niceeval 0.7.1 → 0.8.1

package/src/view/app/components/CodeView.test.tsx

@@ -0,0 +1,142 @@
+// @vitest-environment jsdom
+// cases: docs/engineering/unit-tests/reports/cases.md
+// 「Attempt 详情(view 证据室)」分区——
+// 源码视图是判定与断言的单点:带 loc 的 send 行可点开查看该轮回复(assistant 文本 / thinking),
+// 失败断言行默认展开、展开面直接给 matcher 与 expected / received 的值。
+// 这组测试在 jsdom 里跑真实点击,守住「统一站点管线之后弹窗证据链仍然可用」的行为面
+// (数据侧的字节奇偶由 src/view/site-parity.test.ts 守护,两头合起来 = 线上站点可用)。
+
+import { act } from "react";
+import { createRoot, type Root } from "react-dom/client";
+import { afterEach, beforeAll, describe, expect, it } from "vitest";
+import { CodeView } from "./CodeView.tsx";
+import { makeTranslator } from "../i18n.ts";
+import type { Assertion, CodeSource, TranscriptEvent } from "../types.ts";
+
+declare global {
+  // eslint-disable-next-line no-var
+  var IS_REACT_ACT_ENVIRONMENT: boolean;
+}
+
+beforeAll(() => {
+  globalThis.IS_REACT_ACT_ENVIRONMENT = true;
+});
+
+const t = makeTranslator("en");
+
+const FILE = "evals/a.eval.ts";
+const SOURCE: CodeSource = {
+  path: FILE,
+  content: [
+    "import { defineEval } from \"niceeval\";",
+    "",
+    "await t.send(\"do the task\");",
+    "",
+    "t.check(source, includes(/use cache/));",
+  ].join("\n"),
+};
+
+const EVENTS: TranscriptEvent[] = [
+  { type: "message", role: "user", text: "do the task", loc: { file: FILE, line: 3 } },
+  { type: "thinking", text: "THINKING_MARKER" },
+  { type: "message", role: "assistant", text: "REPLY_TEXT_MARKER" },
+];
+
+const FAILED_ASSERT: Assertion = {
+  name: "Catalog reads use cache",
+  detail: "includes(/use cache/)",
+  severity: "gate",
+  outcome: "failed",
+  score: 0,
+  expected: "/use cache/",
+  received: "RECEIVED_VALUE_MARKER",
+  loc: { file: FILE, line: 5 },
+};
+
+let container: HTMLElement | undefined;
+let root: Root | undefined;
+
+function render(ui: React.ReactElement): HTMLElement {
+  container = document.createElement("div");
+  document.body.appendChild(container);
+  root = createRoot(container);
+  act(() => root!.render(ui));
+  return container;
+}
+
+afterEach(() => {
+  if (root) act(() => root!.unmount());
+  container?.remove();
+  root = undefined;
+  container = undefined;
+});
+
+function click(el: Element): void {
+  act(() => {
+    el.dispatchEvent(new MouseEvent("click", { bubbles: true }));
+  });
+}
+
+/** 按行号取源码行的 DOM 节点(.code-line 里第一个 .ln 是行号)。 */
+function lineEl(host: HTMLElement, n: number): Element {
+  const row = [...host.querySelectorAll(".code-line")].find((el) => el.querySelector(".ln")?.textContent === String(n));
+  expect(row, `source line ${n}`).toBeTruthy();
+  return row!;
+}
+
+describe("CodeView · send 行的回复展开", () => {
+  it("带 loc 的 send 行可点开:回复面板显示 assistant 文本与 thinking;再点收起", () => {
+    const host = render(<CodeView sources={[SOURCE]} events={EVENTS} assertions={[]} t={t} />);
+
+    const sendLine = lineEl(host, 3);
+    expect(sendLine.className).toContain("line-send");
+    // 初始未展开:回复不可见。
+    expect(host.textContent).not.toContain("REPLY_TEXT_MARKER");
+
+    click(sendLine);
+    expect(host.querySelector(".reply-panel")).toBeTruthy();
+    expect(host.textContent).toContain("REPLY_TEXT_MARKER");
+    expect(host.textContent).toContain("THINKING_MARKER");
+
+    click(lineEl(host, 3));
+    expect(host.textContent).not.toContain("REPLY_TEXT_MARKER");
+  });
+
+  it("send 轮没有任何回复事件时,展开面如实显示「无回复」而不是空白", () => {
+    const onlySend: TranscriptEvent[] = [EVENTS[0]!];
+    const host = render(<CodeView sources={[SOURCE]} events={onlySend} assertions={[]} t={t} />);
+    click(lineEl(host, 3));
+    expect(host.querySelector(".reply-empty")?.textContent).toBe(t("code.noReply"));
+  });
+});
+
+describe("CodeView · 断言行的明细展开", () => {
+  it("第一条失败断言默认展开:matcher 与 expected / received 的值直接可见,点行可收起", () => {
+    const host = render(<CodeView sources={[SOURCE]} events={EVENTS} assertions={[FAILED_ASSERT]} t={t} />);
+
+    const assertLine = lineEl(host, 5);
+    expect(assertLine.className).toContain("line-fail");
+    // 默认展开(第一条 failed):不点任何东西就能看到为什么失败。
+    expect(host.textContent).toContain("includes(/use cache/)");
+    expect(host.textContent).toContain("/use cache/");
+    expect(host.textContent).toContain("RECEIVED_VALUE_MARKER");
+
+    click(assertLine);
+    expect(host.textContent).not.toContain("RECEIVED_VALUE_MARKER");
+  });
+
+  it("passed 断言行不默认展开,点开后显示明细", () => {
+    const passed: Assertion = { ...FAILED_ASSERT, outcome: "passed", score: 1, loc: { file: FILE, line: 5 } };
+    delete (passed as { expected?: string }).expected;
+    delete (passed as { received?: string }).received;
+    const host = render(<CodeView sources={[SOURCE]} events={EVENTS} assertions={[passed]} t={t} />);
+
+    const assertLine = lineEl(host, 5);
+    expect(assertLine.className).toContain("line-pass");
+    expect(host.querySelector(".line-detail")).toBeNull();
+
+    click(assertLine);
+    expect(host.querySelector(".line-detail")).toBeTruthy();
+    expect(host.textContent).toContain("includes(/use cache/)");
+  });
+});

package/src/view/app/components/CodeView.tsx

@@ -4,7 +4,7 @@ import type { T } from "../shared.ts";
 import type { Assertion, CodeSource, SourceTurn, TranscriptEvent } from "../types.ts";
 import { highlightTs, indexAsserts, indexTurns, locKey } from "../lib/transcript-data.tsx";
 import { formatScore } from "../lib/format.ts";
-import { InputBlock, ToolBlock, Transcript } from "./Transcript.tsx";
+import { InputBlock, RawEventBlock, ToolBlock, Transcript } from "./Transcript.tsx";
 
 /** soft 断言没过阈值不影响 verdict,颜色上跟 gate 失败(红)区分开,用 warn(黄);
  *  unavailable 用独立第三态(非红非绿)。 */
@@ -209,6 +209,13 @@ export function ReplyPanel({ turn, t }: { turn: SourceTurn; t: T }) {
               <div className="reply-text">{r.text}</div>
             </div>
           );
+        if (r.kind === "user")
+          return (
+            <div key={j} className="reply-user">
+              <span className="reply-role">{t("transcript.user")}</span>
+              <div className="reply-text">{r.text}</div>
+            </div>
+          );
         if (r.kind === "thinking")
           return (
             <details key={j} className="reply-think">
@@ -218,6 +225,13 @@ export function ReplyPanel({ turn, t }: { turn: SourceTurn; t: T }) {
           );
         if (r.kind === "error")
           return <div key={j} className="reply-err">! {r.text}</div>;
+        if (r.kind === "skill")
+          return (
+            <div key={j} className="reply-skill">
+              <span className="reply-role">{t("transcript.skillLoaded")}</span> {r.skill}
+            </div>
+          );
+        if (r.kind === "raw") return <RawEventBlock key={j} raw={r.raw} t={t} />;
         if (r.kind === "tool")
           // 和 Transcript 同一个组件:摘要行显示工具名(入参)→ 出参预览,展开看完整出入参。
           return <ToolBlock key={j} call={r.ev} result={r.result} t={t} />;

package/src/view/app/components/Transcript.tsx

@@ -1,5 +1,5 @@
 import type { T, ToolBlockCall } from "../shared.ts";
-import type { ToolResultEvent, TranscriptEvent } from "../types.ts";
+import type { ObjectRecord, ToolResultEvent, TranscriptEvent } from "../types.ts";
 import { TOOL_VERB, resultBody, toolPrimaryArg } from "../lib/transcript-data.tsx";
 import { prettyJson, previewText, truncate } from "../lib/format.ts";
 
@@ -42,6 +42,15 @@ export function Transcript({ events, t }: { events: TranscriptEvent[]; t: T }) {
             );
           case "input.requested":
             return <InputBlock event={event} t={t} key={index} />;
+          case "skill.loaded":
+            return (
+              <div className="ts-skill" key={index}>
+                <span className="ts-role">{t("transcript.skillLoaded")}</span>
+                <div className="ts-text">{event.skill}</div>
+              </div>
+            );
+          case "view.raw":
+            return <RawEventBlock raw={event.raw} t={t} key={index} />;
           case "compaction":
             return (
               <div className="ts-compaction" key={index}>
@@ -62,6 +71,24 @@ export function Transcript({ events, t }: { events: TranscriptEvent[]; t: T }) {
   );
 }
 
+/** 未识别事件的原样展示:摘要行带原始 type,展开是完整 JSON——不静默丢,方便发现新词汇后续补一等呈现。 */
+export function RawEventBlock({ raw, t }: { raw: ObjectRecord; t: T }) {
+  const type = typeof raw.type === "string" ? raw.type : "?";
+  const body = prettyJson(raw);
+  return (
+    <details className="ts-tool-d ts-raw">
+      <summary className="ts-row">
+        <span className="ts-dot pending" />
+        <span className="ts-tool">{type}</span>
+        <span className="ts-preview">{t("transcript.rawEvent")}</span>
+      </summary>
+      <div className="ts-body">
+        <pre className="attr-pre">{truncate(body, 8000)}</pre>
+      </div>
+    </details>
+  );
+}
+
 export function MessageBlock({ event, t }: { event: Extract<TranscriptEvent, { type: "message" }>; t: T }) {
   const who = event.role === "assistant" ? "assistant" : "user";
   return (

package/src/view/app/i18n.ts

@@ -88,6 +88,8 @@ export type MessageKey =
   | "transcript.inputRequested"
   | "transcript.awaitingInput"
   | "transcript.contextCompacted"
+  | "transcript.skillLoaded"
+  | "transcript.rawEvent"
   | "transcript.running"
   | "transcript.input"
   | "transcript.output"
@@ -218,6 +220,8 @@ const dictionaries: Record<Locale, Dictionary> = {
     "transcript.inputRequested": "input requested",
     "transcript.awaitingInput": "(awaiting input)",
     "transcript.contextCompacted": "context compacted",
+    "transcript.skillLoaded": "skill loaded",
+    "transcript.rawEvent": "unrecognized event, shown as-is",
     "transcript.running": "running...",
     "transcript.input": "input",
     "transcript.output": "output",
@@ -343,6 +347,8 @@ const dictionaries: Record<Locale, Dictionary> = {
     "transcript.inputRequested": "请求输入",
     "transcript.awaitingInput": "(等待输入)",
     "transcript.contextCompacted": "上下文已压缩",
+    "transcript.skillLoaded": "已加载 Skill",
+    "transcript.rawEvent": "未识别事件,原样展示",
     "transcript.running": "运行中...",
     "transcript.input": "输入",
     "transcript.output": "输出",

package/src/view/app/lib/guards.test.ts

@@ -0,0 +1,108 @@
+// cases: docs/engineering/unit-tests/reports/cases.md
+// 「Attempt 详情」:事件流按条目校验、按条目容错;skill.loaded 是一等回复条目。
+// bug: memory/view-unknown-event-type-drops-whole-transcript.md
+import { describe, expect, it } from "vitest";
+import { asEvents } from "./guards.ts";
+import { indexTurns, locKey } from "./transcript-data.tsx";
+import type { TranscriptEvent } from "../types.ts";
+
+const send: TranscriptEvent = {
+  type: "message",
+  role: "user",
+  text: "Migrate every route",
+  loc: { file: "evals/m.eval.ts", line: 37, column: 8 },
+};
+
+describe("asEvents 按条目容错", () => {
+  it("含 skill.loaded 的 events 数组不被判空,全部事件保留", () => {
+    const raw = [
+      send,
+      { type: "skill.loaded", skill: "pdf-export", callId: "c1" },
+      { type: "message", role: "assistant", text: "done" },
+    ];
+    const events = asEvents(raw);
+    expect(events).toHaveLength(3);
+  });
+
+  it("未知事件类型包成 view.raw 原样保留,其余事件照常", () => {
+    const unknown = { type: "future.event", payload: 1 };
+    const events = asEvents([send, unknown, { type: "thinking", text: "hm" }]);
+    expect(events?.map((e) => e.type)).toEqual(["message", "view.raw", "thinking"]);
+    expect(events?.[1]).toEqual({ type: "view.raw", raw: unknown });
+  });
+
+  it("形状不合的已知类型同样包成 view.raw,不静默丢", () => {
+    const malformed = { type: "message", role: "assistant" }; // 缺 text
+    const events = asEvents([malformed]);
+    expect(events).toEqual([{ type: "view.raw", raw: malformed }]);
+  });
+
+  it("非对象条目丢弃;非数组载荷整体拒绝", () => {
+    expect(asEvents(["junk", 42, send])).toHaveLength(1);
+    expect(asEvents({ events: [] })).toBeNull();
+    expect(asEvents("nope")).toBeNull();
+  });
+});
+
+describe("skill.loaded 聚合进 send 的回复", () => {
+  it("indexTurns 聚出 kind: skill 回复并保留 Skill 名", () => {
+    const turns = indexTurns([
+      send,
+      { type: "skill.loaded", skill: "pdf-export", callId: "c1" },
+      { type: "message", role: "assistant", text: "done" },
+    ]);
+    const turn = turns.byKey.get(locKey("evals/m.eval.ts", 37));
+    expect(turn?.replies).toEqual([
+      { kind: "skill", skill: "pdf-export" },
+      { kind: "text", text: "done" },
+    ]);
+  });
+
+  it("view.raw 条目聚成 kind: raw 回复,原始载荷不丢", () => {
+    const turns = indexTurns([send, { type: "view.raw", raw: { type: "future.event", payload: 1 } }]);
+    const turn = turns.byKey.get(locKey("evals/m.eval.ts", 37));
+    expect(turn?.replies).toEqual([{ kind: "raw", raw: { type: "future.event", payload: 1 } }]);
+  });
+});
+
+describe("轮归属按 loc 判定,无 loc 的 user 消息不开新轮", () => {
+  it("send 后紧跟的同文本无 loc 回显被吃掉,回复仍全部聚到 send 行", () => {
+    const turns = indexTurns([
+      send,
+      { type: "message", role: "user", text: send.text },
+      { type: "thinking", text: "plan" },
+      { type: "message", role: "assistant", text: "done" },
+    ]);
+    const turn = turns.byKey.get(locKey("evals/m.eval.ts", 37));
+    expect(turn?.replies).toEqual([
+      { kind: "thinking", text: "plan" },
+      { kind: "text", text: "done" },
+    ]);
+    expect(turns.noloc).toHaveLength(0);
+  });
+
+  it("轮中段的 stop-hook 反馈成为 kind: user 回复,其后的 assistant 回复不脱轮", () => {
+    const turns = indexTurns([
+      send,
+      { type: "message", role: "user", text: send.text },
+      { type: "message", role: "assistant", text: "first" },
+      { type: "message", role: "user", text: "Stop hook feedback: save notes" },
+      { type: "message", role: "assistant", text: "saved" },
+    ]);
+    const turn = turns.byKey.get(locKey("evals/m.eval.ts", 37));
+    expect(turn?.replies).toEqual([
+      { kind: "text", text: "first" },
+      { kind: "user", text: "Stop hook feedback: save notes" },
+      { kind: "text", text: "saved" },
+    ]);
+  });
+
+  it("流首无 loc 的 user 消息(旧工件)仍开 noloc 轮", () => {
+    const turns = indexTurns([
+      { type: "message", role: "user", text: "hi" },
+      { type: "message", role: "assistant", text: "hello" },
+    ]);
+    expect(turns.noloc).toHaveLength(1);
+    expect(turns.noloc[0]?.replies).toEqual([{ kind: "text", text: "hello" }]);
+  });
+});

package/src/view/app/lib/guards.ts

@@ -1,4 +1,4 @@
-import type { CodeSource, ObjectRecord, Span, TranscriptEvent } from "../types.ts";
+import type { CodeSource, KnownTranscriptEvent, ObjectRecord, Span, TranscriptEvent } from "../types.ts";
 
 export function asSources(value: unknown): CodeSource[] | null {
   if (!Array.isArray(value)) return null;
@@ -11,7 +11,15 @@ export function isCodeSource(value: unknown): value is CodeSource {
 
 export function asEvents(value: unknown): TranscriptEvent[] | null {
   if (!Array.isArray(value)) return null;
-  return value.every(isTranscriptEvent) ? value : null;
+  // 事件词汇会演进(skill.loaded 就是先例):未识别或缺字段的条目不整体判空、
+  // 也不静默丢弃——包成 view.raw 原样展示,让新词汇在界面上可被发现、后续补一等呈现。
+  // 全有全无判空会让源码视图的 send 行连回复入口都消失;只有非对象条目没有可展示的结构才丢。
+  const out: TranscriptEvent[] = [];
+  for (const item of value) {
+    if (isKnownTranscriptEvent(item)) out.push(item);
+    else if (isObjectRecord(item)) out.push({ type: "view.raw", raw: item });
+  }
+  return out;
 }
 
 export function asSpans(value: unknown): Span[] | null {
@@ -19,7 +27,7 @@ export function asSpans(value: unknown): Span[] | null {
   return value.every(isSpan) ? value : null;
 }
 
-export function isTranscriptEvent(value: unknown): value is TranscriptEvent {
+export function isKnownTranscriptEvent(value: unknown): value is KnownTranscriptEvent {
   if (!isObjectRecord(value) || typeof value.type !== "string") return false;
   switch (value.type) {
     case "message":
@@ -32,6 +40,8 @@ export function isTranscriptEvent(value: unknown): value is TranscriptEvent {
       return typeof value.callId === "string" && typeof value.name === "string";
     case "subagent.completed":
       return typeof value.callId === "string";
+    case "skill.loaded":
+      return typeof value.skill === "string";
     case "input.requested":
       return isObjectRecord(value.request);
     case "thinking":

package/src/view/app/lib/transcript-data.tsx

@@ -22,6 +22,16 @@ export function indexTurns(events: TranscriptEvent[]): IndexedTurns {
   let cur: SourceTurn | null = null;
   for (const ev of events || []) {
     if (ev.type === "message" && ev.role === "user") {
+      // 同一条 send 会在流里出现两次:runner 在 send 时记带 loc 的一条,agent 原生
+      // transcript 又回显同文本、无 loc 的一条。无 loc 的 user 消息不开新轮——回显
+      // (与当前轮 sent 同文本且回复还没开始)直接吃掉,其它(stop-hook 反馈、skill
+      // 注入等轮内注入)作为回复条目留在当前轮;否则回复全被回显轮抢走,
+      // 带 loc 的 send 行只剩「(无回复)」。
+      if (!ev.loc && cur) {
+        if (cur.replies.length === 0 && (ev.text || "") === cur.sent) continue;
+        cur.replies.push({ kind: "user", text: ev.text || "" });
+        continue;
+      }
       cur = { loc: ev.loc, sent: ev.text || "", replies: [] };
       if (ev.loc) byKey.set(locKey(ev.loc.file, ev.loc.line), cur);
       else noloc.push(cur);
@@ -38,6 +48,10 @@ export function indexTurns(events: TranscriptEvent[]): IndexedTurns {
     } else if (ev.type === "action.result") {
       const tool = toolByCallId.get(ev.callId);
       if (tool) tool.result = ev;
+    } else if (ev.type === "skill.loaded") {
+      cur.replies.push({ kind: "skill", skill: ev.skill });
+    } else if (ev.type === "view.raw") {
+      cur.replies.push({ kind: "raw", raw: ev.raw });
     } else if (ev.type === "input.requested") {
       cur.replies.push({ kind: "input", ev });
     } else if (ev.type === "error") {

package/src/view/app/types.ts

@@ -41,8 +41,11 @@ export interface SourceTurn {
 
 export type Reply =
   | { kind: "text"; text: string }
+  | { kind: "user"; text: string }
   | { kind: "thinking"; text: string }
   | { kind: "error"; text: string }
+  | { kind: "skill"; skill: string }
+  | { kind: "raw"; raw: ObjectRecord }
   | { kind: "tool"; ev: ActionCalledEvent; result?: ActionResultEvent }
   | { kind: "input"; ev: InputRequestedEvent };
 
@@ -55,7 +58,20 @@ export type ToolResultEvent = ActionResultEvent | SubagentCompletedEvent;
 export type ViewJson = JsonValue;
 export type ViewUsage = Usage;
 
-export type TranscriptEvent = StreamEvent;
+/** 标准事件流里前端认识的词汇。 */
+export type KnownTranscriptEvent = StreamEvent;
+
+/**
+ * 未识别或形状不合的事件条目,由 asEvents 包装:不静默丢弃,渲染面原样展示
+ * (带原始 type 标签),让词汇演进在界面上可被发现、后续补一等呈现。
+ * `view.raw` 是 view 内部标记,不属于标准事件流词汇。
+ */
+export interface RawTranscriptEvent {
+  type: "view.raw";
+  raw: ObjectRecord;
+}
+
+export type TranscriptEvent = KnownTranscriptEvent | RawTranscriptEvent;
 
 export interface Indexed<T> {
   byKey: Map<string, T[]>;

package/src/view/artifact-serving.test.ts

@@ -119,7 +119,7 @@ describe("index.ts · copyFetchedArtifacts(--out 静态导出)对 sources.json
     await seedDedupedSnapshot(root);
 
     const out = join(root, "site");
-    await buildView({ input: root, out , allowSensitiveArtifacts: true });
+    await buildView({ input: root, out  });
 
     const scan = await loadViewScan(root);
     const byId = new Map(scan.viewData.snapshots.flatMap((s) => s.results.map((r) => [r.id, r])));