niceeval 0.11.4-canary.20 → 0.12.0

package/INDEX.md

@@ -27,6 +27,7 @@
 - `docs-site/zh/tutorials/criteria-files.mdx` — 隐藏测试判分：让判据文件进缓存 fingerprint:用 loadText 在模块顶层读入隐藏测试、跑测脚本等判据文件，文件内容进入 fingerprint：改一字节自动重跑，不靠人工 --rerun。
 - `docs-site/zh/tutorials/custom-reports.mdx` — 编写自定义报告:用普通 TypeScript 函数计算结果，并同时生成终端与网页报告。
 - `docs-site/zh/tutorials/dataset-fanout.mdx` — 数据驱动测试（dataset fan-out）：用多份数据运行同一套评估用例:从 .eval.ts 文件导出数组或 keyed record，将一套评估逻辑展开为多个 case。用 loadYaml 或 loadJson 读取外部测试集，并获得稳定 ID。
+- `docs-site/zh/tutorials/deploy-report-site.mdx` — 把报告站发布上线:把 .niceeval 直接提交进仓库、配一份排除运行时状态的 .gitignore,再用不装整仓依赖的 CI 脚本和 Vercel CLI,在打 tag 时把报告站发布上线。
 - `docs-site/zh/tutorials/evaluation-kinds.mdx` — 题型、断言、Judge 和成本限制:值断言、作用域断言、LLM-as-judge、严重度怎么折成一个判定、计分制怎么给部分分，以及成本上限怎么写。
 - `docs-site/zh/tutorials/experiments.mdx` — 实验矩阵：用运行矩阵比较 agents 和 models:使用 NiceEval experiments 让同一批评估用例横跨多个 agents、models 和 flags，比较 pass rate、成本和延迟。
 - `docs-site/zh/tutorials/fixtures.mdx` — Sandbox Fixture：用任务评估 coding agents:用 .eval.ts 给 coding agent 准备隔离 workspace、发送真实任务，并用 Sandbox 文件、命令、diff 和 judge 验证结果。

package/dist/i18n/en.d.ts

@@ -162,6 +162,7 @@ export declare const en: {
     "cli.view.noResults": string;
     "cli.view.incompatibleForeign": string;
     "cli.view.urls": string;
+    "codex.envPathManaged": string;
     "context.capabilityMissing": string;
     "context.skipEmpty": string;
     "define.agentNameRequired": string;
@@ -197,11 +198,16 @@ export declare const en: {
     "feedback.human.counts": string;
     "feedback.human.diffHint": string;
     "feedback.human.evalHint": string;
+    "feedback.human.exampleLocator": string;
+    "feedback.human.failureFacts": string;
     "feedback.human.failuresHeader": string;
+    "feedback.human.failuresSoFar": string;
+    "feedback.human.failuresTotalKinds": string;
     "feedback.human.heartbeat": string;
     "feedback.human.inspect": string;
     "feedback.human.keptSandboxesHeader": string;
     "feedback.human.moreActive": string;
+    "feedback.human.moreFailureKinds": string;
     "feedback.human.nextHeader": string;
     "feedback.human.plan": string;
     "feedback.human.planHeader": string;
@@ -217,6 +223,7 @@ export declare const en: {
     "feedback.human.summaryAllReusedLine": string;
     "feedback.human.suppressedFailures": string;
     "feedback.human.trace": string;
+    "feedback.human.warningsHeader": string;
     "feedback.phase.agentSetup": string;
     "feedback.phase.agentEnsure": string;
     "feedback.phase.stateLoad": string;
@@ -269,7 +276,7 @@ export declare const en: {
     "local.dirMissing": string;
     "local.dirNotWritable": string;
     "local.notARepo": string;
-    "local.rootUnsupported": string;
+    "local.userUnsupported": string;
     "report.assertionThreshold": string;
     "report.error": string;
     "report.errored": string;
@@ -354,4 +361,5 @@ export declare const en: {
     "vercel.fileNotFound": string;
     "vercel.rotateFailed": string;
     "vercel.rotated": string;
+    "vercel.userUnsupported": string;
 };

package/dist/i18n/en.js

@@ -228,7 +228,7 @@ export const en = {
     "cli.experimentGroup": " path",
     "cli.fallbackCleanupTimeout": "\ngraceful cleanup timed out; force-cleaning sandboxes...\n",
     "cli.forceCleanupExit": "\nForce-cleaning sandboxes and exiting...\n",
-    "cli.init.done": "Ready: evals/, niceeval.config.ts, and the niceeval agent-rules block in AGENTS.md (points coding agents at node_modules/niceeval/docs-site/zh).\n",
+    "cli.init.done": "Ready: evals/, niceeval.config.ts, and the niceeval agent-rules block in AGENTS.md (tells coding agents to read the bundled docs before writing evals).\n",
     "cli.init.esmHint": "tip: this project's package.json has no \"type\": \"module\". niceeval runs either way, but CommonJS mode disallows top-level await in config/eval files — add \"type\": \"module\" to package.json for the smoothest path.\n",
     "cli.interruptCleanup": "\nInterrupted; cleaning up sandbox containers... (press again to force cleanup and exit)\n",
     "cli.list.header": "Discovered {{count}} evals:\n",
@@ -245,6 +245,8 @@ export const en = {
     "cli.view.noResults": "No results found under {{root}}. Run `niceeval exp` first, then `niceeval view`.\n",
     "cli.view.incompatibleForeign": "{{dir}}: written by {{name}} {{version}} (schemaVersion {{schemaVersion}}); this CLI reads schemaVersion {{supported}}.\nOpen this report with the tool that produced it.\n",
     "cli.view.urls": "niceeval view — open in a browser:\n{{urls}}\n",
+    "codex.envPathManaged": "codexAgent config.env.PATH is not supported: PATH is a Sandbox-managed variable, so silently dropping or overriding it would break hooks and child processes without any error. " +
+        "Prepend directories to it with the Sandbox factory's pathPrepend option instead (see docs/feature/sandbox/library.md).",
     "context.capabilityMissing": "Agent \"{{agent}}\" is not sandbox-backed (built with defineSandboxAgent), so t.{{method}} is unavailable. Use an agent built with defineSandboxAgent, or drop this assertion.\n" +
         "  Docs: node_modules/niceeval/docs-site/zh/tutorials/sandbox-agent.mdx",
     "context.skipEmpty": "skip() requires a non-empty reason.",
@@ -281,11 +283,16 @@ export const en = {
     "feedback.human.counts": "{{total}} total · {{reused}} reused · {{running}} running · {{queued}} queued · {{passed}} passed · {{failed}} failed · {{errored}} errored · {{skipped}} skipped",
     "feedback.human.diffHint": "Diff:    niceeval show {{locator}} --diff",
     "feedback.human.evalHint": "Eval:    niceeval show {{locator}} --source",
+    "feedback.human.exampleLocator": "e.g. {{locator}}",
+    "feedback.human.failureFacts": "facts ×{{count}}",
     "feedback.human.failuresHeader": "FAILURES",
+    "feedback.human.failuresSoFar": "{{count}} so far",
+    "feedback.human.failuresTotalKinds": "{{total}} total · {{kinds}} kinds",
     "feedback.human.heartbeat": "{{elapsed}} elapsed · {{counts}}",
     "feedback.human.inspect": "Inspect: niceeval show {{locator}}",
     "feedback.human.keptSandboxesHeader": "KEPT SANDBOXES",
     "feedback.human.moreActive": "… {{count}} more active",
+    "feedback.human.moreFailureKinds": "+{{count}} more kinds — niceeval view",
     "feedback.human.nextHeader": "NEXT",
     "feedback.human.plan": "{{total}} attempts · {{evals}} evals × {{configs}} configs · concurrency {{concurrency}}",
     "feedback.human.planHeader": "PLAN",
@@ -301,6 +308,7 @@ export const en = {
     "feedback.human.summaryAllReusedLine": "{{passed}} passed · {{failed}} failed · {{errored}} errored  (all {{reused}} reused)",
     "feedback.human.suppressedFailures": "… {{count}} more failures suppressed",
     "feedback.human.trace": "Trace:   niceeval show {{locator}} --execution",
+    "feedback.human.warningsHeader": "WARNINGS",
     "feedback.phase.agentSetup": "agent setup",
     "feedback.phase.agentEnsure": "preparing agent",
     "feedback.phase.stateLoad": "loading experiment state",
@@ -353,7 +361,7 @@ export const en = {
     "local.dirMissing": "local sandbox directory does not exist: {{dir}}. Create it first, or point localSandbox({ dir }) at an existing directory.",
     "local.dirNotWritable": "local sandbox directory is not writable: {{dir}} ({{message}})",
     "local.notARepo": "the current directory (and its parents) is not inside a git repository, so localSandbox() has no deterministic root to run in. cd into the repository you want to evaluate, or pass an explicit directory: localSandbox({ dir: \"/path/to/repo\" }).",
-    "local.rootUnsupported": "the local sandbox provider does not support { root: true } — niceeval does not escalate privileges on your machine. Use a container provider (docker / e2b / vercel) for steps that need root.",
+    "local.userUnsupported": 'the local sandbox provider does not support execution identity overrides, got { user: "{{user}}" } — niceeval does not escalate or switch identity on your machine. Use a container provider (docker / e2b / vercel) for steps that need a different user.',
     "report.assertionThreshold": " (got {{score}} < {{threshold}})",
     "report.error": "error",
     "report.errored": "errored",
@@ -443,4 +451,5 @@ export const en = {
     "vercel.fileNotFound": "File not found: {{path}}",
     "vercel.rotateFailed": "[VercelSandbox] session rotate failed ({{seconds}}s): {{error}}",
     "vercel.rotated": "[VercelSandbox] session rotated after {{seconds}}s -> {{sessionId}}",
+    "vercel.userUnsupported": 'the Vercel Sandbox provider only supports { user: "root" } (mapped to sudo: true) at the command level, got { user: "{{user}}" }. Use a container provider (docker / e2b) for other identities.',
 };

package/dist/i18n/zh-CN.d.ts

@@ -146,7 +146,7 @@ export declare const zhCN: {
     readonly "cli.experimentGroup": "路径";
     readonly "cli.fallbackCleanupTimeout": "\ngraceful 清理超时,强制清理沙箱…\n";
     readonly "cli.forceCleanupExit": "\n强制清理沙箱并退出…\n";
-    readonly "cli.init.done": "已就绪:evals/、niceeval.config.ts,以及 AGENTS.md 里的 niceeval agent 指引区块(指向 node_modules/niceeval/docs-site/zh)。\n";
+    readonly "cli.init.done": "已就绪:evals/、niceeval.config.ts,以及 AGENTS.md 里的 niceeval agent 指引区块(要求 coding agent 写 eval 前先读随包文档)。\n";
     readonly "cli.init.esmHint": "提示:项目 package.json 没有 \"type\": \"module\"。niceeval 两种形态都能跑,但 CommonJS 形态下 config / eval 文件用不了顶层 await——建议在 package.json 加上 \"type\": \"module\"。\n";
     readonly "cli.interruptCleanup": "\n收到中断,正在清理沙箱容器…(再按一次强制清理并退出)\n";
     readonly "cli.list.header": "发现 {{count}} 个 eval:\n";
@@ -162,6 +162,7 @@ export declare const zhCN: {
     readonly "cli.view.noResults": "{{root}} 下没有结果。先 `niceeval exp` 跑一轮,再 `niceeval view`。\n";
     readonly "cli.view.incompatibleForeign": "{{dir}}: 由 {{name}} {{version}} 写入(schemaVersion {{schemaVersion}}),当前 CLI 只读 schemaVersion {{supported}}。\n请用写出它的那个工具查看这份报告。\n";
     readonly "cli.view.urls": "niceeval view 可在浏览器打开：\n{{urls}}\n";
+    readonly "codex.envPathManaged": string;
     readonly "context.capabilityMissing": string;
     readonly "context.skipEmpty": "skip() 需要一个非空理由。";
     readonly "define.agentNameRequired": "defineDirectAgent 需要 name。";
@@ -197,11 +198,16 @@ export declare const zhCN: {
     readonly "feedback.human.counts": "共 {{total}} · 复用 {{reused}} · 运行中 {{running}} · 排队 {{queued}} · 通过 {{passed}} · 失败 {{failed}} · 出错 {{errored}} · 跳过 {{skipped}}";
     readonly "feedback.human.diffHint": "Diff:    niceeval show {{locator}} --diff";
     readonly "feedback.human.evalHint": "Eval:    niceeval show {{locator}} --source";
+    readonly "feedback.human.exampleLocator": "例如 {{locator}}";
+    readonly "feedback.human.failureFacts": "facts ×{{count}}";
     readonly "feedback.human.failuresHeader": "FAILURES";
+    readonly "feedback.human.failuresSoFar": "已发生 {{count}} 条";
+    readonly "feedback.human.failuresTotalKinds": "共 {{total}} 条 · {{kinds}} 种形态";
     readonly "feedback.human.heartbeat": "已运行 {{elapsed}} · {{counts}}";
     readonly "feedback.human.inspect": "Inspect: niceeval show {{locator}}";
     readonly "feedback.human.keptSandboxesHeader": "KEPT SANDBOXES";
     readonly "feedback.human.moreActive": "… 还有 {{count}} 项运行中";
+    readonly "feedback.human.moreFailureKinds": "还有 {{count}} 种形态 — niceeval view";
     readonly "feedback.human.nextHeader": "NEXT";
     readonly "feedback.human.plan": "{{total}} 个 attempt · {{evals}} 个 eval × {{configs}} 个配置 · 并发 {{concurrency}}";
     readonly "feedback.human.planHeader": "PLAN";
@@ -217,6 +223,7 @@ export declare const zhCN: {
     readonly "feedback.human.summaryAllReusedLine": "{{passed}} 通过 · {{failed}} 失败 · {{errored}} 出错  (全部 {{reused}} 条复用)";
     readonly "feedback.human.suppressedFailures": "… 还有 {{count}} 条失败被折叠";
     readonly "feedback.human.trace": "Trace:   niceeval show {{locator}} --execution";
+    readonly "feedback.human.warningsHeader": "WARNINGS";
     readonly "feedback.phase.agentSetup": "agent 预置";
     readonly "feedback.phase.agentEnsure": "正在准备 Agent";
     readonly "feedback.phase.stateLoad": "载入实验状态";
@@ -269,7 +276,7 @@ export declare const zhCN: {
     readonly "local.dirMissing": "本地沙箱目录不存在:{{dir}}。请先创建它,或让 localSandbox({ dir }) 指向一个已存在的目录。";
     readonly "local.dirNotWritable": "本地沙箱目录不可写:{{dir}}({{message}})";
     readonly "local.notARepo": "当前目录(及其上级目录)都不在任何 git 仓库内,localSandbox() 没有确定的 workdir 可用。cd 进要跑的仓库再执行,或显式传 localSandbox({ dir: \"/path/to/repo\" })。";
-    readonly "local.rootUnsupported": "本地 provider 不支持 { root: true }——niceeval 不会在你的机器上提权。需要 root 的步骤请改用容器 provider(docker / e2b / vercel)。";
+    readonly "local.userUnsupported": "本地 provider 不支持切换执行身份,收到 { user: \"{{user}}\" }——niceeval 不会在你的机器上提权或换身份。需要别的身份请改用容器 provider(docker / e2b / vercel)。";
     readonly "report.assertionThreshold": " (得分 {{score}} < {{threshold}})";
     readonly "report.error": "错误";
     readonly "report.errored": "错误";
@@ -354,6 +361,7 @@ export declare const zhCN: {
     readonly "vercel.fileNotFound": "File not found: {{path}}";
     readonly "vercel.rotateFailed": "[VercelSandbox] session rotate failed ({{seconds}}s): {{error}}";
     readonly "vercel.rotated": "[VercelSandbox] session rotated after {{seconds}}s → {{sessionId}}";
+    readonly "vercel.userUnsupported": "Vercel Sandbox provider 的命令级 user 只认 { user: \"root\" }(映射 sudo: true),收到 { user: \"{{user}}\" }。需要别的身份请改用容器 provider(docker / e2b)。";
 };
 export type MessageKey = keyof typeof zhCN;
 export type Messages = globalThis.Record<MessageKey, string>;

package/dist/i18n/zh-CN.js

@@ -223,7 +223,7 @@ export const zhCN = {
     "cli.experimentGroup": "路径",
     "cli.fallbackCleanupTimeout": "\ngraceful 清理超时,强制清理沙箱…\n",
     "cli.forceCleanupExit": "\n强制清理沙箱并退出…\n",
-    "cli.init.done": "已就绪:evals/、niceeval.config.ts,以及 AGENTS.md 里的 niceeval agent 指引区块(指向 node_modules/niceeval/docs-site/zh)。\n",
+    "cli.init.done": "已就绪:evals/、niceeval.config.ts,以及 AGENTS.md 里的 niceeval agent 指引区块(要求 coding agent 写 eval 前先读随包文档)。\n",
     "cli.init.esmHint": "提示:项目 package.json 没有 \"type\": \"module\"。niceeval 两种形态都能跑,但 CommonJS 形态下 config / eval 文件用不了顶层 await——建议在 package.json 加上 \"type\": \"module\"。\n",
     "cli.interruptCleanup": "\n收到中断,正在清理沙箱容器…(再按一次强制清理并退出)\n",
     "cli.list.header": "发现 {{count}} 个 eval:\n",
@@ -240,6 +240,8 @@ export const zhCN = {
     "cli.view.noResults": "{{root}} 下没有结果。先 `niceeval exp` 跑一轮,再 `niceeval view`。\n",
     "cli.view.incompatibleForeign": "{{dir}}: 由 {{name}} {{version}} 写入(schemaVersion {{schemaVersion}}),当前 CLI 只读 schemaVersion {{supported}}。\n请用写出它的那个工具查看这份报告。\n",
     "cli.view.urls": "niceeval view 可在浏览器打开：\n{{urls}}\n",
+    "codex.envPathManaged": "codexAgent config.env.PATH 不支持:PATH 是 Sandbox 受管变量,静默丢弃或覆盖它会让 hooks 与子进程读到错误的可执行文件,且零报错。" +
+        "改用 Sandbox factory 的 pathPrepend 选项前置目录(见 docs/feature/sandbox/library.md)。",
     "context.capabilityMissing": "agent \"{{agent}}\" 不是沙箱型(defineSandboxAgent 构造),t.{{method}} 这类断言只有沙箱型 agent 可用。换用 defineSandboxAgent 构造的 agent,或去掉这条断言。\n" +
         "  文档:node_modules/niceeval/docs-site/zh/tutorials/sandbox-agent.mdx",
     "context.skipEmpty": "skip() 需要一个非空理由。",
@@ -276,11 +278,16 @@ export const zhCN = {
     "feedback.human.counts": "共 {{total}} · 复用 {{reused}} · 运行中 {{running}} · 排队 {{queued}} · 通过 {{passed}} · 失败 {{failed}} · 出错 {{errored}} · 跳过 {{skipped}}",
     "feedback.human.diffHint": "Diff:    niceeval show {{locator}} --diff",
     "feedback.human.evalHint": "Eval:    niceeval show {{locator}} --source",
+    "feedback.human.exampleLocator": "例如 {{locator}}",
+    "feedback.human.failureFacts": "facts ×{{count}}",
     "feedback.human.failuresHeader": "FAILURES",
+    "feedback.human.failuresSoFar": "已发生 {{count}} 条",
+    "feedback.human.failuresTotalKinds": "共 {{total}} 条 · {{kinds}} 种形态",
     "feedback.human.heartbeat": "已运行 {{elapsed}} · {{counts}}",
     "feedback.human.inspect": "Inspect: niceeval show {{locator}}",
     "feedback.human.keptSandboxesHeader": "KEPT SANDBOXES",
     "feedback.human.moreActive": "… 还有 {{count}} 项运行中",
+    "feedback.human.moreFailureKinds": "还有 {{count}} 种形态 — niceeval view",
     "feedback.human.nextHeader": "NEXT",
     "feedback.human.plan": "{{total}} 个 attempt · {{evals}} 个 eval × {{configs}} 个配置 · 并发 {{concurrency}}",
     "feedback.human.planHeader": "PLAN",
@@ -296,6 +303,7 @@ export const zhCN = {
     "feedback.human.summaryAllReusedLine": "{{passed}} 通过 · {{failed}} 失败 · {{errored}} 出错  (全部 {{reused}} 条复用)",
     "feedback.human.suppressedFailures": "… 还有 {{count}} 条失败被折叠",
     "feedback.human.trace": "Trace:   niceeval show {{locator}} --execution",
+    "feedback.human.warningsHeader": "WARNINGS",
     "feedback.phase.agentSetup": "agent 预置",
     "feedback.phase.agentEnsure": "正在准备 Agent",
     "feedback.phase.stateLoad": "载入实验状态",
@@ -348,7 +356,7 @@ export const zhCN = {
     "local.dirMissing": "本地沙箱目录不存在:{{dir}}。请先创建它,或让 localSandbox({ dir }) 指向一个已存在的目录。",
     "local.dirNotWritable": "本地沙箱目录不可写:{{dir}}({{message}})",
     "local.notARepo": "当前目录(及其上级目录)都不在任何 git 仓库内,localSandbox() 没有确定的 workdir 可用。cd 进要跑的仓库再执行,或显式传 localSandbox({ dir: \"/path/to/repo\" })。",
-    "local.rootUnsupported": "本地 provider 不支持 { root: true }——niceeval 不会在你的机器上提权。需要 root 的步骤请改用容器 provider(docker / e2b / vercel)。",
+    "local.userUnsupported": "本地 provider 不支持切换执行身份,收到 { user: \"{{user}}\" }——niceeval 不会在你的机器上提权或换身份。需要别的身份请改用容器 provider(docker / e2b / vercel)。",
     "report.assertionThreshold": " (得分 {{score}} < {{threshold}})",
     "report.error": "错误",
     "report.errored": "错误",
@@ -446,4 +454,5 @@ export const zhCN = {
     "vercel.fileNotFound": "File not found: {{path}}",
     "vercel.rotateFailed": "[VercelSandbox] session rotate failed ({{seconds}}s): {{error}}",
     "vercel.rotated": "[VercelSandbox] session rotated after {{seconds}}s → {{sessionId}}",
+    "vercel.userUnsupported": "Vercel Sandbox provider 的命令级 user 只认 { user: \"root\" }(映射 sudo: true),收到 { user: \"{{user}}\" }。需要别的身份请改用容器 provider(docker / e2b)。",
 };

package/dist/record/attempt-evidence.d.ts

@@ -58,7 +58,7 @@ export interface AttemptEvidence {
     diff: DiffData | null;
     /** OTel spans(`--timing` 把 turn 节点下的 agent/model/tool spans 挂回时间树);没有 trace 就是 null。 */
     trace: import("../types.ts").TraceSpan[] | null;
-    /** 非零 Sandbox 命令的 stdout/stderr 证据(`--execution` 的独立命令卡);原样透传 attempt.commands()。 */
+    /** Sandbox 命令(成功与非零退出都记)的 stdout/stderr 证据(`--execution` 的独立命令卡);原样透传 attempt.commands()。 */
     commands: readonly CommandExitEvidence[] | null;
     artifactPaths: EvidencePaths;
     capabilities: AttemptEvidenceCapabilities;

package/dist/record/open.js

@@ -429,8 +429,33 @@ async function readSnapshotDir(dir, meta, state) {
         ev.attempts.sort((a, b) => a.result.attempt - b.result.attempt || (a.result.startedAt ?? "").localeCompare(b.result.startedAt ?? ""));
     }
     run.attempts = run.evals.flatMap((ev) => ev.attempts);
+    // 回退推导:run.json 缺 configHash(旧存量落盘,exp 写入面在这条修法之前从不写它)时,
+    // 若该快照全部 attempt 的 result.configHash 一致,取之为 Run 的 configHash——让存量记录
+    // 不重跑即痊愈。任一 attempt 缺失或值不一致就如实保持 undefined,不猜一个可能错的值
+    // (见 docs/feature/record/library.md「configHash:配置身份只算一次」)。
+    if (run.configHash === undefined) {
+        const derived = deriveConfigHashFromAttempts(run.attempts);
+        if (derived !== undefined)
+            run.configHash = derived;
+    }
     return run;
 }
+/** `readSnapshotDir` 的 configHash 回退推导:全部 attempt 都带且相等才取用,否则 undefined。 */
+function deriveConfigHashFromAttempts(attempts) {
+    if (attempts.length === 0)
+        return undefined;
+    let hash;
+    for (const attempt of attempts) {
+        const value = attempt.result.configHash;
+        if (value === undefined)
+            return undefined;
+        if (hash === undefined)
+            hash = value;
+        else if (hash !== value)
+            return undefined;
+    }
+    return hash;
+}
 function makeAttempt(run, snapshotDir, attemptDir, record) {
     // 候选 artifact 目录:本 attempt 目录为主;--resume 携带条目的 artifact 留在原快照里,
     // artifactBase(相对结果根 = 快照目录的上两级)指向那里,作为回退。

package/dist/report/assets/series-encoding.d.ts

@@ -0,0 +1,36 @@
+import type { ReactNode } from "react";
+import type { SeriesVariant } from "../presentation.ts";
+/** 槽位色板下标 1..6 → CSS 变量,原样交给 fill / stroke / color。 */
+export declare function seriesColorVar(colorIndex: number): string;
+/** SVG pattern id:niceeval-series-pat-v{2|3|4}-c{1..6}。variant 1 是实心,无 pattern。 */
+export declare function seriesPatternId(colorIndex: number, variant: SeriesVariant): string;
+export declare function seriesFill(colorIndex: number, variant: SeriesVariant): string;
+/** 线型变体 → strokeDasharray;variant 1 为空串(实线)。 */
+export declare function seriesStrokeDasharray(variant: SeriesVariant): string;
+export interface SeriesMarkerShape {
+    readonly path: string;
+    readonly viewBox: string;
+}
+/** 四种 marker 形状(viewBox 0 0 12 12),变体按 docs 槽序表 1–4。 */
+export declare function seriesMarkerShape(variant: SeriesVariant): SeriesMarkerShape;
+export declare function seriesMarker(colorIndex: number, variant: SeriesVariant): {
+    readonly path: string;
+    readonly viewBox: string;
+    readonly fill: string;
+    readonly stroke: string;
+};
+/**
+ * 从 presentation.fill 反推 HTML 柱需要的系列 class。
+ * SVG 用 url(#pattern)/var();HTML 柱不能引用 SVG pattern,改挂 series-cN + fill-vN 类,
+ * 由 styles.css 用 repeating-linear-gradient 画等效图案,颜色仍走 --series 令牌。
+ */
+export declare function seriesClassesFromFill(fill: string): string;
+/** 从 color 呈现或 series 的 stroke/fill 反推 series-cN 类(图例色点回落)。 */
+export declare function seriesClassFromColorVar(color: string): string;
+/**
+ * 页内注入一次的 SVG pattern defs。
+ * 18 个 pattern(3 非实心变体 × 6 色);id 与 seriesFill() 产出的 url(#…) 对齐。
+ * 子元素直接写 var(--niceeval-color-series-N),不走 currentColor
+ * (SVG pattern 内部 currentColor 取自 pattern 自身,引用者传不进来)。
+ */
+export declare function SeriesPatternDefs(): ReactNode;

package/dist/report/assets/series-encoding.js

@@ -0,0 +1,104 @@
+import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
+/** 槽位色板下标 1..6 → CSS 变量,原样交给 fill / stroke / color。 */
+export function seriesColorVar(colorIndex) {
+    if (!Number.isInteger(colorIndex) || colorIndex < 1 || colorIndex > 6) {
+        throw new Error(`colorIndex must be an integer in [1, 6], got ${colorIndex}.`);
+    }
+    return `var(--niceeval-color-series-${colorIndex})`;
+}
+/** SVG pattern id:niceeval-series-pat-v{2|3|4}-c{1..6}。variant 1 是实心,无 pattern。 */
+export function seriesPatternId(colorIndex, variant) {
+    if (variant === 1) {
+        throw new Error("variant 1 is solid fill; it has no pattern id.");
+    }
+    return `niceeval-series-pat-v${variant}-c${colorIndex}`;
+}
+export function seriesFill(colorIndex, variant) {
+    if (variant === 1)
+        return seriesColorVar(colorIndex);
+    return `url(#${seriesPatternId(colorIndex, variant)})`;
+}
+/** 线型变体 → strokeDasharray;variant 1 为空串(实线)。 */
+export function seriesStrokeDasharray(variant) {
+    switch (variant) {
+        case 1:
+            return "";
+        case 2:
+            return "6 4";
+        case 3:
+            return "2 3";
+        case 4:
+            return "8 3 2 3";
+    }
+}
+/** 四种 marker 形状(viewBox 0 0 12 12),变体按 docs 槽序表 1–4。 */
+export function seriesMarkerShape(variant) {
+    switch (variant) {
+        case 1:
+            // 圆
+            return { path: "M6 1.5a4.5 4.5 0 1 0 0.01 0z", viewBox: "0 0 12 12" };
+        case 2:
+            // 方
+            return { path: "M2.5 2.5h7v7h-7z", viewBox: "0 0 12 12" };
+        case 3:
+            // 菱
+            return { path: "M6 1.5 10.5 6 6 10.5 1.5 6z", viewBox: "0 0 12 12" };
+        case 4:
+            // 三角
+            return { path: "M6 1.5 10.5 10.5 1.5 10.5z", viewBox: "0 0 12 12" };
+    }
+}
+export function seriesMarker(colorIndex, variant) {
+    const shape = seriesMarkerShape(variant);
+    const color = seriesColorVar(colorIndex);
+    return {
+        path: shape.path,
+        viewBox: shape.viewBox,
+        fill: color,
+        stroke: color,
+    };
+}
+/**
+ * 从 presentation.fill 反推 HTML 柱需要的系列 class。
+ * SVG 用 url(#pattern)/var();HTML 柱不能引用 SVG pattern,改挂 series-cN + fill-vN 类,
+ * 由 styles.css 用 repeating-linear-gradient 画等效图案,颜色仍走 --series 令牌。
+ */
+export function seriesClassesFromFill(fill) {
+    const pattern = /^url\(#niceeval-series-pat-v([2-4])-c([1-6])\)$/.exec(fill);
+    if (pattern) {
+        const variant = pattern[1];
+        const colorIndex = Number(pattern[2]);
+        return `niceeval-series-c${colorIndex - 1} niceeval-series-fill-v${variant}`;
+    }
+    const solid = /^var\(--niceeval-color-series-([1-6])\)$/.exec(fill);
+    if (solid) {
+        return `niceeval-series-c${Number(solid[1]) - 1}`;
+    }
+    return "niceeval-series-none";
+}
+/** 从 color 呈现或 series 的 stroke/fill 反推 series-cN 类(图例色点回落)。 */
+export function seriesClassFromColorVar(color) {
+    const solid = /^var\(--niceeval-color-series-([1-6])\)$/.exec(color);
+    if (solid)
+        return `niceeval-series-c${Number(solid[1]) - 1}`;
+    return "niceeval-series-none";
+}
+/**
+ * 页内注入一次的 SVG pattern defs。
+ * 18 个 pattern(3 非实心变体 × 6 色);id 与 seriesFill() 产出的 url(#…) 对齐。
+ * 子元素直接写 var(--niceeval-color-series-N),不走 currentColor
+ * (SVG pattern 内部 currentColor 取自 pattern 自身,引用者传不进来)。
+ */
+export function SeriesPatternDefs() {
+    const patterns = [];
+    for (let colorIndex = 1; colorIndex <= 6; colorIndex++) {
+        const color = seriesColorVar(colorIndex);
+        // v2:对角斜线
+        patterns.push(_jsxs("pattern", { id: seriesPatternId(colorIndex, 2), patternUnits: "userSpaceOnUse", width: "6", height: "6", children: [_jsx("rect", { width: "6", height: "6", fill: color, fillOpacity: 0.22 }), _jsx("path", { d: "M-1 1l2-2M0 6l6-6M5 7l2-2", stroke: color, strokeWidth: 1.4, fill: "none" })] }, `v2-c${colorIndex}`));
+        // v3:水平条纹
+        patterns.push(_jsxs("pattern", { id: seriesPatternId(colorIndex, 3), patternUnits: "userSpaceOnUse", width: "6", height: "6", children: [_jsx("rect", { width: "6", height: "6", fill: color, fillOpacity: 0.18 }), _jsx("path", { d: "M0 1.5h6M0 4.5h6", stroke: color, strokeWidth: 1.5, fill: "none" })] }, `v3-c${colorIndex}`));
+        // v4:点阵
+        patterns.push(_jsxs("pattern", { id: seriesPatternId(colorIndex, 4), patternUnits: "userSpaceOnUse", width: "6", height: "6", children: [_jsx("rect", { width: "6", height: "6", fill: color, fillOpacity: 0.16 }), _jsx("circle", { cx: "2", cy: "2", r: "1.15", fill: color }), _jsx("circle", { cx: "5", cy: "5", r: "1.15", fill: color })] }, `v4-c${colorIndex}`));
+    }
+    return (_jsx("svg", { className: "niceeval-series-defs", width: 0, height: 0, "aria-hidden": "true", focusable: "false", style: { position: "absolute", width: 0, height: 0, overflow: "hidden" }, children: _jsx("defs", { children: patterns }) }));
+}

package/dist/report/built-in/result-components.js

@@ -4,6 +4,7 @@ import { Callouts, Chart, Col, Conversation, CopyBlock, DiffView, Grid, Scatter,
 import { HeroCard } from "../components/site-components/index.js";
 import { AttemptSummary } from "../components/attempt-detail/index.js";
 import { experimentListContent } from "../components/entity-lists/content.js";
+import { ExperimentTableView } from "../components/entity-lists/index.js";
 import { attemptAssertionsContent, attemptConversationContent, attemptDiffContent, embedConversationInSource, executionEvidenceUnavailableCallouts, attemptFixPromptContent, attemptNoticesContent, projectedSourceContent, attemptTimelineContent, } from "../components/attempt-detail/content.js";
 import { stabilityMatrixContent } from "../slices/content.js";
 import { formatInstant, formatReportDateTimeRange, } from "../model/format.js";
@@ -34,6 +35,7 @@ function summaryView(result, locale) {
 }
 export const StandardOverviewResultView = defineComponent(async ({ result }, ctx) => {
     const table = experimentListContent(result.experiments);
+    const freshTable = result.freshExperiments ? experimentListContent(result.freshExperiments) : null;
     const hasPassRate = table.columns.some((column) => column.key === "passRate");
     const hasTotalScore = table.columns.some((column) => column.key === "totalScore");
     const defaultSort = hasPassRate === hasTotalScore
@@ -41,7 +43,7 @@ export const StandardOverviewResultView = defineComponent(async ({ result }, ctx
         : hasPassRate
             ? "passRate"
             : "totalScore";
-    return (_jsxs(Col, { children: [_jsx(HeroCard, { title: ctx.report.title, data: result.hero }), _jsx(Callouts, { items: result.notices }), _jsx(Callouts, { items: result.diagnostics }), _jsx(CopyBlock, { content: result.fixPrompt }), summaryView(result, DEFAULT_REPORT_LOCALE), _jsx(Col, { children: result.charts.map((chart) => chart.y === "passRate" ? (_jsx(Scatter, { points: chart.points, x: "costUSD", y: "passRate", point: "experiment", series: "series", connect: chart.connect, pointTarget: experimentTarget, legend: true }, chart.y)) : (_jsx(Scatter, { points: chart.points, x: "costUSD", y: "totalScore", point: "experiment", series: "series", connect: chart.connect, pointTarget: experimentTarget, legend: true }, chart.y))) }), _jsx(TableContentView, { data: table, sort: defaultSort, searchable: true })] }));
+    return (_jsxs(Col, { children: [_jsx(HeroCard, { title: ctx.report.title, data: result.hero }), _jsx(Callouts, { items: result.notices }), _jsx(Callouts, { items: result.diagnostics }), _jsx(CopyBlock, { content: result.fixPrompt }), summaryView(result, DEFAULT_REPORT_LOCALE), _jsx(Col, { children: result.charts.map((chart) => chart.y === "passRate" ? (_jsx(Scatter, { points: chart.points, x: "costUSD", y: "passRate", point: "experiment", series: "series", connect: chart.connect, pointTarget: experimentTarget, legend: true }, chart.y)) : (_jsx(Scatter, { points: chart.points, x: "costUSD", y: "totalScore", point: "experiment", series: "series", connect: chart.connect, pointTarget: experimentTarget, legend: true }, chart.y))) }), freshTable === null ? (_jsx(TableContentView, { data: table, sort: defaultSort, searchable: true })) : (_jsx(ExperimentTableView, { fullContent: table, freshContent: freshTable, sort: defaultSort, searchable: true }))] }));
 });
 StandardOverviewResultView.displayName = "StandardOverviewResultView";
 function stabilityPresentation(result) {
@@ -169,6 +171,14 @@ const TaskUsageKpi = defineComponent({
     },
 });
 TaskUsageKpi.displayName = "TaskUsageKpi";
+/** attempt 级 `ctx.fact()` 运行事实的完整键值表(docs/feature/record/architecture.md#facts运行事实);
+ *  没有 facts 时零输出,不摆空表。 */
+const TaskFactsResultView = defineComponent(({ data }) => {
+    if (data === null || data.facts.length === 0)
+        return null;
+    return (_jsx(Grid, { className: "niceeval-facts-table", children: data.facts.map(({ key, value }) => (_jsx(TaskUsageKpi, { label: key, value: String(value) }, key))) }));
+});
+TaskFactsResultView.displayName = "TaskFactsResultView";
 export const AttemptDetailsResultView = defineComponent(({ result }) => {
     const notices = attemptNoticesContent(result.error, result.diagnostics) ?? [];
     const source = projectedSourceContent(result.source.source, result.source.locator);
@@ -184,6 +194,6 @@ export const AttemptDetailsResultView = defineComponent(({ result }) => {
     const conversation = attemptConversationContent(result.conversation.conversation);
     const embedded = embedConversationInSource(source, conversation);
     const files = attemptDiffContent(result.diff);
-    return (_jsxs(Col, { children: [_jsx(AttemptSummary, { data: result.summary }), _jsx(Callouts, { items: notices }), embedded.source !== null ? (_jsx(SourceView, { data: embedded.source })) : assertions !== null && assertions.rows.length > 0 ? (_jsx(TableContentView, { data: assertions })) : null, _jsx(CopyBlock, { content: attemptFixPromptContent(result.fixPrompt) }), _jsx(Waterfall, { nodes: timeline ?? [], title: { en: "Execution timeline", "zh-CN": "执行时间轴" } }), _jsx(TaskUsageResultView, { data: result.usage }), conversation === null ? (_jsx(Callouts, { items: executionEvidenceUnavailableCallouts })) : embedded.conversation !== null ? (_jsx(Conversation, { data: embedded.conversation })) : (null), _jsx(DiffView, { files: files })] }));
+    return (_jsxs(Col, { children: [_jsx(AttemptSummary, { data: result.summary }), _jsx(Callouts, { items: notices }), embedded.source !== null ? (_jsx(SourceView, { data: embedded.source })) : assertions !== null && assertions.rows.length > 0 ? (_jsx(TableContentView, { data: assertions })) : null, _jsx(CopyBlock, { content: attemptFixPromptContent(result.fixPrompt) }), _jsx(Waterfall, { nodes: timeline ?? [], title: { en: "Execution timeline", "zh-CN": "执行时间轴" } }), _jsx(TaskUsageResultView, { data: result.usage }), _jsx(TaskFactsResultView, { data: result.facts }), conversation === null ? (_jsx(Callouts, { items: executionEvidenceUnavailableCallouts })) : embedded.conversation !== null ? (_jsx(Conversation, { data: embedded.conversation })) : (null), _jsx(DiffView, { files: files })] }));
 });
 AttemptDetailsResultView.displayName = "AttemptDetailsResultView";

package/dist/report/components/attempt-detail/compute.d.ts

@@ -1,5 +1,5 @@
 import type { AttemptEvidence } from "../../../record/attempt-evidence.ts";
-import type { AttemptAssertionsData, AttemptCommandEvidenceData, AttemptConversationData, AttemptDiagnosticsData, AttemptDiffData, AttemptErrorData, AttemptFixPromptData, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, UsageTableData } from "../../model/types.ts";
+import type { AttemptAssertionsData, AttemptCommandEvidenceData, AttemptConversationData, AttemptDiagnosticsData, AttemptDiffData, AttemptErrorData, AttemptFactsData, AttemptFixPromptData, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, UsageTableData } from "../../model/types.ts";
 export declare function attemptSummaryData(evidence: AttemptEvidence): AttemptSummaryData;
 export declare function attemptErrorData(evidence: AttemptEvidence): AttemptErrorData | null;
 export declare function attemptAssertionsData(evidence: AttemptEvidence): AttemptAssertionsData | null;
@@ -35,6 +35,13 @@ export declare function attemptDiagnosticsData(evidence: AttemptEvidence): Attem
  * 没有任何用量事实可摆,与其余叶子同一条"没有 usage 时零输出"规则。
  */
 export declare function usageTableData(evidence: AttemptEvidence): UsageTableData | null;
+/**
+ * attempt 级 `ctx.fact()` 运行事实的完整键值表;`AttemptRecord.facts` 缺失或为空对象时返回
+ * null,不渲染空表(与其余叶子同一条「没有证据时零输出」规则)。
+ * `facts` 落盘就是 `Record<string, string | number | boolean>`,JS 对象的字符串键天然保留
+ * 写入顺序,这里按该顺序投影成数组,不重新排序。
+ */
+export declare function attemptFactsData(evidence: AttemptEvidence): AttemptFactsData | null;
 export declare function attemptTraceData(evidence: AttemptEvidence): AttemptTraceData | null;
 /**
  * `null` 与空清单是两件事:`null` = 这次 attempt 没有 diff 证据(direct agent、发布未带 diff),

package/dist/report/components/attempt-detail/compute.js

@@ -57,7 +57,9 @@ function looksLikeTruncatedCommandTail(message, commands) {
 }
 /** 展示层唯一的命令分类规则;Record 只保存 checked 与 exitCode。 */
 function commandClassification(command) {
-    return command.checked && command.exitCode !== 0 ? "failed" : "observed";
+    if (command.exitCode === 0)
+        return "succeeded";
+    return command.checked ? "failed" : "observed";
 }
 export function attemptErrorData(evidence) {
     const err = evidence.result.error;
@@ -394,6 +396,22 @@ export function usageTableData(evidence) {
         ...(estimatedCostUSD !== null ? { estimatedCostUSD } : {}),
     };
 }
+// ───────────────────────── AttemptFacts ─────────────────────────
+/**
+ * attempt 级 `ctx.fact()` 运行事实的完整键值表;`AttemptRecord.facts` 缺失或为空对象时返回
+ * null,不渲染空表(与其余叶子同一条「没有证据时零输出」规则)。
+ * `facts` 落盘就是 `Record<string, string | number | boolean>`,JS 对象的字符串键天然保留
+ * 写入顺序,这里按该顺序投影成数组,不重新排序。
+ */
+export function attemptFactsData(evidence) {
+    const facts = evidence.result.facts;
+    if (!facts)
+        return null;
+    const entries = Object.entries(facts);
+    if (entries.length === 0)
+        return null;
+    return { facts: entries.map(([key, value]) => ({ key, value })) };
+}
 // ───────────────────────── AttemptTrace ─────────────────────────
 export function attemptTraceData(evidence) {
     const spans = evidence.trace;

package/dist/report/components/attempt-detail/index.js

@@ -6,7 +6,7 @@ import { Callouts, CommandEvidence, Col, Conversation, CopyBlock, DiffView, Grid
 import { formatDurationMs, formatInstant, formatPoints, formatUSD } from "../../model/format.js";
 import { localeText } from "../../model/locale.js";
 import { cx } from "../shared.js";
-import { toAttemptAssertions, toAttemptFixPrompt, toAttemptNotices, toAttemptSource, toAttemptSummary, toAttemptUsage, toCommandEvidence, toConversationTurns, toDiffFiles, toTimelineNodes, } from "../../model/conversions.js";
+import { toAttemptAssertions, toAttemptFacts, toAttemptFixPrompt, toAttemptNotices, toAttemptSource, toAttemptSummary, toAttemptUsage, toCommandEvidence, toConversationTurns, toDiffFiles, toTimelineNodes, } from "../../model/conversions.js";
 import { embedConversationInSource, executionEvidenceUnavailableCallouts } from "./content.js";
 export { validateAssertionsData, validateCommandEvidenceData, validateConversationData, validateDiagnosticsData, validateDiffData, validateErrorData, validateFixPromptData, validateSummaryData, validateTimelineData, validateTraceData, validateUsageData, } from "./validate.js";
 /**
@@ -80,6 +80,27 @@ const AttemptUsage = defineComponent({
     },
 });
 AttemptUsage.displayName = "AttemptUsage";
+/** attempt 级 `ctx.fact()` 运行事实的完整键值表(docs/feature/record/architecture.md#facts运行事实);
+ *  没有 facts 时零输出,不摆空表。 */
+const AttemptFacts = defineComponent({
+    dimensions: () => ({}),
+    web(props) {
+        const d = props.data;
+        if (d === null || d === undefined || d.facts.length === 0)
+            return null;
+        return (_jsx(Grid, { className: cx("niceeval-facts-table", props.className), children: d.facts.map(({ key, value }) => (_jsx(Kpi, { label: key, value: String(value) }, key))) }));
+    },
+    text(props) {
+        const d = props.data;
+        if (d === null || d === undefined || d.facts.length === 0)
+            return "";
+        const lines = ["facts:"];
+        for (const { key, value } of d.facts)
+            lines.push(`  ${key}  ${value}`);
+        return lines.join("\n");
+    },
+});
+AttemptFacts.displayName = "AttemptFacts";
 function evidenceOf(props, ctx) {
     if (props.attempt !== undefined)
         return props.attempt;
@@ -113,7 +134,7 @@ AttemptAssessment.displayName = "AttemptAssessment";
 export const AttemptDetails = defineComponent(async (props, ctx) => {
     const evidence = evidenceOf(props, ctx);
     const hasSource = evidence.capabilities.source;
-    const [notices, source, assertions, summary, fixPrompt, timeline, usage, commandEvidence, conversation, diff] = await Promise.all([
+    const [notices, source, assertions, summary, fixPrompt, timeline, usage, facts, commandEvidence, conversation, diff] = await Promise.all([
         toAttemptNotices(evidence),
         hasSource ? toAttemptSource(evidence) : Promise.resolve(null),
         hasSource ? Promise.resolve(null) : toAttemptAssertions(evidence),
@@ -121,12 +142,13 @@ export const AttemptDetails = defineComponent(async (props, ctx) => {
         toAttemptFixPrompt(evidence),
         toTimelineNodes(evidence),
         toAttemptUsage(evidence),
+        toAttemptFacts(evidence),
         toCommandEvidence(evidence),
         toConversationTurns(evidence),
         toDiffFiles(evidence),
     ]);
     const embedded = embedConversationInSource(source, conversation);
     const commandSections = commandEvidenceSections(commandEvidence);
-    return (_jsxs(Col, { className: props.className, children: [_jsx(AttemptSummary, { data: summary }), _jsxs(Col, { children: [_jsx(Callouts, { items: notices }), _jsx(CommandEvidence, { data: commandSections.before }), embedded.source !== null ? (_jsx(SourceView, { data: embedded.source })) : assertions !== null && assertions.rows.length > 0 ? (_jsx(TableContentView, { data: assertions })) : null] }), _jsx(CopyBlock, { content: fixPrompt }), _jsx(Waterfall, { nodes: timeline, title: { en: "Execution timeline", "zh-CN": "执行时间轴" } }), _jsx(AttemptUsage, { data: usage }), conversation === null ? (_jsx(Callouts, { items: executionEvidenceUnavailableCallouts })) : embedded.conversation !== null ? (_jsx(Conversation, { data: embedded.conversation })) : (null), _jsx(CommandEvidence, { data: commandSections.after }), _jsx(DiffView, { files: diff })] }));
+    return (_jsxs(Col, { className: props.className, children: [_jsx(AttemptSummary, { data: summary }), _jsxs(Col, { children: [_jsx(Callouts, { items: notices }), _jsx(CommandEvidence, { data: commandSections.before }), embedded.source !== null ? (_jsx(SourceView, { data: embedded.source })) : assertions !== null && assertions.rows.length > 0 ? (_jsx(TableContentView, { data: assertions })) : null] }), _jsx(CopyBlock, { content: fixPrompt }), _jsx(Waterfall, { nodes: timeline, title: { en: "Execution timeline", "zh-CN": "执行时间轴" } }), _jsx(AttemptUsage, { data: usage }), _jsx(AttemptFacts, { data: facts }), conversation === null ? (_jsx(Callouts, { items: executionEvidenceUnavailableCallouts })) : embedded.conversation !== null ? (_jsx(Conversation, { data: embedded.conversation })) : (null), _jsx(CommandEvidence, { data: commandSections.after }), _jsx(DiffView, { files: diff })] }));
 });
 AttemptDetails.displayName = "AttemptDetails";

package/dist/report/components/attempt-detail/validate.js

@@ -283,8 +283,9 @@ function commandEvidenceProblem(value, path) {
         return `"${path}.display" must be a string`;
     if (typeof value.exitCode !== "number")
         return `"${path}.exitCode" must be a number`;
-    if (value.classification !== "observed" && value.classification !== "failed")
-        return `"${path}.classification" must be observed | failed`;
+    if (value.classification !== "succeeded" && value.classification !== "observed" && value.classification !== "failed") {
+        return `"${path}.classification" must be succeeded | observed | failed`;
+    }
     if (value.durationMs !== undefined && typeof value.durationMs !== "number")
         return `"${path}.durationMs" must be a number`;
     if (typeof value.stdout !== "string")

package/dist/report/components/entity-lists/compute.d.ts

@@ -11,6 +11,12 @@ export declare function failureSummaryOf(result: EvalResult): {
     summary: string | null;
     more: number;
 };
+/**
+ * 「只看新执行」开关在场的判据(docs/feature/reports/components/summaries/experiment-table.md
+ * 「只看新执行」):Sample 里既没有历史执行也没有过期结论时不画开关——一个永远不改变行集的
+ * 控件只会让人怀疑自己看漏了什么。`ExperimentTable` 与内建默认报告共用同一条判据。
+ */
+export declare function hasHistoricalOrStale(items: readonly ExperimentListItem[]): boolean;
 /** 已选出的 AttemptHandle[] → 列表行；顺序保持传入顺序（不再次按 Sample 去重）。 */
 export declare function attemptRowsOf(attempts: readonly AttemptHandle[]): Promise<AttemptListItem[]>;
 /** `attemptListData(input)`:每个 Attempt 一项,顺序取自 Sample 展平顺序(不重排)。 */