diff --git a/apps/game-server/src/colyseus/GameRoom.ts b/apps/game-server/src/colyseus/GameRoom.ts
index c6332c2..2bafd6f 100644
--- a/apps/game-server/src/colyseus/GameRoom.ts
+++ b/apps/game-server/src/colyseus/GameRoom.ts
@@ -30,6 +30,7 @@ import {
   validateChatMessage,
   validateChatNpcId,
 } from "./npc-chat.js";
+import { getRecentTurns } from "../npc/dialogue-session.js";
 import {
   getColyseusRoom,
   tryClaimMapRoom,
@@ -278,7 +279,8 @@ export class GameRoom extends Room {
         });
         // speakAck before Redis enqueue — fast-lane worker can finish before LPUSH returns otherwise.
         client.send(COLYSEUS_SERVER_MESSAGES.speakAck, { jobId, npcId });
-        const casualStub = previewCasualSpeakStub(text);
+        const recentTurns = getRecentTurns(this.mapRoomId, playerId, npcId, 10);
+        const casualStub = previewCasualSpeakStub(text, recentTurns);
         if (casualStub) {
           emitJobEvent(jobId, "speakPartial", { text: casualStub, npcId });
         }
diff --git a/apps/game-server/src/routes/chat.ts b/apps/game-server/src/routes/chat.ts
index 87cb71d..270f41e 100644
--- a/apps/game-server/src/routes/chat.ts
+++ b/apps/game-server/src/routes/chat.ts
@@ -11,6 +11,7 @@ import {
 } from "../colyseus/npc-chat.js";
 import { getColyseusRoom } from "../colyseus/room-registry.js";
 import type { GameRoom } from "../colyseus/GameRoom.js";
+import { getRecentTurns } from "../npc/dialogue-session.js";
 import { playerIdFromRequest } from "../http/player-id.js";
 import { getOrCreate } from "../room/store.js";
 import { emitJobEvent, subscribeJobEvents } from "../sse/hub.js";
@@ -55,7 +56,8 @@ export function createChatRouter(): Router {
 
     let speakAcquired = Boolean(colyseusRoom);
     try {
-      const casualStub = previewCasualSpeakStub(message);
+      const recentTurns = getRecentTurns(roomId, playerId, npcId, 10);
+      const casualStub = previewCasualSpeakStub(message, recentTurns);
       if (colyseusRoom) {
         registerJob(jobId, colyseusRoom, roomId, undefined, {
           npcId,
diff --git a/apps/game-server/src/speak/casual-stub.test.ts b/apps/game-server/src/speak/casual-stub.test.ts
index e128ff7..b9944ad 100644
--- a/apps/game-server/src/speak/casual-stub.test.ts
+++ b/apps/game-server/src/speak/casual-stub.test.ts
@@ -15,4 +15,12 @@ describe("previewCasualSpeakStub", () => {
   it("returns null for physical message", () => {
     expect(previewCasualSpeakStub("向右走一步")).toBeNull();
   });
+
+  it("returns null when recent turns present", () => {
+    const history = [
+      { role: "player" as const, text: "干嘛呢？" },
+      { role: "npc" as const, text: "在忙" },
+    ];
+    expect(previewCasualSpeakStub("你好，用一句话简短回复", history)).toBeNull();
+  });
 });
diff --git a/apps/web/src/hooks/npcChat/dialogueTurns.test.ts b/apps/web/src/hooks/npcChat/dialogueTurns.test.ts
new file mode 100644
index 0000000..e23744e
--- /dev/null
+++ b/apps/web/src/hooks/npcChat/dialogueTurns.test.ts
@@ -0,0 +1,54 @@
+import { describe, expect, it } from "vitest";
+import { recentDialogueTurnsForNpc } from "./dialogueTurns.js";
+import type { ChatMessage } from "./types.js";
+
+describe("recentDialogueTurnsForNpc", () => {
+  it("pairs player lines with matching npc thread only", () => {
+    const messages: ChatMessage[] = [
+      { id: "1", role: "player", text: "a", npcId: "npc-5" },
+      { id: "2", role: "npc", text: "r1", npcId: "npc-1" },
+      { id: "3", role: "player", text: "b", npcId: "npc-5" },
+      { id: "4", role: "npc", text: "r5", npcId: "npc-5" },
+    ];
+    // Other-NPC replies must not drop pending player lines for npc-5.
+    expect(recentDialogueTurnsForNpc(messages, "npc-5")).toEqual([
+      { role: "player", text: "a" },
+      { role: "player", text: "b" },
+      { role: "npc", text: "r5" },
+    ]);
+  });
+
+  it("excludes interleaved player messages targeting other npcs", () => {
+    const messages: ChatMessage[] = [
+      { id: "1", role: "player", text: "to-1a", npcId: "npc-1" },
+      { id: "2", role: "npc", text: "from-1a", npcId: "npc-1" },
+      { id: "3", role: "player", text: "to-5", npcId: "npc-5" },
+      { id: "4", role: "npc", text: "from-5", npcId: "npc-5" },
+      { id: "5", role: "player", text: "to-1b", npcId: "npc-1" },
+      { id: "6", role: "npc", text: "from-1b", npcId: "npc-1" },
+    ];
+    expect(recentDialogueTurnsForNpc(messages, "npc-1")).toEqual([
+      { role: "player", text: "to-1a" },
+      { role: "npc", text: "from-1a" },
+      { role: "player", text: "to-1b" },
+      { role: "npc", text: "from-1b" },
+    ]);
+    expect(recentDialogueTurnsForNpc(messages, "npc-5")).toEqual([
+      { role: "player", text: "to-5" },
+      { role: "npc", text: "from-5" },
+    ]);
+  });
+
+  it("preserves pending player turn when another NPC replies first", () => {
+    const messages: ChatMessage[] = [
+      { id: "1", role: "player", text: "to-5", npcId: "npc-5" },
+      { id: "2", role: "player", text: "to-1", npcId: "npc-1" },
+      { id: "3", role: "npc", text: "from-1", npcId: "npc-1" },
+      { id: "4", role: "npc", text: "from-5", npcId: "npc-5" },
+    ];
+    expect(recentDialogueTurnsForNpc(messages, "npc-5")).toEqual([
+      { role: "player", text: "to-5" },
+      { role: "npc", text: "from-5" },
+    ]);
+  });
+});
diff --git a/apps/web/src/hooks/npcChat/dialogueTurns.ts b/apps/web/src/hooks/npcChat/dialogueTurns.ts
new file mode 100644
index 0000000..5dfd092
--- /dev/null
+++ b/apps/web/src/hooks/npcChat/dialogueTurns.ts
@@ -0,0 +1,31 @@
+import type { DialogueTurn } from "@aetherlife/shared";
+import type { ChatMessage } from "./types.js";
+
+/** Completed player↔npc turns for one NPC thread (mirrors game-server dialogue-session). */
+export function recentDialogueTurnsForNpc(
+  messages: readonly ChatMessage[],
+  npcId: string,
+  limit = 10,
+): DialogueTurn[] {
+  const turns: DialogueTurn[] = [];
+  let pendingPlayers: DialogueTurn[] = [];
+
+  for (const m of messages) {
+    if (m.role === "error") continue;
+    if (m.role === "player") {
+      // Skip other-NPC lines without clearing pending for this thread.
+      if (m.npcId && m.npcId !== npcId) continue;
+      pendingPlayers.push({ role: "player", text: m.text });
+      continue;
+    }
+    if (m.role === "npc") {
+      // Unrelated NPC replies must not drop an in-flight player line for npcId.
+      if (m.npcId !== npcId) continue;
+      turns.push(...pendingPlayers);
+      pendingPlayers = [];
+      turns.push({ role: "npc", text: m.text });
+    }
+  }
+
+  return turns.slice(-limit);
+}
diff --git a/apps/web/src/hooks/npcChat/index.ts b/apps/web/src/hooks/npcChat/index.ts
index d9066b5..b945ba0 100644
--- a/apps/web/src/hooks/npcChat/index.ts
+++ b/apps/web/src/hooks/npcChat/index.ts
@@ -7,6 +7,7 @@ export type {
   RoomStateShape,
   UseNpcChatOptions,
 } from "./types.js";
+export { recentDialogueTurnsForNpc } from "./dialogueTurns.js";
 export { attitudeGateHintCopy } from "./attitudeGate.js";
 export {
   dequeueNpcSpeak,
diff --git a/apps/web/src/hooks/useNpcChat.ts b/apps/web/src/hooks/useNpcChat.ts
index ada31f8..2ea126a 100644
--- a/apps/web/src/hooks/useNpcChat.ts
+++ b/apps/web/src/hooks/useNpcChat.ts
@@ -41,6 +41,7 @@ import {
   type RoomNpc,
   type RoomStateShape,
   type UseNpcChatOptions,
+  recentDialogueTurnsForNpc,
 } from "./npcChat/index.js";
 
 export type {
@@ -88,6 +89,8 @@ export function useNpcChat(
     };
   }, []);
   const [messages, setMessages] = useState<ChatMessage[]>([]);
+  const messagesRef = useRef<ChatMessage[]>([]);
+  messagesRef.current = messages;
   const [status, setStatus] = useState<ChatStatus>("idle");
   const [roomState, setRoomState] = useState<RoomStateShape | null>(null);
   const [memoryCounts, setMemoryCounts] = useState<Record<string, number>>({});
@@ -134,7 +137,7 @@ export function useNpcChat(
       if (opts?.showPlayerBubble !== false) {
         setMessages((prev) => [
           ...prev,
-          { id: crypto.randomUUID(), role: "player", text },
+          { id: crypto.randomUUID(), role: "player", text, npcId },
         ]);
       }
     },
@@ -497,7 +500,7 @@ export function useNpcChat(
       if (!opts?.skipPlayerBubble) {
         setMessages((prev) => [
           ...prev,
-          { id: crypto.randomUUID(), role: "player", text: trimmed },
+          { id: crypto.randomUUID(), role: "player", text: trimmed, npcId },
         ]);
       }
       inFlightTextRef.current.set(npcId, trimmed);
@@ -518,7 +521,8 @@ export function useNpcChat(
         window.__speakLatencyT0 = performance.now();
       }
 
-      const clientStub = previewCasualSpeakStub(trimmed);
+      const recentTurns = recentDialogueTurnsForNpc(messagesRef.current, npcId);
+      const clientStub = previewCasualSpeakStub(trimmed, recentTurns);
       if (clientStub) {
         setStreamingByNpc((prev) => ({
           ...prev,
diff --git a/docs/ISSUE-LOG.md b/docs/ISSUE-LOG.md
index ba86944..7d74c9b 100644
--- a/docs/ISSUE-LOG.md
+++ b/docs/ISSUE-LOG.md
@@ -202,6 +202,7 @@
 112. **个人传记隔离**：`npc_personal_timeline` 是唯一个人人生时间线存储；**禁止**把个人 biography 写入 `__council__`（C-07）或玩家 speak `npc_memories`（C-05）。Worker/seed 只经 `insertPersonalTimelineEntry` / internal POST；改写路径须 `pnpm --filter @aetherlife/game-server test -- personal-timeline-repository`（含 isolation 源码断言）。
 113. **个人日记须按席位人设写**：weekly/polish/multi/rel/event prompt **必须** `persona_block_for`（speak mirror）；**禁止**无口吻的「人生札记」通稿导致 ENTJ/ESFP 写同款文艺腔（ISSUE-106）。周记须带 `recentBullets`；非廷议双边走 `kind=event` + `min_abs_delta=DYAD_REL_MIN_ABS_DELTA`（|Δ|≥4，禁止无关键词 casual mention）。回归：`pytest tests/test_personal_timeline.py tests/test_personal_timeline_rel07.py -q` · `pnpm --filter @aetherlife/game-server test -- personal-timeline-dyad personal-timeline-weekly`。
 114. **Personal-timeline job 入队须 durable claim**：`claimPersonalTimelineJobId` / worker `claim_personal_timeline_job_id`（SET NX，前缀 `aetherlife:personal-timeline:job-claimed:`）覆盖 polish/weekly/multi/rel/event **以及** dyad pair/ambient-slot；**禁止**仅靠进程内存 debounce（重启会重复 LPUSH → 重复 LLM 行）。传记 UI：fetch 失败须展示 error（勿静默空列表）；已缓存条目在 `personalTimelineSync` 时须后台 refetch。回归：`personal-timeline.claim.test.ts` · `personal-timeline-dyad.test.ts` · `usePersonalTimeline.test.ts` · `pnpm uat:phase27:persona-diary`。
+115. **Speak 多轮连贯（260720-m4b）**：interactive 路径 `llm_social_turn._build_social_messages` **必须**注入 `recent_turns` Human/AI 链（`append_recent_dialogue_messages`）；**禁止** help offer（`player_offers_help` /「我可以帮你」）走 SOCIAL_EDGE deterministic stub；`recent_turns` 非空时 **禁止** CASUAL/SOCIAL_EDGE fast lane（B1 例外：空历史纯问候）。`augment_retrieved_with_dialogue_turns` 须含 `npc:` 行。回归：`pytest tests/test_speak_intent.py tests/test_help_reply_by_npc.py tests/test_llm_social_memory.py tests/test_casual_fast_lane.py tests/test_recall_merge.py -q` · `pnpm --filter @aetherlife/shared test -- speakIntent` · `pnpm agent:verify`。
 
 ## 记录
 
@@ -2808,4 +2809,42 @@ Worker 主循环仅在 npc-turn 队列 **连续 5s 为空** 时才 `BLPOP` chunk
 
 ---
 
+### ISSUE-107 — Speak 多轮对话 stub 断链：「我可以帮你」回复反向
+
+- **状态:** fixed
+- **发现:** 2026-07-20
+- **阶段/范围:** speak intent / `llm_social_turn` / fast lane（worker + `packages/shared`）
+- **严重性:** major（对话连贯 / 人设）
+
+**复现**
+
+1. 对糖果说「干嘛呢？」→ NPC 正常回复想黑系统
+2. 接着说「我可以帮你！」
+3. NPC 回复「好的，我会尽力帮忙。」（像玩家在求助，忽略上文）
+
+**根因**
+
+- ISSUE-018 修复了 `build_turn_messages` 历史，但 interactive 主路径改用 `llm_social_turn._build_social_messages` 后**未注入 `recent_turns`**
+- `infer_social_from_message` 凡含「帮」即 help-request → SOCIAL_EDGE fast lane deterministic stub
+- Fast lane / `_deterministic_social_turn` 不读 session，npc-4 等席别落默认套话
+
+**修复**
+
+- `player_offers_help` 区分 offer vs request（TS/Python parity）
+- `recent_turns` 非空或 offer 时 gate CASUAL/SOCIAL_EDGE fast lane；continuation 短路
+- `_build_social_messages` + `append_recent_dialogue_messages` 注入 Human/AI 链；`augment_retrieved_with_dialogue_turns` 含 npc 行
+
+**验证**
+
+- `pnpm agent:verify`
+- `pnpm uat:speak-help-offer:playwright`（真实 LLM + Playwright；`pnpm dev:stack`）
+- `cd workers/agent-worker && LLM_MOCK=1 uv run pytest tests/test_speak_intent.py tests/test_help_reply_by_npc.py tests/test_llm_social_memory.py tests/test_casual_fast_lane.py tests/test_recall_merge.py tests/test_graph_tools.py -q`
+- `pnpm --filter @aetherlife/shared test -- speakIntent`
+
+**防复发**
+
+- Guardrail #115
+
+---
+
 <!-- 新问题上文追加，保持 ISSUE 编号递增 -->
diff --git a/package.json b/package.json
index b1abbb3..f0329e3 100644
--- a/package.json
+++ b/package.json
@@ -85,6 +85,7 @@
     "uat:phase6:playwright": "node scripts/uat-phase6-playwright.mjs",
     "uat:phase7:playwright": "node scripts/uat-phase7-playwright.mjs",
     "uat:phase8:playwright": "node scripts/uat-phase8-playwright.mjs",
+    "uat:speak-help-offer:playwright": "node scripts/uat-speak-help-offer-playwright.mjs",
     "uat:phase10:playwright": "node scripts/uat-phase10-playwright.mjs",
     "uat:phase11:playwright": "node scripts/uat-phase11-playwright.mjs",
     "uat:phase7:reset-snap": "node scripts/uat-phase7-reset-snap.mjs",
diff --git a/packages/shared/src/casualSpeakStub.ts b/packages/shared/src/casualSpeakStub.ts
index fb353d5..3e9ec20 100644
--- a/packages/shared/src/casualSpeakStub.ts
+++ b/packages/shared/src/casualSpeakStub.ts
@@ -6,8 +6,10 @@ import {
   classifySpeakIntent,
   inferSocialFromMessage,
   isCasualGreetingOnly,
+  playerOffersHelp,
   playerRequestsPhysicalAction,
   SpeakIntent,
+  type DialogueTurn,
   type SpeakIntentValue,
 } from "./speakIntent.js";
 import { stableStringHash } from "./stableStringHash.js";
@@ -54,7 +56,10 @@ function deterministicSocialReply(message: string, speakIntent: SpeakIntentValue
   const inferred = inferSocialFromMessage(msg);
   if (inferred !== null) {
     if (inferred === "rude") return "请不要这样说话。";
-    if (inferred === "help") return "好的，我会尽力帮忙。";
+    if (inferred === "help") {
+      if (playerOffersHelp(msg)) return null;
+      return "好的，我会尽力帮忙。";
+    }
     return `我听到了：${msg.slice(0, 120)}`;
   }
 
@@ -72,8 +77,12 @@ function deterministicSocialReply(message: string, speakIntent: SpeakIntentValue
 }
 
 /** Early speakPartial text for CASUAL deterministic turns. */
-export function previewCasualSpeakStub(message: string): string | null {
-  const intent = classifySpeakIntent(message);
+export function previewCasualSpeakStub(
+  message: string,
+  recentTurns?: readonly DialogueTurn[] | null,
+): string | null {
+  if (recentTurns?.length) return null;
+  const intent = classifySpeakIntent(message, recentTurns);
   if (intent !== SpeakIntent.CASUAL) return null;
   return deterministicSocialReply(message, intent);
 }
@@ -83,8 +92,12 @@ export type CasualFastLanePreview = {
   stub: string;
 };
 
-export function canUseCasualFastLane(message: string): CasualFastLanePreview | null {
-  const intent = classifySpeakIntent(message);
+export function canUseCasualFastLane(
+  message: string,
+  recentTurns?: readonly DialogueTurn[] | null,
+): CasualFastLanePreview | null {
+  if (recentTurns?.length) return null;
+  const intent = classifySpeakIntent(message, recentTurns);
   if (intent !== SpeakIntent.CASUAL) return null;
   const stub = deterministicSocialReply(message, intent);
   if (!stub) return null;
diff --git a/packages/shared/src/index.ts b/packages/shared/src/index.ts
index 19a33a4..dad4633 100644
--- a/packages/shared/src/index.ts
+++ b/packages/shared/src/index.ts
@@ -243,6 +243,7 @@ export {
   playerRequestsInteract,
   playerRequestsMove,
   shouldSkipMemoryContext,
+  type DialogueTurn,
   type SpeakIntentValue,
 } from "./speakIntent.js";
 
diff --git a/packages/shared/src/speakIntent.test.ts b/packages/shared/src/speakIntent.test.ts
index afb8193..8fbe13a 100644
--- a/packages/shared/src/speakIntent.test.ts
+++ b/packages/shared/src/speakIntent.test.ts
@@ -6,6 +6,8 @@ import {
 } from "./casualSpeakStub.js";
 import {
   classifySpeakIntent,
+  inferSocialFromMessage,
+  playerOffersHelp,
   shouldSkipMemoryContext,
   SpeakIntent,
 } from "./speakIntent.js";
@@ -39,9 +41,45 @@ describe("classifySpeakIntent", () => {
     expect(classifySpeakIntent("请帮帮我")).toBe(SpeakIntent.SOCIAL_EDGE);
     expect(classifySpeakIntent("滚开")).toBe(SpeakIntent.SOCIAL_EDGE);
     expect(classifySpeakIntent("你真蠢")).toBe(SpeakIntent.SOCIAL_EDGE);
+    expect(classifySpeakIntent("你真粗鲁")).toBe(SpeakIntent.SOCIAL_EDGE);
     expect(classifySpeakIntent("能请你帮个忙吗")).toBe(SpeakIntent.SOCIAL_EDGE);
   });
 
+  it("continuation short with history routes to narrative", () => {
+    const history = [
+      { role: "player" as const, text: "干嘛呢？" },
+      { role: "npc" as const, text: "在忙" },
+    ];
+    expect(classifySpeakIntent("好的", history)).toBe(SpeakIntent.NARRATIVE);
+    expect(classifySpeakIntent("你好", history)).toBe(SpeakIntent.NARRATIVE);
+  });
+
+  it("help offer is not social edge", () => {
+    expect(playerOffersHelp("我可以帮你！")).toBe(true);
+    expect(playerOffersHelp("我能帮你！")).toBe(true);
+    expect(playerOffersHelp("我愿意帮你")).toBe(true);
+    expect(playerOffersHelp("我想帮你")).toBe(true);
+    expect(playerOffersHelp("我来帮")).toBe(true);
+    expect(playerOffersHelp("让我帮你")).toBe(true);
+    expect(playerOffersHelp("请帮帮我")).toBe(false);
+    expect(inferSocialFromMessage("我可以帮你！")).toBeNull();
+    expect(inferSocialFromMessage("我能帮你！")).toBeNull();
+    expect(inferSocialFromMessage("我愿意帮你")).toBeNull();
+    expect(inferSocialFromMessage("我想帮你")).toBeNull();
+    expect(inferSocialFromMessage("请帮帮我")).toBe("help");
+    expect(classifySpeakIntent("我可以帮你！")).toBe(SpeakIntent.NARRATIVE);
+    expect(classifySpeakIntent("我能帮你！")).toBe(SpeakIntent.NARRATIVE);
+    expect(classifySpeakIntent("我愿意帮你")).toBe(SpeakIntent.NARRATIVE);
+    expect(classifySpeakIntent("我想帮你")).toBe(SpeakIntent.NARRATIVE);
+    expect(classifySpeakIntent("我来帮")).toBe(SpeakIntent.NARRATIVE);
+  });
+
+  it("narrative bang compounds are not help requests", () => {
+    expect(inferSocialFromMessage("帮别人做事")).toBeNull();
+    expect(inferSocialFromMessage("别在这里帮腔")).toBeNull();
+    expect(classifySpeakIntent("帮别人做事")).toBe(SpeakIntent.NARRATIVE);
+  });
+
   it("casual intent", () => {
     expect(classifySpeakIntent("你好")).toBe(SpeakIntent.CASUAL);
     expect(classifySpeakIntent("Hi")).toBe(SpeakIntent.CASUAL);
@@ -99,6 +137,16 @@ describe("casual reply pool", () => {
     expect(previewCasualSpeakStub("你在做什么呢？")).toBeNull();
     expect(previewCasualSpeakStub("你好狂啊～")).toBeNull();
     expect(previewCasualSpeakStub("在啥啊")).toBeNull();
+    expect(previewCasualSpeakStub("我可以帮你！")).toBeNull();
+  });
+
+  it("preview casual stub blocked with recent turns", () => {
+    const history = [
+      { role: "player" as const, text: "干嘛呢？" },
+      { role: "npc" as const, text: "在忙" },
+    ];
+    expect(previewCasualSpeakStub("你好", history)).toBeNull();
+    expect(canUseCasualFastLane("你好，用一句话简短回复", history)).toBeNull();
   });
 
   it("can use casual fast lane b1", () => {
diff --git a/packages/shared/src/speakIntent.ts b/packages/shared/src/speakIntent.ts
index 82c01f6..dfe57e2 100644
--- a/packages/shared/src/speakIntent.ts
+++ b/packages/shared/src/speakIntent.ts
@@ -10,6 +10,9 @@ export const SpeakIntent = {
 
 export type SpeakIntentValue = (typeof SpeakIntent)[keyof typeof SpeakIntent];
 
+/** Short-term dialogue turn — mirrors game-server dialogue-session + worker recent_turns. */
+export type DialogueTurn = { role: "player" | "npc"; text: string };
+
 const RECALL_MARKERS = [
   "记得",
   "还记得",
@@ -74,7 +77,53 @@ const INTERACT_PATTERNS: RegExp[] = [
   /\binteract\b/i,
 ];
 
-const INSULT_MARKERS = ["丑", "滚", "蠢", "有病", "变态", "活该", "什么玩意", "傻", "废物", "去死"];
+const INSULT_MARKERS = [
+  "粗鲁",
+  "丑",
+  "滚",
+  "蠢",
+  "有病",
+  "变态",
+  "活该",
+  "什么玩意",
+  "傻",
+  "废物",
+  "去死",
+  "讨厌",
+  "笨蛋",
+  "侮辱",
+  "辱骂",
+];
+
+const HELP_OFFER_RE = /(?:我可以|我能|我愿意|我想|我来|让我)(?:帮|协助)/;
+/** Avoid narrative false positives (帮别人 / 帮腔) — 帮 + request pronoun/cue only. */
+const HELP_REQUEST_FALLBACK_RE = /帮[我你他她它个一上下把忙]/;
+
+/** Player volunteers to help the NPC — not a help request. Mirrors social_turn.player_offers_help */
+export function playerOffersHelp(message: string): boolean {
+  const msg = (message || "").trim();
+  if (!msg) return false;
+  if (HELP_OFFER_RE.test(msg)) return true;
+  if (msg.startsWith("我帮你")) return true;
+  if (msg.startsWith("帮你")) return true;
+  return false;
+}
+
+function messageImpliesHelpRequest(message: string): boolean {
+  const msg = (message || "").trim();
+  if (!msg || playerOffersHelp(msg)) return false;
+  if (
+    ["帮帮我", "帮个忙", "请帮", "你能帮", "能帮我", "能请你帮", "帮忙"].some((marker) =>
+      msg.includes(marker),
+    )
+  ) {
+    return true;
+  }
+  if (HELP_REQUEST_FALLBACK_RE.test(msg)) return true;
+  if (!msg.includes("请")) return false;
+  const normalized = msg.replace(/回复/g, "").replace(/请假/g, "");
+  return normalized.includes("请");
+}
 
 export function playerRequestsMove(message: string): boolean {
   const text = (message || "").trim();
@@ -103,7 +152,7 @@ export function inferSocialFromMessage(message: string): "rude" | "help" | null
   const msg = (message || "").trim();
   if (!msg) return null;
   if (INSULT_MARKERS.some((marker) => msg.includes(marker))) return "rude";
-  if (msg.includes("帮") || msg.includes("请")) return "help";
+  if (messageImpliesHelpRequest(msg)) return "help";
   return null;
 }
 
@@ -114,11 +163,17 @@ export function isCasualGreetingOnly(message: string): boolean {
   return CASUAL_GREETING_ONLY_RE.test(msg);
 }
 
-export function classifySpeakIntent(message: string): SpeakIntentValue {
+export function classifySpeakIntent(
+  message: string,
+  recentTurns?: readonly DialogueTurn[] | null,
+): SpeakIntentValue {
   const msg = (message || "").trim();
   if (!msg) return SpeakIntent.NARRATIVE;
   if (playerRequestsPhysicalAction(msg)) return SpeakIntent.PHYSICAL;
   if (isRecallQuestion(msg)) return SpeakIntent.RECALL;
+  if (recentTurns?.length && msg.length < 20 && inferSocialFromMessage(msg) === null) {
+    return SpeakIntent.NARRATIVE;
+  }
   if (inferSocialFromMessage(msg) !== null) return SpeakIntent.SOCIAL_EDGE;
   if (isCasualGreetingOnly(msg)) return SpeakIntent.CASUAL;
   if (META_BRIEF_RE.test(msg)) return SpeakIntent.CASUAL;
diff --git a/scripts/uat-speak-help-offer-playwright.mjs b/scripts/uat-speak-help-offer-playwright.mjs
new file mode 100644
index 0000000..cd21dcc
--- /dev/null
+++ b/scripts/uat-speak-help-offer-playwright.mjs
@@ -0,0 +1,133 @@
+/**
+ * Speak coherence UAT (targeted): ensure help offers are not treated as help requests.
+ *
+ * Flow: engage NPC → 「干嘛呢？」→ 「我可以帮你！」
+ * Assert second reply does NOT contain deterministic help-request stub:
+ *   「好的，我会尽力帮忙。」
+ *
+ * Requires: pnpm dev:stack (real LLM). See docs/E2E-POLICY.md
+ */
+import { mkdir } from "node:fs/promises";
+import path from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+
+import { assertE2eRealLlm } from "./lib/e2e-policy.mjs";
+import { engageNpcDialogue } from "./lib/dialogue-engage.mjs";
+import { sendSpeakOverlay } from "./lib/e2e-memory-helpers.mjs";
+import { loadRootEnv } from "./lib/env.mjs";
+
+const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+const OUT_DIR = path.join(ROOT, ".planning/phases/08-multiplayer-room/uat-screenshots");
+
+loadRootEnv(ROOT);
+
+const WEB_BASE = process.env.WEB_URL || "http://localhost:5173";
+const GS = process.env.GAME_SERVER_URL || "http://127.0.0.1:2567";
+const UAT_NPC_ID = process.env.UAT_SPEAK_HELP_OFFER_NPC_ID || "npc-5";
+const UAT_ROOM_ID = process.env.UAT_SPEAK_HELP_OFFER_ROOM_ID || `uat-help-offer-${Date.now()}`;
+const WEB = `${WEB_BASE}${WEB_BASE.includes("?") ? "&" : "?"}room=${encodeURIComponent(UAT_ROOM_ID)}`;
+
+const SPEAK_WAIT_TIMEOUT_MS = Number(process.env.UAT_SPEAK_WAIT_TIMEOUT_MS || 180_000);
+const ENGAGE_TIMEOUT_MS = Number(process.env.UAT_SPEAK_ENGAGE_TIMEOUT_MS || 90_000);
+
+const FORBIDDEN_STUB_RE = /好的，我会尽力帮忙。/;
+const HELP_REQUEST_REPLY_PATTERNS = [
+  FORBIDDEN_STUB_RE,
+  /有什么我能帮你的/,
+  /需要我做什么/,
+  /我会尽力帮忙/,
+];
+const HELP_OFFER_ACCEPTANCE_RE = /感谢|谢谢|好意|感动|愿意|陪|太好了/;
+
+function assertHelpOfferReply(firstReply, secondReply) {
+  const second = (secondReply || "").trim();
+  if (second.length < 10) {
+    throw new Error(`Help-offer reply too short: ${JSON.stringify(second)}`);
+  }
+  for (const pattern of HELP_REQUEST_REPLY_PATTERNS) {
+    if (pattern.test(second)) {
+      throw new Error(
+        `Regression: NPC reply sounds like help-request stub: ${JSON.stringify(second)}`,
+      );
+    }
+  }
+  if (!HELP_OFFER_ACCEPTANCE_RE.test(second)) {
+    throw new Error(
+      `Help-offer reply lacks acceptance/gratitude cue: ${JSON.stringify(second)}`,
+    );
+  }
+  if (firstReply.trim() && second === firstReply.trim()) {
+    throw new Error("Help-offer reply duplicated first NPC line");
+  }
+}
+
+async function screenshot(page, label) {
+  await mkdir(OUT_DIR, { recursive: true });
+  const file = path.join(OUT_DIR, `help-offer-${label}.png`);
+  await page.screenshot({ path: file, fullPage: true });
+  console.log(`  📸 ${path.relative(ROOT, file)}`);
+}
+
+async function health(url, name) {
+  const res = await fetch(`${url}/health`, { signal: AbortSignal.timeout(8000) });
+  if (!res.ok) throw new Error(`${name} /health → ${res.status}`);
+  const body = await res.json();
+  if (body.status !== "ok") throw new Error(`${name} /health body invalid`);
+}
+
+async function main() {
+  assertE2eRealLlm("uat:speak-help-offer:playwright");
+
+  console.log("Speak help-offer UAT (Playwright)");
+  console.log(`WEB=${WEB} GS=${GS} npc=${UAT_NPC_ID} room=${UAT_ROOM_ID}`);
+
+  await health(GS, "game-server");
+  try {
+    const webRes = await fetch(WEB, { signal: AbortSignal.timeout(15_000) });
+    if (!webRes.ok) throw new Error(`Web ${WEB} → ${webRes.status}`);
+  } catch (err) {
+    throw new Error(
+      `Web 不可达 — 请先 pnpm dev:stack（真实 LLM，见 docs/E2E-POLICY.md）: ${err.message}`,
+    );
+  }
+
+  const pwEntry = path.join(ROOT, "scripts", ".pw-deps", "node_modules", "playwright", "index.mjs");
+  const pw = await import(pathToFileURL(pwEntry).href);
+  const chromium = pw.chromium ?? pw.default?.chromium;
+  if (!chromium) throw new Error("playwright 未安装：cd scripts/.pw-deps && npm install");
+
+  const browser = await chromium.launch({ headless: true });
+  const ctx = await browser.newContext({ viewport: { width: 1280, height: 900 } });
+  const page = await ctx.newPage();
+  page.setDefaultTimeout(SPEAK_WAIT_TIMEOUT_MS);
+
+  await page.goto(WEB, { waitUntil: "domcontentloaded", timeout: 60_000 });
+  await page.locator('[data-testid="room-scene"]').waitFor({ state: "visible", timeout: 45_000 });
+  await engageNpcDialogue(page, UAT_NPC_ID, { timeoutMs: ENGAGE_TIMEOUT_MS });
+  await screenshot(page, "00-engaged");
+
+  const speakOpts = {
+    speakTimeoutMs: SPEAK_WAIT_TIMEOUT_MS,
+    engageTimeoutMs: ENGAGE_TIMEOUT_MS,
+    skipEngage: true,
+  };
+
+  const first = await sendSpeakOverlay(page, "干嘛呢？", speakOpts);
+  console.log(`  NPC after 「干嘛呢？」: ${first.reply.slice(0, 80)}…`);
+  await screenshot(page, "01-npc-after-first-message");
+  if (!first.reply.trim()) throw new Error("First NPC reply empty");
+
+  const second = await sendSpeakOverlay(page, "我可以帮你！", speakOpts);
+  await screenshot(page, "02-npc-after-help-offer");
+
+  assertHelpOfferReply(first.reply, second.reply);
+
+  console.log(`✅ Passed: NPC reply after help offer: ${second.reply}`);
+
+  await browser.close();
+}
+
+main().catch((err) => {
+  console.error(`❌ UAT failed: ${err?.message || err}`);
+  process.exit(1);
+});
diff --git a/workers/agent-worker/src/collective/social_turn.py b/workers/agent-worker/src/collective/social_turn.py
index 2a6a71e..3dba5fb 100644
--- a/workers/agent-worker/src/collective/social_turn.py
+++ b/workers/agent-worker/src/collective/social_turn.py
@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import re
 from dataclasses import dataclass
 from typing import Any
 
@@ -45,8 +46,34 @@ class CollectiveApplyResult:
     player_reputation: int | None = None
 
 
+_HELP_OFFER_RE = re.compile(r"(?:我可以|我能|我愿意|我想|我来|让我)(?:帮|协助)")
+# Avoid narrative false positives: 帮别人 / 帮腔 — require 帮 + request pronoun/cue.
+_HELP_REQUEST_FALLBACK_RE = re.compile(r"帮[我你他她它个一上下把忙]")
+
+
+def player_offers_help(message: str) -> bool:
+    """Player volunteers to help the NPC — not a help request."""
+    msg = (message or "").strip()
+    if not msg:
+        return False
+    if _HELP_OFFER_RE.search(msg):
+        return True
+    if msg.startswith("我帮你"):
+        return True
+    if msg.startswith("帮你"):
+        return True
+    return False
+
+
 def _message_implies_help_request(msg: str) -> bool:
-    if "帮" in msg:
+    if player_offers_help(msg):
+        return False
+    if any(
+        marker in msg
+        for marker in ("帮帮我", "帮个忙", "请帮", "你能帮", "能帮我", "能请你帮", "帮忙")
+    ):
+        return True
+    if _HELP_REQUEST_FALLBACK_RE.search(msg):
         return True
     if "请" not in msg:
         return False
diff --git a/workers/agent-worker/src/graph/nodes/llm_social_turn.py b/workers/agent-worker/src/graph/nodes/llm_social_turn.py
index f9fa841..1ad4085 100644
--- a/workers/agent-worker/src/graph/nodes/llm_social_turn.py
+++ b/workers/agent-worker/src/graph/nodes/llm_social_turn.py
@@ -8,10 +8,10 @@
 from typing import Any
 
 import httpx
-from langchain_core.messages import HumanMessage, SystemMessage
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
 
 from src.collective.schemas import SOCIAL_SKIP_KIND, SocialPerception, SocialTurnOut
-from src.collective.social_turn import infer_social_from_message, reconcile_social_perception
+from src.collective.social_turn import infer_social_from_message, player_offers_help, reconcile_social_perception
 from src.config import Settings, get_settings
 from src.graph.action_intent import (
     build_dialogue_context,
@@ -24,6 +24,7 @@
 )
 from src.graph.job_context import get_partial_emit, record_phase_ms
 from src.graph.recall_merge import is_recall_question
+from src.graph.prompt import SOCIAL_DIALOGUE_CONTEXT_APPEND, append_recent_dialogue_messages
 from src.graph.speak_intent import SpeakIntent, is_casual_greeting_only
 from src.graph.stable_string_hash import stable_string_hash
 from src.graph.speak_system_context import (
@@ -160,6 +161,7 @@ def _deterministic_social_turn(
     *,
     speak_intent: str = "",
     npc_id: str = "npc-1",
+    recent_turns: list | None = None,
 ) -> SocialTurnOut | None:
     """Rule-based social + reply — skip social LLM when heuristics are sufficient."""
     msg = player_message.strip()
@@ -167,6 +169,8 @@ def _deterministic_social_turn(
         return None
     inferred = infer_social_from_message(msg)
     if inferred is not None:
+        if (recent_turns or player_offers_help(msg)) and inferred.kind in ("rude", "help"):
+            return None
         if inferred.kind == "rude":
             reply = "请不要这样说话。"
         elif inferred.kind == "help":
@@ -176,6 +180,9 @@ def _deterministic_social_turn(
         else:
             reply = f"我听到了：{msg[:120]}"
         return SocialTurnOut(social=inferred, reply=reply)
+    # With session history, never emit CASUAL/meta stubs — interactive LLM must see turns.
+    if recent_turns:
+        return None
     if speak_intent == SpeakIntent.CASUAL.value or _META_BRIEF_RE.search(msg):
         if not player_requests_physical_action(msg):
             return SocialTurnOut(
@@ -302,12 +309,18 @@ def _build_social_messages(
     state: GraphState,
     *,
     system_append: str = "",
-) -> list[SystemMessage | HumanMessage]:
+) -> list[SystemMessage | HumanMessage | AIMessage]:
+    dialogue_append = SOCIAL_DIALOGUE_CONTEXT_APPEND
+    combined_append = (
+        f"{system_append}\n\n{dialogue_append}".strip()
+        if system_append.strip()
+        else dialogue_append
+    )
     system_text = build_speak_system_context(
         state,
         base_prompt=SOCIAL_SYSTEM_PROMPT,
         memory_suffix=SOCIAL_MEMORY_RECALL_HINT,
-        system_append=system_append,
+        system_append=combined_append,
     )
     player_message = state.get("player_message") or ""
     human = (
@@ -315,10 +328,12 @@ def _build_social_messages(
         "Respond with JSON only. Put \"reply\" as the first key, then \"social\".\n"
         'Example: {"reply":"…","social":{"kind":"ignore","summary":"","delta":0}}'
     )
-    return [
+    messages: list[SystemMessage | HumanMessage | AIMessage] = [
         SystemMessage(content=system_text),
-        HumanMessage(content=human),
     ]
+    append_recent_dialogue_messages(messages, state.get("recent_turns"))
+    messages.append(HumanMessage(content=human))
+    return messages
 
 
 def run_social_turn_llm(
@@ -611,6 +626,7 @@ def llm_social_turn(state: GraphState, *, settings: Settings | None = None) -> G
         player_message,
         speak_intent=speak_intent,
         npc_id=state.get("npc_id") or "npc-1",
+        recent_turns=state.get("recent_turns"),
     )
     if deterministic is not None:
         record_phase_ms("t_social_llm_ms", 0)
diff --git a/workers/agent-worker/src/graph/prompt.py b/workers/agent-worker/src/graph/prompt.py
index 93918b3..58e9cfa 100644
--- a/workers/agent-worker/src/graph/prompt.py
+++ b/workers/agent-worker/src/graph/prompt.py
@@ -20,6 +20,31 @@
 
 RECENT_DIALOGUE_TURN_LIMIT = 10
 
+SOCIAL_DIALOGUE_CONTEXT_APPEND = """\
+Recent dialogue 是真实会话历史。请针对玩家**最新一条**消息作答，结合上文语境；勿复读上一轮几乎相同的回复。
+若玩家主动提出帮助（如「我可以帮你」），须以第一人称接受或回应，**禁止**像玩家向你求助一样说「我会尽力帮你」。"""
+
+
+def append_recent_dialogue_messages(
+    messages: list[SystemMessage | HumanMessage | AIMessage],
+    recent_turns: list | dict[str, str] | None,
+    *,
+    limit: int = RECENT_DIALOGUE_TURN_LIMIT,
+) -> None:
+    """Append Human/AI alternating chain from in-session transcript."""
+    turns = recent_turns if isinstance(recent_turns, list) else []
+    for turn in turns[-limit:]:
+        if not isinstance(turn, dict):
+            continue
+        role = (turn.get("role") or "").strip()
+        text = (turn.get("text") or "").strip()
+        if not text:
+            continue
+        if role == "player":
+            messages.append(HumanMessage(content=text))
+        elif role == "npc":
+            messages.append(AIMessage(content=text))
+
 
 def format_memory_summary(
     *,
@@ -151,16 +176,7 @@ def build_turn_messages(state: GraphState) -> list[SystemMessage | HumanMessage
         SystemMessage(content=system_text)
     ]
 
-    recent = state.get("recent_turns") or []
-    for turn in recent[-RECENT_DIALOGUE_TURN_LIMIT:]:
-        role = (turn.get("role") or "").strip()
-        text = (turn.get("text") or "").strip()
-        if not text:
-            continue
-        if role == "player":
-            messages.append(HumanMessage(content=text))
-        elif role == "npc":
-            messages.append(AIMessage(content=text))
+    append_recent_dialogue_messages(messages, state.get("recent_turns"))
 
     player_message = state.get("player_message") or ""
     human_text = f"Player message: {player_message}\n\nRoom snapshot (JSON):\n{room_json}"
diff --git a/workers/agent-worker/src/graph/recall_merge.py b/workers/agent-worker/src/graph/recall_merge.py
index 81ce3f7..36123a6 100644
--- a/workers/agent-worker/src/graph/recall_merge.py
+++ b/workers/agent-worker/src/graph/recall_merge.py
@@ -318,27 +318,31 @@ def augment_retrieved_with_dialogue_turns(
     retrieved: list[dict[str, Any]] | None,
     recent_turns: list[dict[str, str]] | None,
 ) -> list[dict[str, Any]]:
-    """Merge in-session player turns ahead of DB/embed rows (memory tail may lag)."""
+    """Merge in-session player/npc turns ahead of DB/embed rows (memory tail may lag)."""
     if not recent_turns:
         return list(retrieved or [])
 
     seen: set[str] = set()
     merged: list[dict[str, Any]] = []
-    player_texts: list[str] = []
+    session_texts: list[str] = []
     for turn in reversed(recent_turns):
         role = (turn.get("role") or "").lower()
-        if role != "player":
-            continue
         raw = str(turn.get("text") or "").strip()
         if not raw:
             continue
-        text = raw if raw.lower().startswith("player:") else f"player: {raw}"
+        if role == "player":
+            text = raw if raw.lower().startswith("player:") else f"player: {raw}"
+        elif role == "npc":
+            clipped = raw[:120]
+            text = clipped if clipped.lower().startswith("npc:") else f"npc: {clipped}"
+        else:
+            continue
         if text in seen:
             continue
         seen.add(text)
-        player_texts.append(text)
+        session_texts.append(text)
 
-    for idx, text in enumerate(player_texts):
+    for idx, text in enumerate(session_texts):
         merged.append(
             {
                 "text": text,
diff --git a/workers/agent-worker/src/graph/speak_intent.py b/workers/agent-worker/src/graph/speak_intent.py
index 7f2086f..a8e3f1c 100644
--- a/workers/agent-worker/src/graph/speak_intent.py
+++ b/workers/agent-worker/src/graph/speak_intent.py
@@ -6,12 +6,12 @@
 from enum import Enum
 from typing import TYPE_CHECKING, Any
 
-from src.collective.social_turn import infer_social_from_message
+from src.collective.social_turn import infer_social_from_message, player_offers_help
 
 if TYPE_CHECKING:
     from src.collective.schemas import SocialTurnOut
 from src.graph.action_intent import player_requests_physical_action
-from src.graph.recall_merge import _RECALL_MARKERS, is_recall_question
+from src.graph.recall_merge import is_recall_question
 
 _CASUAL_GREETING_ONLY_RE = re.compile(
     r"^(你好(呀|啊|哦|呐|呢)?|嗨(呀|啊)?|hello|hi|hey|早上好|晚上好|下午好|在吗|在不在)([～!！?？。…]*)?$",
@@ -51,8 +51,7 @@ def classify_speak_intent(
     message: str,
     recent_turns: list | None = None,
 ) -> SpeakIntent:
-    """Order: PHYSICAL → RECALL → SOCIAL_EDGE → CASUAL → NARRATIVE (default)."""
-    del recent_turns  # reserved for future context-aware routing
+    """Order: PHYSICAL → RECALL → continuation → SOCIAL_EDGE → CASUAL → NARRATIVE (default)."""
     msg = (message or "").strip()
     if not msg:
         return SpeakIntent.NARRATIVE
@@ -60,6 +59,8 @@ def classify_speak_intent(
         return SpeakIntent.PHYSICAL
     if is_recall_question(msg):
         return SpeakIntent.RECALL
+    if recent_turns and len(msg) < 20 and infer_social_from_message(msg) is None:
+        return SpeakIntent.NARRATIVE
     if infer_social_from_message(msg) is not None:
         return SpeakIntent.SOCIAL_EDGE
     if is_casual_greeting_only(msg):
@@ -98,10 +99,13 @@ def can_use_casual_fast_lane(
     intent = classify_speak_intent(player_message, recent_turns)
     if intent != SpeakIntent.CASUAL:
         return intent, None
+    if recent_turns:
+        return intent, None
     turn: SocialTurnOut | None = _deterministic_social_turn(
         player_message,
         speak_intent=intent.value,
         npc_id=npc_id,
+        recent_turns=recent_turns,
     )
     if turn is None:
         return intent, None
@@ -120,10 +124,13 @@ def can_use_social_edge_fast_lane(
     intent = classify_speak_intent(player_message, recent_turns)
     if intent != SpeakIntent.SOCIAL_EDGE:
         return intent, None
+    if recent_turns or player_offers_help(player_message):
+        return intent, None
     turn: SocialTurnOut | None = _deterministic_social_turn(
         player_message,
         speak_intent=intent.value,
         npc_id=npc_id,
+        recent_turns=recent_turns,
     )
     if turn is None or turn.social.kind == "ignore":
         return intent, None
diff --git a/workers/agent-worker/tests/test_help_reply_by_npc.py b/workers/agent-worker/tests/test_help_reply_by_npc.py
index 4766f44..6fad39c 100644
--- a/workers/agent-worker/tests/test_help_reply_by_npc.py
+++ b/workers/agent-worker/tests/test_help_reply_by_npc.py
@@ -1,5 +1,5 @@
 from src.graph.nodes.llm_social_turn import _deterministic_social_turn
-from src.graph.speak_intent import can_use_social_edge_fast_lane
+from src.graph.speak_intent import SpeakIntent, can_use_social_edge_fast_lane
 
 
 def test_help_fast_lane_reply_varies_by_npc():
@@ -16,3 +16,44 @@ def test_farm_relay_not_help_fast_lane():
     intent, turn = can_use_social_edge_fast_lane(msg, npc_id="npc-2")
     assert turn is None
     assert intent.value == "physical"
+
+
+def test_help_offer_not_social_edge_fast_lane():
+    intent, turn = can_use_social_edge_fast_lane("我可以帮你！", npc_id="npc-4")
+    assert intent == SpeakIntent.NARRATIVE
+    assert turn is None
+
+
+def test_uat_hack_thread_offer_no_deterministic_stub():
+    history = [
+        {"role": "player", "text": "干嘛呢？"},
+        {"role": "npc", "text": "呀！我正在嚼着草莓棒棒糖，想着怎么才能把这里的系统黑掉"},
+    ]
+    intent, turn = can_use_social_edge_fast_lane("我可以帮你！", recent_turns=history, npc_id="npc-4")
+    assert turn is None
+    stub = _deterministic_social_turn("我可以帮你！", recent_turns=history, npc_id="npc-4")
+    assert stub is None
+
+
+def test_deterministic_casual_blocked_with_history():
+    history = [
+        {"role": "player", "text": "干嘛呢？"},
+        {"role": "npc", "text": "在忙"},
+    ]
+    msg = "你好，用一句话简短回复一下你的计划安排吧"
+    stub = _deterministic_social_turn(
+        msg,
+        speak_intent=SpeakIntent.CASUAL.value,
+        recent_turns=history,
+        npc_id="npc-1",
+    )
+    assert stub is None
+    # First-turn CASUAL still allowed
+    first = _deterministic_social_turn(
+        msg,
+        speak_intent=SpeakIntent.CASUAL.value,
+        recent_turns=[],
+        npc_id="npc-1",
+    )
+    assert first is not None
+    assert first.reply
diff --git a/workers/agent-worker/tests/test_llm_social_memory.py b/workers/agent-worker/tests/test_llm_social_memory.py
index 56c8a93..d175317 100644
--- a/workers/agent-worker/tests/test_llm_social_memory.py
+++ b/workers/agent-worker/tests/test_llm_social_memory.py
@@ -1,3 +1,5 @@
+from langchain_core.messages import AIMessage, HumanMessage
+
 from src.graph.nodes.llm_social_turn import _build_social_messages
 from src.graph.prompt import format_memory_summary
 from src.graph.state import GraphState
@@ -19,9 +21,34 @@ def test_build_social_messages_includes_memory_summary():
     }
     messages = _build_social_messages(state)
     system = messages[0].content
-    human = messages[1].content
+    human = messages[-1].content
     assert "Memory summary:" in system
     assert "门禁密码是 7" in system
     assert "禁止 meta 套话" in system
     assert "必须先写 reply" in system or "reply" in system
+    assert "我会尽力帮你" in system
     assert "Put \"reply\" as the first key" in human
+
+
+def test_build_social_messages_includes_recent_dialogue():
+    state: GraphState = {
+        "room_id": "default",
+        "player_message": "我可以帮你！",
+        "npc_id": "npc-4",
+        "player_id": "p1",
+        "recent_turns": [
+            {"role": "player", "text": "干嘛呢？"},
+            {
+                "role": "npc",
+                "text": "呀！我正在嚼着草莓棒棒糖，想着怎么才能把这里的系统黑掉",
+            },
+        ],
+        "room_snapshot": {"width": 8, "height": 8, "player": {"x": 1, "y": 1}, "npcs": []},
+    }
+    messages = _build_social_messages(state)
+    assert len(messages) == 4
+    assert isinstance(messages[1], HumanMessage)
+    assert messages[1].content == "干嘛呢？"
+    assert isinstance(messages[2], AIMessage)
+    assert "系统黑掉" in messages[2].content
+    assert "我可以帮你！" in messages[-1].content
diff --git a/workers/agent-worker/tests/test_recall_merge.py b/workers/agent-worker/tests/test_recall_merge.py
index 232ac13..7cb8f11 100644
--- a/workers/agent-worker/tests/test_recall_merge.py
+++ b/workers/agent-worker/tests/test_recall_merge.py
@@ -140,6 +140,21 @@ def test_extract_food_preference():
     assert extract_food_preference("npc: 你没告诉过我你喜欢吃什么。") is None
 
 
+def test_augment_dialogue_turns_includes_npc_lines():
+    recent_turns = [
+        {"role": "player", "text": "干嘛呢？"},
+        {
+            "role": "npc",
+            "text": "呀！我正在嚼着草莓棒棒糖，想着怎么才能把这里的系统黑掉，超级有趣的对吧！",
+        },
+    ]
+    merged = augment_retrieved_with_dialogue_turns([], recent_turns)
+    texts = [row["text"] for row in merged]
+    assert any(t.startswith("player:") for t in texts)
+    assert any(t.startswith("npc:") and "系统黑掉" in t for t in texts)
+    assert all(len(t) <= 130 for t in texts)
+
+
 def test_merge_recall_mango_over_watermelon_with_dialogue_turns():
     """Mango disclosed in-session but not yet in DB — dialogue turns must win."""
     db_augmented = [
diff --git a/workers/agent-worker/tests/test_social_turn.py b/workers/agent-worker/tests/test_social_turn.py
index 9fdb676..7e7603d 100644
--- a/workers/agent-worker/tests/test_social_turn.py
+++ b/workers/agent-worker/tests/test_social_turn.py
@@ -9,6 +9,7 @@
     compute_applied_delta,
     infer_social_from_message,
     personality_multiplier,
+    player_offers_help,
     reconcile_social_perception,
     refresh_collective_snapshot,
 )
@@ -131,3 +132,38 @@ def test_infer_help_not_triggered_by_hui_fu_substring():
     assert infer_social_from_message("你好，用一句话简短回复") is None
     assert infer_social_from_message("请帮帮我") is not None
     assert infer_social_from_message("请帮帮我").kind == "help"
+
+
+def test_player_offers_help_detection():
+    assert player_offers_help("我可以帮你！")
+    assert player_offers_help("我能帮你")
+    assert player_offers_help("我愿意帮你")
+    assert player_offers_help("我想帮你")
+    assert player_offers_help("我来帮")
+    assert player_offers_help("让我帮你")
+    assert player_offers_help("我帮你看看")
+    assert not player_offers_help("请帮帮忙")
+    assert not player_offers_help("帮帮我")
+    assert not player_offers_help("你能帮我吗")
+
+
+def test_infer_help_offer_not_request():
+    assert infer_social_from_message("我可以帮你！") is None
+    assert infer_social_from_message("我能帮你") is None
+    assert infer_social_from_message("我愿意帮你") is None
+    assert infer_social_from_message("我想帮你") is None
+    assert infer_social_from_message("我来帮") is None
+    assert infer_social_from_message("让我帮你") is None
+
+
+def test_infer_help_request_still_detected():
+    for msg in ("请帮帮忙", "帮帮我", "你能帮我吗", "能请你帮个忙吗"):
+        inferred = infer_social_from_message(msg)
+        assert inferred is not None
+        assert inferred.kind == "help"
+        assert inferred.summary == "玩家请求帮助"
+
+
+def test_infer_help_not_triggered_by_narrative_bang_compounds():
+    assert infer_social_from_message("帮别人做事") is None
+    assert infer_social_from_message("别在这里帮腔") is None
diff --git a/workers/agent-worker/tests/test_speak_intent.py b/workers/agent-worker/tests/test_speak_intent.py
index af30e48..4317e66 100644
--- a/workers/agent-worker/tests/test_speak_intent.py
+++ b/workers/agent-worker/tests/test_speak_intent.py
@@ -38,6 +38,12 @@ def test_social_edge_intent():
     assert classify_speak_intent("能请你帮个忙吗") == SpeakIntent.SOCIAL_EDGE
 
 
+def test_help_offer_not_social_edge():
+    assert classify_speak_intent("我可以帮你！") == SpeakIntent.NARRATIVE
+    assert classify_speak_intent("我来帮") == SpeakIntent.NARRATIVE
+    assert classify_speak_intent("让我帮你") == SpeakIntent.NARRATIVE
+
+
 def test_casual_intent():
     assert classify_speak_intent("你好") == SpeakIntent.CASUAL
     assert classify_speak_intent("Hi") == SpeakIntent.CASUAL
@@ -73,6 +79,24 @@ def test_can_use_casual_fast_lane_b1():
     assert turn.reply
 
 
+def test_can_use_casual_fast_lane_b1_empty_history():
+    intent, turn = can_use_casual_fast_lane("你好", recent_turns=[])
+    assert intent == SpeakIntent.CASUAL
+    assert turn is not None
+
+
+def test_can_use_casual_fast_lane_blocked_with_history():
+    history = [{"role": "player", "text": "干嘛呢？"}, {"role": "npc", "text": "在忙"}]
+    intent, turn = can_use_casual_fast_lane("你好", recent_turns=history)
+    assert intent == SpeakIntent.NARRATIVE
+    assert turn is None
+
+
+def test_continuation_short_with_history():
+    history = [{"role": "player", "text": "干嘛呢？"}, {"role": "npc", "text": "在忙"}]
+    assert classify_speak_intent("好的", recent_turns=history) == SpeakIntent.NARRATIVE
+
+
 def test_can_use_casual_fast_lane_recall_blocked():
     intent, turn = can_use_casual_fast_lane("你好，还记得密码吗")
     assert intent == SpeakIntent.RECALL
@@ -113,6 +137,13 @@ def test_can_use_social_edge_fast_lane_help():
     assert turn.social.kind == "help"
 
 
+def test_can_use_social_edge_fast_lane_help_blocked_with_history():
+    history = [{"role": "player", "text": "干嘛呢？"}, {"role": "npc", "text": "在忙"}]
+    intent, turn = can_use_social_edge_fast_lane("请帮帮忙", recent_turns=history)
+    assert intent == SpeakIntent.SOCIAL_EDGE
+    assert turn is None
+
+
 def test_can_use_social_edge_fast_lane_narrative_blocked():
     intent, turn = can_use_social_edge_fast_lane("故宫在哪里，给我讲讲历史")
     assert intent == SpeakIntent.NARRATIVE