TouhouLittleMaid-Self-Talk

新增功能

• 新增结构化人设模式：在人设编辑页可在「整段人设模式」与「结构化人设模式」之间切换

  • 结构化模式将人设拆分为身份、性格、背景、爱好四栏并行编辑
  • 身份与性格每轮固定注入，背景与爱好按当前情境选择性注入片段
  • 两种模式的内容独立保存，切换模式不会造成人设丢失

• 新增「AI 拆分」：作为结构化人设模式的附属功能，可在人设编辑页使用一次额外的模型请求，将已有的整段人设自动整理为身份、性格、背景、爱好四栏

  • 整理期间可以离开页面，完成后会私信通知，重新打开该女仆的人设页即可选择载入

• 新增预制自言自语与欢迎语功能（实验性，默认关闭）：开启后，女仆真实生成的自话与欢迎语会按名称、人设、语言与环境信息保存，后续触发满足环境条件时按可配置的概率直接复用复用旧回复，减少真实的模型 API 请求

  • 仅支持自话与欢迎语，不包含互聊
  • 可在 AI 聊天界面的 💬 设置中按玩家或单只女仆开启
  • 最低可用条数与使用预制内容的概率可在「预制复用」面板中调整，服务器管理员可设置默认值并决定是否允许玩家个性化调整

• 预制功能支持自定义条目：允许添加自定义的预制内容供女仆使用，详见预制功能使用文档

• 新增预制语音缓存（实验性，默认关闭）：作为预制内容的配套功能。开启后，自话与欢迎语语音也会作为预制内容被保存

  • 在使用预制内容时可以直接播放本地保存的音频，不再调用 TTS 供应商
  • 支持在 custom/ 中放置自定义音频，详见预制功能使用文档

• 环境上下文新增五个条目：

  • 新增条目：
    • 女仆饰品
    • 女仆 Curios 饰品
    • 主人 Curios 饰品
    • 女仆通过神龛复活
    • 女仆自身击杀生物
  • 前三个默认进入随机池，未安装 Curios 或女仆 Curios 时默认不启用
  • 后两个默认必定进入上下文，只记录女仆自身

• 新增「思考气泡」本地开关：关闭后可隐藏自己看到的所有女仆在自话与互聊时的等待、工具提示与思考气泡，不影响最终回答，属于仅客户端设置，不影响其他玩家

优化与修复

• 修复检索模式下中英混排关键词无法命中的问题
• 同时修复长时间运行下检索缓存与独立聊天档案记录持续累积的问题
• 修复自话超时后迟到的回复仍可能上屏、写入档案或广播的问题
• 修复互聊的两处异常收尾问题：
  • 语音合成在发起阶段报错时，归档、广播与后续续接不再被整段跳过
  • 配对冲突时会正确释放上一条对话，不再残留占用
• 修复语言标签与主人附近判定：
  • ZH_CN 等大小写写法不再导致同一请求中英模板混用，zh_tw / zh_hk 会保留地区语义
  • 主人在另一维度且坐标相近时，不会再被误判为「主人在附近」
• 修复 Forge 版本玩家设置在自话与互聊关闭记录损坏或清空后，旧记录可能被重新迁移回来的问题
• 提高自定义 Prompt 上限：由 500 字符提高到 2048 字符

此版本服务端不兼容之前版本的客户端，若在服务器中更新，那么在服务器中游玩的玩家都需要更新

New Features

• Added structured persona mode: switch between "Full-text Persona" and "Structured Persona" on the persona editing page.

  • Structured mode provides four fields that can be edited side by side: identity, personality, background, and hobbies.
  • Identity and personality are included in every request, while relevant excerpts from background and hobbies are selectively included based on the current context.
  • Each mode saves its content separately, so switching modes does not erase either persona.

• Added "AI Split" as a companion feature to structured persona mode: on the persona editing page, use one additional model request to automatically organize an existing full-text persona into identity, personality, background, and hobbies.

  • You can leave the page while processing is in progress. A private message will notify you when it is complete; reopen that maid's persona page to choose whether to load the result.

• Added prebuilt self-talk and welcome messages (experimental, disabled by default): when enabled, self-talk and welcome messages generated by the model are saved along with the maid's name, persona, language, and environmental information. When a later trigger meets the environmental conditions, a saved reply can be reused directly with a configurable probability, reducing model API requests.

  • Supports self-talk and welcome messages only; inter-maid chat is not included.
  • Can be enabled per player or per maid through the 💬 settings in the AI chat interface.
  • The minimum number of eligible entries and the probability of using prebuilt content can be adjusted in the "Prebuilt Reuse" panel. Server administrators can set defaults and decide whether players may customize these values.

• Prebuilt content supports custom entries: add your own prebuilt content for maids to use. See the Prebuilt Content Guide for details.

• Added prebuilt voice caching (experimental, disabled by default) as a companion feature to prebuilt content. When enabled, audio for self-talk and welcome messages is also saved for reuse.

  • When prebuilt content is used, locally saved audio can be played directly without calling the TTS provider.
  • Custom audio can be placed in custom/. See the Prebuilt Content Guide for details.

• Added five environmental context entries:

  • New entries:
    • Maid accessories
    • Maid Curios accessories
    • Owner Curios accessories
    • Maid revived through a shrine
    • Mobs killed by the maid herself
  • The first three are included in the random selection pool by default; the relevant Curios entries are disabled by default if Curios or the maid Curios integration is not installed.
  • The last two are always included in context by default and only record events involving the maid herself.

• Added a local "Thinking Bubbles" toggle: turning it off hides the waiting, tool hint, and thinking bubbles you see for all maids during self-talk and inter-maid chat. Final replies are unaffected. This is a client-only setting and does not affect other players.

Improvements and Fixes

• Fixed mixed Chinese and English keywords failing to match in retrieval mode.
• Also fixed retrieval caches and independent chat archive records continuing to accumulate during long-running sessions.
• Fixed late replies after a self-talk timeout potentially still appearing on screen, being saved to the archive, or being broadcast.
• Fixed two issues with inter-maid chat cleanup:
  • If speech synthesis fails during startup, archiving, broadcasting, and subsequent conversation turns are no longer skipped entirely.
  • Pairing conflicts now correctly release the previous conversation without leaving it locked.
• Fixed language tags and nearby-owner detection:
  • Case variants such as ZH_CN no longer cause Chinese and English templates to be mixed within the same request, and zh_tw / zh_hk retain their regional distinctions.
  • An owner in another dimension with similar coordinates is no longer incorrectly considered "nearby."
• Fixed an issue in the Forge version where old player settings could be migrated back after the stored records for disabled self-talk and inter-maid chat settings were corrupted or cleared.
• Increased the custom prompt limit from 500 to 2,048 characters.

Servers running this version are incompatible with clients running earlier versions. If you update a server, all players on that server must also update.