From 3cbc4fa62d4b30d9a782bf779f59e9c16750bdcf Mon Sep 17 00:00:00 2001
From: Xinyue-QwQc <255773902+Xinyue-QwQc@users.noreply.github.com>
Date: Tue, 7 Apr 2026 13:26:06 +0800
Subject: [PATCH] Fix L0 JSON output routing for reasoning models

---
 modules/story-summary/vector/llm/atom-extraction.js | 4 ++++
 modules/story-summary/vector/llm/llm-service.js     | 2 +-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/modules/story-summary/vector/llm/atom-extraction.js b/modules/story-summary/vector/llm/atom-extraction.js
index 412e4adf..30ff52b5 100644
--- a/modules/story-summary/vector/llm/atom-extraction.js
+++ b/modules/story-summary/vector/llm/atom-extraction.js
@@ -83,6 +83,8 @@ 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
 // ============================================================================
@@ -221,6 +223,8 @@ async function extractAtomsForRoundWithRetry(userMessage, aiMessage, aiFloor, op
             const response = await callLLM([
                 { role: 'system', content: SYSTEM_PROMPT },
                 { role: 'user', content: input },
+                // Force the model to continue a JSON object instead of planning in a reasoning channel.
+                { role: 'assistant', content: JSON_PREFILL },
             ], {
                 temperature: 0.3,
                 max_tokens: 600,
diff --git a/modules/story-summary/vector/llm/llm-service.js b/modules/story-summary/vector/llm/llm-service.js
index 44e86f09..f27f6efc 100644
--- a/modules/story-summary/vector/llm/llm-service.js
+++ b/modules/story-summary/vector/llm/llm-service.js
@@ -125,7 +125,7 @@ export async function callLLM(messages, options = {}) {
         max_tokens,
         stream: false,
     };
-    if (model.includes('Qwen3')) {
+    if (model.toLowerCase().includes('qwen3')) {
         body.enable_thinking = false;
     }