chore: 发布 v2.1.9 / Release v2.1.9

fix: 设置面板两项统一靠左对齐 / Align the two settings entries to the left edge
- 「语言与外观」与「助手设置」两块的内容改为顶到最左边，左边缘齐平，看起来是一整块 - 标题仍在：由各自的 h3 显示，收起的只是 ComfyUI 重复渲染出来的那个名称列
2026-10-29 00:54:45 +08:00 · 2026-10-29 00:51:33 +08:00 · 2026-10-29 00:02:32 +08:00 · 2026-10-26 05:42:02 +08:00 · 2026-10-26 05:37:36 +08:00 · 2026-10-26 04:27:40 +08:00
33 changed files with 4334 additions and 762 deletions
@@ -0,0 +1,27 @@
+name: Publish to Comfy Registry
+
+# 只在并入 main 时发布：分支上的改动（哪怕改了 pyproject.toml）不会触发，
+# 因此"功能做完再合进 main"的开发节奏是安全的——合进来时才对外发布一次。
+# 触发条件是 pyproject.toml 变动，实际等价于"版本号升了才发"；
+# 版本号没升时 Registry 会报 Version already exists，本次 job 会失败，属预期。
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - "pyproject.toml"
+
+permissions:
+  contents: read
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    steps:
+      # publish-node-action 自带 checkout（默认 skip_checkout=false），不要再写一遍
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@main
+        with:
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
@@ -29,4 +29,8 @@ jobs:
        run: npm test

      - name: Run backend tests
-        run: python -m unittest discover -s tests -p "test_*.py"
+        # server.py 依赖 aiohttp（由 ComfyUI 运行时提供，不作为插件依赖声明），
+        # test_upstream_gate 导入它，因此 CI 上要单独装一次。
+        run: |
+          pip install aiohttp
+          python -m unittest discover -s tests -p "test_*.py"
@@ -1,5 +1,158 @@
 # 更新记录

+## v2.1.9（2026-10-29）
+
+### 助手设置里可以自动获取模型列表
+
+- OpenAI 兼容和 Ollama 两段的「模型名称」旁边多了一个「获取模型」按钮：拿输入框里当前填的地址问一次服务端有哪些模型（OpenAI 兼容问 `/models`，Ollama 问 `/api/tags`），拉到的名字填进输入框自带的下拉——点输入框里的箭头就能选，名称也照旧可以手写。
+- 下拉用浏览器原生的方式挂在输入框上，不弹任何窗口：助手设置长在 ComfyUI 的「设置」面板里，浮层会被它挡住点不动。
+- 再点一次按钮就是重新拉取，按钮上显示拿到了几个；地址改了会自动清空，列表只在本次会话内有效、不写进配置文件。
+- 这次请求用你在 API Key 输入框里填的密钥；输入框留空就用已经保存的那一份（密钥仍不会回传给浏览器），只会随这一次请求发出。
+- 服务端返回看不懂的内容时（比如缺密钥时 DeepSeek 只回一行纯文本 `Authentication Fails`、或网关甩回来一整页登录页），界面直接显示那句原话和状态码，不再崩一段 `Expecting value: line 1 column 1 (char 0)`。「保存并测试连接」和翻译 / 优化走的是同一套解析，一起修好了。
+- 地址照旧走既有的白名单：远端必须 https、内网与链路本地地址驳回、Ollama 只认本机。
+
+### 设置面板两项统一靠左对齐
+
+- 「语言与外观」与「助手设置」两块的内容顶到最左边，左边缘齐平，看起来是一整块。标题仍在，由各自的标题行显示，收起的只是 ComfyUI 重复渲染出来的那个名称列。
+
+## v2.1.8（2026-10-26）
+
+### 新增「适应高度」开关
+
+- 提示词一多，英文标签区就卡在写死的 360 像素上限里，只能内部滚动，把节点拉大也没用。侧边栏「标签管理」的明细表上方（「颜色」按钮右边）多了一个「适应高度」滑动开关，跟着所选节点走——切到别的节点就显示那个节点自己的状态。
+- 打开后：面板不再被固定上限锁住，多出来的高度全部给英文标签区，提示词能多显示很多行；开关打开时节点会自动撑高一次（按被压住的内容量撑），之后想拉多大多大。
+- 关掉就回到原来的自动高度。开关状态跟着节点保存，重开工作流还在。
+- 默认关闭，提示词少的时候和以前完全一样。
+
+### 「优化」按模式分派，新增自然语言规则
+
+- 节点上的「优化为 Anima」改名为「优化」：标签模式下和以前一样，整理英文标签流、只接受英文；模式切到「自然语言」后改用新的自然语言规则，输出完整英文句子，中文也可以直接优化（会先译成英文，状态栏会提示一句）。
+- 助手设置里 OpenAI 兼容与 Ollama 各多一条「自然语言规则」，可以分别自定义，也能单独恢复默认。
+- 模式切到「自然语言」时「按 Anima 顺序排序」自动隐藏——整段句子没有标签流可排；已经排过序时按钮会保留，方便点「取消排序」。
+- 手动选的「自然语言」也算，不用非得自动检测出来；模式选择现在跟着节点保存，重开工作流不用重选。
+
+## v2.1.5（2026-10-13）
+
+### 可信度徽章改成按词条整理程度分档
+
+- 明细表的可信度只显示「未判定 / 低 / 中 / 高」四档，去掉「待确认」这类后缀，看起来干净了。
+- 判据：已收录、**已校验**、且填了分类（不是「未分类」）才算**高**；已收录但没填分类、或还没标成已校验的算**中**；机器翻译还没存进词库的算**低**；词库里查不到的仍是**未判定**。
+- 鼠标停在徽章上会说明为什么是这一档，填了分类的会把分类名带出来。
+- 「高」「中」两档补上中文译文——之前这两档一直显示的是英文。
+
+### 双击已收录标签就能改词条
+
+- 明细表里已收录的标签，双击英文那格会打开「新增或编辑个人标签」窗口，和在词库里点「编辑」是同一个：中文、别名、分类、模型、权重、备注都能改。
+- 未收录、待确认的机器翻译，以及 `BREAK`、`<lora:…>` 这类特殊语法，双击仍是原来的定位到提示词。
+- 未收录行去掉「搜索候选」按钮——词库里本来就搜不出这个词，点了只会打开空结果。
+
+### 自然语言片段双击也开编辑弹窗
+
+- 双击自然语言片段不再只提示「只有标签支持单独权重」，改成和标签一样打开编辑弹窗，只是只有「颜色」和「编辑」两页（权重和词条信息对整段文本没有意义）。
+- 自然语言片段可以像标签一样设置单独配色。
+
+## v2.1.0（2026-10-12）
+
+### 助手设置搬进 ComfyUI 设置面板
+
+- 侧边栏只保留「标签管理 / 词库 / 词包 / 收藏」四个页签，助手设置改到 ComfyUI 顶部「设置 → 双语提示词检查器 → 助手设置」，打开设置面板就能直接改，不用再进侧边栏找。
+- 「外观」和「语言」从助手设置里分出来，单独成为「语言与外观」一项，排在助手设置上面；下拉改完立即保存并生效，不用再点保存按钮。
+- 「清除已保存的 API 密钥」「清除已保存的百度密钥」从复选框改成滑动开关，打开后旁边显示「保存后生效」，再关掉可以取消。
+
+### OpenAI 兼容与 Ollama 在翻译服务里直接可选
+
+- 翻译服务下拉改成四个平级选项：纯词库 ／ 百度翻译 ／ OpenAI 兼容接口 ／ Ollama，不再需要先选「AI」再在下面选后端。
+- OpenAI 兼容和 Ollama 各有自己的地址、模型、密钥、温度、超时和三条翻译规则，来回切换不会互相覆盖，切换后原来填的密钥也还在。
+- Ollama 一栏没有密钥输入框。
+- Ollama 区块新增「使用 Ollama 本地预设」按钮，一键填入本机默认地址；OpenAI 兼容区块的「使用 LM Studio 本地预设」按钮不变。
+- 选纯词库或百度翻译时，会多出一个「AI 后端」下拉，决定「翻译并优化」和「优化为 Anima」用哪一套。
+- 升级后旧配置会按当前选择的后端搬进对应的一套，另一套保留默认值。
+
+### API 地址限制
+
+- OpenAI 兼容地址只接受 https，或本机地址（`127.0.0.1` / `localhost` / `::1`），方便 LM Studio 这类本地服务；http 只在填本机地址时允许。
+- Ollama 地址只接受本机地址。
+- 地址里不能带用户名和密码，也不能带 `?` 之后的参数。
+- 远端地址不允许是内网、保留或链路本地地址（例如 `192.168.x.x`、`10.x.x.x`、`169.254.169.254`）。
+
+### 百度翻译不再显示翻译规则
+
+- 翻译规则只有 AI 会用上，百度翻译和纯词库下面不再出现规则输入框。
+
+### 暗色模式下的表单配色
+
+- 输入框和下拉框改用 ComfyUI 自己的配色变量，不再另外写死一套，暗色主题下和设置面板里其它控件保持一致。
+
+## v2.0.6（2026-10-10）
+
+### 明细表可以按关键词查找
+
+- 明细表上方新增查找框，输入即筛选出当前节点里命中的标签，英文原文和中文解释都参与匹配。
+- 命中的片段会高亮，右侧显示命中数量，按 `Esc` 或点 × 一键清空。
+- 查找与「全部 / 未收录 / 机器翻译 / 个人词库 / 收藏」**叠加生效**，可以先筛来源再筛关键词。
+
+### 添加标签会插到选中标签之后
+
+- 在节点里、以及侧边栏搜索词库添加标签时，新标签插在当前选中标签的后面，不再一律追加到提示词末尾；没选中标签时才加到末尾。
+- 连续添加多个标签会依次往后排，添加完新标签自动处于选中状态。
+
+### Anima 排序可以一键取消
+
+- 排序写入后按钮变成「取消 Anima 顺序排序」，点击把提示词恢复到排序前的样子，同时退出分类表视图。
+- 提示词本来就是 Anima 顺序时，取消只退出分类视图，不动提示词内容。
+- 排序后又手动改过提示词的，取消入口会自动失效，避免回退把后来的改动一起吞掉。
+- 排序与优化的预览弹窗里，「确认写入英文」改成「确认写入新提示词」。
+
+### 侧边栏与词库
+
+- 词库来源下拉新增「自然语言词库」，只显示保存时选了该词库的条目；分类下拉里「自然语言」固定排在「全部」下面，其余仍按拼音排序。
+- 标签管理的节点下拉、以及保存收藏时选节点的弹窗，会显示在画布上双击改过的节点名（后面保留节点号，方便区分同名节点）。
+- 去掉侧边栏「标签管理」里的「管理面板」按钮——它和顶部「词库」页签是同一回事，工具栏只留「刷新词库」。
+
+## v2.0.0（2026-10-10）
+
+### 新增「重复检查」节点
+
+- 独立节点，检测一段提示词里的重复标签，把重复数量报出来；无重复时节点保持默认配色，有重复时转成淡橙色提示。
+- 开关「只保留第一个」默认关闭，打开后只保留每个标签第一次出现的写法。
+- 提示词直通下游，重复项另外输出，接预览任意可看。判定口径与双语节点完全一致（大小写、空格、下划线、权重、强调括号都视为同一个标签，`<lora:…>` 与 `BREAK`/`AND` 不参与）。
+
+### 重复标签可以设「唯一」，并新增重复栏
+
+- 「重复标签：X」后面给每个标签加了独立的「唯一」开关：打开后只有第一次出现会进实际输出，其余在界面上划线保留、**原文一个字不动**，支持 `Ctrl+Z` 撤回。
+- 英文标签区、「已隐藏」那一栏下面新增**重复栏**：一个标签一个胶囊，单击依次跳到每一次出现并联动选中（当前那处标红），已经定位时双击就删掉当前这一处。
+- 重复分组抽成 `parser.js` 的 `duplicateGroups()`，问题提示行、重复栏、重复检查节点共用同一份口径，不会再各算一套。
+
+### 其它
+
+- 标签卡上的隐藏（眼睛）按钮不再跟着英文行走，固定在整张卡片的右上角内侧；鼠标移到卡片任意位置（含中文行）都会显示它。
+
+## v1.2.18（2026-10-05）
+
+- **双击标签卡打开的编辑框支持一次改多个标签**：多选后双击，编辑范围是「最前一个选中标签的开头」到「最后一个选中标签的结尾」，中间没选中的标签也一起带进来；只在选中多于一个时给标签卡和明细表行加描边。
+- **「编辑」页签改成整段改写**：把这一段原文整块剪切出来改、改完粘回去（新增 `replacePromptSpan()`），不再是逐个标签拼装。清空即删除这一段，接缝处多余的逗号会被收掉，纯换行原样保留，多行排版不会被压平。
+- 标签字段抽成 `buildTagFields()`，编辑框与侧边栏「新增／编辑个人标签」弹窗共用同一份字段，不会再两处走样；弹窗与面板的样式细节同步调整。
+
+## v1.2.16（2026-10-01）
+
+- **快捷添加会插到选中标签后面**：选中某个标签时，新增的标签按它的结束位置插进去，不再一律丢到末尾；没选中或选中失效时仍落到末尾，判定与「插入换行」共用同一套。
+- **隐藏区的长文本胶囊不再被撑成大圆角块**：胶囊里的中文译文缺了单行限制，遇到长自然语言会在里面折行把高度顶起来。中英文两段都改成单行截断，高度恒定一行，超出部分显示省略号。
+
+## v1.2.15（2026-09-30）
+
+- **双击标签卡打开的编辑框新增「词条信息」页签**：现在三个页签——权重、颜色、词条信息。命中的词库词条会整条带过来改（分类、中文、来源、备注都能改），没命中就按当前标签卡起个头新建。词条字段抽成了 `buildTagFields()`，这个页签和侧边栏「新增／编辑个人标签」弹窗共用同一份字段，不会再两处走样。
+- 编辑框里的「Anima 格式」说明挪进权重页，「清除权重」按钮也只在权重页显示（之前在别的页签上也会出现，容易误点）。页签状态由原来的 `onColorTab` 布尔改成 `activeTab` 三态，底部共用一个「应用」按钮按当前页签分派。
+- Comfy Registry 的插件页面补上图标与横幅：`[tool.comfy]` 新增 `Icon`（400×400 方形）和 `Banner`（941×941 原图），在 Registry 节点页和 ComfyUI-Manager 里显示，方便按图找到本插件。
+
+## v1.2.14（2026-09-29）
+
+- **标签弹窗的「分类」改成可筛选下拉**（与「适合模型」同一套控件）：候选取词库里**真实出现过**的分类，按词条数降序，每项标出条数；输入即筛选，`↑` `↓` 切换、回车确认、`Esc` 收起，按 `↓` 或再点输入框可重新唤出列表。「自定义」恒为列表第一项且不会被过滤掉，选中后输入框右侧挂青绿色标识，表示「这个值不是词库现成的分类」；手打的值命中候选时自动摘掉标识，编辑历史标签带进来的陌生值当场也算自定义。新建时分类留空、由后端落成「自定义」，避免一打开窗口就挂着标识。
+- **标签卡新增两种上色方式**：过滤栏的「颜色」改成就地展开的面板，除原有的默认 / 按收录 / 随机外，新增**按词库**（每个词库一色）与**按 Anima**（按 Anima 官方分类上色）。配色写入 `data/token_colors.json` 的 `pack_colors` / `anima_colors` 并新增写接口；双击标签设的自定义色优先级最高。
+- **识别并标记「疑似自然语言」**：英文片段里空格超过 3 个时判定为疑似自然语言——这是**形状上的怀疑**，跟「有没有被词库收录」是两回事，两者互不替代。节点卡片与明细表都会着色标出；只在默认／按收录这类诊断模式下显示，「按词库」「随机」这类配色方案不会被它盖住。
+- 新增自然语言词库：从原词库里拆出独立的 `pack_id`，让自然语言片段有专门的来源标注。
+- 修复「优化为 Anima」的结果被转义符、句点和权重小数打断：清洗表补全角引号；清掉模型把结果写成 JSON／代码字符串时残留的反斜杠（字面 `\n` `\t` 先还原成换行，其余孤立反斜杠删除）；末尾孤立句点直接删除。
+- 修复权重小数让整串标签被误判成自然语言段落：判定句末标点前先中和数字句点（如 `1.5`）与闭括号后的句点。
+
 ## v1.2.10（2026-09-27）

 - **界面重构：英文输出与中文同步编辑合并成一张双语卡片**。节点里不再有上下两个编辑区，每个标签上面英文、下面中文，横排自动折行，节点一下子矮了一半。随之移除中文编辑框、编辑态、草稿、暂存、同步预览，以及「编辑文本／翻译／翻译并优化／展开编辑／同步到英文输出」这批只在中文区才需要的按钮；「优化为 Anima」改为直接处理英文提示词，出结果后先弹预览确认再写回。
@@ -23,6 +176,7 @@
 - **修复插件前端扩展被加载两遍**：`pyproject.toml` 的 `[tool.comfy] web = "js"` 与 `__init__.py` 的 `WEB_DIRECTORY = "./js"` 会被 ComfyUI 分别登记（key 一个是 `project.name`、一个是模块名），`/extensions` 接口因此把同一批 js 列两次，前端各加载一次并报 `Extension named ... already registered.`。已移除 `web` 字段，前端资源统一由 `WEB_DIRECTORY` 提供。
 - 上游连接判定改用 `node.isInputConnected(slot)`（老版本无此方法时退回读 `input.link`），消除新版前端里 `input.link is deprecated` 的控制台刷屏。

+
 ## v1.2.0（2026-09-23）

 - 翻译服务与 AI 后端解耦（schema v3）：翻译/解释可选纯词库 / 百度翻译 / AI；"翻译并优化"与"优化为 Anima"恒走 AI。AI 与百度密钥独立留存，切换翻译服务互不清空——只有 AI 后端/地址变更或换机器时才清 AI Key，百度密钥只在显式清除时才清。
@@ -31,7 +185,6 @@
 - 修复 `clear_ai_api_key` / `clear_baidu_secret_key` 在 payload 未同时包含对应密钥字段时不生效的嵌套条件 bug。
 - 修复节点高度棘轮上涨：ComfyUI `text` 控件的 `computedHeight` 是单调累加器，每次写值都会让节点变大一点。改用原生 `text` 控件的 `scrollHeight@auto` 作为稳定基线，不再读 `widget.y`，节点高度在多次写值后锁定不变。
 - `inspectorNodeTargetHeight` 签名简化为 `{baseHeight, targetWidgetHeight, fallbackBaseHeight, minimumHeight}`，移除旧 `widgetY`/`nodeHeight`/`currentWidgetHeight` 等参数。
-
 - 新增标签拖拽排序：英文标签区可按住标签直接拖动，明细表每行左侧提供 `⠿` 手柄（含插入指示线与边缘自动滚动），选中标签后支持 `Alt+↑` / `Alt+↓` 键盘微调；三处入口共用同一套 `Ctrl+Z` 撤销历史。
 - 排序按“提取—插入”建模：保留原有逗号、空格与换行排版，多行结构不被合并；自然语言块整体移动，内部不拆分；重复标签按出现位置独立移动。
 - 明细表筛选视图（未知／机器译／个人词库／收藏）与自然语言、指令模式下不启用拖拽，避免局部顺序被误解为完整顺序。
@@ -52,16 +205,13 @@
 - 收藏数据存在 ComfyUI 用户目录 `user/bilingual-prompt-inspector/saved_prompts/`（`index.json` + 独立图片文件），不写入插件目录，重装插件不丢数据；单图上限 5 MB，最多 2000 条。
 - 收藏不记录工作流信息；「载入节点」先弹窗确认再覆盖节点文本，不进 `Ctrl+Z` 撤销栈。侧边栏「新建」在没有选中节点时提供节点选择器兜底。
 - 新增 `tests/test_saved_prompt_store.py`：覆盖图片魔数校验、超限拒绝、删除连带删图、导出导入往返（保留原顺序）、重复与非法条目跳过、非法时间戳回退等 14 个用例。
-
 - 新增外观模式配置：助手设置里可切换「跟随 ComfyUI ／ 暗色 ／ 亮色」，默认跟随——ComfyUI 用什么主题插件就用什么主题。此前插件把配色硬编码成暗色值，在 ComfyUI 亮色模式下文字与背景几乎无法分辨。
 - 实现上新增一组 `--bpi-*` CSS 变量（面板、文字、边框、按钮、输入框、表格、弹窗），按 `:root`（亮色默认）、`.dark-theme`（跟随 ComfyUI 暗色）、`[data-bpi-theme]`（强制暗／亮）四套定义；配置存在 `assistant_store` 的 `appearance` 字段，切换后即时生效无需刷新。
 - 节点中文同步编辑区的「标签管理」「助手设置」入口已移除（与侧边栏管理面板重复），代码里保留注释说明如何恢复；助手设置里的翻译服务说明改为整行显示，不再挤在表单左列。
-
 - 用户界面英文化：为符合 ComfyUI-Manager 注册要求，节点显示名改为 `Prompt Translator & Manager`，`@title` / `@description` / `RETURN_NAMES` / `CATEGORY` 与约 540 处 JS 界面文字（按钮、提示、状态消息、输入框占位符、语法诊断）全部改为英文。
 - 中文通过 locale 保留：新增 `locales/zh/main.json`（`nodeDefs` 结构），在 ComfyUI 里把界面语言切到中文后，节点名、描述与输入输出名仍显示中文「提示词翻译与管理（英文输出）」——由 ComfyUI 的 `build_translations()` 与前端 `mergeCustomNodesI18n()` 自动合并，无需额外配置。
 - 英文化时有四类内容刻意保留中文：代码注释；匹配用户中文输入的正则模式（如「翻译要求／正文」结构化指令识别）；词库中文分类名匹配表（改掉会导致分类排序失效）；以及「未收录」「自然语言片段（待翻译或确认）」等在中文列显示的翻译数据。
 - 更新 5 个测试文件的断言以匹配新的英文字符串（`test_anima_sorter.mjs`、`test_clear_controls.mjs`、`test_dictionary_tools.mjs`、`test_parser.mjs`、`test_project_link.mjs`）。
-
 - 新增界面语言设置：助手设置页新增「Language」下拉，可选「跟随 ComfyUI 语言设置」（默认）／「中文」／「英文」，随其它设置一起存在 `assistant_store` 的 `language` 字段（auto / zh / en），并在 `localStorage` 留一份供首屏同步取值。
 - ComfyUI 自带的 locale 只翻译 `nodeDefs`（节点名、描述、输入输出名），面板里的按钮、状态栏、tooltip 管不到，因此新增插件自建的 i18n 层 `js/i18n.js`：源码继续保留英文原文（上架要求），运行时按原文查中文表，查不到就原样显示英文，不会出现半中半英或空文案。
 - 文案统一走 `t()`：`element()` / `button()` / `field()` 三个 DOM 工厂与全部 `title` / `placeholder` / `textContent` 赋值点都接入翻译（共 200+ 条文案）。带变量的文案（如 `deleted “xxx”; press Ctrl+Z to undo`）把「“…”」片段和数字折叠成 `{}` 占位符再回填，模板字符串也能整句翻译。
@@ -71,7 +221,7 @@
 - 补齐第二批翻译（约 200 条）：词包卡片信息（个人词库／大词库／社区词包的条数、版本、来源、授权）、标签管理的统计行「共 N 个标签｜已识别｜未收录｜待确认」、模式提示、各类操作回执（移动／隐藏／恢复／翻译／导入）、语法诊断全部文案（括号、权重、重复标签、冲突组、输入模式判定理由）、Anima 分类名、以及所有请求失败提示。
 - 变量的递归翻译：嵌在文案里的变量本身也可能是界面词（如冲突组名 `Solo vs multiple subjects`、模式名 `Tag`），回填时会再翻译一层，避免出现「中文句子里夹一个英文词」。
 - 为让变量能被稳定折叠，`parser.js` 的诊断模板与词包信息模板给动态片段加上「“...”」包裹（版本号 `1.0.0`、来源、授权等），否则带点号的版本会被拆成多个变量而匹配不到译表。
- 统计行改为分段翻译再拼接：词库页的 `Showing … | Selected … | Enabled packs … | Built-in … | Large … | Personal … | Pending …` 与节点里的 `Common dictionary … items | Large …` 原来是整句做键，大词库有「未安装 / 按需（N 条）/ 已停用」三种形态，整句键必然漏掉其中几种。现在每段各自翻译后再用 ` | ` 拼起来，新增形态不会再漏。
+- 统计行改为分段翻译再拼接：词库页的 `Showing … | Selected … | Enabled packs … | Built-in … | Large … | Personal … | Pending …` 与节点里的 `Common dictionary … items | Large …` 原来是整句做键，大词库有「未安装 / 按需（N 条）/ 已停用」三种形态，整句键必然漏掉其中几种。现在每段各自翻译后再用 `|` 拼起来，新增形态不会再漏。
 - 修掉一类「译表里有、界面上却还是英文」的隐蔽 bug：模板里的自由文本变量（标签名、词包名、错误信息、文件名、操作名）没有用「“...”」包裹时不会被折叠成 `{}`，整句变成带具体内容的字面量，永远匹配不到译表。已统一给 27 处这类变量加上包裹，并把 4 处误用 ASCII 直引号的地方（`"Re-translated \"...\""` 等）换成全角引号；译表里 6 条把直引号烤进键名的条目（`"Processed \"{}\""` 等）同步改成纯 `{}`。
 - 补齐第三批翻译（22 条）：置信度徽章（`Low · pending` / `Medium · pending` / `Medium · session-only`）、收藏星标按钮（★/☆）、`⚠ Not indexed:` 标签提示、Anima 格式说明、`Node #N · N tags` 收藏备注、`Running …` / `Translating unknown tags N/M` / `Translation complete …` 三条运行状态、`Querying dictionary…` / `Searching for more…` / `Paused, waiting for confirmation…`、权重越界诊断、以及词包导入预览与删除确认。
 - 两处组合句改分段翻译：词包导入预览 `Tags N | Sample: …`（示例内容是数据，不该参与翻译）与标签管理节点标题 `Node #N: …`（提示词原文不翻译）。
@@ -26,9 +26,23 @@ include credentials or private prompt content.

 # 版本号与发布（改版本前必读）

+## 一键升版本（平时用这个就够了）
+
+    python tools/bump_version.py 1.2.16
+
+脚本会把下面 5 处全部改好，并在 `CHANGELOG.md` 顶部插入新的一节（正文留「- （待补充）」，自己填）。
+只做字符串替换，`pyproject.toml` 里的 description / Icon / Banner 原样保留。
+任何一处没匹配上就整体中止，不会留下改了一半的文件。
+
+脚本跑完之后剩下的事（commit / push / 打 tag）自己来做：
+
+    git add -A && git commit -m 'chore: 发布 v1.2.16 / Release v1.2.16'
+    git push origin main
+    git tag -a v1.2.16 -m 'v1.2.16' && git push origin v1.2.16
+
 ## 需要同步修改的 5 个文件

-升版本号时**这 5 处必须一起改**，漏一处就会出现"上架版本是 1.2.1、关于框还显示 1.2.0"这类不一致。
+升版本号时**这 5 处必须一起改**（即上面脚本做的事，手动改时照这张表逐项核对），漏一处就会出现"上架版本是 1.2.1、关于框还显示 1.2.0"这类不一致。

 | # | 文件 | 位置 | 当前写法 | 说明 |
 |---|------|------|----------|------|
@@ -51,26 +65,30 @@ include credentials or private prompt content.

 ## 发布顺序

-```bash
-git push origin main
-git tag -a v1.2.0 -m "v1.2.0"      # 版本号换成当次版本
-git push origin v1.2.0
-```
-
-然后在 GitHub 上基于该标签建 Release（内容取自 CHANGELOG 对应章节），最后：
+Registry 的发布由 `.github/workflows/publish_action.yml` **自动完成**：并入 main 且 `pyproject.toml` 有变动
+（实际就是版本号升了）时触发。所以在分支上怎么改都不会发布，只有合进 main 那一刻才对外发一次。

 ```bash
-comfy node publish     # 需要 registry.comfy.org 生成的 API key
+git push origin main                 # 版本号已升 → 自动发布
+git tag -a v1.2.10 -m "v1.2.10"      # 标签只用于 GitHub Release 与回溯，Registry 不读它
+git push origin v1.2.10
 ```

+版本号没升又想重发时，到 GitHub 的 Actions → "Publish to Comfy Registry" → Run workflow（预留了 `workflow_dispatch`）。
+本地 `comfy node publish` 只在 Actions 不可用时才用——它打的包是 **git 已跟踪的文件**，没 commit 的改动不会进包。
+
 ## 发布前必须确认（不可撤销）

 - `pyproject.toml` 的 **`name`（`prompt-translator-manager`）发布后永久不可改**，且全局唯一
 - **`version` 发布后不可覆盖**：1.2.0 发出去后，哪怕只改一个标点也只能发 1.2.1
 - 确认 `.comfyignore` 没有误伤 `data/`（内置词库）、`assets/`（README 截图）、`locales/`（中文节点名）——排除了插件会直接跑不起来

-## 自动发布（可选）
+## 自动发布（已启用）

-仓库里目前只有 `.github/workflows/tests.yml`，没有发布 workflow。想实现"升版本号 → 推 main → 自动发布"，可加
-`.github/workflows/publish_action.yml`（用 `Comfy-Org/publish-node-action@main`），并把 registry 的 API key
-存为仓库 secret `REGISTRY_ACCESS_TOKEN`。该 workflow 只在 `pyproject.toml` 变动时触发。
+`.github/workflows/publish_action.yml` 用 `Comfy-Org/publish-node-action@main`，凭据是仓库 secret
+`REGISTRY_ACCESS_TOKEN`（在 registry.comfy.org 进自己的 publisher 页面生成 API Key 后，填到
+Settings → Secrets and Variables → Actions → New repository secret）。
+
+- 触发条件：`push` 到 `main` 且 `pyproject.toml` 有变动，或手动 `workflow_dispatch`。
+- **所以每次要发布都必须 bump 版本号**：没升版本时 Registry 回 `Version already exists`，job 会报红，这是预期行为不是故障。
+- 反过来，改了 `pyproject.toml` 里除 `version` 以外的字段也会触发发布，同样会因版本已存在而失败——非必要别动那些字段。
@@ -1,4 +1,4 @@
-# ComfyUI 双语提示词管理器 v1.2.10
+# ComfyUI 双语提示词管理器 v2.1.9

 > ComfyUI 双语提示词节点：把英文提示词拆成可点击的标签，边写边看中文解释、调顺序、试效果。

@@ -67,14 +67,16 @@
 <td>提示词里的换行会渲染成行尾的 <code>↵</code> 块，标签区真的断行，中英两栏行数对齐。换行块本身也能拖动。</td>
 </tr>
 </table>
-![Demo — 编辑](https://raw.githubusercontent.com/mengshengzhijie/tuchuan/master/assets/ComfyUI-Bilingual-Prompt-Inspector/demo-editing.gif)
+![Demo — edit](https://raw.githubusercontent.com/mengshengzhijie/tuchuan/master/assets/ComfyUI-Bilingual-Prompt-Inspector/demo-editing.gif)

 ### 翻译与整理

 | 能力 | 说明 |
 |---|:-:|
-| **四种处理方式** | 仅翻译 / 翻译并优化 / 优化为 Anima / 按 Anima 官方顺序整理（不新增标签，只重排分段） |
-| **自动识别模式** | 自动区分纯标签、纯自然语言、混合文本和翻译指令，不用手选模式 |
+| **四种处理方式** | 仅翻译 / 翻译并优化 / 优化 / 按 Anima 官方顺序整理（不新增标签，只重排分段） |
+| **优化按模式分派** | 标签模式下「优化」整理现有英文标签流（只吃英文）；自然语言模式下改写为英文自然语句，中文会先译成英文。两种模式各用一套规则，可在助手设置里分别自定义 |
+| **自动识别模式** | 自动区分纯标签、纯自然语言、混合文本和翻译指令，也可以在明细表上方手动指定；指定的模式跟着节点保存 |
+| **自然语言模式** | 模式切到「自然语言」后，「按 Anima 顺序排序」自动隐藏——整段句子没有标签流可排 |
 | **提示词诊断** | 自动标出可疑权重、重复标签、语法问题 |
 | **快捷添加** | 底部的快捷输入框：写中文自动翻译成英文标签后追加，写英文标签直接追加；`Enter` 添加、`Shift+Enter` 换行、`Ctrl+Enter` 翻译并优化 |

@@ -85,7 +87,7 @@
 | **多来源词库** | 内置词库 + 个人词库 + 社区词库包，每条解释都会标注来源 |
 | **概念搜索** | 搜「丝袜」能返回连裤袜、长筒袜、过膝袜等相关标签，支持滚动分页加载 |
 | **大型 Danbooru 词库（可选）** | 按需安装，覆盖面更广，不装也不影响其它功能 |
-| **翻译服务可切换** | 纯词库（离线、免费）/ 百度翻译 / AI（OpenAI 兼容 API 或 Ollama）。翻译与解释可选任意一种；「翻译并优化」「优化为 Anima」恒定走 AI |
+| **翻译服务可切换** | 纯词库（离线、免费）／ 百度翻译 ／ OpenAI 兼容 API ／ Ollama。翻译与解释可选任意一种；「翻译并优化」「优化」恒定走 AI |
 | **密钥互不干扰** | AI 与百度密钥独立留存，切换翻译服务不会清空；均不写入工作流、不返回浏览器 |

 ### 上游提示词接管
@@ -122,7 +124,9 @@

 ### 侧边栏管理面板

-标签管理、词库维护、词库包、助手设置、收藏五个分页都收在侧边栏的「双语提示词管理」面板里，节点本身保持简洁。面板里的改动会实时同步到页面上所有同类节点，不需要刷新页面。
+标签管理、词库维护、词库包、收藏四个分页收在侧边栏的「双语提示词管理」面板里，节点本身保持简洁。面板里的改动会实时同步到页面上所有同类节点，不需要刷新页面。
+
+助手设置不在这个面板里：它在 ComfyUI 的「设置」面板中，「设置 → 双语提示词检查器」下有两项 ——「语言与外观」和「助手设置」，表单都直接铺在设置面板右侧。

 ---

@@ -132,7 +136,7 @@

 节点上可以大致分成四块：

-1. **英文原文（实际输出）**——节点真正输出给下游的文本。上面有「编辑文本 / 完成」「清空」「插入换行」「收藏」「优化为 Anima」「按 Anima 顺序排序」几个按钮。
+1. **英文原文（实际输出）**——节点真正输出给下游的文本。上面有「编辑文本 / 完成」「清空」「插入换行」「收藏」「优化」「按 Anima 顺序排序」几个按钮。
 2. **中文对照区**——英文标签逐条对应的中文，只用于阅读，不进入输出，也不会传给下游。
 3. **明细表**——每行是一个标签：来源、翻译、权重、分类，右侧按行状态给出操作（翻译 / 确认并存 / 重新翻译 / 编辑中文 / 收藏 / 复制 / 删除个人词条）。上方有搜索框、模式选择、颜色模式、筛选（全部 / 未知 / 机器翻译 / 个人词库 / 收藏）和问题提示。
 4. **快捷添加**——底部输入框，中文进英文出（自动翻译后追加），英文则直接追加。
@@ -141,15 +145,24 @@

 ## 安装

-**用 ComfyUI-Manager（推荐）**：在 Manager 的节点列表里搜索 `Bilingual Prompt Inspector` 安装；也可以走 Manager 的「Install via Git URL」，填入下面的仓库地址。
+**用 ComfyUI-Manager（推荐）**：在 Manager 的节点列表里搜索 `Bilingual Prompt Inspector` 安装；也可以走 Manager 的「Install via Git URL」，填入仓库的仓库地址。
+
+**使用comfy-cli安装**：
+
+```bash
+cd ComfyUI/
+comfy node install prompt-translator-manager
+```

 **手动安装**：

 ```bash
 cd ComfyUI/custom_nodes/
-git clone https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.git
+git clone --depth 1 https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.git
 ```

+不过这里如果有动手能力的话推荐使用git安转，因为现在就我和我的朋友在使用，有时候修复bug和制作新功能我是总结下来然后一步一步的修复。直到把这次收录的bug和建议给完成之后才bump版本。
+
 重启 ComfyUI，在节点菜单的 **text/prompt** 分类里找到 **提示词翻译与管理（英文输出）**（英文界面下显示为 Prompt Translator & Manager），也可以直接搜索「提示词」。

 两个常见坑：
@@ -168,7 +181,7 @@ git clone https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.
 3. 在标签视图里逐个检查：点标签看中文解释，双击调权重，拖动手柄排顺序。
 4. 想确认某个标签到底有没有用？点眼睛图标隐藏它，重新出图对比，再一键恢复。

-![Demo — 上手](https://raw.githubusercontent.com/mengshengzhijie/tuchuan/master/assets/ComfyUI-Bilingual-Prompt-Inspector/demo-run.gif)
+![Demo — run](https://raw.githubusercontent.com/mengshengzhijie/tuchuan/master/assets/ComfyUI-Bilingual-Prompt-Inspector/demo-run.gif)

 ---

@@ -176,13 +189,19 @@ git clone https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.

 ### 翻译服务配置

-默认走纯词库，不联网也能用。需要机器翻译或 AI 优化时，打开侧边栏「双语提示词管理」面板的「助手设置」分页。面板分三段：
+默认走纯词库，不联网也能用。需要机器翻译或 AI 优化时，打开「设置 → 双语提示词检查器 → 助手设置」。这一项分三段：

 | 分段 | 内容 |
 |---|---|
-| **翻译服务** | 翻译 / 解释按此选择：纯词库（离线、免费）／ 百度翻译（需自行申请开放平台的 APP ID 与密钥）／ AI（走下面的 AI 配置） |
+| **翻译服务** | 翻译 / 解释按此选择：纯词库（离线、免费）／ 百度翻译（需自行申请开放平台的 APP ID 与密钥）／ OpenAI 兼容接口 ／ Ollama（本地）。选后面两个时，它们本身就是下面的 AI 服务 |
 | **百度翻译** | 独立配置 APP ID 与密钥，切换翻译服务不会清空 |
-| **AI 服务** | 后端（OpenAI 兼容 / Ollama）、API 地址与模型、可选 API Key、温度、超时。「翻译并优化」「优化为 Anima」恒定走这里 |
+| **AI 服务** | 「翻译并优化」「优化」恒定走这里。翻译服务选了 OpenAI 兼容 / Ollama 时，下面直接就是那个后端的配置；选纯词库 / 百度翻译时，这里多出一个「AI 后端」下拉，决定优化用哪一套。两个后端各存一套配置，切换不清空： |
+| ↳ OpenAI 兼容 | API 地址、模型、可选 API Key、温度、超时，另有「使用 LM Studio 本地预设」一键填入 `1234/v1`。远端地址必须 https，本机回环（127.0.0.1 / localhost）才允许 http；内网、保留、链路本地地址一律拒绝 |
+| ↳ Ollama | API 地址、模型、温度、超时，没有密钥，另有「使用 Ollama 本地预设」一键填入 `11434`。地址只能是本机回环地址，其它一律拒绝 |
+
+模型名称旁边的「获取模型」按钮会拿输入框里当前填的地址问一次服务端有哪些模型（OpenAI 兼容问 `/models`，Ollama 问 `/api/tags`），拉到的名字直接填进输入框自带的下拉里——点输入框里的箭头就能选，也可以完全不管它自己手写。再点一次按钮就是重新拉取。列表只在本次会话内有效、地址一改就作废，不写进配置文件。密钥以输入框里填的为准，输入框留空就用你已经保存的那一份；无论哪种都只随这一次请求发出，不会随着设置一起回传给浏览器。服务端返回看不懂的内容（纯文本报错、登录页之类）时会原话告诉你，不会甩一段 JSON 报错。
+
+两个后端还各有一套规则（翻译 / 翻译并优化 / 优化 / 自然语言）：本地小模型与云端大模型要写的提示词往往不一样，分开存可以随时切换后端而不用重改规则。其中「自然语言」那套只在自然语言模式下被「优化」使用，结果按完整英文句子保留标点，不套 Anima 的标点限制。地址不合规时会在发出请求前直接报错。

 ### 大型词库（可选）

@@ -190,7 +209,7 @@ git clone https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.

 ### 界面语言与外观

-界面文案跟随 ComfyUI 的语言设置，也可以在侧边栏面板的「助手设置」分页里固定为中文或英文。外观同样可在「跟随 ComfyUI / 暗色 / 亮色」之间切换。
+界面文案跟随 ComfyUI 的语言设置，也可以在「设置 → 双语提示词检查器 → 语言与外观」里固定为中文或英文。外观同样可在「跟随 ComfyUI / 暗色 / 亮色」之间切换，两项改完立即生效并单独保存。

 ---

@@ -204,7 +223,7 @@ git clone https://github.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector.
 <details>
 <summary><b>点翻译提示要先配置翻译服务</b></summary>

-纯词库模式下只能查词库，不会翻译。想翻译请在侧边栏面板的「助手设置」分页里选一个翻译服务（百度或 AI）。
+纯词库模式下只能查词库，不会翻译。想翻译请在「设置 → 双语提示词检查器 → 助手设置」里选一个翻译服务（百度翻译、OpenAI 兼容接口或 Ollama）。
 </details>

 <details>
@@ -31,3 +31,24 @@ unauthenticated ComfyUI instance directly to the public internet.
 External translation and optimization requests are sent only to the provider configured
 by the user. Review that provider's privacy policy, logging behavior, content policy,
 and cost before use.
+
+## Outbound request targets
+
+AI requests are sent only to the address stored in the local configuration file, and that
+address is validated before a request is sent:
+
+- OpenAI-compatible endpoints must use HTTPS. Plain HTTP is accepted only for loopback
+  addresses on the same machine (127.0.0.0/8, ::1, localhost), so a local runtime such as
+  LM Studio keeps working without sending the API key in clear text across the network.
+- Remote OpenAI-compatible endpoints must be public. An address written directly as a
+  private, reserved or link-local IP (192.168.x.x, 10.x.x.x, 172.16.x.x, 169.254.169.254,
+  ...) is rejected even over HTTPS. Domain names are accepted; they are not resolved during
+  validation.
+- Ollama endpoints must be a loopback address on the same machine.
+
+Credentials embedded in the URL (`user:pass@`) and query strings are rejected as well, and
+only the `http` and `https` schemes are allowed. The address is re-validated immediately
+before each request, so editing the configuration file by hand cannot bypass these rules.
+
+The large Danbooru dictionary is a local read-only SQLite file under the extension's data
+directory. Despite the `connect(` call, it opens a file on disk and opens no socket.
@@ -1,13 +1,15 @@
-from .nodes import BilingualPromptInspector
+from .nodes import BilingualPromptInspector, DuplicateChecker
 from . import server  # noqa: F401 - importing registers local API routes


 NODE_CLASS_MAPPINGS = {
    "BilingualPromptInspector": BilingualPromptInspector,
+    "DuplicateChecker": DuplicateChecker,
 }

 NODE_DISPLAY_NAME_MAPPINGS = {
    "BilingualPromptInspector": "Prompt Translator & Manager",
+    "DuplicateChecker": "重复检查",
 }

 WEB_DIRECTORY = "./js"
@@ -46,17 +46,30 @@ DEFAULT_OPTIMIZATION_RULE = """你是 Anima 模型英文提示词优化器。输
 对于明确为成年人的合法成人内容，准确保留并规范相关成人标签，不弱化、不含糊替换，也不主动增加原文没有的成人内容；不得将未成年人或年龄不明确的人物成人化。
 标签使用英文逗号分隔；自然语言只放在末尾并使用英文句点分隔。仅使用 Anima 支持的英文标点和语法：逗号、句点、圆括号与冒号权重、@ 画师前缀、下划线、连字符、撇号、百分号，以及尖括号 LoRA 语法。"""

+DEFAULT_NATURAL_RULE = """你是面向现代生图模型（Krea、FLUX、Qwen-Image 等）的英文自然语言提示词助手。把用户输入整理成一段通顺、具体、可直接使用的英文描述；只返回最终结果，不输出 Markdown、标题、引号、解释、警告或免责声明。
+输入是中文、英文或中英混合都可以：含中文时先忠实译为英文再整理表达，纯英文时只做表达优化，不改变原意。输入是逗号分隔的标签时，先理解这些标签的含义，再改写成一段连贯的英文自然语言描述。
+保留主体、动作与状态、外貌与服装、环境与背景、光线、镜头与构图、风格与画材、氛围与情绪；原文没有给出的要素不要凭空添加，也不要删掉原文已经写明的限制。
+严格保留人物数量、成年人身份、角色名、作品名、否定关系、左右位置和安全等级，不得擅自换人、增加人物或改变主题。
+对于明确为成年人的合法成人内容，准确保留相关描述，不弱化、不含糊替换，也不主动增加原文没有的成人内容；不得将未成年人或年龄不明确的人物成人化。
+用完整的英文句子输出，按语法正常使用逗号、句点和连接词；不要写成逗号分隔的标签列表，也不要使用括号权重、@ 画师前缀或尖括号语法。"""
+
 DEFAULT_CONFIG = {
-    "schema_version": 3,
-    # 翻译/解释走哪个服务；优化永远走 AI，与此无关
-    "translate_service": "dictionary",  # dictionary | baidu | ai
-    # AI 后端独立配置，不随翻译服务切换而清空
+    "schema_version": 4,
+    # 翻译/解释走哪个服务；OpenAI 兼容与 Ollama 本身就是两个平级选项
+    "translate_service": "dictionary",  # dictionary | baidu | openai_compatible | ollama
+    # 选词库/百度时翻译不走 AI，但「翻译并优化」「优化」恒走 AI —— 这时用它决定读哪一套
    "ai_provider": "openai_compatible",  # openai_compatible | ollama
-    "ai_base_url": "",
-    "ai_model": "",
-    "ai_api_key": "",
-    "ai_temperature": 0.2,
-    "ai_timeout_seconds": 90,
+    # OpenAI 兼容：远端必须 https，本机回环 http 也放行（本地 LM Studio 之类）
+    "openai_base_url": "",
+    "openai_model": "",
+    "openai_api_key": "",
+    "openai_temperature": 0.2,
+    "openai_timeout_seconds": 90,
+    # Ollama：只允许本机回环地址，没有密钥
+    "ollama_base_url": "http://127.0.0.1:11434",
+    "ollama_model": "",
+    "ollama_temperature": 0.2,
+    "ollama_timeout_seconds": 90,
    # 百度翻译独立配置，不随翻译服务切换而清空
    "baidu_appid": "",
    "baidu_secret_key": "",
@@ -66,11 +79,24 @@ DEFAULT_CONFIG = {
    "appearance": "auto",  # auto | dark | light
    # 插件面板自己的界面语言：auto 跟随 ComfyUI 的 Comfy.Locale，zh / en 强制指定
    "language": "auto",  # auto | zh | en
-    "translation_rule": DEFAULT_TRANSLATION_RULE,
-    "translate_optimize_rule": DEFAULT_TRANSLATE_OPTIMIZE_RULE,
-    "optimization_rule": DEFAULT_OPTIMIZATION_RULE,
+    # 每个后端一套规则：本地小模型和云端大模型要写的提示词不一样
+    "openai_translation_rule": DEFAULT_TRANSLATION_RULE,
+    "openai_translate_optimize_rule": DEFAULT_TRANSLATE_OPTIMIZE_RULE,
+    "openai_optimization_rule": DEFAULT_OPTIMIZATION_RULE,
+    "openai_natural_rule": DEFAULT_NATURAL_RULE,
+    "ollama_translation_rule": DEFAULT_TRANSLATION_RULE,
+    "ollama_translate_optimize_rule": DEFAULT_TRANSLATE_OPTIMIZE_RULE,
+    "ollama_optimization_rule": DEFAULT_OPTIMIZATION_RULE,
+    "ollama_natural_rule": DEFAULT_NATURAL_RULE,
 }

+# 规则按后端分成两套，遍历时统一用这两个前缀
+AI_PROVIDER_PREFIXES = {"openai_compatible": "openai", "ollama": "ollama"}
+# translate_service 里属于 AI 的两个值，与 ai_provider 共用同一套前缀
+AI_SERVICES = frozenset(AI_PROVIDER_PREFIXES)
+TRANSLATE_SERVICES = frozenset({"dictionary", "baidu"}) | AI_SERVICES
+RULE_KEYS = ("translation_rule", "translate_optimize_rule", "optimization_rule", "natural_rule")
+
 BAIDU_TRANSLATE_ENDPOINT = "https://fanyi-api.baidu.com/api/trans/vip/translate"

 # 百度「通用文本翻译」按账号档位给 QPS：标准版 1 QPS，个人认证 10 QPS，企业认证更高。
@@ -78,6 +104,8 @@ BAIDU_TRANSLATE_ENDPOINT = "https://fanyi-api.baidu.com/api/trans/vip/translate"
 BAIDU_QPS_DEFAULT = 1.0
 BAIDU_QPS_MIN = 0.1
 BAIDU_QPS_MAX = 50.0
+# 百度单块最多 600 字，90 秒与原默认值一致（AI 拆成两套后不再借用它的超时）
+BAIDU_TIMEOUT_SECONDS = 90


 def normalize_baidu_qps(value):
@@ -110,6 +138,59 @@ def translation_direction(text):
    return "to_english" if re.search(r"[\u3400-\u9fff]", str(text or "")) else "to_chinese"


+def is_loopback_host(hostname):
+    """只允许真正的回环地址：127.0.0.0/8、::1、localhost。
+
+    比「私有网段」窄得多 —— 私有网段里还有路由器、NAS、同事的机器，
+    SSRF 正是靠它们横向探测，所以这里一律不放行。
+    """
+    host = str(hostname or "").strip().lower().rstrip(".")
+    if host in {"localhost", "127.0.0.1", "::1"}:
+        return True
+    try:
+        return ip_address(host).is_loopback
+    except ValueError:
+        return False
+
+
+def is_public_address(hostname):
+    """主机名直接写成 IP 时，只放行公网地址；域名不做解析，一律放行。
+
+    挡的是 https://192.168.x.x、https://10.x.x.x、https://169.254.169.254（云元数据）这类
+    「虽然是 https、但仍然指向内网」的直连写法。
+    """
+    try:
+        return ip_address(hostname).is_global
+    except ValueError:
+        return True
+
+
+def normalize_ai_base_url(value, *, allow_remote=False):
+    """校验 AI 地址，返回去掉尾斜杠的原文；不合规直接抛错，不发请求。
+
+    allow_remote=False（Ollama）：只能是本机回环地址。
+    allow_remote=True（OpenAI 兼容）：https 一律放行，http 只放行本机回环。
+    """
+    base = str(value or "").strip().rstrip("/")
+    if not base:
+        return ""
+    parsed = urlparse(base)
+    if parsed.scheme not in {"http", "https"} or not parsed.netloc:
+        raise ValueError("API 地址必须是完整的 http:// 或 https:// 地址")
+    if parsed.username or parsed.password:
+        raise ValueError("API 地址中不能包含用户名或密码")
+    if parsed.query or parsed.fragment:
+        raise ValueError("API 地址中不能包含查询参数或片段")
+    loopback = is_loopback_host(parsed.hostname)
+    if not allow_remote and not loopback:
+        raise ValueError("Ollama 地址只能是本机地址（127.0.0.1 / localhost / ::1）")
+    if parsed.scheme == "http" and not loopback:
+        raise ValueError("非本机 API 地址必须使用 HTTPS，避免密钥明文传输")
+    if allow_remote and not loopback and not is_public_address(parsed.hostname):
+        raise ValueError("远端 API 地址不能是内网、保留或链路本地地址")
+    return base
+
+
 def openai_chat_endpoint(base_url):
    """Return a chat-completions endpoint for an OpenAI-compatible base URL."""
    base = str(base_url or "").strip().rstrip("/")
@@ -124,6 +205,68 @@ def openai_chat_endpoint(base_url):
    return f"{base}/chat/completions"


+def openai_models_endpoint(base_url):
+    """Return a model-list endpoint for an OpenAI-compatible base URL."""
+    base = str(base_url or "").strip().rstrip("/")
+    if not base:
+        return ""
+    if base.endswith("/models"):
+        return base
+    if base.endswith("/chat/completions"):
+        base = base[: -len("/chat/completions")].rstrip("/")
+    return f"{base}/models"
+
+
+def response_snippet(body):
+    """错误应答常常是一整页 HTML 或带换行的纯文本；压成一行并截断，免得提示框被撑爆。"""
+    return re.sub(r"\s+", " ", str(body or "")).strip()[:160]
+
+
+def upstream_error_message(body, status, who):
+    """把上游错误应答压成一句能看懂的中文。
+
+    各家返回的不是齐步走的：OpenAI 系给 JSON，DeepSeek 缺密钥时回的是纯文本一行
+    “Authentication Fails”，还有网关直接甩一页 HTML。所以先看原文能不能解析成 JSON，
+    解析不了就退化成原文片段，绝不在取错误信息的路上把请求整条打崩。
+    `who` 是“Ollama” / “API” 这一类的来源名，只在连原文都没有时用来兜底。
+    """
+    message = None
+    try:
+        data = json.loads(body)
+    except (TypeError, UnicodeDecodeError, json.JSONDecodeError):
+        data = None
+    if isinstance(data, dict):
+        detail = data.get("error")
+        if isinstance(detail, dict):
+            detail = detail.get("message")
+        message = str(detail).strip() if detail else None
+    hint = "；请检查 API 密钥是否已填写、是否有效" if status in {401, 403} else ""
+    return f"{message or response_snippet(body) or f'{who} 请求失败'}（HTTP {status}{hint}）"
+
+
+def parse_json_body(body, who):
+    """解析成功应答；不是 JSON 或不是字典时给一句能看懂的提示，而不是抛堆栈。"""
+    try:
+        data = json.loads(body)
+    except (TypeError, UnicodeDecodeError, json.JSONDecodeError) as error:
+        raise ValueError(f"{who} 返回的不是 JSON（可能是登录页或错误页）：{response_snippet(body)}") from error
+    if not isinstance(data, dict):
+        raise ValueError(f"{who} 返回的内容无法识别：{response_snippet(body)}")
+    return data
+
+
+def ollama_tags_endpoint(base_url):
+    """Return the local-model list endpoint for an Ollama base URL."""
+    base = str(base_url or "").strip().rstrip("/")
+    if not base:
+        return ""
+    if base.endswith("/v1"):
+        base = base[:-3].rstrip("/")
+    if base.endswith("/api/chat"):
+        base = base[: -len("/api/chat")].rstrip("/")
+    return f"{base}/api/tags"
+
+
 def baidu_translate_params(appid, secret_key, query, to_lang, from_lang="auto", salt=None):
    """组装百度通用文本翻译接口参数；sign = MD5(appid + q + salt + 密钥)。"""
    appid = str(appid or "")
@@ -157,6 +300,30 @@ def baidu_split_query(text, max_chars=600, max_lines=30):
    return chunks


+# 权重的句点只出现在「冒号后的小数」里：(tag:1.2)、[tag:1.2]、<lora:name:0.8>。
+# 断句用的句点则是「词 + 句点 + 空白」，形态完全不同，所以先把权重片段整段挑出来保护。
+WEIGHT_FRAGMENT = re.compile(r"[(\[<][^\n]*?:\s*-?\d+(?:\.\d+)?\s*[)\]>]")
+
+
+def replace_sentence_periods(text):
+    """把权重之外的句点换成逗号。
+
+    模型按规则会用句点把末尾自然语言接在标签后面，但那会让整串标签被前端当成自然语言段。
+    权重片段整段跳过，片段外再要求句点前后都不是数字，双保险。
+    """
+    def to_commas(segment):
+        return re.sub(r"(?<!\d)\s*\.\s*(?!\d)", ", ", segment)
+
+    parts = []
+    cursor = 0
+    for match in WEIGHT_FRAGMENT.finditer(text):
+        parts.append(to_commas(text[cursor:match.start()]))
+        parts.append(match.group(0))
+        cursor = match.end()
+    parts.append(to_commas(text[cursor:]))
+    return "".join(parts)
+
+
 def sanitize_anima_prompt(value):
    text = str(value or "").strip()
    fenced = re.fullmatch(r"```(?:text|plaintext)?\s*([\s\S]*?)\s*```", text, re.IGNORECASE)
@@ -165,17 +332,23 @@ def sanitize_anima_prompt(value):
        "，": ",", "、": ",", "；": ",", ";": ",", "。": ".",
        "！": ".", "!": ".", "？": ".", "?": ".", "：": ":",
        "（": "(", "）": ")", "＜": "<", "＞": ">", "—": "-", "–": "-", "…": ".",
-        "“": "", "”": "", "‘": "'", "’": "'", '"': "", "`": "",
+        "“": "", "”": "", "＂": "", "‘": "'", "’": "'", "＇": "'", '"': "", "`": "",
    }
    for source, target in replacements.items():
        text = text.replace(source, target)
+    # 模型常把结果写成 JSON 或代码字符串，留下 \" \n 这类转义；先把字面换行/制表还原成换行，
+    # 再去掉其余孤立反斜杠，否则一个转义符就会让整条提示词判废。
+    text = re.sub(r"\\+[nrt]", "\n", text)
+    text = text.replace("\\", "")
    text = re.sub(r"(?m)^\s*(?:[-*•]+|\d+[.)])\s*", "", text)
    text = re.sub(r"[\r\n]+", ", ", text)
    text = re.sub(r"\s*,\s*", ", ", text)
+    text = replace_sentence_periods(text)
    text = re.sub(r",(?:\s*,)+", ",", text)
-    text = re.sub(r"(?<!\d)\s*\.\s*(?!\d)", ". ", text)
-    text = re.sub(r",\s*\.\s*", ". ", text)
    text = re.sub(r"\s+", " ", text).strip(" ,")
+    # 模型常在结果末尾多留一个句点。规则里的句点是用来分隔末尾自然语言的，后面没有内容时
+    # 它没有意义，却会让前端把整串标签误判成自然语言段落，所以直接去掉。
+    text = re.sub(r"\s*[.]+\s*$", "", text)
    if re.search(r"[\u3400-\u9fff]", text):
        raise ValueError("助手返回的 Anima 提示词仍包含中文，请调整规则后重试")
    allowed_punctuation = set(",.():@_-'<>%+/&")
@@ -187,6 +360,35 @@ def sanitize_anima_prompt(value):
    return text


+def sanitize_natural_prompt(value):
+    """自然语言优化结果的清洗。
+
+    与 sanitize_anima_prompt 的区别：这里要保留完整句子，所以既不把句点改成逗号，
+    也不做 Anima 标点白名单收窄——只挡代码块、转义符、列表符号和残留中文。
+    """
+    text = str(value or "").strip()
+    fenced = re.fullmatch(r"```(?:text|plaintext)?\s*([\s\S]*?)\s*```", text, re.IGNORECASE)
+    text = fenced.group(1).strip() if fenced else text
+    replacements = {
+        "，": ", ", "、": ", ", "；": "; ", "。": ". ",
+        "！": "! ", "？": "? ", "：": ": ",
+        "（": "(", "）": ")", "＜": "<", "＞": ">", "—": " - ", "–": " - ",
+        "“": '"', "”": '"', "＂": '"', "‘": "'", "’": "'", "＇": "'", "`": "",
+    }
+    for source, target in replacements.items():
+        text = text.replace(source, target)
+    text = re.sub(r"\\+[nrt]", "\n", text)
+    text = text.replace("\\", "")
+    text = re.sub(r"(?m)^\s*(?:[-*•]+|\d+[.)])\s*", "", text)
+    text = re.sub(r"[\r\n]+", " ", text)
+    text = re.sub(r"\s+", " ", text).strip()
+    if re.search(r"[\u3400-\u9fff]", text):
+        raise ValueError("助手返回的自然语言提示词仍包含中文，请调整规则后重试")
+    if not text:
+        raise ValueError("助手返回的自然语言提示词为空")
+    return text
+
+
 def _default_config_dir():
    try:
        import folder_paths
@@ -253,17 +455,6 @@ class AssistantStore:
    def _normalize_base_url(value):
        return str(value or "").strip().rstrip("/")

-    @staticmethod
-    def _is_local_hostname(hostname):
-        host = str(hostname or "").strip().lower().rstrip(".")
-        if host in {"localhost", "host.docker.internal"} or host.endswith(".local") or "." not in host:
-            return True
-        try:
-            address = ip_address(host)
-            return address.is_private or address.is_loopback or address.is_link_local
-        except ValueError:
-            return False
-
    @classmethod
    def _credential_binding(cls, provider, base_url):
        identity = f"{str(provider or '').strip().lower()}\0{cls._normalize_base_url(base_url)}"
@@ -273,7 +464,7 @@ class AssistantStore:
        return {
            **DEFAULT_CONFIG,
            "installation_id": self.installation_id,
-            "ai_credential_binding": "",
+            "openai_credential_binding": "",
            "baidu_credential_binding": "",
        }

@@ -304,6 +495,40 @@ class AssistantStore:
        value["schema_version"] = 3
        return value

+    @classmethod
+    def _migrate_v3(cls, value):
+        """v3（AI 单槽）→ v4（OpenAI / Ollama 两套独立配置）。
+
+        旧的单槽值整份搬到「当前所选后端」那套，另一套留默认；
+        规则原本不分后端，两套都继承，之后可以各自改。
+        """
+        if value.get("schema_version", 0) >= 4:
+            return value
+        provider = str(value.get("ai_provider", "openai_compatible")).strip().lower()
+        prefix = AI_PROVIDER_PREFIXES.get(provider, "openai")
+        for source, target in (
+            ("ai_base_url", "base_url"),
+            ("ai_model", "model"),
+            ("ai_temperature", "temperature"),
+            ("ai_timeout_seconds", "timeout_seconds"),
+        ):
+            if str(value.get(source) or "").strip() != "":
+                value[f"{prefix}_{target}"] = value[source]
+        if prefix == "openai" and str(value.get("ai_api_key") or "").strip():
+            value["openai_api_key"] = value["ai_api_key"]
+            # 绑定的 provider 标记从 openai_compatible 变成 openai，旧值必然对不上；
+            # 直接按新前缀重算，否则密钥会被当成「换了地址」清掉
+            value["openai_credential_binding"] = cls._credential_binding("openai", value.get("openai_base_url", ""))
+        for rule in RULE_KEYS:
+            existing = str(value.get(rule) or "").strip()
+            if existing:
+                for target in ("openai", "ollama"):
+                    value.setdefault(f"{target}_{rule}", existing)
+        for key in ("ai_base_url", "ai_model", "ai_temperature", "ai_timeout_seconds", "ai_api_key", "ai_credential_binding"):
+            value.pop(key, None)
+        value["schema_version"] = 4
+        return value
+
    def config(self):
        value = {}
        should_rewrite = False
@@ -320,20 +545,27 @@ class AssistantStore:
        if value and value.get("schema_version", 0) < 3:
            value = self._migrate_v2(value)
            should_rewrite = True
-        if value.get("translation_rule") in {LEGACY_DEFAULT_TRANSLATION_RULE, V17_DEFAULT_TRANSLATION_RULE}:
-            value["translation_rule"] = DEFAULT_TRANSLATION_RULE
+        if value and value.get("schema_version", 0) < 4:
+            value = self._migrate_v3(value)
            should_rewrite = True
-        if value.get("optimization_rule") == LEGACY_DEFAULT_OPTIMIZATION_RULE:
-            value["optimization_rule"] = DEFAULT_OPTIMIZATION_RULE
-            should_rewrite = True
-        if "translate_optimize_rule" not in value:
-            value["translate_optimize_rule"] = DEFAULT_TRANSLATE_OPTIMIZE_RULE
+        # 旧值 "ai" 已拆成两个平级服务：按当时的后端落到 openai_compatible / ollama
+        if value and value.get("translate_service") == "ai":
+            provider = str(value.get("ai_provider") or "openai_compatible").strip().lower()
+            value["translate_service"] = provider if provider in AI_SERVICES else "openai_compatible"
            should_rewrite = True
+        # 旧版默认规则已废弃：两套各查一次，命中就换成现在的默认
+        for prefix in ("openai", "ollama"):
+            if value.get(f"{prefix}_translation_rule") in {LEGACY_DEFAULT_TRANSLATION_RULE, V17_DEFAULT_TRANSLATION_RULE}:
+                value[f"{prefix}_translation_rule"] = DEFAULT_TRANSLATION_RULE
+                should_rewrite = True
+            if value.get(f"{prefix}_optimization_rule") == LEGACY_DEFAULT_OPTIMIZATION_RULE:
+                value[f"{prefix}_optimization_rule"] = DEFAULT_OPTIMIZATION_RULE
+                should_rewrite = True
        result = {**self._default_persisted_config(), **value, "installation_id": self.installation_id}
-        # 密钥只在「绑定的后端/地址变了」或「换机器」时清，切翻译服务绝不清
-        if result.get("ai_api_key") and result.get("ai_credential_binding") != self._credential_binding(result["ai_provider"], result["ai_base_url"]):
-            result["ai_api_key"] = ""
-            result["ai_credential_binding"] = ""
+        # 密钥只在「地址变了」或「换机器」时清，切翻译服务 / 切后端绝不清
+        if result.get("openai_api_key") and result.get("openai_credential_binding") != self._credential_binding("openai", result["openai_base_url"]):
+            result["openai_api_key"] = ""
+            result["openai_credential_binding"] = ""
            should_rewrite = True
        if result.get("baidu_secret_key") and result.get("baidu_credential_binding") != self._credential_binding("baidu", ""):
            result["baidu_secret_key"] = ""
@@ -351,15 +583,18 @@ class AssistantStore:

    def public_config(self):
        config = self.config()
+        defaults = {}
+        for prefix in ("openai", "ollama"):
+            defaults[f"default_{prefix}_translation_rule"] = DEFAULT_TRANSLATION_RULE
+            defaults[f"default_{prefix}_translate_optimize_rule"] = DEFAULT_TRANSLATE_OPTIMIZE_RULE
+            defaults[f"default_{prefix}_optimization_rule"] = DEFAULT_OPTIMIZATION_RULE
+            defaults[f"default_{prefix}_natural_rule"] = DEFAULT_NATURAL_RULE
        return {
-            key: config[key] for key in DEFAULT_CONFIG if key not in ("ai_api_key", "baidu_secret_key")
+            key: config[key] for key in DEFAULT_CONFIG if key not in ("openai_api_key", "baidu_secret_key")
        } | {
-            "ai_api_key_configured": bool(config.get("ai_api_key")),
+            "openai_api_key_configured": bool(config.get("openai_api_key")),
            "baidu_secret_key_configured": bool(config.get("baidu_secret_key")),
-            "default_translation_rule": DEFAULT_TRANSLATION_RULE,
-            "default_translate_optimize_rule": DEFAULT_TRANSLATE_OPTIMIZE_RULE,
-            "default_optimization_rule": DEFAULT_OPTIMIZATION_RULE,
-        }
+        } | defaults

    @staticmethod
    def _clean_rule(value, fallback):
@@ -375,45 +610,46 @@ class AssistantStore:
            raise ValueError("设置必须是对象")
        current = self.config()
        translate_service = str(payload.get("translate_service", current["translate_service"])).strip().lower()
-        if translate_service not in {"dictionary", "baidu", "ai"}:
-            raise ValueError("翻译服务仅支持词库、百度翻译或 AI")
+        if translate_service not in TRANSLATE_SERVICES:
+            raise ValueError("翻译服务仅支持词库、百度翻译、OpenAI 兼容接口或 Ollama")
        ai_provider = str(payload.get("ai_provider", current["ai_provider"])).strip().lower()
        if ai_provider not in {"openai_compatible", "ollama"}:
            raise ValueError("AI 后端仅支持 OpenAI 兼容接口或 Ollama")
-        ai_base_url = self._normalize_base_url(payload.get("ai_base_url", current["ai_base_url"]))
-        if ai_base_url:
-            parsed = urlparse(ai_base_url)
-            if parsed.scheme not in {"http", "https"} or not parsed.netloc:
-                raise ValueError("API 地址必须是完整的 http:// 或 https:// 地址")
-            if parsed.username or parsed.password:
-                raise ValueError("API 地址中不能包含用户名或密码")
-            if parsed.query or parsed.fragment:
-                raise ValueError("API 地址中不能包含查询参数或片段")
-            if parsed.scheme == "http" and not self._is_local_hostname(parsed.hostname):
-                raise ValueError("非本机 API 地址必须使用 HTTPS，避免 API Key 明文传输")
-        ai_model = str(payload.get("ai_model", current["ai_model"])).strip()
-        if len(ai_model) > 300:
-            raise ValueError("模型名称过长")
-        try:
-            ai_temperature = float(payload.get("ai_temperature", current["ai_temperature"]))
-            ai_timeout_seconds = int(payload.get("ai_timeout_seconds", current["ai_timeout_seconds"]))
-        except (TypeError, ValueError) as error:
-            raise ValueError("温度和超时时间必须是数字") from error
-        if not 0 <= ai_temperature <= 2:
-            raise ValueError("温度必须在 0 到 2 之间")
-        if not 5 <= ai_timeout_seconds <= 600:
-            raise ValueError("超时时间必须在 5 到 600 秒之间")
-        # AI key：只在「AI 后端/地址变了」或显式清除时清，切翻译服务绝不清
-        ai_endpoint_changed = ai_provider != current["ai_provider"] or ai_base_url != self._normalize_base_url(current["ai_base_url"])
-        ai_api_key = "" if ai_endpoint_changed else current.get("ai_api_key", "")
-        if payload.get("clear_ai_api_key"):
-            ai_api_key = ""
-        elif "ai_api_key" in payload:
-            incoming_key = str(payload.get("ai_api_key") or "").strip()
+        # 选了 OpenAI / Ollama 当翻译服务时，它就是 AI 后端，两者不允许脱节
+        if translate_service in AI_SERVICES:
+            ai_provider = translate_service
+        # 两个后端各一套：地址规则不同（Ollama 只认本机，OpenAI 远端必须 https）
+        openai_base_url = normalize_ai_base_url(payload.get("openai_base_url", current["openai_base_url"]), allow_remote=True)
+        ollama_base_url = normalize_ai_base_url(payload.get("ollama_base_url", current["ollama_base_url"]))
+        if not ollama_base_url:
+            ollama_base_url = DEFAULT_CONFIG["ollama_base_url"]
+        models = {}
+        numbers = {}
+        for prefix in ("openai", "ollama"):
+            models[prefix] = str(payload.get(f"{prefix}_model", current[f"{prefix}_model"])).strip()
+            if len(models[prefix]) > 300:
+                raise ValueError("模型名称过长")
+            try:
+                numbers[prefix] = (
+                    float(payload.get(f"{prefix}_temperature", current[f"{prefix}_temperature"])),
+                    int(payload.get(f"{prefix}_timeout_seconds", current[f"{prefix}_timeout_seconds"])),
+                )
+            except (TypeError, ValueError) as error:
+                raise ValueError("温度和超时时间必须是数字") from error
+            if not 0 <= numbers[prefix][0] <= 2:
+                raise ValueError("温度必须在 0 到 2 之间")
+            if not 5 <= numbers[prefix][1] <= 600:
+                raise ValueError("超时时间必须在 5 到 600 秒之间")
+        # OpenAI key：只在「地址变了」或显式清除时清，切翻译服务 / 切后端绝不清
+        openai_api_key = "" if openai_base_url != self._normalize_base_url(current["openai_base_url"]) else current.get("openai_api_key", "")
+        if payload.get("clear_openai_api_key"):
+            openai_api_key = ""
+        elif "openai_api_key" in payload:
+            incoming_key = str(payload.get("openai_api_key") or "").strip()
            if incoming_key and incoming_key != "••••••••":
                if len(incoming_key) > 4096:
                    raise ValueError("API Key 过长")
-                ai_api_key = incoming_key
+                openai_api_key = incoming_key
        baidu_appid = str(payload.get("baidu_appid", current.get("baidu_appid", ""))).strip()
        if len(baidu_appid) > 128:
            raise ValueError("百度 APP ID 过长")
@@ -428,7 +664,7 @@ class AssistantStore:
                    raise ValueError("百度密钥过长")
                baidu_secret = incoming_secret
        baidu_qps = normalize_baidu_qps(payload.get("baidu_qps", current.get("baidu_qps", BAIDU_QPS_DEFAULT)))
-        ai_credential_binding = self._credential_binding(ai_provider, ai_base_url) if ai_api_key else ""
+        openai_credential_binding = self._credential_binding("openai", openai_base_url) if openai_api_key else ""
        baidu_credential_binding = self._credential_binding("baidu", "") if baidu_secret else ""
        appearance = str(payload.get("appearance", current.get("appearance", "auto"))).strip().lower()
        if appearance not in {"auto", "dark", "light"}:
@@ -437,37 +673,34 @@ class AssistantStore:
        if language not in {"auto", "zh", "en"}:
            raise ValueError("界面语言仅支持 auto / zh / en")
        result = {
-            "schema_version": 3,
+            "schema_version": 4,
            "translate_service": translate_service,
            "ai_provider": ai_provider,
-            "ai_base_url": ai_base_url,
-            "ai_model": ai_model,
-            "ai_api_key": ai_api_key,
-            "ai_temperature": ai_temperature,
-            "ai_timeout_seconds": ai_timeout_seconds,
+            "openai_base_url": openai_base_url,
+            "openai_model": models["openai"],
+            "openai_api_key": openai_api_key,
+            "openai_temperature": numbers["openai"][0],
+            "openai_timeout_seconds": numbers["openai"][1],
+            "ollama_base_url": ollama_base_url,
+            "ollama_model": models["ollama"],
+            "ollama_temperature": numbers["ollama"][0],
+            "ollama_timeout_seconds": numbers["ollama"][1],
            "baidu_appid": baidu_appid,
            "baidu_secret_key": baidu_secret,
            "baidu_qps": baidu_qps,
            "appearance": appearance,
            "language": language,
-            "translation_rule": self._clean_rule(payload.get("translation_rule"), current["translation_rule"]),
-            "translate_optimize_rule": self._clean_rule(payload.get("translate_optimize_rule"), current["translate_optimize_rule"]),
-            "optimization_rule": self._clean_rule(payload.get("optimization_rule"), current["optimization_rule"]),
-            "installation_id": self.installation_id,
-            "ai_credential_binding": ai_credential_binding,
            "baidu_credential_binding": baidu_credential_binding,
+            "openai_credential_binding": openai_credential_binding,
+            "installation_id": self.installation_id,
        }
+        # 规则两套，各取各自的当前值兜底
+        for prefix in ("openai", "ollama"):
+            for rule in RULE_KEYS:
+                result[f"{prefix}_{rule}"] = self._clean_rule(payload.get(f"{prefix}_{rule}"), current[f"{prefix}_{rule}"])
        self._write_json_atomic(self.config_path, result)
        return self.public_config()

-    def reset_rules(self):
-        config = self.config()
-        config["translation_rule"] = DEFAULT_TRANSLATION_RULE
-        config["translate_optimize_rule"] = DEFAULT_TRANSLATE_OPTIMIZE_RULE
-        config["optimization_rule"] = DEFAULT_OPTIMIZATION_RULE
-        self._write_json_atomic(self.config_path, config)
-        return self.public_config()
-

 def dictionary_translate(text, dictionary_store):
    parts = [part.strip() for part in re.split(r"[,，、\r\n]+", str(text or "")) if part.strip()]
@@ -488,7 +721,8 @@ def dictionary_translate(text, dictionary_store):
            large = dictionary_store.search_large_tags(lookup_part, limit=30)
            exact = [tag for tag in large if lookup_part == tag.get("chinese")]
        if exact:
-            exact.sort(key=lambda tag: (tag.get("pack_id") == "personal", int(tag.get("post_count", 0))), reverse=True)
+            # 个人词库与自然语言词库都算本机自建，优先于内置词库（与 dictionary_store.USER_PACK_IDS 一致）
+            exact.sort(key=lambda tag: (tag.get("pack_id") in ("personal", "natural"), int(tag.get("post_count", 0))), reverse=True)
            english = exact[0]["english"]
            output.append(f"({english}:{weighted.group(2)})" if weighted else english)
        else:
@@ -22,5 +22,18 @@
    "#5b7a8a",
    "#8a8a5b",
    "#5b8a5b"
-  ]
+  ],
+  "pack_colors": {
+    "base": "#8a6b5b",
+    "anima": "#7a8a5b",
+    "characters": "#5b8a72",
+    "poses": "#8a5b7a",
+    "camera": "#8a725b",
+    "clothing": "#5b8a8a",
+    "adult": "#5b8a5b",
+    "personal": "#8a8a5b",
+    "natural": "#5b8a72",
+    "unindexed": "#5b6e8a"
+  },
+  "anima_colors": {}
 }
@@ -12,6 +12,11 @@ from pathlib import Path

 SCHEMA_VERSION = 1
 PACK_SCHEMA_VERSION = 1
+# 个人词库与自然语言词库共用 user_tags.json，靠 natural 标记分成两个 pack，
+# 这样按词库筛选、配色都能把它们分开处理。
+PERSONAL_PACK_ID = "personal"
+NATURAL_PACK_ID = "natural"
+USER_PACK_IDS = (PERSONAL_PACK_ID, NATURAL_PACK_ID)
 MAX_TEXT_LENGTH = 500
 MAX_ALIASES = 30
 MAX_LARGE_LOOKUP_TERMS = 200
@@ -101,6 +106,8 @@ def validate_tag(value, *, default_source="user"):
        "notes": notes,
        "source": source,
        "verified": bool(value.get("verified", default_source == "user")),
+        # 自然语言词库：与个人词库共用一个文件，靠这个标记区分来源
+        "natural": bool(value.get("natural", False)),
    }


@@ -246,7 +253,13 @@ class DictionaryStore:
        user = self.user_tags()
        effective = {normalize_key(tag["english"]): tag for tag in builtin}
        for tag in user:
-            clean = {**tag, "pack_id": "personal", "pack_name": "个人词库", "pack_source": "本机用户"}
+            natural = bool(tag.get("natural"))
+            clean = {
+                **tag,
+                "pack_id": NATURAL_PACK_ID if natural else PERSONAL_PACK_ID,
+                "pack_name": "自然语言词库" if natural else "个人词库",
+                "pack_source": "本机用户",
+            }
            effective[normalize_key(tag["english"])] = clean
        return {
            "schema_version": SCHEMA_VERSION,
@@ -33,11 +33,13 @@
 - `server.py`
  - 注册助手、个人词库、词库包和大型词库的本地接口。
  - 执行同源检查、临时会话令牌校验、请求体限制、助手频率限制和并发限制。
-  - 只在用户主动调用时访问配置的百度翻译或 AI（OpenAI 兼容 / Ollama）服务；"翻译并优化"与"优化为 Anima"恒走 AI，翻译/解释按 `translate_service` 选择。
+  - 只在用户主动调用时访问配置的百度翻译或 AI（OpenAI 兼容 / Ollama）服务；"翻译并优化"与"优化"恒走 AI，翻译/解释按 `translate_service` 选择。
  - `UpstreamGate` 负责上游输入的挂起：按节点 id 保存一个 future，只能被 `/bpi/upstream/{ack,resume,cancel}` 唤醒。等待循环每 0.5 秒检查一次 ComfyUI 中断，无人接管时到点自动透传。唤醒一律走 `loop.call_soon_threadsafe`——future 属于执行线程的 loop，而路由跑在主线程。
 - `assistant_store.py`
-  - 保存翻译服务类型（schema v3：`translate_service` = dictionary / baidu / ai）、AI 后端与地址模型、百度 APP ID 与密钥、三套规则和独立留存的 AI Key 与百度密钥。
-  - AI Key 与百度密钥互不干扰：切换翻译服务不清任一密钥；AI Key 只在 AI 后端/地址变更或换机器时清，百度密钥只在显式清除时清。
+  - 保存翻译服务类型（schema v4：`translate_service` = dictionary / baidu / openai_compatible / ollama；后两个同时是 AI 后端，选前两个时由 `ai_provider` 决定优化走哪一套）、两套 AI 配置（OpenAI 兼容 / Ollama 各一套地址、模型、温度、超时与四条规则）、百度 APP ID 与密钥，以及独立留存的 AI Key 与百度密钥。
+  - 四条规则是翻译 / 翻译并优化 / 优化 / 自然语言。前三条产出 Anima 标签流，走 `sanitize_anima_prompt`（句点收成逗号、标点只放行白名单）；自然语言那条产出完整英文句子，走 `sanitize_natural_prompt`，两者不能混用。
+  - 读到旧值 `translate_service = "ai"` 时按当时的 `ai_provider` 拆成 `openai_compatible` / `ollama` 并回写。
+  - AI Key 与百度密钥互不干扰：切换翻译服务或 AI 后端不清任一密钥；AI Key 只在地址变更或换机器时清，百度密钥只在显式清除时清。
  - 旧 v2 配置（单一 `provider`）自动迁移至 v3，密钥与绑定原样搬入 `ai_*` 槽。
  - API Key 不通过读取接口返回前端；公开配置只返回 `ai_api_key_configured` 与 `baidu_secret_key_configured` 布尔值。
  - 负责翻译方向判断、纯词库翻译和 Anima 输出标点清理。
@@ -58,8 +60,11 @@
  - `openSavePromptDialog` 用 multipart 一次提交名称、备注与图片文件，正文只读展示，保存成功后回调刷新侧边栏。
  - `openManagerPanel` 通过 ComfyUI 侧边栏 API 打开管理面板，可指定定位分页与目标节点。
 - `js/manager_panel.js`
-  - 在 ComfyUI 侧边栏注册「双语提示词管理」面板，包含标签管理、词库、词库包、助手设置、收藏五个分页。
-  - 承担原节点内管理弹窗的全部职责：词条增删改、批量操作、个人词库导入导出、社区包与大型词库管理、助手服务与 API Key 配置。
+  - 在 ComfyUI 侧边栏注册「双语提示词管理」面板，包含标签管理、词库、词库包、收藏四个分页。
+  - 承担原节点内管理弹窗的全部职责：词条增删改、批量操作、个人词库导入导出、社区包与大型词库管理。
+  - 助手设置不在侧边栏：同一文件的 `registerExtension` 用 `settings` 在 ComfyUI「设置」面板的「双语提示词检查器」分类下注册两项 ——「语言与外观」（`appearanceControl()`）与「助手设置」（`assistantSettingsControl()`），表单都直接内嵌在设置面板右侧，配置仍写入 `assistant_store.py`。面板反复渲染会重建控件容器，所以每项的整块 DOM 只建一次后复用，避免输入到一半被清空。
+  - 语言与外观只提交 `appearance` / `language` 两个字段：`assistant_store.update()` 对 payload 里没给的字段一律取当前值，所以不会动翻译服务与密钥；助手设置的 payload 反过来已经不含这两项。
+  - AI 两个后端各一整套配置（地址 / 模型 / 温度 / 超时 / 三条规则），键名前缀 `openai_` 与 `ollama_`，`ai_provider` 只决定当前读哪一套；切换后端不清空任何一套。地址白名单在 `assistant_store.normalize_ai_base_url()`：OpenAI 远端必须 https、http 仅限本机回环，Ollama 只认本机回环，且 `server.py` 发请求前会再校验一次（防手改配置文件）。
  - “标签管理”分页通过节点下拉列表挂载指定节点的标签翻译与词库搜索界面（DOM 直接移植，逻辑仍归属各节点闭包），节点删除时自动摘下。
  - 修改后通过 `panelSyncHub` 通知当前页面所有节点刷新。
  - “收藏”分页按需挂载并监听 `bpi:saved-prompts-changed`：卡片支持搜索、复制、删除、导出导入；「新建」没有选中节点时弹出节点选择器；「载入节点」先确认再调目标节点的 `_bilingualPromptInspector.setText`，覆盖不进 undo 栈。
@@ -68,7 +73,7 @@
  - 标签排序是结构化编辑：英文标签区支持指针拖拽（4px 阈值区分点击与拖动，拖动后吞掉 click/dblclick），明细表行首提供 `⠿` 手柄（带插入指示线与边缘自动滚动），`Alt+↑/↓` 键盘微调；三个入口都汇入同一个 `moveTokenTo` → `movePromptToken` → undo 栈管线。
  - 标签隐藏（临时排除出实际输出）只是“打标记”：标签留在文本里，卡片与明细行加删除线并灰化褪色，拖动/双击改权重/改色照常；剔除发生在输出时——`syncEffectiveOutput` 把去掉隐藏标签的文本同步进隐藏 widget `effective_text`（并置 `use_effective`），`nodes.py` 按它输出，旧工作流没有这两个 widget 时默认直通。明细表行内眼睛图标、英文标签 hover 浮出的眼睛角标在隐藏后变为眼睛斜线（取消隐藏）；明细表下方与英文标签区下方的“已隐藏”条负责逐个/全部恢复；清单存在 `node.properties.bpiHiddenTags`（随工作流保存），每次编辑把清单快照写进 undo 条目，撤销时文本与清单一起回滚。旧版工作流里当时被物理删除的隐藏标签，恢复时仍按旧锚点插回文本。
  - 标签翻译与词库搜索（标签管理）不再渲染进节点，其 DOM 由侧边栏“标签管理”分页挂载，节点高度恒为折叠尺寸。
-  - 管理功能收敛为「标签管理」「助手设置」等入口按钮，打开侧边栏面板对应分页并定位到本节点。
+  - 管理功能收敛为「标签管理」等入口按钮，打开侧边栏面板对应分页并定位到本节点。
  - 上游输入条（`.bpi-source-bar`）只在 `prompt` 端口真的连上时才渲染，未连线时零高度、界面与旧版一致。条内含接管开关、导入策略下拉和状态文案，等待确认时额外出现「继续运行」「放弃」。开关与策略存在 `node.properties.bpiUpstream`（随工作流保存）；放行时把节点当前文本 POST 回后端，因此输出始终等于放行那一刻节点上的内容。
 - `js/parser.js`
  - 非破坏性切分标签，保留原始字符位置。
@@ -113,7 +118,8 @@ ComfyUI 原生 text 控件（工作流真值）

 ```text
 绿色草稿框
-  → 用户点击仅翻译 / 翻译并优化 / 优化为 Anima
+  → 用户点击仅翻译 / 翻译并优化 / 优化
+  → 自然语言模式下「优化」改用 optimize_natural，走自然语言规则
  → /bpi/assistant/run
  → 纯词库、百度翻译或 AI（OpenAI 兼容 / Ollama）
  → 结果写回绿色暂存区
@@ -16,9 +16,9 @@

 ## 3. 翻译、优化和同步必须分阶段

-决定：设置“仅翻译”“翻译并优化”“优化为 Anima”“同步到英文输出”四个独立动作。
+决定：设置“仅翻译”“翻译并优化”“优化”“同步到英文输出”四个独立动作。

-原因：翻译与提示词创作不是同一任务。用户需要看见模型做了什么，并决定是否覆盖实际输出。每个动作使用独立规则，避免一个通用提示词同时承担忠实翻译、扩写和格式整理而行为混乱。
+原因：翻译与提示词创作不是同一任务。用户需要看见模型做了什么，并决定是否覆盖实际输出。每个动作使用独立规则，避免一个通用提示词同时承担忠实翻译、扩写和格式整理而行为混乱。“优化”会按当前输入模式分派不同规则，见第 19 条。

 ## 4. Anima 排序只在按下按钮后执行

@@ -116,3 +116,25 @@
 - GitHub 首个公开正式 Release 使用 `v1.0.0`。
 - 后续发布应使用语义化版本号，不要把内部演进代次和公开版本号混为一谈。
 - 兼容修复递增补丁号；向后兼容的新功能递增次版本；破坏工作流或数据格式时才递增主版本，并提供迁移说明。
+
+## 19. “优化”按输入模式分派，而不是再加一个按钮
+
+决定：节点顶部只保留一个“优化”按钮。标签模式下走 Anima 优化规则（只接受英文），自然语言模式下走独立的自然语言规则（中文先译后整理）；模式切到“自然语言”时“按 Anima 顺序排序”隐藏。
+
+原因：Krea、FLUX 一类新模型直接吃自然语言，再加一个按钮只会让顶部操作越排越长。模式下拉已经能判定当前是标签流还是整段句子，用它分派比让用户记两个按钮省事；排序只重排标签流，对整段句子没有意义，留着反而误导。
+
+两套规则必须分开：`sanitize_anima_prompt` 会把句点收成逗号并只放行 Anima 标点白名单，完整句子喂进去会被判废，所以自然语言结果改走 `sanitize_natural_prompt`。
+
+代价：已经排过序时要保留排序按钮，否则“取消排序”的入口会跟着一起消失。另外 mixed（标签 + 自然语言）不算自然语言模式——里面还有标签流可排。
+
+## 20. 模型列表由服务端代问，且只在本次会话内缓存
+
+决定：助手设置的「模型名称」输入框保持自由输入，旁边另加一个「获取模型」按钮。按钮把输入框里当前填的地址和密钥发给插件自己的 `/bpi/assistant/models`，由服务端去问那个 AI 后端（OpenAI 兼容问 `/models`，Ollama 问 `/api/tags`），把名字填进输入框自带的原生下拉（`<datalist>`）；按钮显示「获取模型（N）」，再点一次就是重新拉取。
+
+不弹窗是有原因的：助手设置长在 ComfyUI 的「设置」面板里，任何浮层都会被这块面板压住、点不动，逼着用户先关掉设置面板。下拉是浏览器原生的输入框附件，既没有层级问题，也不用自己处理定位与溢出。
+
+原因：地址白名单（远端必须 https、内网与链路本地地址驳回、Ollama 只认本机回环）是整个插件的安全边界。让浏览器直接去拉，这份校验就得在前端重写一遍，而且可以绕过；交给服务端代发，还能顺带解决输入框里那份还没保存的密钥往哪发：填了就用填的，留空就用已存的那份，都只用于这一次请求，返回给浏览器的始终只有模型名字。
+
+列表不持久化有两点考虑：各家网关的模型随时增减，写进配置文件第二天就是过期数据；地址一改缓存必须作废，把「这次问的地址」和列表绑在同一个闭包里最省事，也不用为它设计一份需要清理的存储格式。
+
+代价：第一次点按钮要多等一次 `/models` 往返；各家返回结构不一样（`data[].id` 与 `models[].name`），得分别解析。面板一律不去猜哪个模型能聊天——embedding、rerank 之类照常列出，免得误伤名字奇怪但确实可用的模型。
@@ -0,0 +1,173 @@
+"""重复检查：切分、归一、去重都在这里，节点执行与前端实时提示共用同一份口径。
+
+规则与前端 js/parser.js 保持一致（splitPrompt / analyzeSegment / normalizeKey），
+改任一侧都要同步另一侧，否则节点上显示的重复数会和实际去掉的数量对不上。
+"""
+
+import re
+
+# 段与段之间的间隙里只取第一个分隔符本身：段后的空白归下一段自己带（raw 里有前导空白），
+# 两边都管就会在删掉中间段时丢掉空格（「standing, sitting」变成「standing,sitting」）。
+_GAP_SEPARATOR = re.compile(r"^[ \t]*([,\r\n])")
+# 间隙里只有分隔符和空白（比如两个标签之间隔了一个空行）时原样保留，
+# 否则说明中间夹着被删掉的段，那点空白是被删段的前导，不能再计入。
+_GAP_BLANK = re.compile(r"^[ \t]*[,\r\n][ \t\r\n]*$")
+_SPECIAL_SEGMENT = re.compile(r"^<[^>]+>$")
+_WEIGHTED = re.compile(r"^\((.+):\s*(-?(?:\d+(?:\.\d*)?|\.\d+))\)$", re.DOTALL)
+_WHITESPACE = re.compile(r"\s+")
+
+
+def strip_emphasis(value):
+    current = str(value or "").strip()
+    while len(current) >= 2 and _is_pair(current[0], current[-1]):
+        current = current[1:-1].strip()
+    return current
+
+
+def _is_pair(open_char, close_char):
+    return (
+        (open_char == "(" and close_char == ")")
+        or (open_char == "[" and close_char == "]")
+        or (open_char == "{" and close_char == "}")
+    )
+
+
+def normalize_key(value):
+    return _WHITESPACE.sub(" ", str(value or "").strip().lower().replace("_", " "))
+
+
+def split_prompt(text):
+    """按顶层逗号与换行切分；括号里、尖括号里、转义字符后面的分隔符都不算切分点。"""
+    value = str(text or "")
+    segments = []
+    start = 0
+    round_depth = 0
+    square_depth = 0
+    curly_depth = 0
+    angle_depth = 0
+    escaped = False
+    for index, char in enumerate(value):
+        if escaped:
+            escaped = False
+            continue
+        if char == "\\":
+            escaped = True
+            continue
+        if char == "<":
+            angle_depth += 1
+        elif char == ">" and angle_depth > 0:
+            angle_depth -= 1
+        elif angle_depth == 0:
+            if char == "(":
+                round_depth += 1
+            elif char == ")" and round_depth > 0:
+                round_depth -= 1
+            elif char == "[":
+                square_depth += 1
+            elif char == "]" and square_depth > 0:
+                square_depth -= 1
+            elif char == "{":
+                curly_depth += 1
+            elif char == "}" and curly_depth > 0:
+                curly_depth -= 1
+        if (
+            round_depth == 0
+            and square_depth == 0
+            and curly_depth == 0
+            and angle_depth == 0
+            and char in (",", "\n", "\r")
+        ):
+            raw = value[start:index]
+            if raw.strip():
+                segments.append({"raw": raw, "start": start, "end": index})
+            start = index + 1
+    tail = value[start:]
+    if tail.strip():
+        segments.append({"raw": tail, "start": start, "end": len(value)})
+    return segments
+
+
+def analyze_segment(raw):
+    """剥掉权重与强调括号，得到用来判断重复的词条；纯特殊语法（<lora:...>）原样保留。"""
+    display = str(raw or "").strip()
+    if not display:
+        return None
+    if _SPECIAL_SEGMENT.match(display):
+        return {"term": display, "syntax": "special"}
+    term = display
+    weighted = _WEIGHTED.match(display)
+    if weighted:
+        term = weighted.group(1).strip()
+    term = strip_emphasis(term)
+    return {
+        "term": term,
+        "syntax": "operator" if term.upper() in ("BREAK", "AND") else "tag",
+    }
+
+
+def find_duplicates(text):
+    """返回重复报告：多余出现几处、重复词条、以及重复段的下标（保留第一次出现）。"""
+    value = str(text or "")
+    segments = split_prompt(value)
+    seen = {}
+    order = []
+    drop = []
+    for index, segment in enumerate(segments):
+        info = analyze_segment(segment["raw"])
+        if not info or info["syntax"] != "tag":
+            continue
+        key = normalize_key(info["term"])
+        if not key:
+            continue
+        if key in seen:
+            seen[key]["count"] += 1
+            drop.append(index)
+        else:
+            seen[key] = {"term": info["term"], "count": 1}
+            order.append(key)
+    return {
+        "extra": len(drop),
+        "terms": [seen[key] for key in order if seen[key]["count"] > 1],
+        "drop": drop,
+        "segments": segments,
+    }
+
+
+def remove_segments(text, segments, drop):
+    """删掉指定段并把剩下的段按原文的连接符接回去，不重写整段提示词的格式。"""
+    value = str(text or "")
+    doomed = set(drop or ())
+    keep = [index for index in range(len(segments)) if index not in doomed]
+    if not keep:
+        return ""
+    parts = []
+    previous_end = None
+    for index in keep:
+        segment = segments[index]
+        if previous_end is None:
+            parts.append(segment["raw"].strip())
+        else:
+            gap = value[previous_end:segment["start"]]
+            if _GAP_BLANK.match(gap):
+                # 只去掉首尾的空格，换行要留着（段落之间的空行是有意义的）
+                parts.append(gap.strip(" \t") or ",")
+            else:
+                matched = _GAP_SEPARATOR.match(gap)
+                parts.append(matched.group(1) if matched else ",")
+            parts.append(segment["raw"])
+        previous_end = segment["end"]
+    return "".join(parts)
+
+
+def dedupe_prompt(text):
+    """只保留第一次出现的标签，返回去重后的文本。"""
+    value = str(text or "")
+    report = find_duplicates(value)
+    if not report["drop"]:
+        return value
+    return remove_segments(value, report["segments"], report["drop"])
+
+
+def format_duplicates(report):
+    """重复项预览：一行一个「词条 ×次数」，没有重复时给空串。"""
+    return "\n".join(f"{item['term']} ×{item['count']}" for item in report["terms"])
@@ -1,6 +1,7 @@
 import { app } from "../../scripts/app.js";
 import { normalizePreferences } from "./dictionary_tools.js";
 import { matchModelNames } from "./model_suggest.js";
+import { ANIMA_SLOTS } from "./anima_sorter.js";
 import {
  LANGUAGE_CHANGED_EVENT,
  applyLanguageToDom,
@@ -23,6 +24,14 @@ const MANAGER_TAB_ID = "bpi-manager";
 const MANAGER_OPEN_EVENT = "bpi:open-manager";
 // 后端把上游传来的提示词推给前端时使用的事件名
 const UPSTREAM_ARRIVED_EVENT = "bpi/upstream-arrived";
+// 按词库配色时，「没命中任何词库」也当成一个来源，与后端 token_color_store.UNINDEXED_PACK_ID 对齐
+const UNINDEXED_PACK_ID = "unindexed";
+// 骰子：点一下从 random_pool 里随机抽一个颜色
+const DICE_ICON = '<svg viewBox="0 0 14 14" width="13" height="13" aria-hidden="true">'
+  + '<rect x="1" y="1" width="12" height="12" rx="3" fill="none" stroke="currentColor" stroke-width="1.2"/>'
+  + '<circle cx="4.5" cy="4.5" r="1.15" fill="currentColor"/><circle cx="9.5" cy="4.5" r="1.15" fill="currentColor"/>'
+  + '<circle cx="7" cy="7" r="1.15" fill="currentColor"/>'
+  + '<circle cx="4.5" cy="9.5" r="1.15" fill="currentColor"/><circle cx="9.5" cy="9.5" r="1.15" fill="currentColor"/></svg>';

 let sessionTokenPromise = null;
 let dictionaryPromise = null;
@@ -75,6 +84,17 @@ async function postUpstreamAction(action, nodeId, text = null) {
  return result;
 }

+async function checkDuplicateTags(text) {
+  const response = await bpiFetch(`${API_ROOT}/duplicate-check`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ text: String(text ?? "") }),
+  });
+  const result = await response.json().catch(() => null);
+  if (!response.ok || !result?.success) throw new Error(result?.error || `Duplicate check failed (${response.status})`);
+  return result.data;
+}
+
 function loadPreferences() {
  try {
    return normalizePreferences(JSON.parse(localStorage.getItem(PREFERENCES_KEY) || "{}"));
@@ -179,6 +199,18 @@ async function getAssistantConfig() {
  return result.data;
 }

+// 拿表单当前填的地址和密钥问一次「你有哪些模型」；后端不读已保存的密钥，也不写配置
+async function listAssistantModels(payload) {
+  const response = await bpiFetch(`${API_ROOT}/assistant/models`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(payload),
+  });
+  const result = await response.json().catch(() => null);
+  if (!response.ok || !result?.success) throw new Error(result?.error || "Failed to fetch models");
+  return Array.isArray(result.data?.models) ? result.data.models : [];
+}
+
 // 读 data/token_colors.json；用户可以不改源码只改这个文件管理节点标签卡的颜色池。
 // 文件不在或格式不对时后端回退内置默认，这里只管透传。
 let tokenColorsPromise = null;
@@ -198,6 +230,20 @@ function loadTokenColors(force = false) {
  return tokenColorsPromise;
 }

+// 只传要更新的那一类；没传的那类后端保持原样
+async function saveTokenColors({ pack_colors: packColors, anima_colors: animaColors } = {}) {
+  const response = await bpiFetch(`${API_ROOT}/token-colors/colors`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ pack_colors: packColors, anima_colors: animaColors }),
+  });
+  const result = await response.json();
+  if (!response.ok || !result.success) throw new Error(result.error || "Failed to save colors");
+  // 缓存失效：保存完节点那边要重新拉一次才能看到新配色
+  tokenColorsPromise = null;
+  return result.data;
+}
+
 async function saveAssistantConfig(payload) {
  const response = await bpiFetch(`${API_ROOT}/assistant/config`, {
    method: "POST",
@@ -689,24 +735,243 @@ function openSavePromptDialog({ name = "", text = "", note = null, models = [],
  return { close };
 }

+// 分类候选：只收词库里真实出现过的 category，用得多的排前面。
+// 不预置任何「应该有但还没用过」的分类；中英文混着也不翻译，原样列出。
+function tagCategoryCounts(data) {
+  const counts = new Map();
+  for (const tag of data?.tags ?? []) {
+    const name = String(tag?.category ?? "").trim();
+    if (!name) continue;
+    counts.set(name, (counts.get(name) ?? 0) + 1);
+  }
+  return [...counts]
+    .map(([name, count]) => ({ name, count }))
+    .sort((left, right) => right.count - left.count || left.name.localeCompare(right.name, "zh-CN"));
+}
+
+// 插件自己写死的默认分类：词库里存的其实是英文，显示时才翻成中文，存回去再还原，
+// 免得同一个「未分类」在词库里分成两条。别的分类都是词库数据，不进翻译层。
+const UNCATEGORIZED = "Uncategorized";
+const categoryName = (value) => (value === UNCATEGORIZED ? t("Uncategorized") : value);
+const isUncategorized = (value) => value === UNCATEGORIZED || value === t("Uncategorized");
+
+// 「分类」输入框：下拉列出已有分类，边打字边收窄，第一项恒为「自定义」。
+// 青绿色的「自定义」标识只表示「这个值不是词库里的现成分类」—— 选了自定义才会挂上，
+// 手打的值一旦命中某个候选就自动摘掉；编辑历史标签时带进来的陌生值也当场算自定义。
+function buildCategoryPicker({ initial = "", categories = [] } = {}) {
+  const root = element("div", "bpi-autocomplete");
+  const wrap = element("div", "bpi-input-wrap");
+  const input = element("input");
+  input.id = "bpi-field-category";
+  input.name = "category";
+  input.autocomplete = "off";
+  input.value = categoryName(initial ?? "");
+  setPlaceholder(input, "Type to filter; ↑↓ to choose");
+  const badge = element("span", "bpi-custom-badge", "Custom");
+  badge.hidden = true;
+  const menu = element("div", "bpi-ac-menu");
+  menu.hidden = true;
+  wrap.append(input, badge);
+  root.append(wrap, menu);
+
+  let catalog = categories;
+  let rows = [];
+  let active = -1;
+  let custom = false;
+
+  const setCustom = (flag) => {
+    custom = flag;
+    badge.hidden = !flag;
+  };
+  const closeMenu = () => {
+    menu.hidden = true;
+    active = -1;
+  };
+  const renderMenu = () => {
+    menu.replaceChildren();
+    if (!rows.length) {
+      closeMenu();
+      return;
+    }
+    rows.forEach((row, index) => {
+      const item = element("div", `bpi-ac-item${index === active ? " bpi-ac-active" : ""}${row.custom ? " bpi-ac-custom" : ""}`);
+      // 分类名是词库数据，不进翻译层；只有插件自己的默认分类翻成中文
+      const name = element("span", "bpi-ac-name");
+      name.textContent = row.custom ? t("Custom") : categoryName(row.name);
+      const hint = element("span", "bpi-ac-kind");
+      hint.textContent = row.custom ? (input.value.trim() || t("Manual input")) : t(`${row.count} items`);
+      item.append(name, hint);
+      // 用 mousedown：等到 click 时输入框已经失焦、菜单早被关掉了
+      item.addEventListener("mousedown", (event) => {
+        event.preventDefault();
+        commit(row);
+      });
+      menu.appendChild(item);
+    });
+    menu.hidden = false;
+  };
+  const update = ({ reveal = true } = {}) => {
+    const needle = input.value.trim().toLowerCase();
+    const matches = catalog.filter((entry) => !needle || categoryName(entry.name).toLowerCase().includes(needle));
+    rows = [{ custom: true }, ...matches];
+    if (!reveal) {
+      menu.hidden = true;
+      return;
+    }
+    // 打出来的字正好等于某个候选时，默认就落在它身上，省一次↓
+    const exact = rows.findIndex((row) => !row.custom && categoryName(row.name) === input.value.trim());
+    active = exact > 0 ? exact : 0;
+    renderMenu();
+  };
+  const commit = (row) => {
+    if (!row) return;
+    if (row.custom) {
+      setCustom(true);
+    } else {
+      input.value = categoryName(row.name);
+      setCustom(false);
+    }
+    closeMenu();
+    input.focus();
+  };
+
+  input.addEventListener("input", () => {
+    if (custom && catalog.some((entry) => categoryName(entry.name) === input.value.trim())) setCustom(false);
+    update();
+  });
+  input.addEventListener("focus", () => update());
+  input.addEventListener("click", () => {
+    if (menu.hidden) update();
+  });
+  input.addEventListener("keydown", (event) => {
+    if (event.key === "ArrowDown" || event.key === "ArrowUp") {
+      event.preventDefault();
+      if (menu.hidden) {
+        update(); // 关着的时候先把列表开出来
+        return;
+      }
+      const step = event.key === "ArrowDown" ? 1 : -1;
+      active = (active + step + rows.length) % rows.length;
+      renderMenu();
+      return;
+    }
+    if (event.key === "Enter") {
+      // 菜单开着才拦回车；关着的时候回车照旧提交整个表单
+      if (menu.hidden) return;
+      event.preventDefault();
+      event.stopPropagation();
+      commit(rows[Math.max(0, active)]);
+      return;
+    }
+    if (event.key === "Escape" && !menu.hidden) {
+      event.stopPropagation();
+      closeMenu();
+    }
+  });
+
+  // Nodes 2.0 会在捕获阶段拦掉指针事件，所以点空白关闭也挂 document + 捕获
+  const onOutside = (event) => {
+    if (!root.contains(event.target)) closeMenu();
+  };
+  document.addEventListener("mousedown", onOutside, true);
+
+  return {
+    root,
+    destroy: () => document.removeEventListener("mousedown", onOutside, true),
+    // 存回词库的是英文原值：显示用的「未分类」要还原，不然会多出一个中文分类
+    value: () => (isUncategorized(input.value.trim()) ? UNCATEGORIZED : input.value.trim()),
+    // 候选是异步取来的，取到之后不能把用户的菜单硬弹开
+    setCategories: (list) => {
+      catalog = list;
+      const current = input.value.trim();
+      if (current && !catalog.some((entry) => categoryName(entry.name) === current)) setCustom(true);
+      // 菜单本来开着就重画，没开就别突然弹出来打断用户
+      update({ reveal: !menu.hidden });
+    },
+  };
+}
+
+// 「标签字段」那一整套：独立的「新增／编辑个人标签」弹窗和节点里双击标签卡后的
+// 「编辑」页签共用同一份，免得两边各抄一遍字段。
+// withTarget = false 时不给「保存到哪个词库」的下拉：双击进来的一定是标签
+// （自然语言片段进不了这个入口），恒存个人词库，让选项出现反而是噪音。
+// readonlyEnglish：英文只显示、不给输入框。改它并不会改提示词，后端按英文做主键，
+// 只会另存一条新词条、旧的还留着——要改提示词里的文字得走节点的「编辑」页。
+function buildTagFields(parent, initial = {}, { withTarget = true, readonlyEnglish = false } = {}) {
+  const englishValue = String(initial?.english ?? "");
+  let english;
+  if (readonlyEnglish) {
+    const label = element("label", "", "English Tag");
+    const text = element("span", "bpi-field-static", englishValue);
+    parent.append(label, text);
+    // 兼容下面按 input 用法读写的地方：只提供 value / focus 两个口子
+    english = { value: englishValue, focus: () => {} };
+  } else {
+    english = field(parent, "English Tag", "english", initial?.english, "e.g.: looking at viewer");
+  }
+  // 中文名称的示例要给中文：这格本来就是填中文的，示例还写英文会让人以为该填英文
+  const chinese = field(parent, "Chinese Name", "chinese", initial?.chinese, "e.g.: apple");
+  const aliases = field(parent, "Chinese Aliases", "aliases", initial?.aliases?.join("，"), "Comma-separated");
+  const categoryLabel = element("label", "", "Category");
+  categoryLabel.htmlFor = "bpi-field-category";
+  // 分类留空也行：后端会把空值落成「自定义」，所以新建时不要预先塞值、免得一开窗就挂着标识
+  const categoryPicker = buildCategoryPicker({ initial: initial?.category ?? "" });
+  parent.append(categoryLabel, categoryPicker.root);
+  // 候选按词库实际用词实时汇总；取不到就退化成普通手输，这行本来也能打字
+  loadDictionary()
+    .then((data) => categoryPicker.setCategories(tagCategoryCounts(data)))
+    .catch(() => {});
+  const models = field(parent, "Models", "models", initial?.models?.join("，") ?? "general, anima", "Comma-separated");
+  const weight = field(parent, "Recommended Weight", "recommended_weight", initial?.recommended_weight ?? "", "Optional");
+  const notes = field(parent, "Notes", "notes", initial?.notes, "Optional");
+  let dictionary = null;
+  if (withTarget) {
+    const dictionaryLabel = element("label", "", "Save to Dictionary");
+    dictionaryLabel.htmlFor = "bpi-field-dictionary";
+    dictionary = element("select", "bpi-mode");
+    dictionary.id = dictionaryLabel.htmlFor;
+    dictionary.name = "dictionary";
+    for (const [value, label] of [["personal", "Personal dictionary"], ["natural", "Natural Language Dictionary"]]) {
+      const option = element("option", "", label);
+      option.value = value;
+      dictionary.appendChild(option);
+    }
+    dictionary.value = initial?.natural ? "natural" : "personal";
+    parent.append(dictionaryLabel, dictionary);
+  }
+  return {
+    destroy: () => categoryPicker.destroy(),
+    focus: () => (initial?.english ? chinese : english).focus(),
+    collect: () => ({
+      english: english.value,
+      chinese: chinese.value,
+      aliases: aliases.value,
+      category: categoryPicker.value(),
+      models: models.value,
+      recommended_weight: weight.value,
+      notes: notes.value,
+      natural: dictionary ? dictionary.value === "natural" : Boolean(initial?.natural),
+      source: initial?.source === "bpi-assistant" ? initial.source : "user",
+      verified: initial?.verified ?? true,
+    }),
+  };
+}
+
 function openTagDialog(initial, onSaved) {
  const shade = element("div", "bpi-modal-shade");
  const modal = element("div", "bpi-modal");
  const title = element("h3", "", initial?.english ? "Add or Edit Personal Tag" : "Add Personal Tag");
  const form = element("form", "bpi-form");
-  const english = field(form, "English Tag", "english", initial?.english, "e.g.: looking at viewer");
-  const chinese = field(form, "Chinese Name", "chinese", initial?.chinese, "e.g.: looking at viewer");
-  const aliases = field(form, "Chinese Aliases", "aliases", initial?.aliases?.join("，"), "Comma-separated");
-  const category = field(form, "Category", "category", initial?.category ?? "Custom", "e.g.: pose, camera, style");
-  const models = field(form, "Models", "models", initial?.models?.join("，") ?? "general, anima", "Comma-separated");
-  const weight = field(form, "Recommended Weight", "recommended_weight", initial?.recommended_weight ?? "", "Optional");
-  const notes = field(form, "Notes", "notes", initial?.notes, "Optional");
+  const fields = buildTagFields(form, initial);
  const error = element("div", "bpi-status");
  error.dataset.kind = "error";
  error.style.gridColumn = "1 / -1";
  form.appendChild(error);
  const actions = element("div", "bpi-modal-actions");
-  const close = () => shade.remove();
+  const close = () => {
+    fields.destroy();
+    shade.remove();
+  };
  actions.append(
    button("Cancel", close),
    button("Save to Personal Dictionary", () => form.requestSubmit(), "bpi-primary"),
@@ -722,24 +987,235 @@ function openTagDialog(initial, onSaved) {
    event.preventDefault();
    error.textContent = "";
    try {
-      await saveTag({
-        english: english.value,
-        chinese: chinese.value,
-        aliases: aliases.value,
-        category: category.value,
-        models: models.value,
-        recommended_weight: weight.value,
-        notes: notes.value,
-        source: initial?.source === "bpi-assistant" ? initial.source : "user",
-        verified: initial?.verified ?? true,
-      });
+      await saveTag(fields.collect());
      close();
      await onSaved();
    } catch (saveError) {
      error.textContent = saveError.message;
    }
  });
-  setTimeout(() => (initial?.english ? chinese : english).focus(), 0);
+  setTimeout(() => fields.focus(), 0);
+}
+
+// 颜色区能配色的「来源」：内置 / 社区词库包 + 大词库 + 两个本机词库 + 未收录。
+// 后三者不在 packs 里（个人词库和自然语言词库共用 user_tags.json，未收录压根没有词条），得手工补。
+function colorTargets(data) {
+  const userTags = data?.user ?? [];
+  const items = (count) => t(`${count ?? 0} items`);
+  const targets = (data?.packs ?? []).map((pack) => ({ id: pack.id, name: pack.name, meta: items(pack.count) }));
+  const large = data?.large_dictionary;
+  if (large?.available) {
+    targets.push({ id: "danbooru_large", name: large.name, meta: items(large.count) });
+  }
+  targets.push({ id: "personal", name: t("Personal dictionary"), meta: items(userTags.filter((tag) => !tag.natural).length) });
+  targets.push({ id: "natural", name: t("Natural Language Dictionary"), meta: items(userTags.filter((tag) => tag.natural).length) });
+  targets.push({ id: UNINDEXED_PACK_ID, name: t("Not indexed"), meta: t("Tags matching no dictionary") });
+  return targets;
+}
+
+// #RGB / #RRGGBB 统一成小写 6 位；认不出来返回空串（= 不着色）
+function normalizeHexColor(value) {
+  const text = String(value ?? "").trim();
+  if (/^#[0-9a-fA-F]{6}$/.test(text)) return text.toLowerCase();
+  if (/^#[0-9a-fA-F]{3}$/.test(text)) return "#" + text.slice(1).split("").map((char) => char + char).join("").toLowerCase();
+  return "";
+}
+
+// 按 Anima 配色时的分类名：跟 anima_sorter.js 的 ANIMA_SLOTS 一一对应。
+// 网格卡片窄，用短名，ANIMA_SLOTS 里的英文全称放 meta。
+const ANIMA_SHORT_LABELS = {
+  quality: "Quality",
+  people: "People count",
+  character: "Character",
+  copyright: "Copyright",
+  artist: "Artist",
+  appearance: "Appearance",
+  action: "Expression",
+  camera: "Camera",
+  style: "Style",
+  environment: "Scene",
+  natural: "Natural language",
+  uncertain: "Uncertain",
+};
+
+function animaColorTargets() {
+  return ANIMA_SLOTS.map((slot) => ({
+    id: slot.id,
+    name: t(ANIMA_SHORT_LABELS[slot.id] ?? slot.label),
+    meta: slot.label,
+  }));
+}
+
+// 一块配色网格：按词库一块、按 Anima 一块，共用这套卡片。改动后由 onCommit 落盘。
+function buildColorGrid({ targets, initial, presets, dicePool, error, onCommit }) {
+  const grid = element("div", "bpi-color-grid");
+  const picked = new Map();
+  const setters = [];
+  for (const target of targets) {
+    picked.set(target.id, initial?.[target.id] || "");
+    const card = element("div", "bpi-color-card");
+    const chip = element("span", "bpi-color-chip");
+    const head = element("div", "bpi-color-card-head");
+    head.append(chip, element("span", "", target.name));
+    const swatchRow = element("div", "bpi-color-swatches");
+    const swatches = [];
+    for (const hex of presets) {
+      const dot = element("i");
+      dot.style.background = hex;
+      dot.dataset.color = hex;
+      setTitle(dot, hex);
+      swatchRow.appendChild(dot);
+      swatches.push(dot);
+    }
+    const input = element("input");
+    setPlaceholder(input, "#6b9b78");
+    const dice = element("button", "bpi-color-dice");
+    dice.type = "button";
+    dice.innerHTML = DICE_ICON;
+    setTitle(dice, "Random color");
+    const row = element("div", "bpi-color-row");
+    row.append(input, dice, button("None", () => {
+      apply("");
+      onCommit();
+    }));
+
+    function apply(value) {
+      const color = normalizeHexColor(value);
+      picked.set(target.id, color);
+      chip.style.background = color || "transparent";
+      for (const dot of swatches) dot.classList.toggle("bpi-swatch-on", Boolean(color) && dot.dataset.color === color);
+      input.value = color;
+      error.textContent = "";
+    }
+    swatchRow.addEventListener("click", (event) => {
+      const hex = event.target?.dataset?.color;
+      if (!hex) return;
+      apply(hex);
+      onCommit();
+    });
+    dice.addEventListener("click", () => {
+      apply(dicePool[Math.floor(Math.random() * dicePool.length)]);
+      onCommit();
+    });
+    input.addEventListener("change", () => {
+      const value = input.value.trim();
+      if (!value) {
+        apply("");
+        onCommit();
+        return;
+      }
+      if (!normalizeHexColor(value)) {
+        error.textContent = t("Invalid color; use #RGB or #RRGGBB");
+        return;
+      }
+      apply(value);
+      onCommit();
+    });
+    apply(picked.get(target.id));
+    setters.push(apply);
+
+    card.append(head, element("div", "bpi-color-meta", target.meta), swatchRow, row);
+    grid.appendChild(card);
+  }
+  return {
+    root: grid,
+    values() {
+      const payload = {};
+      for (const [id, color] of picked) {
+        if (color) payload[id] = color;
+      }
+      return payload;
+    },
+    reset() {
+      for (const set of setters) set("");
+    },
+  };
+}
+
+// 颜色面板：节点过滤栏点「颜色」就在过滤栏下面就地展开，不做弹窗。
+// 跟左边那排筛选按钮一个操作逻辑：点一下立刻生效；按词库 / 按 Anima 才展开对应的配色网格。
+function buildColorPanel({ data, colors, mode, onChange }) {
+  const root = element("div", "bpi-color-panel");
+  const modes = [
+    ["default", t("Default"), ["var(--bpi-border-2)"]],
+    ["status", t("By inclusion"), ["#c08a4a", "#8a6bb0"]],
+    ["random", t("Random"), ["#5b8a72", "#8a5b7a", "#5b6e8a"]],
+    ["by_pack", t("By dictionary"), ["#6b9b78", "#6b7b9b", "#9b6b8b"]],
+    ["by_anima", t("By Anima"), ["#9b8b6b", "#9b6b8b", "#6b9b9b"]],
+  ];
+  // 旧存档里可能有不在选项里的值，回落到默认，别让按钮组一个都不亮
+  let currentMode = modes.some(([value]) => value === mode) ? mode : "default";
+  let live = colors;
+  const error = element("div", "bpi-color-error");
+  const presets = (colors?.presets ?? []).length ? colors.presets : ["#6b9b78"];
+  const dicePool = (colors?.random_pool ?? []).length ? colors.random_pool : presets;
+
+  const modeRow = element("div", "bpi-color-modes");
+  const modeButtons = new Map();
+  for (const [value, label, dots] of modes) {
+    const item = element("button", "bpi-color-mode");
+    item.type = "button";
+    for (const dot of dots) {
+      const mark = element("i", "bpi-color-mode-dot");
+      mark.style.background = dot;
+      item.appendChild(mark);
+    }
+    item.appendChild(element("span", "", label));
+    // 模式本身不落盘，只有配色才写文件
+    item.addEventListener("click", () => setMode(value));
+    modeButtons.set(value, item);
+    modeRow.appendChild(item);
+  }
+
+  async function commit() {
+    try {
+      live = await saveTokenColors({ pack_colors: packGrid.values(), anima_colors: animaGrid.values() });
+      error.textContent = "";
+    } catch (saveError) {
+      error.textContent = saveError.message;
+      return;
+    }
+    onChange(currentMode, live);
+  }
+
+  const gridOptions = { presets, dicePool, error, onCommit: commit };
+  const packGrid = buildColorGrid({ ...gridOptions, targets: colorTargets(data), initial: colors?.pack_colors });
+  const animaGrid = buildColorGrid({ ...gridOptions, targets: animaColorTargets(), initial: colors?.anima_colors });
+
+  const gridActions = (reset) => {
+    const actions = element("div", "bpi-color-panel-actions");
+    actions.appendChild(button("Restore defaults", () => {
+      reset();
+      commit();
+    }));
+    return actions;
+  };
+  const packWrap = element("div");
+  packWrap.append(
+    element("div", "bpi-color-note", t("Pick a color for each dictionary; uncolored ones keep the default.")),
+    packGrid.root,
+    gridActions(() => packGrid.reset()),
+  );
+  const animaWrap = element("div");
+  animaWrap.append(
+    element("div", "bpi-color-note", t("Pick a color for each Anima category; unset ones keep the default.")),
+    animaGrid.root,
+    gridActions(() => animaGrid.reset()),
+  );
+
+  const setMode = (value, notify = true) => {
+    currentMode = value;
+    for (const [id, item] of modeButtons) item.classList.toggle("bpi-color-mode-active", id === value);
+    packWrap.hidden = value !== "by_pack";
+    animaWrap.hidden = value !== "by_anima";
+    error.textContent = "";
+    if (notify) onChange(value, live);
+  };
+
+  root.append(modeRow, packWrap, animaWrap, error);
+  // 首次只是把当前模式点亮、决定网格显不显示，不该顺带触发一次重绘和同步
+  setMode(currentMode, false);
+  return root;
 }

 function openManagerPanel(section = "", nodeId = null) {
@@ -758,53 +1234,55 @@ function injectBpiStyles() {
  const style = document.createElement("style");
  style.id = "bpi-styles";
  style.textContent = `
-    :root{--bpi-surface:rgba(255,255,255,.92);--bpi-surface-2:#f5f5f5;--bpi-surface-3:#eaeaea;--bpi-surface-table:#f0f0f0;--bpi-surface-modal:#fff;--bpi-text:#333;--bpi-text-muted:#666;--bpi-text-faint:#999;--bpi-border:#ccc;--bpi-border-2:#bbb;--bpi-border-3:#e0e0e0;--bpi-button-bg:#e8e8e8;--bpi-button-text:#333;--bpi-input-bg:#fff;--bpi-input-text:#333;--bpi-head-bg:#e8e8e8;--bpi-editor-bg:#fff}
-    .dark-theme{--bpi-surface:rgba(10,12,18,.72);--bpi-surface-2:#171b22;--bpi-surface-3:#14181f;--bpi-surface-table:#11151b;--bpi-surface-modal:#20252d;--bpi-text:#e6edf7;--bpi-text-muted:#aab5c5;--bpi-text-faint:#7f8a9a;--bpi-border:#3a4250;--bpi-border-2:#4c5668;--bpi-border-3:#29313d;--bpi-button-bg:#2c3340;--bpi-button-text:#e8edf5;--bpi-input-bg:#171b22;--bpi-input-text:#eef3fb;--bpi-head-bg:#252c36;--bpi-editor-bg:#111923}
-    [data-bpi-theme="dark"]{--bpi-surface:rgba(10,12,18,.72);--bpi-surface-2:#171b22;--bpi-surface-3:#14181f;--bpi-surface-table:#11151b;--bpi-surface-modal:#20252d;--bpi-text:#e6edf7;--bpi-text-muted:#aab5c5;--bpi-text-faint:#7f8a9a;--bpi-border:#3a4250;--bpi-border-2:#4c5668;--bpi-border-3:#29313d;--bpi-button-bg:#2c3340;--bpi-button-text:#e8edf5;--bpi-input-bg:#171b22;--bpi-input-text:#eef3fb;--bpi-head-bg:#252c36;--bpi-editor-bg:#111923}
-    [data-bpi-theme="light"]{--bpi-surface:rgba(255,255,255,.92);--bpi-surface-2:#f5f5f5;--bpi-surface-3:#eaeaea;--bpi-surface-table:#f0f0f0;--bpi-surface-modal:#fff;--bpi-text:#333;--bpi-text-muted:#666;--bpi-text-faint:#999;--bpi-border:#ccc;--bpi-border-2:#bbb;--bpi-border-3:#e0e0e0;--bpi-button-bg:#e8e8e8;--bpi-button-text:#333;--bpi-input-bg:#fff;--bpi-input-text:#333;--bpi-head-bg:#e8e8e8;--bpi-editor-bg:#fff}
-    .bpi-panel{box-sizing:border-box;width:100%;height:100%;min-height:var(--bpi-collapsed-height,390px);max-height:var(--bpi-collapsed-height,390px);padding:8px;display:flex;flex-direction:column;gap:7px;color:var(--bpi-text);font:12px/1.4 Arial,sans-serif;background:var(--bpi-surface);border:1px solid var(--bpi-border);border-radius:8px;overflow-y:auto;scrollbar-gutter:stable}
-    .bpi-english-section{flex:none;border:1px solid #3e7197;border-radius:7px;background:var(--bpi-surface);overflow:hidden}.bpi-english-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:#cbe8ff;font-weight:700;flex-wrap:wrap}.bpi-english-hint{margin-left:auto;color:#91b4cf;font-size:10px;font-weight:400}.bpi-english-token-view{box-sizing:border-box;width:100%;height:auto;min-height:92px;max-height:360px;overflow:auto;scrollbar-gutter:stable;padding:8px 9px;outline:none;color:var(--bpi-text);font:12px/1.85 Consolas,monospace;white-space:normal}.bpi-english-token-view:focus{box-shadow:inset 0 0 0 1px #5aa6d8}.bpi-english-editor{box-sizing:border-box;width:100%;height:120px;min-height:92px;max-height:420px;overflow-y:auto;resize:vertical;border:0;border-top:1px solid #3e7197;background:var(--bpi-editor-bg);color:var(--bpi-text);padding:9px;outline:none;font:12px/1.55 Consolas,monospace}.bpi-english-editor:focus{box-shadow:inset 0 0 0 1px #5aa6d8}.bpi-english-token{font-family:Consolas,monospace;color:var(--bpi-text)}.bpi-english-token.bpi-linked{background:#147fc3;border-color:#76c9ff;color:#fff;box-shadow:0 0 0 1px rgba(118,201,255,.32)}
-    .bpi-mirror-section{flex:none;border:1px solid #488739;border-radius:7px;background:var(--bpi-surface);overflow:visible}.bpi-section-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:#cbe9c8;font-weight:700;flex-wrap:wrap}.bpi-section-hint{margin-left:auto;color:#91ad93;font-size:10px;font-weight:400}.bpi-mirror-actions{display:flex;gap:4px;align-items:center;flex-wrap:wrap}.bpi-chinese-mirror{box-sizing:border-box;width:100%;height:auto;min-height:92px;max-height:360px;overflow:auto;resize:none;scrollbar-gutter:stable;padding:8px 9px;outline:none;color:var(--bpi-text);line-height:1.85;cursor:text;white-space:normal}.bpi-chinese-mirror:focus{box-shadow:inset 0 0 0 1px #70b35f}.bpi-category-table{display:grid;grid-template-columns:minmax(110px,150px) minmax(0,1fr);border:1px solid #315f3a;border-radius:6px;overflow:hidden;background:var(--bpi-surface-3)}.bpi-category-row{display:contents}.bpi-category-name,.bpi-category-content{padding:6px 8px;border-top:1px solid #31513a}.bpi-category-row:first-child .bpi-category-name,.bpi-category-row:first-child .bpi-category-content{border-top:0}.bpi-category-name{background:var(--bpi-head-bg);border-right:1px solid #31513a;color:#9ee8a7;font-weight:700}.bpi-category-content{min-width:0}.bpi-chinese-editor{box-sizing:border-box;width:100%;height:auto;min-height:110px;max-height:600px;overflow-y:auto;resize:vertical;border:0;border-top:1px solid #3b693c;background:var(--bpi-editor-bg);color:var(--bpi-text);padding:9px;outline:none;font:12px/1.65 Arial,sans-serif}.bpi-chinese-editor:focus{box-shadow:inset 0 0 0 1px #70b35f}.bpi-hidden{display:none!important}.bpi-mirror-empty{color:var(--bpi-text-muted)}.bpi-mirror-token{display:inline-flex;align-items:center;border:1px solid transparent;border-radius:5px;padding:0 3px;margin:1px 0;cursor:pointer;transition:background .12s,border-color .12s,color .12s}.bpi-mirror-token:hover{background:#345a3b;border-color:#5c8d63;color:#fff}.bpi-mirror-token.bpi-linked{background:#315f83;border-color:#76a9ff;color:#fff}
+    :root{--bpi-field-bg:#fff;--bpi-field-border:#bbb;--bpi-switch-off:#c9ced6;--bpi-switch-knob:#fff;--bpi-switch-off-border:#b0b6bf;--bpm-accent:#45926f;--bpm-accent-hover:#3a7d5e;--bpm-accent-border:#35755a;--bpi-custom-bg:#e6f6f1;--bpi-custom-border:#7fc9b0;--bpi-custom-text:#0f6e56;--bpi-surface:rgba(255,255,255,.92);--bpi-surface-2:#f5f5f5;--bpi-surface-3:#eaeaea;--bpi-surface-table:#f0f0f0;--bpi-surface-modal:#fff;--bpi-text:#333;--bpi-text-muted:#666;--bpi-text-faint:#999;--bpi-border:#ccc;--bpi-border-2:#bbb;--bpi-border-3:#e0e0e0;--bpi-button-bg:#e8e8e8;--bpi-button-text:#333;--bpi-input-bg:#fff;--bpi-input-text:#333;--bpi-head-bg:#e8e8e8;--bpi-danger-bg:#c0392b;--bpi-danger-bg-hover:#a53023;--bpi-danger-border:#a53023;--bpi-danger-text:#fff;--bpi-break-bg:#e6f6f1;--bpi-break-bg-hover:#0f6e56;--bpi-break-border:#7fc9b0;--bpi-break-text:#0f6e56;--bpi-primary-bg:#2b6cb0;--bpi-primary-bg-hover:#245a94;--bpi-primary-border:#1e4f85;--bpi-primary-text:#fff;--bpi-en-head-text:#17557f;--bpi-en-hint-text:#5b7f9e;--bpi-zh-head-text:#1f5c22;--bpi-zh-hint-text:#5b7d5e;--bpi-cat-name-text:#1e5a28;--bpi-info-text:#1f5f96;--bpi-link-text:#1a5f96;--bpi-zh-text:#18508a;--bpi-status-error:#b3261e;--bpi-status-ok:#1a7f45;--bpi-status-busy:#8a5a00;--bpi-badge-warn:#8a5a00;--bpi-badge-source:#0f6e56;--bpi-badge-suspected:#1c5f80;--bpi-badge-machine:#5b3fa8;--bpi-star-on:#a8750a;--bpi-star-off:#6b7280;--bpi-quick-head-text:#4a3a7a;--bpi-quick-hint-text:#6b5a8f;--bpi-handle-text:#4a6a85;--bpi-handle-hover:#1f5f96;--bpi-mini-hover-bg:#d8dee7;--bpi-handle-hover-bg:#e2e8ef;--bpi-chip-hide-bg:#eef1f5;--bpi-waiting-bg:#fdf1dc;--bpi-waiting-border:#d9a95a;--bpi-hidden-chip-hover-bg:#e8eef5;--bpi-issues-bg:#fdf6e8;--bpi-issues-border:#d9b98a;--bpi-issue-text:#7a5410;--bpi-row-hover-overlay:rgba(47,108,176,.06);--bpi-row-pinned-overlay:#cfe3f7;--bpi-row-hover-ring:rgba(47,108,176,.55);--bpi-row-pinned-ring:#1a5f96;--bpi-row-hover-text:#1f1f1f;--bpi-row-hover-shadow:none;--bpi-row-selected-bg:#dcebf8;--bpi-en-section-bg:rgba(43,108,176,.06);--bpi-quick-section-bg:rgba(107,85,148,.06);--bpi-dup-overlay:rgba(179,38,30,.12);--bpi-editor-bg:#fff;--bpi-hit-bg:#fbe49c;--bpi-hit-text:#7a5410}
+    .dark-theme{--bpi-field-bg:#0b0d10;--bpi-field-border:#333a45;--bpi-switch-off:#39414f;--bpi-switch-knob:#e8edf5;--bpi-switch-off-border:#4c5668;--bpi-custom-bg:#17342c;--bpi-custom-border:#3f8f77;--bpi-custom-text:#7fdcbd;--bpi-surface:rgba(10,12,18,.72);--bpi-surface-2:#171b22;--bpi-surface-3:#14181f;--bpi-surface-table:#11151b;--bpi-surface-modal:#20252d;--bpi-text:#e6edf7;--bpi-text-muted:#aab5c5;--bpi-text-faint:#7f8a9a;--bpi-border:#3a4250;--bpi-border-2:#4c5668;--bpi-border-3:#29313d;--bpi-button-bg:#2c3340;--bpi-button-text:#e8edf5;--bpi-input-bg:#171b22;--bpi-input-text:#eef3fb;--bpi-head-bg:#252c36;--bpi-danger-bg:#65313b;--bpi-danger-bg-hover:#7d3d49;--bpi-danger-border:#8d4653;--bpi-danger-text:#ffd9de;--bpi-break-bg:#17342c;--bpi-break-bg-hover:#0f6e56;--bpi-break-border:#3f8f77;--bpi-break-text:#7fdcbd;--bpi-primary-bg:#285f8e;--bpi-primary-bg-hover:#346f9f;--bpi-primary-border:#3f83ba;--bpi-primary-text:#eaf3fb;--bpi-en-head-text:#cbe8ff;--bpi-en-hint-text:#91b4cf;--bpi-zh-head-text:#cbe9c8;--bpi-zh-hint-text:#91ad93;--bpi-cat-name-text:#9ee8a7;--bpi-info-text:#87bfea;--bpi-link-text:#76c9ff;--bpi-zh-text:#9ed0ff;--bpi-status-error:#ff8b93;--bpi-status-ok:#7fdda2;--bpi-status-busy:#ffd27a;--bpi-badge-warn:#e2ca78;--bpi-badge-source:#86d5c4;--bpi-badge-suspected:#7fc4e0;--bpi-badge-machine:#cbb1ff;--bpi-star-on:#ffd45f;--bpi-star-off:#8f9aaa;--bpi-quick-head-text:#d9ccff;--bpi-quick-hint-text:#a795c6;--bpi-handle-text:#7194ad;--bpi-handle-hover:#a9c8e4;--bpi-mini-hover-bg:#3c4b60;--bpi-handle-hover-bg:#2b3a4d;--bpi-chip-hide-bg:#1d2937;--bpi-waiting-bg:rgba(70,52,25,.5);--bpi-waiting-border:#8a5f16;--bpi-hidden-chip-hover-bg:#31404f;--bpi-issues-bg:rgba(55,39,25,.55);--bpi-issues-border:#5b4a36;--bpi-issue-text:#e6c792;--bpi-row-hover-overlay:rgba(76,167,232,.34);--bpi-row-pinned-overlay:#314b68;--bpi-row-hover-ring:rgba(76,167,232,.7);--bpi-row-pinned-ring:#76c9ff;--bpi-row-hover-text:#fff;--bpi-row-hover-shadow:0 0 7px rgba(109,190,255,.6);--bpi-row-selected-bg:#314b68;--bpi-en-section-bg:rgba(76,167,232,.08);--bpi-quick-section-bg:rgba(167,149,198,.08);--bpi-dup-overlay:rgba(255,139,147,.22);--bpi-editor-bg:#111923;--bpi-hit-bg:#5a4413;--bpi-hit-text:#ffe6a8}
+    [data-bpi-theme="dark"]{--bpi-field-bg:#0b0d10;--bpi-field-border:#333a45;--bpi-switch-off:#39414f;--bpi-switch-knob:#e8edf5;--bpi-switch-off-border:#4c5668;--bpi-custom-bg:#17342c;--bpi-custom-border:#3f8f77;--bpi-custom-text:#7fdcbd;--bpi-surface:rgba(10,12,18,.72);--bpi-surface-2:#171b22;--bpi-surface-3:#14181f;--bpi-surface-table:#11151b;--bpi-surface-modal:#20252d;--bpi-text:#e6edf7;--bpi-text-muted:#aab5c5;--bpi-text-faint:#7f8a9a;--bpi-border:#3a4250;--bpi-border-2:#4c5668;--bpi-border-3:#29313d;--bpi-button-bg:#2c3340;--bpi-button-text:#e8edf5;--bpi-input-bg:#171b22;--bpi-input-text:#eef3fb;--bpi-head-bg:#252c36;--bpi-danger-bg:#65313b;--bpi-danger-bg-hover:#7d3d49;--bpi-danger-border:#8d4653;--bpi-danger-text:#ffd9de;--bpi-break-bg:#17342c;--bpi-break-bg-hover:#0f6e56;--bpi-break-border:#3f8f77;--bpi-break-text:#7fdcbd;--bpi-primary-bg:#285f8e;--bpi-primary-bg-hover:#346f9f;--bpi-primary-border:#3f83ba;--bpi-primary-text:#eaf3fb;--bpi-en-head-text:#cbe8ff;--bpi-en-hint-text:#91b4cf;--bpi-zh-head-text:#cbe9c8;--bpi-zh-hint-text:#91ad93;--bpi-cat-name-text:#9ee8a7;--bpi-info-text:#87bfea;--bpi-link-text:#76c9ff;--bpi-zh-text:#9ed0ff;--bpi-status-error:#ff8b93;--bpi-status-ok:#7fdda2;--bpi-status-busy:#ffd27a;--bpi-badge-warn:#e2ca78;--bpi-badge-source:#86d5c4;--bpi-badge-suspected:#7fc4e0;--bpi-badge-machine:#cbb1ff;--bpi-star-on:#ffd45f;--bpi-star-off:#8f9aaa;--bpi-quick-head-text:#d9ccff;--bpi-quick-hint-text:#a795c6;--bpi-handle-text:#7194ad;--bpi-handle-hover:#a9c8e4;--bpi-mini-hover-bg:#3c4b60;--bpi-handle-hover-bg:#2b3a4d;--bpi-chip-hide-bg:#1d2937;--bpi-waiting-bg:rgba(70,52,25,.5);--bpi-waiting-border:#8a5f16;--bpi-hidden-chip-hover-bg:#31404f;--bpi-issues-bg:rgba(55,39,25,.55);--bpi-issues-border:#5b4a36;--bpi-issue-text:#e6c792;--bpi-row-hover-overlay:rgba(76,167,232,.34);--bpi-row-pinned-overlay:#314b68;--bpi-row-hover-ring:rgba(76,167,232,.7);--bpi-row-pinned-ring:#76c9ff;--bpi-row-hover-text:#fff;--bpi-row-hover-shadow:0 0 7px rgba(109,190,255,.6);--bpi-row-selected-bg:#314b68;--bpi-en-section-bg:rgba(76,167,232,.08);--bpi-quick-section-bg:rgba(167,149,198,.08);--bpi-dup-overlay:rgba(255,139,147,.22);--bpi-editor-bg:#111923;--bpi-hit-bg:#5a4413;--bpi-hit-text:#ffe6a8}
+    [data-bpi-theme="light"]{--bpi-field-bg:#fff;--bpi-field-border:#bbb;--bpi-switch-off:#c9ced6;--bpi-switch-knob:#fff;--bpi-switch-off-border:#b0b6bf;--bpi-custom-bg:#e6f6f1;--bpi-custom-border:#7fc9b0;--bpi-custom-text:#0f6e56;--bpi-surface:rgba(255,255,255,.92);--bpi-surface-2:#f5f5f5;--bpi-surface-3:#eaeaea;--bpi-surface-table:#f0f0f0;--bpi-surface-modal:#fff;--bpi-text:#333;--bpi-text-muted:#666;--bpi-text-faint:#999;--bpi-border:#ccc;--bpi-border-2:#bbb;--bpi-border-3:#e0e0e0;--bpi-button-bg:#e8e8e8;--bpi-button-text:#333;--bpi-input-bg:#fff;--bpi-input-text:#333;--bpi-head-bg:#e8e8e8;--bpi-danger-bg:#c0392b;--bpi-danger-bg-hover:#a53023;--bpi-danger-border:#a53023;--bpi-danger-text:#fff;--bpi-break-bg:#e6f6f1;--bpi-break-bg-hover:#0f6e56;--bpi-break-border:#7fc9b0;--bpi-break-text:#0f6e56;--bpi-primary-bg:#2b6cb0;--bpi-primary-bg-hover:#245a94;--bpi-primary-border:#1e4f85;--bpi-primary-text:#fff;--bpi-en-head-text:#17557f;--bpi-en-hint-text:#5b7f9e;--bpi-zh-head-text:#1f5c22;--bpi-zh-hint-text:#5b7d5e;--bpi-cat-name-text:#1e5a28;--bpi-info-text:#1f5f96;--bpi-link-text:#1a5f96;--bpi-zh-text:#18508a;--bpi-status-error:#b3261e;--bpi-status-ok:#1a7f45;--bpi-status-busy:#8a5a00;--bpi-badge-warn:#8a5a00;--bpi-badge-source:#0f6e56;--bpi-badge-suspected:#1c5f80;--bpi-badge-machine:#5b3fa8;--bpi-star-on:#a8750a;--bpi-star-off:#6b7280;--bpi-quick-head-text:#4a3a7a;--bpi-quick-hint-text:#6b5a8f;--bpi-handle-text:#4a6a85;--bpi-handle-hover:#1f5f96;--bpi-mini-hover-bg:#d8dee7;--bpi-handle-hover-bg:#e2e8ef;--bpi-chip-hide-bg:#eef1f5;--bpi-waiting-bg:#fdf1dc;--bpi-waiting-border:#d9a95a;--bpi-hidden-chip-hover-bg:#e8eef5;--bpi-issues-bg:#fdf6e8;--bpi-issues-border:#d9b98a;--bpi-issue-text:#7a5410;--bpi-row-hover-overlay:rgba(47,108,176,.06);--bpi-row-pinned-overlay:#cfe3f7;--bpi-row-hover-ring:rgba(47,108,176,.55);--bpi-row-pinned-ring:#1a5f96;--bpi-row-hover-text:#1f1f1f;--bpi-row-hover-shadow:none;--bpi-row-selected-bg:#dcebf8;--bpi-en-section-bg:rgba(43,108,176,.06);--bpi-quick-section-bg:rgba(107,85,148,.06);--bpi-dup-overlay:rgba(179,38,30,.12);--bpi-editor-bg:#fff;--bpi-hit-bg:#fbe49c;--bpi-hit-text:#7a5410}
+    .bpi-panel{box-sizing:border-box;width:100%;height:100%;min-height:var(--bpi-collapsed-height,390px);max-height:var(--bpi-collapsed-height,390px);padding:8px;display:flex;flex-direction:column;gap:7px;color:var(--bpi-text);font:12px/1.4 Arial,sans-serif;background:var(--bpi-surface);border:1px solid var(--bpi-border);border-radius:8px;overflow-y:auto;scrollbar-gutter:stable}.bpi-panel.bpi-fill{max-height:none}.bpi-panel.bpi-fill > .bpi-english-section{display:flex;flex-direction:column;flex:1 1 auto;min-height:0}.bpi-panel.bpi-fill > .bpi-english-section > *{flex:none}.bpi-panel.bpi-fill > .bpi-english-section > .bpi-english-token-view{flex:1 1 auto;min-height:92px;max-height:none}.bpi-panel.bpi-fill > .bpi-english-section > .bpi-english-editor{flex:1 1 auto;min-height:120px}
+    .bpi-english-section{flex:none;border:1px solid #3e7197;border-radius:7px;background-color:var(--bpi-en-section-bg);overflow:hidden}.bpi-english-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:var(--bpi-en-head-text);font-weight:700;flex-wrap:wrap}.bpi-english-hint{margin-left:auto;color:var(--bpi-en-hint-text);font-size:10px;font-weight:400}.bpi-english-token-view{box-sizing:border-box;width:100%;height:auto;min-height:92px;max-height:360px;overflow:auto;scrollbar-gutter:stable;padding:8px 9px;outline:none;color:var(--bpi-text);font:12px/1.85 Consolas,monospace;white-space:normal}.bpi-english-token-view:focus{box-shadow:inset 0 0 0 1px #5aa6d8}.bpi-english-editor{box-sizing:border-box;width:100%;height:120px;min-height:92px;max-height:420px;overflow-y:auto;resize:vertical;border:0;border-top:1px solid #3e7197;background-color:transparent;color:var(--bpi-text);padding:9px;outline:none;font:12px/1.55 Consolas,monospace}.bpi-english-editor:focus{box-shadow:inset 0 0 0 1px #5aa6d8}.bpi-english-token{font-family:Consolas,monospace;color:var(--bpi-text)}.bpi-english-token.bpi-linked{background:#147fc3;border-color:#76c9ff;color:#fff;box-shadow:0 0 0 1px rgba(118,201,255,.32)}
+    .bpi-mirror-section{flex:none;border:1px solid #488739;border-radius:7px;background:var(--bpi-surface);overflow:visible}.bpi-section-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:var(--bpi-zh-head-text);font-weight:700;flex-wrap:wrap}.bpi-section-hint{margin-left:auto;color:var(--bpi-zh-hint-text);font-size:10px;font-weight:400}.bpi-mirror-actions{display:flex;gap:4px;align-items:center;flex-wrap:wrap}.bpi-chinese-mirror{box-sizing:border-box;width:100%;height:auto;min-height:92px;max-height:360px;overflow:auto;resize:none;scrollbar-gutter:stable;padding:8px 9px;outline:none;color:var(--bpi-text);line-height:1.85;cursor:text;white-space:normal}.bpi-chinese-mirror:focus{box-shadow:inset 0 0 0 1px #70b35f}.bpi-category-table{display:grid;grid-template-columns:minmax(110px,150px) minmax(0,1fr);border:1px solid #315f3a;border-radius:6px;overflow:hidden;background:var(--bpi-surface-3)}.bpi-category-row{display:contents}.bpi-category-name,.bpi-category-content{padding:6px 8px;border-top:1px solid #31513a}.bpi-category-row:first-child .bpi-category-name,.bpi-category-row:first-child .bpi-category-content{border-top:0}.bpi-category-name{background:var(--bpi-head-bg);border-right:1px solid #31513a;color:var(--bpi-cat-name-text);font-weight:700}.bpi-category-content{min-width:0}.bpi-chinese-editor{box-sizing:border-box;width:100%;height:auto;min-height:110px;max-height:600px;overflow-y:auto;resize:vertical;border:0;border-top:1px solid #3b693c;background:var(--bpi-editor-bg);color:var(--bpi-text);padding:9px;outline:none;font:12px/1.65 Arial,sans-serif}.bpi-chinese-editor:focus{box-shadow:inset 0 0 0 1px #70b35f}.bpi-hidden{display:none!important}.bpi-mirror-empty{color:var(--bpi-text-muted)}.bpi-mirror-token{display:inline-flex;align-items:center;border:1px solid transparent;border-radius:5px;padding:0 3px;margin:1px 0;cursor:pointer;transition:background .12s,border-color .12s,color .12s}.bpi-mirror-token:hover{background:#345a3b;border-color:#5c8d63;color:#fff}.bpi-mirror-token.bpi-linked{background:#315f83;border-color:#76a9ff;color:#fff}
    .bpi-mirror-token.bpi-linked,.bpi-english-token.bpi-linked{background:#147fc3;border-color:#76c9ff;color:#fff;box-shadow:0 0 0 1px rgba(118,201,255,.32)}.bpi-english-token-view .bpi-mirror-separator{color:#7194ad}
    .bpi-toolbar,.bpi-search-line,.bpi-summary{display:flex;gap:6px;align-items:center;flex-wrap:wrap}
    .bpi-toolbar{justify-content:space-between}.bpi-toolbar-group{display:flex;gap:5px;align-items:center;flex-wrap:wrap}
    .bpi-button{border:1px solid var(--bpi-border-2);border-radius:5px;padding:4px 8px;background:var(--bpi-button-bg);color:var(--bpi-button-text);cursor:pointer;font-size:11px;line-height:1.25}
-    .bpi-button:hover{background:var(--bpi-border-3);border-color:#6c83a8}.bpi-button:disabled{opacity:.45;cursor:default}.bpi-button.bpi-primary{background:#285f8e;border-color:#3f83ba}.bpi-button.bpi-danger{background:#65313b;border-color:#8d4653}
+    .bpi-button:hover{background:var(--bpi-border-3);border-color:#6c83a8}.bpi-button:disabled{opacity:.45;cursor:default}.bpi-button.bpi-primary,.bpi-mini.bpi-primary{background:var(--bpi-primary-bg);border-color:var(--bpi-primary-border);color:var(--bpi-primary-text)}.bpi-button.bpi-primary:hover,.bpi-mini.bpi-primary:hover{background:var(--bpi-primary-bg-hover);border-color:var(--bpi-primary-border);color:var(--bpi-primary-text)}.bpi-button.bpi-danger{background:var(--bpi-danger-bg);border-color:var(--bpi-danger-border);color:var(--bpi-danger-text)}.bpi-button.bpi-danger:hover,.bpi-mini.bpi-danger:hover{background:var(--bpi-danger-bg-hover);border-color:var(--bpi-danger-border);color:var(--bpi-danger-text)}.bpi-mini.bpi-danger{background:var(--bpi-danger-bg);border-color:var(--bpi-danger-border);color:var(--bpi-danger-text)}
    .bpi-search{flex:1;min-width:140px;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 7px;outline:none}.bpi-search:focus{border-color:#4ca7e8}.bpi-searching{padding:5px 8px;color:var(--bpi-text-muted);font-size:10px}
-    .bpi-mode{border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 6px;font-size:11px;outline:none}.bpi-mode:focus{border-color:#4ca7e8}.bpi-mode-info{color:#87bfea}
-    .bpi-summary{color:var(--bpi-text-muted);font-size:11px;min-height:17px}.bpi-status{margin-left:auto}.bpi-status[data-kind="error"]{color:#ff8b93}.bpi-status[data-kind="ok"]{color:#7fdda2}.bpi-status[data-kind="busy"]{color:#ffd27a}
-    .bpi-issues{display:none;max-height:105px;overflow:auto;border:1px solid #5b4a36;border-radius:6px;background:rgba(55,39,25,.55);padding:4px 6px}.bpi-issues.bpi-visible{display:block}.bpi-issue{padding:2px 4px;color:#e6c792}.bpi-issue[data-severity="error"]{color:#ff9299}.bpi-issue[data-severity="info"]{color:#86c8eb}.bpi-issue::before{content:"⚠ ";}.bpi-issue[data-severity="error"]::before{content:"⛔ ";}.bpi-issue[data-severity="info"]::before{content:"ℹ ";}
-    .bpi-search-label{color:var(--bpi-text-muted);font-size:10px;white-space:nowrap}.bpi-results{display:none;max-height:220px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-2)}.bpi-results.bpi-visible{display:block}.bpi-result{display:grid;grid-template-columns:minmax(0,1fr) minmax(0,1fr) minmax(150px,auto);gap:7px;padding:5px 7px;border-bottom:1px solid var(--bpi-border-3);cursor:pointer}.bpi-result:last-child{border-bottom:0}.bpi-result:hover{background:var(--bpi-surface-3)}.bpi-result-en{color:var(--bpi-text)}.bpi-result-zh{color:#9ed0ff}.bpi-category{color:var(--bpi-text-muted);font-size:10px;white-space:normal}.bpi-search-reason{display:block;color:#77b6df;margin-top:2px}.bpi-search-more{display:block;margin:7px auto}.bpi-results>.bpi-searching{text-align:center}
-    .bpi-table{position:relative;height:300px;min-height:180px;max-height:360px;flex:1 1 300px;overflow-x:hidden;overflow-y:scroll;scrollbar-gutter:stable;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-table)}.bpi-split-handle{position:absolute;top:0;bottom:0;left:var(--bpi-split,40%);width:9px;margin-left:-4px;cursor:col-resize;z-index:4}.bpi-split-handle:hover{background:rgba(76,167,232,.35)}.bpi-head,.bpi-row{display:grid;grid-template-columns:var(--bpi-split,minmax(0,2fr)) minmax(0,1fr);gap:1px}.bpi-head{position:sticky;top:0;z-index:2;background:var(--bpi-head-bg);color:var(--bpi-text-muted);font-weight:700}.bpi-head>div,.bpi-cell{padding:6px 8px}.bpi-head>div+div,.bpi-cell+.bpi-cell{border-left:1px solid var(--bpi-border-2)}
-    .bpi-row{border-top:1px solid var(--bpi-border-3);cursor:pointer}.bpi-row:hover,.bpi-row.bpi-pinned{background:#314b68}.bpi-row:hover .bpi-cell,.bpi-row.bpi-pinned .bpi-cell{color:#fff}.bpi-row.bpi-unknown{background:rgba(112,70,28,.18)}.bpi-row.bpi-machine{background:rgba(85,62,131,.22)}.bpi-row.bpi-has-warning{box-shadow:inset 3px 0 #d39845}.bpi-row.bpi-has-error{box-shadow:inset 3px 0 #e35b66}
-    .bpi-cell{min-width:0;word-break:break-word;display:flex;align-items:flex-start;gap:5px}.bpi-en{color:var(--bpi-text);font-family:Consolas,monospace}.bpi-zh{color:#9ed0ff}.bpi-row:hover .bpi-en,.bpi-row:hover .bpi-zh,.bpi-row.bpi-pinned .bpi-en,.bpi-row.bpi-pinned .bpi-zh{color:#fff;text-shadow:0 0 7px rgba(109,190,255,.6)}
-    .bpi-badge{flex:none;border:1px solid var(--bpi-border-2);border-radius:8px;padding:0 5px;color:var(--bpi-text-muted);font-size:9px;line-height:15px}.bpi-source{border-color:#3e6f67;color:#86d5c4}.bpi-confidence-high{border-color:#3f785c;color:#85d5a6}.bpi-confidence-medium{border-color:#8a7538;color:#e2ca78}.bpi-confidence-low,.bpi-confidence-none{border-color:#7a4c55;color:#ee9ca8}.bpi-unknown .bpi-badge{color:#ffc277;border-color:#85602f}.bpi-machine .bpi-badge{color:#cbb1ff;border-color:#6b5594}.bpi-inline-actions{display:flex;gap:3px;margin-left:auto;flex-wrap:wrap}.bpi-mini{border:1px solid var(--bpi-border-2);border-radius:4px;background:var(--bpi-button-bg);color:var(--bpi-button-text);padding:1px 5px;cursor:pointer;font-size:9px}.bpi-mini:hover{background:#3c4b60}.bpi-inline-editor{min-width:90px;flex:1;border:1px solid #4ca7e8;border-radius:4px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:3px 5px}.bpi-empty{padding:24px;text-align:center;color:var(--bpi-text-faint)}
+    .bpi-mode{border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 6px;font-size:11px;outline:none}.bpi-mode:focus{border-color:#4ca7e8}.bpi-mode-info{color:var(--bpi-info-text)}
+    .bpi-summary{color:var(--bpi-text-muted);font-size:11px;min-height:17px}.bpi-status{margin-left:auto}.bpi-status[data-kind="error"]{color:var(--bpi-status-error)}.bpi-status[data-kind="ok"]{color:var(--bpi-status-ok)}.bpi-status[data-kind="busy"]{color:var(--bpi-status-busy)}
+    .bpi-issues{display:none;max-height:105px;overflow:auto;border:1px solid var(--bpi-issues-border);border-radius:6px;background:var(--bpi-issues-bg);padding:4px 6px}.bpi-issues.bpi-visible{display:block}.bpi-issue{padding:2px 4px;color:var(--bpi-issue-text)}.bpi-issue[data-severity="error"]{color:var(--bpi-status-error)}.bpi-issue[data-severity="info"]{color:var(--bpi-info-text)}.bpi-issue::before{content:"⚠ ";}.bpi-issue[data-severity="error"]::before{content:"⛔ ";}.bpi-issue[data-severity="info"]::before{content:"ℹ ";}
+    .bpi-search-label{color:var(--bpi-text-muted);font-size:10px;white-space:nowrap}.bpi-results{display:none;max-height:220px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-2)}.bpi-results.bpi-visible{display:block}.bpi-result{display:grid;grid-template-columns:minmax(0,1fr) minmax(0,1fr) minmax(150px,auto);gap:7px;padding:5px 7px;border-bottom:1px solid var(--bpi-border-3);cursor:pointer}.bpi-result:last-child{border-bottom:0}.bpi-result:hover{background:var(--bpi-surface-3)}.bpi-result-en{color:var(--bpi-text)}.bpi-result-zh{color:var(--bpi-zh-text)}.bpi-category{color:var(--bpi-text-muted);font-size:10px;white-space:normal}.bpi-search-reason{display:block;color:var(--bpi-info-text);margin-top:2px}.bpi-search-more{display:block;margin:7px auto}.bpi-results>.bpi-searching{text-align:center}
+    .bpi-table{position:relative;height:300px;min-height:180px;max-height:360px;flex:1 1 300px;overflow-x:hidden;overflow-y:scroll;scrollbar-gutter:stable;border:1px solid var(--bpi-border-2);border-radius:0 0 6px 6px;background:var(--bpi-surface-table)}.bpi-split-handle{position:absolute;top:0;bottom:0;left:var(--bpi-split,40%);width:9px;margin-left:-4px;cursor:col-resize;z-index:4}.bpi-split-handle:hover{background:rgba(76,167,232,.35)}.bpi-head,.bpi-row{display:grid;grid-template-columns:var(--bpi-split,minmax(0,2fr)) minmax(0,1fr);gap:1px}.bpi-head{position:sticky;top:0;z-index:2;background:var(--bpi-head-bg);color:var(--bpi-text-muted);font-weight:700}.bpi-head>div,.bpi-cell{padding:6px 8px}.bpi-head>div+div,.bpi-cell+.bpi-cell{border-left:1px solid var(--bpi-border-2)}
+    .bpi-row{border-top:1px solid var(--bpi-border-3);cursor:pointer}.bpi-row:hover{background-image:linear-gradient(var(--bpi-row-hover-overlay),var(--bpi-row-hover-overlay));outline:1px dashed var(--bpi-row-hover-ring);outline-offset:-1px}.bpi-row.bpi-pinned{background-image:linear-gradient(var(--bpi-row-pinned-overlay),var(--bpi-row-pinned-overlay));outline:1px solid var(--bpi-row-pinned-ring);outline-offset:-1px;font-weight:500}.bpi-row:hover .bpi-cell,.bpi-row.bpi-pinned .bpi-cell{color:var(--bpi-row-hover-text)}.bpi-row.bpi-unknown{background-color:rgba(112,70,28,.18)}.bpi-row.bpi-machine{background-color:rgba(85,62,131,.22)}.bpi-row.bpi-suspected-natural{background-color:rgba(58,132,160,.20)}.bpi-row.bpi-has-warning{box-shadow:inset 3px 0 #d39845}.bpi-row.bpi-has-error{box-shadow:inset 3px 0 #e35b66}
+    .bpi-cell{min-width:0;word-break:break-word;display:flex;align-items:flex-start;gap:5px}.bpi-en{color:var(--bpi-text);font-family:Consolas,monospace}.bpi-zh{color:var(--bpi-zh-text)}.bpi-row:hover .bpi-en,.bpi-row:hover .bpi-zh,.bpi-row.bpi-pinned .bpi-en,.bpi-row.bpi-pinned .bpi-zh{color:var(--bpi-row-hover-text);text-shadow:var(--bpi-row-hover-shadow)}
+    .bpi-badge{flex:none;border:1px solid var(--bpi-border-2);border-radius:8px;padding:0 5px;color:var(--bpi-text-muted);font-size:9px;line-height:15px}.bpi-source{border-color:#3e6f67;color:var(--bpi-badge-source)}.bpi-suspected{border-color:#3a84a0;color:var(--bpi-badge-suspected)}.bpi-confidence-high{border-color:#3f785c;color:var(--bpi-status-ok)}.bpi-confidence-medium{border-color:#8a7538;color:var(--bpi-badge-warn)}.bpi-confidence-low,.bpi-confidence-none{border-color:#7a4c55;color:var(--bpi-status-error)}.bpi-unknown .bpi-badge{color:var(--bpi-badge-warn);border-color:#85602f}.bpi-machine .bpi-badge{color:var(--bpi-badge-machine);border-color:#6b5594}.bpi-inline-actions{display:flex;gap:3px;margin-left:auto;flex-wrap:wrap}.bpi-mini{border:1px solid var(--bpi-border-2);border-radius:4px;background:var(--bpi-button-bg);color:var(--bpi-button-text);padding:1px 5px;cursor:pointer;font-size:9px}.bpi-mini:hover{background:var(--bpi-mini-hover-bg)}.bpi-inline-editor{min-width:90px;flex:1;border:1px solid #4ca7e8;border-radius:4px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:3px 5px}.bpi-empty{padding:24px;text-align:center;color:var(--bpi-text-faint)}
    .bpi-filters{display:flex;gap:4px;align-items:center;flex-wrap:wrap}
-.bpi-batch-bar{flex:none;gap:8px;padding:6px 8px;margin-bottom:6px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3)}.bpi-filter-active{background:#365f84;border-color:#5792c4;color:#fff}.bpi-star{color:#8f9aaa}.bpi-star.bpi-starred{color:#ffd45f;border-color:#8f772f}.bpi-result.bpi-result-selected{background:#314b68;outline:1px solid #5688b5}.bpi-result-star{font-size:14px;color:#8995a8;align-self:center}.bpi-result-star.bpi-starred{color:#ffd45f}
-    .bpi-quick-section{flex:none;border:1px solid #6b5594;border-radius:7px;background:var(--bpi-surface);overflow:hidden}.bpi-quick-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:#d9ccff;font-weight:700;flex-wrap:wrap}.bpi-quick-hint{margin-left:auto;color:#a795c6;font-size:10px;font-weight:400}.bpi-quick-row{display:flex;gap:6px;align-items:flex-start;padding:6px 8px}.bpi-quick-input{flex:1;min-width:0;box-sizing:border-box;height:44px;min-height:44px;max-height:140px;overflow-y:auto;resize:vertical;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 7px;outline:none;font:12px/1.55 Arial,sans-serif}.bpi-quick-input:focus{box-shadow:inset 0 0 0 1px #a98be0}.bpi-quick-row .bpi-button{flex:none}
-    .bpi-drag-handle{flex:none;cursor:grab;color:#7194ad;font-size:11px;line-height:1;user-select:none;padding:1px 2px;border-radius:3px}.bpi-drag-handle:hover{color:#a9c8e4;background:#2b3a4d}.bpi-drag-handle:active{cursor:grabbing}
+    .bpi-table-search{flex:none;display:flex;align-items:center;gap:6px;box-sizing:border-box;padding:4px 6px;margin-bottom:-7px;border:1px solid var(--bpi-border-2);border-bottom:0;border-radius:6px 6px 0 0;background-color:var(--bpi-head-bg)}.bpi-table-search-icon{flex:none;display:inline-flex;color:var(--bpi-text-faint)}.bpi-search.bpi-table-search-input{flex:1;min-width:0;padding:3px 6px}.bpi-table-search-count{flex:none;color:var(--bpi-text-muted);font-size:10px;white-space:nowrap}.bpi-table-search-clear{flex:none;display:none;border:0;background-color:transparent;color:var(--bpi-text-muted);cursor:pointer;font-size:13px;line-height:1;padding:1px 3px;border-radius:3px}.bpi-table-search-clear.bpi-visible{display:block}.bpi-table-search-clear:hover{color:var(--bpi-text);background-color:var(--bpi-border-3)}.bpi-hit{background-color:var(--bpi-hit-bg);color:var(--bpi-hit-text);border-radius:2px}
+.bpi-batch-bar{flex:none;gap:8px;padding:6px 8px;margin-bottom:6px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3)}.bpi-filter-active{background:#365f84;border-color:#5792c4;color:#fff}.bpi-star{color:var(--bpi-star-off)}.bpi-star.bpi-starred{color:var(--bpi-star-on);border-color:#8f772f}.bpi-result.bpi-result-selected{background:var(--bpi-row-selected-bg);outline:1px solid var(--bpi-link-text)}.bpi-result-star{font-size:14px;color:var(--bpi-star-off);align-self:center}.bpi-result-star.bpi-starred{color:var(--bpi-star-on)}
+    .bpi-quick-section{flex:none;border:1px solid #6b5594;border-radius:7px;background-color:var(--bpi-quick-section-bg);overflow:hidden}.bpi-quick-head{display:flex;align-items:center;gap:7px;padding:5px 8px;background:var(--bpi-head-bg);color:var(--bpi-quick-head-text);font-weight:700;flex-wrap:wrap}.bpi-quick-hint{margin-left:auto;color:var(--bpi-quick-hint-text);font-size:10px;font-weight:400}.bpi-quick-target{padding:0 8px 6px;font-size:11px;color:var(--bpi-text-muted)}.bpi-quick-target-linked{color:var(--bpi-link-text)}.bpi-quick-row{display:flex;gap:6px;align-items:flex-start;padding:6px 8px}.bpi-quick-input{flex:1;min-width:0;box-sizing:border-box;height:44px;min-height:44px;max-height:140px;overflow-y:auto;resize:vertical;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 7px;outline:none;font:12px/1.55 Arial,sans-serif}.bpi-quick-input:focus{box-shadow:inset 0 0 0 1px #a98be0}.bpi-quick-row .bpi-button{flex:none}
+    .bpi-drag-handle{flex:none;cursor:grab;color:var(--bpi-handle-text);font-size:11px;line-height:1;user-select:none;padding:1px 2px;border-radius:3px}.bpi-drag-handle:hover{color:var(--bpi-handle-hover);background:var(--bpi-handle-hover-bg)}.bpi-drag-handle:active{cursor:grabbing}
    .bpi-row.bpi-dragging,.bpi-english-token.bpi-dragging{opacity:.45}.bpi-english-token.bpi-dragging{cursor:grabbing}
    .bpi-drop-indicator{position:absolute;left:3px;right:3px;height:0;border-top:2px solid #76c9ff;box-shadow:0 0 5px rgba(118,201,255,.55);z-index:3;pointer-events:none}
-    .bpi-break-row{display:flex;align-items:center;gap:8px;border-top:1px solid var(--bpi-border-3);background:#0f6e56;color:#bff0dd;padding:6px 10px;cursor:default}
+    .bpi-break-row{display:flex;align-items:center;gap:8px;border-top:1px solid var(--bpi-border-3);background:var(--bpi-break-bg);color:var(--bpi-break-text);padding:6px 10px;cursor:default}
+    .bpi-row.bpi-break-row:hover{background-color:var(--bpi-break-bg-hover);background-image:none;outline:0;color:#fff}
    .bpi-break-row.bpi-dragging{opacity:.45}
-    .bpi-break-row .bpi-drag-handle{color:#bff0dd}.bpi-break-row .bpi-drag-handle:hover{color:#fff;background:#0a5240}
+    .bpi-break-row .bpi-drag-handle{color:var(--bpi-break-text)}.bpi-break-row .bpi-drag-handle:hover{color:#fff;background:var(--bpi-break-bg-hover)}
    .bpi-break-row-label{flex:1;font-size:12px;font-weight:500;letter-spacing:.5px}
    .bpi-mirror-token{user-select:none}.bpi-english-token.bpi-drop-before,.bpi-mirror-token.bpi-drop-before{box-shadow:-2px 0 0 0 #76c9ff,0 0 5px rgba(118,201,255,.4)}.bpi-english-token.bpi-drop-after,.bpi-mirror-token.bpi-drop-after{box-shadow:2px 0 0 0 #76c9ff,0 0 5px rgba(118,201,255,.4)}
-    .bpi-english-token{position:relative}.bpi-hide-btn{flex:none;cursor:pointer;color:#7194ad;font-size:11px;line-height:1;user-select:none;padding:1px 2px;border-radius:3px;display:inline-flex;align-items:center}.bpi-hide-btn:hover{color:#9ed0ff;background:#2b3a4d}.bpi-eye-icon{display:inline-flex;align-items:center}
+    .bpi-english-token{position:relative}.bpi-hide-btn{flex:none;cursor:pointer;color:var(--bpi-handle-text);font-size:11px;line-height:1;user-select:none;padding:1px 2px;border-radius:3px;display:inline-flex;align-items:center}.bpi-hide-btn:hover{color:var(--bpi-link-text);background:var(--bpi-handle-hover-bg)}.bpi-eye-icon{display:inline-flex;align-items:center}
 .bpi-card-hidden{opacity:.55;filter:grayscale(.7)}
 .bpi-card-hidden:hover{opacity:.75}
 .bpi-card-hidden .bpi-english-token{text-decoration:line-through}
 .bpi-card-hidden .bpi-token-card-zh{color:var(--bpi-text-faint)}
 .bpi-card-hidden .bpi-chip-hide{display:inline-flex}
-.bpi-chip-hidden-mark{color:#e8b06a}
+.bpi-chip-hidden-mark{color:var(--bpi-badge-warn)}
 .bpi-row-hidden{opacity:.6}
 .bpi-row-hidden:hover{opacity:.8}
 .bpi-hidden-term{text-decoration:line-through;color:var(--bpi-text-faint)}
-.bpi-hide-btn-off{color:#e8b06a}
-.bpi-eye-off{color:#e8b06a}
-    .bpi-chip-hide{position:absolute;top:-5px;right:-5px;display:none;cursor:pointer;padding:1px 2px;border-radius:999px;background:#1d2937;border:1px solid #4c5668;color:#9ed0ff;line-height:1;z-index:2}.bpi-english-token:hover .bpi-chip-hide{display:inline-flex}.bpi-chip-hide:hover{background:#285f8e;border-color:#76c9ff;color:#fff}
+.bpi-hide-btn-off{color:var(--bpi-badge-warn)}
+.bpi-eye-off{color:var(--bpi-badge-warn)}
+    .bpi-chip-hide{position:absolute;top:2px;right:3px;display:none;cursor:pointer;padding:1px 2px;border-radius:999px;background:var(--bpi-chip-hide-bg);border:1px solid var(--bpi-border-2);color:var(--bpi-link-text);line-height:1;z-index:2}.bpi-token-card:hover .bpi-chip-hide{display:inline-flex}.bpi-chip-hide:hover{background:var(--bpi-primary-bg);border-color:var(--bpi-primary-border);color:#fff}
 .bpi-token-line{display:flex;flex-wrap:wrap;align-items:flex-end;gap:2px 0}
-    .bpi-token-card{display:inline-flex;flex-direction:column;align-items:flex-start;max-width:100%;border:1px solid var(--bpi-border-3);border-radius:5px;background:var(--bpi-surface-3);padding:1px 5px;margin:2px 3px 2px 0;vertical-align:top;cursor:pointer}
+    .bpi-token-card{position:relative;display:inline-flex;flex-direction:column;align-items:flex-start;max-width:100%;border:1px solid var(--bpi-border-3);border-radius:5px;background:var(--bpi-surface-3);padding:1px 5px;margin:2px 3px 2px 0;vertical-align:top;cursor:pointer}
    .bpi-token-card:hover{border-color:var(--bpi-border-2)}
    .bpi-token-card.bpi-dragging{opacity:.45;cursor:grabbing}
    .bpi-token-card .bpi-english-token{margin:0;padding:0}
@@ -813,37 +1291,41 @@ function injectBpiStyles() {
    .bpi-token-card.bpi-card-linked .bpi-english-token{background:transparent;border-color:transparent;box-shadow:none;color:#fff}
    .bpi-token-card-zh{min-height:15px;max-width:200px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;color:var(--bpi-text-muted);font:11px/1.35 Arial,sans-serif}
    .bpi-category-content{display:flex;flex-wrap:wrap;gap:2px 3px}
-.bpi-break-chip{position:relative;display:inline-flex;align-items:center;gap:3px;border:1px dashed #6c83a8;border-radius:5px;background:#1d2937;color:#9ed0ff;padding:0 5px;margin:1px 0;cursor:grab;user-select:none;font-family:Consolas,monospace}
-.bpi-break-chip:hover{background:#285f8e;border-color:#76c9ff;color:#fff}
+.bpi-break-chip{position:relative;display:inline-flex;align-items:center;gap:3px;border:1px dashed var(--bpi-break-border);border-radius:5px;background:var(--bpi-break-bg);color:var(--bpi-break-text);padding:0 5px;margin:1px 0;cursor:grab;user-select:none;font-family:Consolas,monospace}
+.bpi-break-chip:hover,.bpi-break-chip.bpi-drop-target{background:var(--bpi-break-bg-hover);border-color:var(--bpi-break-text);color:#fff}
 .bpi-break-chip.bpi-dragging{opacity:.45;cursor:grabbing}
-.bpi-break-chip.bpi-drop-target{background:#285f8e;border-color:#76c9ff;color:#fff}
-.bpi-break-x{display:none;cursor:pointer;color:#9ed0ff;line-height:1}
+.bpi-break-x{display:none;cursor:pointer;line-height:1}
 .bpi-break-chip:hover .bpi-break-x{display:inline-flex}
 .bpi-break-x:hover{color:#fff}
 .bpi-source-bar{display:flex;gap:7px;align-items:center;flex-wrap:wrap;padding:5px 7px;margin-bottom:7px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3)}
-.bpi-source-bar.bpi-source-waiting{border-color:#8a5f16;background:rgba(70,52,25,.5)}
-.bpi-source-label{color:#87bfea;font-size:11px}
+.bpi-source-bar.bpi-source-waiting{border-color:var(--bpi-waiting-border);background:var(--bpi-waiting-bg)}
+.bpi-source-label{color:var(--bpi-info-text);font-size:11px}
 .bpi-source-state{color:var(--bpi-text-muted);font-size:11px}
-.bpi-source-state.bpi-source-waiting-text{color:#ffd27a}
+.bpi-source-state.bpi-source-waiting-text{color:var(--bpi-badge-warn)}
 .bpi-source-toggle{display:inline-flex;gap:5px;align-items:center;font-size:11px;color:var(--bpi-text);cursor:pointer;user-select:none}
 .bpi-source-toggle input{margin:0;accent-color:#3f83ba}
 .bpi-source-select{border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:3px 5px;font-size:11px}
 .bpi-source-select:disabled{opacity:.45}
 .bpi-save-drop{position:relative;display:flex;flex-direction:column;gap:5px;align-items:center;justify-content:center;min-height:86px;border:1px dashed var(--bpi-border-2);border-radius:7px;background:var(--bpi-surface-3);color:var(--bpi-text-faint);font-size:11px;text-align:center;cursor:pointer}
-.bpi-save-drop.bpi-dragover{border-color:#4ca7e8;color:#9ed0ff}
+.bpi-save-drop.bpi-dragover{border-color:#4ca7e8;color:var(--bpi-link-text)}
    .bpi-autocomplete{position:relative}
    .bpi-chip-row{display:flex;flex-wrap:wrap;gap:5px;align-items:center;box-sizing:border-box;width:100%;min-height:30px;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);padding:5px 6px}
    .bpi-form .bpi-chip-row input{flex:1;min-width:90px;width:auto;border:0;outline:none;background:transparent;color:var(--bpi-input-text);padding:2px 0;font:12px Arial,sans-serif}
    .bpi-chip{display:inline-flex;align-items:center;gap:4px;max-width:100%;padding:2px 5px 2px 8px;border:1px solid var(--bpi-border-2);border-radius:11px;background:var(--bpi-surface-3);color:var(--bpi-text);font:11px Arial,sans-serif}
    .bpi-chip-label{overflow:hidden;text-overflow:ellipsis;white-space:nowrap;max-width:160px}
    .bpi-chip-x{border:0;background:transparent;color:var(--bpi-text-muted);cursor:pointer;font:12px Arial,sans-serif;line-height:1;padding:0 1px}
-    .bpi-chip-x:hover{color:#ff8b93}
+    .bpi-chip-x:hover{color:var(--bpi-status-error)}
    .bpi-ac-menu{position:absolute;left:0;right:0;top:calc(100% + 4px);z-index:95;max-height:188px;overflow:auto;padding:3px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-modal);box-shadow:0 6px 16px rgba(0,0,0,.25)}
    .bpi-ac-menu[hidden]{display:none}
    .bpi-ac-item{display:flex;align-items:center;gap:8px;padding:5px 7px;border-radius:4px;color:var(--bpi-text);font:12px Arial,sans-serif;cursor:pointer}
    .bpi-ac-item:hover,.bpi-ac-item.bpi-ac-active{background:var(--bpi-surface-3)}
    .bpi-ac-name{overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
    .bpi-ac-kind{margin-left:auto;flex:none;color:var(--bpi-text-faint);font-size:10px}
+    .bpi-ac-custom{color:var(--bpi-custom-text);border-bottom:1px solid var(--bpi-border-3);border-radius:4px 4px 0 0}
+    .bpi-input-wrap{display:flex;align-items:center;gap:5px;box-sizing:border-box;width:100%;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);padding:5px 6px}
+    .bpi-input-wrap:focus-within{border-color:#4ca7e8}
+    .bpi-form .bpi-input-wrap input{flex:1;min-width:0;width:auto;border:0;outline:none;background:transparent;color:var(--bpi-input-text);padding:0;font:12px Arial,sans-serif}
+    .bpi-custom-badge{flex:none;border:1px solid var(--bpi-custom-border);border-radius:8px;background:var(--bpi-custom-bg);color:var(--bpi-custom-text);font-size:10px;line-height:15px;padding:0 5px}
    .bpi-form-hint{color:var(--bpi-text-faint);font-size:10px}
    .bpi-model-tags{display:flex;flex-wrap:wrap;gap:4px}
    .bpi-model-tag{padding:1px 7px;border:1px solid var(--bpi-border-2);border-radius:10px;background:var(--bpi-surface-3);color:var(--bpi-text-muted);font-size:10px}
@@ -858,61 +1340,132 @@ function injectBpiStyles() {
 .bpi-fav-meta{display:flex;flex-wrap:wrap;gap:5px;align-items:center;color:var(--bpi-text-faint);font-size:10px}
 .bpi-fav-actions{display:flex;flex-wrap:wrap;gap:5px;margin-top:2px}
    .bpi-hidden-bar{display:flex;align-items:center;gap:6px;flex-wrap:wrap;padding:5px 2px;border-top:1px solid var(--bpi-border-3)}.bpi-hidden-bar.bpi-hidden{display:none}.bpi-hidden-label{flex:none;color:var(--bpi-text-muted);font-size:11px}
-    .bpi-hidden-chip{display:inline-flex;align-items:center;gap:5px;max-width:280px;overflow:hidden;border:1px dashed var(--bpi-border-2);border-radius:999px;padding:2px 8px;background:var(--bpi-surface-3);color:var(--bpi-text-muted);cursor:pointer;font-size:11px}.bpi-hidden-chip:hover{background:#31404f;border-color:#76c9ff;color:#d7e6f7}.bpi-hidden-en{text-decoration:line-through;white-space:nowrap}.bpi-hidden-zh{color:var(--bpi-text-muted);font-size:10px}.bpi-hidden-restore{color:#76c9ff;font-size:12px}
+    .bpi-hidden-chip{display:inline-flex;align-items:center;gap:5px;max-width:280px;overflow:hidden;border:1px dashed var(--bpi-border-2);border-radius:999px;padding:2px 8px;background:var(--bpi-surface-3);color:var(--bpi-text-muted);cursor:pointer;font-size:11px}.bpi-hidden-chip:hover{background:var(--bpi-hidden-chip-hover-bg);border-color:var(--bpi-link-text);color:var(--bpi-row-hover-text)}.bpi-hidden-en{text-decoration:line-through;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;min-width:0}.bpi-hidden-zh{color:var(--bpi-text-muted);font-size:10px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;min-width:0}.bpi-hidden-restore{color:var(--bpi-link-text);font-size:12px}
    .bpi-hidden-all{margin-left:auto}
-    .bpi-modal-shade{position:fixed;inset:0;z-index:10020;background:rgba(0,0,0,.58);display:flex;align-items:center;justify-content:center}.bpi-modal{width:min(520px,calc(100vw - 32px));max-height:calc(100vh - 40px);overflow:auto;background:var(--bpi-surface-modal);color:var(--bpi-text);border:1px solid var(--bpi-border-2);border-radius:10px;padding:16px;box-shadow:0 18px 55px rgba(0,0,0,.55)}.bpi-modal h3{margin:0 0 12px;font-size:16px}.bpi-form{display:grid;grid-template-columns:92px minmax(0,1fr);gap:9px;align-items:center}.bpi-form label{color:var(--bpi-text-muted)}.bpi-form input,.bpi-form select,.bpi-form textarea{box-sizing:border-box;width:100%;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:6px 7px}.bpi-form textarea{min-height:112px;resize:vertical;font:11px/1.45 Arial,sans-serif}.bpi-form input:focus,.bpi-form select:focus,.bpi-form textarea:focus{outline:none;border-color:#4ca7e8}.bpi-modal-actions{display:flex;justify-content:flex-end;gap:7px;margin-top:14px;flex-wrap:wrap}.bpi-weight-presets{display:flex;gap:5px;flex-wrap:wrap;margin-top:10px}
+    .bpi-modal-shade{position:fixed;inset:0;z-index:10020;background:rgba(0,0,0,.58);display:flex;align-items:center;justify-content:center}.bpi-modal{width:min(520px,calc(100vw - 32px));max-height:calc(100vh - 40px);overflow:auto;background:var(--bpi-surface-modal);color:var(--bpi-text);border:1px solid var(--bpi-border-2);border-radius:10px;padding:16px;box-shadow:0 18px 55px rgba(0,0,0,.55)}.bpi-modal h3{margin:0 0 12px;font-size:16px}.bpi-form{display:grid;grid-template-columns:92px minmax(0,1fr);gap:9px;align-items:center}.bpi-form [hidden]{display:none!important}.bpi-form label{color:var(--bpi-text-muted)}.bpi-form input,.bpi-form select,.bpi-form textarea{box-sizing:border-box;width:100%;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:6px 7px}.bpi-form textarea{min-height:112px;resize:vertical;font:11px/1.45 Arial,sans-serif}.bpi-form input:focus,.bpi-form select:focus,.bpi-form textarea:focus{outline:none;border-color:#4ca7e8}.bpi-modal-actions{display:flex;justify-content:flex-end;gap:7px;margin-top:14px;flex-wrap:wrap}.bpi-weight-presets{display:flex;gap:5px;flex-wrap:wrap;margin-top:10px}
 .bpi-editor-tabs{display:flex;gap:0;border-bottom:1px solid var(--bpi-border);margin-bottom:12px}
 .bpi-editor-tab{background:none;border:none;border-bottom:2px solid transparent;padding:5px 12px;color:var(--bpi-text-muted);font:12px Arial,sans-serif;cursor:pointer}
 .bpi-editor-tab:hover{color:var(--bpi-text)}
-.bpi-editor-tab.bpi-editor-tab-active{color:#76c9ff;border-bottom-color:#76c9ff}
+.bpi-editor-tab.bpi-editor-tab-active{color:var(--bpi-link-text);border-bottom-color:var(--bpi-link-text)}
 .bpi-editor-panel{display:block}
+.bpi-editor-divider{height:1px;background:var(--bpi-border-3);margin:12px 0}
+.bpi-form .bpi-field-static{box-sizing:border-box;padding:6px 0;color:var(--bpi-text);font-family:Consolas,monospace;word-break:break-all}
+.bpi-raw-editor{box-sizing:border-box;width:100%;min-height:72px;max-height:220px;resize:vertical;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:5px 7px;outline:none;font:12px/1.55 Consolas,monospace}
+.bpi-raw-editor:focus{border-color:#4ca7e8}
+.bpi-token-card.bpi-card-multi{position:relative}
+.bpi-row.bpi-row-multi{position:relative}
+/* 多选的虚线框做成会流动的「蚂蚁线」：四边各一条虚线渐变，靠 background-position 走一个周期，
+   看着就是虚线绕着选中的提示词转。用伪元素画，不占元素自己的 background-image（行选中态要用它） */
+.bpi-card-multi::after,.bpi-row-multi::after{content:"";position:absolute;border-radius:inherit;pointer-events:none;background-image:repeating-linear-gradient(90deg,var(--bpi-link-text) 0 4px,transparent 4px 8px),repeating-linear-gradient(90deg,var(--bpi-link-text) 0 4px,transparent 4px 8px),repeating-linear-gradient(180deg,var(--bpi-link-text) 0 4px,transparent 4px 8px),repeating-linear-gradient(180deg,var(--bpi-link-text) 0 4px,transparent 4px 8px);background-size:100% 1px,100% 1px,1px 100%,1px 100%;background-position:0 0,0 100%,0 0,100% 0;background-repeat:repeat-x,repeat-x,repeat-y,repeat-y;animation:bpi-ants .6s linear infinite}
+.bpi-card-multi::after{inset:-2px}
+.bpi-row-multi::after{inset:0}
+@keyframes bpi-ants{to{background-position:8px 0,-8px 100%,0 -8px,100% 8px}}
 .bpi-color-presets{display:flex;gap:6px;flex-wrap:wrap;margin-bottom:10px}
 .bpi-color-swatch{width:26px;height:26px;border-radius:4px;border:1px solid var(--bpi-border);cursor:pointer;display:inline-block}
 .bpi-color-swatch:hover{border-color:var(--bpi-border-2)}
-.bpi-color-swatch.bpi-color-swatch-active{border-color:#76c9ff;box-shadow:0 0 0 1px #76c9ff}
+.bpi-color-swatch.bpi-color-swatch-active{border-color:var(--bpi-link-text);box-shadow:0 0 0 1px var(--bpi-link-text)}
 .bpi-color-swatch-none.bpi-color-swatch-active{border-style:solid;color:var(--bpi-text)}
 .bpi-color-custom{display:flex;align-items:center;gap:6px}
 .bpi-color-custom input{width:150px}
 .bpi-color-swatch-none{display:inline-flex;align-items:center;justify-content:center;background:var(--bpi-surface-3);border-style:dashed;border-color:var(--bpi-border-2);color:var(--bpi-text-muted);font:11px Arial,sans-serif}
+    .bpi-color-grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(196px,1fr));gap:6px}
+    .bpi-color-card{border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-2);padding:7px}
+    .bpi-color-card-head{display:flex;align-items:center;gap:6px;font-weight:700;color:var(--bpi-text);word-break:break-word}
+    .bpi-color-chip{width:13px;height:13px;border-radius:3px;border:1px solid var(--bpi-border-2);background:transparent;flex:none}
+    .bpi-color-meta{font-size:10px;color:var(--bpi-text-muted);margin-top:2px}
+    .bpi-color-swatches{display:grid;grid-template-columns:repeat(4,1fr);gap:5px;margin-top:6px}
+    .bpi-color-swatches i{height:20px;border-radius:4px;border:1px solid var(--bpi-border);cursor:pointer;display:block}
+    .bpi-color-swatches i.bpi-swatch-on{border-color:var(--bpi-link-text);box-shadow:0 0 0 1px var(--bpi-link-text)}
+    .bpi-color-row{display:flex;align-items:center;gap:5px;margin-top:6px}
+    .bpi-color-row input{width:92px;border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:3px 5px;font-size:11px}
+    .bpi-color-dice{width:24px;height:24px;display:inline-flex;align-items:center;justify-content:center;border:1px solid var(--bpi-border-2);border-radius:5px;background:transparent;color:var(--bpi-text-muted);cursor:pointer;padding:0;flex:none}
+    .bpi-color-dice:hover{color:var(--bpi-text)}
+    .bpi-color-error{color:var(--bpi-status-error);font-size:11px;margin-top:6px}
+    .bpi-color-panel{border:1px solid var(--bpi-border-2);border-radius:7px;background:var(--bpi-surface-3);padding:8px;margin:5px 0}
+    .bpi-color-panel-actions{display:flex;justify-content:flex-end;margin-top:8px}
+    .bpi-color-modes{display:flex;gap:6px;flex-wrap:wrap}
+    .bpi-color-mode{display:inline-flex;align-items:center;gap:5px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-2);color:var(--bpi-text);padding:5px 9px;cursor:pointer;font-size:12px;font-family:inherit}
+    .bpi-color-mode:hover{border-color:var(--bpi-text-muted)}
+    .bpi-color-mode-active{border-color:var(--bpi-link-text);box-shadow:0 0 0 1px var(--bpi-link-text)}
+    .bpi-color-mode-dot{width:9px;height:9px;border-radius:50%;display:inline-block;border:1px solid var(--bpi-border)}
+    .bpi-color-note{font-size:11px;color:var(--bpi-text-muted);margin:10px 0 6px}
+    .bpi-color-open{border-color:var(--bpi-link-text);box-shadow:0 0 0 1px var(--bpi-link-text)}
 .bpi-filter-sep{width:1px;height:18px;background:var(--bpi-border);margin:0 4px;align-self:center}
-.bpi-color-label{display:inline-flex;align-items:center;gap:5px;font-size:12px}
-.bpi-dropdown{position:relative;display:inline-flex}
 .bpi-dropdown-trigger{display:inline-flex;align-items:center;gap:6px;padding:3px 9px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-button-bg);color:var(--bpi-button-text);font:12px Arial,sans-serif;line-height:1.2;cursor:pointer;user-select:none}
 .bpi-dropdown-trigger:hover{border-color:var(--bpi-text-faint)}
 .bpi-dropdown-caret{font-size:9px;color:var(--bpi-text-muted);transition:transform .12s}
-.bpi-dropdown-open .bpi-dropdown-caret{transform:rotate(180deg)}
-.bpi-dropdown-menu{position:absolute;top:calc(100% + 5px);left:0;z-index:90;min-width:150px;padding:4px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-modal);box-shadow:0 6px 16px rgba(0,0,0,.25)}
-.bpi-dropdown-menu.bpi-hidden{display:none}
-.bpi-dropdown-item{display:flex;align-items:center;gap:8px;padding:5px 8px;border-radius:4px;color:var(--bpi-text);font:12px Arial,sans-serif;white-space:nowrap;cursor:pointer}
-.bpi-dropdown-item:hover{background:var(--bpi-surface-3)}
-.bpi-dropdown-item-active{color:#76c9ff}
-.bpi-dropdown-dots{display:inline-flex;gap:3px}
-.bpi-dropdown-dot{width:9px;height:9px;border-radius:50%;display:inline-block}
-.bpi-dropdown-check{margin-left:auto;font-size:11px;color:#76c9ff}.bpi-about-modal{width:min(460px,calc(100vw - 32px))}.bpi-about-name{margin-bottom:7px;color:var(--bpi-text);font-weight:700}.bpi-config-note{color:var(--bpi-text-muted);font-size:10px;word-break:break-all}
+.bpi-about-modal{width:min(460px,calc(100vw - 32px))}.bpi-about-name{margin-bottom:7px;color:var(--bpi-text);font-weight:700}.bpi-config-note{color:var(--bpi-text-muted);font-size:10px;word-break:break-all}
+/* 同一句说明变成警告时用：只换颜色，不动字号字重，免得换行数变了把下面的表单顶走 */
+.bpi-config-note-warn{color:var(--bpi-status-error)}
+/* 「获取模型」：输入框照旧独占一行宽度，按钮挤在右边不换行 */
+.bpi-model-line{display:flex;gap:6px;align-items:stretch;min-width:0}
+.bpi-model-line input{flex:1 1 auto;min-width:0}
+.bpi-model-line .bpi-button{flex:none;white-space:nowrap;background-color:var(--bpm-accent);border-color:var(--bpm-accent-border);color:#fff}
+.bpi-model-line .bpi-button:hover{background-color:var(--bpm-accent-hover);border-color:var(--bpm-accent-border);color:#fff}
 .bpi-help{position:relative;display:inline-flex;align-items:center;justify-content:center;width:13px;height:13px;margin-left:5px;border:1px solid var(--bpi-border);border-radius:50%;color:var(--bpi-text-muted);font:10px/1 Arial,sans-serif;vertical-align:middle;cursor:help;user-select:none}
 .bpi-help:hover,.bpi-help:focus{color:var(--bpi-text);border-color:var(--bpi-border-2)}
 .bpi-help-pop{position:absolute;left:50%;bottom:calc(100% + 6px);transform:translateX(-50%);z-index:80;display:none;box-sizing:border-box;width:290px;max-width:70vw;padding:7px 9px;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-modal);color:var(--bpi-text-muted);font:11px/1.65 Arial,sans-serif;white-space:pre-line;text-align:left;cursor:default}
-.bpi-help:hover .bpi-help-pop,.bpi-help:focus .bpi-help-pop{display:block}.bpi-preview-text{box-sizing:border-box;width:100%;min-height:130px;max-height:300px;resize:vertical;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3);color:var(--bpi-text);padding:8px;font:12px/1.5 monospace}.bpi-sort-groups{max-height:230px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3);padding:6px;margin:8px 0}.bpi-sort-group{display:grid;grid-template-columns:145px minmax(0,1fr);gap:7px;padding:4px;border-top:1px solid var(--bpi-border-3)}.bpi-sort-group:first-child{border-top:0}.bpi-sort-group strong{color:#9fd3ff}
-    .bpi-about-footer{flex:none;min-height:20px;display:flex;align-items:center;justify-content:space-between;gap:8px;padding:1px 4px 0;color:var(--bpi-text-faint);font-size:9px}.bpi-about-button{border:0;background:transparent;color:var(--bpi-text-muted);padding:1px 3px;cursor:pointer;font-size:9px}.bpi-about-button:hover{color:#b8dcfa;text-decoration:underline}
+.bpi-help:hover .bpi-help-pop,.bpi-help:focus .bpi-help-pop{display:block}.bpi-preview-text{box-sizing:border-box;width:100%;min-height:130px;max-height:300px;resize:vertical;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3);color:var(--bpi-text);padding:8px;font:12px/1.5 monospace}.bpi-sort-groups{max-height:230px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-3);padding:6px;margin:8px 0}.bpi-sort-group{display:grid;grid-template-columns:145px minmax(0,1fr);gap:7px;padding:4px;border-top:1px solid var(--bpi-border-3)}.bpi-sort-group:first-child{border-top:0}.bpi-sort-group strong{color:var(--bpi-zh-text)}
+    .bpi-about-footer{flex:none;min-height:20px;display:flex;align-items:center;justify-content:space-between;gap:8px;padding:1px 4px 0;color:var(--bpi-text-faint);font-size:9px}.bpi-about-button{border:0;background:transparent;color:var(--bpi-text-muted);padding:1px 3px;cursor:pointer;font-size:9px}.bpi-about-button:hover{color:var(--bpi-link-text);text-decoration:underline}
    .bpi-pack-section{border:1px solid var(--bpi-border-2);border-radius:7px;background:var(--bpi-surface-3);padding:7px;margin-bottom:8px}.bpi-pack-toolbar{display:flex;gap:6px;align-items:center;justify-content:space-between;flex-wrap:wrap;margin-bottom:6px}.bpi-pack-toolbar strong{color:var(--bpi-text)}.bpi-pack-list{display:grid;grid-template-columns:repeat(auto-fit,minmax(225px,1fr));gap:5px;max-height:166px;overflow:auto}.bpi-pack-card{border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-2);padding:6px;display:grid;grid-template-columns:auto minmax(0,1fr) auto;gap:6px;align-items:start}.bpi-pack-card.bpi-pack-disabled{opacity:.58}.bpi-pack-name{font-weight:700;color:var(--bpi-text)}.bpi-pack-meta{font-size:10px;color:var(--bpi-text-muted);margin-top:2px;word-break:break-word}.bpi-pack-controls{display:flex;gap:3px;flex-wrap:wrap;justify-content:flex-end}.bpi-switch{margin-top:3px}.bpi-pack-personal{border-color:#397062}.bpi-community-form{display:grid;grid-template-columns:105px minmax(0,1fr);gap:8px;align-items:center}.bpi-community-form input{border:1px solid var(--bpi-border-2);border-radius:5px;background:var(--bpi-input-bg);color:var(--bpi-input-text);padding:6px}.bpi-community-preview{max-height:150px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:5px;padding:5px;margin-top:8px;color:var(--bpi-text-muted)}
-    .bpi-import-conflicts{max-height:240px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;margin-top:10px}.bpi-import-conflict{display:grid;grid-template-columns:1fr 1fr 1fr;gap:6px;padding:6px;border-top:1px solid var(--bpi-border-3)}.bpi-import-conflict:first-child{border-top:0}.bpi-import-stat{display:flex;gap:12px;flex-wrap:wrap;color:var(--bpi-text)}.bpi-danger-text{color:#ff9ba3}
+    .bpi-import-conflicts{max-height:240px;overflow:auto;border:1px solid var(--bpi-border-2);border-radius:6px;margin-top:10px}.bpi-import-conflict{display:grid;grid-template-columns:1fr 1fr 1fr;gap:6px;padding:6px;border-top:1px solid var(--bpi-border-3)}.bpi-import-conflict:first-child{border-top:0}.bpi-import-stat{display:flex;gap:12px;flex-wrap:wrap;color:var(--bpi-text)}.bpi-danger-text{color:var(--bpi-status-error)}
    .bpm-panel{box-sizing:border-box;width:100%;height:100%;min-height:280px;display:flex;flex-direction:column;gap:7px;padding:8px;color:var(--bpi-text);font:12px/1.4 Arial,sans-serif;overflow:hidden}
    .bpm-head{flex:none;display:flex;align-items:center;gap:7px;flex-wrap:wrap}.bpm-head strong{color:var(--bpi-text)}.bpm-head .bpi-status{margin-left:auto}
-    .bpm-tabs{flex:none;display:flex;gap:4px}.bpm-tab{flex:1;border:1px solid var(--bpi-border-2);border-radius:5px;padding:4px 6px;background:var(--bpi-button-bg);color:var(--bpi-button-text);cursor:pointer;font-size:11px;line-height:1.25}.bpm-tab:hover{background:var(--bpi-border-3)}.bpm-tab.bpm-tab-active{background:#285f8e;border-color:#3f83ba;color:#fff}
+    .bpm-tabs{flex:none;display:flex;gap:4px}.bpm-tab{flex:1;border:1px solid var(--bpi-border-2);border-radius:5px;padding:4px 6px;background:var(--bpi-button-bg);color:var(--bpi-button-text);cursor:pointer;font-size:11px;line-height:1.25}.bpm-tab:hover{background:var(--bpi-border-3)}.bpm-tab.bpm-tab-active{background:var(--bpm-accent);border-color:var(--bpm-accent-border);color:#fff}
    .bpm-section{flex:1;min-height:0;display:flex;flex-direction:column;gap:6px}.bpm-section[hidden]{display:none}
    .bpm-toolbar{flex:none;display:flex;gap:4px;align-items:center;flex-wrap:wrap}.bpm-toolbar .bpi-search{min-width:120px}.bpm-toolbar .bpi-mode{flex:1;min-width:90px}
    .bpm-summary{flex:none;color:var(--bpi-text-muted);font-size:11px;min-height:15px;word-break:break-all}
    .bpm-rows{flex:1;min-height:120px;overflow-y:auto;scrollbar-gutter:stable;border:1px solid var(--bpi-border-2);border-radius:6px;background:var(--bpi-surface-table)}.bpm-tagmanager-host{flex:1;min-height:0;overflow-y:auto;scrollbar-gutter:stable;display:flex;flex-direction:column;gap:7px;padding-right:2px}.bpm-tagmanager-host .bpi-details-body{display:flex;flex-direction:column;gap:7px}.bpm-tagmanager-host .bpi-table{height:auto;min-height:220px;max-height:none;flex:0 1 auto}
    .bpm-row{border-top:1px solid var(--bpi-border-3);padding:6px 8px;display:flex;flex-direction:column;gap:3px}.bpm-row:first-child{border-top:0}.bpm-row:hover{background:var(--bpi-surface-3)}
    .bpm-row-top{display:flex;align-items:flex-start;gap:5px}.bpm-row-top input[type="checkbox"]{margin-top:2px}.bpm-row-en{color:var(--bpi-text);font-family:Consolas,monospace;word-break:break-word;min-width:0;flex:1}
-    .bpm-row-zh{color:#9ed0ff;word-break:break-word}
+    .bpm-row-zh{color:var(--bpi-zh-text);word-break:break-word}
    .bpm-row-meta{display:flex;align-items:center;gap:5px;flex-wrap:wrap;color:var(--bpi-text-muted);font-size:10px}
    .bpm-actions{display:flex;gap:4px;flex-wrap:wrap}
    .bpm-footer{flex:none;display:flex;gap:4px;align-items:center;justify-content:space-between;flex-wrap:wrap}.bpm-footer .bpi-toolbar-group{justify-content:flex-start}
    .bpm-section .bpi-pack-list{flex:1;min-height:0;max-height:none;grid-template-columns:1fr}
    .bpm-section .bpi-form{grid-template-columns:78px minmax(0,1fr)}.bpm-section .bpi-form textarea{min-height:84px}
-    .bpm-assistant-actions{display:flex;gap:7px;flex-wrap:wrap;margin-top:10px}
    .bpm-loading{padding:24px;text-align:center;color:var(--bpi-text-faint)}
+    .bpi-form-actions{display:flex;justify-content:flex-end;gap:7px;margin-top:12px;flex-wrap:wrap}
+    /* 助手设置在 ComfyUI「设置」面板里：原生那一行是「标签 + 控件」，把标签收掉、控件占满整行 */
+    .bpi-settings-form{box-sizing:border-box;width:100%;padding:2px 0 6px}.bpi-settings-form h3{margin:0 0 10px;font-size:15px}
+    /* 设置面板每一项 = 名称列 + 控件列，名称列宽度跟着名字长短走：插件这两项一个叫
+       「语言与外观」一个叫「助手设置」，字数不同，右边的内容就被推得差了一格。
+       让名称列不占位、控件列占满，两块都从最左边开始，标题由各自的 h3 显示。
+       只在「我们的块不在第一列」时才生效，结构对不上时不会把内容自己藏掉。 */
+    *:has(> :not(:first-child) > .bpi-settings-form){align-items:flex-start}
+    *:has(> :not(:first-child) > .bpi-settings-form) > :first-child{display:none}
+    *:has(> :not(:first-child) > .bpi-settings-form) > :not(:first-child){flex:1 1 0%;min-width:0;justify-content:flex-start}
+    /* ComfyUI 给控件列自带 justify-end：宽度一占满就会把内容推到最右，这里改回靠左 */
+    *:has(> .bpi-settings-form){flex:1 1 0%;min-width:0}
+    .bpi-switch-row{display:flex;align-items:center;gap:8px;cursor:pointer;user-select:none;width:fit-content}
+    .bpi-switch-row input{position:absolute;width:0;height:0;opacity:0;margin:0}
+    .bpi-switch-track{position:relative;flex:0 0 auto;width:34px;height:18px;border-radius:9px;background-color:var(--bpi-switch-off);border:1px solid var(--bpi-switch-off-border);transition:background-color .16s ease,border-color .16s ease}
+    .bpi-switch-thumb{position:absolute;top:2px;left:2px;width:12px;height:12px;border-radius:50%;background-color:var(--bpi-switch-knob);transition:transform .16s ease}
+    .bpi-switch-row:hover .bpi-switch-track{border-color:var(--bpi-border-2)}
+    .bpi-switch-row input:checked + .bpi-switch-track{background-color:var(--bpi-danger-bg);border-color:var(--bpi-danger-border)}
+    .bpi-switch-row input:checked + .bpi-switch-track .bpi-switch-thumb{transform:translateX(16px)}
+    .bpi-switch-row input:focus-visible + .bpi-switch-track{outline:2px solid var(--bpi-row-pinned-ring);outline-offset:1px}
+    .bpi-switch-text{font-size:12px;color:var(--bpi-text-muted)}
+    .bpi-switch-row input:checked ~ .bpi-switch-text{color:var(--bpi-status-error)}
+    .bpi-switch-hint{display:none;font-size:11px;color:var(--bpi-status-error)}
+    .bpi-switch-row input:checked ~ .bpi-switch-hint{display:inline}.bpi-switch-row.bpi-switch-accent input:checked + .bpi-switch-track{background-color:var(--bpm-accent);border-color:var(--bpm-accent-border)}.bpi-switch-row.bpi-switch-accent input:checked ~ .bpi-switch-text{color:var(--bpi-text)}
+    /* 设置面板里的输入框 / 下拉：直接用 ComfyUI 的字段底色与描边，取不到才用兜底值，
+       免得暗色下我们的蓝灰底比面板亮一截、显得突兀 */
+    .bpi-settings-form .bpi-form input,.bpi-settings-form .bpi-form select,.bpi-settings-form .bpi-form textarea{background-color:var(--p-textfield-background,var(--p-inputtext-background,var(--bpi-field-bg)));border-color:var(--p-textfield-border-color,var(--p-inputtext-border-color,var(--bpi-field-border)))}
+    .bpi-settings-form .bpi-form input:focus,.bpi-settings-form .bpi-form select:focus,.bpi-settings-form .bpi-form textarea:focus{border-color:var(--p-primary-color,#4ca7e8)}
+    .bpm-panel .bpi-button.bpi-primary,.bpm-panel .bpi-mini.bpi-primary{background:var(--bpm-accent);border-color:var(--bpm-accent-border);color:#fff}
+    .bpm-panel .bpi-button.bpi-primary:hover,.bpm-panel .bpi-mini.bpi-primary:hover{background:var(--bpm-accent-hover);border-color:var(--bpm-accent-border);color:#fff}
+    .bpm-panel .bpi-filter-active{background:var(--bpm-accent);border-color:var(--bpm-accent-border);color:#fff}
+    .bpi-modal .bpi-button.bpi-primary,.bpi-modal .bpi-mini.bpi-primary{background:var(--bpm-accent);border-color:var(--bpm-accent-border);color:#fff}
+    .bpi-modal .bpi-button.bpi-primary:hover,.bpi-modal .bpi-mini.bpi-primary:hover{background:var(--bpm-accent-hover);border-color:var(--bpm-accent-border);color:#fff}
+    .bpi-dup-bar{display:flex;align-items:center;gap:6px;flex-wrap:wrap;padding:5px 2px;border-top:1px solid var(--bpi-border-3)}.bpi-dup-bar.bpi-hidden{display:none}.bpi-dup-label{flex:none;color:var(--bpi-text-muted);font-size:11px}
+    .bpi-dup-chip{display:inline-flex;align-items:center;gap:5px;max-width:260px;overflow:hidden;border:1px dashed var(--bpi-border-2);border-radius:999px;padding:2px 8px;background-color:var(--bpi-surface-3);color:var(--bpi-text);cursor:pointer;font-size:11px;user-select:none}
+    .bpi-dup-chip:hover{background-color:var(--bpi-hidden-chip-hover-bg);border-color:var(--bpi-link-text)}.bpi-dup-en{font-family:Consolas,monospace;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;min-width:0}.bpi-dup-count{flex:none;color:var(--bpi-text-muted);font-size:10px}.bpi-dup-index{flex:none;color:var(--bpi-status-error);font-size:10px;font-weight:700}.bpi-dup-chip-armed{border-style:solid;border-color:var(--bpi-status-error);color:var(--bpi-status-error)}
+    .bpi-dup-off{opacity:.6}.bpi-dup-off .bpi-en,.bpi-dup-off .bpi-english-token{text-decoration:line-through;color:var(--bpi-text-faint)}
+    .bpi-row.bpi-dup-hit,.bpi-token-card.bpi-dup-hit{background-image:linear-gradient(var(--bpi-dup-overlay),var(--bpi-dup-overlay));outline:1px solid var(--bpi-status-error);outline-offset:-1px}
+    .bpi-token-card.bpi-dup-hit .bpi-english-token{background-color:transparent;border-color:transparent;box-shadow:none}.bpi-unique-toggle{margin-left:6px;font-size:10px}
+    .bpi-dup-chip-unique{border-style:solid;border-color:var(--bpi-custom-border);background-color:var(--bpi-custom-bg)}.bpi-dup-chip-unique:hover{border-color:var(--bpi-link-text)}.bpi-dup-flag{flex:none;color:var(--bpi-custom-text);font-size:10px}
+    .bpi-dup-space{flex:1 1 32px;min-width:32px;align-self:stretch;cursor:default}
  `;
  document.head.appendChild(style);
 }
@@ -931,10 +1484,13 @@ export {
  LANGUAGE_CHANGED_EVENT,
  MANAGER_TAB_ID,
  MANAGER_OPEN_EVENT,
+  UNINDEXED_PACK_ID,
  UPSTREAM_ARRIVED_EVENT,
  applyBpiAppearance,
  applyLanguageToDom,
+  buildTagFields,
  button,
+  checkDuplicateTags,
  createSavedPrompt,
  deleteCommunityPack,
  deletePersonalTag,
@@ -954,11 +1510,13 @@ export {
  importTags,
  bulkUpdateTags,
  injectBpiStyles,
+  listAssistantModels,
  listModelNames,
  listSavedPrompts,
  loadDictionary,
  loadPreferences,
  lookupLargeDictionary,
+  buildColorPanel,
  openManagerPanel,
  openSavePromptDialog,
  openTagDialog,
@@ -968,6 +1526,7 @@ export {
  saveAssistantConfig,
  savePreferences,
  saveTag,
+  saveTokenColors,
  searchLargeDictionary,
  setLanguageMode,
  setLargeDictionaryEnabled,
@@ -0,0 +1,190 @@
+import { app } from "../../scripts/app.js";
+import { checkDuplicateTags, t } from "./bpi_shared.js";
+
+const NODE_NAME = "DuplicateChecker";
+const STYLE_ID = "bpi-duplicate-check-style";
+const PILL_HEIGHT = 26;
+const REFRESH_DELAY = 250;
+// 半透明橙：叠在画布上，亮色深色主题都能用，也不用去猜当前主题
+const AMBER_TINT = "rgba(214,158,46,0.14)";
+
+function injectStyles() {
+  if (document.getElementById(STYLE_ID)) return;
+  const style = document.createElement("style");
+  style.id = STYLE_ID;
+  style.textContent = [
+    // holder 才是交给画布的 widget 元素：节点拖大时它被拉伸，胶囊自己锁死高度不跟着变
+    `.bpi-dup-holder{display:flex;align-items:center;justify-content:flex-start;width:100%;min-height:${PILL_HEIGHT}px;overflow:hidden}`,
+    `.bpi-dup-pill{display:inline-flex;align-items:center;gap:6px;padding:3px 10px;border-radius:999px;font-size:12px;line-height:1.5;height:${PILL_HEIGHT}px;box-sizing:border-box;flex:none;max-width:100%;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}`,
+    ".bpi-dup-pill .bpi-dup-dot{width:7px;height:7px;border-radius:50%;flex:none}",
+    ".bpi-dup-pill .bpi-dup-text{overflow:hidden;text-overflow:ellipsis}",
+    ".bpi-dup-pill.bpi-dup-warn{background:rgba(214,158,46,.18);color:#8a5a00}",
+    ".bpi-dup-pill.bpi-dup-warn .bpi-dup-dot{background:#8a5a00}",
+    ".bpi-dup-pill.bpi-dup-idle{background:rgba(128,128,128,.14);opacity:.75}",
+    ".dark-theme .bpi-dup-pill.bpi-dup-warn{color:#e8b85c}",
+    ".dark-theme .bpi-dup-pill.bpi-dup-warn .bpi-dup-dot{background:#e8b85c}",
+  ].join("");
+  document.head.appendChild(style);
+}
+
+function upstreamWidget(node) {
+  const linkId = node?.inputs?.[0]?.link;
+  if (linkId == null) return null;
+  const graph = app.graph;
+  const links = graph?.links;
+  const info = Array.isArray(links) ? links[linkId] : links?.[linkId];
+  if (!info) return null;
+  const origin = (graph?._nodes ?? []).find((item) => String(item.id) === String(info.origin_id));
+  // 上游节点的文本 widget 是画布上唯一能读到的值；读不到就只靠执行结果更新。
+  // 缓存 widget 本身，这样每帧只要读一次 value，不用再遍历整张图。
+  return (
+    (origin?.widgets ?? []).find((item) => item.name === "text" && typeof item.value === "string") ??
+    null
+  );
+}
+
+function setPill(node, text, tone, terms = []) {
+  const state = node._duplicateCheck;
+  if (!state) return;
+  state.pill.className = `bpi-dup-pill ${tone === "warn" ? "bpi-dup-warn" : "bpi-dup-idle"}`;
+  state.textNode.textContent = text;
+  state.dot.style.display = tone === "warn" ? "" : "none";
+  state.pill.title = terms.length ? terms.join("、") : "";
+}
+
+function applyResult(node, count, terms = []) {
+  const state = node._duplicateCheck;
+  if (!state) return;
+  state.count = count;
+  if (!state.connected) {
+    setPill(node, t("Waiting for prompt"), "idle");
+  } else if (count > 0) {
+    setPill(node, t(`Duplicates ${count}`), "warn", terms);
+  } else {
+    setPill(node, t("No duplicates"), "idle");
+  }
+  node.bgcolor = state.connected && count > 0 ? AMBER_TINT : undefined;
+  node.setDirtyCanvas?.(true, true);
+}
+
+async function runCheck(node, text) {
+  const state = node._duplicateCheck;
+  if (!state) return;
+  const ticket = (state.ticket = (state.ticket ?? 0) + 1);
+  try {
+    const data = await checkDuplicateTags(text);
+    if (state.ticket !== ticket) return;
+    applyResult(node, data?.count ?? 0, (data?.terms ?? []).map((item) => item.term));
+  } catch {
+    // 查询失败就保持上一次的状态，别把节点闪成「无重复」误导人
+  }
+}
+
+function scheduleCheck(node, text) {
+  const state = node._duplicateCheck;
+  if (!state || text === state.checkedText) return;
+  state.checkedText = text;
+  clearTimeout(state.timer);
+  state.timer = setTimeout(() => runCheck(node, text), REFRESH_DELAY);
+}
+
+function syncConnection(node) {
+  const state = node._duplicateCheck;
+  if (!state) return;
+  state.connected = node.inputs?.[0]?.link != null;
+  state.source = state.connected ? upstreamWidget(node) : null;
+  if (!state.connected) {
+    state.checkedText = null;
+    applyResult(node, 0, []);
+    return;
+  }
+  const text = state.source?.value;
+  if (typeof text !== "string") {
+    applyResult(node, 0, []);
+    return;
+  }
+  scheduleCheck(node, text);
+}
+
+app.registerExtension({
+  name: "ComfyUI.BilingualPromptInspector.DuplicateCheck",
+  async beforeRegisterNodeDef(nodeType, nodeData) {
+    if (nodeData.name !== NODE_NAME) return;
+
+    const originalCreated = nodeType.prototype.onNodeCreated;
+    nodeType.prototype.onNodeCreated = function () {
+      originalCreated?.apply(this, arguments);
+      injectStyles();
+      // widget.label / input.label 都是直接赋值，不走翻译层，所以这里直接写中文
+      const toggle = this.widgets?.find((widget) => widget.name === "keep_first");
+      if (toggle) toggle.label = "只保留第一个";
+      if (this.inputs?.[0]) this.inputs[0].label = "提示词";
+      const holder = document.createElement("div");
+      holder.className = "bpi-dup-holder";
+      const pill = document.createElement("div");
+      pill.className = "bpi-dup-pill bpi-dup-idle";
+      const dot = document.createElement("span");
+      dot.className = "bpi-dup-dot";
+      const textNode = document.createElement("span");
+      textNode.className = "bpi-dup-text";
+      pill.append(dot, textNode);
+      holder.append(pill);
+      this._duplicateCheck = { pill, dot, textNode, count: 0, connected: false };
+      try {
+        this.addDOMWidget("bpi_duplicate_status", "bpi-duplicate-status", holder, {
+          getMinHeight: () => PILL_HEIGHT,
+          getMaxHeight: () => PILL_HEIGHT,
+          hideOnZoom: false,
+          serialize: false,
+        });
+      } catch (error) {
+        console.error("[DuplicateChecker] Failed to create status pill", error);
+        this._duplicateCheck = null;
+        return;
+      }
+      setPill(this, t("Waiting for prompt"), "idle");
+      requestAnimationFrame(() => {
+        // 胶囊：一行状态加一个开关。等 DOM 布局算完再压尺寸，免得把开关挤没了
+        this.setSize?.([236, Math.min(this.size?.[1] ?? 88, 96)]);
+        syncConnection(this);
+      });
+    };
+
+    const originalConnectionsChange = nodeType.prototype.onConnectionsChange;
+    nodeType.prototype.onConnectionsChange = function () {
+      originalConnectionsChange?.apply(this, arguments);
+      syncConnection(this);
+    };
+
+    // canvas 模式靠每帧比对上游文本来保持实时；Nodes 2.0 不跑这里，
+    // 那种情况下由连线变化与执行结果兜底。
+    const originalDrawForeground = nodeType.prototype.onDrawForeground;
+    nodeType.prototype.onDrawForeground = function () {
+      originalDrawForeground?.apply(this, arguments);
+      const state = this._duplicateCheck;
+      if (!state?.connected) return;
+      const text = state.source?.value;
+      if (typeof text === "string") scheduleCheck(this, text);
+    };
+
+    const originalExecuted = nodeType.prototype.onExecuted;
+    nodeType.prototype.onExecuted = function (message) {
+      originalExecuted?.apply(this, arguments);
+      const state = this._duplicateCheck;
+      if (!state) return;
+      state.connected = true;
+      state.checkedText = null;
+      // 后端 ui 的值会被 execution.py 展开成列表，所以这里取第一个元素
+      const counts = message?.duplicate_count;
+      const count = Number(Array.isArray(counts) ? counts[0] : counts);
+      applyResult(this, Number.isFinite(count) ? count : 0, message?.duplicate_terms ?? []);
+    };
+
+    const originalRemoved = nodeType.prototype.onRemoved;
+    nodeType.prototype.onRemoved = function () {
+      clearTimeout(this._duplicateCheck?.timer);
+      this._duplicateCheck = null;
+      originalRemoved?.apply(this, arguments);
+    };
+  },
+});
@@ -44,12 +44,29 @@ const ZH = {
  Discard: "丢弃",
  Resume: "继续",
  Weight: "权重",
-  Color: "颜色",
+  "Weight / color": "权重 / 颜色",
  Apply: "应用",
  Default: "默认",
  Random: "随机",
  None: "无",
  "By inclusion": "按收录",
+  "By dictionary": "按词库",
+  "By Anima": "按 Anima",
+  Quality: "质量",
+  Copyright: "作品",
+  Expression: "表情",
+  Camera: "镜头",
+  Scene: "场景",
+  "Pick a color for each Anima category; unset ones keep the default.": "给每个 Anima 分类挑一个颜色；没配色的保持默认底色。",
+  "This color has the lowest priority: it only shows in Default mode; other color modes override it.": "这里改的颜色优先级最低：只有「默认」上色模式才显示，按收录、随机、按词库、按 Anima 都会盖住它。",
+  Colors: "颜色",
+  "Pick a color for each dictionary; uncolored ones keep the default.": "给每个词库挑一个颜色；没配色的保持默认底色。",
+  "Restore defaults": "恢复默认",
+  "Random color": "随机颜色",
+  "Colors saved": "颜色已保存",
+  "Tags matching no dictionary": "没命中任何词库的标签",
+  "Invalid color; use #RGB or #RRGGBB": "颜色格式不对；请用 #RGB 或 #RRGGBB",
+  "{} items": "{} 条",
  "Invalid color; use #hex or rgb(r,g,b)": "颜色格式不对；请用 #hex 或 rgb(r,g,b)",
  Category: "分类",
  English: "英文",
@@ -71,7 +88,6 @@ const ZH = {
  "Chinese explanation (read-only)": "中文解释（只读）",
  "Category table view | click tags to link; category names are not written to the prompt":
    "分类表视图 | 点击标签联动；分类名不会写入提示词",
-  "Click tags to link; select then press Delete": "点击标签联动；选中后按 Delete",
  "Enter the English prompt; after editing, it will be shown as selectable, deletable tags.":
    "输入英文提示词；编辑完成后会显示成可选中、可删除的标签。",
  "Enter a prompt in the English text box above; per-tag bilingual mapping will appear here.":
@@ -87,17 +103,21 @@ const ZH = {
  "Highest": "最高",
  Translate: "翻译",
  "Translate & optimize": "翻译并优化",
-  "Optimize to Anima": "优化为 Anima",
+  Optimize: "优化",
+  "Fit height": "适应高度",
+  "Let the prompt area grow with the node instead of being capped": "让提示词区跟着节点一起变高，不再卡在固定的高度上限",
+  "Back to the automatic height; the prompt area stops following the node": "回到自动高度，提示词区不再跟着节点变高",
  "Re-translate": "重新翻译",
  "Translate all unknown": "翻译全部未收录",
  "Sort by Anima order": "按 Anima 顺序排序",
-  "Confirm write English": "确认写入英文",
+  "Cancel Anima order sort": "取消 Anima 顺序排序",
+  "Cancel Anima sort": "取消 Anima 排序",
+  "Confirm write new prompt": "确认写入新提示词",
  "Confirm & save": "确认并保存",
  "Save to dict": "存入词库",
  "Save to Personal Dictionary": "保存到个人词库",
  "Add tag": "添加标签",
  "New Tag": "新建标签",
-  "Search candidates": "搜索候选",
  "Dictionary search": "词库搜索",
  "Enter Chinese, English, alias, or concept; click a result to insert the English tag":
    "输入中文、英文、别名或概念；点击结果即可插入英文标签",
@@ -128,6 +148,10 @@ const ZH = {
  "Personal dictionary": "个人词库",
  "No inspector node on canvas": "画布上没有检查器节点",
  "No matching items under the current filter.": "当前筛选条件下没有匹配项。",
+  "Find in the current prompt": "在当前提示词里查找",
+  "Clear search": "清除查找",
+  "No tag matches {}": "没有匹配“…”的标签",
+  "Filter the detail list by keyword; matches the English original or the Chinese explanation; the prompt itself never changes": "按关键词筛明细表；英文原文和中文解释都参与匹配，提示词本身不会变",
  "No unknown tags currently": "当前没有未收录标签",
  "Edit directly": "直接编辑",

@@ -155,8 +179,6 @@ const ZH = {
  "No content change": "内容没有变化",
  "Natural-language segments only support whole-segment editing; tags can still be deleted individually":
    "自然语言片段只支持整段编辑；标签仍可单独删除",
-  "Only tags support individual weights; use text editing for natural language":
-    "只有标签支持单独权重；自然语言请用文本编辑",
  "Weight must be between 0 and 3": "权重必须在 0 到 3 之间",
  "Use “Edit text” to process it": "请用「编辑文本」处理",
  "Click “Edit text”, then use the inspector's own “Translate” or “Translate & optimize”":
@@ -177,6 +199,8 @@ const ZH = {
  "deleted {}; press Ctrl+Z to undo": "已删除「{}」；按 Ctrl+Z 可撤销",
  "{}; press Ctrl+Z to undo": "{}；按 Ctrl+Z 可撤销",
  "moved {}; press Ctrl+Z to undo": "已移动「{}」；按 Ctrl+Z 可撤销",
+  "rewrote {}; press Ctrl+Z to undo": "已改写「{}」；按 Ctrl+Z 可撤销",
+  "rewrote {} tags; press Ctrl+Z to undo": "已改写 {} 个标签；按 Ctrl+Z 可撤销",
  "hidden {} (excluded from actual output); can be restored anytime in the hidden section; press Ctrl+Z to undo":
    "已隐藏「{}」（不计入实际输出）；随时可在隐藏区恢复；按 Ctrl+Z 可撤销",
  "restored {} to its original position; press Ctrl+Z to undo": "已把「{}」恢复到原位置；按 Ctrl+Z 可撤销",
@@ -208,6 +232,8 @@ const ZH = {
  "These tags stay visible above (strikethrough) but will not enter the actual output; click a tag to unhide":
    "这些标签仍显示在上方（加删除线），但不会进入实际输出；点击标签取消隐藏",
  "Click to unhide {}": "点击取消隐藏「{}」",
+  "Click to unhide {} — {}": "点击取消隐藏「{}」——{}",
+  "Click to restore {} — {} to its original position": "点击把「{}」恢复到原位置——{}",
  "Unhide all {} hidden tags": "全部取消隐藏（{} 个）",
  "hidden ({})": "已隐藏（{}）",
  "Delete personal override and restore built-in explanation": "删除个人覆盖并恢复内置解释",
@@ -222,8 +248,8 @@ const ZH = {
  "Aliases: {}": "别名：{}",
  "{} ↔ {} | natural language supports only whole-segment editing; can drag to reorder the whole segment":
    "「{}」↔「{}」｜自然语言只支持整段编辑；可拖动整段调整顺序",
-  "{} ↔ {} | click to link; double-click to edit weight; select then press Delete; drag to reorder or Alt+↑/↓ to nudge":
-    "「{}」↔「{}」｜点击联动；双击编辑权重；选中后按 Delete；拖动排序或 Alt+↑/↓ 微调",
+  "{} ↔ {} | click to link; Ctrl+click to select multiple; double-click to open the editor; select then press Delete; drag to reorder or Alt+↑/↓ to nudge":
+    "「{}」↔「{}」｜点击联动；Ctrl+点可多选；双击打开编辑；选中后按 Delete；拖动排序或 Alt+↑/↓ 微调",
  "Insert line break": "插入换行",
  "Delete line break": "删除换行",
  "Move line break": "移动换行",
@@ -264,6 +290,10 @@ const ZH = {
  "Type Chinese to translate it into English tags, or type English tags to add them directly":
    "输入中文会自动翻译为英文标签；直接输入英文则原样添加",
  Add: "添加",
+  "Insert after {}": "插到「{}」之后",
+  "Add to the end (no tag selected)": "添加到末尾（未选中标签）",
+  "The new tag goes right after {}": "新标签会插到「{}」之后",
+  "No tag is selected, so the new tag goes at the end of the prompt": "没有选中标签，新标签会添加到提示词末尾",
  "Enter Chinese or an English tag first": "请先输入中文或英文标签",
  "Added {} to the prompt": "已把「{}」添加到提示词",
  "Translating {}…": "正在翻译「{}」…",
@@ -289,6 +319,7 @@ const ZH = {
  Models: "适用模型",
  "Recommended Weight": "推荐权重",
  "e.g.: looking at viewer": "例如：looking at viewer",
+  "e.g.: apple": "例如：苹果",
  "e.g.: pose, camera, style": "例如：pose、camera、style",

  // ── 弹窗：保存 / 编辑收藏 ────────────────────────────────────────────
@@ -298,6 +329,13 @@ const ZH = {
  "Model limit reached": "模型数量已达上限",
  "Type keywords": "输入关键词",
  "Press Enter to add; ↑↓ to choose": "回车添加，↑↓ 切换候选",
+  "Type to filter; ↑↓ to choose": "输入即筛选，↑↓ 切换候选",
+  "Manual input": "手动输入",
+  "Tag saved": "词条已保存",
+  "Entry info": "词条信息",
+  "Edits go to the personal dictionary and override same-name entries.": "改动存入个人词库，会覆盖同名词条的显示。",
+  "Nothing was saved: the Chinese name is still the unknown placeholder.": "没有保存：中文名称还是「未收录」，这样存进去只会得到一条残词条。",
+  "Enter a real Chinese name, then apply again.": "先填一个真正的中文名称，再点应用。",
  "Edit Favorite": "编辑收藏",
  "Keep the current image, or pick a new one to replace it": "不换图就保持原样，选一张新的会替换掉原来的",

@@ -317,7 +355,6 @@ const ZH = {
  Packs: "词包",
  Favorites: "收藏",
  "Assistant Settings": "助手设置",
-  "Manager panel": "管理面板",
  "Add or select a BilingualPromptInspector node to see its tag manager here.":
    "添加或选中画布上的 BilingualPromptInspector 节点，即可在此看到它的标签管理。",
  "Tag translation, dictionary search and per-tag operations edit the selected node below.":
@@ -364,18 +401,17 @@ const ZH = {
  "Recognizing tag categories…": "正在识别标签分类…",
  "Loading dictionary…": "正在加载词库…",
  "Loading settings...": "正在加载设置…",
-  "Settings load on first visit...": "首次进入时加载设置…",

-  // ── 助手设置 ────────────────────────────────────────────────────────
+  // ── 助手设置（ComfyUI「设置」面板里内嵌的那一项）─────────────────────
  "Translation Service": "翻译服务",
  "Dictionary only (no API)": "仅词库（不需要接口）",
  "Baidu Translate": "百度翻译",
-  "AI (OpenAI-compatible / Ollama)": "AI（OpenAI 兼容 / Ollama）",
-  "Translation/explanation uses this choice. \"Translate & Optimize\" and \"Optimize to Anima\" always use AI, unrelated here. To test AI connection, temporarily set translation service to AI then click test (key preserved).":
-    "翻译／解释按这里的选择走。「翻译并优化」和「优化为 Anima」始终走 AI，与此无关。要测试 AI 连接，请临时把翻译服务切到 AI 再点测试（密钥会保留）。",
+  "Only this service's settings are shown below. \"Translate & Optimize\" and \"Optimize\" always use AI: choose OpenAI-compatible API or Ollama above to configure or test them. Switching service never clears saved keys.":
+    "下面只显示当前所选服务的设置。「翻译并优化」和「优化」始终走 AI：要配置或测试它们，请在上方选择「OpenAI 兼容接口」或「Ollama」。切换服务不会清掉已保存的密钥。",
  "APP ID": "APP ID",
  "Secret Key": "密钥",
  "Clear saved Baidu secret": "清除已保存的百度密钥",
+  "Takes effect after saving": "保存后生效",
  "Baidu Cloud \"General Text Translation\" APP ID": "百度智能云「通用文本翻译」的 APP ID",
  "Baidu Cloud \"General Text Translation\" Secret Key": "百度智能云「通用文本翻译」的密钥",
  "QPS (requests/second)": "QPS（每秒请求数）",
@@ -387,14 +423,26 @@ const ZH = {
  "OpenAI-compatible API": "OpenAI 兼容接口",
  "Ollama (local)": "Ollama（本地）",
  "API URL": "API 地址",
-  "Optional for Ollama; e.g.: https://host/v1": "Ollama 可选；例如 https://host/v1",
+  "provider model ID": "服务商的模型 ID",
+  "e.g.: qwen3:8b": "例如：qwen3:8b",
+  "Optional for local LM Studio": "本地 LM Studio 可选",
+  "Remote addresses must use https; http is only accepted for this machine. Private, reserved and link-local addresses are rejected. Keys are never sent in clear text over the network.":
+    "远端地址必须使用 https；只有本机地址可以用 http。内网、保留和链路本地地址一律拒绝。密钥绝不会明文发到网络上。",
+  "Ollama only accepts a loopback address on this machine; anything else is rejected before a request is sent.":
+    "Ollama 只能填本机回环地址；其它地址在发出请求前就会被拒绝。",
  "Model Name": "模型名称",
-  "e.g.: qwen3:8b or provider model ID": "例如：qwen3:8b 或服务商的模型 ID",
+  "Fetch Models": "获取模型",
+  "Fetch Models ({})": "获取模型（{}）",
+  "Fetching...": "获取中…",
+  "Fetched {} models": "已获取 {} 个模型",
+  Close: "关闭",
  "Use LM Studio Local Preset": "使用 LM Studio 本地预设",
  "Connects to localhost:1234 by default; no external API required.": "默认连接 localhost:1234，不需要外部接口。",
  "LM Studio default URL filled; select a loaded model ID, then Save & Test": "已填入 LM Studio 默认地址；选择已加载的模型 ID 后保存并测试",
+  "Use Ollama Local Preset": "使用 Ollama 本地预设",
+  "Connects to localhost:11434 by default; no external API required.": "默认连接 localhost:11434，不需要外部接口。",
+  "Ollama default URL filled; select a pulled model ID, then Save & Test": "已填入 Ollama 默认地址；选择已拉取的模型 ID 后保存并测试",
  "API Key": "API 密钥",
-  "Optional for local Ollama / LM Studio": "本地 Ollama / LM Studio 可选",
  "Clear saved API Key": "清除已保存的 API 密钥",
  "Saved; leave empty to keep": "已保存；留空则保持不变",
  Temperature: "温度",
@@ -415,11 +463,13 @@ const ZH = {
  // ── 语言设置（本功能新增） ────────────────────────────────────────────
  Language: "语言",
  Chinese: "中文",
+  English: "英文",
  "Chinese (Simplified)": "简体中文",
  "Follow ComfyUI language setting": "跟随 ComfyUI 语言设置",
  "Choose the language for this plugin's own panels. \"Follow ComfyUI\" reads the ComfyUI language setting; \"Chinese\" and \"English\" force one. Applied to open panels right away; reload the page if anything still shows the old language.":
    "选择本插件面板的语言。「跟随 ComfyUI 语言设置」会随 ComfyUI 的语言切换；「中文」「英文」为强制指定。已打开的面板会立即更新；若个别位置仍是旧语言，刷新页面即可。",
-  "Language saved": "语言设置已保存",
+  "Appearance & Language": "语言与外观",
+  "Appearance and language saved": "外观与语言已保存",

  // ── 标签来源 / 分类 / 模式 ──────────────────────────────────────────
  "Inspector assistant": "检查器助手",
@@ -431,6 +481,14 @@ const ZH = {
  "Uncategorized": "未分类",
  "Not indexed": "未收录",
  "Natural-language segment (pending translation or confirmation)": "自然语言片段（待翻译或确认）",
+  "Suspected natural language": "疑似自然语言",
+  Suspected: "疑似",
+  "Save to Dictionary": "保存的词库",
+  "Natural Language Dictionary": "自然语言词库",
+  "Not indexed · suspected natural language": "未收录 · 疑似自然语言",
+  // 句尾的 3 也是变量：VAR_PATTERN 同时吃「“…”」和数字
+  "{} ↔ {} | suspected natural language: more than {} spaces; confirm or split it into tags":
+    "{} ↔ {}｜疑似自然语言：空格超过 {} 个；请确认或拆成标签",
  "Custom": "自定义",
  "Pending": "待确认",
  Syntax: "语法",
@@ -444,6 +502,9 @@ const ZH = {
  "English prompt tag view": "英文提示词标签视图",
  "Anima official sort": "Anima 官方顺序",
  "Anima official order preview": "Anima 官方顺序预览",
+  "Restore the prompt as it was before the Anima sort": "恢复到 Anima 排序前的提示词（分类视图一起退出）",
+  "Restored the prompt from before the Anima sort": "已恢复到 Anima 排序前的提示词",
+  "Exited the Anima category view": "已退出 Anima 分类视图",
  "Empty input": "输入为空",
  "Copied": "已复制",

@@ -502,6 +563,8 @@ const ZH = {
  "Save Personal": "保存为个人",
  "Reset Built-in": "重置为内置",
  "Select all ({} items)": "全选（{} 条）",
+  "Edit {}": "编辑「{}」",
+  "Edit {} tags": "编辑 {} 个标签",
  "pack:": "词包：",
  "Copied {}": "已复制 {}",
  "Processed {}": "已处理「{}」",
@@ -539,6 +602,8 @@ const ZH = {
  "Added {}": "新增 {}",
  "Conflicts {}": "冲突 {}",
  "Duplicates {}": "重复 {}",
+  "No duplicates": "无重复",
+  "Waiting for prompt": "等待提示词",
  "Invalid {}": "无效 {}",
  "Import complete: added {}, overwrote {}, skipped {} | auto-backed up": "导入完成：新增 {}、覆盖 {}、跳过 {}｜已自动备份",
  "Import complete: {} added": "导入完成：新增 {} 条",
@@ -564,6 +629,9 @@ const ZH = {
  "Translation Rule": "翻译规则",
  "Translate & Optimize Rule": "翻译并优化规则",
  "Optimize Rule": "优化规则",
+  "Natural Language Rule": "自然语言规则",
+  "Optimize with the rule for the current mode: tag mode tidies the tag flow, natural language mode rewrites it as English sentences":
+    "按当前模式对应的规则优化：标签模式整理标签流，自然语言模式改写为英文自然语句",
  "Connection OK: {}": "连接正常：{}",
  // 确认对话框的默认确认按钮（confirmDialog 的 confirmLabel 缺省值），
  // 不是「连接正常」的 OK——那个是整句键 "Connection OK: {}"。
@@ -577,8 +645,6 @@ const ZH = {
  "Delete {}? Reference image will be deleted too.": "删除「{}」？参考图也会一并删除。",
  "Node #{}": "节点 #{}",
  "Node #{}: {}": "节点 #{}：{}",
-  "Bilingual Prompt Inspector: Tag Manager, Dictionary, Packs & Assistant Settings":
-    "提示词翻译与管理：标签管理、词库、词包与助手设置",

  // ── 语法诊断（parser.js）────────────────────────────────────────────
  "Manual override": "手动覆盖",
@@ -604,6 +670,27 @@ const ZH = {
  "Bracket mismatch: extra or misplaced {}": "括号不匹配：多了或位置不对的 {}",
  "Unclosed bracket: missing {}": "括号未闭合：缺少 {}",
  "Duplicate tag: {}": "重复标签：{}",
+  // 重复标签的「唯一」开关 + 重复栏
+  Unique: "唯一",
+  "Keep only the first {} in the actual output; later occurrences stay here with a strikethrough":
+    "只让第一次出现的「{}」进入实际输出；其余照常显示在界面上（划线）",
+  "Tags repeated in this prompt; click a chip to step through its occurrences, double-click to delete the one you stepped to; click empty space here to clear the selection":
+    "提示词里重复的标签：单击胶囊依次跳到每一次出现，双击删掉跳到的那一处；点这里空白处取消选中",
+  "{} appears {} times; click to jump to the first one": "「{}」出现了 {} 次；单击跳到第一次",
+  "{} appears {} times; {} is on, only the first reaches the actual output": "「{}」出现了 {} 次；已开「{}」：只有第一次进入实际输出",
+  "{} appears {} times; now at {}, click for the next one, double-click to delete it":
+    "「{}」出现了 {} 次；当前第 {} 处，单击跳下一处，双击删除这一处",
+  "{} occurrence {} of {}; click again for the next one, double-click to delete it":
+    "「{}」第 {} / {} 处；再点一次跳下一处，双击删除这一处",
+  "Click a duplicate chip once to step to it before double-clicking to delete":
+    "先单击一次重复胶囊定位到某一处，再双击才会删除",
+  "{} now unique: only the first occurrence enters the actual output; press Ctrl+Z to undo":
+    "「{}」已设为唯一：只有第一次出现进入实际输出；按 Ctrl+Z 可撤销",
+  "{} no longer unique: all occurrences enter the actual output again; press Ctrl+Z to undo":
+    "「{}」已取消唯一：所有重复项重新进入实际输出；按 Ctrl+Z 可撤销",
+  "Query: {} | {} is on: only the first occurrence reaches the actual output": "查询「{}」｜已开「{}」：只有第一次出现进入实际输出",
+  "keep only the first {}": "只保留第一次出现的「{}」",
+  "restore duplicates of {}": "恢复「{}」的重复项",
  "Invalid weight format: {}": "权重格式无效：{}",
  "Full-width separator detected; Anima tag flow usually uses ASCII commas": "检测到全角分隔符；Anima 标签流一般用半角逗号",
  "Consecutive commas or empty tags detected": "检测到连续逗号或空标签",
@@ -682,9 +769,15 @@ const ZH = {

  // ── 第三批：自由文本变量包裹后的组合形态与遗漏项 ──────────────────────
  "Low": "低",
-  "Low · pending": "低 · 待确认",
-  "Medium · pending": "中 · 待确认",
-  "Medium · session-only": "中 · 仅本次会话",
+  "Medium": "中",
+  "High": "高",
+  "Not in any dictionary, so it cannot be judged": "词库里查不到，没法判断",
+  "Machine translation, not saved to the dictionary or confirmed yet": "机器翻译，还没存进词库，也没人工确认",
+  "Syntax kept as-is; nothing to translate": "语法保留项，本来就不用翻译",
+  "In the dictionary, but not marked as verified": "词库里有，但还没标成已校验",
+  "In the dictionary, verified, category {}": "已在词库里，已校验，分类「{}」",
+  "In the dictionary, but no category set": "词库里有，但还没填分类",
+  "Changed for this session only; not saved to the dictionary": "只改了本次会话，没存进词库",
  "★ Unfavorite": "★ 取消收藏",
  "☆ Favorite": "☆ 收藏",
  "Paused, waiting for confirmation…": "已暂停，等待确认…",
@@ -713,7 +806,8 @@ const ZH = {

  "Click a tag to link it with the detail table below; select then press Delete": "点击标签可与下方明细表联动；选中后按 Delete 删除",
  "English prompt is empty; nothing to optimize": "英文提示词是空的，没有可优化的内容",
-  "“Optimize to Anima” only processes English; translate Chinese in the sidebar first": "「优化为 Anima」只处理英文；中文请先在侧边栏翻译",
+  "“Optimize” only processes English in tag mode; translate Chinese in the sidebar first": "「优化」在标签模式下只处理英文；中文请先在侧边栏翻译，或把模式切到「自然语言」",
+  "(Chinese detected; translating to English first)": "（检测到中文，会先译成英文）",
  "{} done; please confirm in the preview window": "「{}」完成；请在预览窗口确认",
  "{}; please fix it in the preview before confirming": "「{}」；请在预览里改好再确认",
 };
@@ -24,6 +24,7 @@ import {
  bulkUpdateTags,
  applyBpiAppearance,
  injectBpiStyles,
+  listAssistantModels,
  listSavedPrompts,
  loadDictionary,
  loadPreferences,
@@ -48,7 +49,6 @@ const SECTIONS = [
  ["tag-manager", "Tag Manager"],
  ["tags", "Dictionary"],
  ["packs", "Packs"],
-  ["assistant", "Assistant Settings"],
  ["favorites", "Favorites"],
 ];

@@ -58,7 +58,6 @@ const manager = {
  selected: new Set(),
  section: "tags",
  loaded: false,
-  assistantLoaded: false,
  favoritesLoaded: false,
  favorites: [],
  favoritesChangeHandler: null,
@@ -136,6 +135,9 @@ function rebuildSourceOptions() {
    ["all", "All Packs"],
    ["builtin", "Built-in pack enabled"],
    ["personal", "Personal"],
+    // 保存时选了「自然语言词库」的条目：和个人词库共用文件，只靠 natural 标记区分，
+    // 与分类叫「自然语言」的标签是两回事（后者是 Anima 分类名）。
+    ["natural", "Natural Language Dictionary"],
    ...(manager.data.large_dictionary?.available ? [["large", "Danbooru Large Dict"]] : []),
    ["machine", "Pending Machine"],
    ["favorites", "Favorites"],
@@ -149,6 +151,13 @@ function rebuildSourceOptions() {
  refs.source.value = options.some(([value]) => value === current) ? current : "all";
 }

+// 自然语言跟普通标签分类不是一个层级，按拼音排会掉到很后面；
+// 词库里这个分类名中英两种写法都有（Anima 分类是英文，手填的多是中文），都认。
+function isNaturalCategory(name) {
+  const value = String(name ?? "").trim().toLowerCase();
+  return value === "自然语言" || value === "natural language";
+}
+
 function rebuildCategoryOptions() {
  const current = refs.category.value;
  refs.category.replaceChildren();
@@ -156,7 +165,7 @@ function rebuildCategoryOptions() {
  all.value = "all";
  refs.category.appendChild(all);
  const categories = [...new Set(managerRows().map((row) => row.tag.category).filter(Boolean))].sort((a, b) => a.localeCompare(b, "zh-CN"));
-  for (const category of categories) {
+  for (const category of [...categories.filter(isNaturalCategory), ...categories.filter((name) => !isNaturalCategory(name))]) {
    const option = element("option", "", category);
    option.value = category;
    refs.category.appendChild(option);
@@ -218,9 +227,10 @@ function renderRows() {
  const category = refs.category.value;
  const source = refs.source.value;
  const rows = managerRows().filter((row) => {
+    if (source === "natural" && !row.tag.natural) return false;
    if (source === "favorites" && !isFavorite(row.tag.english)) return false;
    if (source.startsWith("pack:") && row.tag.pack_id !== source.slice(5)) return false;
-    if (!["all", "favorites"].includes(source) && !source.startsWith("pack:") && row.kind !== source) return false;
+    if (!["all", "favorites", "natural"].includes(source) && !source.startsWith("pack:") && row.kind !== source) return false;
    if (category !== "all" && row.tag.category !== category) return false;
    if (query) {
      const haystack = [row.tag.english, row.tag.chinese, row.tag.category, ...(row.tag.aliases ?? [])].join("\n").toLowerCase();
@@ -575,10 +585,58 @@ function openImportPreview(payload) {
  modal.addEventListener("mousedown", (event) => event.stopPropagation());
 }

-async function buildAssistantSection(section) {
-  section.replaceChildren();
+// 设置面板每一项 = 名称列 + 控件列，名称列宽度跟着设置项名字长短走：
+// 插件这两项一个「语言与外观」一个「助手设置」，字数不同，右边的内容就被推得差了一格。
+// CSS 里已经用 :has() 处理过一次，这里再兜一次——面板重建后内联样式会丢，
+// 而设置面板每次打开都会重新调用下面的 control，正好顺势对齐一次。
+// 只认「恰好两列、且第二列装着我们内容」的那一行，结构对不上就不动，免得误伤别的元素。
+function alignSettingRow(node) {
+  let row = node.parentElement;
+  while (row && row !== document.body) {
+    if (row.children.length === 2 && row.children[1].contains(node)) break;
+    row = row.parentElement;
+  }
+  if (!row || row === document.body) return;
+  const [name, control] = row.children;
+  name.style.display = "none";
+  control.style.flex = "1 1 0%";
+  control.style.minWidth = "0";
+  // ComfyUI 给控件列自带 justify-end：宽度一占满，它反而把内容整个推到右边。
+  // 改回靠左，并让包着我们的那一层也占满，否则内容仍会缩在右侧。
+  control.style.justifyContent = "flex-start";
+  if (node.parentElement && node.parentElement !== control) {
+    node.parentElement.style.flex = "1 1 0%";
+    node.parentElement.style.minWidth = "0";
+  }
+}
+
+function alignWhenMounted(node, frames = 30) {
+  if (node.isConnected) {
+    alignSettingRow(node);
+    return;
+  }
+  if (frames > 0) requestAnimationFrame(() => alignWhenMounted(node, frames - 1));
+}
+
+// 外观 / 语言从助手设置里拆出来，和它并列挂在「设置」面板的插件分类下。
+// 改动立刻生效并单独保存，不需要去点助手设置的保存按钮。
+let appearanceSettingsNode = null;
+
+function appearanceControl() {
+  if (appearanceSettingsNode) {
+    alignWhenMounted(appearanceSettingsNode);
+    return appearanceSettingsNode;
+  }
+  appearanceSettingsNode = element("div", "bpi-settings-form");
+  appearanceSettingsNode.appendChild(element("h3", "", "Appearance & Language"));
  const loading = element("div", "bpm-loading", "Loading settings...");
-  section.appendChild(loading);
+  appearanceSettingsNode.appendChild(loading);
+  buildAppearanceForm(appearanceSettingsNode, loading);
+  alignWhenMounted(appearanceSettingsNode);
+  return appearanceSettingsNode;
+}
+
+async function buildAppearanceForm(host, loading) {
  let config;
  try {
    config = await getAssistantConfig();
@@ -587,110 +645,11 @@ async function buildAssistantSection(section) {
    loading.classList.add("bpi-danger-text");
    return;
  }
-  manager.assistantLoaded = true;
-  applyBpiAppearance(config.appearance);
-  section.replaceChildren();
+  loading.remove();
  const form = element("form", "bpi-form");
-  form.addEventListener("submit", (event) => event.preventDefault());
-  const error = element("div", "bpi-status");
-  error.style.gridColumn = "1 / -1";
+  const status = element("div", "bpi-status");
+  status.style.gridColumn = "1 / -1";

-  // 翻译服务：翻译/解释按此选择；“翻译并优化”“优化为 Anima”恒走 AI
-  const serviceLabel = element("label", "", "Translation Service");
-  const service = element("select");
-  for (const [value, label] of [
-    ["dictionary", "Dictionary only (no API)"],
-    ["baidu", "Baidu Translate"],
-    ["ai", "AI (OpenAI-compatible / Ollama)"],
-  ]) {
-    const option = element("option", "", label);
-    option.value = value;
-    service.appendChild(option);
-  }
-  service.value = config.translate_service;
-  form.append(serviceLabel, service);
-  const serviceNote = element("div", "bpi-config-note", 'Translation/explanation uses this choice. "Translate & Optimize" and "Optimize to Anima" always use AI, unrelated here. To test AI connection, temporarily set translation service to AI then click test (key preserved).');
-  serviceNote.style.gridColumn = "1 / -1";
-  form.append(serviceNote);
-
-  // 百度翻译（独立留存，切翻译服务不清）
-  const baiduHead = element("div", "bpm-toolbar");
-  baiduHead.style.gridColumn = "1 / -1";
-  baiduHead.append(element("strong", "", "Baidu Translate"));
-  form.append(baiduHead);
-  const baiduAppId = field(form, "APP ID", "assistant-baidu-appid", config.baidu_appid, 'Baidu Cloud "General Text Translation" APP ID');
-  const baiduSecretKey = field(form, "Secret Key", "assistant-baidu-secret", "", config.baidu_secret_key_configured ? "Saved; leave empty to keep" : 'Baidu Cloud "General Text Translation" Secret Key');
-  baiduSecretKey.type = "password";
-  const clearBaiduLine = element("label", "bpm-toolbar");
-  const clearBaiduSecretKey = element("input");
-  clearBaiduSecretKey.type = "checkbox";
-  clearBaiduLine.append(clearBaiduSecretKey, element("span", "", "Clear saved Baidu secret"));
-  const baiduSpacer = element("span");
-  form.append(baiduSpacer, clearBaiduLine);
-  const baiduQps = field(form, "QPS (requests/second)", "assistant-baidu-qps", config.baidu_qps ?? 1, "1");
-  baiduQps.type = "number";
-  baiduQps.min = "0.1";
-  baiduQps.max = "50";
-  baiduQps.step = "0.5";
-  const baiduQpsHelp = [
-    "Baidu rate-limits General Text Translation by account tier. Enter your tier's number and the plugin queues requests to match:",
-    "· Unverified / Standard: 1",
-    "· Personal verified: 10",
-    "· Enterprise verified: higher, see the Baidu AI Cloud console",
-    "Find your tier under Baidu AI Cloud → Machine Translation → General Text Translation.",
-    "Too high and Baidu still returns 54003 (too many requests); too low is only slower, never an error.",
-    "If unsure, use 1.",
-  ].join("\n");
-  form.querySelector('label[for="bpi-field-assistant-baidu-qps"]')?.appendChild(helpMark(baiduQpsHelp));
-  const baiduQpsNote = element("div", "bpi-config-note", "How many requests per second Baidu may receive at most. Batch translation queues by this value and never exceeds it.");
-  baiduQpsNote.style.gridColumn = "1 / -1";
-  form.append(baiduQpsNote);
-
-  // AI 服务（优化恒走它；翻译在选 AI 时也走它）
-  const aiHead = element("div", "bpm-toolbar");
-  aiHead.style.gridColumn = "1 / -1";
-  aiHead.append(element("strong", "", "AI Service (Translate & Optimize / Optimize / AI Translate)"));
-  form.append(aiHead);
-  const aiProviderLabel = element("label", "", "AI Backend");
-  const aiProvider = element("select");
-  for (const [value, label] of [["openai_compatible", "OpenAI-compatible API"], ["ollama", "Ollama (local)"]]) {
-    const option = element("option", "", label);
-    option.value = value;
-    aiProvider.appendChild(option);
-  }
-  aiProvider.value = config.ai_provider;
-  form.append(aiProviderLabel, aiProvider);
-  const baseUrl = field(form, "API URL", "assistant-base-url", config.ai_base_url, "Optional for Ollama; e.g.: https://host/v1");
-  const model = field(form, "Model Name", "assistant-model", config.ai_model, "e.g.: qwen3:8b or provider model ID");
-  const presetLine = element("div", "bpm-toolbar");
-  presetLine.style.gridColumn = "1 / -1";
-  const lmStudioPreset = button("Use LM Studio Local Preset", () => {
-    aiProvider.value = "openai_compatible";
-    baseUrl.value = "http://127.0.0.1:1234/v1";
-    setText(error, "LM Studio default URL filled; select a loaded model ID, then Save & Test");
-    error.dataset.kind = "ok";
-  });
-  presetLine.append(lmStudioPreset, element("span", "bpi-config-note", "Connects to localhost:1234 by default; no external API required."));
-  form.append(presetLine);
-  const apiKey = field(form, "API Key", "assistant-api-key", "", config.ai_api_key_configured ? "Saved; leave empty to keep" : "Optional for local Ollama / LM Studio");
-  apiKey.type = "password";
-  const clearLine = element("label", "bpm-toolbar");
-  const clearApiKey = element("input");
-  clearApiKey.type = "checkbox";
-  clearLine.append(clearApiKey, element("span", "", "Clear saved API Key"));
-  const apiKeySpacer = element("span");
-  form.append(apiKeySpacer, clearLine);
-  const temperature = field(form, "Temperature", "assistant-temperature", config.ai_temperature, "0–2");
-  temperature.type = "number";
-  temperature.min = "0";
-  temperature.max = "2";
-  temperature.step = "0.05";
-  const timeout = field(form, "Timeout (s)", "assistant-timeout", config.ai_timeout_seconds, "5–600");
-  timeout.type = "number";
-  timeout.min = "5";
-  timeout.max = "600";
-
-  // 外观模式
  const appearanceLabel = element("label", "", "Appearance");
  const appearanceSelect = element("select");
  for (const [value, label] of [
@@ -718,31 +677,350 @@ async function buildAssistantSection(section) {
    languageSelect.appendChild(option);
  }
  languageSelect.value = config.language || "auto";
-  form.append(languageLabel, languageSelect);
  const languageNote = element("div", "bpi-config-note", "Choose the language for this plugin's own panels. \"Follow ComfyUI\" reads the ComfyUI language setting; \"Chinese\" and \"English\" force one. Applied to open panels right away; reload the page if anything still shows the old language.");
  languageNote.style.gridColumn = "1 / -1";
-  form.append(languageNote);
+  form.append(languageLabel, languageSelect, languageNote, status);
+  host.appendChild(form);

-  // 规则
-  const addRule = (labelText, value) => {
+  // 只提交这两项：后端 update 对没给的字段一律保留原值，不会动翻译服务和密钥
+  const save = async () => {
+    try {
+      const saved = await saveAssistantConfig({ appearance: appearanceSelect.value, language: languageSelect.value });
+      applyBpiAppearance(saved.appearance);
+      setLanguageMode(saved.language);
+      setText(status, "Appearance and language saved");
+      status.dataset.kind = "ok";
+    } catch (saveError) {
+      setText(status, saveError.message);
+      status.dataset.kind = "error";
+    }
+  };
+  appearanceSelect.addEventListener("change", save);
+  languageSelect.addEventListener("change", save);
+}
+
+// 助手设置直接长在 ComfyUI「设置」面板里（入口见下方 registerExtension 的 settings），
+// 这一个函数返回的就是设置项右侧那块控件。
+// 面板每次渲染都会重建控件容器，所以整块 DOM 只建一次后复用，输入到一半的内容不会丢。
+let assistantSettingsNode = null;
+
+function assistantSettingsControl() {
+  if (assistantSettingsNode) {
+    alignWhenMounted(assistantSettingsNode);
+    return assistantSettingsNode;
+  }
+  assistantSettingsNode = element("div", "bpi-settings-form");
+  assistantSettingsNode.appendChild(element("h3", "", "Assistant Settings"));
+  const loading = element("div", "bpm-loading", "Loading settings...");
+  assistantSettingsNode.appendChild(loading);
+  buildAssistantForm(assistantSettingsNode, loading);
+  alignWhenMounted(assistantSettingsNode);
+  return assistantSettingsNode;
+}
+
+async function buildAssistantForm(host, loading) {
+  let config;
+  try {
+    config = await getAssistantConfig();
+  } catch (error) {
+    setText(loading, error.message);
+    loading.classList.add("bpi-danger-text");
+    return;
+  }
+  applyBpiAppearance(config.appearance);
+  loading.remove();
+  const form = element("form", "bpi-form");
+  form.addEventListener("submit", (event) => event.preventDefault());
+  const error = element("div", "bpi-status");
+  error.style.gridColumn = "1 / -1";
+
+  // 翻译服务：翻译/解释按此选择；“翻译并优化”“优化”恒走 AI
+  const serviceLabel = element("label", "", "Translation Service");
+  const service = element("select");
+  for (const [value, label] of [
+    ["dictionary", "Dictionary only (no API)"],
+    ["baidu", "Baidu Translate"],
+    ["openai_compatible", "OpenAI-compatible API"],
+    ["ollama", "Ollama (local)"],
+  ]) {
+    const option = element("option", "", label);
+    option.value = value;
+    service.appendChild(option);
+  }
+  // 后端已把旧的 "ai" 拆成两个具体服务；这里再兜一次，避免下拉落空
+  service.value = config.translate_service === "ai" ? config.ai_provider || "openai_compatible" : config.translate_service;
+  form.append(serviceLabel, service);
+  const serviceNote = element("div", "bpi-config-note", 'Only this service\'s settings are shown below. "Translate & Optimize" and "Optimize" always use AI: choose OpenAI-compatible API or Ollama above to configure or test them. Switching service never clears saved keys.');
+  serviceNote.style.gridColumn = "1 / -1";
+  form.append(serviceNote);
+
+  // 按翻译服务分组显示：隐藏的字段不提交也不清空，切换前后值都在
+  const baiduOnly = [];
+  // AI 两个后端各一组：地址规则不同（Ollama 只认本机，OpenAI 远端必须 https），
+  // 规则也各一套 —— 本地小模型和云端大模型要写的提示词不一样
+  const openaiOnly = [];
+  const ollamaOnly = [];
+  // AI 整段（标题 + 后端下拉 + 两套具体配置 + 规则）只在翻译服务就是 AI 时出现；
+  // 选词库 / 百度时一段都不显示，但优化仍会走上次选过的那个后端
+  const aiOnly = [];
+  // 翻译服务里属于 AI 的两个值：选了它们，翻译服务自己就是 AI 后端
+  const AI_SERVICES = new Set(["openai_compatible", "ollama"]);
+  const currentBackend = () => (AI_SERVICES.has(service.value) ? service.value : aiProvider.value);
+  const addField = (list, ...args) => {
+    const input = field(form, ...args);
+    list.push(form.querySelector(`label[for="${input.id}"]`), input);
+    return input;
+  };
+  // 模型名称这一行：输入框照旧可以手写，旁边的按钮去问一次服务端有哪些模型，填进输入框自带的下拉里。
+  // 用原生 datalist 而不是弹窗：助手设置长在 ComfyUI 的「设置」面板里，浮层会被它盖住且点不动。
+  const addModelRow = (list, name, value, placeholder, options) => {
+    const label = element("label", "", "Model Name");
+    label.htmlFor = `bpi-field-${name}`;
+    const input = element("input");
+    input.id = label.htmlFor;
+    input.name = name;
+    input.value = value ?? "";
+    setPlaceholder(input, placeholder);
+    // datalist 本身不渲染，挂在 form 里跟着这块设置 DOM 一起活；选项只在本次会话内有效
+    const dataList = element("datalist");
+    dataList.id = `bpi-models-${name}`;
+    input.setAttribute("list", dataList.id);
+    let models = null;
+    const trigger = button("Fetch Models", () => pick());
+    const line = element("div", "bpi-model-line");
+    line.append(input, trigger);
+    form.append(label, line, dataList);
+    list.push(label, line, dataList);
+    const countLabel = () => (models ? `Fetch Models (${models.length})` : "Fetch Models");
+    const forget = () => {
+      models = null;
+      dataList.textContent = "";
+      setText(trigger, countLabel());
+    };
+    // 地址改了就地作废：下拉里不能留着上一个地址的模型
+    options.baseUrl().addEventListener("change", forget);
+    const pick = async () => {
+      trigger.disabled = true;
+      setText(trigger, models ? countLabel() : "Fetching...");
+      try {
+        const url = options.baseUrl().value.trim();
+        const keyInput = options.apiKey();
+        models = await listAssistantModels({
+          backend: options.backend,
+          base_url: url,
+          api_key: keyInput ? keyInput.value.trim() : "",
+        });
+        dataList.textContent = "";
+        for (const model of models) {
+          const option = element("option");
+          option.value = model; // 模型名是服务端原样返回值，不参与翻译
+          dataList.appendChild(option);
+        }
+        setText(trigger, countLabel());
+        input.focus();
+        setText(error, `Fetched ${models.length} models`);
+        error.dataset.kind = "ok";
+      } catch (fetchError) {
+        setText(trigger, countLabel());
+        setText(error, fetchError.message);
+        error.dataset.kind = "error";
+      } finally {
+        trigger.disabled = false;
+      }
+    };
+    return input;
+  };
+  // 清除已保存的密钥：开关样式，打开后要保存才生效。返回的 input 上挂了整行（`.row`）
+  const clearSwitch = (labelText) => {
+    const row = element("label", "bpi-switch-row");
+    const input = element("input");
+    input.type = "checkbox";
+    const track = element("span", "bpi-switch-track");
+    track.appendChild(element("span", "bpi-switch-thumb"));
+    row.append(input, track, element("span", "bpi-switch-text", labelText), element("span", "bpi-switch-hint", "Takes effect after saving"));
+    input.row = row;
+    return input;
+  };
+
+  // 百度翻译（独立留存，切翻译服务不清）
+  const baiduHead = element("div", "bpm-toolbar");
+  baiduHead.style.gridColumn = "1 / -1";
+  baiduHead.append(element("strong", "", "Baidu Translate"));
+  form.append(baiduHead);
+  baiduOnly.push(baiduHead);
+  const baiduAppId = addField(baiduOnly, "APP ID", "assistant-baidu-appid", config.baidu_appid, 'Baidu Cloud "General Text Translation" APP ID');
+  const baiduSecretKey = addField(baiduOnly, "Secret Key", "assistant-baidu-secret", "", config.baidu_secret_key_configured ? "Saved; leave empty to keep" : 'Baidu Cloud "General Text Translation" Secret Key');
+  baiduSecretKey.type = "password";
+  const clearBaiduSecretKey = clearSwitch("Clear saved Baidu secret");
+  const baiduSpacer = element("span");
+  form.append(baiduSpacer, clearBaiduSecretKey.row);
+  baiduOnly.push(baiduSpacer, clearBaiduSecretKey.row);
+  const baiduQps = addField(baiduOnly, "QPS (requests/second)", "assistant-baidu-qps", config.baidu_qps ?? 1, "1");
+  baiduQps.type = "number";
+  baiduQps.min = "0.1";
+  baiduQps.max = "50";
+  baiduQps.step = "0.5";
+  const baiduQpsHelp = [
+    "Baidu rate-limits General Text Translation by account tier. Enter your tier's number and the plugin queues requests to match:",
+    "· Unverified / Standard: 1",
+    "· Personal verified: 10",
+    "· Enterprise verified: higher, see the Baidu AI Cloud console",
+    "Find your tier under Baidu AI Cloud → Machine Translation → General Text Translation.",
+    "Too high and Baidu still returns 54003 (too many requests); too low is only slower, never an error.",
+    "If unsure, use 1.",
+  ].join("\n");
+  form.querySelector('label[for="bpi-field-assistant-baidu-qps"]')?.appendChild(helpMark(baiduQpsHelp));
+  const baiduQpsNote = element("div", "bpi-config-note", "How many requests per second Baidu may receive at most. Batch translation queues by this value and never exceeds it.");
+  baiduQpsNote.style.gridColumn = "1 / -1";
+  form.append(baiduQpsNote);
+  baiduOnly.push(baiduQpsNote);
+
+  // AI 服务（优化恒走它；翻译在选 AI 时也走它）
+  const aiHead = element("div", "bpm-toolbar");
+  aiHead.style.gridColumn = "1 / -1";
+  aiHead.append(element("strong", "", "AI Service (Translate & Optimize / Optimize / AI Translate)"));
+  form.append(aiHead);
+  aiOnly.push(aiHead);
+  const aiProviderLabel = element("label", "", "AI Backend");
+  const aiProvider = element("select");
+  for (const [value, label] of [["openai_compatible", "OpenAI-compatible API"], ["ollama", "Ollama (local)"]]) {
+    const option = element("option", "", label);
+    option.value = value;
+    aiProvider.appendChild(option);
+  }
+  aiProvider.value = config.ai_provider;
+  form.append(aiProviderLabel, aiProvider);
+  aiOnly.push(aiProviderLabel, aiProvider);
+
+  // OpenAI 兼容：远端必须 https，本机回环才允许 http
+  const openaiHead = element("div", "bpm-toolbar");
+  openaiHead.style.gridColumn = "1 / -1";
+  openaiHead.append(element("strong", "", "OpenAI-compatible API"));
+  form.append(openaiHead);
+  openaiOnly.push(openaiHead);
+  const openaiBaseUrl = addField(openaiOnly, "API URL", "assistant-openai-base-url", config.openai_base_url, "https://host/v1");
+  const openaiModel = addModelRow(openaiOnly, "assistant-openai-model", config.openai_model, "provider model ID", {
+    backend: "openai_compatible",
+    baseUrl: () => openaiBaseUrl,
+    apiKey: () => openaiApiKey,
+  });
+  const presetLine = element("div", "bpm-toolbar");
+  presetLine.style.gridColumn = "1 / -1";
+  const lmStudioPreset = button("Use LM Studio Local Preset", () => {
+    aiProvider.value = "openai_compatible";
+    openaiBaseUrl.value = "http://127.0.0.1:1234/v1";
+    setText(error, "LM Studio default URL filled; select a loaded model ID, then Save & Test");
+    error.dataset.kind = "ok";
+  });
+  presetLine.append(lmStudioPreset, element("span", "bpi-config-note", "Connects to localhost:1234 by default; no external API required."));
+  form.append(presetLine);
+  openaiOnly.push(presetLine);
+  const openaiApiKey = addField(openaiOnly, "API Key", "assistant-openai-api-key", "", config.openai_api_key_configured ? "Saved; leave empty to keep" : "Optional for local LM Studio");
+  openaiApiKey.type = "password";
+  const clearApiKey = clearSwitch("Clear saved API Key");
+  const apiKeySpacer = element("span");
+  form.append(apiKeySpacer, clearApiKey.row);
+  openaiOnly.push(apiKeySpacer, clearApiKey.row);
+  const openaiTemperature = addField(openaiOnly, "Temperature", "assistant-openai-temperature", config.openai_temperature, "0–2");
+  openaiTemperature.type = "number";
+  openaiTemperature.min = "0";
+  openaiTemperature.max = "2";
+  openaiTemperature.step = "0.05";
+  const openaiTimeout = addField(openaiOnly, "Timeout (s)", "assistant-openai-timeout", config.openai_timeout_seconds, "5–600");
+  openaiTimeout.type = "number";
+  openaiTimeout.min = "5";
+  openaiTimeout.max = "600";
+  const openaiUrlNote = element("div", "bpi-config-note", "Remote addresses must use https; http is only accepted for this machine. Private, reserved and link-local addresses are rejected. Keys are never sent in clear text over the network.");
+  openaiUrlNote.style.gridColumn = "1 / -1";
+  form.append(openaiUrlNote);
+  openaiOnly.push(openaiUrlNote);
+
+  // Ollama：只允许本机回环，没有密钥
+  const ollamaHead = element("div", "bpm-toolbar");
+  ollamaHead.style.gridColumn = "1 / -1";
+  ollamaHead.append(element("strong", "", "Ollama (local)"));
+  form.append(ollamaHead);
+  ollamaOnly.push(ollamaHead);
+  const ollamaBaseUrl = addField(ollamaOnly, "API URL", "assistant-ollama-base-url", config.ollama_base_url, "http://127.0.0.1:11434");
+  const ollamaModel = addModelRow(ollamaOnly, "assistant-ollama-model", config.ollama_model, "e.g.: qwen3:8b", {
+    backend: "ollama",
+    baseUrl: () => ollamaBaseUrl,
+    apiKey: () => null,
+  });
+  const ollamaPresetLine = element("div", "bpm-toolbar");
+  ollamaPresetLine.style.gridColumn = "1 / -1";
+  const ollamaPreset = button("Use Ollama Local Preset", () => {
+    aiProvider.value = "ollama";
+    ollamaBaseUrl.value = "http://127.0.0.1:11434";
+    setText(error, "Ollama default URL filled; select a pulled model ID, then Save & Test");
+    error.dataset.kind = "ok";
+  });
+  ollamaPresetLine.append(ollamaPreset, element("span", "bpi-config-note", "Connects to localhost:11434 by default; no external API required."));
+  form.append(ollamaPresetLine);
+  ollamaOnly.push(ollamaPresetLine);
+  const ollamaTemperature = addField(ollamaOnly, "Temperature", "assistant-ollama-temperature", config.ollama_temperature, "0–2");
+  ollamaTemperature.type = "number";
+  ollamaTemperature.min = "0";
+  ollamaTemperature.max = "2";
+  ollamaTemperature.step = "0.05";
+  const ollamaTimeout = addField(ollamaOnly, "Timeout (s)", "assistant-ollama-timeout", config.ollama_timeout_seconds, "5–600");
+  ollamaTimeout.type = "number";
+  ollamaTimeout.min = "5";
+  ollamaTimeout.max = "600";
+  const ollamaUrlNote = element("div", "bpi-config-note", "Ollama only accepts a loopback address on this machine; anything else is rejected before a request is sent.");
+  ollamaUrlNote.style.gridColumn = "1 / -1";
+  form.append(ollamaUrlNote);
+  ollamaOnly.push(ollamaUrlNote);
+
+  // 规则：只有 AI 会读，两个后端各一套（server.py 按当前后端取对应的那套）
+  const addRule = (list, labelText, value) => {
    const labelElement = element("label", "", labelText);
    const textarea = element("textarea");
    textarea.value = value;
    form.append(labelElement, textarea);
-    return { labelElement, textarea };
+    list.push(labelElement, textarea);
+    return textarea;
+  };
+  const rules = {
+    openai_compatible: {
+      translation: addRule(openaiOnly, "Translation Rule", config.openai_translation_rule),
+      translateOptimize: addRule(openaiOnly, "Translate & Optimize Rule", config.openai_translate_optimize_rule),
+      optimization: addRule(openaiOnly, "Optimize Rule", config.openai_optimization_rule),
+      natural: addRule(openaiOnly, "Natural Language Rule", config.openai_natural_rule),
+    },
+    ollama: {
+      translation: addRule(ollamaOnly, "Translation Rule", config.ollama_translation_rule),
+      translateOptimize: addRule(ollamaOnly, "Translate & Optimize Rule", config.ollama_translate_optimize_rule),
+      optimization: addRule(ollamaOnly, "Optimize Rule", config.ollama_optimization_rule),
+      natural: addRule(ollamaOnly, "Natural Language Rule", config.ollama_natural_rule),
+    },
  };
-  const translationRule = addRule("Translation Rule", config.translation_rule);
-  const translateOptimizeRule = addRule("Translate & Optimize Rule", config.translate_optimize_rule);
-  const optimizationRule = addRule("Optimize Rule", config.optimization_rule);
  form.appendChild(error);

-  section.appendChild(form);
-  section.appendChild(element("div", "bpi-config-note", "Settings are stored in the ComfyUI user directory. AI key and Baidu secret are retained independently; switching service never clears either. AI key is only cleared when AI backend/URL changes or on a different machine. Neither is written to workflows or returned to the browser."));
-  const actions = element("div", "bpm-assistant-actions");
+  const applyServiceVisibility = () => {
+    const value = service.value;
+    // 只有翻译服务本身就是 AI 时才展开 AI 整段；选词库 / 百度时一段都不显示
+    for (const node of baiduOnly) node.hidden = value !== "baidu";
+    for (const node of aiOnly) node.hidden = !AI_SERVICES.has(value);
+    for (const node of openaiOnly) node.hidden = value !== "openai_compatible";
+    for (const node of ollamaOnly) node.hidden = value !== "ollama";
+  };
+  service.addEventListener("change", applyServiceVisibility);
+  aiProvider.addEventListener("change", applyServiceVisibility);
+  applyServiceVisibility();
+
+  host.appendChild(form);
+  host.appendChild(element("div", "bpi-config-note", "Settings are stored in the ComfyUI user directory. AI key and Baidu secret are retained independently; switching service never clears either. AI key is only cleared when AI backend/URL changes or on a different machine. Neither is written to workflows or returned to the browser."));
+  const actions = element("div", "bpi-form-actions");
+  // 只恢复当前所选后端那套规则，另一套不动
  const resetButton = button("Reset to Default Rules", () => {
-    translationRule.textarea.value = config.default_translation_rule;
-    translateOptimizeRule.textarea.value = config.default_translate_optimize_rule;
-    optimizationRule.textarea.value = config.default_optimization_rule;
+    const backend = currentBackend();
+    const prefix = backend === "ollama" ? "ollama" : "openai";
+    const active = rules[backend] ?? rules.openai_compatible;
+    active.translation.value = config[`default_${prefix}_translation_rule`];
+    active.translateOptimize.value = config[`default_${prefix}_translate_optimize_rule`];
+    active.optimization.value = config[`default_${prefix}_optimization_rule`];
+    active.natural.value = config[`default_${prefix}_natural_rule`];
    setText(error, "Defaults restored in editor; click Save to apply");
    error.dataset.kind = "ok";
  });
@@ -752,7 +1030,6 @@ async function buildAssistantSection(section) {
    error.dataset.kind = "busy";
    try {
      config = await saveAssistantConfig(payload());
-      setLanguageMode(config.language);
      const result = await testAssistantConnection();
      setText(error, `Connection OK: “${result.message}”`);
      error.dataset.kind = "ok";
@@ -767,8 +1044,6 @@ async function buildAssistantSection(section) {
    saveButton.disabled = true;
    try {
      config = await saveAssistantConfig(payload());
-      applyBpiAppearance(config.appearance);
-      setLanguageMode(config.language);
      setText(error, "Assistant settings saved");
      error.dataset.kind = "ok";
    } catch (saveError) {
@@ -779,25 +1054,34 @@ async function buildAssistantSection(section) {
    }
  }, "bpi-primary");
  actions.append(resetButton, testButton, saveButton);
-  section.appendChild(actions);
+  host.appendChild(actions);
+  // 两套 AI 配置和两套规则都提交（后端对没给的字段才保留原值，这里给全最省心）；
+  // 外观 / 语言已在「设置」面板里独立成项（appearanceControl），这里既不提交也不应用
  const payload = () => ({
    translate_service: service.value,
-    ai_provider: aiProvider.value,
-    ai_base_url: baseUrl.value.trim(),
-    ai_model: model.value.trim(),
-    ai_api_key: apiKey.value.trim(),
-    clear_ai_api_key: clearApiKey.checked,
+    ai_provider: currentBackend(),
+    openai_base_url: openaiBaseUrl.value.trim(),
+    openai_model: openaiModel.value.trim(),
+    openai_api_key: openaiApiKey.value.trim(),
+    clear_openai_api_key: clearApiKey.checked,
+    openai_temperature: openaiTemperature.value,
+    openai_timeout_seconds: openaiTimeout.value,
+    ollama_base_url: ollamaBaseUrl.value.trim(),
+    ollama_model: ollamaModel.value.trim(),
+    ollama_temperature: ollamaTemperature.value,
+    ollama_timeout_seconds: ollamaTimeout.value,
    baidu_appid: baiduAppId.value.trim(),
    baidu_secret_key: baiduSecretKey.value.trim(),
    clear_baidu_secret_key: clearBaiduSecretKey.checked,
    baidu_qps: baiduQps.value.trim(),
-    ai_temperature: temperature.value,
-    ai_timeout_seconds: timeout.value,
-    appearance: appearanceSelect.value,
-    language: languageSelect.value,
-    translation_rule: translationRule.textarea.value,
-    translate_optimize_rule: translateOptimizeRule.textarea.value,
-    optimization_rule: optimizationRule.textarea.value,
+    openai_translation_rule: rules.openai_compatible.translation.value,
+    openai_translate_optimize_rule: rules.openai_compatible.translateOptimize.value,
+    openai_optimization_rule: rules.openai_compatible.optimization.value,
+    openai_natural_rule: rules.openai_compatible.natural.value,
+    ollama_translation_rule: rules.ollama.translation.value,
+    ollama_translate_optimize_rule: rules.ollama.translateOptimize.value,
+    ollama_optimization_rule: rules.ollama.optimization.value,
+    ollama_natural_rule: rules.ollama.natural.value,
  });
 }

@@ -809,7 +1093,6 @@ function setSection(section) {
    refs.sections?.querySelector(`.bpm-section[data-section="${id}"]`)?.toggleAttribute("hidden", id !== section);
  }
  if (section === "tag-manager") rebuildTagManagerSection();
-  if (section === "assistant" && !manager.assistantLoaded) buildAssistantSection(refs.assistantSection);
  if (section === "favorites" && !manager.favoritesLoaded) {
    manager.favoritesLoaded = true;
    refs.sections.appendChild(buildFavoritesSection());
@@ -1109,8 +1392,21 @@ function inspectorNodes() {
  return (app.graph?._nodes ?? []).filter((node) => node._bilingualPromptInspector?.detailsBody);
 }

+// 用户在画布上双击节点顶部改名后，列表里显示他自己起的名字；
+// 名字后面保留节点号，免得两个节点改成同一个名字时认不出来。
+function customNodeTitle(node) {
+  const current = String(node.title ?? "").trim();
+  // defaultNodeTitle 是节点创建时记下的出厂标题；拿不到就退回节点类型的默认标题，
+  // 两者都拿不到时不做判断，免得把出厂标题当成用户改过的名字。
+  const factory = String(node._bilingualPromptInspector?.defaultNodeTitle ?? node.constructor?.title ?? "").trim();
+  if (!current || !factory || current === factory) return null;
+  return current;
+}
+
 function tagNodeLabel(node) {
  const text = String(node.widgets?.find((widget) => widget.name === "text")?.value ?? "").trim();
+  const custom = customNodeTitle(node);
+  if (custom) return `${custom} · ${t(`Node #${node.id}`)}`;
  return `${t(`Node #${node.id}`)}: ${text ? text.slice(0, 30) : t("(empty prompt)")}`;
 }

@@ -1186,8 +1482,6 @@ function build() {
  refs.sections.appendChild(refs.tagManagerSection);
  refs.sections.appendChild(buildTagsSection());
  refs.sections.appendChild(buildPacksSection());
-  refs.assistantSection = buildAssistantPlaceholder();
-  refs.sections.appendChild(refs.assistantSection);
  root.append(head, refs.tabs, refs.sections);
  root.addEventListener("mousedown", (event) => event.stopPropagation());
  root.addEventListener("wheel", (event) => event.stopPropagation(), { passive: true });
@@ -1303,14 +1597,6 @@ function buildPacksSection() {
  return section;
 }

-function buildAssistantPlaceholder() {
-  const section = element("div", "bpm-section");
-  section.dataset.section = "assistant";
-  section.hidden = true;
-  section.appendChild(element("div", "bpm-loading", "Settings load on first visit..."));
-  return section;
-}
-
 function renderTab(container) {
  if (!root) build();
  container.replaceChildren(root);
@@ -1341,6 +1627,24 @@ getAssistantConfig()

 app.registerExtension({
  name: "ComfyUI.BilingualPromptInspector.ManagerPanel",
+  // 助手设置整体挂在 ComfyUI 的「设置」面板里：项名和分类由前端原生渲染，
+  // 不走插件的 t()，直接写中文；控件本体是 assistantSettingsControl() 返回的那块 DOM。
+  settings: [
+    {
+      id: "BilingualPromptInspector.Appearance",
+      name: "语言与外观",
+      type: appearanceControl,
+      defaultValue: "",
+      category: ["双语提示词检查器", "语言与外观"],
+    },
+    {
+      id: "BilingualPromptInspector.AssistantSettings",
+      name: "助手设置",
+      type: assistantSettingsControl,
+      defaultValue: "",
+      category: ["双语提示词检查器", "助手设置"],
+    },
+  ],
  async setup() {
    if (!app.extensionManager?.registerSidebarTab) {
      console.warn("[BilingualPromptInspector] Current frontend does not support sidebar; panel unavailable");
@@ -1352,7 +1656,7 @@ app.registerExtension({
      // 侧边栏标签显示的就是插件名，直接写中文：注册发生在启动时，
      // 语言设置要等后端配置返回才落地，走 t() 会先渲染成英文且不会再刷新。
      title: "提示词翻译与管理",
-      tooltip: "提示词翻译与管理：标签管理、词库、词包与助手设置",
+      tooltip: "提示词翻译与管理：标签管理、词库、词包与收藏",
      type: "custom",
      render: renderTab,
      destroy: destroyTab,
@@ -20,6 +20,10 @@ function stripEmphasis(value) {
  return current;
 }

+// 没命中任何词库时中文列填的占位文案。它只是「还没翻译」的提示，不是一条真正的译文，
+// 所以任何往词库回写的地方都要先拿它比对，别把占位当成中文名存进去。
+export const UNKNOWN_CHINESE = "未收录";
+
 export function analyzeSegment(raw) {
  const display = raw.trim();
  if (!display) return null;
@@ -184,6 +188,43 @@ export function replacePromptTokenWeight(text, token, weight) {
  };
 }

+/**
+ * Replace a raw slice of the prompt with whatever the user typed.
+ *
+ * Used by the "cut a piece out, edit it, paste it back" editor: the slice is a
+ * span of the ORIGINAL text (usually first-selected.start .. last-selected.end),
+ * so whatever sits between two selected tags comes along in the editor and is
+ * replaced with it.  No parsing or validation happens here — commas the user
+ * typed simply get re-split by the next parse.
+ *
+ * Clearing the slice means deleting it, so the seam is tidied to avoid leaving
+ * `a, , b` behind: the comma on the left is dropped, or the one on the right
+ * when the slice started at the very beginning.  A plain line break is left
+ * alone so multi-line layouts survive.
+ *
+ * The slice keeps its own leading / trailing whitespace in the text (that is
+ * where the space after a comma lives), so it is put back around what the user
+ * typed — otherwise the seam would come out as `a,b`.
+ */
+export function replacePromptSpan(text, start, end, replacement) {
+  const value = String(text ?? "");
+  const from = Math.max(0, Math.min(value.length, Number.isInteger(start) ? start : 0));
+  const to = Math.max(from, Math.min(value.length, Number.isInteger(end) ? end : from));
+  const span = value.slice(from, to);
+  const typed = String(replacement ?? "").trim();
+  const insert = typed ? `${span.match(/^\s*/)[0]}${typed}${span.match(/\s*$/)[0]}` : "";
+  if (span === insert) return { text: value, cursor: to, changed: false };
+  let left = value.slice(0, from);
+  let right = value.slice(to);
+  if (!insert) {
+    const trimmedLeft = left.replace(/[ \t]*,\r?\n?[ \t]*$/, "");
+    if (trimmedLeft !== left) left = trimmedLeft;
+    else right = right.replace(/^[\s,]+/, "");
+  }
+  const nextText = left + insert + right;
+  return { text: nextText, cursor: Math.min(from + insert.length, nextText.length), changed: true };
+}
+
 /**
 * Move one parsed prompt token to a new position in the token order.
 *
@@ -409,6 +450,15 @@ function looksLikeTagSuffix(raw) {
  return true;
 }

+/**
+ * 标签里的空格超过 3 个（不含 3 个）时，它多半是被当成标签用的自然语言短语，
+ * 界面上标成「疑似自然语言」提醒确认。真正的 Danbooru 标签很少超过 3 个空格。
+ */
+export function looksSuspiciousNatural(raw) {
+  const value = String(raw ?? "").trim();
+  return (value.match(/ /g) ?? []).length > 3;
+}
+
 function trailingTagStart(segments, afterIndex) {
  let start = segments.length;
  while (start > afterIndex && looksLikeTagSuffix(segments[start - 1].raw)) start -= 1;
@@ -431,7 +481,10 @@ export function detectInputMode(text, requestedMode = "auto") {

  const words = value.match(/[A-Za-z]+(?:[-'][A-Za-z]+)*/g) ?? [];
  const commaSegments = splitPrompt(value);
-  const hasSentenceEnd = /[。！？.!?](?:\s|$)/.test(value);
+  // 权重小数和版本号里的句点（":1.2"、"1."）以及权重括号后的句点（"(...:1.2)."）都不是句子结束。
+  // 先把它们中和掉再判断，否则整串标签会只因为一个小数点被当成自然语言段落。
+  const sentenceEndCandidate = value.replace(/\d\s*[。！？.!?]|[)\]][。！？.!?]/g, (match) => match.replace(/[。！？.!?]/, ""));
+  const hasSentenceEnd = /[。！？.!?](?:\s|$)/.test(sentenceEndCandidate);
  const hasNaturalGrammar = /\b(?:she|he|they|who|which|is|are|was|were|wearing|holding|standing|sitting|with|while|under|inside|outside|in the|on the)\b/i.test(value);
  const longCommaSegment = commaSegments.some((item) => (item.raw.match(/[A-Za-z]+(?:[-'][A-Za-z]+)*/g) ?? []).length >= 10);
  const strongNaturalStart = commaSegments.findIndex((item) => looksLikeNaturalClause(item.raw));
@@ -448,6 +501,15 @@ export function detectInputMode(text, requestedMode = "auto") {
  if (commaSegments.length >= 3 && naturalStart === 0 && naturalEnd < commaSegments.length) {
    return { mode: "mixed", reason: "Natural language followed by tag list", confidence: "high", naturalStart, naturalEnd };
  }
+  // 上面没成立通常是尾部定界没认出标签流：trailingTagStart 还要求后缀里有 masterpiece /
+  // best quality 那类明星信号词，标签写得偏冷门时一个都命中不了。这时不该退化成整段自然语言——
+  // 开头那句是自然语言，后面明明是标签流，至少把开头的句子切出来。
+  if (naturalStart === 0 && commaSegments.length >= 4 && naturalEnd >= commaSegments.length && longCommaSegment) {
+    const tailTags = commaSegments.slice(1).filter((item) => looksLikeTagSuffix(item.raw)).length;
+    if (tailTags >= 4) {
+      return { mode: "mixed", reason: "Natural language followed by tag list", confidence: "medium", naturalStart, naturalEnd: naturalStart + 1 };
+    }
+  }
  if (commaSegments.length >= 3 && naturalStart > 0) {
    const shortBefore = commaSegments.slice(0, naturalStart).filter((item) =>
      (item.raw.match(/[A-Za-z]+(?:[-'][A-Za-z]+)*/g) ?? []).length <= 6
@@ -559,13 +621,30 @@ export function validateTranslationResult(source, translated, options = {}) {
  return { ok: true, text: result };
 }

+// 插件写死的默认分类：分到它等于还没分类。parser 是纯逻辑模块，不引 UI 那层，
+// 所以这个常量和 bpi_shared.js 里的那份各存一份，改要一起改。
+const UNCATEGORIZED_CATEGORY = "Uncategorized";
+
+// 可信度：徽章只显示「未判定 / 低 / 中 / 高」四档，为什么是这一档写在 tooltip 里。
+// 判据从粗到细：查不到 → 还没入库 → 没校验 → 没分类 → 有分类。
 function confidenceFor(status, entry) {
-  if (status === "machine") return { confidence: "low", confidenceLabel: "Low · pending" };
-  if (status === "unverified") return { confidence: "medium", confidenceLabel: "Medium · pending" };
-  if (status === "unknown") return { confidence: "none", confidenceLabel: "Not judged" };
-  if (status === "special") return { confidence: "high", confidenceLabel: "High" };
-  if (entry?.source === "user" || entry?.verified !== false) return { confidence: "high", confidenceLabel: "High" };
-  return { confidence: "medium", confidenceLabel: "Medium" };
+  if (status === "unknown") {
+    return { confidence: "none", confidenceLabel: "Not judged", confidenceReason: "Not in any dictionary, so it cannot be judged" };
+  }
+  if (status === "machine") {
+    return { confidence: "low", confidenceLabel: "Low", confidenceReason: "Machine translation, not saved to the dictionary or confirmed yet" };
+  }
+  if (status === "special") {
+    return { confidence: "high", confidenceLabel: "High", confidenceReason: "Syntax kept as-is; nothing to translate" };
+  }
+  if (entry?.verified === false) {
+    return { confidence: "medium", confidenceLabel: "Medium", confidenceReason: "In the dictionary, but not marked as verified" };
+  }
+  const category = String(entry?.category ?? "").trim();
+  if (category && category !== UNCATEGORIZED_CATEGORY) {
+    return { confidence: "high", confidenceLabel: "High", confidenceReason: `In the dictionary, verified, category “${category}”` };
+  }
+  return { confidence: "medium", confidenceLabel: "Medium", confidenceReason: "In the dictionary, but no category set" };
 }

 export function parsePrompt(text, dictionaryIndex, machineTranslations = new Map(), options = {}) {
@@ -590,7 +669,7 @@ export function parsePrompt(text, dictionaryIndex, machineTranslations = new Map
    const dictionaryEntry = dictionaryIndex.get(key);
    const machine = machineTranslations.get(key);
    let status = "unknown";
-    let chinese = "未收录";
+    let chinese = UNKNOWN_CHINESE;
    if (dictionaryEntry) {
      status = dictionaryEntry.verified === false ? "unverified" : "verified";
      chinese = dictionaryEntry.chinese;
@@ -608,6 +687,9 @@ export function parsePrompt(text, dictionaryIndex, machineTranslations = new Map
      chinese = "自然语言片段（待翻译或确认）";
    }
    const confidence = confidenceFor(status, dictionaryEntry);
+    const suspectedNatural = analyzed.syntax === "tag" &&
+      segment.segmentKind !== "natural" &&
+      looksSuspiciousNatural(segment.raw);
    return {
      ...segment,
      ...analyzed,
@@ -615,6 +697,7 @@ export function parsePrompt(text, dictionaryIndex, machineTranslations = new Map
      key,
      chinese,
      status,
+      suspectedNatural,
      ...confidence,
      inputMode: detected.mode,
      entry: dictionaryEntry ?? null,
@@ -658,6 +741,23 @@ const CONFLICT_GROUPS = [
  { left: ["full body"], right: ["extreme close-up", "close-up"], label: "Full body vs close-up" },
 ];

+// 重复分组：按 token.key 聚拢，只保留出现 2 次以上的组，组内按原文顺序。
+// 问题行的「重复标签」提示和「唯一」开关都从这里取数，两边口径不会打架。
+// 口径沿用 key：忽略大小写、连续空格、强调括号与权重（(1girl:1.3) ≡ 1girl）；
+// 下划线按一个空格算，所以 sm_ile ≠ smile。BREAK 与 <lora:…> 不算重复标签。
+export function duplicateGroups(tokens) {
+  const groups = new Map();
+  for (const token of tokens ?? []) {
+    if (!token.key || token.syntax !== "tag") continue;
+    const list = groups.get(token.key) ?? [];
+    list.push(token);
+    groups.set(token.key, list);
+  }
+  return [...groups]
+    .filter(([, list]) => list.length > 1)
+    .map(([key, list]) => ({ key, tokens: list }));
+}
+
 export function analyzePromptSyntax(text, tokens, modeInfo = detectInputMode(text)) {
  const value = String(text ?? "");
  const issues = bracketIssues(value);
@@ -677,8 +777,8 @@ export function analyzePromptSyntax(text, tokens, modeInfo = detectInputMode(tex
      issues.push(issue("warning", "unusual-weight", `Weight ${token.weight} outside common range 0–3`, [token.key]));
    }
  }
-  for (const [key, matches] of keys) {
-    if (matches.length > 1) issues.push(issue("warning", "duplicate", `Duplicate tag: “${matches[0].term}”`, [key]));
+  for (const { key, tokens: matches } of duplicateGroups(tokens ?? [])) {
+    issues.push(issue("warning", "duplicate", `Duplicate tag: “${matches[0].term}”`, [key]));
  }

  const present = new Set(keys.keys());
@@ -1,3 +1,4 @@
+from .duplicate_check import dedupe_prompt, find_duplicates, format_duplicates
 from .server import announce_upstream, current_prompt_id, upstream_gate


@@ -52,3 +53,43 @@ class BilingualPromptInspector:
        released = await upstream_gate.wait(node_id, future)
        # 前端放行时带回的是节点上编辑后的文本；无人接管（无浏览器或 API 调用）时原样透传。
        return (released if isinstance(released, str) else prompt,)
+
+
+class DuplicateChecker:
+    """
+    @title: 重复检查
+    @nickname: Duplicate Check
+    @description: 检查提示词里的重复标签：节点上用淡橙色提示有几处重复，「提示词」输出直通下游，「重复项」输出可接预览任意查看。开关打开时只保留第一次出现的标签。
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "text": ("STRING", {"forceInput": True}),
+            },
+            "optional": {
+                "keep_first": ("BOOLEAN", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING", "STRING")
+    RETURN_NAMES = ("提示词", "重复项")
+    FUNCTION = "check"
+    CATEGORY = "text/prompt"
+    DESCRIPTION = "Experimental duplicate tag checker: flags repeated tags with a soft amber tint, passes the prompt through, and exposes the duplicated tags for preview."
+
+    def check(self, text, keep_first=False):
+        value = str(text or "")
+        report = find_duplicates(value)
+        # 开关没开就原样直通，连格式都不重写，免得改了用户没要求改的东西。
+        prompt = dedupe_prompt(value) if keep_first else value
+        # ui 里每个值都必须是列表：execution.py 会把它们展开拼接成一个列表，
+        # 传裸 int 会直接抛 "'int' object is not iterable"。
+        return {
+            "ui": {
+                "duplicate_count": [report["extra"]],
+                "duplicate_terms": [item["term"] for item in report["terms"]],
+            },
+            "result": (prompt, format_duplicates(report)),
+        }
@@ -1,6 +1,6 @@
 {
  "name": "comfyui-bilingual-prompt-inspector",
-  "version": "1.2.10",
+  "version": "2.1.9",
  "description": "ComfyUI bilingual prompt inspection, translation, Anima sorting, and on-demand Danbooru dictionary extension.",
  "private": true,
  "type": "module",
@@ -1,7 +1,8 @@
 [project]
 name = "prompt-translator-manager"
-version = "1.2.10"
-description = "ComfyUI prompt node with translation and management: bilingual tag view, drag-to-reorder, temporary hiding, multiple translation services and a sidebar dictionary manager."
+version = "2.1.9"
+# 与 Comfy Registry 页面上显示的描述保持一致： publish 会用这里的值覆盖网页上的手改
+description = "ComfyUI prompt node with translation and management: bilingual tag view, drag-to-reorder, temporary hiding, multiple translation services and a sidebar dictionary manager.\nComfyUI 提示节点，支持翻译和管理：双语标签视图、拖拽排序、临时隐藏、多翻译服务和侧边栏词典管理器。"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"
 dependencies = []
@@ -23,6 +24,6 @@ requires-comfyui = ">=0.3.40"
 # 和 __init__.py 的 WEB_DIRECTORY（key = 模块名），两者指向同一个 js/ 目录时
 # /extensions 接口会把每个 js 列两遍，前端加载两次 → "Extension already registered"。
 # 前端资源统一由 __init__.py 的 WEB_DIRECTORY = "./js" 提供（各版本 ComfyUI 都支持）。
-# 图标与横幅（可选，填好图片的公开 URL 后取消注释即可）
-# Icon = "https://..."    # 方形，≤400×400，PNG/SVG/JPG/GIF
-# Banner = "https://..."  # 21:9 横幅，PNG/SVG/JPG/GIF
+# 图标与横幅
+Icon = "https://raw.githubusercontent.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector/main/assets/icon.png"      # 方形 400×400（缩过的小图）
+Banner = "https://raw.githubusercontent.com/mengshengzhijie/ComfyUI-Bilingual-Prompt-Inspector/main/assets/banner.png"  # 原图 941×941（未缩放），站点会自行裁切并加背景虚化
@@ -14,31 +14,44 @@ from comfy import model_management
 from server import PromptServer

 from .assistant_store import (
+    AI_PROVIDER_PREFIXES,
+    AI_SERVICES,
    BAIDU_ERROR_MESSAGES,
    BAIDU_QPS_DEFAULT,
+    BAIDU_TIMEOUT_SECONDS,
    BAIDU_TRANSLATE_ENDPOINT,
    AssistantStore,
    baidu_split_query,
    baidu_translate_params,
    dictionary_translate,
+    normalize_ai_base_url,
+    ollama_tags_endpoint,
    openai_chat_endpoint,
+    openai_models_endpoint,
+    parse_json_body,
    sanitize_anima_prompt,
+    sanitize_natural_prompt,
    translation_direction,
+    upstream_error_message,
 )
 from .dictionary_store import DictionaryStore, normalize_key
+from .duplicate_check import find_duplicates
 from .saved_prompt_store import (
    MAX_IMAGE_BYTES,
    SavedPromptStore,
 )
+from .token_color_store import TokenColorStore


 store = DictionaryStore()
+token_color_store = TokenColorStore()
 assistant_store = AssistantStore()
 assistant_store.config()  # 进程启动时立即清除与当前安装身份不匹配的旧配置和凭据

 _SESSION_TOKEN = secrets.token_urlsafe(32)
 _MAX_JSON_BYTES = 8 * 1024 * 1024
 _MAX_ASSISTANT_JSON_BYTES = 256 * 1024
+_MODEL_LIST_TIMEOUT_SECONDS = 20
 _ASSISTANT_RATE_WINDOW = 60.0
 _ASSISTANT_RATE_LIMIT = 30
 _assistant_requests = deque()
@@ -167,7 +180,7 @@ async def _baidu_translate(text, to_lang):
    if not chunks:
        raise ValueError("待翻译文本为空")
    qps = config.get("baidu_qps") or BAIDU_QPS_DEFAULT
-    timeout = ClientTimeout(total=config["ai_timeout_seconds"])
+    timeout = ClientTimeout(total=BAIDU_TIMEOUT_SECONDS)
    results = []
    async with ClientSession(timeout=timeout) as session:
        for chunk in chunks:
@@ -191,15 +204,21 @@ async def _baidu_translate(text, to_lang):


 async def _call_ai_assistant(action, text, instruction, config, connection_test=False):
-    """翻译/优化都走这里。优化类恒走 AI；翻译类在 translate_service=ai 时也走 AI。"""
-    if not config.get("ai_model"):
-        raise ValueError("尚未填写 AI 模型名称（“翻译并优化”“优化为 Anima”以及 AI 翻译都需要）")
-    ai_provider = config["ai_provider"]
+    """翻译/优化都走这里。优化类恒走 AI；翻译类在翻译服务就是某个 AI 后端时也走 AI。"""
+    # 两个后端各一套配置和规则。选 OpenAI / Ollama 当翻译服务时它就是后端；
+    # 选词库 / 百度时翻译不走 AI，优化仍走，这时用 ai_provider 决定读哪一套
+    prefix = AI_PROVIDER_PREFIXES.get(config.get("translate_service")) or AI_PROVIDER_PREFIXES.get(
+        config.get("ai_provider", "openai_compatible"), "openai"
+    )
+    if not config.get(f"{prefix}_model"):
+        raise ValueError("尚未填写 AI 模型名称（“翻译并优化”“优化”以及 AI 翻译都需要）")
    rule = {
-        "translate": config["translation_rule"],
-        "explain": config["translation_rule"],
-        "translate_optimize": config["translate_optimize_rule"],
-        "optimize": config["optimization_rule"],
+        "translate": config[f"{prefix}_translation_rule"],
+        "explain": config[f"{prefix}_translation_rule"],
+        "translate_optimize": config[f"{prefix}_translate_optimize_rule"],
+        "optimize": config[f"{prefix}_optimization_rule"],
+        # 自然语言模式下的「优化」：输入可以是中文，输出是英文自然语句，不走 Anima 标点规则
+        "optimize_natural": config[f"{prefix}_natural_rule"],
    }[action]
    if connection_test:
        rule = "你是连接测试助手。"
@@ -213,55 +232,61 @@ async def _call_ai_assistant(action, text, instruction, config, connection_test=
        rule = f"{rule}\n\n本次任务：无论输入语言为何，最终必须输出经过优化的英文 Anima 提示词。"
    elif not connection_test and action == "optimize":
        if re.search(r"[\u3400-\u9fff]", text):
-            raise ValueError("“优化为 Anima”只接受英文；中文或混合文本请使用“翻译并优化”")
+            raise ValueError("“优化”在标签模式下只接受英文；中文或混合文本请使用“翻译并优化”，或把模式切到“自然语言”")
        rule = f"{rule}\n\n本次任务：整理现有英文内容，不承担翻译。"
+    elif not connection_test and action == "optimize_natural":
+        # 自然语言模式允许中文输入：先译后整理，输出一段英文自然语句
+        direction = "先把中文忠实译为英文，再整理表达" if translation_direction(text) == "to_english" else "整理现有英文内容，不承担翻译"
+        rule = f"{rule}\n\n本次任务方向：{direction}。"
    if instruction:
        rule = f"{rule}\n\n用户本次附加要求优先遵循：{instruction.strip()}"
-    timeout = ClientTimeout(total=config["ai_timeout_seconds"])
-    if ai_provider == "ollama":
-        base = (config.get("ai_base_url") or "http://127.0.0.1:11434").rstrip("/")
+    timeout = ClientTimeout(total=config[f"{prefix}_timeout_seconds"])
+    if prefix == "ollama":
+        # Ollama 只允许本机回环地址：这里再校验一次，挡住手改配置文件的情况
+        base = normalize_ai_base_url(config["ollama_base_url"])
        if base.endswith("/v1"):
            base = base[:-3].rstrip("/")
        endpoint = base if base.endswith("/api/chat") else f"{base}/api/chat"
        payload = {
-            "model": config["ai_model"],
+            "model": config["ollama_model"],
            "stream": False,
            "messages": [{"role": "system", "content": rule}, {"role": "user", "content": text}],
-            "options": {"temperature": config["ai_temperature"]},
+            "options": {"temperature": config["ollama_temperature"]},
        }
        async with ClientSession(timeout=timeout) as session:
            async with session.post(endpoint, json=payload) as response:
-                data = await response.json(content_type=None)
+                body = await response.text()
                if response.status >= 400:
-                    raise ValueError(data.get("error") or f"Ollama 请求失败：HTTP {response.status}")
-        result = data.get("message", {}).get("content")
+                    raise ValueError(upstream_error_message(body, response.status, "Ollama"))
+        result = parse_json_body(body, "Ollama").get("message", {}).get("content")
    else:
-        if not config.get("ai_base_url"):
+        # OpenAI 兼容：远端必须 https，本机回环才允许 http
+        base = normalize_ai_base_url(config["openai_base_url"], allow_remote=True)
+        if not base:
            raise ValueError("尚未填写 OpenAI 兼容 API 地址")
        headers = {"Content-Type": "application/json"}
-        if config.get("ai_api_key"):
-            headers["Authorization"] = f"Bearer {config['ai_api_key']}"
+        if config.get("openai_api_key"):
+            headers["Authorization"] = f"Bearer {config['openai_api_key']}"
        payload = {
-            "model": config["ai_model"],
-            "temperature": config["ai_temperature"],
+            "model": config["openai_model"],
+            "temperature": config["openai_temperature"],
            "messages": [{"role": "system", "content": rule}, {"role": "user", "content": text}],
        }
        async with ClientSession(timeout=timeout) as session:
-            async with session.post(openai_chat_endpoint(config["ai_base_url"]), headers=headers, json=payload) as response:
-                data = await response.json(content_type=None)
+            async with session.post(openai_chat_endpoint(base), headers=headers, json=payload) as response:
+                body = await response.text()
                if response.status >= 400:
-                    detail = data.get("error") if isinstance(data, dict) else None
-                    if isinstance(detail, dict):
-                        detail = detail.get("message")
-                    raise ValueError(detail or f"API 请求失败：HTTP {response.status}")
-        choices = data.get("choices") if isinstance(data, dict) else None
+                    raise ValueError(upstream_error_message(body, response.status, "API"))
+        choices = parse_json_body(body, "API").get("choices")
        result = choices[0].get("message", {}).get("content") if choices else None
    result = _clean_assistant_result(result)
    if not result:
        raise ValueError("模型没有返回文本结果")
    if len(result) > 50000:
        raise ValueError("模型返回内容异常过长")
-    if action in {"translate_optimize", "optimize"}:
+    if action == "optimize_natural":
+        result = sanitize_natural_prompt(result)
+    elif action in {"translate_optimize", "optimize"}:
        result = sanitize_anima_prompt(result)
    return result

@@ -269,13 +294,13 @@ async def _call_ai_assistant(action, text, instruction, config, connection_test=
 async def _call_configured_assistant(action, text, instruction="", connection_test=False):
    config = assistant_store.config()
    translate_service = config["translate_service"]
-    # “翻译并优化”“优化为 Anima”恒走 AI；翻译/解释在 translate_service=ai 时也走 AI
-    needs_ai = action in {"translate_optimize", "optimize"} or (
-        action in {"translate", "explain"} and translate_service == "ai"
+    # “翻译并优化”“优化”恒走 AI；翻译/解释在翻译服务就是某个 AI 后端时也走 AI
+    needs_ai = action in {"translate_optimize", "optimize", "optimize_natural"} or (
+        action in {"translate", "explain"} and translate_service in AI_SERVICES
    )
    if needs_ai:
        return await _call_ai_assistant(action, text, instruction, config, connection_test)
-    # 以下只处理 translate/explain 且 translate_service != "ai"
+    # 以下只处理 translate/explain 且翻译服务不是 AI
    if translate_service == "dictionary":
        if connection_test:
            return "纯词库模式可用"
@@ -288,7 +313,7 @@ async def _call_configured_assistant(action, text, instruction="", connection_te
            return "百度翻译连接成功"
        to_lang = "zh" if action == "explain" or translation_direction(text) == "to_chinese" else "en"
        return await _baidu_translate(text, to_lang)
-    raise ValueError("翻译服务配置异常：请在助手设置里选择词库、百度或 AI")
+    raise ValueError("翻译服务配置异常：请在助手设置里选择词库、百度翻译、OpenAI 兼容接口或 Ollama")


@PromptServer.instance.routes.get("/bpi/session")
@@ -315,12 +340,54 @@ async def save_assistant_config(request):
        return error_response(error)


-@protected_route("post", "/bpi/assistant/reset-rules")
-async def reset_assistant_rules(_request):
+async def _list_backend_models(prefix, base_url, api_key):
+    """向 AI 后端问一次「你有哪些模型」；地址沿用现有白名单，问完不留任何东西。"""
+    timeout = ClientTimeout(total=_MODEL_LIST_TIMEOUT_SECONDS)
+    async with ClientSession(timeout=timeout) as session:
+        if prefix == "ollama":
+            # Ollama 只认本机回环地址，这里再校验一次，挡住手改配置文件的情况
+            base = normalize_ai_base_url(base_url)
+            if not base:
+                raise ValueError("尚未填写 Ollama 地址")
+            async with session.get(ollama_tags_endpoint(base)) as response:
+                # 先读原文再看状态码：错误应答常常不是 JSON，按 JSON 硬解析会把整个请求打崩
+                body = await response.text()
+                if response.status >= 400:
+                    raise ValueError(upstream_error_message(body, response.status, "Ollama"))
+            names = [item.get("name") for item in parse_json_body(body, "Ollama").get("models") or []]
+        else:
+            base = normalize_ai_base_url(base_url, allow_remote=True)
+            if not base:
+                raise ValueError("尚未填写 OpenAI 兼容 API 地址")
+            headers = {"Authorization": f"Bearer {api_key}"} if api_key else {}
+            async with session.get(openai_models_endpoint(base), headers=headers) as response:
+                body = await response.text()
+                if response.status >= 400:
+                    raise ValueError(upstream_error_message(body, response.status, "API"))
+            names = [item.get("id") for item in parse_json_body(body, "API").get("data") or []]
+    models = sorted({str(name).strip() for name in names if str(name).strip()})
+    if not models:
+        raise ValueError("服务端没有返回任何模型；也可以直接在输入框里手写模型名称")
+    return models
+
+
+@protected_route("post", "/bpi/assistant/models")
+async def list_assistant_models(request):
+    """用表单里当前填的地址问一次模型列表，只为给下拉候选用：不写配置，也不把密钥返回浏览器。"""
    try:
-        return web.json_response({"success": True, "data": assistant_store.reset_rules()})
-    except OSError as error:
-        return error_response(error, 500)
+        payload = await _read_json(request, _MAX_ASSISTANT_JSON_BYTES)
+        if not isinstance(payload, dict):
+            raise ValueError("请求内容格式不正确")
+        _consume_assistant_rate_limit()
+        prefix = "ollama" if payload.get("backend") == "ollama" else "openai"
+        # 密钥输入框留空 = 用户已经存过、这次不想再敲一遍，就拿已存的那份去问；仍只用于这一次请求
+        api_key = str(payload.get("api_key") or "").strip() or str(assistant_store.config().get(f"{prefix}_api_key") or "").strip()
+        models = await _list_backend_models(prefix, str(payload.get("base_url") or "").strip(), api_key)
+        return web.json_response({"success": True, "data": {"models": models}})
+    except AssistantLimitError as error:
+        return error_response(error, 429)
+    except Exception as error:
+        return error_response(error)


@protected_route("post", "/bpi/assistant/test")
@@ -349,7 +416,7 @@ async def run_assistant(request):
        action = str(payload.get("action", "translate")) if isinstance(payload, dict) else "translate"
        text = str(payload.get("text", "")) if isinstance(payload, dict) else ""
        instruction = str(payload.get("instruction", "")) if isinstance(payload, dict) else ""
-        if action not in {"translate", "explain", "translate_optimize", "optimize"}:
+        if action not in {"translate", "explain", "translate_optimize", "optimize", "optimize_natural"}:
            raise ValueError("未知的助手操作")
        if not text.strip():
            raise ValueError("输入内容为空")
@@ -388,26 +455,26 @@ async def get_dictionary(_request):
        return error_response(error, 500)


-# 节点标签卡的颜色配置：用户可以直接编辑 data/token_colors.json 增删色块，
-# 不用改源码。文件不在或格式不对时回退内置默认，不影响功能。
-_DEFAULT_TOKEN_COLORS = {
-    "presets": ["#6b9b78", "#9b8b6b", "#6b7b9b", "#9b6b8b", "#8b9b6b", "#6b9b9b", "#9b7b6b", "#7b6b9b"],
-    "random_pool": ["#5b8a72", "#8a725b", "#5b6e8a", "#8a5b7a", "#7a8a5b", "#5b8a8a", "#8a6b5b", "#6b5b8a"],
-}
-
-
+# 节点标签卡的颜色配置：色块池只读（用户可以直接编辑 data/token_colors.json），
+# 按词库配色由侧边栏「颜色」区写回同一个文件。
@protected_route("get", "/bpi/token-colors")
 async def get_token_colors(_request):
-    from pathlib import Path
-    path = Path(__file__).resolve().parent / "data" / "token_colors.json"
    try:
-        if path.exists():
-            data = json.loads(path.read_text(encoding="utf-8"))
-            if isinstance(data, dict) and isinstance(data.get("presets"), list) and isinstance(data.get("random_pool"), list):
-                return web.json_response({"success": True, "data": data})
-    except (OSError, ValueError):
-        pass
-    return web.json_response({"success": True, "data": _DEFAULT_TOKEN_COLORS})
+        return web.json_response({"success": True, "data": token_color_store.config()})
+    except OSError as error:
+        return error_response(error, 500)
+
+
+@protected_route("post", "/bpi/token-colors/colors")
+async def save_token_colors(request):
+    try:
+        payload = await _read_json(request, _MAX_JSON_BYTES)
+        if not isinstance(payload, dict):
+            raise ValueError("配色内容必须是对象")
+        data = token_color_store.save_colors(payload.get("pack_colors"), payload.get("anima_colors"))
+        return web.json_response({"success": True, "data": data})
+    except (OSError, ValueError) as error:
+        return error_response(error)


@protected_route("post", "/bpi/large/lookup")
@@ -852,3 +919,21 @@ def collect_model_names():
 async def list_model_names(_request):
    # 首次调用要遍历 models 目录，放到线程里免得卡住事件循环
    return web.json_response({"success": True, "data": await asyncio.to_thread(collect_model_names)})
+
+
+# ---------------------------------------------------------------------------
+# 重复检查：节点上输入一变就要立刻提示，和节点执行走同一个 find_duplicates
+# ---------------------------------------------------------------------------
+
+
+@protected_route("post", "/bpi/duplicate-check")
+async def check_duplicate_tags(request):
+    try:
+        payload = await _read_json(request)
+    except (ValueError, web.HTTPBadRequest) as error:
+        return error_response(error)
+    report = find_duplicates(payload.get("text") or "")
+    return web.json_response(
+        {"success": True, "data": {"count": report["extra"], "terms": report["terms"]}},
+        headers={"Cache-Control": "no-store"},
+    )
@@ -6,6 +6,7 @@ from pathlib import Path

 from assistant_store import (
    AssistantStore,
+    DEFAULT_NATURAL_RULE,
    DEFAULT_OPTIMIZATION_RULE,
    DEFAULT_TRANSLATE_OPTIMIZE_RULE,
    DEFAULT_TRANSLATION_RULE,
@@ -15,9 +16,15 @@ from assistant_store import (
    baidu_split_query,
    baidu_translate_params,
    dictionary_translate,
+    normalize_ai_base_url,
+    ollama_tags_endpoint,
    openai_chat_endpoint,
+    openai_models_endpoint,
+    parse_json_body,
    sanitize_anima_prompt,
+    sanitize_natural_prompt,
    translation_direction,
+    upstream_error_message,
 )
 from dictionary_store import DictionaryStore

@@ -37,41 +44,108 @@ class AssistantStoreTests(unittest.TestCase):
        with tempfile.TemporaryDirectory() as root:
            store = AssistantStore(Path(root) / "config")
            public = store.update({
-                "translate_service": "ai",
+                "translate_service": "openai_compatible",
                "ai_provider": "openai_compatible",
-                "ai_base_url": "https://example.test/v1",
-                "ai_model": "example-model",
-                "ai_api_key": "secret-value",
-                "translation_rule": "自定义仅翻译",
-                "translate_optimize_rule": "自定义翻译并优化",
-                "optimization_rule": "自定义优化",
+                "openai_base_url": "https://example.test/v1",
+                "openai_model": "example-model",
+                "openai_api_key": "secret-value",
+                "openai_translation_rule": "自定义仅翻译",
+                "openai_translate_optimize_rule": "自定义翻译并优化",
+                "openai_optimization_rule": "自定义优化",
            })
-            self.assertTrue(public["ai_api_key_configured"])
-            self.assertNotIn("ai_api_key", public)
-            self.assertEqual(store.config()["ai_api_key"], "secret-value")
-            self.assertEqual(store.config()["translation_rule"], "自定义仅翻译")
-            self.assertEqual(store.config()["translate_optimize_rule"], "自定义翻译并优化")
+            self.assertTrue(public["openai_api_key_configured"])
+            self.assertNotIn("openai_api_key", public)
+            self.assertEqual(store.config()["openai_api_key"], "secret-value")
+            self.assertEqual(store.config()["openai_translation_rule"], "自定义仅翻译")
+            self.assertEqual(store.config()["openai_translate_optimize_rule"], "自定义翻译并优化")
+            # 另一套规则不受影响
+            self.assertEqual(store.config()["ollama_translation_rule"], DEFAULT_TRANSLATION_RULE)
            self.assertNotIn("explanation_rule", public)

+    def test_openai_and_ollama_slots_are_independent(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.update({
+                "ai_provider": "openai_compatible",
+                "openai_base_url": "https://example.test/v1",
+                "openai_model": "cloud-model",
+                "openai_api_key": "secret-value",
+                "openai_translation_rule": "云端规则",
+            })
+            store.update({
+                "ai_provider": "ollama",
+                "ollama_model": "qwen3:8b",
+                "ollama_translation_rule": "本地规则",
+            })
+            config = store.config()
+            self.assertEqual(config["openai_model"], "cloud-model")
+            self.assertEqual(config["openai_api_key"], "secret-value")
+            self.assertEqual(config["openai_translation_rule"], "云端规则")
+            self.assertEqual(config["ollama_model"], "qwen3:8b")
+            self.assertEqual(config["ollama_base_url"], "http://127.0.0.1:11434")
+            self.assertEqual(config["ollama_translation_rule"], "本地规则")
+
    def test_dictionary_translation_preserves_weight(self):
        with tempfile.TemporaryDirectory() as root:
            dictionary = self.make_dictionary(root)
            self.assertEqual(dictionary_translate("长发，(微笑:1.2)", dictionary), "long hair, (smile:1.2)")

-    def test_api_key_is_bound_to_provider_and_base_url(self):
+    def test_api_key_is_bound_to_base_url(self):
        with tempfile.TemporaryDirectory() as root:
            store = AssistantStore(Path(root) / "config")
            store.update({
-                "translate_service": "ai",
+                "translate_service": "openai_compatible",
                "ai_provider": "openai_compatible",
-                "ai_base_url": "https://first.example/v1",
-                "ai_model": "example-model",
-                "ai_api_key": "secret-value",
+                "openai_base_url": "https://first.example/v1",
+                "openai_model": "example-model",
+                "openai_api_key": "secret-value",
            })
-            self.assertEqual(store.config()["ai_api_key"], "secret-value")
-            public = store.update({"ai_base_url": "https://second.example/v1"})
-            self.assertFalse(public["ai_api_key_configured"])
-            self.assertEqual(store.config()["ai_api_key"], "")
+            self.assertEqual(store.config()["openai_api_key"], "secret-value")
+            public = store.update({"openai_base_url": "https://second.example/v1"})
+            self.assertFalse(public["openai_api_key_configured"])
+            self.assertEqual(store.config()["openai_api_key"], "")
+
+    def test_api_key_survives_provider_switch(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.update({"openai_base_url": "https://example.test/v1", "openai_model": "m", "openai_api_key": "secret-value"})
+            # 切到 Ollama 再切回来，密钥还在
+            store.update({"ai_provider": "ollama"})
+            self.assertEqual(store.config()["openai_api_key"], "secret-value")
+            store.update({"ai_provider": "openai_compatible"})
+            self.assertTrue(store.public_config()["openai_api_key_configured"])
+
+    def test_legacy_ai_translate_service_is_split_by_provider(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.config_dir.mkdir(parents=True, exist_ok=True)
+            store.config_path.write_text(json.dumps({
+                "installation_id": store.installation_id,
+                "schema_version": 4,
+                "translate_service": "ai",
+                "ai_provider": "ollama",
+                "ollama_model": "qwen3:8b",
+            }, ensure_ascii=False), encoding="utf-8")
+            config = store.config()
+            self.assertEqual(config["translate_service"], "ollama")
+            self.assertEqual(config["ai_provider"], "ollama")
+
+    def test_ai_translate_service_pins_the_matching_provider(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.update({
+                "translate_service": "openai_compatible",
+                "openai_base_url": "https://example.test/v1",
+                "openai_model": "m",
+                "openai_api_key": "secret-value",
+            })
+            public = store.update({"translate_service": "ollama"})
+            self.assertEqual(public["translate_service"], "ollama")
+            self.assertEqual(public["ai_provider"], "ollama")
+            back = store.update({"translate_service": "openai_compatible"})
+            self.assertEqual(back["translate_service"], "openai_compatible")
+            self.assertEqual(back["ai_provider"], "openai_compatible")
+            self.assertTrue(back["openai_api_key_configured"])

    def test_new_installation_id_invalidates_saved_configuration(self):
        with tempfile.TemporaryDirectory() as root:
@@ -79,25 +153,115 @@ class AssistantStoreTests(unittest.TestCase):
            first_marker = Path(root) / "first-installation-id"
            first = AssistantStore(config_dir, install_id_path=first_marker)
            first.update({
-                "translate_service": "ai",
+                "translate_service": "openai_compatible",
                "ai_provider": "openai_compatible",
-                "ai_base_url": "https://example.test/v1",
-                "ai_model": "example-model",
-                "ai_api_key": "secret-value",
+                "openai_base_url": "https://example.test/v1",
+                "openai_model": "example-model",
+                "openai_api_key": "secret-value",
            })
            second = AssistantStore(config_dir, install_id_path=Path(root) / "second-installation-id")
            config = second.config()
            self.assertEqual(config["translate_service"], "dictionary")
-            self.assertEqual(config["ai_base_url"], "")
-            self.assertEqual(config["ai_api_key"], "")
+            self.assertEqual(config["openai_base_url"], "")
+            self.assertEqual(config["openai_api_key"], "")

-    def test_remote_http_endpoint_is_rejected_but_local_http_is_allowed(self):
+    def test_openai_base_url_requires_https_or_loopback(self):
        with tempfile.TemporaryDirectory() as root:
            store = AssistantStore(Path(root) / "config")
            with self.assertRaisesRegex(ValueError, "HTTPS"):
-                store.update({"ai_provider": "openai_compatible", "ai_base_url": "http://public.example/v1"})
-            result = store.update({"ai_provider": "ollama", "ai_base_url": "http://127.0.0.1:11434"})
-            self.assertEqual(result["ai_base_url"], "http://127.0.0.1:11434")
+                store.update({"openai_base_url": "http://public.example/v1"})
+            with self.assertRaisesRegex(ValueError, "完整的 http"):
+                store.update({"openai_base_url": "public.example/v1"})
+            self.assertEqual(store.update({"openai_base_url": "https://public.example/v1"})["openai_base_url"], "https://public.example/v1")
+            # 本机回环的 http 放行：本地 LM Studio 之类
+            self.assertEqual(store.update({"openai_base_url": "http://127.0.0.1:1234/v1"})["openai_base_url"], "http://127.0.0.1:1234/v1")
+
+    def test_ollama_base_url_must_be_loopback(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            self.assertEqual(store.config()["ollama_base_url"], "http://127.0.0.1:11434")
+            self.assertEqual(store.update({"ollama_base_url": "http://localhost:11434"})["ollama_base_url"], "http://localhost:11434")
+            self.assertEqual(store.update({"ollama_base_url": "http://[::1]:11434"})["ollama_base_url"], "http://[::1]:11434")
+            for blocked in ("http://192.168.1.20:11434", "http://10.0.0.5:11434", "https://ollama.example.test"):
+                with self.assertRaisesRegex(ValueError, "Ollama 地址只能是本机地址"):
+                    store.update({"ollama_base_url": blocked})
+
+    def test_remote_https_rejects_private_and_reserved_addresses(self):
+        # https 也不能拐进内网：云元数据、私有网段、链路本地一律拒绝
+        for blocked in (
+            "https://169.254.169.254/latest",
+            "https://192.168.1.20/v1",
+            "https://10.0.0.5/v1",
+            "https://172.16.0.9/v1",
+            "https://[fd00::1]/v1",
+        ):
+            with self.assertRaisesRegex(ValueError, "内网、保留或链路本地"):
+                normalize_ai_base_url(blocked, allow_remote=True)
+        # 公网域名与公网 IP 正常放行
+        self.assertEqual(normalize_ai_base_url("https://api.example.test/v1", allow_remote=True), "https://api.example.test/v1")
+        self.assertEqual(normalize_ai_base_url("https://8.8.8.8/v1", allow_remote=True), "https://8.8.8.8/v1")
+        # 本机回环不受这条限制
+        self.assertEqual(normalize_ai_base_url("http://127.0.0.1:1234/v1", allow_remote=True), "http://127.0.0.1:1234/v1")
+
+    def test_private_https_openai_url_is_rejected_on_save(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            with self.assertRaisesRegex(ValueError, "内网、保留或链路本地"):
+                store.update({"openai_base_url": "https://192.168.1.20/v1"})
+            self.assertEqual(store.update({"openai_base_url": "https://api.example.test/v1"})["openai_base_url"], "https://api.example.test/v1")
+
+    def test_normalize_ai_base_url_rejects_credentials_and_query(self):
+        with self.assertRaisesRegex(ValueError, "用户名或密码"):
+            normalize_ai_base_url("https://user:pass@example.test/v1", allow_remote=True)
+        with self.assertRaisesRegex(ValueError, "查询参数"):
+            normalize_ai_base_url("https://example.test/v1?x=1", allow_remote=True)
+        self.assertEqual(normalize_ai_base_url("https://example.test/v1/", allow_remote=True), "https://example.test/v1")
+        self.assertEqual(normalize_ai_base_url(""), "")
+
+    def test_v3_config_migrates_into_selected_provider_slot(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.config_dir.mkdir(parents=True, exist_ok=True)
+            store.config_path.write_text(json.dumps({
+                "installation_id": store.installation_id,
+                "schema_version": 3,
+                "translate_service": "ollama",
+                "ai_provider": "ollama",
+                "ai_base_url": "http://127.0.0.1:11434",
+                "ai_model": "qwen3:8b",
+                "ai_temperature": 0.7,
+                "ai_timeout_seconds": 120,
+                "translation_rule": "我的规则",
+                "translate_optimize_rule": "我的优化规则",
+                "optimization_rule": "我的整理规则",
+            }, ensure_ascii=False), encoding="utf-8")
+            config = store.config()
+            self.assertEqual(config["ollama_base_url"], "http://127.0.0.1:11434")
+            self.assertEqual(config["ollama_model"], "qwen3:8b")
+            self.assertEqual(config["ollama_temperature"], 0.7)
+            self.assertEqual(config["ollama_timeout_seconds"], 120)
+            self.assertEqual(config["openai_model"], "")  # 另一套留默认
+            # 规则原本不分后端，两套都继承
+            self.assertEqual(config["ollama_translation_rule"], "我的规则")
+            self.assertEqual(config["openai_translation_rule"], "我的规则")
+            self.assertEqual(config["schema_version"], 4)
+
+    def test_v3_migration_keeps_openai_api_key(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(Path(root) / "config")
+            store.config_dir.mkdir(parents=True, exist_ok=True)
+            store.config_path.write_text(json.dumps({
+                "installation_id": store.installation_id,
+                "schema_version": 3,
+                "ai_provider": "openai_compatible",
+                "ai_base_url": "https://example.test/v1",
+                "ai_model": "example-model",
+                "ai_api_key": "secret-value",
+            }, ensure_ascii=False), encoding="utf-8")
+            config = store.config()
+            # 绑定的 provider 标记变了，密钥不能因此被当成「换了地址」清掉
+            self.assertEqual(config["openai_api_key"], "secret-value")
+            self.assertTrue(store.public_config()["openai_api_key_configured"])

    def test_dictionary_translation_reports_unknown(self):
        with tempfile.TemporaryDirectory() as root:
@@ -124,13 +288,80 @@ class AssistantStoreTests(unittest.TestCase):
        self.assertEqual(translation_direction("杰作, best quality"), "to_english")
        self.assertEqual(
            sanitize_anima_prompt("```text\nmasterpiece，(smile：1.2)；@artist！\n```"),
-            "masterpiece, (smile:1.2), @artist.",
+            "masterpiece, (smile:1.2), @artist",
+        )
+        # 模型把结果包成 JSON/代码字符串时会残留转义符，清洗后不应判废
+        self.assertEqual(
+            sanitize_anima_prompt('masterpiece, \\"blue eyes\\", 1girl'),
+            "masterpiece, blue eyes, 1girl",
+        )
+        self.assertEqual(
+            sanitize_anima_prompt("masterpiece\\n\\t1girl, ＂blue dress＂"),
+            "masterpiece, 1girl, blue dress",
+        )
+        # 模型用来断句的句点统一换成逗号，权重里的小数点必须原样保留
+        self.assertEqual(
+            sanitize_anima_prompt("masterpiece, 1girl, dragon slayer ornstein."),
+            "masterpiece, 1girl, dragon slayer ornstein",
+        )
+        self.assertEqual(
+            sanitize_anima_prompt("masterpiece, 1girl. A boy stands in the rain."),
+            "masterpiece, 1girl, A boy stands in the rain",
+        )
+        self.assertEqual(
+            sanitize_anima_prompt("(smile:1.2), (@rei (sanbonzakura):1.3), 1girl. A knight stands."),
+            "(smile:1.2), (@rei (sanbonzakura):1.3), 1girl, A knight stands",
+        )
+        self.assertEqual(
+            sanitize_anima_prompt("<lora:detail:0.8>, 1girl. A knight stands amid ruins, red banners."),
+            "<lora:detail:0.8>, 1girl, A knight stands amid ruins, red banners",
        )
        with self.assertRaisesRegex(ValueError, "不符合 Anima"):
            sanitize_anima_prompt("masterpiece [smile]")
        with self.assertRaisesRegex(ValueError, "仍包含中文"):
            sanitize_anima_prompt("masterpiece, 杰作")

+    def test_natural_prompt_keeps_sentence_punctuation(self):
+        # 与 Anima 清洗的核心区别：句点是句子的一部分，不能被换成逗号
+        self.assertEqual(
+            sanitize_natural_prompt("A woman stands in the rain. She looks tired."),
+            "A woman stands in the rain. She looks tired.",
+        )
+        # 逗号是句子内部的停顿，原样保留
+        self.assertEqual(
+            sanitize_natural_prompt("A woman stands in the rain, tired and cold."),
+            "A woman stands in the rain, tired and cold.",
+        )
+        # 全角标点统一成半角，代码块围栏和转义符照旧清掉
+        self.assertEqual(
+            sanitize_natural_prompt('```text\nA woman stands in the rain，tired。\n```'),
+            "A woman stands in the rain, tired.",
+        )
+        self.assertEqual(
+            sanitize_natural_prompt('A woman \\"stands\\" in the rain'),
+            'A woman "stands" in the rain',
+        )
+        # 模型爱用列表；自然语言结果里列表符号要去掉，换行收成空格
+        self.assertEqual(
+            sanitize_natural_prompt("1. A woman stands in the rain\n2. She looks tired"),
+            "A woman stands in the rain She looks tired",
+        )
+        with self.assertRaisesRegex(ValueError, "仍包含中文"):
+            sanitize_natural_prompt("一名女子站在雨中")
+
+    def test_natural_rule_is_persisted_per_backend(self):
+        with tempfile.TemporaryDirectory() as root:
+            store = AssistantStore(config_dir=root, install_id_path=Path(root) / "install")
+            store.update({"openai_natural_rule": "我的自然语言规则", "ollama_natural_rule": "本地规则"})
+            config = store.config()
+            self.assertEqual(config["openai_natural_rule"], "我的自然语言规则")
+            self.assertEqual(config["ollama_natural_rule"], "本地规则")
+            # 另一套不受影响，且默认值能通过 public_config 取回
+            self.assertEqual(config["openai_optimization_rule"], DEFAULT_OPTIMIZATION_RULE)
+            public = store.public_config()
+            self.assertEqual(public["default_openai_natural_rule"], DEFAULT_NATURAL_RULE)
+            self.assertEqual(public["default_ollama_natural_rule"], DEFAULT_NATURAL_RULE)
+
    def test_legacy_default_rules_are_upgraded(self):
        with tempfile.TemporaryDirectory() as root:
            store = AssistantStore(Path(root) / "config")
@@ -141,9 +372,10 @@ class AssistantStoreTests(unittest.TestCase):
                "optimization_rule": LEGACY_DEFAULT_OPTIMIZATION_RULE,
            }, ensure_ascii=False), encoding="utf-8")
            config = store.config()
-            self.assertEqual(config["translation_rule"], DEFAULT_TRANSLATION_RULE)
-            self.assertEqual(config["translate_optimize_rule"], DEFAULT_TRANSLATE_OPTIMIZE_RULE)
-            self.assertEqual(config["optimization_rule"], DEFAULT_OPTIMIZATION_RULE)
+            for prefix in ("openai", "ollama"):
+                self.assertEqual(config[f"{prefix}_translation_rule"], DEFAULT_TRANSLATION_RULE)
+                self.assertEqual(config[f"{prefix}_translate_optimize_rule"], DEFAULT_TRANSLATE_OPTIMIZE_RULE)
+                self.assertEqual(config[f"{prefix}_optimization_rule"], DEFAULT_OPTIMIZATION_RULE)

    def test_v17_default_translation_rule_is_upgraded_without_overwriting_custom_rules(self):
        with tempfile.TemporaryDirectory() as root:
@@ -154,10 +386,11 @@ class AssistantStoreTests(unittest.TestCase):
                "provider": "dictionary",
                "translation_rule": V17_DEFAULT_TRANSLATION_RULE,
            }, ensure_ascii=False), encoding="utf-8")
-            self.assertEqual(store.config()["translation_rule"], DEFAULT_TRANSLATION_RULE)
+            self.assertEqual(store.config()["openai_translation_rule"], DEFAULT_TRANSLATION_RULE)

-            store.update({"translation_rule": "我的自定义翻译规则"})
-            self.assertEqual(store.config()["translation_rule"], "我的自定义翻译规则")
+            store.update({"openai_translation_rule": "我的自定义翻译规则"})
+            self.assertEqual(store.config()["openai_translation_rule"], "我的自定义翻译规则")
+            self.assertEqual(store.config()["ollama_translation_rule"], DEFAULT_TRANSLATION_RULE)

    def test_baidu_provider_saves_and_masks_secret(self):
        with tempfile.TemporaryDirectory() as root:
@@ -172,13 +405,13 @@ class AssistantStoreTests(unittest.TestCase):
            self.assertTrue(public["baidu_secret_key_configured"])
            self.assertNotIn("baidu_secret_key", public)
            self.assertEqual(store.config()["baidu_secret_key"], "baidu-secret")
-            self.assertFalse(public["ai_api_key_configured"])
+            self.assertFalse(public["openai_api_key_configured"])

    def test_baidu_secret_is_retained_when_translate_service_changes(self):
        with tempfile.TemporaryDirectory() as root:
            store = AssistantStore(Path(root) / "config")
            store.update({"translate_service": "baidu", "baidu_appid": "appid", "baidu_secret_key": "baidu-secret"})
-            public = store.update({"translate_service": "ai", "ai_provider": "openai_compatible", "ai_base_url": "https://example.test/v1", "ai_model": "m"})
+            public = store.update({"translate_service": "openai_compatible", "ai_provider": "openai_compatible", "openai_base_url": "https://example.test/v1", "openai_model": "m"})
            self.assertTrue(public["baidu_secret_key_configured"])
            self.assertEqual(store.config()["baidu_secret_key"], "baidu-secret")

@@ -215,6 +448,38 @@ class AssistantStoreTests(unittest.TestCase):
            path.write_text(json.dumps(raw, ensure_ascii=False), encoding="utf-8")
            self.assertEqual(store.config()["baidu_qps"], 1.0)

+    def test_model_list_endpoints_reuse_the_saved_base_url_shape(self):
+        self.assertEqual(openai_models_endpoint("https://api.openai.com/v1"), "https://api.openai.com/v1/models")
+        self.assertEqual(openai_models_endpoint("http://127.0.0.1:1234/v1/"), "http://127.0.0.1:1234/v1/models")
+        self.assertEqual(openai_models_endpoint("https://x.cn/v1/chat/completions"), "https://x.cn/v1/models")
+        self.assertEqual(openai_models_endpoint("https://x.cn/models"), "https://x.cn/models")
+        self.assertEqual(ollama_tags_endpoint("http://127.0.0.1:11434"), "http://127.0.0.1:11434/api/tags")
+        self.assertEqual(ollama_tags_endpoint("http://127.0.0.1:11434/v1"), "http://127.0.0.1:11434/api/tags")
+        self.assertEqual(ollama_tags_endpoint("http://127.0.0.1:11434/api/chat"), "http://127.0.0.1:11434/api/tags")
+        self.assertEqual(openai_models_endpoint(""), "")
+        self.assertEqual(ollama_tags_endpoint(""), "")
+
+    def test_error_bodies_that_are_not_json_still_read_as_a_sentence(self):
+        # DeepSeek 缺密钥时回的是纯文本一行，不是 JSON；解析不了也不能把整条请求打崩
+        # 第三个参数和 server.py 里的调用保持一致：整条路径改签名时这里会先炸出来
+        self.assertIn("Authentication Fails", upstream_error_message("Authentication Fails (governor)", 401, "API"))
+        self.assertIn("请检查 API 密钥", upstream_error_message("Authentication Fails (governor)", 401, "API"))
+        self.assertEqual(
+            upstream_error_message('{"error":{"message":"Invalid API key"}}', 401, "API"),
+            "Invalid API key（HTTP 401；请检查 API 密钥是否已填写、是否有效）",
+        )
+        self.assertEqual(upstream_error_message("404 page not found", 404, "API"), "404 page not found（HTTP 404）")
+        self.assertEqual(upstream_error_message("", 502, "Ollama"), "Ollama 请求失败（HTTP 502）")
+        self.assertTrue(len(upstream_error_message("<html>" + "x" * 500 + "</html>", 403, "API")) < 220)
+
+    def test_non_json_success_body_reports_instead_of_crashing(self):
+        self.assertEqual(parse_json_body('{"data": []}', "API"), {"data": []})
+        with self.assertRaises(ValueError) as raised:
+            parse_json_body("<html>login</html>", "API")
+        self.assertIn("返回的不是 JSON", str(raised.exception))
+        with self.assertRaises(ValueError):
+            parse_json_body("[1, 2, 3]", "API")
+
    def test_baidu_params_and_query_splitting(self):
        params = baidu_translate_params("appid", "secret", "你好", "zh", salt="12345")
        self.assertEqual(params["sign"], hashlib.md5("appid你好12345secret".encode("utf-8")).hexdigest())
@@ -41,6 +41,17 @@ class DictionaryStoreTests(unittest.TestCase):
        self.assertTrue(store.user_path.exists())
        self.assertEqual(store.user_tags()[0]["english"], "keep me")

+    def test_natural_language_pack_name_in_snapshot(self):
+        self.assertFalse(validate_tag({"english": "smile", "chinese": "微笑"})["natural"])
+        self.assertTrue(validate_tag({"english": "smile", "chinese": "微笑", "natural": True})["natural"])
+        temp, store = self.make_store()
+        self.addCleanup(temp.cleanup)
+        store.upsert({"english": "smile", "chinese": "微笑"})
+        store.upsert({"english": "a knight stands in the rain", "chinese": "一名骑士站在雨中", "natural": True})
+        names = {tag["english"]: tag["pack_name"] for tag in store.snapshot()["tags"]}
+        self.assertEqual(names["smile"], "个人词库")
+        self.assertEqual(names["a knight stands in the rain"], "自然语言词库")
+
    def test_normalize_matches_spaces_and_underscores(self):
        self.assertEqual(normalize_key(" Looking_At_Viewer "), "looking at viewer")

@@ -124,6 +135,17 @@ class DictionaryStoreTests(unittest.TestCase):
        self.assertEqual(effective["smile"]["chinese"], "个人微笑")
        self.assertEqual(effective["smile"]["pack_id"], "personal")

+    def test_natural_language_dictionary_is_a_separate_pack(self):
+        temp, store = self.make_pack_store()
+        self.addCleanup(temp.cleanup)
+        store.upsert({"english": "smile", "chinese": "个人微笑"})
+        store.upsert({"english": "a woman is standing in the rain", "chinese": "一个女人站在雨中", "natural": True})
+        effective = {tag["english"]: tag for tag in store.snapshot()["tags"]}
+        self.assertEqual(effective["smile"]["pack_id"], "personal")
+        self.assertEqual(effective["smile"]["pack_name"], "个人词库")
+        self.assertEqual(effective["a woman is standing in the rain"]["pack_id"], "natural")
+        self.assertEqual(effective["a woman is standing in the rain"]["pack_name"], "自然语言词库")
+
    def test_community_pack_lifecycle(self):
        temp, store = self.make_pack_store()
        self.addCleanup(temp.cleanup)
@@ -0,0 +1,78 @@
+import assert from "node:assert/strict";
+
+import {
+  analyzePromptSyntax,
+  buildDictionaryIndex,
+  duplicateGroups,
+  normalizeKey,
+  parsePrompt,
+  removePromptToken,
+} from "../js/parser.js";
+import { setLanguageMode, t } from "../js/i18n.js";
+
+const dictionary = buildDictionaryIndex([
+  { english: "1girl", chinese: "一个女孩", verified: true },
+  { english: "smile", chinese: "微笑", verified: true },
+]);
+
+// 同一个标签出现 3 次：权重写法不同的那次也算同一 key
+const groups = duplicateGroups(parsePrompt("1girl, smile, (1girl:1.3), 1girl", dictionary));
+assert.equal(groups.length, 1);
+assert.equal(groups[0].key, normalizeKey("1girl"));
+assert.equal(groups[0].tokens.length, 3);
+// 顺序按原文，且三次都是普通标签
+assert.equal(groups[0].tokens[0].raw.trim(), "1girl");
+assert.equal(groups[0].tokens[1].raw.trim(), "(1girl:1.3)");
+assert.ok(groups[0].tokens.every((token) => token.syntax === "tag"));
+
+// key 归一：大小写差异算同一个标签（下划线是按空格算的，"sm_ile" 不等于 "smile"）
+assert.equal(duplicateGroups(parsePrompt("smile, Smile", dictionary)).length, 1);
+
+// BREAK 与 <lora:…> 不算重复标签
+assert.equal(duplicateGroups(parsePrompt("BREAK, BREAK", dictionary)).length, 0);
+assert.equal(duplicateGroups(parsePrompt("<lora:a:1>, <lora:a:1>", dictionary)).length, 0);
+
+// 问题行的「重复标签」提示与分组口径一致（同一个标签只提示一次）
+const text = "1girl, 1girl, smile";
+const issues = analyzePromptSyntax(text, parsePrompt(text, dictionary));
+assert.equal(issues.filter((item) => item.code === "duplicate").length, 1);
+assert.deepEqual(issues.find((item) => item.code === "duplicate").tokenKeys, [normalizeKey("1girl")]);
+
+// 「唯一」的输出剔除：按出现次序删，第一次出现的那次要留着
+let output = "1girl, smile, (1girl:1.3), 1girl";
+for (;;) {
+  const parsed = parsePrompt(output, dictionary);
+  const target = duplicateGroups(parsed)
+    .filter((group) => group.key === normalizeKey("1girl"))
+    .flatMap((group) => group.tokens.slice(1))[0];
+  if (!target) break;
+  const step = removePromptToken(output, target);
+  assert.ok(step.changed);
+  output = step.text;
+}
+assert.equal(output.trim(), "1girl, smile");
+
+// 界面文案：英文原文 + 中文译表双写，变量折成 {} 后要能命中
+setLanguageMode("en");
+assert.equal(t("Unique"), "Unique");
+assert.equal(t("Duplicates {}").replace("{}", "3"), "Duplicates 3");
+setLanguageMode("zh");
+assert.equal(t("Unique"), "唯一");
+assert.equal(t("Duplicates 3"), "重复 3");
+assert.equal(t("Keep only the first “1girl” in the actual output; later occurrences stay here with a strikethrough"),
+  "只让第一次出现的「1girl」进入实际输出；其余照常显示在界面上（划线）");
+assert.equal(t("Tags repeated in this prompt; click a chip to step through its occurrences, double-click to delete the one you stepped to; click empty space here to clear the selection"),
+  "提示词里重复的标签：单击胶囊依次跳到每一次出现，双击删掉跳到的那一处；点这里空白处取消选中");
+assert.equal(t("“1girl” appears 3 times; click to jump to the first one"), "「1girl」出现了 3 次；单击跳到第一次");
+// 开了「唯一」的胶囊是另一个提示语，里面嵌套的「Unique」要跟着翻成「唯一」
+assert.equal(t("“1girl” appears 3 times; “Unique” is on, only the first reaches the actual output"),
+  "「1girl」出现了 3 次；已开「唯一」：只有第一次进入实际输出");
+assert.equal(t("“1girl” occurrence 2 of 3; click again for the next one, double-click to delete it"),
+  "「1girl」第 2 / 3 处；再点一次跳下一处，双击删除这一处");
+assert.equal(t("“1girl” now unique: only the first occurrence enters the actual output; press Ctrl+Z to undo"),
+  "「1girl」已设为唯一：只有第一次出现进入实际输出；按 Ctrl+Z 可撤销");
+// 「Unique」在译文里会被二次翻译，tooltip 读起来是一致的中文
+assert.equal(t("Query: “1girl” | “Unique” is on: only the first occurrence reaches the actual output"),
+  "查询「1girl」｜已开「唯一」：只有第一次出现进入实际输出");
+
+console.log("test_duplicates: ok");
@@ -34,6 +34,27 @@ assert.equal(t("restored 3 hidden tags; press Ctrl+Z to undo"), "已恢复 3 个
 assert.equal(t("items 3 | recognized 1 | unknown 2"), "共 3 个标签｜已识别 1｜未收录 2");
 assert.equal(t("items 3 | recognized 1 | unknown 2 | pending 1"), "共 3 个标签｜已识别 1｜未收录 2｜待确认 1");
 assert.equal(t("Personal"), "个人");
+// 过滤栏颜色下拉的第四项：按标签所在词库上色
+assert.equal(t("By dictionary"), "按词库");
+// 侧边栏「颜色」区：分区名、按钮、词条数、未收录项
+assert.equal(t("Colors"), "颜色");
+assert.equal(t("Restore defaults"), "恢复默认");
+assert.equal(t("Random color"), "随机颜色");
+assert.equal(t("Not indexed"), "未收录");
+assert.equal(t("3 items"), "3 条");
+assert.equal(t("952 items"), "952 条");
+// 按 Anima 配色的分类短名（跟 anima_sorter.js 的槽位一一对应）
+assert.equal(t("By Anima"), "按 Anima");
+assert.equal(t("Quality"), "质量");
+assert.equal(t("Expression"), "表情");
+assert.equal(t("Camera"), "镜头");
+assert.equal(t("Scene"), "场景");
+assert.equal(t("Copyright"), "作品");
+// 双击改色面板底部那句优先级说明
+assert.equal(
+  t("This color has the lowest priority: it only shows in Default mode; other color modes override it."),
+  "这里改的颜色优先级最低：只有「默认」上色模式才显示，按收录、随机、按词库、按 Anima 都会盖住它。",
+);
 assert.equal(t("Highest"), "最高");
 assert.equal(t("12 items | Local user | Always enabled, overrides same-name entries"), "12 条｜本地用户｜始终启用，覆盖同名条目");

@@ -64,7 +85,15 @@ assert.equal(
 // 自由文本变量必须用「"…"」包起来才能折叠成 {}，否则整句永远匹配不上。
 // 这里锁死「运行时真实形态」，防止以后有人把引号去掉或换成 ASCII 直引号。
 for (const [source, expected] of [
-  ["“Optimize to Anima” done; please confirm in the preview window", "「优化为 Anima」完成；请在预览窗口确认"],
+  ["“Optimize” done; please confirm in the preview window", "「优化」完成；请在预览窗口确认"],
+  ["Optimize", "优化"],
+  ["Fit height", "适应高度"],
+  ["Natural Language Rule", "自然语言规则"],
+  ["(Chinese detected; translating to English first)", "（检测到中文，会先译成英文）"],
+  [
+    "Optimize with the rule for the current mode: tag mode tidies the tag flow, natural language mode rewrites it as English sentences",
+    "按当前模式对应的规则优化：标签模式整理标签流，自然语言模式改写为英文自然语句",
+  ],
  ["applied “Translate & optimize”; press Ctrl+Z to undo", "已应用 翻译并优化；按 Ctrl+Z 可撤销"],
  ["Undone: “Sort to Anima order”", "已撤销：Sort to Anima order"],
  ["Redone: “Sort to Anima order”", "已重做：Sort to Anima order"],
@@ -77,6 +106,12 @@ for (const [source, expected] of [
  ["Translation complete 5 items, rejected abnormal results 2 items", "翻译完成 5 条，已拒绝异常结果 2 条"],
  ["uses “12,345”", "使用 12,345 次"],
  ["tag: “1girl” | Anima format: (Tag:Weight)", "标签：1girl｜Anima 格式：(标签:权重)"],
+  ["Suspected natural language", "疑似自然语言"],
+  ["Not indexed · suspected natural language", "未收录 · 疑似自然语言"],
+  [
+    "“overgrown with a variety of red wildflowers” ↔ “未收录” | suspected natural language: more than 3 spaces; confirm or split it into tags",
+    "overgrown with a variety of red wildflowers ↔ 未收录｜疑似自然语言：空格超过 3 个；请确认或拆成标签",
+  ],
  ["Node #3 · 12  tags", "节点 #3 · 12 个标签"],
  ["Translating: “1girl”", "正在翻译：1girl"],
  ["Re-translated “1girl”", "已重新翻译「1girl」"],
@@ -97,11 +132,19 @@ for (const [source, expected] of [
  assert.equal(t(source), expected, `未命中：${source}`);
 }

-// 置信度徽章：这几个是 parser.js 直接产出的界面文本
-assert.equal(t("Low · pending"), "低 · 待确认");
-assert.equal(t("Medium · pending"), "中 · 待确认");
-assert.equal(t("Medium · session-only"), "中 · 仅本次会话");
+// 置信度：徽章只显示档位，细分原因放在 tooltip，两边都是 parser.js 直接产出的
+assert.equal(t("Not judged"), "未判定");
 assert.equal(t("Low"), "低");
+assert.equal(t("Medium"), "中");
+assert.equal(t("High"), "高");
+assert.equal(t("Not in any dictionary, so it cannot be judged"), "词库里查不到，没法判断");
+assert.equal(t("Machine translation, not saved to the dictionary or confirmed yet"), "机器翻译，还没存进词库，也没人工确认");
+assert.equal(t("Syntax kept as-is; nothing to translate"), "语法保留项，本来就不用翻译");
+assert.equal(t("In the dictionary, but not marked as verified"), "词库里有，但还没标成已校验");
+assert.equal(t("In the dictionary, but no category set"), "词库里有，但还没填分类");
+assert.equal(t("Changed for this session only; not saved to the dictionary"), "只改了本次会话，没存进词库");
+// 带分类名的那条：变量被 normalizeKey 折成 {}，键里也是 {}
+assert.equal(t("In the dictionary, verified, category “发型”"), "已在词库里，已校验，分类「发型」");

 // 收藏星标按钮（带符号的整串）
 assert.equal(t("★ Unfavorite"), "★ 取消收藏");
@@ -117,15 +160,19 @@ assert.equal(t("(weight 1.5)"), "（权重 1.5）");
 // 插件名：侧边栏标签写死中文，面板/关于框走译表，三处必须同一个名字
 assert.equal(t("Bilingual Prompt Manager"), "提示词翻译与管理");
 assert.equal(t("Bilingual Prompt Inspector"), "提示词翻译与管理");
+// 设置面板里「语言与外观」那一项
+assert.equal(t("Appearance & Language"), "语言与外观");
+assert.equal(t("Appearance and language saved"), "外观与语言已保存");
+// 两个本地预设按钮各有一套文案，别混
+assert.equal(t("Use LM Studio Local Preset"), "使用 LM Studio 本地预设");
+assert.equal(t("Use Ollama Local Preset"), "使用 Ollama 本地预设");
+assert.equal(t("Takes effect after saving"), "保存后生效");
+assert.equal(t("Ollama default URL filled; select a pulled model ID, then Save & Test"), "已填入 Ollama 默认地址；选择已拉取的模型 ID 后保存并测试");
 // 确认对话框：默认确认按钮是 "OK"，必须译为「确认」而不是「正常」
 assert.equal(t("OK"), "确认");
 assert.equal(t("Cancel"), "取消");
 assert.equal(t("This will overwrite node #3 prompt. Cannot be undone with Ctrl+Z."), "这会覆盖节点 #3 的提示词，Ctrl+Z 无法撤销。");
 assert.equal(t("Delete “Fluffy”? Reference image will be deleted too."), "删除「Fluffy」？参考图也会一并删除。");
-assert.equal(
-  t("Bilingual Prompt Inspector: Tag Manager, Dictionary, Packs & Assistant Settings"),
-  "提示词翻译与管理：标签管理、词库、词包与助手设置",
-);
 assert.equal(t("restored 3 hidden tags; press Ctrl+Z to undo"), "已恢复 3 个隐藏标签；按 Ctrl+Z 可撤销");
 assert.equal(t("Node #3"), "节点 #3");

@@ -6,8 +6,10 @@ import {
  buildDictionaryIndex,
  detectInputMode,
  extractInstructionBody,
+  looksSuspiciousNatural,
  parsePrompt,
  removePromptToken,
+  replacePromptSpan,
  replacePromptTokenWeight,
  splitMixedText,
  splitNaturalText,
@@ -19,7 +21,7 @@ import {

 const source = "masterpiece, (looking at viewer:1.2), <lora:test:0.8>\nBREAK, hand_on_hip";
 const dictionary = buildDictionaryIndex([
-  { english: "masterpiece", chinese: "杰作", verified: true },
+  { english: "masterpiece", chinese: "杰作", verified: true, category: "画质" },
  { english: "looking at viewer", chinese: "看向镜头", verified: true },
  { english: "hand on hip", chinese: "叉腰", verified: true },
 ]);
@@ -90,6 +92,26 @@ assert.equal(tagsAroundTokens[7].segmentKind, "natural");
 assert.equal(tagsAroundTokens[8].term, "best quality");
 assert.equal(tagsAroundTokens.at(-1).term, "newest");

+// 开头的自然语言句很长、后面的标签又偏冷门（没有 masterpiece/best quality 那类信号词）时，
+// 仍要把开头那句切出来，而不是把整段当成一个自然语言段
+const coldTailSource = "A knight in dark gold armor holds a polearm spear amid overgrown ruins with red banners, cracked white stone, and red wildflowers under dramatic cinematic lighting\nultra-HD, masterwork, detailed, fs_ornstein, dark gold armor, breastplate, full armor, holding polearm, overgrown ruins, white and red theme";
+const coldTailMode = detectInputMode(coldTailSource);
+assert.equal(coldTailMode.mode, "mixed");
+assert.equal(coldTailMode.naturalStart, 0);
+const coldTailTokens = parsePrompt(coldTailSource, dictionary);
+assert.equal(coldTailTokens.filter((token) => token.segmentKind === "natural").length, 1);
+assert.equal(coldTailTokens.at(-1).term, "white and red theme");
+
+// 空格超过 3 个（不含 3 个）就标成「疑似自然语言」
+assert.equal(looksSuspiciousNatural("white and red theme"), false);
+assert.equal(looksSuspiciousNatural("dragon slayer ornstein"), false);
+assert.equal(looksSuspiciousNatural("overgrown with a variety of red wildflowers"), true);
+assert.equal(looksSuspiciousNatural("and red wildflowers under dramatic cinematic lighting"), true);
+const suspectedTokens = parsePrompt("1girl, white and red theme, overgrown with a variety of red wildflowers", dictionary);
+assert.equal(suspectedTokens[0].suspectedNatural, false);
+assert.equal(suspectedTokens[1].suspectedNatural, false);
+assert.equal(suspectedTokens[2].suspectedNatural, true);
+
 const commaRichNatural = "A woman walks home after the rain, feeling tired and cold, carrying a small umbrella, while the city grows quiet.";
 assert.equal(detectInputMode(commaRichNatural).mode, "natural");
 const shortClauseNatural = "She came home, tired, cold, wet, alone, at night.";
@@ -113,6 +135,12 @@ assert.equal(
  "tags",
 );
 assert.equal(detectInputMode("A woman sits on a bed, asking for a hug.").mode, "natural");
+// 权重里的句点不能把整串标签判定成自然语言
+const weightedTagStream = "8K, ultra detailed, masterpiece, best quality, newest, absurdres, highres, very aesthetic, sharp focus, ultra-HD, amazing quality, masterwork, detailed, high detail, great atmosphere, amazing ambience, perfect composition, intricate, dynamic, rich color, 1boy, male focus, fs_ornstein, dragon slayer ornstein";
+assert.equal(detectInputMode(`${weightedTagStream}, (@rei (sanbonzakura):1. )`).mode, "tags");
+assert.equal(detectInputMode(`${weightedTagStream}, (@rei (sanbonzakura):1.\t`).mode, "tags");
+assert.equal(detectInputMode(`${weightedTagStream}, (@rei (sanbonzakura):1.2). `).mode, "tags");
+assert.equal(parsePrompt(`${weightedTagStream}, (@rei (sanbonzakura):1.2)`, dictionary).length, 25);
 assert.equal(detectInputMode("whatever", "natural").reason, "Manual override");
 const instructed = "翻译要求：转换为 Anima 标签\n正文：一位成年女性站在雨中";
 assert.equal(detectInputMode(instructed).mode, "instruction");
@@ -155,6 +183,23 @@ assert.equal(replacePromptTokenWeight(weightSource, weightTokens[1], 4).changed,
 const naturalWeightToken = parsePrompt("A woman is standing in the rain.", dictionary)[0];
 assert.equal(replacePromptTokenWeight("A woman is standing in the rain.", naturalWeightToken, 1.2).changed, false);

+// 「剪切出来改完再粘回去」：整段替换，段首段尾的空白是原文排版，要跟着补回去
+const spanSource = "masterpiece, best quality, solo";
+const spanTokens = parsePrompt(spanSource, dictionary);
+assert.deepEqual(replacePromptSpan(spanSource, spanTokens[1].start, spanTokens[1].end, "high quality, ultra detailed"), {
+  text: "masterpiece, high quality, ultra detailed, solo",
+  cursor: 41,
+  changed: true,
+});
+assert.equal(replacePromptSpan(spanSource, spanTokens[1].start, spanTokens[2].end, "1girl").text, "masterpiece, 1girl");
+// 清空 = 删除这一段，接缝处的逗号只留一个；开头那段没有左逗号就吃右边那个
+assert.equal(replacePromptSpan(spanSource, spanTokens[1].start, spanTokens[2].end, "").text, "masterpiece");
+assert.equal(replacePromptSpan(spanSource, spanTokens[0].start, spanTokens[0].end, "").text, "best quality, solo");
+// 原样写回不算改动
+assert.equal(replacePromptSpan(spanSource, spanTokens[1].start, spanTokens[1].end, "best quality").changed, false);
+// 换行不是逗号，删掉标签后要留下，别把多行排版压平
+assert.equal(replacePromptSpan("a\n1girl\n, b", 2, 7, "").text, "a\nb");
+
 const diagnosticSource = "1girl, solo, 2girls, standing, sitting, red hair, red_hair, (smile:abc), [open";
 const diagnosticTokens = parsePrompt(diagnosticSource, dictionary);
 const diagnostics = analyzePromptSyntax(diagnosticSource, diagnosticTokens);
@@ -165,6 +210,13 @@ assert.ok(diagnostics.some((item) => item.code === "conflict"));

 const knownToken = parsePrompt("masterpiece", dictionary)[0];
 assert.equal(knownToken.confidence, "high");
+// 可信度看词条整理得怎么样：填了分类才高，没填分类（含默认的「未分类」）只是中
+const plain = buildDictionaryIndex([{ english: "masterpiece", chinese: "杰作", verified: true }]);
+assert.equal(parsePrompt("masterpiece", plain)[0].confidence, "medium");
+const uncategorized = buildDictionaryIndex([{ english: "masterpiece", chinese: "杰作", verified: true, category: "Uncategorized" }]);
+assert.equal(parsePrompt("masterpiece", uncategorized)[0].confidence, "medium");
+const unverified = buildDictionaryIndex([{ english: "masterpiece", chinese: "杰作", verified: false, category: "画质" }]);
+assert.equal(parsePrompt("masterpiece", unverified)[0].confidence, "medium");
 const machineToken = parsePrompt("cinematic", dictionary, new Map([["cinematic", { text: "电影感", source: "prompt-assistant" }]]))[0];
 assert.equal(machineToken.confidence, "low");

@@ -0,0 +1,76 @@
+import json
+import tempfile
+import unittest
+from pathlib import Path
+
+from token_color_store import (
+    DEFAULT_PRESETS,
+    DEFAULT_RANDOM_POOL,
+    TokenColorStore,
+    normalize_color,
+)
+
+
+class TokenColorStoreTests(unittest.TestCase):
+    def make_store(self, payload=None):
+        temp = tempfile.TemporaryDirectory()
+        self.addCleanup(temp.cleanup)
+        store = TokenColorStore(Path(temp.name))
+        if payload is not None:
+            store.path.write_text(json.dumps(payload), encoding="utf-8")
+        return store
+
+    def test_normalize_color_accepts_short_and_long_hex(self):
+        self.assertEqual(normalize_color("#abc"), "#aabbcc")
+        self.assertEqual(normalize_color("#AABBCC"), "#aabbcc")
+        self.assertIsNone(normalize_color("#abcde"))
+        self.assertIsNone(normalize_color("red"))
+        self.assertIsNone(normalize_color(None))
+
+    def test_missing_file_falls_back_to_defaults(self):
+        config = self.make_store().config()
+        self.assertEqual(config["presets"], DEFAULT_PRESETS)
+        self.assertEqual(config["random_pool"], DEFAULT_RANDOM_POOL)
+        self.assertEqual(config["pack_colors"], {})
+        self.assertEqual(config["anima_colors"], {})
+
+    def test_broken_or_partial_file_still_reads_pools(self):
+        config = self.make_store({"presets": ["#112233"], "random_pool": [], "pack_colors": "bad"}).config()
+        self.assertEqual(config["presets"], ["#112233"])
+        self.assertEqual(config["random_pool"], [])
+        self.assertEqual(config["pack_colors"], {})
+
+    def test_save_colors_writes_back_and_keeps_pools(self):
+        store = self.make_store()
+        config = store.save_colors(pack_colors={"personal": "#ABC", "camera": None})
+        self.assertEqual(config["pack_colors"], {"personal": "#aabbcc"})
+        self.assertEqual(config["presets"], DEFAULT_PRESETS)
+        written = json.loads(store.path.read_text(encoding="utf-8"))
+        self.assertEqual(written["pack_colors"], {"personal": "#aabbcc"})
+
+    def test_save_colors_rejects_bad_input(self):
+        store = self.make_store()
+        with self.assertRaises(ValueError):
+            store.save_colors(pack_colors={"personal": "not-a-color"})
+        with self.assertRaises(ValueError):
+            store.save_colors(anima_colors={"bad id!": "#112233"})
+        with self.assertRaises(ValueError):
+            store.save_colors(pack_colors=[])
+
+    def test_empty_value_clears_existing_pack_color(self):
+        store = self.make_store({"pack_colors": {"personal": "#112233"}})
+        self.assertEqual(store.config()["pack_colors"], {"personal": "#112233"})
+        self.assertEqual(store.save_colors(pack_colors={"personal": ""})["pack_colors"], {})
+
+    def test_anima_colors_are_saved_alongside_pack_colors(self):
+        store = self.make_store({"pack_colors": {"personal": "#112233"}})
+        config = store.save_colors(anima_colors={"quality": "#ABC", "expression": ""})
+        self.assertEqual(config["anima_colors"], {"quality": "#aabbcc"})
+        # 只传了一类，另一类保持原样
+        self.assertEqual(config["pack_colors"], {"personal": "#112233"})
+        self.assertEqual(store.config()["anima_colors"], {"quality": "#aabbcc"})
+
+    def test_save_colors_without_payload_is_rejected(self):
+        store = self.make_store()
+        with self.assertRaises(ValueError):
+            store.save_colors()
@@ -0,0 +1,130 @@
+"""节点标签卡的配色配置。
+
+色块池（presets / random_pool）只读，用户可以直接编辑 data/token_colors.json 增删色块；
+按词库配色（pack_colors）由界面写回同一个文件，格式不对时回退默认，不影响功能。
+"""
+
+import json
+import os
+import re
+import tempfile
+from pathlib import Path
+
+
+DEFAULT_PRESETS = [
+    "#6b9b78", "#9b8b6b", "#6b7b9b", "#9b6b8b",
+    "#8b9b6b", "#6b9b9b", "#9b7b6b", "#7b6b9b",
+]
+DEFAULT_RANDOM_POOL = [
+    "#5b8a72", "#8a725b", "#5b6e8a", "#8a5b7a",
+    "#7a8a5b", "#5b8a8a", "#8a6b5b", "#6b5b8a",
+]
+# 未命中任何词库的标签也当成一个可配色的来源，key 固定为 unindexed
+UNINDEXED_PACK_ID = "unindexed"
+
+MAX_COLOR_ENTRIES = 200
+PACK_ID_PATTERN = re.compile(r"[A-Za-z0-9_-]{1,64}")
+HEX_COLOR_PATTERN = re.compile(r"#[0-9a-fA-F]{6}")
+SHORT_HEX_COLOR_PATTERN = re.compile(r"#[0-9a-fA-F]{3}")
+
+
+def normalize_color(value):
+    """把 #RGB / #RRGGBB 统一成小写的 #rrggbb，认不出来返回 None。"""
+    text = str(value or "").strip()
+    if HEX_COLOR_PATTERN.fullmatch(text):
+        return text.lower()
+    if SHORT_HEX_COLOR_PATTERN.fullmatch(text):
+        return "#" + "".join(char * 2 for char in text[1:]).lower()
+    return None
+
+
+class TokenColorStore:
+    def __init__(self, data_dir=None):
+        self.data_dir = Path(data_dir or Path(__file__).resolve().parent / "data")
+        self.path = self.data_dir / "token_colors.json"
+
+    @staticmethod
+    def defaults():
+        return {
+            "presets": list(DEFAULT_PRESETS),
+            "random_pool": list(DEFAULT_RANDOM_POOL),
+            "pack_colors": {},
+            "anima_colors": {},
+        }
+
+    def config(self):
+        """读取配置：色块池原样透传，缺字段或文件坏了就回退默认。"""
+        payload = self._read()
+        if payload is None:
+            return self.defaults()
+        config = self.defaults()
+        for key in ("presets", "random_pool"):
+            values = payload.get(key)
+            if isinstance(values, list):
+                config[key] = values
+        for key in ("pack_colors", "anima_colors"):
+            values = payload.get(key)
+            if isinstance(values, dict):
+                config[key] = self._clean_map(values)
+        return config
+
+    @staticmethod
+    def _clean_map(raw):
+        cleaned = {}
+        for key, value in raw.items():
+            color = normalize_color(value)
+            if color and PACK_ID_PATTERN.fullmatch(str(key)):
+                cleaned[str(key)] = color
+        return cleaned
+
+    @staticmethod
+    def _validated_map(raw, label):
+        if not isinstance(raw, dict):
+            raise ValueError(f"{label}必须是对象")
+        if len(raw) > MAX_COLOR_ENTRIES:
+            raise ValueError(f"{label}最多 {MAX_COLOR_ENTRIES} 项")
+        cleaned = {}
+        for raw_key, value in raw.items():
+            key = str(raw_key).strip()
+            if not PACK_ID_PATTERN.fullmatch(key):
+                raise ValueError(f"分类ID格式不正确：{key}")
+            if value in (None, ""):
+                continue
+            color = normalize_color(value)
+            if color is None:
+                raise ValueError(f"颜色必须是 #RGB 或 #RRGGBB 十六进制值：{value}")
+            cleaned[key] = color
+        return cleaned
+
+    def save_colors(self, pack_colors=None, anima_colors=None):
+        """整体覆盖某一类配色：传空值表示该项恢复默认。只更新传进来的那一类。"""
+        if pack_colors is None and anima_colors is None:
+            raise ValueError("没有要保存的配色")
+        config = self.config()
+        if pack_colors is not None:
+            config["pack_colors"] = self._validated_map(pack_colors, "按词库配色")
+        if anima_colors is not None:
+            config["anima_colors"] = self._validated_map(anima_colors, "按 Anima 配色")
+        self._write(config)
+        return config
+
+    def _read(self):
+        if not self.path.exists():
+            return None
+        try:
+            payload = json.loads(self.path.read_text(encoding="utf-8"))
+        except (OSError, ValueError):
+            return None
+        return payload if isinstance(payload, dict) else None
+
+    def _write(self, payload):
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        fd, temp_name = tempfile.mkstemp(prefix="token_colors_", suffix=".json", dir=self.path.parent)
+        try:
+            with os.fdopen(fd, "w", encoding="utf-8") as handle:
+                json.dump(payload, handle, ensure_ascii=False, indent=2)
+                handle.write("\n")
+            os.replace(temp_name, self.path)
+        finally:
+            if os.path.exists(temp_name):
+                os.unlink(temp_name)
@@ -0,0 +1,95 @@
+"""一键跑全部测试（JS + Python）。
+
+用法（在插件目录下，或者任何地方都行）：
+    X:\\webui\\comfyui\\ComfyUI\\.venv\\Scripts\\python.exe tools\\main_test.py
+
+会跑两轮：
+    1. tests\\*.mjs      —— JS 测试
+    2. tests\\test_*.py  —— Python 测试（PYTHONPATH 自己设好，不用手动加）
+
+两轮都会实时输出，能看着它一条条走完。最后给一个汇总，
+有失败的会单独列出来。全部通过退出码 0，有失败 1。
+
+注意：必须用 ComfyUI 的 venv 跑（server.py 要 aiohttp），
+用别的 Python 会出现 "No module named 'aiohttp'" 且用例数偏少。
+"""
+
+import os
+import pathlib
+import subprocess
+import sys
+
+# 这个脚本放在 tools/ 下（该目录被 .comfyignore 排除，不进发布包），
+# 所以插件根目录在上一级。放根目录时也能用，这里两种位置都认。
+ROOT = pathlib.Path(__file__).resolve().parent
+if not (ROOT / "tests").is_dir():
+    ROOT = ROOT.parent
+TESTS = ROOT / "tests"
+
+# 不设成行缓冲的话，接管道（python main_test.py | tail）时我们自己的输出会被攒到最后，
+# 跟子进程的实时输出插在一起，看起来顺序全乱。
+sys.stdout.reconfigure(line_buffering=True)
+
+
+def find_node():
+    """找 node 可执行文件。
+
+    managed node 的目录号会随环境升级变（22.22.2-3 → 22.22.2-6），
+    写死路径会某天突然全部 FAIL，所以这里动态取最新的那个。
+    """
+    versions = pathlib.Path.home() / ".workbuddy" / "binaries" / "node" / "versions"
+    if versions.is_dir():
+        candidates = [p for p in versions.iterdir() if p.is_dir() and p.name[:1].isdigit()]
+        for candidate in sorted(candidates, reverse=True):
+            exe = candidate / ("node.exe" if os.name == "nt" else "node")
+            if exe.exists():
+                return str(exe)
+    return "node"
+
+
+def run_js():
+    files = sorted(TESTS.glob("*.mjs"))
+    if not files:
+        print("[js] 没有找到 tests/*.mjs，跳过")
+        return []
+    node = find_node()
+    print(f"\n=== JS 测试（{len(files)} 个）用 {node} ===")
+    failed = []
+    for path in files:
+        result = subprocess.run([node, str(path)], cwd=ROOT)
+        if result.returncode == 0:
+            print(f"  PASS {path.name}")
+        else:
+            print(f"  FAIL {path.name}")
+            failed.append(path.name)
+    return failed
+
+
+def run_python():
+    print(f"\n=== Python 测试 用 {sys.executable} ===")
+    env = dict(os.environ, PYTHONPATH=str(ROOT))
+    result = subprocess.run(
+        [sys.executable, "-m", "unittest", "discover", "-s", "tests", "-p", "test_*.py"],
+        cwd=ROOT,
+        env=env,
+    )
+    return [] if result.returncode == 0 else ["python 测试（见上面的输出）"]
+
+
+def main():
+    if not TESTS.is_dir():
+        sys.exit(f"[test] 找不到 {TESTS}")
+
+    failed = run_js() + run_python()
+
+    print("\n" + "=" * 46)
+    if failed:
+        print(f"有 {len(failed)} 项失败：")
+        for name in failed:
+            print(f"  - {name}")
+        sys.exit(1)
+    print("全部通过")
+
+
+if __name__ == "__main__":
+    main()
Author	SHA1	Message	Date
梦圣智杰	575ab1814f	chore: 发布 v2.1.9 / Release v2.1.9	2026-10-29 00:54:45 +08:00
梦圣智杰	7759128dac	fix: 设置面板两项统一靠左对齐 / Align the two settings entries to the left edge - 「语言与外观」与「助手设置」两块的内容改为顶到最左边，左边缘齐平，看起来是一整块 - 标题仍在：由各自的 h3 显示，收起的只是 ComfyUI 重复渲染出来的那个名称列	2026-10-29 00:51:33 +08:00
梦圣智杰	2cf72b1a10	feat: 助手设置可自动获取模型列表 / Fetch model list automatically in assistant settings - OpenAI 兼容与 Ollama 两段的「模型名称」旁新增「获取模型」按钮，拉到的名字填进输入框自带的下拉，点箭头即可选择，名称也照旧可以手写 - 下拉挂在输入框上、不弹窗，ComfyUI 设置面板不用关也能用；按键即重新拉取，按钮显示拿到了几个 - 密钥以输入框里填的为准，留空就用已保存的那一份，只随这一次请求发出 - 列表只在本次会话内有效，地址一改自动清空，不写进配置文件 - 服务端返回非 JSON 的内容（纯文本报错、登录页之类）时，界面直接显示原文与状态码	2026-10-29 00:02:32 +08:00
梦圣智杰	194a6a8206	chore: 发布 v2.1.8 / Release v2.1.8	2026-10-26 05:42:02 +08:00
梦圣智杰	d38de335bd	feat: 新增「适应高度」开关，提示词区不再被高度上限卡住 / Add a fit-height switch so the prompt area is no longer height-capped - 提示词多的时候英文区卡在写死的高度上限里只能内部滚动，把节点拉大也没用；侧边栏「标签管理」的明细表上方新增「适应高度」滑动开关 - 打开后提示词区跟着节点一起变高，打开时节点自动撑高一次，之后可以随意拖动，不会被自动拽回原来的大小 - 开关状态跟着节点保存，切到别的节点显示那个节点自己的状态，切回来不变	2026-10-26 05:37:36 +08:00
梦圣智杰	f2d3de3659	feat: 「优化」按模式分派，新增自然语言规则 / Optimize follows input mode with a new natural language rule - 「优化为 Anima」改名「优化」：标签模式整理标签流，自然语言模式输出英文句子，中文也能直接用 - 助手设置新增一条「自然语言规则」，两个 AI 后端各一套 - 自然语言模式下隐藏「按 Anima 排序」，模式选择跟着节点保存	2026-10-26 04:27:40 +08:00
梦圣智杰	464a6c79dc	fix: 拖大重复检查节点时状态胶囊不再被撑变形 / Keep the duplicate pill at its own size when the node is resized - 手动拖大「重复检查」节点时，节点上「重复 N」的胶囊不再跟着变高变成一个巨大的圆角块，始终保持原本大小并垂直居中	2026-10-13 07:46:11 +08:00
梦圣智杰	e2b2834298	chore: 发布 v2.1.5 / Release v2.1.5	2026-10-13 03:10:54 +08:00
梦圣智杰	ae96399bdd	feat: 可信度按词条整理程度分档 / Rate confidence by how complete an entry is - 明细表的可信度徽章只显示「未判定 / 低 / 中 / 高」四档，去掉「待确认」这类后缀 - 已收录、已校验、且填了分类（不是「未分类」）才算高；已收录但没填分类，或还没标成已校验的，算中 - 机器翻译还没存进词库的算低；词库里查不到的仍是未判定 - 鼠标悬停在徽章上会说明为什么是这一档，填了分类的会把分类名带出来 - 「高」「中」补上中文译文，之前这两档一直显示的是英文	2026-10-13 02:20:02 +08:00
梦圣智杰	84bbc97cba	feat: 明细表双击已收录标签即可编辑词条 / Double-click an indexed tag to edit its entry - 明细表里已收录的标签，双击英文那格会打开「新增或编辑个人标签」窗口，和在词库里点「编辑」是同一个，中文、别名、分类、模型、权重、备注都能改 - 未收录、待确认的机器翻译，以及 BREAK、<lora:…> 这类特殊语法，双击仍是原来的定位到提示词 - 未收录行去掉「搜索候选」按钮：词库里本来就搜不出这个词，点了只会打开空结果	2026-10-13 02:02:26 +08:00
梦圣智杰	b09ce43898	fix: 自然语言片段双击也能开编辑弹窗 / Natural-language segments now open the editor too - 双击自然语言片段不再只提示「只有标签支持单独权重」，改为和标签一样打开编辑弹窗 - 自然语言片段只提供「颜色」和「编辑」两页，权重与词条信息不再出现 - 自然语言片段可以像标签一样设置单独配色	2026-10-13 01:13:36 +08:00
梦圣智杰	5bb839964a	fix: 未收录标签不再被存成残词条 / Stop saving the not-indexed placeholder as a tag name - 双击未收录标签打开「词条信息」时，「中文名称」改为留空，自己填上中文再点应用才会存入个人词库 - 中文名称还是「未收录」时点应用不会保存，下方说明变红提示，2 秒后自动恢复 - 明细表里直接改中文时，不再把「未收录」占位带进输入框	2026-10-12 11:35:36 +08:00
梦圣智杰	1233c9f0c9	feat:添加测试启动文件 / Add test startup file	2026-10-12 07:45:32 +08:00
梦圣智杰	ca916f5b82	chore: 发布 v2.1.0 / Release v2.1.0	2026-10-12 07:09:39 +08:00
梦圣智杰	6565157f2a	fix: 暗色下输入框与下拉框改用 ComfyUI 的配色 / Dark mode fields now use ComfyUI's own colors	2026-10-12 07:00:13 +08:00
梦圣智杰	737458163b	feat: AI 分成 OpenAI 与 Ollama 两套 / AI split into OpenAI and Ollama configs - 助手设置移到 ComfyUI「设置 → 双语提示词检查器 → 助手设置」，表单直接铺在设置面板右侧；侧边栏只保留标签管理、词库、词包、收藏四个页签。 - 外观与语言从助手设置里分出来，单独成为「语言与外观」一项，下拉改完立即保存并生效。 - OpenAI 兼容与 Ollama 各有独立的地址、模型、密钥、超时和三条翻译规则，切换后端互不覆盖，已填的密钥不会被清掉。 - API 地址只接受 https 或本机地址，Ollama 只能用本机地址，地址中不能带用户名密码或查询参数，远端不接受内网与保留地址。	2026-10-12 06:53:08 +08:00
梦圣智杰	d6c51f9de2	chore: 发布 v2.0.6 / Release v2.0.6	2026-10-10 23:05:14 +08:00
梦圣智杰	b8218f7e10	feat: 明细表可按关键词查找当前提示词 / Filter the detail table by keyword - 明细表上方新增查找框，输入即筛选出当前节点里命中的标签，英文原文和中文解释都参与匹配 - 命中片段会高亮，右侧显示命中数量，按 Esc 或点 × 一键清空 - 查找与「全部 / 未收录 / 机器翻译 / 个人词库 / 收藏」叠加生效。	2026-10-10 23:00:01 +08:00
梦圣智杰	01123cef5c	chore: 去掉跳词库的「管理面板」按钮 / Remove the redundant manager panel button - 侧边栏「标签管理」里的「管理面板」按钮和顶部「词库」页签是一回事，去掉后工具栏只留「刷新词库」 - 画布上节点面板里的同一个按钮也一并去掉，两边共用同一份面板	2026-10-10 19:09:04 +08:00
梦圣智杰	a8b7aca8f8	feat: 词库可按自然语言词库来源筛选 / Filter the dictionary by the natural language source - 来源下拉新增「自然语言词库」，只显示保存时选了该词库的条目；「个人」不受影响，仍包含它们 - 分类下拉里「自然语言」固定排在「全部」下面，其余分类仍按拼音排序	2026-10-10 18:33:18 +08:00
梦圣智杰	c72422e648	feat: 侧边栏节点列表显示画布上改过的节点名 / Show custom node titles in the sidebar node list - 标签管理的节点下拉、以及保存收藏时选节点的弹窗，会显示在画布上双击改过的节点名，后面保留节点号以便区分同名节点 - 没改过名的节点仍显示节点号加提示词摘要	2026-10-10 18:17:19 +08:00
梦圣智杰	ff8395302a	feat: Anima 排序可一键取消 / Cancel the Anima sort with one click - 排序写入后按钮变成「取消 Anima 顺序排序」，点击把提示词恢复到排序前的样子，同时退出分类表视图。 - 提示词本来就是 Anima 顺序时，取消只退出分类视图，不动提示词内容。 - 排序后又手动改过提示词，取消入口会自动失效，避免回退把后来的改动一起吞掉。 - 排序与优化的预览弹窗里，「确认写入英文」改为「确认写入新提示词」。	2026-10-10 18:03:11 +08:00
梦圣智杰	cab56b2f60	feat: 词库搜索添加的标签插到选中标签之后 / Insert searched tags after the selected tag - 在节点和侧边栏里搜索词库添加标签时，新标签会插在当前选中标签的后面，不再总是追加到提示词末尾；没有选中标签时才加到末尾。 - 连续添加多个标签会依次往后排，添加完新标签自动处于选中状态。	2026-10-10 17:43:17 +08:00
梦圣智杰	cd55defc87	chore: 发布 v2.0.0 / Release v2.0.0	2026-10-10 01:33:23 +08:00
梦圣智杰	79dc337c0e	feat: 合并重复检查节点到主线 / Merge duplicate-check node into main	2026-10-10 01:00:09 +08:00
梦圣智杰	642837842c	feat: 隐藏按钮挪到方框右上角内 / Move hide button inside the card corner - 标签卡上的隐藏（眼睛）按钮不再跟着英文行走，固定在整张卡片的右上角内侧 - 鼠标移到卡片任意位置（含中文行）都会显示该按钮	2026-10-10 00:30:18 +08:00
梦圣智杰	13c2fecce2	feat: 重复标签可设「唯一」并新增重复栏 / Unique toggle and duplicate bar for repeated tags - 重复分组抽成 parser.js 的 duplicateGroups()，问题提示行与界面共用同一份口径 - 「重复标签：X」后面加每个标签独立的「唯一」开关：只保留第一次出现进入实际输出，其余在界面上划线保留，不删原文，支持 Ctrl+Z - 英文标签区「已隐藏」下方新增重复栏，一个标签一个胶囊，单击依次定位到每一次出现并联动选中（标红），已定位时双击删掉当前那一处 - 新增 tests/test_duplicates.mjs，14 个 JS 测试全绿	2026-10-06 17:16:49 +08:00
梦圣智杰	7cb8d80b02	feat: 新增重复检查节点 / Add duplicate tag checker node - 检测提示词里的重复标签，节点用淡橙色提示重复数量，无重复时保持默认配色 - 开关「只保留第一个」默认关闭，打开后只保留第一次出现的标签 - 输出提示词直通下游，重复项可接预览任意查看；判定口径沿用 parser.js	2026-10-05 22:42:56 +08:00
梦圣智杰	ed0ab5eee2	chore: 发布 v1.2.18 / Release v1.2.18	2026-10-05 09:19:41 +08:00
梦圣智杰	d45a9d9e63	feat: 标签编辑弹窗整合 / Consolidate tag editor	2026-10-05 08:27:34 +08:00
梦圣智杰	35ed25d69c	chore: 界面细节调整 / UI polish	2026-10-05 08:22:37 +08:00
梦圣智杰	d3e6d7edce	chore: 发布 v1.2.16 / Release v1.2.16	2026-10-01 02:50:30 +08:00
梦圣智杰	43a049b43f	feat: 快捷添加支持按选中位置插入 / Quick add inserts after the selected tag - 选中标签时用 token.end 偏移插到它后面，没选中或选中失效则落到末尾，与「插入换行」同一套判定	2026-10-01 02:00:12 +08:00
梦圣智杰	78924e8244	fix: 隐藏区长文本胶囊不再被撑成大圆角块 / Keep long hidden-tag chips on one line - 胶囊里的中文译文缺少 nowrap，长自然语言会在里面换行把高度撑起来， - 中英文两段都加单行截断，胶囊高度恒定一行，超出部分显示省略号	2026-10-01 01:35:38 +08:00
梦圣智杰	905d165348	chore: 发布 v1.2.15 / Release v1.2.15	2026-09-30 01:57:49 +08:00
梦圣智杰	3806ae7ae1	chore:添加了在ComfyUI注册表上面的图标/Added an icon on the ComfyUI registry	2026-09-30 01:53:44 +08:00
梦圣智杰	45eb64a5bf	feat: 双击标签卡的编辑框加「词条信息」页签 / Add an entry-info tab to the tag editor - 标签字段抽成 buildTagFields，弹窗与页签共用同一份 - 新页签展开完整词条 - 命中词库就整条带回来自改，没命中按当前标签卡起个头 - 「Anima 格式」说明挪进权重页，「清除权重」只在权重页显示 - 页签状态由 onColorTab 布尔改成 activeTab 三态，共用底部 Apply 分派	2026-09-29 21:07:39 +08:00
梦圣智杰	c19242700e	chore: 发布 v1.2.14 / Release v1.2.14	2026-09-29 03:05:30 +08:00
梦圣智杰	6fe88a6bae	feat: 标签弹窗的分类改成可筛选下拉 / Turn the category field into a filterable dropdown - 候选取词库里真实出现过的分类，按词条数降序，每项标出条数 - 「自定义」恒为列表第一项且不被过滤掉，选中后输入框右侧挂青绿色标识 - 手打的值命中候选时自动摘掉标识；编辑历史标签带进来的陌生值当场算自定义 - 输入即筛选，↑↓ 切换、回车确认、Esc 收起，↓ 或再点输入框可重新唤出列表 - 新建时分类留空，由后端落成「自定义」，避免一开窗就挂着标识	2026-09-29 02:52:21 +08:00
梦圣智杰	238612920c	feat: 标签卡支持按词库、按 Anima 上色，并标记疑似自然语言 / Color tag cards by dictionary or Anima, flag suspected natural language - 过滤栏「颜色」改成就地展开面板：默认 / 按收录 / 随机 / 按词库 / 按 Anima - 配色写入 data/token_colors.json（pack_colors / anima_colors），新增写接口 - 新增「疑似自然语言」种类（空格 > 3），节点卡片与明细表着色 - 自然语言词库拆出独立 pack_id - 双击设的自定义色补上优先级说明	2026-09-29 02:18:01 +08:00
梦圣智杰	84d3430576	feat: 识别并标记「疑似自然语言」，新增自然语言词库 / Flag suspected natural language and add a natural language dictionary	2026-09-29 00:22:12 +08:00
梦圣智杰	125654864c	fix: 优化为 Anima 不再被转义符、句点和权重小数打断 / Fix Anima optimization output rejected by escapes, stray periods and decimal weights - assistant_store：清洗表补全角引号；清掉模型写成 JSON/代码字符串时残留的反斜杠（字面 \n \t 先还原成换行，其余孤立反斜杠删除） - assistant_store：末尾孤立句点直接删除 - parser：判定句末标点前先中和数字句点与闭括号后的句点，权重小数不再让整串标签被当成自然语言段落 - 测试：新增转义残留、句点替换、权重保留等断言；既有「@artist！」用例同步改为清洗后不带句点	2026-09-28 22:49:04 +08:00
梦圣智杰	97c87043d7	ci: 后端测试前安装 aiohttp / Install aiohttp before backend tests	2026-09-27 21:23:17 +08:00
梦圣智杰	36daee2bed	ci: 接入 Comfy Registry 自动发布 / Add registry publish workflow	2026-09-27 21:05:12 +08:00