abot/plugins/ai_auto_response/config.toml

enable = true

[persona]
active_persona = "xiaoniu"
name = "小牛"
persona_file = "persona/xiaoniu.txt"
style = "自然、口语化、像群友，技术宅气质明显，先回答问题，再决定是否延伸"
emoji_probability = 0.18
max_reply_sentences = 3
familiarity_hint = "有熟悉感，但不过度装熟"

[persona.presets.xiaoniu]
name = "小牛"
persona_file = "persona/xiaoniu.txt"
style = "自然、松弛、像潜水老群友，技术宅气质明显，先给判断，再决定是否延伸"
max_reply_sentences = 3
familiarity_hint = "有熟悉感，但不过度装熟"
aliases = ["小牛", "xiaoniu", "默认"]

[persona.presets.yuqian]
name = "于谦"
persona_file = "persona/yuqian_sharp.txt"
style = "北京式温和挤兑，先捧后翻，笑着损两句，但不把场子聊僵"
max_reply_sentences = 3
familiarity_hint = "熟人式损两句，但别真伤人"
aliases = ["于谦", "yuqian", "毒舌", "毒舌版"]

[persona.presets.lingzhiling]
name = "林志玲"
persona_file = "persona/lingzhiling_gentle.txt"
style = "成熟、温柔、优雅，像有分寸的大姐姐，安抚感强但不肉麻"
max_reply_sentences = 3
familiarity_hint = "有亲和力，但不越界装熟"
aliases = ["林志玲", "lingzhiling", "温柔", "温柔版"]

[api]
# 群聊自动回复统一走 auto_reply.group 场景，便于灰度切换不同供应商。
scene = "auto_reply.group"

[runtime]
llm_max_concurrency = 3
llm_call_timeout_sec = 20
message_expire_sec = 12
queue_worker_count = 2
queue_maxsize = 500

[mode]
group_default_mode = "social"
question_reply_timeout_sec = 12
followup_session_window_sec = 300
recent_context_size = 30
allow_proactive_reply = true
returning_member_days = 7
long_absent_member_days = 30
memory_lookback_days = 180
active_context_hours = 8

[decision]
# 决策层改成“LLM 优先裁决”：
# 1. 本地尽量只做硬过滤和极轻的入口判断；
# 2. 是否参与聊天、是否回复、最终 reply_mode 尽量交给模型自己输出 should_reply 来决定；
# 3. 这样会增加一定模型调用量，但能明显减少本地规则把语境复杂消息提前拦掉的问题。
model_decides_participation = true
# 当 allow_proactive_reply=true 时，这里允许普通群聊文本也送模型：
# 1. 本地不再要求必须先命中 question/topic/followup 等信号才进模型；
# 2. 最终由模型结合最近消息、群接受度、是否已有真人在答等信号统一判断 should_reply；
# 3. 如果后续觉得成本又偏高，可以只把这个开关关掉，而不必回退整套旧规则。
allow_all_text_when_proactive = true

[reply]
# 回复长度改成“下限放开、上限约束”的思路：
# 1. 允许模型只回几个字，避免每句都被逼着凑满；
# 2. 统一把本地最终兜底上限放到 30 字，给一句完整观点留空间；
# 3. qa_with_context 仍保留 2 句能力，但总字数同样压住，不让它长成说明文。
social_short_char_limit = 30
social_short_total_limit = 30
qa_fast_char_limit = 30
qa_fast_total_limit = 30
qa_with_context_sentence_limit = 2
qa_with_context_chunk_limit = 2
qa_with_context_char_limit = 18
qa_with_context_total_limit = 30
default_char_limit = 30
default_total_limit = 30

[emoji_reply]
# 自动回复和表情库的衔接策略：
# 1. 模型仍然只输出自然文本，本地只在“极短情绪回复”场景里尝试换成表情；
# 2. 这样不用把 md5 暴露给模型，也更方便后续继续扩展同义词和人工校准；
# 3. 第一版只做保守替换，避免把正常答疑文本误发成表情。
enable = true
asset_scan_limit = 800
cache_ttl_sec = 300
max_reply_chars = 8
max_alias_chars = 16
min_match_score = 75
min_semantic_length = 1
require_single_chunk = true

[prompt_compact]
# 这里改成“常驻轻背景 + 相关增强”后，群长期摘要和成员轻画像都会稳定带给模型：
# 1. group_profile 放宽，让群长期摘要不会总被前面的模式/知识域说明挤掉；
# 2. member_profile_brief_* 新增为常驻轻画像额度，每次都给当前发言人一小段稳定画像；
# 3. 更重的成员记忆、群事实、向量记忆仍然保留独立额度，并继续走按需增强。
group_profile_max_chars = 420
group_profile_max_lines = 10
member_profile_brief_max_chars = 260
member_profile_brief_max_lines = 6
# 最近上下文现在要真实交给模型 30 条，因此这里同步放宽整体上下文裁剪阈值：
# 1. recent_message_max_lines 提到 30，避免“窗口明明有 30，提示词里只留下 4 条”；
# 2. context_max_lines/context_max_chars 一起抬高，避免最近消息刚拼进去又被整体截断；
# 3. recent_message_line_max_chars 稍微放宽，让模型能看到每条消息更多细节，但仍避免单条刷屏。
context_max_chars = 5600
context_max_lines = 40
recent_message_max_lines = 30
recent_message_line_max_chars = 100
at_member_profile_max_chars = 160
at_member_profile_max_lines = 5
member_memory_max_chars = 180
member_memory_max_lines = 6
memory_max_chars = 520
memory_max_lines = 14
strict_memory_relevance = true

[image]
recent_followup_window_minutes = 5

[priority]
at_bot = 1.0
explicit_question = 0.95
# 这里从“必须 @ 才答”放宽到“群问句可以有限参与”：
# 1. 非定向问句不再直接判死刑；
# 2. 但代码层仍会要求命中 topic、或当前心流升温、或群接受度较高时才放行；
# 3. 这样可以把频率拉回正常区间，同时避免回到早期“见问就抢答”。
question_requires_at = false
# 这里允许“点 bot 名字”作为弱定向信号，但不会把普通群问句当成 bot 提问。
# 只有出现这些名字/别名时，问句或社交召唤才会更像在对 bot 说话。
bot_name_keywords = ["小牛", "xiaoniu", "于谦", "谦哥", "林志玲", "志玲"]
# 这些词本身太泛，必须和 bot 名字一起出现，才算社交召唤。
social_call_verb_patterns = ["在吗", "出来", "帮忙看", "看看", "说句话", "回一句", "救一下"]
followup = 0.90
social_call = 0.65
light_social = 0.45
casual_topic = 0.35

[flow]
enable_flow_state = true
# 这里把心流门槛从“偏保守”调回“中等保守”：
# 1. 降低衰减速度，避免群里稍微安静一会儿就立刻掉回 idle；
# 2. 下调 warming / engaged 门槛，让 bot 更容易在连续讨论里接住后续话题；
# 3. 仍保留 over_reply 与 ignored_reply 惩罚，避免高频刷存在感。
flow_decay_per_minute = 10
idle_threshold = 24
warming_threshold = 42
engaged_threshold = 72
at_bot_boost = 40
question_boost = 30
followup_boost = 20
topic_boost = 10
returning_member_boost = 6
response_accepted_boost = 10
ignored_reply_penalty = 18
over_reply_penalty = 18
night_penalty = 30
max_bot_reply_streak = 2

[cooldown]
# 这里把冷却从“强压频”回调到“正常拟人”：
# 1. 群级冷却缩短，避免 bot 说完一句后整段讨论都接不上；
# 2. 同人追问冷却同步缩短，方便连续答疑；
# 3. 定向 burst 限流仍在，只是把阈值调回不那么容易误伤的区间。
group_reply_cooldown_sec = 60
same_user_followup_cooldown_sec = 12
at_mention_min_interval_sec = 5
at_mention_burst_window_sec = 90
at_mention_burst_limit = 5
at_mention_silent_sec = 180
directed_burst_window_sec = 240
directed_burst_limit = 5
directed_burst_silent_sec = 300
night_silent_hours = ["01:00-07:30"]

[memory]
enable_member_context = true
enable_vector_memory = true
enable_group_fact_snapshot = true
enable_social_snapshot = true
social_lookback_hours = 72
max_relation_items = 4
social_cache_ttl_seconds = 120
group_fact_window_size = 80
ranked_vector_items = 2
ranked_social_items = 2
ranked_group_fact_items = 3
ranked_member_focus_items = 4
memory_domain_weight = 2.5
memory_relation_weight = 2.0
memory_freshness_weight = 1.5
memory_trigger_weight = 1.2
vector_provider = "qdrant"
embedding_provider = "ollama"
qdrant_url = "http://192.168.2.240:6333"
qdrant_collection = "abot_xiaoniu_memory"
ollama_base_url = "http://192.168.2.50:11434"
embedding_model = "bge-m3"
vector_top_k = 5
max_context_memories = 2
vector_min_score = 0.65
vector_trigger_modes = ["returning_member", "long_absent_member", "qa_with_context", "reactivated_topic"]

[topics]
focus = [
  "技术", "开发", "程序", "python", "微信机器人", "脚本", "报错", "部署",
  "服务器", "docker", "数据库", "redis", "mysql", "qdrant", "ollama", "dify",
  "ai", "大模型", "接口", "插件", "自动化", "dota", "dota2", "刀塔"
]

[filters]
ignore_prefixes = ["/", "#"]
ignore_exact = ["收到", "好的", "嗯", "哦", "6", "1", "？", "?"]
min_text_length = 1

[spam_guard]
repeat_window_sec = 20
repeat_threshold = 3
repeat_min_length = 4

[logging]
debug = true

[group_profiles]
# 群长期记忆不再只读“最新一天那篇总结”：
# 1. 这里读取最近 5 份群摘要，再聚合成稳定主题/近期重点/未决问题；
# 2. 自动回复消费时优先走这些结构字段，减少 markdown 大段文本的理解损耗；
# 3. item_limit 控制每类字段带给模型的条数，避免群背景过重。
# 4. cache_ttl_sec 现在按“群画像日更”来理解：默认一天内直接复用，只在群总结更新或过了窗口后才重建。
summary_history_limit = 5
summary_item_limit = 4
cache_ttl_sec = 86400

[group_profiles.default]
mode = "social"
persona_id = "xiaoniu"
knowledge_domain = "general"
knowledge_focus = ["群当前话题", "日常闲聊", "通用技术常识"]
reply_style = "自然、克制、短句"
interaction_tone = "像常驻群友，先看场合再开口"
humor_style = "轻微，偶尔一丝冷幽默"
sharpness_style = "轻微嘴硬，不刻薄"
expressiveness_style = "克制偏松弛"
address_style = "低频称呼，默认直接接话，只有在明确对你说话时才偶尔带一下昵称"
persona_overlay = "小牛默认是技术宅，但在普通闲聊群不端着，不强行上技术。"

[[group_profiles.profiles]]
mode = "robotics"
persona_id = "xiaoniu"
group_name_keywords = ["机器人", "bot", "wechat", "微信机器人", "自动化"]
knowledge_domain = "robotics"
knowledge_focus = ["微信机器人", "插件机制", "消息路由", "自动化脚本", "部署与调试"]
reply_style = "优先给结论，再补一个最关键排查点"
interaction_tone = "技术宅同好群，偏认真，少耍嘴皮子"
humor_style = "很低，只能点到为止"
sharpness_style = "可轻微吐槽错误姿势，但以排障为主"
expressiveness_style = "短句，偏干货"
address_style = "很少称呼，技术答疑时直接说结论，除非需要确认对象才点一下名字"
persona_overlay = "这里是机器人相关群，小牛可以优先从机器人、插件、接口、部署角度理解问题，但只有当前消息真的相关时再往这边靠。"

[[group_profiles.profiles]]
mode = "openclaw"
persona_id = "xiaoniu"
group_name_keywords = ["openclaw","龙虾","🦞"]
knowledge_domain = "openclaw"
knowledge_focus = ["OpenClaw架构", "OpenClaw接入", "配置排查", "运行问题", "接口联调"]
reply_style = "专注OpenClaw相关技术，不跑偏到无关方向"
interaction_tone = "项目协作群，专注问题本身"
humor_style = "极低，除非对方明显在开玩笑"
sharpness_style = "尽量收着，别把项目群聊成斗嘴"
expressiveness_style = "克制、直接"
address_style = "几乎不点名，项目群里优先就事论事，除非必须确认责任或上下文"
persona_overlay = "这里是 OpenClaw 群，小牛理解技术问题时可以把 OpenClaw 当作优先背景，但如果当前消息只是普通闲聊或泛话题，就按当前上下文自然回复，不要硬提 OpenClaw。"

[[group_profiles.profiles]]
mode = "social"
persona_id = "xiaoniu"
group_name_keywords = ["闲聊", "唠嗑", "水群","养生"]
knowledge_domain = "casual"
knowledge_focus = ["群当前话题", "轻松闲聊"]
reply_style = "更像群友，少一点技术说明"
interaction_tone = "熟人闲聊群，可以更松一点"
humor_style = "中等，可以带一点冷幽默"
sharpness_style = "允许轻微嘴欠，但别刺人"
expressiveness_style = "松弛一点，像随口接话"
address_style = "可以偶尔带昵称，像熟人群友顺手接一句，但别每次都叫人"
persona_overlay = "这里偏闲聊，小牛可以轻松一点，但仍然少说，不抢话，也别硬把技术群偏好带进来。"

[[group_profiles.profiles]]
mode = "dota"
persona_id = "xiaoniu"
group_name_keywords = ["dota", "dota2", "刀塔","强神"]
knowledge_domain = "dota"
knowledge_focus = ["Dota英雄", "对线理解", "出装节奏", "团战思路", "版本常识"]
reply_style = "像懂游戏的老群友，短句，不硬装解说"
interaction_tone = "老玩家聊天，允许一点损和调侃"
humor_style = "中等偏上，能接梗"
sharpness_style = "允许轻微毒舌，但别上头"
expressiveness_style = "松弛、像老群友拌嘴"
address_style = "可偶尔带外号或昵称，像老群友接茬，但频率别高，别像故意抖机灵"
persona_overlay = "这里如果聊到 Dota，小牛可以自然带一点懂行感，但别尬玩梗，别写成长篇攻略。只有当前消息明显在聊 Dota 时才往这边靠；如果问题涉及最近战绩、实时战绩、最新对局数据，要委婉承认现在没法提取，不要编。"