niceeval 0.11.4-canary.20 → 0.12.0

package/src/i18n/zh-CN.ts

@@ -281,7 +281,7 @@ 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",
@@ -300,6 +300,9 @@ 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",
@@ -338,11 +341,16 @@ export const zhCN = {
     "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",
@@ -359,6 +367,7 @@ 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",
@@ -413,7 +422,7 @@ export const zhCN = {
   "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",
@@ -516,6 +525,8 @@ 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)。",
 } as const;
 
 export type MessageKey = keyof typeof zhCN;

package/src/record/attempt-evidence.ts

@@ -91,7 +91,7 @@ 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/src/record/open.ts

@@ -505,9 +505,30 @@ async function readSnapshotDir(dir: string, meta: RunMeta, state: ScanState): Pr
     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: readonly AttemptHandle[]): string | undefined {
+  if (attempts.length === 0) return undefined;
+  let hash: string | undefined;
+  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: Run, snapshotDir: string, attemptDir: string, record: EvalResult): AttemptHandle {
   // 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/src/record/results.test.ts

@@ -207,6 +207,38 @@ describe("openRecord · 实验 → 快照 → eval → attempt 分层", () => {
   });
 });
 
+describe("openRecord · configHash 回退推导", () => {
+  // cases: docs/engineering/testing/unit/record.md「configHash 读取面回退推导」
+  // bug: memory/exp-runjson-missing-confighash-breaks-current-sample.md
+  it("run.json 缺 configHash 时,全部 attempt 的 result.configHash 一致才回退推导为 Run 的值;不一致则保持缺失", async () => {
+    const root = await makeRoot();
+    // meta() 不设置 configHash——模拟 exp 写入面在这条修法之前遗留的存量落盘。
+    const consistentDir = await writeSnapshot(
+      root,
+      "compare_a",
+      "2026-07-01T08-00-00-000Z-a1b2",
+      meta({ experimentId: "compare/a", agent: "codex", startedAt: "2026-07-01T08:00:00.000Z" }),
+    );
+    await writeResultFile(consistentDir, "q1/a1", record({ id: "q1", attempt: 1, configHash: "same-hash" }));
+    await writeResultFile(consistentDir, "q2/a1", record({ id: "q2", attempt: 1, configHash: "same-hash" }));
+
+    const inconsistentDir = await writeSnapshot(
+      root,
+      "compare_b",
+      "2026-07-01T08-00-00-000Z-c3d4",
+      meta({ experimentId: "compare/b", agent: "codex", startedAt: "2026-07-01T08:00:00.000Z" }),
+    );
+    await writeResultFile(inconsistentDir, "q1/a1", record({ id: "q1", attempt: 1, configHash: "hash-a" }));
+    await writeResultFile(inconsistentDir, "q2/a1", record({ id: "q2", attempt: 1, configHash: "hash-b" }));
+
+    const results = await openRecord(root);
+    const a = results.experiments.find((e) => e.id === "compare/a")!;
+    const b = results.experiments.find((e) => e.id === "compare/b")!;
+    expect(a.latestRun.configHash).toBe("same-hash");
+    expect(b.latestRun.configHash).toBeUndefined();
+  });
+});
+
 describe("record · migratedFrom", () => {
   it("读写保留旧 opaque carryEpoch 的迁移来源", async () => {
     const root = await makeRoot();
@@ -524,6 +556,34 @@ describe("createWriter", () => {
     expect(JSON.parse(await readFile(join(snap.dir, "run.json"), "utf-8")).completedAt).toBeUndefined();
   });
 
+  // cases: docs/engineering/testing/unit/record.md「configHash 写入面」
+  // bug: memory/exp-runjson-missing-confighash-breaks-current-sample.md
+  it("configHash:writeAttemptFor 的隐式声明按 WriterOptions.configHashes 取值(exp 路径不必显式声明)", async () => {
+    const root = await makeRoot();
+    const writer = createWriter(root, {
+      producer: { name: "niceeval", version: "0.12.0" },
+      configHashes: new Map([["compare/a", "cfg-hash-v1"]]),
+    });
+    await writer.writeAttemptFor({
+      id: "q1",
+      experimentId: "compare/a",
+      agent: "codex",
+      verdict: "passed",
+      attempt: 0,
+      startedAt: "2026-07-01T08:00:00.000Z",
+      durationMs: 10,
+      assertions: [],
+    });
+    await finishAll(writer);
+
+    const dir = writer.snapshotDirs().find((s) => s.experimentId === "compare/a")!.dir;
+    const meta = JSON.parse(await readFile(join(dir, "run.json"), "utf-8"));
+    expect(meta.configHash).toBe("cfg-hash-v1");
+
+    const results = await openRecord(root);
+    expect(results.experiments[0]!.latestRun.configHash).toBe("cfg-hash-v1");
+  });
+
   it("没有清单的 Run 不生成 manifests.json(读取面如实为缺失,不合成一份空清单)", async () => {
     const root = await makeRoot();
     const writer = createWriter(root, { producer: { name: "niceeval", version: "0.12.0" } });
@@ -611,8 +671,8 @@ describe("createWriter", () => {
     expect(JSON.parse(await readFile(copied, "utf-8"))).toEqual(manifest);
   });
 
-  // cases: docs/engineering/testing/unit/record.md「Usage、facts 与失败命令证据落盘」
-  it("commands:落成 commands.json(不内联进 result.json),artifacts 含 commands;懒加载读回原样往返;没有非零命令的 attempt 恒 null;publish 缺省携带", async () => {
+  // cases: docs/engineering/testing/unit/record.md「命令证据落盘」
+  it("commands:落成 commands.json(不内联进 result.json),artifacts 含 commands;懒加载读回原样往返;没有 Sandbox 命令的 attempt 恒 null;publish 缺省携带", async () => {
     const root = await makeRoot();
     const writer = createWriter(root, { producer: { name: "niceeval", version: "0.12.0" } });
     const evidence = [
@@ -657,18 +717,19 @@ describe("createWriter", () => {
     expect(JSON.parse(await readFile(copied, "utf-8"))).toEqual(evidence);
   });
 
-  it("commands:超 256 KiB 的失败输出不参与逐值截断,全量原样往返(起因在前段、summary 在尾部两端都要在)", async () => {
-    const { ARTIFACT_VALUE_MAX_BYTES } = await import("./truncate.ts");
+  it("commands:stdout/stderr 各自超 64 KiB 时独立截断并打结构化 truncated 标记,未超限的流原样往返", async () => {
+    const { COMMAND_STREAM_MAX_BYTES } = await import("./truncate.ts");
     const root = await makeRoot();
     const writer = createWriter(root, { producer: { name: "niceeval", version: "0.12.0" } });
-    // 真实形状:头部是失败起因,中段是几百 KB 噪声,尾部是 runner 的 summary——截哪一端都毁掉另一半。
-    const hugeStdout = `E   assert 429 == 200\n${"collecting …\n".repeat(Math.ceil(ARTIFACT_VALUE_MAX_BYTES / 13))}2 failed, 14 passed in 3.41s\n`;
-    const hugeStderr = `Prepared 5 packages\n${"e".repeat(ARTIFACT_VALUE_MAX_BYTES + 500)}\nInstalled 5 packages`;
-    expect(Buffer.byteLength(hugeStdout, "utf-8")).toBeGreaterThan(ARTIFACT_VALUE_MAX_BYTES);
+    // 真实形状:stdout 超限、stderr 未超限——两条流各自独立判定,不因一条超限就一起截。
+    const hugeStdout = `E   assert 429 == 200\n${"collecting …\n".repeat(Math.ceil(COMMAND_STREAM_MAX_BYTES / 13))}2 failed, 14 passed in 3.41s\n`;
+    const shortStderr = "Prepared 5 packages\nInstalled 5 packages";
+    const hugeStdoutBytes = Buffer.byteLength(hugeStdout, "utf-8");
+    expect(hugeStdoutBytes).toBeGreaterThan(COMMAND_STREAM_MAX_BYTES);
 
     const snap = await writer.run({ experimentId: "huge/output", agent: "bub", startedAt: "2026-07-11T08:00:00.000Z" });
     const evidence = [
-      { timingNodeId: "n1", phase: "eval.run" as const, display: "uv run pytest", exitCode: 1, checked: true, stdout: hugeStdout, stderr: hugeStderr },
+      { timingNodeId: "n1", phase: "eval.run" as const, display: "uv run pytest", exitCode: 1, checked: true, stdout: hugeStdout, stderr: shortStderr },
     ];
     await snap.writeAttempt({ id: "q1", verdict: "errored", attempt: 1, durationMs: 10, assertions: [] }, { commands: evidence });
     await finishAll(writer);
@@ -676,10 +737,10 @@ describe("createWriter", () => {
     const results = await openRecord(root);
     const q1 = results.experiments[0].latestRun.attempts[0]!;
     const [readBack] = (await q1.commands())!;
-    expect(readBack.stdout).toBe(hugeStdout); // 逐字节相等,没有 marker 行、没有丢尾部 summary
-    expect(readBack.stderr).toBe(hugeStderr);
-    expect(JSON.stringify(readBack)).not.toContain("[niceeval] truncated");
-    expect((readBack as { truncated?: unknown }).truncated).toBeUndefined(); // 不再产出 truncated 字段
+    expect(Buffer.byteLength(readBack.stdout, "utf-8")).toBeLessThanOrEqual(COMMAND_STREAM_MAX_BYTES + 100);
+    expect(readBack.stdout).toContain(`[niceeval] truncated ${hugeStdoutBytes} →`);
+    expect(readBack.stderr).toBe(shortStderr); // 未超限的流逐字节原样往返,不被同伴截断牵连
+    expect(readBack.truncated).toEqual([{ path: "stdout", originalBytes: hugeStdoutBytes }]);
   });
 
   it("每个 Run 各自独立封口:两个 Experiment 各自不同的 completedAt 与 diagnostics 不串味,空 diagnostics 省略字段", async () => {

package/src/record/truncate.ts

@@ -1,30 +1,38 @@
 // 落盘大值截断(见 docs/feature/record/architecture.md「大值截断」)。
 // 运行时全量,落盘截断:落点唯一在 snap.writeAttempt(writer.ts)——不在 adapter、不在 OTLP
 // 解析、不在事件归一化里做。适用范围:events.json 的事件字段与 trace.json 的 span 属性里的
-// 任意字符串值;commands.json 不在适用范围内(失败诊断的完整语义单位,全量原样落盘)。
+// 任意字符串值,按 256 KiB 上限;commands.json 的 stdout/stderr 各自按 64 KiB 上限独立截断
+// (见「commandsjson」)——上限更小是因为现在成功命令也记录,体量不再天然有界于失败命令数。
 // 没有 flag、没有配置项;截断永远不影响判决(落盘是证据,不是评分输入)。
 
-import type { JsonValue, StreamEvent, TraceSpan, Truncation } from "../types.ts";
+import type { CommandExitEvidence, JsonValue, StreamEvent, TraceSpan, Truncation } from "../types.ts";
 
-/** 每个字符串值的落盘上限(UTF-8 字节)。 */
+/** events.json / trace.json 每个字符串值的落盘上限(UTF-8 字节)。 */
 export const ARTIFACT_VALUE_MAX_BYTES = 256 * 1024;
 
+/** commands.json 每条命令 stdout / stderr 各自的落盘上限(UTF-8 字节)。 */
+export const COMMAND_STREAM_MAX_BYTES = 64 * 1024;
+
 const encoder = new TextEncoder();
 
 /** 截断到 maxBytes(按 UTF-8 字符边界回退,不切断多字节字符),末尾追加人可读 marker。 */
-function truncateString(value: string, originalBytes: number): string {
+function truncateToBytes(value: string, originalBytes: number, maxBytes: number): string {
   // 按字节预算切:先按「字节数 ≥ 字符数」粗切,再逐步回退到不超预算的字符边界。
-  let sliceLen = Math.min(value.length, ARTIFACT_VALUE_MAX_BYTES);
+  let sliceLen = Math.min(value.length, maxBytes);
   let head = value.slice(0, sliceLen);
-  while (encoder.encode(head).length > ARTIFACT_VALUE_MAX_BYTES && sliceLen > 0) {
+  while (encoder.encode(head).length > maxBytes && sliceLen > 0) {
     // 超出量 ÷ 4(UTF-8 最长 4 字节)是安全的最小回退步长;循环兜底处理边界。
-    sliceLen -= Math.max(1, Math.ceil((encoder.encode(head).length - ARTIFACT_VALUE_MAX_BYTES) / 4));
+    sliceLen -= Math.max(1, Math.ceil((encoder.encode(head).length - maxBytes) / 4));
     head = value.slice(0, sliceLen);
   }
   const kept = encoder.encode(head).length;
   return `${head}\n[niceeval] truncated ${originalBytes} → ${kept} bytes`;
 }
 
+function truncateString(value: string, originalBytes: number): string {
+  return truncateToBytes(value, originalBytes, ARTIFACT_VALUE_MAX_BYTES);
+}
+
 /**
  * 深度遍历一个 JSON 值,截断其中超限的字符串;返回(可能新建的)值与收集到的截断记录。
  * `path` 用点分段(数组下标同样入段),事件里从字段名起、span 属性里从 attribute key 起。
@@ -89,3 +97,20 @@ export function truncateSpans(spans: readonly TraceSpan[]): TraceSpan[] {
     return { ...span, attributes, truncated: [...(span.truncated ?? []), ...out] };
   });
 }
+
+/** commands.json 落盘前的截断:stdout / stderr 各自独立按 `COMMAND_STREAM_MAX_BYTES` 截断。 */
+export function truncateCommands(commands: readonly CommandExitEvidence[]): CommandExitEvidence[] {
+  return commands.map((command) => {
+    const out: Truncation[] = [];
+    const next: { stdout?: string; stderr?: string } = {};
+    for (const path of ["stdout", "stderr"] as const) {
+      const value = command[path];
+      const bytes = encoder.encode(value).length;
+      if (bytes <= COMMAND_STREAM_MAX_BYTES) continue;
+      out.push({ path, originalBytes: bytes });
+      next[path] = truncateToBytes(value, bytes, COMMAND_STREAM_MAX_BYTES);
+    }
+    if (out.length === 0) return command;
+    return { ...command, ...next, truncated: [...(command.truncated ?? []), ...out] };
+  });
+}

package/src/record/writer.ts

@@ -17,7 +17,7 @@ import { RESULT_FILE, RUN_FILE, artifactFileOf, attemptDirOf, experimentDirOf }
 import { encodeAttemptLocator } from "./locator.ts";
 import { MANIFESTS_FILE, type RunManifests } from "./manifest.ts";
 import { hashEvalSource, normalizeEvalSource } from "./source-hash.ts";
-import { truncateEvents, truncateSpans } from "./truncate.ts";
+import { truncateCommands, truncateEvents, truncateSpans } from "./truncate.ts";
 import type { Producer, RunMeta } from "./types.ts";
 
 export interface WriterOptions {
@@ -48,6 +48,13 @@ export interface WriterOptions {
    * `RunDeclaration.manifests`,后者优先。
    */
   manifests?: ReadonlyMap<string, RunManifests>;
+  /**
+   * 本次 invocation 规划期算出的 Run 级配置身份,按 experimentId 分组(见
+   * `runner/types.ts` 的 `InvocationShape.configHashes`)。隐式 run 声明
+   * (`writeAttemptFor`)按 experimentId 取用;显式 `run()` 声明用 `RunDeclaration.configHash`,
+   * 后者优先——`niceeval accept` 走的正是这条显式路径。
+   */
+  configHashes?: ReadonlyMap<string, string>;
 }
 
 /** 快照级声明:一个 experiment 声明一次,这些字段不塞进每条 attempt。 */
@@ -130,6 +137,14 @@ export interface RunWriter {
     timings?: TimingActivity[];
     sandboxBuilds?: SandboxBuildRecord[];
     name?: LocalizedText;
+    /**
+     * 本 Run 携带合入的 eval id(见 `experiment:complete` 事件的 `carriedResults`)。并入
+     * `experiment.selectedEvalIds`,使这份快照声明的覆盖题集包含携带条目——`selectedEvalIds`
+     * 的语义是「这份快照声明覆盖的题集」(本次选择 ∪ 携带条目),不止是本次选择器求值的结果
+     * (见 docs/feature/record/architecture.md「selectedEvalIds」)。省略或没有 `experiment`
+     * 声明时不产生任何效果。
+     */
+    carriedEvalIds?: readonly string[];
   }): Promise<void>;
 }
 
@@ -152,6 +167,21 @@ export interface Writer {
   snapshotWriters(): Promise<{ experimentId: string; writer: RunWriter }[]>;
 }
 
+/**
+ * `finish()` 的 `carriedEvalIds` 并入 `experiment.selectedEvalIds`;没有 `experiment` 声明
+ * 或没有携带 id 时原样返回,不凭空合成一份 `ExperimentRunInfo`。
+ */
+function mergeCarriedEvalIds(
+  experiment: ExperimentRunInfo | undefined,
+  carriedEvalIds: readonly string[] | undefined,
+): ExperimentRunInfo | undefined {
+  if (experiment === undefined || !carriedEvalIds?.length) return experiment;
+  const merged = new Set(experiment.selectedEvalIds);
+  for (const id of carriedEvalIds) merged.add(id);
+  if (merged.size === experiment.selectedEvalIds.length) return experiment;
+  return { ...experiment, selectedEvalIds: [...merged] };
+}
+
 interface SnapshotState {
   /** 快照的权威 meta(不含 completedAt;knownEvalIds 随重复声明累加)。 */
   meta: RunMeta;
@@ -173,6 +203,7 @@ export function createWriter(root: string, opts: WriterOptions): Writer {
   const created: { experimentId: string; dir: string }[] = [];
 
   async function buildSnapshot(decl: RunDeclaration): Promise<SnapshotHandle> {
+    const configHash = decl.configHash ?? opts.configHashes?.get(decl.experimentId);
     const meta: RunMeta = {
       format: RECORD_FORMAT,
       schemaVersion: RECORD_SCHEMA_VERSION,
@@ -184,7 +215,7 @@ export function createWriter(root: string, opts: WriterOptions): Writer {
       agent: decl.agent,
       ...(decl.model !== undefined ? { model: decl.model } : {}),
       startedAt: decl.startedAt,
-      ...(decl.configHash !== undefined ? { configHash: decl.configHash } : {}),
+      ...(configHash !== undefined ? { configHash } : {}),
       ...(decl.knownEvalIds?.length ? { knownEvalIds: [...new Set(decl.knownEvalIds)] } : {}),
       ...(decl.name !== undefined ? { name: decl.name } : {}),
     };
@@ -221,13 +252,14 @@ export function createWriter(root: string, opts: WriterOptions): Writer {
         state.finished = true;
         const completedAt = finishOpts?.completedAt ?? state.declCompletedAt ?? new Date().toISOString();
         const name = finishOpts?.name ?? state.declName;
+        const experiment = mergeCarriedEvalIds(state.meta.experiment, finishOpts?.carriedEvalIds);
         const finalMeta: RunMeta = {
           format: state.meta.format,
           schemaVersion: state.meta.schemaVersion,
           producer: state.meta.producer,
           runId: state.meta.runId,
           experimentId: state.meta.experimentId,
-          ...(state.meta.experiment !== undefined ? { experiment: state.meta.experiment } : {}),
+          ...(experiment !== undefined ? { experiment } : {}),
           agent: state.meta.agent,
           ...(state.meta.model !== undefined ? { model: state.meta.model } : {}),
           startedAt: state.meta.startedAt,
@@ -394,12 +426,11 @@ async function writeAttemptFiles(
 
   const writes: Promise<void>[] = [];
   // 大值截断只发生在这里(序列化的那一刻):events 的事件字段与 trace 的 span 属性里的任意
-  // 字符串值按 ARTIFACT_VALUE_MAX_BYTES 截断并留结构化 truncated 标记;运行时(断言 / o11y
-  // 派生)看到的始终是完整值。commands、sources 与 diff 不截断——失败命令的 stdout/stderr 是
-  // 一个完整的诊断语义单位,截哪一端都毁掉另一半(见 docs/feature/record/architecture.md
-  // 的证据 registry)。
+  // 字符串值按 ARTIFACT_VALUE_MAX_BYTES 截断,commands 的 stdout/stderr 各自按
+  // COMMAND_STREAM_MAX_BYTES 截断,均留结构化 truncated 标记;运行时(断言 / o11y 派生)看到的
+  // 始终是完整值。sources 与 diff 不截断(见 docs/feature/record/architecture.md 的证据 registry)。
   if (hasCommands)
-    writes.push(writeFile(join(attemptDir, artifactFileOf("commands")), JSON.stringify(artifacts!.commands!), "utf-8"));
+    writes.push(writeFile(join(attemptDir, artifactFileOf("commands")), JSON.stringify(truncateCommands(artifacts!.commands!)), "utf-8"));
   if (hasEvents)
     writes.push(writeFile(join(attemptDir, "events.json"), JSON.stringify(truncateEvents(artifacts!.events!)), "utf-8"));
   if (hasSources) writes.push(writeSourcesRef(snapDir, attemptDir, artifacts!.sources!, sourceStore));

package/src/report/assets/series-encoding.tsx

@@ -0,0 +1,177 @@
+// 24 视觉身份 = 六色 × 四变体的可直接使用 SVG/CSS 值。
+// 契约:docs/feature/reports/library/presentation.md「实验颜色与维度呈现」、
+// docs/feature/reports/components/README.md「视觉编码容量（24 个身份）」。
+// pattern id 全局唯一、颜色走主题令牌 --niceeval-color-series-N,换 basalt/chalk 自动换色。
+
+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 function seriesColorVar(colorIndex: number): string {
+  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: number, variant: SeriesVariant): string {
+  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: number, variant: SeriesVariant): string {
+  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: SeriesVariant): string {
+  switch (variant) {
+    case 1:
+      return "";
+    case 2:
+      return "6 4";
+    case 3:
+      return "2 3";
+    case 4:
+      return "8 3 2 3";
+  }
+}
+
+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 function seriesMarkerShape(variant: SeriesVariant): SeriesMarkerShape {
+  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: number, variant: SeriesVariant): {
+  readonly path: string;
+  readonly viewBox: string;
+  readonly fill: string;
+  readonly stroke: string;
+} {
+  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: string): string {
+  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: string): string {
+  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(): ReactNode {
+  const patterns: ReactNode[] = [];
+  for (let colorIndex = 1; colorIndex <= 6; colorIndex++) {
+    const color = seriesColorVar(colorIndex);
+    // v2:对角斜线
+    patterns.push(
+      <pattern
+        key={`v2-c${colorIndex}`}
+        id={seriesPatternId(colorIndex, 2)}
+        patternUnits="userSpaceOnUse"
+        width="6"
+        height="6"
+      >
+        <rect width="6" height="6" fill={color} fillOpacity={0.22} />
+        <path
+          d="M-1 1l2-2M0 6l6-6M5 7l2-2"
+          stroke={color}
+          strokeWidth={1.4}
+          fill="none"
+        />
+      </pattern>,
+    );
+    // v3:水平条纹
+    patterns.push(
+      <pattern
+        key={`v3-c${colorIndex}`}
+        id={seriesPatternId(colorIndex, 3)}
+        patternUnits="userSpaceOnUse"
+        width="6"
+        height="6"
+      >
+        <rect width="6" height="6" fill={color} fillOpacity={0.18} />
+        <path d="M0 1.5h6M0 4.5h6" stroke={color} strokeWidth={1.5} fill="none" />
+      </pattern>,
+    );
+    // v4:点阵
+    patterns.push(
+      <pattern
+        key={`v4-c${colorIndex}`}
+        id={seriesPatternId(colorIndex, 4)}
+        patternUnits="userSpaceOnUse"
+        width="6"
+        height="6"
+      >
+        <rect width="6" height="6" fill={color} fillOpacity={0.16} />
+        <circle cx="2" cy="2" r="1.15" fill={color} />
+        <circle cx="5" cy="5" r="1.15" fill={color} />
+      </pattern>,
+    );
+  }
+  return (
+    <svg
+      className="niceeval-series-defs"
+      width={0}
+      height={0}
+      aria-hidden="true"
+      focusable="false"
+      style={{ position: "absolute", width: 0, height: 0, overflow: "hidden" }}
+    >
+      <defs>{patterns}</defs>
+    </svg>
+  );
+}