree)!;
     expect(scatter.rows.every((r) => r.x.value === null && r.y.value === null)).toBe(true);
     expect(scatter.rows.filter((r) => r.x.value !== null && r.y.value !== null)).toHaveLength(0);
   });
 
   it("场景 5:failed/errored/skipped 混合 —— ExperimentList 展开区结果摘要正确(eval 级折叠计票)", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, richContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, richContext());
     const codex = experimentListOf(tree)!.find((e) => e.experimentId === "compare/codex-mid")!;
     // algebra/x errored、algebra/y skipped、algebra/z passed → eval 级折叠 {passed:1, errored:1, skipped:1}
     expect(codex.verdicts).toEqual({ passed: 1, failed: 0, errored: 1, skipped: 1 });
@@ -374,7 +374,7 @@ describe("必测场景:resolved data 事实", () => {
   });
 
   it("场景 6:同 experiment 多 eval、多 attempts —— 展开层与 attempt locator 正确", async () => {
-    const tree = await resolvedTreeOf(CostPassRateComparison, richContext());
+    const tree = await resolvedTreeOf(ExperimentComparison, richContext());
     const bub = experimentListOf(tree)!.find((e) => e.experimentId === "compare/bub-low")!;
     expect(bub.evalRows.map((e) => e.evalId).sort()).toEqual(["algebra/x", "algebra/y"]);
     const multi = bub.evalRows.find((e) => e.evalId === "algebra/x")!;
@@ -394,12 +394,12 @@ describe("必测场景:resolved data 事实", () => {
 describe("built-in 与 public-copy 渲染出的事实相同", () => {
   for (const { name, ctx } of SCENARIOS) {
     it(`text / web 事实一致 — ${name}`, async () => {
-      const builtinText = await renderReportToText(CostPassRateComparison, ctx(), { width: 100 });
+      const builtinText = await renderReportToText(ExperimentComparison, ctx(), { width: 100 });
       const userText = await renderReportToText(publicCopy, ctx(), { width: 100 });
       // 同一机制、同一 options:text 输出逐字相等(最强的事实级证据)。
       expect(userText).toBe(builtinText);
 
-      const builtinHtml = await renderReportToStaticHtml(CostPassRateComparison, ctx());
+      const builtinHtml = await renderReportToStaticHtml(ExperimentComparison, ctx());
       const userHtml = await renderReportToStaticHtml(publicCopy, ctx());
       // 事实比较(不做整段 HTML snapshot):散点点键、attempt 深链集合相同。
       expect(scatterKeysOf(userHtml)).toEqual(scatterKeysOf(builtinHtml));
@@ -411,7 +411,7 @@ describe("built-in 与 public-copy 渲染出的事实相同", () => {
     const ctx = richContext();
     // renderReportToStaticHtml 默认 attemptHref = view 的 `#/attempt/@<locator>`(不透明单段路由,
     // 与 view/data.ts renderReportSlot 不传 attemptHref 时同一条默认路径)。
-    const builtinHtml = await renderReportToStaticHtml(CostPassRateComparison, ctx);
+    const builtinHtml = await renderReportToStaticHtml(ExperimentComparison, ctx);
     const userHtml = await renderReportToStaticHtml(publicCopy, richContext());
     const builtinHrefs = hrefsOf(builtinHtml);
     const userHrefs = hrefsOf(userHtml);
@@ -437,31 +437,28 @@ describe("built-in 与 public-copy 渲染出的事实相同", () => {
 
 describe("散点空态 / 单点 / 多点行为", () => {
   it("0 可画点(空 Selection):text 说 No data、web 走 nre-scatter-empty nre-missing", async () => {
-    const text = await renderReportToText(CostPassRateComparison, emptyContext(), { width: 100 });
+    const text = await renderReportToText(ExperimentComparison, emptyContext(), { width: 100 });
     expect(text).toContain("No data to plot");
     expect(text).not.toContain("better → upper right");
-    const html = await renderReportToStaticHtml(CostPassRateComparison, emptyContext());
+    const html = await renderReportToStaticHtml(ExperimentComparison, emptyContext());
     expect(html).toContain("nre-scatter-empty");
     expect(html).toContain("nre-missing");
   });
 
-  it("恰好 1 可画点:text/web 都说「至少 2 个实验」而不是画孤点", async () => {
-    const text = await renderReportToText(CostPassRateComparison, oneDrawableContext(), { width: 100 });
-    expect(text).toContain("At least 2 experiments needed to compare");
-    expect(text).not.toContain("better → upper right");
-    const html = await renderReportToStaticHtml(CostPassRateComparison, oneDrawableContext());
-    expect(html).toContain("nre-scatter-empty");
-    // 单点不是「缺数据」空态:散点空态本身不带 nre-missing 变体类(与 0 可画点的
-    // "nre-scatter-empty nre-missing" 区分开)。ExperimentList 展开区自己的缺成本格子
-    // 也合法渲染 nre-missing,不属于这条散点专属的断言范围。
+  it("恰好 1 可画点:text/web 都正常画单点", async () => {
+    const text = await renderReportToText(ExperimentComparison, oneDrawableContext(), { width: 100 });
+    expect(text).toContain("better → upper right");
+    const html = await renderReportToStaticHtml(ExperimentComparison, oneDrawableContext());
+    expect(html).toContain("nre-scatter-point");
+    expect(html).not.toContain("nre-scatter-empty");
     expect(html).not.toContain("nre-scatter-empty nre-missing");
-    expect(html).not.toContain("<svg");
+    expect(html).toContain("<svg");
   });
 
   it("≥2 可画点:正常绘图(text 有 better → upper right,web 有 svg)", async () => {
-    const text = await renderReportToText(CostPassRateComparison, richContext(), { width: 100 });
+    const text = await renderReportToText(ExperimentComparison, richContext(), { width: 100 });
     expect(text).toContain("better → upper right");
-    const html = await renderReportToStaticHtml(CostPassRateComparison, richContext());
+    const html = await renderReportToStaticHtml(ExperimentComparison, richContext());
     expect(html).toContain("<svg");
     expect(html).not.toContain("nre-scatter-empty");
   });
@@ -472,7 +469,7 @@ describe("散点空态 / 单点 / 多点行为", () => {
 describe("场景 7:en / zh-CN 数据 resolve 一次、chrome 分别本地化", () => {
   it("resolve 一次的同一棵树,en / zh-CN 各渲染一遍:data 事实相同,只有 chrome 文案不同", async () => {
     // 关键:build + resolveReportTree 只做一次,拿到唯一的 resolved 树;两个 locale 复用它。
-    const node = (await CostPassRateComparison.build(richContext())) as ReportNode;
+    const node = (await ExperimentComparison.build(richContext())) as ReportNode;
     const resolved = await resolveReportTree(node);
     validateReportTree(resolved);
 
@@ -518,7 +515,7 @@ describe("场景 7:en / zh-CN 数据 resolve 一次、chrome 分别本地化", (
     // 注意:spy 挂在 ExperimentList.data 本身,不是 compute.experimentListData —— components.tsx
     // 用 `Object.assign(component, { data: experimentListData })` 把函数值复制成一个普通属性,
     // 那份拷贝在 components.tsx 求值时(模块加载期)就定死了,晚于此的 vi.spyOn(compute, …) патch
-    // 不到已经复制走的引用。直接调用方(built-ins/cost-pass-rate-comparison.tsx)读的是
+    // 不到已经复制走的引用。直接调用方(built-ins/experiment-comparison.tsx)读的是
     // `ExperimentList.data` 这个属性,所以要在这个属性上打桩。MetricScatter 不受影响:它的
     // resolve() 住在 components.tsx 里,调用的是 scatterData 这个 import 绑定本身(ESM 实时绑定),
     // vi.spyOn(compute, "scatterData") 打得中。
@@ -526,7 +523,7 @@ describe("场景 7:en / zh-CN 数据 resolve 一次、chrome 分别本地化", (
     try {
       // ExperimentList 没有 resolve 面,.data() 直接在 build() 里 await——计算发生在 build,
       // 不是 resolveReportTree;MetricScatter 仍是 selection-form,resolve 阶段才调它的 .data()。
-      const node = (await CostPassRateComparison.build(richContext())) as ReportNode;
+      const node = (await ExperimentComparison.build(richContext())) as ReportNode;
       expect(experimentListSpy).toHaveBeenCalledTimes(1);
       const resolved = await resolveReportTree(node);
       expect(scatterSpy).toHaveBeenCalledTimes(1);
@@ -603,38 +600,38 @@ describe("场景 9:最小用户报告只放 <ExperimentList> —— 不计算散
 describe("架构不变量:renderer 不认识 built-in", () => {
   const RENDERER_FILES = ["./report.ts", "./web.ts", "./tree.ts"];
   for (const rel of RENDERER_FILES) {
-    it(`${rel} 源码里没有 "CostPassRateComparison" / "built-ins" 分支`, () => {
+    it(`${rel} 源码里没有 "ExperimentComparison" / "built-ins" 分支`, () => {
       const src = readFileSync(new URL(rel, import.meta.url), "utf8");
-      expect(src).not.toContain("CostPassRateComparison");
+      expect(src).not.toContain("ExperimentComparison");
       expect(src).not.toContain("built-ins");
     });
   }
 });
 
-// ═════════════════════════ 8. 宿主选择:show 索引与可替换报告槽分开 ═════════════════════════
+// ═════════════════════════ 8. 宿主选择:show / view 共用默认 definition ═════════════════════════
 
-describe("宿主选择:裸 show 是诊断索引,--report 与 view 走报告 definition", () => {
-  it("show / view 源码:裸 show 选择专用索引,--report 才装载 definition", () => {
+describe("宿主选择:裸 show / view 共用 ExperimentComparison,--report 替换 definition", () => {
+  it("show / view 源码:Selection 先合成,随后选择内置或用户 definition", () => {
     const showSrc = readFileSync(new URL("../show/index.ts", import.meta.url), "utf8");
-    expect(showSrc).toMatch(/flags\.report === undefined[\s\S]*showIndexText\(selection, io\.width\)/);
-    expect(showSrc).toMatch(/const definition = await loadReportFile\(cwd, flags\.report\)/);
-    expect(showSrc).not.toMatch(/import \{ CostPassRateComparison \}/);
+    expect(showSrc).toMatch(/import \{ ExperimentComparison \} from "\.\.\/\.\.\/dist\/report\/built-ins\/index\.js"/);
+    expect(showSrc).toMatch(/flags\.report === undefined \? ExperimentComparison : await loadReportFile\(cwd, flags\.report\)/);
+    expect(showSrc).not.toContain("showIndexText(selection");
     // Selection 在报告槽之前、与 report 无关地由 selectCurrentResults 合成。
     expect(showSrc).toMatch(/selectCurrentResults\(results, \{ experiment: flags\.experiment, patterns \}\)/);
 
     const viewSrc = readFileSync(new URL("../view/data.ts", import.meta.url), "utf8");
-    // view 的报告槽:report 有值走 loadReportFile,否则 CostPassRateComparison;Selection 同样先于报告槽合成。
-    expect(viewSrc).toMatch(/report\s*\?[\s\S]*loadReportFile[\s\S]*:\s*\(await import\([^)]*built-ins[^)]*\)\)\.CostPassRateComparison/);
+    // view 的报告槽:report 有值走 loadReportFile,否则 ExperimentComparison;Selection 同样先于报告槽合成。
+    expect(viewSrc).toMatch(/report\s*\?[\s\S]*loadReportFile[\s\S]*:\s*\(await import\([^)]*built-ins[^)]*\)\)\.ExperimentComparison/);
     expect(viewSrc).toMatch(/selectCurrentResults\(results, \{ experiment: opts\.experiment, patterns \}\)/);
   });
 
-  it("CostPassRateComparison 与 public-copy 作为显式报告时事实完全相同", async () => {
+  it("ExperimentComparison 与 public-copy 作为显式报告时事实完全相同", async () => {
     for (const { ctx } of SCENARIOS) {
-      const bareText = await renderReportToText(CostPassRateComparison, ctx(), { width: 100 });
+      const bareText = await renderReportToText(ExperimentComparison, ctx(), { width: 100 });
       const reportText = await renderReportToText(publicCopy, ctx(), { width: 100 });
       expect(reportText).toBe(bareText);
 
-      const bareHtml = await renderReportToStaticHtml(CostPassRateComparison, ctx());
+      const bareHtml = await renderReportToStaticHtml(ExperimentComparison, ctx());
       const reportHtml = await renderReportToStaticHtml(publicCopy, ctx());
       expect(scatterKeysOf(reportHtml)).toEqual(scatterKeysOf(bareHtml));
       expect(hrefsOf(reportHtml)).toEqual(hrefsOf(bareHtml));

package/src/report/built-ins/experiment-comparison.tsx

@@ -0,0 +1,19 @@
+// ExperimentComparison:niceeval view 裸跑时的内置 experiment 比较报告。
+//
+// 名字描述整份报告回答的问题,不绑定其中某一张图。报告由成本 × 通过率散点和固定列
+// ExperimentList 组成；它是一份普通 ReportDefinition,没有 renderer 私有通道。
+
+import { Col } from "../primitives.tsx";
+import { ExperimentList, MetricScatter } from "../components.tsx";
+import { costUSD, passRate } from "../metrics.ts";
+import { defineReport } from "../report.ts";
+
+export const ExperimentComparison = defineReport(async ({ selection }) => {
+  const experiments = await ExperimentList.data(selection);
+  return (
+    <Col>
+      <MetricScatter selection={selection} points="experiment" series="agent" x={costUSD} y={passRate} />
+      <ExperimentList items={experiments} filter />
+    </Col>
+  );
+});

package/src/report/built-ins/index.ts

@@ -1,2 +1,2 @@
 // 内置报告目录:一个文件一个按展示内容命名的普通 ReportDefinition,显式导出值,不建字符串 registry。
-export { CostPassRateComparison } from "./cost-pass-rate-comparison.tsx";
+export { ExperimentComparison } from "./experiment-comparison.tsx";

package/src/report/components.tsx

@@ -90,6 +90,8 @@ export interface GroupSummaryProps {
  */
 export interface ExperimentListProps {
   items: ExperimentListItem[];
+  /** web 面在比较表前显示实验过滤框；text 面忽略。 */
+  filter?: boolean;
   /** chrome 文案 locale;省略时随宿主上下文(宿主外默认 "en")。 */
   locale?: ReportLocale;
   className?: string;
@@ -254,7 +256,7 @@ EvalList.displayName = "EvalList";
 
 /**
  * Attempt 列表:实体列表的叶子层,每项一个 Attempt,固定展示判定、断言、error、Judge 评语
- * (assertions 的 detail/evidence)与证据引用(locator + 证据能力标记)。它不预设只看失败;
+ * (assertions 的 detail/evidence)与证据引用(locator)。它不预设只看失败;
  * 报告作者过滤 `AttemptListItem[]`、用 `.slice()` 限量,`total` 让渲染面如实报告剩余数量。
  */
 export const AttemptList: ReportComponent<AttemptListProps> & { data: typeof attemptListData } = Object.assign(

package/src/report/compute.ts

@@ -11,7 +11,6 @@
 // - core 中立:只认 Metric / Dimension 接口,不出现具体 agent 名的分支。
 
 import type {
-  AttemptEvidenceCapabilities,
   AttemptListItem,
   AttemptLocator,
   DeltaData,
@@ -248,34 +247,6 @@ export async function tableData<const M extends readonly Metric[]>(
 // (docs/feature/reports/library.md「实体列表」)。AttemptListItem 是三者共用的叶子形状——
 // ExperimentList / EvalList 的下钻数组直接复用它,不各自精简一份。
 
-/**
- * 一个 attempt 的证据能力标记,只读已有的瘦身摘要位(`hasSources` / `hasEvents` / `hasTrace`)
- * 加一次 `diff()` 懒加载——不调用完整的 `loadAttemptEvidence`(那还会额外装配 Eval 源码标注
- * 与 ExecutionTree,这里只要四个布尔位)。三位定义与 `AttemptEvidence.capabilities` 完全一致:
- *
- * - `eval`:`hasSources` 为真。真正的门槛是 `AnnotatedEvalSource` 非空且有内容(源码行或断言
- *   非空),但 sources 存在时 `buildAnnotatedEvalSource` 恒产出至少一行源码,断言又永远来自
- *   `result.assertions`(不需要 IO)——`hasSources` 是这个条件的精确瘦身版,不是近似。
- * - `execution`:`hasEvents` 为真,与 `loadAttemptEvidence` 的 `events !== null && events.length > 0`
- *   同一个判定,写入面按同一条规则算的 `hasEvents` 就是它的瘦身镜像。
- * - `timing`:`hasEvents && hasTrace` 为真——`execution` 成立且这次运行接入过 OTel
- *   (`ExecutionTree.timingAvailable` = `spans.length > 0`,`hasTrace` 正是这个判定的瘦身镜像)。
- * - `diff`:唯一需要 IO 的一位,`attempt.diff()` 非空且至少改动/删除了一个文件——落盘没有
- *   与 `hasEvents` 对应的瘦身摘要位,只能懒加载判定;成本是每个 item 一次 `diff()` 读取,
- *   不是完整证据装配的四路 Promise.all。
- */
-async function attemptCapabilities(attempt: AttemptHandle): Promise<AttemptEvidenceCapabilities> {
-  const result = attempt.result;
-  const diff = await attempt.diff();
-  const diffCapable = diff !== null && (Object.keys(diff.generatedFiles).length > 0 || diff.deletedFiles.length > 0);
-  return {
-    eval: result.hasSources === true,
-    execution: result.hasEvents === true,
-    timing: result.hasEvents === true && result.hasTrace === true,
-    diff: diffCapable,
-  };
-}
-
 /** 自由文本(error / 断言 detail / evidence)的发布消毒钩子;身份字段(name/severity/loc)不经它。 */
 function redactAssertions(assertions: AssertionResult[], redact: (text: string) => string): AssertionResult[] {
   if (assertions.length === 0) return assertions;
@@ -287,7 +258,7 @@ function redactAssertions(assertions: AssertionResult[], redact: (text: string)
 }
 
 /** AttemptList / ExperimentList / EvalList 共用的叶子构造:一个 Item → 一个 AttemptListItem。 */
-async function attemptListItemOf(item: Item, redact: (text: string) => string): Promise<AttemptListItem> {
+function attemptListItemOf(item: Item, redact: (text: string) => string): AttemptListItem {
   const result = item.attempt.result;
   const cost = attemptCostUSD(result);
   return {
@@ -301,7 +272,6 @@ async function attemptListItemOf(item: Item, redact: (text: string) => string):
     durationMs: result.durationMs,
     ...(cost !== null ? { costUSD: cost } : {}),
     locator: locatorOf(item),
-    capabilities: await attemptCapabilities(item.attempt),
   };
 }
 
@@ -320,7 +290,7 @@ export async function attemptListData(
   const { snapshots } = resolveInput(input);
   const redact = opts?.redact ?? identityRedact;
   const items = collectItems(snapshots);
-  return Promise.all(items.map((item) => attemptListItemOf(item, redact)));
+  return items.map((item) => attemptListItemOf(item, redact));
 }
 
 /** `EvalList.data(selection)`:每个 `experimentId + evalId` 一项,按 evalId 再按 experimentId 升序。 */
@@ -339,7 +309,7 @@ export async function evalListData(input: SnapshotsInput): Promise<EvalListItem[
     const sorted = [...group].sort((a, b) => a.attempt.result.attempt - b.attempt.result.attempt);
     const verdict = foldEvalVerdict(sorted.map((item) => item.attempt.result));
     const representative = sorted.find((item) => item.attempt.result.verdict === verdict) ?? sorted[0]!;
-    const attempts = await Promise.all(sorted.map((item) => attemptListItemOf(item, identityRedact)));
+    const attempts = sorted.map((item) => attemptListItemOf(item, identityRedact));
     out.push({
       evalId: evalIdOf(sorted[0]!),
       experimentId: experimentIdOf(sorted[0]!),
@@ -370,7 +340,7 @@ export async function experimentListData(input: SnapshotsInput): Promise<Experim
       const sorted = [...evalItems].sort((a, b) => a.attempt.result.attempt - b.attempt.result.attempt);
       const verdict = foldEvalVerdict(sorted.map((item) => item.attempt.result));
       const representative = sorted.find((item) => item.attempt.result.verdict === verdict) ?? sorted[0]!;
-      const attempts = await Promise.all(sorted.map((item) => attemptListItemOf(item, identityRedact)));
+      const attempts = sorted.map((item) => attemptListItemOf(item, identityRedact));
       evalRows.push({
         evalId,
         verdict,
@@ -400,7 +370,14 @@ export async function experimentListData(input: SnapshotsInput): Promise<Experim
       evalRows,
     });
   }
-  out.sort((a, b) => a.experimentId.localeCompare(b.experimentId));
+  // ExperimentList 是默认实验比较表:初始态按成功率从高到低,缺数据沉底;
+  // 同分时按 experiment id 稳定排序。web 增强可临时重排,text 面沿用同一基准顺序。
+  out.sort((a, b) => {
+    if (a.passRate.value === null && b.passRate.value === null) return a.experimentId.localeCompare(b.experimentId);
+    if (a.passRate.value === null) return 1;
+    if (b.passRate.value === null) return -1;
+    return b.passRate.value - a.passRate.value || a.experimentId.localeCompare(b.experimentId);
+  });
   return out;
 }
 
@@ -772,4 +749,3 @@ function deltaDisplay(metric: Metric, delta: number | null): string {
   const text = displayValue(metric, delta); // 负号由格式化自带
   return delta > 0 ? `+${text}` : text;
 }
-