fix(web): keep rectification interview planning out of the spoken answer

Thinking is disabled, so Chinese process talk after tools arrives as
text-delta. Widen the leak-net so tool-label echoes and internal field
names stay out of answer.delta, and keep the user-facing follow-up.

Co-authored-by: Cursor <cursoragent@cursor.com>

docs/BUG_HISTORY.md

@@ -5415,7 +5415,7 @@
 - 修复：保持 `thinking: disabled` 与 16384 正文预算。对已加载 Case 的 `text-delta` 按段落拆成过程自述与口语结论：自述进 `thinking.delta`，结论进 `answer.delta` 并作为落盘正文。界面与咨询对齐：折叠「思考」（13px 次要色，有正文后默认收起）和完整「回复」。
 - 验证：`frontend/tests/rectification-spoken-answer.test.ts`、`frontend/tests/rectification-v9-stream.test.ts`、`frontend/tests/rectification-agentic-entry.test.ts`、`frontend/tests/chat-stream-layout.test.ts`；`tsc --noEmit` 通过。口语/思考拆分的 `emit` 类型必须与 `RunV9AgentTurnInput.emit` 同为 `Promise<void> | void`，否则 Docker `next build` 会在 publish 阶段失败。
 - 防复发：纠正组答不得在不拆开正文预算的情况下重开 provider thinking 来充当思考 UI。`answer.delta` 不得承载 skill/schema/batch 过程自述。思考与回复必须分层渲染，不得共用同一段正文样式。
-- 相关记录：BUG-345、BUG-354、BUG-356
+- 相关记录：BUG-345、BUG-354、BUG-356、BUG-373
 - 复发自：BUG-345（关 thinking 后自述改走正文）、BUG-354（纠正与咨询一并关闭 provider thinking）
 - 修复版本：`f03a2706`
 
@@ -5531,6 +5531,22 @@
 - 复发自：无
 - 修复版本：未修复
 
+## BUG-373 | 生时纠正把中文过程自述当成回答，工具完成标签和 occupation_note 漏进正文
+
+- 状态：resolved
+- 首次发现：2026-08-25
+- 最近更新：2026-08-25
+- 影响面：`POST /api/rectification/agent`、`splitRectificationSpokenAndThinking`、已落盘 Turn 水合
+- 用户现象：工具进度（读取校正记录、整理证据、比较候选）之后，回复气泡里先出现大段自我规划：方法覆盖、occupation_note、open_question、还不能出牌、本轮对照了哪些分盘，最后才是对用户的一句追问。看起来像思考过程被当成推理结果。
+- 触发条件：provider thinking 关闭；模型在工具完成后的最终 `text-delta` 里先写中文过程自述，再写对用户的问题。
+- 根因：BUG-368 维持 `thinking: disabled`。漏检网只认旧的 `datePrecision` / `用户提到` / `让我调用 batch` 句式。新的规划句复制了公开工具完成标签和服务器字段名，整段被当成口语并落盘。末尾「本轮对照了…」与界面已有的技法句重复。
+- 修复：漏检网补上工具完成标签回声、occupation_note / method_followup_plan / open_question / not_separated 等内部字段，以及方法覆盖、不可分宽度、让我继续访谈等规划句。混有过程句和对用户提问时只保留口语。不改已哈希 Skill `10.0.11`。
+- 验证：`frontend/tests/rectification-spoken-answer.test.ts` 锁定规划段进思考、入职追问留在口语；`frontend/tests/rectification-v9-stream.test.ts` 锁定最终 `text-delta` 只把口语发成 `answer.delta`。
+- 防复发：禁止把工具完成标签、内部字段名或「本轮对照了」写进 `answer.delta`。thinking 关闭时漏检网必须覆盖中文规划句，不能只认英文 / batch / datePrecision。对用户说话的追问不得被一起丢掉。
+- 相关记录：BUG-357、BUG-359、BUG-368、BUG-360
+- 复发自：BUG-357（关 thinking 后用正则从正文切思考；句式换了就漏）
+- 修复版本：待发布
+
 ## BUG-372 | 生时纠正同一轮重复工具调用把已成功写入打成失败
 
 - 状态：resolved

frontend/src/lib/rectification-agentic/v9/spoken-answer.ts

@@ -6,20 +6,40 @@
  * - drops process-only `text-delta` leaks so they cannot complete as the reply
  */
 
+import {
+  RECTIFICATION_ACTIVITY_PROGRESS_LABELS,
+  RECTIFICATION_TOOL_DONE_LABELS,
+} from "../../rectification-activity-labels.ts";
+
 const CJK_RE = /[\u4e00-\u9fff]/;
-const INTERNAL_TOKEN_RE = /\b(?:datePrecision|occurredFrom|occurredTo|proposedKind|education_start|missing_evidence|SKILL\.md|rectification-[a-z0-9-]+|focusId|evidenceId|display_date_label)\b/;
-const PROCESS_ZH_RE = /skill\s*规则|不得猜补|让我(?:调用|记录|batch|提交)|我(?:决定|倾向|batch|需要用)|权衡：|内部矛盾|思维链|调用 batch|批量工具|写入(?:这些)?证据|datePrecision|occurredFrom|occurredTo/;
-const THIRD_PERSON_USER_RE = /用户(?:在上|提到|先(?:说|提到)|说|自己|的核心|想表达|原话|的最终|对年份|提供了)/;
+const INTERNAL_TOKEN_RE = /\b(?:datePrecision|occurredFrom|occurredTo|proposedKind|education_start|missing_evidence|SKILL\.md|rectification-[a-z0-9-]+|focusId|evidenceId|display_date_label|occupation_note|method_followup_plan|open_question|next_action|next_user_action|not_separated|propose_allowed|selection_allowed|information_gain|event_probe|session_outcome|unique_minute_path|confirmation_allowed|collect_method_evidence|candidate_contrast|deferred_followup)\b/;
+const PROCESS_ZH_RE = /skill\s*规则|不得猜补|让我(?:调用|记录|batch|提交|继续|用)|我(?:决定|倾向|batch|需要用|需要继续|继续收集|继续访谈|自然地|用自然语言)|权衡：|内部矛盾|思维链|调用 batch|批量工具|写入(?:这些)?证据|datePrecision|occurredFrom|occurredTo|方法覆盖|还不能出牌|不得出牌|本轮对照了|不可分宽度|重新计算了候选|带评分日期|当前还应继续收集|根据 method_followup/;
+const THIRD_PERSON_USER_RE = /^用户|用户(?:在上|提到|先(?:说|提到)|说|自己|的核心|想表达|原话|的最终|对年份|提供了)/;
+
+const ACTIVITY_ECHO_LABELS = [
+  ...Object.values(RECTIFICATION_TOOL_DONE_LABELS),
+  ...Object.values(RECTIFICATION_ACTIVITY_PROGRESS_LABELS),
+];
 
 export type SplitSpokenAndThinking = Readonly<{
   thinking: string;
   spoken: string;
 }>;
 
+function isActivityEcho(text: string): boolean {
+  const trimmed = text.trim().replace(/[。.…]+$/u, "");
+  if (!trimmed) return false;
+  return ACTIVITY_ECHO_LABELS.some((label) => {
+    const bare = label.replace(/[。.…]+$/u, "");
+    return trimmed === bare || trimmed === `正在${bare}`;
+  });
+}
+
 export function isRectificationProcessNarration(text: string): boolean {
   const trimmed = text.trim();
   if (!trimmed) return false;
   if (/[A-Za-z]{4,}/.test(trimmed) && !CJK_RE.test(trimmed)) return true;
+  if (isActivityEcho(trimmed)) return true;
   if (INTERNAL_TOKEN_RE.test(trimmed)) return true;
   if (PROCESS_ZH_RE.test(trimmed)) return true;
   if (THIRD_PERSON_USER_RE.test(trimmed)) return true;
@@ -31,17 +51,31 @@ export function nextStableChannelDelta(published: string, next: string): string
   return next.slice(published.length);
 }
 
+function splitSentences(text: string): string[] {
+  const parts = text.split(/(?<=[。！？])\s*/u).map((part) => part.trim()).filter(Boolean);
+  return parts.length > 0 ? parts : [text];
+}
+
+function splitUnits(text: string): string[] {
+  const lines = text.split(/\n/).map((line) => line.trim()).filter(Boolean);
+  if (lines.length > 1) return lines.flatMap(splitSentences);
+  return splitSentences(text);
+}
+
+function isMixed(units: readonly string[]): boolean {
+  return units.length > 1
+    && units.some(isRectificationProcessNarration)
+    && units.some((unit) => !isRectificationProcessNarration(unit));
+}
+
 function splitParagraphs(text: string): string[] {
   return text
     .split(/\n{2,}/)
     .flatMap((block) => {
       const trimmed = block.trim();
       if (!trimmed) return [];
-      const lines = trimmed.split(/\n/).map((line) => line.trim()).filter(Boolean);
-      if (lines.length <= 1) return [trimmed];
-      const hasProcess = lines.some(isRectificationProcessNarration);
-      const hasSpoken = lines.some((line) => !isRectificationProcessNarration(line));
-      return hasProcess && hasSpoken ? lines : [trimmed];
+      const units = splitUnits(trimmed);
+      return isMixed(units) ? units : [trimmed];
     });
 }
 

frontend/tests/rectification-spoken-answer.test.ts

@@ -89,6 +89,30 @@ test("finalizing does not promote process-only self-talk into the spoken answer"
   });
 });
 
+test("Chinese interview planning after tools stays out of the spoken answer", () => {
+  const processTalk = [
+    "读取校正记录",
+    "整理多条事件证据",
+    "比较候选时间",
+    "职业类型已经记录，并重新计算了候选。虽然职业领域现在有信息了，但账户里仍然只有 3 件带评分日期的事件，还不足以拉开候选范围区间（当前不可分宽度 8 分钟）。",
+    "方法覆盖上，职业这一层已经通过 occupation_note 补齐了（不计分），但还需要更多带日期的经历来区分候选。当前还应继续收集事件。",
+    "我用自然语言再问一件带大概年份的经历。根据 method_followup_plan，职业已经覆盖。当前 open_question 为空，next_action 为空。",
+    "候选区分尚未充分（not_separated），应继续收集。",
+    "让我继续访谈，问一件能帮助区分候选的职业前事。",
+    "本轮对照了Gochara、D1 本命盘、D10 事业分盘。",
+  ].join("\n\n");
+  const spoken = "职业类型已经记下。接下来想请你回想一下这份工作的时间段——**你大概是在哪一年入职的？**又是什么时候离开的？只要个大概年份就行。";
+
+  const split = splitRectificationSpokenAndThinking(`${processTalk}\n\n${spoken}`);
+  assert.equal(split.spoken, spoken);
+  assert.match(split.thinking, /occupation_note/);
+  assert.match(split.thinking, /读取校正记录/);
+  assert.match(split.thinking, /本轮对照了/);
+  assert.doesNotMatch(split.thinking, /你大概是在哪一年入职/);
+  assert.equal(isRectificationProcessNarration("读取校正记录"), true);
+  assert.equal(isRectificationProcessNarration(spoken), false);
+});
+
 test("leaked process text on the answer channel is not mixed into native thinking", () => {
   const processTalk = "用户在上一轮里提供了两件带日期的经历。我需要用批量工具写入这些证据。用户";
   const spoken = "记下了升学这两件。接下来有没有一件带大概年份的工作变化？";

frontend/tests/rectification-v9-stream.test.ts

@@ -1301,6 +1301,43 @@ test("does not publish intermediate tool-step text as answer.delta", async () =>
   assert.equal(emitted.some((event) => event.type === "thinking.delta"), false);
 });
 
+test("Chinese interview planning in the terminal text-delta is not the spoken answer", async () => {
+  const spoken = "职业类型已经记下。接下来想请你回想一下这份工作的时间段——**你大概是在哪一年入职的？**只要个大概年份就行。";
+  const { options, emitted } = runOptions({
+    buildAgent: async () => fakeAgentStream([
+      chunk("start"),
+      chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }),
+      chunk("tool-result", { toolName: "skill" }),
+      chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-read-case" }),
+      chunk("tool-call", { toolName: "rectification-compare-candidates", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-compare-candidates" }),
+      chunk("text-delta", {
+        text: [
+          "读取校正记录",
+          "比较候选时间",
+          "方法覆盖上，职业这一层已经通过 occupation_note 补齐了（不计分）。",
+          "让我继续访谈，问一件能帮助区分候选的职业前事。",
+          "本轮对照了Gochara、D1 本命盘、D10 事业分盘。",
+          spoken,
+        ].join("\n\n"),
+      }),
+      chunk("finish"),
+    ]) as never,
+  });
+  const result = await runV9AgentTurn(options);
+  assert.equal(result.ok, true);
+  assert.equal(result.answerText, spoken);
+  assert.deepEqual(
+    emitted.filter((event) => event.type === "answer.delta"),
+    [{ type: "answer.delta", text: spoken }],
+  );
+  const publicText = JSON.stringify(emitted);
+  assert.doesNotMatch(publicText, /occupation_note/);
+  assert.doesNotMatch(publicText, /让我继续访谈/);
+  assert.doesNotMatch(publicText, /本轮对照了/);
+});
+
 test("does not reset the whole attempt after duplicate_focus", async () => {
   let buildCount = 0;
   const accounting = fakeAccounting({