优化斗鱼日报LLM输入提纯并清理think输出\n\n- 为斗鱼日报链路统一接入think和推理内容清洗\n- 将提交给LLM的材料改为更聚焦现场弹幕的提纯结构\n- 提高热点窗口原声样本量，避免窗口样本长期不足\n- 刷新日报缓存版本，确保新提示词和新材料立即生效

plugins/douyu/danmu_summary.py

@@ -759,6 +759,22 @@ class DouyuDanmuSummaryHelper:
             seen.add(content)
             if len(selected) >= limit:
                 break
+        # 固定位置采样只能快速抓到“窗口骨架”，但在弹幕量大时通常不足以凑满 limit。
+        # 这里继续顺序补样本，把同一热点窗口里更多真实原声带给 LLM，
+        # 减少模型只看到 4-5 条孤立短句、难以还原现场氛围的问题。
+        if len(selected) < limit:
+            for item in items:
+                content = str(item.get("content") or "").strip()
+                if not content or content in seen:
+                    continue
+                selected.append({
+                    "time": str(item.get("timestamp_text") or ""),
+                    "nickname": str(item.get("nickname") or ""),
+                    "content": content[:80],
+                })
+                seen.add(content)
+                if len(selected) >= limit:
+                    break
         return selected
 
     @classmethod

plugins/douyu/main.py

@@ -30,6 +30,7 @@ from utils.decorator.plugin_decorators import plugin_stats_decorator
 from utils.decorator.points_decorator import plugin_points_cost
 from utils.markdown_to_image import convert_md_str_to_image, html_to_image
 from utils.robot_cmd.robot_command import Feature, PermissionStatus, GroupBotManager
+from utils.string_utils import remove_reasoning_content
 from wechat_ipad import WechatAPIClient
 from wechat_ipad.models.appmsg_xml import DOUYU_MESSAGE_XML
 
@@ -513,9 +514,11 @@ class DouyuRedisManager:
 class DouyuPlugin(MessagePluginInterface):
     # 报告缓存版本号：
     # 1. 版本升级后会自动让历史缓存失效，避免继续复用旧文本/旧图片；
-    # 2. 本次将版本提升到 8，背景画像链路删掉了部分过程控制和提示噪音，
-    #    需要刷新旧日报缓存，确保新版 prompt 使用新的精简上下文。
-    _DAILY_REPORT_CACHE_VERSION = 8
+    # 2. 本次将版本提升到 9：
+    #    - LLM 输入材料从“整份大 payload”改成“提纯后的现场材料”；
+    #    - 同时统一清洗 <think> / reasoning 输出；
+    #    因此需要刷新旧缓存，确保新版结果真实命中新链路。
+    _DAILY_REPORT_CACHE_VERSION = 9
     FEATURE_KEY = "DOUYU_MONITOR"
     FEATURE_DESCRIPTION = "🎮 斗鱼开播提醒 [订阅斗鱼 房间号, 取消订阅斗鱼 房间号]"
 
@@ -764,7 +767,7 @@ class DouyuPlugin(MessagePluginInterface):
         1. 模型把 JSON 包在 ```json 代码块里；
         2. 模型前后补了少量解释文字。
         """
-        raw = str(text or "").strip()
+        raw = remove_reasoning_content(str(text or "").strip())
         if not raw:
             return None
         if raw.startswith("```"):
@@ -2172,6 +2175,7 @@ class DouyuPlugin(MessagePluginInterface):
     def _build_daily_report_prompt(self, payload: Dict[str, Any]) -> Tuple[str, str]:
         meta = payload.get("report_meta", {}) or {}
         room_context_prompt = self._build_room_context_prompt_block(payload)
+        prompt_material = self._build_llm_prompt_material(payload, include_operator=True)
         system_prompt = (
             "你是斗鱼直播日报助手。请基于给定的结构化弹幕材料，输出一份适合发群的中文日报。"
             "要求简洁、自然、信息密度高，不要编造，不要使用代码块。"
@@ -2186,13 +2190,15 @@ class DouyuPlugin(MessagePluginInterface):
             "5. 单独列出 2-3 个热点时段。\n"
             "6. 整体控制在 600 字以内。\n\n"
             f"{room_context_prompt}"
-            f"材料如下：\n{json.dumps(payload, ensure_ascii=False, indent=2)}"
+            "下面是已经提纯给 LLM 的材料，优先依据现场弹幕片段、热点窗口和共识梗来写，不要被大段统计信息带偏。\n"
+            f"材料如下：\n{json.dumps(prompt_material, ensure_ascii=False, indent=2)}"
         )
         return system_prompt, user_prompt
 
     def _build_danmu_summary_prompt(self, payload: Dict[str, Any]) -> Tuple[str, str]:
         meta = payload.get("report_meta", {}) or {}
         room_context_prompt = self._build_room_context_prompt_block(payload)
+        prompt_material = self._build_llm_prompt_material(payload, include_operator=False)
         system_prompt = (
             "你是直播弹幕总结助手。请只根据给定材料，总结这场直播的弹幕内容与氛围。"
             "不要输出运营数据，不要编造，不要写空话套话。"
@@ -2209,7 +2215,8 @@ class DouyuPlugin(MessagePluginInterface):
             f"主播：{meta.get('nickname') or meta.get('room_name') or meta.get('room_id')}\n"
             f"日期：{meta.get('anchor_day', '')}\n"
             f"{room_context_prompt}"
-            f"材料：\n{json.dumps(payload, ensure_ascii=False, indent=2)}"
+            "下面是已经提纯给 LLM 的现场材料，请优先使用原声弹幕、热点窗口和复读梗，不要写成词频复述。\n"
+            f"材料：\n{json.dumps(prompt_material, ensure_ascii=False, indent=2)}"
         )
         return system_prompt, user_prompt
 
@@ -2222,6 +2229,7 @@ class DouyuPlugin(MessagePluginInterface):
         """
         meta = payload.get("report_meta", {}) or {}
         room_context_prompt = self._build_room_context_prompt_block(payload)
+        prompt_material = self._build_llm_prompt_material(payload, include_operator=False)
         system_prompt = (
             "你是斗鱼直播间的粉丝向整活日报编辑。"
             "请只根据提供的真实弹幕材料，输出一份开心、欢乐、带一点恶搞气质的中文总结。"
@@ -2239,10 +2247,173 @@ class DouyuPlugin(MessagePluginInterface):
             f"主播：{meta.get('nickname') or meta.get('room_name') or meta.get('room_id')}\n"
             f"日期：{meta.get('anchor_day', '')}\n"
             f"{room_context_prompt}"
-            f"材料：\n{json.dumps(payload, ensure_ascii=False, indent=2)}"
+            "下面是已经提纯给 LLM 的现场材料，请优先抓原声弹幕、热点窗口和集体起哄片段，少写空泛概括。\n"
+            f"材料：\n{json.dumps(prompt_material, ensure_ascii=False, indent=2)}"
         )
         return system_prompt, user_prompt
 
+    def _build_llm_prompt_material(
+        self,
+        payload: Dict[str, Any],
+        *,
+        include_operator: bool = False,
+    ) -> Dict[str, Any]:
+        """
+        为日报 LLM 单独构建“提纯后的材料”。
+        设计目标：
+        1. 把真正有现场感的弹幕内容提到前面，降低大体量统计 JSON 对模型注意力的干扰；
+        2. 保留足够的梗、热点时段和原声样本，方便模型写出更像直播间回放的内容；
+        3. 运营指标只在确实需要的日报正文场景里保留精简版，不再无差别塞给所有任务。
+        """
+        meta = payload.get("report_meta", {}) or {}
+        room_context = payload.get("room_context", {}) or {}
+        sessions = payload.get("sessions", []) or []
+        representative_messages = payload.get("representative_messages", []) or []
+        raw_window_samples = payload.get("raw_window_samples", []) or []
+        merged_templates = payload.get("merged_templates", []) or []
+        repeated_messages = payload.get("repeated_messages", []) or []
+        top_terms = payload.get("top_terms", []) or []
+        burst_terms = payload.get("burst_terms", []) or []
+        peak_buckets = payload.get("peak_buckets", []) or []
+
+        material: Dict[str, Any] = {
+            "report_meta": {
+                "room_id": str(meta.get("room_id") or "").strip(),
+                "anchor_day": str(meta.get("anchor_day") or "").strip(),
+                "nickname": str(meta.get("nickname") or "").strip(),
+                "room_name": str(meta.get("room_name") or "").strip(),
+                "session_count": int(meta.get("session_count", 0) or 0),
+                "message_count": int(meta.get("message_count", 0) or 0),
+                "unique_user_count": int(meta.get("unique_user_count", 0) or 0),
+            },
+            "room_context": {
+                "domain": str(room_context.get("domain") or "").strip(),
+                "inferred_domains": self._normalize_text_list(room_context.get("inferred_domains"))[:6],
+                "identity_summary": str(room_context.get("identity_summary") or "").strip(),
+                "career_background": str(room_context.get("career_background") or "").strip(),
+                "related_people": self._normalize_text_list(room_context.get("related_people"))[:10],
+                "storyline_keywords": self._normalize_text_list(room_context.get("storyline_keywords"))[:10],
+                "style_hints": self._normalize_text_list(room_context.get("style_hints"))[:6],
+            },
+            "session_overview": [
+                {
+                    "session_id": str(item.get("session_id") or "").strip(),
+                    "segments": item.get("segments", []) or [],
+                    "message_count": int(item.get("message_count", 0) or 0),
+                    "organized_message_count": int(item.get("organized_message_count", 0) or 0),
+                }
+                for item in sessions[:4]
+            ],
+            "high_frequency_topics": {
+                "top_terms": [
+                    {"term": str(item.get("term") or "").strip(), "count": int(item.get("count", 0) or 0)}
+                    for item in top_terms[:16]
+                    if str(item.get("term") or "").strip()
+                ],
+                "burst_terms": [
+                    {"text": str(item.get("text") or "").strip(), "count": int(item.get("count", 0) or 0)}
+                    for item in burst_terms[:12]
+                    if str(item.get("text") or "").strip()
+                ],
+                "merged_templates": [
+                    {
+                        "text": str(item.get("text") or "").strip()[:80],
+                        "count": int(item.get("count", 0) or 0),
+                        "user_count": int(item.get("user_count", 0) or 0),
+                    }
+                    for item in merged_templates[:12]
+                    if str(item.get("text") or "").strip()
+                ],
+                "repeated_messages": [
+                    {
+                        "text": str(item.get("text") or "").strip()[:80],
+                        "count": int(item.get("count", 0) or 0),
+                        "user_count": int(item.get("user_count", 0) or 0),
+                    }
+                    for item in repeated_messages[:12]
+                    if str(item.get("text") or "").strip()
+                ],
+            },
+            # 这里把“现场材料”单独抽成一大块，方便模型先看原声，再看统计摘要。
+            "live_scene_material": {
+                "representative_messages": [
+                    {
+                        "time": str(item.get("time") or "").strip(),
+                        "nickname": str(item.get("nickname") or "").strip(),
+                        "content": str(item.get("content") or "").strip()[:90],
+                    }
+                    for item in representative_messages[:24]
+                    if str(item.get("content") or "").strip()
+                ],
+                "hot_window_samples": [
+                    {
+                        "start_time": str(window.get("start_time") or "").strip(),
+                        "message_count": int(window.get("message_count", 0) or 0),
+                        "user_count": int(window.get("user_count", 0) or 0),
+                        "top_terms": [
+                            str(term.get("term") or "").strip()
+                            for term in (peak_buckets[index].get("top_terms", []) or [])[:5]
+                            if str(term.get("term") or "").strip()
+                        ] if index < len(peak_buckets) else [],
+                        "samples": [
+                            {
+                                "time": str(sample.get("time") or "").strip(),
+                                "nickname": str(sample.get("nickname") or "").strip(),
+                                "content": str(sample.get("content") or "").strip()[:90],
+                            }
+                            for sample in (window.get("samples", []) or [])[:8]
+                            if str(sample.get("content") or "").strip()
+                        ],
+                    }
+                    for index, window in enumerate(raw_window_samples[:8])
+                ],
+            },
+        }
+
+        if include_operator:
+            operator = payload.get("operator_metrics", {}) or {}
+            material["operator_focus"] = {
+                "fans_badge_user_count": int(operator.get("fans_badge_user_count", 0) or 0),
+                "high_room_level_user_count": int(operator.get("high_room_level_user_count", 0) or 0),
+                "high_fans_level_user_count": int(operator.get("high_fans_level_user_count", 0) or 0),
+                "active_users_5plus": int(operator.get("active_users_5plus", 0) or 0),
+                "active_users_10plus": int(operator.get("active_users_10plus", 0) or 0),
+                "top_badges": [
+                    {
+                        "badge_name": str(item.get("badge_name") or "").strip(),
+                        "user_count": int(item.get("user_count", 0) or 0),
+                        "message_count": int(item.get("message_count", 0) or 0),
+                    }
+                    for item in (operator.get("top_badges", []) or [])[:6]
+                    if str(item.get("badge_name") or "").strip()
+                ],
+                "top_active_users": [
+                    {
+                        "nickname": str(item.get("nickname") or item.get("uid") or "").strip(),
+                        "message_count": int(item.get("message_count", 0) or 0),
+                        "fans_name": str(item.get("fans_name") or "").strip(),
+                        "fans_level": int(item.get("fans_level", 0) or 0),
+                        "room_level": int(item.get("room_level", 0) or 0),
+                    }
+                    for item in (operator.get("top_active_users", []) or [])[:8]
+                ],
+            }
+
+        return material
+
+    @staticmethod
+    def _clean_daily_report_llm_text(text: str) -> str:
+        """
+        统一清理日报类 LLM 输出中的思考内容。
+        重点处理：
+        1. <think> / <thinking> / <reasoning> 标签；
+        2. 某些模型额外吐出来的“思考过程”“分析过程”等段落；
+        3. 清理后顺手压缩多余空行，避免图片模板里出现大片空白。
+        """
+        cleaned = remove_reasoning_content(str(text or "").strip())
+        cleaned = re.sub(r"\n{3,}", "\n\n", cleaned).strip()
+        return cleaned
+
     def _build_funny_scene_lines(self, payload: Dict[str, Any], limit: int = 5) -> List[str]:
         """
         组装“弹幕名场面”兜底素材。
@@ -2980,7 +3151,9 @@ class DouyuPlugin(MessagePluginInterface):
                 tag=f"douyu_danmu_summary_{(payload.get('report_meta', {}) or {}).get('room_id', '')}",
             )
             if result:
-                return result.strip()
+                cleaned = self._clean_daily_report_llm_text(result)
+                if cleaned:
+                    return cleaned
             logger.warning(
                 f"斗鱼弹幕总结 LLM 生成失败: model={self._daily_report_llm_client.model}, "
                 f"last_error={self._daily_report_llm_client.last_error}"
@@ -3003,10 +3176,11 @@ class DouyuPlugin(MessagePluginInterface):
                 tag=f"douyu_fans_daily_report_{(payload.get('report_meta', {}) or {}).get('room_id', '')}",
             )
             if result:
-                text = result.strip()
+                text = self._clean_daily_report_llm_text(result)
                 if len(text) > self._daily_report_max_length:
                     return text[: self._daily_report_max_length - 20].rstrip() + "\n...(已截断)"
-                return text
+                if text:
+                    return text
             logger.warning(
                 f"斗鱼粉丝日报 LLM 生成失败: model={self._daily_report_llm_client.model}, "
                 f"last_error={self._daily_report_llm_client.last_error}"
@@ -3135,10 +3309,11 @@ class DouyuPlugin(MessagePluginInterface):
                 tag=f"douyu_daily_report_{(payload.get('report_meta', {}) or {}).get('room_id', '')}",
             )
             if result:
-                text = result.strip()
+                text = self._clean_daily_report_llm_text(result)
                 if len(text) > self._daily_report_max_length:
                     return text[: self._daily_report_max_length - 20].rstrip() + "\n...(已截断)"
-                return text
+                if text:
+                    return text
             logger.warning(
                 f"斗鱼每日报告 LLM 生成失败: model={self._daily_report_llm_client.model}, "
                 f"last_error={self._daily_report_llm_client.last_error}"