Fix vector recall pending user message

modules/story-summary/generate/prompt.js

@@ -1,17 +1,10 @@
 // ═══════════════════════════════════════════════════════════════════════════
 // Story Summary - Prompt Injection (Final Clean Version)
-// - 注入只在 GENERATION_STARTED 发生（由 story-summary.js 调用）
+// - 仅负责“构建注入文本”，不负责写入 extension_prompts
-// - 向量关闭：注入全量总结（世界/事件/弧光）
+// - 注入发生在 story-summary.js 的 CHAT_COMPLETION_PROMPT_READY（最终 messages 已裁剪）
-// - 向量开启：召回 + 预算装配注入
-// - 没有“快速注入”写入 extension_prompts，避免覆盖/残留/竞态
 // ═══════════════════════════════════════════════════════════════════════════
 
 import { getContext } from "../../../../../../extensions.js";
-import {
-    extension_prompts,
-    extension_prompt_types,
-    extension_prompt_roles,
-} from "../../../../../../../script.js";
 import { xbLog } from "../../../core/debug-core.js";
 import { getSummaryStore } from "../data/store.js";
 import { getVectorConfig, getSummaryPanelConfig, getSettings } from "../data/config.js";
@@ -19,7 +12,6 @@ import { recallMemory, buildQueryText } from "../vector/recall.js";
 import { getChunksByFloors, getAllChunkVectors, getAllEventVectors, getMeta } from "../vector/chunk-store.js";
 
 const MODULE_ID = "summaryPrompt";
-const SUMMARY_PROMPT_KEY = "LittleWhiteBox_StorySummary";
 
 // ─────────────────────────────────────────────────────────────────────────────
 // 召回失败提示节流（避免连续生成刷屏）
@@ -137,7 +129,8 @@ function formatArcLine(a) {
 
 // 完整 chunk 输出（不截断）
 function formatChunkFullLine(c) {
-    const speaker = c.isUser ? "{{user}}" : "{{char}}";
+    const { name1, name2 } = getContext();
+    const speaker = c.isUser ? (name1 || "用户") : (name2 || "角色");
     return `› #${c.floor + 1} [${speaker}] ${String(c.text || "").trim()}`;
 }
 
@@ -299,23 +292,19 @@ function buildNonVectorPrompt(store) {
     );
 }
 
-export async function injectNonVectorPrompt(postToFrame = null) {
+export function buildNonVectorPromptText() {
     if (!getSettings().storySummary?.enabled) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return "";
-        return;
     }
 
-    const { chat } = getContext();
     const store = getSummaryStore();
     if (!store?.json) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return "";
-        return;
     }
 
     let text = buildNonVectorPrompt(store);
     if (!text.trim()) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return "";
-        return;
     }
 
     // wrapper（沿用面板设置）
@@ -323,21 +312,7 @@ export async function injectNonVectorPrompt(postToFrame = null) {
     if (cfg.trigger?.wrapperHead) text = cfg.trigger.wrapperHead + "\n" + text;
     if (cfg.trigger?.wrapperTail) text = text + "\n" + cfg.trigger.wrapperTail;
 
-    const lastIdx = store.lastSummarizedMesId ?? 0;
+    return text;
-    let depth = (chat?.length || 0) - lastIdx - 1;
-    if (depth < 0) depth = 0;
-    if (cfg.trigger?.forceInsertAtEnd) depth = 10000;
-
-    extension_prompts[SUMMARY_PROMPT_KEY] = {
-        value: text,
-        position: extension_prompt_types.IN_CHAT,
-        depth,
-        role: extension_prompt_roles.SYSTEM,
-    };
-
-    if (postToFrame) {
-        postToFrame({ type: "RECALL_LOG", text: "\n[Non-vector] Injected full summary prompt.\n" });
-    }
 }
 
 // ─────────────────────────────────────────────────────────────
@@ -706,19 +681,17 @@ async function attachEvidenceToCausalEvents(causalEvents, eventVectorMap, chunkV
 // ✅ 向量模式：召回 + 注入（供 story-summary.js 在 GENERATION_STARTED 调用）
 // ─────────────────────────────────────────────────────────────────────────────
 
-export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
+export async function buildVectorPromptText(excludeLastAi = false, hooks = {}) {
-    const { postToFrame = null, echo = null } = hooks;
+    const { postToFrame = null, echo = null, pendingUserMessage = null } = hooks;
     if (!getSettings().storySummary?.enabled) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return { text: "", logText: "" };
-        return;
     }
 
     const { chat } = getContext();
     const store = getSummaryStore();
 
     if (!store?.json) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return { text: "", logText: "" };
-        return;
     }
 
     const allEvents = store.json.events || [];
@@ -726,14 +699,12 @@ export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
     const length = chat?.length || 0;
 
     if (lastIdx >= length) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return { text: "", logText: "" };
-        return;
     }
 
     const vectorCfg = getVectorConfig();
     if (!vectorCfg?.enabled) {
-        // 向量没开，不该走这条
+        return { text: "", logText: "" };
-        return;
     }
 
     const { chatId } = getContext();
@@ -745,7 +716,10 @@ export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
 
     try {
         const queryText = buildQueryText(chat, 2, excludeLastAi);
-        recallResult = await recallMemory(queryText, allEvents, vectorCfg, { excludeLastAi });
+        recallResult = await recallMemory(queryText, allEvents, vectorCfg, {
+            excludeLastAi,
+            pendingUserMessage,
+        });
 
         recallResult = {
             ...recallResult,
@@ -808,9 +782,7 @@ export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
             });
         }
 
-        // 清空本次注入，避免残留误导
+        return { text: "", logText: `\n[Vector Recall Failed]\n${String(e?.stack || e?.message || e)}\n` };
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
-        return;
     }
 
     // 成功但结果为空：也提示，并清空注入（不降级）
@@ -831,8 +803,7 @@ export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
                 text: "\n[Vector Recall Empty]\nNo recall candidates / vectors not ready.\n",
             });
         }
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
+        return { text: "", logText: "\n[Vector Recall Empty]\nNo recall candidates / vectors not ready.\n" };
-        return;
     }
 
     // 拼装向量 prompt
@@ -844,50 +815,17 @@ export async function recallAndInjectPrompt(excludeLastAi = false, hooks = {}) {
         meta
     );
 
-    // 写入 extension_prompts（真正注入）
+    // wrapper（沿用面板设置）——必须补回，否则语义回退
-    await writePromptToExtensionPrompts(promptText, store, chat);
+    const cfg = getSummaryPanelConfig();
+    let finalText = String(promptText || "");
+    if (cfg.trigger?.wrapperHead) finalText = cfg.trigger.wrapperHead + "\n" + finalText;
+    if (cfg.trigger?.wrapperTail) finalText = finalText + "\n" + cfg.trigger.wrapperTail;
 
     // 发给涌现窗口：召回报告 + 装配报告
     if (postToFrame) {
         const recallLog = recallResult.logText || "";
         postToFrame({ type: "RECALL_LOG", text: recallLog + (injectionLogText || "") });
     }
-}
+
-
+    return { text: finalText, logText: (recallResult.logText || "") + (injectionLogText || "") };
-// ─────────────────────────────────────────────────────────────────────────────
-// 写入 extension_prompts（统一入口）
-// ─────────────────────────────────────────────────────────────────────────────
-
-async function writePromptToExtensionPrompts(text, store, chat) {
-    const cfg = getSummaryPanelConfig();
-    let finalText = String(text || "");
-
-    if (cfg.trigger?.wrapperHead) finalText = cfg.trigger.wrapperHead + "\n" + finalText;
-    if (cfg.trigger?.wrapperTail) finalText = finalText + "\n" + cfg.trigger.wrapperTail;
-
-    if (!finalText.trim()) {
-        delete extension_prompts[SUMMARY_PROMPT_KEY];
-        return;
-    }
-
-    const lastIdx = store.lastSummarizedMesId ?? 0;
-    let depth = (chat?.length || 0) - lastIdx - 1;
-    if (depth < 0) depth = 0;
-
-    if (cfg.trigger?.forceInsertAtEnd) depth = 10000;
-
-    extension_prompts[SUMMARY_PROMPT_KEY] = {
-        value: finalText,
-        position: extension_prompt_types.IN_CHAT,
-        depth,
-        role: extension_prompt_roles.SYSTEM,
-    };
-}
-
-// ─────────────────────────────────────────────────────────────────────────────
-// 清理 prompt（供 story-summary.js 调用）
-// ─────────────────────────────────────────────────────────────────────────────
-
-export function clearSummaryExtensionPrompt() {
-    delete extension_prompts[SUMMARY_PROMPT_KEY];
 }

modules/story-summary/story-summary.js

@@ -1,9 +1,12 @@
 // ═══════════════════════════════════════════════════════════════════════════
-// Story Summary - 主入口（干净版）
+// Story Summary - 主入口（最终版）
-// - 注入只在 GENERATION_STARTED 发生
+//
-// - 向量关闭：注入全量总结（L3+L2+Arcs）
+// 稳定目标：
-// - 向量开启：召回 + 1万预算装配注入
+// 1) "聊天时隐藏已总结" 永远只隐藏"已总结"部分，绝不影响未总结部分
-// - 删除所有 updateSummaryExtensionPrompt() 调用，避免覆盖/残留/竞态
+// 2) 关闭隐藏 = 暴力全量 unhide，确保立刻恢复
+// 3) 开启隐藏 / 改Y / 切Chat / 收新消息：先全量 unhide，再按边界重新 hide
+// 4) Prompt 注入位置稳定：永远插在"最后一条 user 消息"之前
+// 5) 注入不依赖 extension_prompts/depth，不受 ST 裁剪/隐藏参照系影响
 // ═══════════════════════════════════════════════════════════════════════════
 
 import { getContext } from "../../../../../extensions.js";
@@ -23,11 +26,10 @@ import {
     clearSummaryData,
 } from "./data/store.js";
 
-// prompt injection (ONLY on generation started)
+// prompt text builder
 import {
-    recallAndInjectPrompt,
+    buildVectorPromptText,
-    clearSummaryExtensionPrompt,
+    buildNonVectorPromptText,
-    injectNonVectorPrompt,
 } from "./generate/prompt.js";
 
 // summary generation
@@ -91,6 +93,13 @@ let pendingFrameMessages = [];
 let eventsRegistered = false;
 let vectorGenerating = false;
 let vectorCancelled = false;
+let vectorAbortController = null;
+
+let pendingInjectText = "";
+
+// ★ 用户消息缓存（解决 GENERATION_STARTED 时 chat 尚未包含用户消息的问题）
+let lastSentUserMessage = null;
+let lastSentTimestamp = 0;
 
 let hideApplyTimer = null;
 const HIDE_APPLY_DEBOUNCE_MS = 250;
@@ -118,6 +127,18 @@ async function executeSlashCommand(command) {
     }
 }
 
+function getLastMessageId() {
+    const { chat } = getContext();
+    const len = Array.isArray(chat) ? chat.length : 0;
+    return Math.max(-1, len - 1);
+}
+
+async function unhideAllMessages() {
+    const last = getLastMessageId();
+    if (last < 0) return;
+    await executeSlashCommand(`/unhide 0-${last}`);
+}
+
 // ═══════════════════════════════════════════════════════════════════════════
 // 生成状态管理
 // ═══════════════════════════════════════════════════════════════════════════
@@ -302,6 +323,8 @@ async function handleGenerateVectors(vectorCfg) {
 
     vectorGenerating = true;
     vectorCancelled = false;
+    vectorAbortController?.abort?.();
+    vectorAbortController = new AbortController();
 
     const fingerprint = getEngineFingerprint(vectorCfg);
     const isLocal = vectorCfg.engine === "local";
@@ -362,6 +385,8 @@ async function handleGenerateVectors(vectorCfg) {
     postToFrame({ type: "VECTOR_GEN_PROGRESS", phase: "L1", current: 0, total: l1Total });
     postToFrame({ type: "VECTOR_GEN_PROGRESS", phase: "L2", current: l2Completed, total: l2Total });
 
+    let rateLimitWarned = false;
+
     const allTasks = [...l1Batches, ...l2Batches];
     const l1Vectors = new Array(l1Texts.length);
     const l2VectorItems = [];
@@ -371,6 +396,7 @@ async function handleGenerateVectors(vectorCfg) {
     async function worker() {
         while (taskIndex < allTasks.length) {
             if (vectorCancelled) break;
+            if (vectorAbortController?.signal?.aborted) break;
 
             const i = taskIndex++;
             if (i >= allTasks.length) break;
@@ -378,7 +404,7 @@ async function handleGenerateVectors(vectorCfg) {
             const task = allTasks[i];
 
             try {
-                const vectors = await embed(task.texts, vectorCfg);
+                const vectors = await embed(task.texts, vectorCfg, { signal: vectorAbortController.signal });
 
                 if (task.phase === "L1") {
                     for (let j = 0; j < vectors.length; j++) {
@@ -404,7 +430,23 @@ async function handleGenerateVectors(vectorCfg) {
                     });
                 }
             } catch (e) {
+                if (e?.name === "AbortError") {
+                    xbLog.warn(MODULE_ID, "向量生成已取消（AbortError）");
+                    break;
+                }
+
                 xbLog.error(MODULE_ID, `${task.phase} batch 向量化失败`, e);
+
+                const msg = String(e?.message || e);
+                const isRateLike = /429|403|rate|limit|quota/i.test(msg);
+                if (isRateLike && !rateLimitWarned) {
+                    rateLimitWarned = true;
+                    executeSlashCommand("/echo severity=warning 向量生成遇到速率/配额限制，已进入自动重试。");
+                }
+
+                vectorCancelled = true;
+                vectorAbortController?.abort?.();
+                break;
             }
         }
     }
@@ -415,6 +457,13 @@ async function handleGenerateVectors(vectorCfg) {
             .map(() => worker())
     );
 
+    if (vectorCancelled || vectorAbortController?.signal?.aborted) {
+        postToFrame({ type: "VECTOR_GEN_PROGRESS", phase: "L1", current: -1, total: 0 });
+        postToFrame({ type: "VECTOR_GEN_PROGRESS", phase: "L2", current: -1, total: 0 });
+        vectorGenerating = false;
+        return;
+    }
+
     if (allChunks.length > 0 && l1Vectors.filter(Boolean).length > 0) {
         const chunkVectorItems = allChunks
             .map((chunk, idx) => (l1Vectors[idx] ? { chunkId: chunk.chunkId, vector: l1Vectors[idx] } : null))
@@ -433,6 +482,7 @@ async function handleGenerateVectors(vectorCfg) {
 
     vectorGenerating = false;
     vectorCancelled = false;
+    vectorAbortController = null;
 
     xbLog.info(MODULE_ID, `向量生成完成: L1=${l1Vectors.filter(Boolean).length}, L2=${l2VectorItems.length}`);
 }
@@ -463,8 +513,6 @@ async function maybeAutoBuildChunks() {
         if (!isLocalModelLoaded(modelId)) return;
     }
 
-    xbLog.info(MODULE_ID, `auto L1 chunks: pending=${status.pending}`);
-
     try {
         await buildIncrementalChunks({ vectorConfig: cfg });
     } catch (e) {
@@ -579,7 +627,6 @@ async function sendSavedConfigToFrame() {
         const savedConfig = await CommonSettingStorage.get(SUMMARY_CONFIG_KEY, null);
         if (savedConfig) {
             postToFrame({ type: "LOAD_PANEL_CONFIG", config: savedConfig });
-            xbLog.info(MODULE_ID, "已从服务器加载面板配置");
         }
     } catch (e) {
         xbLog.warn(MODULE_ID, "加载面板配置失败", e);
@@ -633,7 +680,7 @@ function openPanelForMessage(mesId) {
     const store = getSummaryStore();
     const totalFloors = chat.length;
 
-    sendFrameBaseData(store, totalFloors); // 不需要 await，fire-and-forget
+    sendFrameBaseData(store, totalFloors);
     sendFrameFullData(store, totalFloors);
     setSummaryGenerating(summaryGenerating);
 
@@ -646,9 +693,9 @@ function openPanelForMessage(mesId) {
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
-// Hide/Unhide：向量模式联动（"已总结"的定义切换）
+// Hide/Unhide
-// - 非向量：boundary = lastSummarizedMesId（LLM总结边界）
+// - 非向量：boundary = lastSummarizedMesId
-// - 向量：boundary = meta.lastChunkFloor（已向量化）
+// - 向量：boundary = meta.lastChunkFloor（若为 -1 则回退到 lastSummarizedMesId）
 // ═══════════════════════════════════════════════════════════════════════════
 
 async function getHideBoundaryFloor(store) {
@@ -658,16 +705,21 @@ async function getHideBoundaryFloor(store) {
     }
 
     const { chatId } = getContext();
-    if (!chatId) return -1;
+    if (!chatId) return store?.lastSummarizedMesId ?? -1;
 
     const meta = await getMeta(chatId);
-    return meta?.lastChunkFloor ?? -1;
+    const v = meta?.lastChunkFloor ?? -1;
+    if (v >= 0) return v;
+    return store?.lastSummarizedMesId ?? -1;
 }
 
 async function applyHideState() {
     const store = getSummaryStore();
     if (!store?.hideSummarizedHistory) return;
 
+    // 先全量 unhide，杜绝历史残留
+    await unhideAllMessages();
+
     const boundary = await getHideBoundaryFloor(store);
     if (boundary < 0) return;
 
@@ -685,15 +737,12 @@ function applyHideStateDebounced() {
 }
 
 async function clearHideState() {
-    const store = getSummaryStore();
+    // 暴力全量 unhide，确保立刻恢复
-    const boundary = await getHideBoundaryFloor(store);
+    await unhideAllMessages();
-    if (boundary < 0) return;
-
-    await executeSlashCommand(`/unhide 0-${boundary}`);
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
-// 自动总结（保持原逻辑；不做 prompt 注入）
+// 自动总结
 // ═══════════════════════════════════════════════════════════════════════════
 
 async function maybeAutoRunSummary(reason) {
@@ -716,7 +765,6 @@ async function maybeAutoRunSummary(reason) {
     const pending = chat.length - lastSummarized - 1;
     if (pending < (trig.interval || 1)) return;
 
-    xbLog.info(MODULE_ID, `自动触发剧情总结: reason=${reason}, pending=${pending}`);
     await autoRunSummaryWithRetry(chat.length - 1, { api: cfgAll.api, gen: cfgAll.gen, trigger: trig });
 }
 
@@ -739,10 +787,8 @@ async function autoRunSummaryWithRetry(targetMesId, configForRun) {
                         lastSummarizedMesId: endMesId,
                     },
                 });
-                postToFrame({
+
-                    type: "SUMMARY_STATUS",
+                applyHideStateDebounced();
-                    statusText: `已更新至 ${endMesId + 1} 楼 · ${merged.events?.length || 0} 个事件`,
-                });
                 updateFrameStatsAfterSummary(endMesId, merged);
             },
         });
@@ -756,7 +802,6 @@ async function autoRunSummaryWithRetry(targetMesId, configForRun) {
     }
 
     setSummaryGenerating(false);
-    xbLog.error(MODULE_ID, "自动总结失败（已重试3次）");
     await executeSlashCommand("/echo severity=error 剧情总结失败（已自动重试 3 次）。请稍后再试。");
 }
 
@@ -829,7 +874,6 @@ function handleFrameMessage(event) {
             postToFrame({ type: "SUMMARY_STATUS", statusText: "已停止" });
             break;
 
-        // vector UI
         case "VECTOR_DOWNLOAD_MODEL":
             handleDownloadLocalModel(data.modelId);
             break;
@@ -856,24 +900,21 @@ function handleFrameMessage(event) {
 
         case "VECTOR_GENERATE":
             if (data.config) saveVectorConfig(data.config);
-            clearSummaryExtensionPrompt(); // 防残留
             handleGenerateVectors(data.config);
             break;
 
         case "VECTOR_CLEAR":
-            clearSummaryExtensionPrompt(); // 防残留
             handleClearVectors();
             break;
 
         case "VECTOR_CANCEL_GENERATE":
             vectorCancelled = true;
+            try { vectorAbortController?.abort?.(); } catch {}
             break;
 
-        // summary actions
         case "REQUEST_CLEAR": {
             const { chat, chatId } = getContext();
             clearSummaryData(chatId);
-            clearSummaryExtensionPrompt(); // 防残留
             postToFrame({
                 type: "SUMMARY_CLEARED",
                 payload: { totalFloors: Array.isArray(chat) ? chat.length : 0 },
@@ -920,16 +961,13 @@ function handleFrameMessage(event) {
 
             const oldCount = store.keepVisibleCount ?? 3;
             const newCount = Math.max(0, Math.min(50, parseInt(data.count) || 3));
-
             if (newCount === oldCount) break;
 
             store.keepVisibleCount = newCount;
             saveSummaryStore();
 
             (async () => {
-                // 先清掉原隐藏，再按新 keepCount 重算隐藏
                 if (store.hideSummarizedHistory) {
-                    await clearHideState();
                     await applyHideState();
                 }
                 const { chat } = getContext();
@@ -941,8 +979,6 @@ function handleFrameMessage(event) {
         case "SAVE_PANEL_CONFIG":
             if (data.config) {
                 CommonSettingStorage.set(SUMMARY_CONFIG_KEY, data.config);
-                clearSummaryExtensionPrompt(); // 配置变化立即清除注入，避免残留
-                xbLog.info(MODULE_ID, "面板配置已保存到服务器");
             }
             break;
 
@@ -963,7 +999,6 @@ async function handleManualGenerate(mesId, config) {
     }
 
     setSummaryGenerating(true);
-    xbLog.info(MODULE_ID, `开始手动总结 mesId=${mesId}`);
 
     await runSummaryGeneration(mesId, config, {
         onStatus: (text) => postToFrame({ type: "SUMMARY_STATUS", statusText: text }),
@@ -981,14 +1016,7 @@ async function handleManualGenerate(mesId, config) {
                 },
             });
 
-            postToFrame({
-                type: "SUMMARY_STATUS",
-                statusText: `已更新至 ${endMesId + 1} 楼 · ${merged.events?.length || 0} 个事件`,
-            });
-
-            // 隐藏逻辑：统一走 boundary（vector on/off 自动切换定义）
             applyHideStateDebounced();
-
             updateFrameStatsAfterSummary(endMesId, merged);
         },
     });
@@ -997,7 +1025,7 @@ async function handleManualGenerate(mesId, config) {
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
-// 事件处理器（不做 prompt 注入）
+// 消息事件
 // ═══════════════════════════════════════════════════════════════════════════
 
 async function handleChatChanged() {
@@ -1008,7 +1036,10 @@ async function handleChatChanged() {
     initButtonsForAll();
 
     const store = getSummaryStore();
-    applyHideStateDebounced();
+    
+    if (store?.hideSummarizedHistory) {
+        await applyHideState();
+    }
 
     if (frameReady) {
         await sendFrameBaseData(store, newLength);
@@ -1022,6 +1053,7 @@ async function handleMessageDeleted() {
 
     await rollbackSummaryIfNeeded();
     await syncOnMessageDeleted(chatId, newLength);
+    applyHideStateDebounced();
 }
 
 async function handleMessageSwiped() {
@@ -1030,6 +1062,7 @@ async function handleMessageSwiped() {
 
     await syncOnMessageSwiped(chatId, lastFloor);
     initButtonsForAll();
+    applyHideStateDebounced();
 }
 
 async function handleMessageReceived() {
@@ -1043,9 +1076,7 @@ async function handleMessageReceived() {
     await syncOnMessageReceived(chatId, lastFloor, message, vectorConfig);
     await maybeAutoBuildChunks();
 
-    // 向量模式下，lastChunkFloor 会持续推进；如果勾选隐藏，自动扩展隐藏范围
     applyHideStateDebounced();
-
     setTimeout(() => maybeAutoRunSummary("after_ai"), 1000);
 }
 
@@ -1057,6 +1088,7 @@ function handleMessageSent() {
 async function handleMessageUpdated() {
     await rollbackSummaryIfNeeded();
     initButtonsForAll();
+    applyHideStateDebounced();
 }
 
 function handleMessageRendered(data) {
@@ -1066,7 +1098,20 @@ function handleMessageRendered(data) {
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
-// ✅ 唯一注入入口：GENERATION_STARTED
+// 用户消息缓存（供向量召回使用）
+// ═══════════════════════════════════════════════════════════════════════════
+
+function handleMessageSentForRecall() {
+    const { chat } = getContext();
+    const lastMsg = chat?.[chat.length - 1];
+    if (lastMsg?.is_user) {
+        lastSentUserMessage = lastMsg.mes;
+        lastSentTimestamp = Date.now();
+    }
+}
+
+// ═══════════════════════════════════════════════════════════════════════════
+// Prompt 注入
 // ═══════════════════════════════════════════════════════════════════════════
 
 async function handleGenerationStarted(type, _params, isDryRun) {
@@ -1076,17 +1121,62 @@ async function handleGenerationStarted(type, _params, isDryRun) {
     const excludeLastAi = type === "swipe" || type === "regenerate";
     const vectorCfg = getVectorConfig();
 
-    // 向量模式：召回 + 预算装配
+    pendingInjectText = "";
+
+    // ★ 判断是否使用缓存的用户消息（30秒内有效）
+    let pendingUserMessage = null;
+    if (type === "normal" && lastSentUserMessage && (Date.now() - lastSentTimestamp < 30000)) {
+        pendingUserMessage = lastSentUserMessage;
+    }
+    // 用完清空
+    lastSentUserMessage = null;
+    lastSentTimestamp = 0;
+
     if (vectorCfg?.enabled) {
-        await recallAndInjectPrompt(excludeLastAi, {
+        const r = await buildVectorPromptText(excludeLastAi, {
             postToFrame,
-            echo: executeSlashCommand, // recall failure notification
+            echo: executeSlashCommand,
+            pendingUserMessage,
         });
+        pendingInjectText = r?.text || "";
         return;
     }
 
-    // 非向量模式：全量总结注入（不召回）
+    pendingInjectText = buildNonVectorPromptText() || "";
-    await injectNonVectorPrompt(postToFrame);
+}
+
+function clearPendingInject() {
+    pendingInjectText = "";
+}
+
+function findInsertIndexBeforeLastUserMessage(chat) {
+    if (!Array.isArray(chat) || !chat.length) return 0;
+    for (let i = chat.length - 1; i >= 0; i--) {
+        if (chat[i]?.role === "user") return i;
+    }
+    // 没有 user，退化为末尾前一位
+    return Math.max(0, chat.length - 1);
+}
+
+function handleChatCompletionPromptReady(eventData) {
+    try {
+        if (!pendingInjectText?.trim()) return;
+        if (!eventData || eventData.dryRun) return;
+        if (!Array.isArray(eventData.chat)) return;
+
+        // 永远插在最后一条 user 消息之前，保证三段结构稳定
+        const insertAt = findInsertIndexBeforeLastUserMessage(eventData.chat);
+
+        eventData.chat.splice(insertAt, 0, {
+            role: "assistant",
+            content: pendingInjectText,
+        });
+
+        clearPendingInject();
+    } catch (e) {
+        xbLog.error(MODULE_ID, "Prompt inject failed", e);
+        clearPendingInject();
+    }
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
@@ -1097,8 +1187,6 @@ function registerEvents() {
     if (eventsRegistered) return;
     eventsRegistered = true;
 
-    xbLog.info(MODULE_ID, "模块初始化");
-
     CacheRegistry.register(MODULE_ID, {
         name: "待发送消息队列",
         getSize: () => pendingFrameMessages.length,
@@ -1121,26 +1209,28 @@ function registerEvents() {
     eventSource.on(event_types.MESSAGE_DELETED, () => setTimeout(handleMessageDeleted, 50));
     eventSource.on(event_types.MESSAGE_RECEIVED, () => setTimeout(handleMessageReceived, 150));
     eventSource.on(event_types.MESSAGE_SENT, () => setTimeout(handleMessageSent, 150));
+    eventSource.on(event_types.MESSAGE_SENT, handleMessageSentForRecall);
     eventSource.on(event_types.MESSAGE_SWIPED, () => setTimeout(handleMessageSwiped, 100));
     eventSource.on(event_types.MESSAGE_UPDATED, () => setTimeout(handleMessageUpdated, 100));
     eventSource.on(event_types.MESSAGE_EDITED, () => setTimeout(handleMessageUpdated, 100));
     eventSource.on(event_types.USER_MESSAGE_RENDERED, (data) => setTimeout(() => handleMessageRendered(data), 50));
     eventSource.on(event_types.CHARACTER_MESSAGE_RENDERED, (data) => setTimeout(() => handleMessageRendered(data), 50));
 
-    // ✅ 只在生成开始时注入
+    // 注入链路
     eventSource.on(event_types.GENERATION_STARTED, handleGenerationStarted);
+    eventSource.on(event_types.GENERATION_STOPPED, clearPendingInject);
+    eventSource.on(event_types.GENERATION_ENDED, clearPendingInject);
+    eventSource.on(event_types.CHAT_COMPLETION_PROMPT_READY, handleChatCompletionPromptReady);
 }
 
 function unregisterEvents() {
-    xbLog.info(MODULE_ID, "模块清理");
     CacheRegistry.unregister(MODULE_ID);
     eventsRegistered = false;
 
     $(".xiaobaix-story-summary-btn").remove();
     hideOverlay();
 
-    // 禁用时清理注入，避免残留
+    clearPendingInject();
-    clearSummaryExtensionPrompt();
 }
 
 // ═══════════════════════════════════════════════════════════════════════════
@@ -1151,9 +1241,6 @@ $(document).on("xiaobaix:storySummary:toggle", (_e, enabled) => {
     if (enabled) {
         registerEvents();
         initButtonsForAll();
-
-        // 开启时清一次，防止旧注入残留
-        clearSummaryExtensionPrompt();
     } else {
         unregisterEvents();
     }
@@ -1164,12 +1251,6 @@ $(document).on("xiaobaix:storySummary:toggle", (_e, enabled) => {
 // ═══════════════════════════════════════════════════════════════════════════
 
 jQuery(() => {
-    if (!getSettings().storySummary?.enabled) {
+    if (!getSettings().storySummary?.enabled) return;
-        clearSummaryExtensionPrompt();
-        return;
-    }
     registerEvents();
-
-    // 初始化也清一次，保证干净（注入只在生成开始发生）
-    clearSummaryExtensionPrompt();
 });

modules/story-summary/vector/recall.js

@@ -158,7 +158,7 @@ function buildExpDecayWeights(n, beta) {
 // Query 构建
 // ═══════════════════════════════════════════════════════════════════════════
 
-function buildQuerySegments(chat, count, excludeLastAi) {
+function buildQuerySegments(chat, count, excludeLastAi, pendingUserMessage = null) {
     if (!chat?.length) return [];
 
     let messages = chat;
@@ -166,6 +166,18 @@ function buildQuerySegments(chat, count, excludeLastAi) {
         messages = messages.slice(0, -1);
     }
 
+    // ★ 如果有待处理的用户消息且 chat 中最后一条不是它，追加虚拟消息
+    if (pendingUserMessage) {
+        const lastMsg = messages[messages.length - 1];
+        const lastMsgText = lastMsg?.mes?.trim() || "";
+        const pendingText = pendingUserMessage.trim();
+
+        // 避免重复（如果 chat 已包含该消息则不追加）
+        if (lastMsgText !== pendingText) {
+            messages = [...messages, { is_user: true, name: "用户", mes: pendingUserMessage }];
+        }
+    }
+
     return messages.slice(-count).map((m, idx, arr) => {
         const speaker = m.name || (m.is_user ? '用户' : '角色');
         const clean = stripNoise(m.mes);
@@ -673,12 +685,13 @@ export async function recallMemory(queryText, allEvents, vectorConfig, options =
 export async function recallMemory(queryText, allEvents, vectorConfig, options = {}) {
     const T0 = performance.now();
     const { chat } = getContext();
+    const store = getSummaryStore();
     const { pendingUserMessage = null } = options;
 
     if (!allEvents?.length) {
         return { events: [], chunks: [], elapsed: 0, logText: 'No events.' };
     }
-
+
     const segments = buildQuerySegments(chat, CONFIG.QUERY_MSG_COUNT, !!options.excludeLastAi, pendingUserMessage);
 
     let queryVector, weights;