dsh-cait 1.0.5 → 1.1.3

package/README.md

@@ -16,7 +16,7 @@
 <p align="center">
   <a href="https://www.npmjs.com/package/dsh-cait"><img alt="npm version" src="https://img.shields.io/npm/v/dsh-cait.svg" /></a>
   <img alt="DSH 0.2.x" src="https://img.shields.io/badge/DSH-0.2.x-4d6bfe.svg" />
-  <img alt="Tests: 15 suites" src="https://img.shields.io/badge/tests-15%20suites-brightgreen.svg" />
+  <img alt="Tests: 29 suites" src="https://img.shields.io/badge/tests-29%20suites-brightgreen.svg" />
   <img alt="No API key required" src="https://img.shields.io/badge/API%20key-not%20required-6b7280.svg" />
 </p>
 
@@ -34,7 +34,7 @@ translated document.
 | | |
 |---|---|
 | **Native to DSH** | A real sidebar panel, a main workbench, DSH theme tokens, and the same store for humans and agents. No iframe, no separate app. |
-| **TM that explains itself** | Multiple TM libraries, mounted-only lookup, exact / near-exact / high / medium / low fuzzy grades, and automatic write-back on confirm. |
+| **TM that explains itself** | Multiple TM libraries, mounted-only lookup, 101% context (ICE) matches, exact / near-exact / high / medium / low fuzzy grades, and automatic write-back on confirm. |
 | **Glossary-aware editing** | Multiple term bases, source-language filtering, inline highlighting, click-to-insert terms, and per-term case sensitivity. |
 | **A document workflow** | Tabs, close/reopen, recent-document cleanup, duplicate-import reuse, and progress per document. |
 | **Real format coverage** | TXT, Markdown, DOCX, EPUB, TMX, CSV, SRT/VTT, PO, XLIFF 1.2/2.0/SDLXLIFF, HTML, JSON, XLSX, PPTX, and RTF. |
@@ -60,6 +60,10 @@ manual reload.
 ### Translation memory
 
 - Multiple TM libraries with id, name, mount state, and a default write library.
+- **Context match (ICE) at 101%.** Each confirmation records the two neighbouring
+  source sentences. A later segment with the same neighbours scores 101% and is
+  graded `context`; identical text with different neighbours stays at 100%. A key
+  keeps up to 8 context variants and always keeps the context-free one.
 - Exact, near-exact, and three fuzzy bands:
   - `near_exact` compares a key that ignores digits, punctuation, and whitespace;
     it does not depend on a score cutoff.
@@ -74,7 +78,8 @@ manual reload.
   `bootstrap().tm.count` is the mounted count and `bootstrap().tm.total` is the
   all-library count. Use `cat_lookup_tm` or the library modal for full data.
 - TMX 1.4 `<tu>/<tuv>/<seg>` import/export. Inline TMX placeholders are flattened
-  to plain text with an explicit warning.
+  to plain text with an explicit warning. Context travels as
+  `<prop type="x-context">`, so a CAIT round-trip keeps 101% matches.
 
 ### Glossary
 
@@ -83,11 +88,71 @@ manual reload.
 - CSV/TSV import, per-term notes, source-language filtering, and a
   `caseSensitive` flag honored by both the workbench and `cat_lookup_glossary`.
 - Language codes are canonicalized (`EN` → `en`, `ZH_cn` → `zh-CN`).
+- Entries can be edited or deleted in any library, mounted or not; mount state
+  only decides what the workbench searches and writes.
+- **Terminology mining**: *Mine terms from the document* in the glossary toolbar
+  lists candidates with occurrence counts, sample wording from memory, and a
+  marker when the term is already known. Counting finds the candidates and ranks
+  them (Latin n-grams, repeated CJK substrings filtered by boundary entropy);
+  a model only judges a shortlist, which is the part models are good at.
+
+### Library workspace
+
+`Documents` / `Libraries` / `Align` in the topbar switches between the translator
+and the library manager. The library view needs no open document.
+
+- One table for translation memory, one for glossaries, with paging
+  (20/50/100/200, remembered), language filters, and a search that can be scoped
+  to any combination of libraries.
+- A side panel edits the selected entry in place, adds entries, and shows when a
+  draft is unsaved.
+- Translation memory exports to **TMX**; glossaries export to **TBX**, **CSV**, or
+  **JSON**. Import and clear stay available in the same toolbar.
+- The right panel's library button opens this workspace with the panel's search
+  text already applied, so a lookup in the panel continues in the full list.
+
+### Alignment
+
+`Documents` / `Libraries` / `Align` — the third view builds a translation
+memory from two finished files: the source document and its translation, as
+separate files in any supported format and any pair of languages.
+
+- Cross-language scoring: shared anchors (numbers, `{0}` placeholders, URLs,
+  capitalised names), length normalised per language, and paragraph structure.
+  Word overlap is deliberately not used, because it cannot compare two languages.
+- Blocked dynamic programming pairs 1:1, 1:0, 0:1, 2:1 and 1:2; rows are graded
+  high / to review / doubtful and start selected only when they are high.
+- Each row can be merged with the next, split on line breaks, or removed before
+  writing; the target library is chosen in the toolbar, and identical entries are
+  skipped rather than duplicated.
 
 ### Workbench
 
 - Paragraph/sentence segmentation with a per-segment target editor.
-- Draft saving, confirm-and-next, previous/next navigation, and progress.
+- Segment states: **pending** → **translated** → **confirmed**, with the origin of
+  each translation (`tm`, `mt`, `agent`, `manual`) kept alongside the text. The
+  rail filters by state and search covers source, target, and suggestions.
+- **Segment paging**: the rail pages through the filtered segments with a page
+  size of 20/50/100/200 and a box to jump straight to a segment number. Jumping
+  clears the filter when the target segment would otherwise be hidden.
+- **Pre-translate** fills the whole document from exact TM matches and machine
+  translation in small blocks, shows `done/total`, and can be stopped between
+  blocks. Confirmed and already-translated segments are never overwritten.
+- **Stacked review**: source, the machine/agent baseline, and the editable target
+  are laid out top to bottom, so each block gets the full pane width.
+  **Confirm all** accepts every segment awaiting review after one confirmation
+  dialog.
+- Draft saving, confirm-and-next, previous/next navigation, and progress. The
+  selected segment and the draft are persisted (debounced) so a restart resumes
+  where you left off.
+- **Repeated segments are mirrored**: confirming a sentence fills every other
+  unconfirmed copy with the same source (`origin: auto`), editing a draft keeps
+  them in sync, and *Fill repeated segments* does the whole document at once.
+  Confirmed segments are never touched.
+- Advisory QA markers in the rail and an export-time summary: missing
+  placeholders/tags, mismatched numbers, unused glossary terms, empty targets,
+  an untouched copy of the source, and one source with several translations.
+  Nothing is blocked and nothing is stored.
 - Document tabs; closing a tab does not delete the document.
 - **Open existing document** reopens a closed document.
 - Interface language switch in **Settings → Interface**: *Follow DSH language*
@@ -100,9 +165,12 @@ manual reload.
 - Re-importing identical content in the same format reuses the existing document
   instead of creating a duplicate. The fingerprint covers languages, source,
   format, and exchange metadata.
-- Responsive density: the workbench measures its DSH panel and applies a
-  `--cait-u` multiplier of 1×–1.8×. A 1600×900 panel renders at ~1.32× and a
-  2560×1440 panel at 1.8×; short windows stay at 1× rather than shrinking.
+- Responsive density: the workbench measures its DSH panel and applies a damped
+  `--cait-u` multiplier of 1×–1.45×. A 1600×900 panel renders at ~1.17× and a
+  maximized 4K panel stops at 1.45×, which keeps large windows close to the host
+  UI and leaves room for the three review columns.
+- The workbench follows the host light/dark theme, including native controls
+  (`color-scheme`), so selects, checkboxes, and scrollbars match the panel.
 
 ### Preprocessing
 
@@ -141,7 +209,7 @@ provider/model/code/status; the client receives only `MT_PROVIDER_FAILED`.
 | CSV | ✅ | ✅ | Comma, semicolon, tab, or pipe delimiter auto-detected; source/target column by header name or 1-based number; existing bilingual export |
 | DOCX | ✅ | ✅ | Body paragraphs; table blocks are excluded from import. Export patches the original package, keeping styles, numbering, images, headers, tables, and layout. |
 | EPUB | ✅ | ✅ | EPUB 3 spine order; XHTML text blocks with inline markup kept as tokens and patched back into the retained package. DRM-protected, fixed-layout, and scripted EPUBs are not supported. |
-| TMX 1.4 | ✅ | ✅ | Common `<tu>/<tuv>/<seg>` subset; inline markup flattened with warning |
+| TMX 1.4 | ✅ | ✅ | Common `<tu>/<tuv>/<seg>` subset; inline markup flattened with warning; context kept in `<prop type="x-context">` |
 | SRT / WebVTT | ✅ | ✅ | One cue per segment, timings/settings/identifier retained; semantic tags kept; invalid ranges rejected |
 | PO | ✅ | ✅ | Contexts, plural forms, translator comments, obsolete `#~` entries |
 | XLIFF 1.2 | ✅ | ✅ | Inline tags become `{n}` tokens; missing tokens block confirmation |
@@ -195,7 +263,7 @@ In DSH Desktop (or any profile with the `plugin_manager` tool):
 plugin_manager install_bundle target: dsh-cait
 ```
 
-Pin a release with `dsh-cait@1.0.5`, or ask for the newest one with
+Pin a release with `dsh-cait@1.1.3`, or ask for the newest one with
 `dsh-cait@latest`. The manager resolves the registry package, applies the bundle
 patch, and activates the plugin — no hand-written profile rows.
 
@@ -276,8 +344,9 @@ confirmation. Other structural edits remain best-effort and surface
 
 If the snapshot is corrupt or has an unknown version, CAIT logs a warning, moves
 it to `store.json.corrupt-<timestamp>`, and starts empty instead of failing
-activation. Version-1 snapshots are migrated to version 2 with default TM and
-glossary libraries.
+activation. Version-1 and version-2 snapshots are migrated to version 3 with
+default TM and glossary libraries; a version-2 draft that had text but was never
+confirmed becomes `translated` with `origin: manual`.
 
 ## Agent tools
 
@@ -290,6 +359,20 @@ glossary libraries.
 | `cat_list_documents` | List documents and progress |
 | `cat_switch_document` | Select the active document |
 | `cat_suggest_mt` | MT via the selected DSH route, or the explicit offline fallback |
+| `cat_pre_translate` | Fill one block from TM + MT, skipping confirmed/translated segments; returns `next` |
+| `cat_get_segments` | Bounded segment window with status, origin, and suggestions |
+| `cat_confirm_segments` | Confirm a list of indices, a window, or everything awaiting review |
+| `cat_propose_translation` | Leave a suggestion plus a reason without touching the confirmed target |
+| `cat_align` | Align two texts and report high / to-review / doubtful pairs |
+| `cat_consistency_report` | Group repeated sources that carry different translations |
+| `cat_term_candidates` | Mining shortlist: frequency, sample sentences, existing memory wording |
+| `cat_add_terms` | Add accepted terminology to a library in one batch |
+
+An agent can therefore draft a whole document without clicking in the UI:
+`cat_pre_translate` with `from: next` until `next` is `null`, then
+`cat_get_segments` to review, then `cat_confirm_segments` to accept. For
+terminology it reads `cat_term_candidates`, judges the shortlist, and writes the
+result with `cat_add_terms`.
 
 The workbench and tools share one `CatStore`, so an agent can translate the same
 document the user has open.
@@ -300,7 +383,7 @@ document the user has open.
 flowchart LR
   UI["DSH Web UI<br/>(sidebar + workbench)"] --> Client["client.js<br/>generated Web bundle"]
   Client --> API["/cait/api/rpc<br/>same-origin JSON"]
-  Tools["7 agent tools"] --> Store["CatStore"]
+  Tools["15 agent tools"] --> Store["CatStore"]
   API --> Store
   Store --> Snapshot[("store.json<br/>+ sources/")]
 ```
@@ -335,6 +418,8 @@ cait/
 │  ├─ errors.js
 │  ├─ store-utils.js
 │  ├─ libraries.js          # TM/glossary registries + mounts
+│  ├─ library-view.js       # Library paging/search + export entry points
+│  ├─ library-export.js     # Glossary TBX / CSV / JSON builders
 │  ├─ tm.js
 │  ├─ glossary.js
 │  ├─ persistence.js
@@ -343,6 +428,8 @@ cait/
 │  ├─ html.js               # Shared HTML/XHTML run scanner + patcher
 │  ├─ epub.js               # EPUB package read / spine import / patch export
 │  ├─ preprocess.js         # Import-time source cleaning stage
+│  ├─ pre-translate.js      # Block pre-translation methods
+│  ├─ qa.js                 # Derived QA report (never persisted)
 │  ├─ xlsx.js
 │  ├─ pptx.js
 │  ├─ segmenter.js
@@ -383,11 +470,20 @@ node _dev/test-libraries.mjs
 node _dev/test-mounts.mjs
 node _dev/test-ui-scale.mjs
 node _dev/test-mt.mjs
+node _dev/test-mt-routes.mjs
+node _dev/test-three-state.mjs
+node _dev/test-pretranslate.mjs
+node _dev/test-qa.mjs
+node _dev/test-library-crud.mjs
+node _dev/test-library-view.mjs
+node _dev/test-pagination.mjs
+node _dev/test-theme-tokens.mjs
+node _dev/test-e2e-1.1.mjs
 ```
 
 The import/export fixtures are `_dev/sample.docx`, `_dev/sample-table.docx`, and
 `_dev/sample-format.docx`; regenerate them with the matching
-`python _dev/make-*.py` scripts. All 15 suites should pass.
+`python _dev/make-*.py` scripts. All 29 suites should pass.
 
 ### Releasing
 
@@ -444,13 +540,18 @@ GitHub with a release.
 - HTML/RTF export does not round-trip original styling; JSON translates string
   values only.
 - TM matching is lexical, not semantic; no stemming, synonyms, or embeddings.
-- Library rename/delete and per-tab mount state are not implemented yet.
+- One translation per source text per context: a repeated sentence keeps a
+  separate variant for each surrounding pair, capped at 8 variants per key, and
+  reuses the generic entry when the neighbours differ.
+- Context is the two neighbouring source sentences only; document, customer, or
+  project metadata is not part of the match.
+- TBX and JSON glossary exports are one-way; only TMX and CSV import back.
 - The workbench UI ships in Chinese, English, and Japanese, switchable in
   Settings; this README ships in English, Simplified Chinese, and Japanese, and
   the English version is the source of truth.
 
-Planned next: image/vision input through a multimodal DSH route, richer library
-management, style-preserving XLSX output, and XLIFF `<alt-trans>` / notes support.
+Planned next: image/vision input through a multimodal DSH route,
+style-preserving XLSX output, and XLIFF `<alt-trans>` / notes support.
 
 ## Changelog and versioning
 
@@ -458,9 +559,15 @@ See [CHANGELOG.md](CHANGELOG.md). `v1.0.0` was the first public release;
 `v1.0.1` adds the Japanese UI and the in-app interface language switch,
 `v1.0.2` makes *Follow DSH language* honour the DSH locale, `v1.0.3`
 publishes the package to npm, `v1.0.4` moves later releases to the
-provenance-publishing workflow, and `v1.0.5` adds EPUB, import preprocessing,
-and the PDF policy. The project was developed as `0.1.0` and
-renumbered for the GitHub release.
+provenance-publishing workflow, `v1.0.5` adds EPUB, import preprocessing,
+and the PDF policy, `v1.1.0` adds agent-driven pre-translation, the
+three-state segment model, stacked review, and advisory QA, `v1.1.1` adds
+the library workspace with TMX/TBX/CSV/JSON export, segment paging, and the
+panel-to-library jump, `v1.1.2` adds context-aware memory with 101% (ICE)
+matches, and `v1.1.3` adds two-file alignment, automatic propagation of
+repeated segments, terminology mining, and the proposal primitive agents use.
+The project
+was developed as `0.1.0` and renumbered for the GitHub release.
 
 > **License:** Apache License 2.0 — see [LICENSE](LICENSE). Redistributions must
 > retain the attribution in [NOTICE](NOTICE).

package/README.zh-CN.md

@@ -16,12 +16,12 @@
 <p align="center">
   <a href="https://www.npmjs.com/package/dsh-cait"><img alt="npm 版本" src="https://img.shields.io/npm/v/dsh-cait.svg" /></a>
   <img alt="DSH 0.2.x" src="https://img.shields.io/badge/DSH-0.2.x-4d6bfe.svg" />
-  <img alt="15 套测试" src="https://img.shields.io/badge/tests-15%20suites-brightgreen.svg" />
+  <img alt="29 套测试" src="https://img.shields.io/badge/tests-29%20suites-brightgreen.svg" />
   <img alt="无需 API Key" src="https://img.shields.io/badge/API%20key-not%20required-6b7280.svg" />
 </p>
 
 CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从左侧栏入口打开，
-工作台与七个 Agent 工具共用一个数据存储，覆盖从导入源文件到导出译文的完整流程。
+工作台与十五个 Agent 工具共用一个数据存储，覆盖从导入源文件到导出译文的完整流程。
 
 > **已在 DSH `0.2.2-rc.1` 中验证。** CAIT 使用 DSH 自己的模型路由和凭据，
 > 不会要求用户填写 API Key。内部回退提供方仍叫 `mock`，界面显示为
@@ -32,11 +32,11 @@ CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从
 | | |
 |---|---|
 | **原生集成 DSH** | 真正的侧栏入口、主工作台、DSH 主题令牌，以及人和 Agent 共用的同一份数据。不是 iframe，也不是另一个独立应用。 |
-| **可解释的翻译记忆** | 支持多个 TM 库、仅检索已挂载库、精确/近似精确/高/中/低模糊等级，确认句段时自动写回。 |
+| **可解释的翻译记忆** | 支持多个 TM 库、仅检索已挂载库、上下文 101%、精确/近似精确/高/中/低模糊等级，确认句段时自动写回。 |
 | **术语感知编辑** | 支持多个术语库、按源语言过滤、行内高亮、点击插入术语，以及按术语设置是否区分大小写。 |
 | **完整文档流程** | 文档标签页、关闭/重新打开、最近文档清理、重复导入复用、逐文档进度。 |
 | **真实格式覆盖** | TXT、Markdown、DOCX、EPUB、TMX、CSV、SRT/VTT、PO、XLIFF 1.2/2.0/SDLXLIFF、HTML、JSON、XLSX、PPTX、RTF。 |
-| **面向 Agent** | 七个工具共享当前文档和存储，Agent 可以与用户同时翻译。 |
+| **面向 Agent** | 十个工具共享当前文档和存储，Agent 可以与用户同时翻译。 |
 
 ## 快速开始
 
@@ -56,6 +56,9 @@ CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从
 ### 翻译记忆
 
 - 支持多个 TM 库，每个库有 id、名称、挂载状态和默认写入库。
+- **上下文匹配（ICE）为 101%。** 每次确认都会记录前后两句源文；之后相邻句相同的
+  句段得 101%，等级为 `context`；文本相同但相邻句不同仍为 100%。每个 key 最多保留
+  8 条上下文变体，并且始终保留无上下文的那一条。
 - 支持精确、近似精确和三个模糊等级：
   - `near_exact` 使用忽略数字、标点和空白后的 key 比较，不依赖分数阈值。
   - `high_fuzzy` 为 `>= 0.85`，`medium_fuzzy` 为 `0.75–0.85`，
@@ -68,7 +71,8 @@ CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从
   `bootstrap().tm.count` 是已挂载数量，`bootstrap().tm.total` 是所有库的总数。
   完整数据请使用 `cat_lookup_tm` 或库面板。
 - 支持 TMX 1.4 `<tu>/<tuv>/<seg>` 导入/导出。TMX 内联占位符会扁平化为
-  纯文本，并给出明确警告。
+  纯文本，并给出明确警告。上下文以 `<prop type="x-context">` 往返，所以 CAIT
+  自己导出的文件再导入后仍保留 101% 匹配。
 
 ### 术语库
 
@@ -77,11 +81,57 @@ CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从
 - 支持 CSV/TSV 导入、术语备注、源语言过滤，以及工作台和
   `cat_lookup_glossary` 都遵循的 `caseSensitive` 标记。
 - 语言代码会规范化（`EN` → `en`，`ZH_cn` → `zh-CN`）。
+- 任何库（无论是否挂载）里的条目都能编辑或删除；挂载状态只决定工作台搜索
+  和写入哪些库。
+- **术语提取**：术语库工具栏的「从文档提取术语」会列出候选词，带出现次数、
+  记忆库里已有的译法和「已有」标记。统计负责找候选并排序（拉丁 n-gram；
+  中日韩重复子串用左右邻接熵过滤），模型只判断一份短名单——这正是模型擅长
+  的部分。
+
+### 库页面
+
+顶栏的「文档 / 库 / 对齐」在三个工作区之间切换，库页面不需要打开文档。
+
+- TM 一张表、术语一张表，支持分页（20/50/100/200，会记住）、语言筛选，以及
+  可限定在任意若干库内的搜索。
+- 右侧边栏直接编辑选中的条目、新增条目，并在有未保存改动时提示。
+- TM 导出 **TMX**；术语库导出 **TBX**、**CSV** 或 **JSON**。导入与清空仍在同一
+  个工具栏里。
+- 右侧面板的库按钮会把面板当前的搜索词一起带进库页面，在面板里查到的东西
+  可以接着在完整列表里看。
+
+### 对齐
+
+顶栏的「文档 / 库 / 对齐」第三个视图用两份成品文件建记忆库：源文一份、
+译文一份，可以是完全不同的语言，格式任选。
+
+- 跨语言打分：共享锚点（数字、`{0}` 占位符、URL、首字母大写的名字）、按语言
+  归一化的长度比例、段落结构。刻意不用字面相似度，因为它比不了两种语言。
+- 分块动态规划可以配对 1:1、1:0、0:1、2:1、1:2；每行分高置信 / 待确认 / 可疑
+  三档，只有高置信默认勾选。
+- 写入前每行都能与下一行合并、按换行拆开或删除；目标库在工具栏里选，完全
+  相同的条目会被跳过而不是重复写入。
 
 ### 工作台
 
 - 按段落/句子切分，并逐句编辑译文。
-- 支持保存草稿、确认并下一句、上一句/下一句导航和进度。
+- 句段三态：**待译** → **待审** → **已确认**，同时记录每句译文的来源
+  （`tm`、`mt`、`agent`、`manual`）。句段栏可按状态筛选，搜索覆盖原文、
+  译文和建议。
+- **句段分页**：句段栏按 20/50/100/200 的页大小分页，并有一个输入框可直接
+  跳到某个编号。如果目标句段被当前筛选隐藏，跳转时会先清空筛选。
+- **预翻译** 用 TM 精确命中加机器翻译把整篇铺底：按小块推进，显示
+  `已完成/总数`，可在块与块之间停止；已确认或已有译文的句段永远不会被覆盖。
+- **上下排列的审校区**：原文、机器/agent 基线、可编辑译文从上到下依次排列，
+  每一块都占满面板宽度。**全部确认** 在一次确认弹窗后接受所有待审句段。
+- 支持保存草稿、确认并下一句、上一句/下一句导航和进度。当前句段与草稿会
+  防抖落盘，重启后能回到原位。
+- **重复句段自动同步**：确认一句后，本文档里其它相同原文的未确认句段会填上
+  同一译文（来源 `auto`），编辑草稿也会同步；「铺满重复句段」可一次性处理整篇。
+  已确认的句段永远不动。
+- 句段栏里的 QA 提醒与导出前的汇总：占位符/标签缺失、数字不一致、
+  术语未使用、空译文、译文仍是原文、同一原文多种译法。只提醒，不阻挡导出，
+  也不写盘。
 - 文档标签页；关闭标签页不会删除文档。
 - **打开已有文档** 可重新打开已关闭的文档。
 - **设置 → 界面** 提供界面语言切换：跟随 DSH 语言（默认）、简体中文、English、
@@ -91,9 +141,11 @@ CAIT 为 DSH Web UI 和 DSH 桌面端增加一个 CAT 翻译工作台。它从
   不会影响 TM、术语库或其他文档。
 - 同格式、同内容的再次导入会复用已有文档，而不是新建重复文档。
   指纹覆盖语言、源文、格式和交换元数据。
-- 响应式密度：工作台测量自身 DSH 面板，并应用 1×–1.8× 的 `--cait-u`
-  缩放系数。1600×900 面板约为 1.32×，2560×1440 面板为 1.8×；
-  窗口较矮时保持 1×，不会继续缩小。
+- 响应式密度：工作台测量自身 DSH 面板，并应用经过衰减的 1×–1.45× `--cait-u`
+  缩放系数。1600×900 面板约为 1.17×，最大化的 4K 面板停在 1.45×：
+  大窗口贴近 DSH 自带界面，不再比宿主明显放大。
+- 工作台跟随 DSH 的亮/暗主题，包括原生控件（`color-scheme`），
+  下拉框、复选框和滚动条都与面板一致。
 
 ### 预处理
 
@@ -128,7 +180,7 @@ provider/model/code/status，客户端只会收到 `MT_PROVIDER_FAILED`。
 | CSV | ✅ | ✅ | 自动识别逗号、分号、制表符或竖线分隔；可通过表头名称或从 1 开始的列号指定源/目标列；现有双语导出 |
 | DOCX | ✅ | ✅ | 正文段落；导入时排除表格块。导出会修改原包，保留样式、编号、图片、页眉、表格和版式。 |
 | EPUB | ✅ | ✅ | 按 EPUB 3 spine 顺序读取 XHTML 文本块，内联标记保留为 token 并回写到保留的源包。不支持 DRM、固定版式和脚本化 EPUB。 |
-| TMX 1.4 | ✅ | ✅ | 常见 `<tu>/<tuv>/<seg>` 子集；内联标记扁平化并警告 |
+| TMX 1.4 | ✅ | ✅ | 常见 `<tu>/<tuv>/<seg>` 子集；内联标记扁平化并警告；上下文用 `<prop type="x-context">` 往返 |
 | SRT / WebVTT | ✅ | ✅ | 一个字幕 cue 对应一个句段；保留时间轴/设置/标识；保留语义标签；非法时间范围会拒绝 |
 | PO | ✅ | ✅ | 支持上下文、复数形式、译者注释和废弃 `#~` 条目 |
 | XLIFF 1.2 | ✅ | ✅ | 内联标签转换为 `{n}` token；缺少 token 时阻止确认 |
@@ -178,7 +230,7 @@ CAIT 以 DSH bundle 的形式安装；发布的 npm 包名为 `dsh-cait`，插
 plugin_manager install_bundle target: dsh-cait
 ```
 
-需要固定版本时写 `dsh-cait@1.0.5`，需要最新版时写 `dsh-cait@latest`。管理器会解析 registry
+需要固定版本时写 `dsh-cait@1.1.3`，需要最新版时写 `dsh-cait@latest`。管理器会解析 registry
 上的包、应用 bundle patch 并激活插件，不需要手工在配置档里加行。
 
 使用 CLI 管理的配置档：
@@ -255,7 +307,8 @@ TM 返回条数、MT 路由/模型/温度，以及界面语言（`uiLang`：`aut
 
 如果快照损坏或版本未知，CAIT 会记录警告，将其移动到
 `store.json.corrupt-<timestamp>`，然后以空状态启动，而不是让插件激活失败。
-版本 1 快照会迁移到版本 2，并创建默认 TM 和术语库。
+版本 1、2 快照会迁移到版本 3，并创建默认 TM 和术语库；版本 2 里「有译文但未
+确认」的草稿会变成 `translated`，来源记为 `manual`。
 
 ## Agent 工具
 
@@ -268,6 +321,19 @@ TM 返回条数、MT 路由/模型/温度，以及界面语言（`uiLang`：`aut
 | `cat_list_documents` | 列出文档和进度 |
 | `cat_switch_document` | 选择活动文档 |
 | `cat_suggest_mt` | 通过所选 DSH 路由调用 MT，或使用明确的离线回退 |
+| `cat_pre_translate` | 用 TM + MT 填一块，跳过已确认/已有译文的句段，返回 `next` |
+| `cat_get_segments` | 取一段有界句段窗口，含状态、来源和建议 |
+| `cat_confirm_segments` | 确认索引列表、一段窗口，或全部待审句段 |
+| `cat_propose_translation` | 只留下建议和理由，不动已确认的译文 |
+| `cat_align` | 对齐两段文本，报告高置信 / 待确认 / 可疑的配对 |
+| `cat_consistency_report` | 按原文分组列出译文不一致的句段 |
+| `cat_term_candidates` | 术语候选短名单：频次、例句、记忆库里已有的译法 |
+| `cat_add_terms` | 一次性把选中的术语写进术语库 |
+
+因此 Agent 可以完全不点界面就把整篇铺底：反复调用 `cat_pre_translate`
+（`from: next`）直到 `next` 为 `null`，用 `cat_get_segments` 查看，再用
+`cat_confirm_segments` 批量接受。做术语时读 `cat_term_candidates`，判断这份
+短名单，再用 `cat_add_terms` 写入。
 
 工作台和工具共用一个 `CatStore`，因此 Agent 可以翻译用户当前打开的同一份文档。
 
@@ -277,7 +343,7 @@ TM 返回条数、MT 路由/模型/温度，以及界面语言（`uiLang`：`aut
 flowchart LR
   UI["DSH Web UI<br/>(侧栏 + 工作台)"] --> Client["client.js<br/>生成的 Web bundle"]
   Client --> API["/cait/api/rpc<br/>同源 JSON"]
-  Tools["7 个 Agent 工具"] --> Store["CatStore"]
+  Tools["15 个 Agent 工具"] --> Store["CatStore"]
   API --> Store
   Store --> Snapshot[("store.json<br/>+ sources/")]
 ```
@@ -411,18 +477,25 @@ GitHub release 发布。
 - XLSX 导入/导出是两列表格；不保留单元格样式、公式、合并单元格、图片和额外工作表。
 - HTML/RTF 导出不往返原始样式；JSON 只翻译字符串值。
 - TM 匹配是词法级的，不是语义级；没有词干、同义词或向量。
-- 库重命名/删除和逐标签页挂载状态尚未实现。
+- 每个上下文下，同一原文存一条译文：重复的句子会为每种相邻句组合保留一条变体，
+  每个 key 上限 8 条；相邻句不同时回落到无上下文的通用条目。
+- 上下文只取前后两句源文，不包含文档、客户或项目元数据。
+- TBX 与 JSON 术语导出是单向的；只有 TMX 和 CSV 能再导入回来。
 - 工作台界面提供中文、English、日本語三种语言，可在设置中切换；README 提供英文、简体中文和日文三个版本，以英文版为准。
 
-后续计划：通过多模态 DSH 路由支持图片/视觉输入、更完整的库管理、
-保留样式的 XLSX 输出，以及 XLIFF `<alt-trans>` / 注释支持。
+后续计划：通过多模态 DSH 路由支持图片/视觉输入、保留样式的 XLSX 输出，
+以及 XLIFF `<alt-trans>` / 注释支持。
 
 ## 变更日志与版本
 
 见 [CHANGELOG.md](CHANGELOG.md)。`v1.0.0` 是首个公开发布版本，`v1.0.1` 增加了
 日语界面和界面内语言切换，`v1.0.2` 让「跟随 DSH 语言」真正跟随 DSH 的语言设置，
 `v1.0.3` 把包发布到 npm，`v1.0.4` 起后续版本改由带 provenance 的发布流水线产出，
-`v1.0.5` 增加 EPUB、导入预处理和 PDF 政策；项目在开发阶段使用 `0.1.0`，
+`v1.0.5` 增加 EPUB、导入预处理和 PDF 政策，`v1.1.0` 增加 agent 驱动的预翻译、
+句段三态、上下排列的审校区和轻量 QA，`v1.1.1` 增加带 TMX/TBX/CSV/JSON 导出的
+库页面、句段分页，以及从右侧面板跳进库页面，`v1.1.2` 增加带 101%（ICE）
+匹配的上下文记忆，`v1.1.3` 增加双文件对齐、重复句段自动同步、术语提取，
+以及 agent 写建议用的原语；项目在开发阶段使用 `0.1.0`，
 发布到 GitHub 前统一改为 `1.0.0`。
 
 > **License：** Apache License 2.0，见 [LICENSE](LICENSE)。再分发时需保留