fix(story-summary): stabilize L0 extraction on upstream main

modules/story-summary/vector/llm/atom-extraction.js

@@ -17,8 +17,9 @@ const MODULE_ID = 'atom-extraction';
 const CONCURRENCY = 10;
 const RETRY_COUNT = 2;
 const RETRY_DELAY = 500;
-const DEFAULT_TIMEOUT = 20000;
+const DEFAULT_TIMEOUT = 40000;
 const STAGGER_DELAY = 80;
+const DEBUG_RAW_PREVIEW_LEN = 800;
 
 let batchCancelled = false;
 
@@ -81,14 +82,18 @@ const SYSTEM_PROMPT = `你是场景摘要器。从一轮对话中提取1-2个场
 输出：
 {"anchors":[{"scene":"火山口上艾拉举起圣剑刺穿古龙的心脏，龙血溅满铠甲，古龙轰然倒地，艾拉跪倒在滚烫的岩石上痛哭，完成了她不得不做的弑杀","edges":[{"s":"艾拉","t":"古龙","r":"以圣剑刺穿心脏"}],"where":"火山口"}]}`;
 
-const JSON_PREFILL = '{"anchors":[';
-
 // ============================================================================
 // 睡眠工具
 // ============================================================================
 
 const sleep = (ms) => new Promise(r => setTimeout(r, ms));
 
+function previewText(text, maxLen = DEBUG_RAW_PREVIEW_LEN) {
+    const raw = String(text ?? '').replace(/\s+/g, ' ').trim();
+    if (!raw) return '(empty)';
+    return raw.length > maxLen ? `${raw.slice(0, maxLen)} ...(truncated)` : raw;
+}
+
 const ACTION_STRIP_WORDS = [
     '突然', '非常', '有些', '有点', '轻轻', '悄悄', '缓缓', '立刻',
     '马上', '然后', '并且', '而且', '开始', '继续', '再次', '正在',
@@ -206,7 +211,7 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
     const aiText = filterText(aiMessage.mes);
     parts.push(`<assistant>\n${aiText}\n</assistant>`);
 
-    const input = `<round>\n${parts.join('\n')}\n</round>`;
+    const input = `<round>\n${parts.join('\n')}\n</round>\n请读取上述 <round> 内容，提取 1-2 个场景锚点，并严格按 JSON 输出。\n不要解释，不要续写，不要角色扮演，不要输出 JSON 以外的任何内容。`;
 
     for (let attempt = 0; attempt <= RETRY_COUNT; attempt++) {
         if (batchCancelled) return [];
@@ -215,7 +220,6 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
             const response = await callLLM([
                 { role: 'system', content: SYSTEM_PROMPT },
                 { role: 'user', content: input },
-                { role: 'assistant', content: JSON_PREFILL },
             ], {
                 temperature: 0.3,
                 max_tokens: 600,
@@ -223,6 +227,7 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
             });
 
             const rawText = String(response || '');
+            xbLog.info(MODULE_ID, `floor ${aiFloor} attempt ${attempt} rawText(len=${rawText.length}): ${previewText(rawText)}`);
             if (!rawText.trim()) {
                 if (attempt < RETRY_COUNT) {
                     await sleep(RETRY_DELAY);
@@ -231,11 +236,11 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
                 return null;
             }
 
-            const fullJson = JSON_PREFILL + rawText;
+            xbLog.info(MODULE_ID, `floor ${aiFloor} attempt ${attempt} parseSource(len=${rawText.length}): ${previewText(rawText)}`);
 
             let parsed;
             try {
-                parsed = parseJson(fullJson);
+                parsed = parseJson(rawText);
             } catch (e) {
                 xbLog.warn(MODULE_ID, `floor ${aiFloor} JSON解析失败 (attempt ${attempt})`);
                 if (attempt < RETRY_COUNT) {
@@ -248,6 +253,7 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
             // 兼容：优先 anchors，回退 atoms
             const rawAnchors = parsed?.anchors;
             if (!rawAnchors || !Array.isArray(rawAnchors)) {
+                xbLog.warn(MODULE_ID, `floor ${aiFloor} attempt ${attempt} 缺少有效 anchors，parsed=${previewText(JSON.stringify(parsed))}`);
                 if (attempt < RETRY_COUNT) {
                     await sleep(RETRY_DELAY);
                     continue;
@@ -261,6 +267,12 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
                 .map((a, idx) => anchorToAtom(a, aiFloor, idx))
                 .filter(Boolean);
 
+            xbLog.info(MODULE_ID, `floor ${aiFloor} attempt ${attempt} anchors=${rawAnchors.length} atoms=${atoms.length}`);
+
+            if (rawAnchors.length === 0) {
+                return [];
+            }
+
             return atoms;
 
         } catch (e) {
@@ -373,4 +385,3 @@ export async function batchExtractAtoms(chat, onProgress) {
 
     return allAtoms;
 }
-

modules/story-summary/vector/llm/llm-service.js

@@ -29,12 +29,13 @@ function b64UrlEncode(str) {
 
 /**
  * 统一LLM调用 - 走酒馆后端（非流式）
- * assistant prefill 用 bottomassistant 参数传递
+ * 临时改为标准 messages 调用，避免 bottomassistant prefill 兼容性问题。
  */
 export async function callLLM(messages, options = {}) {
     const {
         temperature = 0.2,
         max_tokens = 500,
+        timeout = 40000,
     } = options;
 
     const mod = getStreamingModule();
@@ -45,14 +46,7 @@ export async function callLLM(messages, options = {}) {
         throw new Error('L0 requires siliconflow API key');
     }
 
-    // 分离 assistant prefill
+    const topMessages = [...messages].filter(msg => msg?.role !== 'assistant');
-    let topMessages = [...messages];
-    let assistantPrefill = '';
-
-    if (topMessages.length > 0 && topMessages[topMessages.length - 1]?.role === 'assistant') {
-        const lastMsg = topMessages.pop();
-        assistantPrefill = lastMsg.content || '';
-    }
 
     const top64 = b64UrlEncode(JSON.stringify(topMessages));
     const uniqueId = generateUniqueId('l0');
@@ -74,13 +68,14 @@ export async function callLLM(messages, options = {}) {
         args.enable_thinking = 'false';
     }
 
-    // ★ 用 bottomassistant 参数传递 prefill
-    if (assistantPrefill) {
-        args.bottomassistant = assistantPrefill;
-    }
-
     try {
-        const result = await mod.xbgenrawCommand(args, '');
+        const timeoutPromise = new Promise((_, reject) => {
+            setTimeout(() => reject(new Error(`L0 request timeout after ${timeout}ms`)), timeout);
+        });
+        const result = await Promise.race([
+            mod.xbgenrawCommand(args, ''),
+            timeoutPromise,
+        ]);
         return String(result ?? '');
     } catch (e) {
         xbLog.error(MODULE_ID, 'LLM调用失败', e);

modules/story-summary/vector/pipeline/state-integration.js

@@ -184,7 +184,7 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress, options
     // ★ 限流检测：连续失败 N 次后暂停并降速
     let consecutiveFailures = 0;
     let rateLimited = false;
-    const RATE_LIMIT_THRESHOLD = 3;       // 连续失败多少次触发限流保护
+    const RATE_LIMIT_THRESHOLD = 6;       // 连续失败多少次触发限流保护
     const RATE_LIMIT_WAIT_MS = 60000;      // 限流后等待时间（60 秒）
     const RETRY_INTERVAL_MS = 1000;        // 降速模式下每次请求间隔（1 秒）
     const RETRY_CONCURRENCY = 1;           // ★ 降速模式下的并发数（默认1，建议不要超过5）
@@ -201,7 +201,7 @@ export async function incrementalExtractAtoms(chatId, chat, onProgress, options
         }
 
         try {
-            const atoms = await extractAtomsForRound(pair.userMsg, pair.aiMsg, floor, { timeout: 20000 });
+            const atoms = await extractAtomsForRound(pair.userMsg, pair.aiMsg, floor, { timeout: 40000 });
 
             if (extractionCancelled) return;