dsh-cait 1.0.3 → 1.0.5

package/CHANGELOG.md

@@ -1,352 +1,410 @@
-# Changelog
-
-All notable changes to CAIT are documented here. The project follows
-[Semantic Versioning](https://semver.org/) and the format is based on
-[Keep a Changelog](https://keepachangelog.com/).
-
-## [1.0.3] - 2026-11-05
-
-npm distribution.
-
-### Added
-
-- The package can now be installed from the npm registry:
-  `plugin_manager install_bundle target: dsh-cait` in DSH Desktop, or
-  `dsh plugin --profile <name> add dsh-cait`. Pin a release with
-  `dsh-cait@1.0.3`; the GitHub and local-checkout routes still work.
-- `engines.dsh` (`>=0.2.2-rc.1`), so the plugin manager refuses an older host
-  with a compatibility report instead of a broken mount.
-- `publishConfig.access: public`, an `author` field, and the `dsh-plugin` /
-  `cordis` keywords for the registry listing.
-- `prepublishOnly` rebuilds the client bundle and syntax-checks it, so a
-  release cannot ship a stale `client.js`.
-- `.github/workflows/publish.yml` publishes a GitHub release through npm
-  Trusted Publishing with `--provenance`; it checks the tag against
-  `package.json`, runs the whole test suite, and never stores an npm token.
-- `@deepseek-ai/schemastery` is a `devDependency` as well as a peer dependency,
-  so a fresh checkout can run the tests.
-
-### Changed
-
-- `README.ja.md` ships in the package (`files`).
-- The README install section leads with the npm route and keeps the GitHub and
-  local-checkout routes as alternatives.
-- The repository no longer tracks `node_modules/`; packed tarballs are ignored.
-
-### Removed
-
-- `"private": true`, which blocked `npm publish`.
-
-## [1.0.2] - 2026-11-04
-
-Interface-language fix.
-
-### Added
-
-- Japanese README (`README.ja.md`), linked from the language switcher of all
-  three READMEs and listed in the project layout.
-
-### Fixed
-
-- **Follow DSH language** now follows the DSH locale instead of staying on
-  English. The client plugin declared only `slots` in its inject list, so
-  Cordis rejected the `ctx.locale` access with a throw that the registration
-  `try`/`catch` swallowed: the DSH-bound translator stayed `null`, every lookup
-  fell through to the English dictionary, and a Simplified-Chinese DSH looked
-  English. `locale` is now a declared dependency.
-- `auto` resolves the locale service snapshot itself (`zh-CN` → `zh`,
-  `ja-JP` → `ja`, `en-US` → `en`) before consulting the DSH-bound translator,
-  so a locale tag without a registered fallback chain no longer degrades to
-  English. The export `resolveUiLang` covers this mapping.
-- The workbench and the sidebar label subscribe to DSH locale changes, so
-  switching the DSH language while `auto` is selected re-renders CAIT
-  immediately instead of requiring a reload.
-
-## [1.0.1] - 2026-11-04
-
-UI languages and license metadata.
-
-### Added
-
-- Japanese (`ja`) UI dictionary. The workbench now ships three interface
-  languages: Simplified Chinese, English, and Japanese.
-- **Settings → Interface → Interface language**, with *Follow DSH language*
-  (default), 简体中文, English, and 日本語. The choice is stored as
-  `settings.uiLang` in the CAIT store, applies to the workbench and the sidebar
-  label immediately, and only affects the CAIT interface — document source and
-  target languages are untouched.
-- `NOTICE`, the attribution supplement for Apache-2.0 (section 4(d)), and an
-  explicit `license` field in `package.json`.
-
-### Changed
-
-- `LICENSE` appendix copyright line now reads `Copyright 2026 Jovi Lecca`.
-- Version bumped to `1.0.1` (`package.json`, `CAIT_VERSION`, and the exported
-  document metadata move together).
-- Verified against DSH `0.2.2-rc.1`; the README badges and development notes use
-  that version.
-
-### Fixed
-
-- `saveSettings` rejects an unsupported interface language with
-  `INVALID_UI_LANG` instead of silently storing it.
-
-## [1.0.0] - 2026-11-04
-
-First public release. CAIT was developed as `0.1.0` and renumbered to `1.0.0`
-for the first GitHub release.
-
-The npm package and Cordis entry are named `dsh-cait`; the UI entry, data
-directory (`$DSH_HOME/cait`), and API route (`/cait/api`) keep the CAIT name.
-
-### Highlights
-
-- A complete CAT workbench inside DeepSeek Harness: sidebar entry, main panel,
-  segment list/editor, TM/glossary/MT side panel, and export flow.
-- Seven agent tools that share the exact store used by the UI.
-- Multiple TM and glossary libraries with mounted-only lookup and default write
-  targets.
-- Format-preserving DOCX and PPTX export, plus a broad text, subtitle,
-  localization, and Office format matrix.
-- Persistence with atomic snapshots, corrupt-file recovery, strict confirm
-  writes, and revision polling between agents and the open workbench.
-
-### Added
-
-#### Plugin and UI
-
-- Host + Client bundle structure (`index.js`, `client.js`, `cordis.patch.yml`,
-  `package.json`, locale metadata, icon).
-- New two-bubble A/文 icon: a bilingual speech-bubble pair that reads as
-  source/target, used both for the package icon and the DSH sidebar mark.
-- Left-sidebar entry `cait` and main panel `cait`; all styling uses DSH
-  `--dsw-alias-*` theme tokens.
-- English and Chinese UI dictionaries; plugin metadata locales.
-- Workbench with document tabs, segment rail, source highlighting, glossary
-  chips, target editor, confirm/next/previous/draft actions, progress, and
-  responsive density scaling from 1× to 1.8×.
-- Document tabs with close, reopen, new-page import, duplicate-import reuse,
-  and a recent-document cleanup dialog for one or all closed documents.
-- Settings modal for source/target languages, fuzzy threshold, TM result limit,
-  MT route/model/temperature, and library mounts.
-- Lightweight **库 / Libraries** modal to create, mount, and choose default TM
-  and glossary libraries.
-- Visual save-failure/stale banner and a two-second revision poll so
-  agent-triggered document switches and confirmations appear in an open
-  workbench without discarding a dirty draft.
-
-#### Translation memory and glossary
-
-- Multiple TM libraries with id, name, mount state, default write library, and
-  per-library deduplication.
-- Graded TM matching: `exact`, independent `near_exact` (ignores digits,
-  punctuation, and whitespace), `high_fuzzy >= 0.85`,
-  `medium_fuzzy 0.75–0.85`, `low_fuzzy < 0.75` above the configurable floor.
-- Lexical fuzzy scoring: character-bigram Dice (72%) + token Jaccard (28%);
-  Latin words and CJK characters are tokenized separately.
-- Multiple glossary libraries, source-language filtering, live highlighting,
-  click-to-insert terms, case-sensitive terms, CSV/TSV import, and notes.
-- Mounted-only lookups in the UI and tools; unmounted write targets are rejected
-  with `TM_NOT_MOUNTED` / `GLOSSARY_NOT_MOUNTED`.
-- Clear operations default to mounted libraries, accept explicit mounted ids, and
-  require `all: true` plus `confirm: true` to clear every library.
-- Canonical language tags via `Intl.getCanonicalLocales` (`EN` → `en`,
-  `ZH_cn` → `zh-CN`).
-
-#### Agent tools
-
-- `cat_get_segment` — current segment, TM matches, glossary hits.
-- `cat_submit_translation` — confirm target, optional mounted `targetTmId`,
-  write to TM, advance.
-- `cat_lookup_tm` — mounted-library graded lookup.
-- `cat_lookup_glossary` — mounted glossary query with language/case handling.
-- `cat_list_documents` — document list and progress.
-- `cat_switch_document` — switch active document.
-- `cat_suggest_mt` — MT through the selected DSH route, or explicit offline
-  glossary-substitution fallback.
-
-#### MT
-
-- MT provider-route selector built from `ctx.llm.listProviders()` and
-  `ctx.llm.listModels()`, optionally filtered by `ctx.loader.entries()`.
-- Calls through `ctx.llm.stream()`; concatenates `text-delta` chunks only.
-- No CAIT-managed `baseUrl` or API key; DSH resolves credentials.
-- Explicit offline fallback labelled **术语替换（离线） / Glossary substitution
-  (offline)** (internal id `mock`) when no route/model is available.
-- Host-only detailed logging of stream failures; clients receive the generic
-  `MT_PROVIDER_FAILED`.
-
-#### Persistence and configuration
-
-- Schemastery `dataDir` config; default `$DSH_HOME/cait` or `~/.dsh/cait`.
-- Versioned JSON snapshot (`version: 2`) containing documents, segments, open-tab
-  ids, TM, glossary, libraries, mounts, and non-secret settings.
-- Atomic temp-file + rename writes; strict confirm writes that roll back on disk
-  failure and return `PERSISTENCE_FAILED` (503); best-effort structural writes
-  with a visible storage-error banner.
-- Corrupt/unknown-version snapshots are backed up as
-  `store.json.corrupt-<timestamp>` and replaced with empty state.
-- v1 → v2 migration to default TM/glossary libraries.
-- DOCX and PPTX source sidecars under `sources/<document-id>.*` for
-  format-preserving export; sidecars are removed when the document is deleted.
-- Open tabs and `activeDocId` persist; draft text and per-document cursor do not.
-- Duplicate import guard based on SHA-256 of languages, source, format, and
-  exchange metadata.
-
-#### Formats
-
-- TXT, Markdown, DOCX body text, TMX 1.4, CSV, SRT, WebVTT, PO, XLIFF 1.2,
-  basic XLIFF 2.0, basic SDLXLIFF, HTML, JSON string paths, XLSX, PPTX, and RTF
-  import.
-- Target-only TXT/Markdown/HTML/RTF/SRT/WebVTT export; bilingual
-  Markdown/CSV/TMX; document exports for PO, XLIFF, XLSX, and PPTX;
-  current-document JSON; structure-preserving JSON; DOCX source-package export.
-- CSV column mapping by header name or 1-based number.
-- SRT/VTT cue timing/settings/identifier metadata; semantic inline tags retained;
-  invalid ranges rejected; untimed cues follow the previous cue end + 2 s.
-- PO contexts, plural forms, translator comments, and obsolete `#~` entries.
-- XLIFF inline placeholders converted to `{n}` tokens and expanded back on
-  export; confirmation rejects dropped tokens with `PLACEHOLDER_MISMATCH`.
-- DOCX export patches translated text into the original package, preserving
-  paragraph/run properties, styles, numbering, images, headers, tables, and
-  layout.
-- PPTX export patches translated body text into the retained source package,
-  preserving slides, masters, shapes, images, and layout.
-- XLSX two-column Source/Target workbook import/export.
-- JSON original-template preservation through `formatMeta.jsonTemplate`.
-
-#### API and security
-
-- Same-origin JSON bridge at `GET /cait/api/health` and
-  `POST /cait/api/rpc`.
-- `ctx.connection.requestRejection` browser/Origin check.
-- 32 MB body limit (`413 BODY_TOO_LARGE`), JSON content-type check, and
-  allow-listed handlers.
-- `clearTM` / `clearGlossary` require `confirm: true` (`409
-  CONFIRMATION_REQUIRED` otherwise).
-- No API key in the browser, snapshot, RPC response, or logs.
-
-### Changed
-
-- Renamed the UI fallback from “mock” to **术语替换（离线） / Glossary
-  substitution (offline)** while keeping internal ids compatible.
-- Replaced CAIT-managed OpenAI-compatible HTTP calls with DSH provider routes.
-- Split the 1,718-line `lib/store.js` into a `CatStore` facade plus
-  `libraries.js`, `tm.js`, `glossary.js`, `persistence.js`,
-  `import-export.js`, `store-utils.js`, and `errors.js`.
-- Split the Web UI into `client/helpers.js`, `rpc.js`, `styles.js`,
-  `components.js`, `locales.js`, and `entry.js`; `client.js` is now generated by
-  `scripts/build-client.mjs`.
-- JSON export was renamed from a purported “project backup” to
-  `*.document.json` (current document only); structured JSON export is a
-  separate `json-tree` format.
-- Subtitle export now renumbers cues in file order and maintains a monotonic
-  cursor for untimed cues instead of overlapping `index * 2s`.
-- Removed the never-incremented `usageCount` TM metric.
-- `cat_submit_translation` now returns `tm.count` as the mounted count and
-  `tm.total` as the all-library count.
-- Open documents persist `openDocIds` and `activeDocId`; settings persist, but
-  drafts, cursor, and library-modal UI state do not.
-- API key and endpoint fields were removed from settings; `publicSettings()`
-  exposes only route/model/temperature.
-- README rewritten for GitHub; this changelog added.
-
-### Fixed
-
-- Mount checkbox was disabled for every unmounted library when only one was
-  mounted; unmounted libraries are now selectable and only the last mounted one
-  is protected.
-- `loadDemo()` TM entries were not persisted; demo state now survives reload.
-- Save errors were swallowed and confirmations could report success without a
-  disk write; strict confirm + rollback + banner now surface the failure.
-- Agent-driven document switches/confirmations were invisible in an open
-  workbench; revision polling and stale/draft handling fixed this.
-- DOCX table paragraphs were flattened into the body text despite the docs; table
-  blocks are now excluded during import.
-- TMX inline `bpt/ept/ph` markup was silently dropped; imports now return a
-  `TMX_INLINE_MARKUP_LOST` warning with a count.
-- Glossary language matching was case-sensitive in hits but case-insensitive in
-  the tool; both paths now canonicalize.
-- XLIFF placeholders, PO plurals/comments/obsolete blocks, and subtitle semantic
-  tags were dropped in earlier adapters; all are preserved now.
-- Subtitle zero/negative-duration cues were silently rewritten; they now fail
-  with an explicit parse error.
-- Duplicate imports always created a new document; identical content now reuses
-  the existing document.
-- Closed documents could not be cleaned up; the recent-documents dialog can now
-  delete one or all closed documents.
-- `cat_lookup_glossary` parameter text still claimed case-insensitive matching
-  after case-sensitive support was added; corrected.
-- PPTX export no longer silently emits an unformatted deck when the source
-  package is missing; it returns `PPTX_SOURCE_REQUIRED`.
-- JSON `formatMeta.jsonTemplate`, XLSX metadata, and PPTX source sidecars were
-  not preserved through reload in the first integration pass; fixed and tested.
-- Save-label text was invisible in the settings modal because the generic button
-  reset overrode `.cait-btn-primary`; fixed and covered by the mount test.
-
-### Security
-
-- Removed the CAIT-managed `baseUrl` and API key path, eliminating the SSRF /
-  keyless-regex / remote-error-body findings from the audit.
-- MT failures expose only `MT_PROVIDER_FAILED`; provider/model/code/status stay in
-  the Host log.
-- All RPC calls retain the DSH Origin/browser-trust check; clear methods require
-  explicit confirmation.
-- Snapshot excludes API keys by design; MT credentials remain inside DSH.
-
-### Tests
-
-- 13 suites: `test-core`, `test-imports`, `test-formats`,
-  `test-second-tier`, `test-docx-export`, `test-api`, `test-plugin`,
-  `test-persistence`, `test-documents`, `test-libraries`, `test-mounts`,
-  `test-ui-scale`, and `test-mt`.
-- Coverage includes TM grading and boundaries, multi-library mounts,
-  mounted-only lookup, unmounted-target rejection, clear scope/confirmation,
-  corruption fallback, v1 migration, strict persistence rollback, demo reload,
-  document tabs/dedupe/delete, DOCX/XLSX/PPTX source preservation, CSV mapping,
-  SRT/VTT timelines, PO plurals/comments/obsolete, XLIFF 1.2/2.0 placeholders,
-  second-tier round-trips, API 413/403/404/409/415, and UI helper logic.
-- Fixtures: `sample.docx`, `sample-table.docx`, `sample-format.docx` with
-  `make-*.py` regeneration scripts.
-
-### Documentation
-
-- Replaced the long development README with a GitHub-oriented README: badges,
-  quick start, feature overview, format matrix, install paths, configuration,
-  tools, architecture, development, security, and limits.
-- Added this changelog and set `package.json` version to `1.0.0`.
-- Added a complete Simplified Chinese README (`README.zh-CN.md`) and a language
-  switcher; the English README now contains English-only prose and UI labels.
-
-### Known limitations
-
-- No PDF, EPUB, or image/OCR import yet; no PDF layout-preserving export.
-- DOCX: tables, headers/footers, footnotes, fields, alt text, and tracked
-  deletions are not imported; mixed-run translation inherits the first run style.
-- PPTX: body text only; notes, charts, SmartArt, and fields are not imported.
-- XLSX: two-column table only; styles, formulas, merged cells, images, and extra
-  sheets are not preserved.
-- HTML/RTF export does not round-trip original styling; JSON only translates
-  string values.
-- TM matching is lexical, not semantic.
-- Library rename/delete and per-tab mount state are not part of v1.0.0.
-- Persistence is one snapshot with last-write-wins semantics, not a multi-user
-  journal.
-- No browser pixel/click test is included; the generated client is validated by
-  `node --check`, helper tests, and plugin slot activation.
-
-## Development history
-
-| Date | Milestone |
-| --- | --- |
-| 2026-11-03 | Plugin skeleton, Host/Client split, CAT store, segmentation, TM, glossary, three tools, workbench UI, TXT/MD/DOCX/TMX import/export, mock MT. |
-| 2026-11-03 | Persistence layer (`dataDir`, v1 snapshot, atomic writes, corrupt fallback), `tmId`/`glossaryId` reservations, reload/corruption tests. |
-| 2026-11-03 | Graded `matchType` (`exact`/`near_exact`/three fuzzy bands) with locale labels and boundary tests. |
-| 2026-11-03 | Four additional agent tools (`cat_lookup_glossary`, `cat_list_documents`, `cat_switch_document`, `cat_suggest_mt`). |
-| 2026-11-03 | Multi-library TM/glossary mounts, `targetTmId`, v1→v2 migration, library modal, mounted-only semantics. |
-| 2026-11-03 | Read-only audit and hardening: mount checkbox, MT route reuse, clear confirmation, mounted write boundaries, strict persistence, revision polling, language/metadata fixes, store/client split. |
-| 2026-11-04 | Document tabs, close/reopen, duplicate fingerprint, recent-document cleanup. |
-| 2026-11-04 | Source-format-preserving DOCX export with sidecar storage and fixtures. |
-| 2026-11-04 | First-tier exchange formats: CSV mapping, SRT, WebVTT, PO, XLIFF 1.2, metadata persistence, timeline fixes. |
-| 2026-11-04 | Second-tier formats: HTML, JSON, XLSX, PPTX, RTF; binary upload/export; source-preserving PPTX. |
-| 2026-11-04 | Responsive UI density 1×–1.8×; button contrast fix; v1.0.0 release. |
+# Changelog
+
+All notable changes to CAIT are documented here. The project follows
+[Semantic Versioning](https://semver.org/) and the format is based on
+[Keep a Changelog](https://keepachangelog.com/).
+
+## [1.0.5] - 2026-11-06
+
+EPUB support, an import preprocessing stage, and format hardening.
+
+### Added
+
+- EPUB 3 import and export. Import walks the OPF spine, turns XHTML text blocks
+  into segments, and keeps inline tags as placeholder tokens; export patches the
+  translations back into the retained source package, preserving chapters,
+  styles, images, and navigation. Documents created from plain text export as a
+  fresh EPUB 3 package. DRM, fixed-layout, and scripted EPUBs are out of scope.
+- Import preprocessing: whitespace normalization, invisible/control character
+  stripping, and blank-line collapsing before segmentation. Each step has a
+  settings toggle, applies to file imports and `createDocument`, and is stored
+  with the other settings.
+- A shared HTML/XHTML run scanner (`lib/html.js`) used by both HTML and EPUB, so
+  headings and inline markup survive import and are patched back on export.
+- CSV delimiter detection (comma, semicolon, tab, pipe) and richer RTF escapes
+  (`\cell`, `\row`, `\page`, `\sect`, symbol characters).
+
+### Changed
+
+- PDF import is explicitly out of scope. All three READMEs explain why PDF is a
+  poor fit for AI-assisted translation and recommend converting it with pandoc,
+  Acrobat, or LibreOffice first.
+- The supported-formats table, project layout, roadmap, and “Known limits”
+  sections now describe EPUB, preprocessing, and the PDF policy.
+
+### Fixed
+
+- Settings persistence now stores `uiLang` and the preprocessing flags instead of
+  falling back to defaults on reload.
+- The EPUB `mimetype` entry is written uncompressed, as the OCF specification
+  requires, and every other entry stays deflated.
+- HTML import no longer flattens headings and inline markup into blank-line
+  paragraphs.
+
+### Development
+
+- New `_dev/test-epub.mjs` and `_dev/test-preprocess.mjs` suites; the full suite
+  count is 15.
+
+## [1.0.4] - 2026-11-05
+
+Provenance publishing.
+
+### Added
+
+- `README.md`, `README.zh-CN.md` and `README.ja.md` carry the live npm badge, so
+  the package page and the repository show the published version.
+
+### Changed
+
+- Releases from this version on are produced by `.github/workflows/publish.yml`,
+  which attaches npm provenance and validates the GitHub trusted publisher.
+  `npm publish` from a local checkout still works and still asks for an OTP.
+
+## [1.0.3] - 2026-11-05
+
+npm distribution.
+
+### Added
+
+- The package can now be installed from the npm registry:
+  `plugin_manager install_bundle target: dsh-cait` in DSH Desktop, or
+  `dsh plugin --profile <name> add dsh-cait`. Pin a release with
+  `dsh-cait@1.0.3`; the GitHub and local-checkout routes still work.
+- `engines.dsh` (`>=0.2.2-rc.1`), so the plugin manager refuses an older host
+  with a compatibility report instead of a broken mount.
+- `publishConfig.access: public`, an `author` field, and the `dsh-plugin` /
+  `cordis` keywords for the registry listing.
+- `prepublishOnly` rebuilds the client bundle and syntax-checks it, so a
+  release cannot ship a stale `client.js`.
+- `.github/workflows/publish.yml` publishes a GitHub release through npm
+  Trusted Publishing with `--provenance`; it checks the tag against
+  `package.json`, runs the whole test suite, and never stores an npm token.
+- `@deepseek-ai/schemastery` is a `devDependency` as well as a peer dependency,
+  so a fresh checkout can run the tests.
+
+### Changed
+
+- `README.ja.md` ships in the package (`files`).
+- The README install section leads with the npm route and keeps the GitHub and
+  local-checkout routes as alternatives.
+- The repository no longer tracks `node_modules/`; packed tarballs are ignored.
+
+### Removed
+
+- `"private": true`, which blocked `npm publish`.
+
+## [1.0.2] - 2026-11-04
+
+Interface-language fix.
+
+### Added
+
+- Japanese README (`README.ja.md`), linked from the language switcher of all
+  three READMEs and listed in the project layout.
+
+### Fixed
+
+- **Follow DSH language** now follows the DSH locale instead of staying on
+  English. The client plugin declared only `slots` in its inject list, so
+  Cordis rejected the `ctx.locale` access with a throw that the registration
+  `try`/`catch` swallowed: the DSH-bound translator stayed `null`, every lookup
+  fell through to the English dictionary, and a Simplified-Chinese DSH looked
+  English. `locale` is now a declared dependency.
+- `auto` resolves the locale service snapshot itself (`zh-CN` → `zh`,
+  `ja-JP` → `ja`, `en-US` → `en`) before consulting the DSH-bound translator,
+  so a locale tag without a registered fallback chain no longer degrades to
+  English. The export `resolveUiLang` covers this mapping.
+- The workbench and the sidebar label subscribe to DSH locale changes, so
+  switching the DSH language while `auto` is selected re-renders CAIT
+  immediately instead of requiring a reload.
+
+## [1.0.1] - 2026-11-04
+
+UI languages and license metadata.
+
+### Added
+
+- Japanese (`ja`) UI dictionary. The workbench now ships three interface
+  languages: Simplified Chinese, English, and Japanese.
+- **Settings → Interface → Interface language**, with *Follow DSH language*
+  (default), 简体中文, English, and 日本語. The choice is stored as
+  `settings.uiLang` in the CAIT store, applies to the workbench and the sidebar
+  label immediately, and only affects the CAIT interface — document source and
+  target languages are untouched.
+- `NOTICE`, the attribution supplement for Apache-2.0 (section 4(d)), and an
+  explicit `license` field in `package.json`.
+
+### Changed
+
+- `LICENSE` appendix copyright line now reads `Copyright 2026 Jovi Lecca`.
+- Version bumped to `1.0.1` (`package.json`, `CAIT_VERSION`, and the exported
+  document metadata move together).
+- Verified against DSH `0.2.2-rc.1`; the README badges and development notes use
+  that version.
+
+### Fixed
+
+- `saveSettings` rejects an unsupported interface language with
+  `INVALID_UI_LANG` instead of silently storing it.
+
+## [1.0.0] - 2026-11-04
+
+First public release. CAIT was developed as `0.1.0` and renumbered to `1.0.0`
+for the first GitHub release.
+
+The npm package and Cordis entry are named `dsh-cait`; the UI entry, data
+directory (`$DSH_HOME/cait`), and API route (`/cait/api`) keep the CAIT name.
+
+### Highlights
+
+- A complete CAT workbench inside DeepSeek Harness: sidebar entry, main panel,
+  segment list/editor, TM/glossary/MT side panel, and export flow.
+- Seven agent tools that share the exact store used by the UI.
+- Multiple TM and glossary libraries with mounted-only lookup and default write
+  targets.
+- Format-preserving DOCX and PPTX export, plus a broad text, subtitle,
+  localization, and Office format matrix.
+- Persistence with atomic snapshots, corrupt-file recovery, strict confirm
+  writes, and revision polling between agents and the open workbench.
+
+### Added
+
+#### Plugin and UI
+
+- Host + Client bundle structure (`index.js`, `client.js`, `cordis.patch.yml`,
+  `package.json`, locale metadata, icon).
+- New two-bubble A/文 icon: a bilingual speech-bubble pair that reads as
+  source/target, used both for the package icon and the DSH sidebar mark.
+- Left-sidebar entry `cait` and main panel `cait`; all styling uses DSH
+  `--dsw-alias-*` theme tokens.
+- English and Chinese UI dictionaries; plugin metadata locales.
+- Workbench with document tabs, segment rail, source highlighting, glossary
+  chips, target editor, confirm/next/previous/draft actions, progress, and
+  responsive density scaling from 1× to 1.8×.
+- Document tabs with close, reopen, new-page import, duplicate-import reuse,
+  and a recent-document cleanup dialog for one or all closed documents.
+- Settings modal for source/target languages, fuzzy threshold, TM result limit,
+  MT route/model/temperature, and library mounts.
+- Lightweight **库 / Libraries** modal to create, mount, and choose default TM
+  and glossary libraries.
+- Visual save-failure/stale banner and a two-second revision poll so
+  agent-triggered document switches and confirmations appear in an open
+  workbench without discarding a dirty draft.
+
+#### Translation memory and glossary
+
+- Multiple TM libraries with id, name, mount state, default write library, and
+  per-library deduplication.
+- Graded TM matching: `exact`, independent `near_exact` (ignores digits,
+  punctuation, and whitespace), `high_fuzzy >= 0.85`,
+  `medium_fuzzy 0.75–0.85`, `low_fuzzy < 0.75` above the configurable floor.
+- Lexical fuzzy scoring: character-bigram Dice (72%) + token Jaccard (28%);
+  Latin words and CJK characters are tokenized separately.
+- Multiple glossary libraries, source-language filtering, live highlighting,
+  click-to-insert terms, case-sensitive terms, CSV/TSV import, and notes.
+- Mounted-only lookups in the UI and tools; unmounted write targets are rejected
+  with `TM_NOT_MOUNTED` / `GLOSSARY_NOT_MOUNTED`.
+- Clear operations default to mounted libraries, accept explicit mounted ids, and
+  require `all: true` plus `confirm: true` to clear every library.
+- Canonical language tags via `Intl.getCanonicalLocales` (`EN` → `en`,
+  `ZH_cn` → `zh-CN`).
+
+#### Agent tools
+
+- `cat_get_segment` — current segment, TM matches, glossary hits.
+- `cat_submit_translation` — confirm target, optional mounted `targetTmId`,
+  write to TM, advance.
+- `cat_lookup_tm` — mounted-library graded lookup.
+- `cat_lookup_glossary` — mounted glossary query with language/case handling.
+- `cat_list_documents` — document list and progress.
+- `cat_switch_document` — switch active document.
+- `cat_suggest_mt` — MT through the selected DSH route, or explicit offline
+  glossary-substitution fallback.
+
+#### MT
+
+- MT provider-route selector built from `ctx.llm.listProviders()` and
+  `ctx.llm.listModels()`, optionally filtered by `ctx.loader.entries()`.
+- Calls through `ctx.llm.stream()`; concatenates `text-delta` chunks only.
+- No CAIT-managed `baseUrl` or API key; DSH resolves credentials.
+- Explicit offline fallback labelled **术语替换（离线） / Glossary substitution
+  (offline)** (internal id `mock`) when no route/model is available.
+- Host-only detailed logging of stream failures; clients receive the generic
+  `MT_PROVIDER_FAILED`.
+
+#### Persistence and configuration
+
+- Schemastery `dataDir` config; default `$DSH_HOME/cait` or `~/.dsh/cait`.
+- Versioned JSON snapshot (`version: 2`) containing documents, segments, open-tab
+  ids, TM, glossary, libraries, mounts, and non-secret settings.
+- Atomic temp-file + rename writes; strict confirm writes that roll back on disk
+  failure and return `PERSISTENCE_FAILED` (503); best-effort structural writes
+  with a visible storage-error banner.
+- Corrupt/unknown-version snapshots are backed up as
+  `store.json.corrupt-<timestamp>` and replaced with empty state.
+- v1 → v2 migration to default TM/glossary libraries.
+- DOCX and PPTX source sidecars under `sources/<document-id>.*` for
+  format-preserving export; sidecars are removed when the document is deleted.
+- Open tabs and `activeDocId` persist; draft text and per-document cursor do not.
+- Duplicate import guard based on SHA-256 of languages, source, format, and
+  exchange metadata.
+
+#### Formats
+
+- TXT, Markdown, DOCX body text, TMX 1.4, CSV, SRT, WebVTT, PO, XLIFF 1.2,
+  basic XLIFF 2.0, basic SDLXLIFF, HTML, JSON string paths, XLSX, PPTX, and RTF
+  import.
+- Target-only TXT/Markdown/HTML/RTF/SRT/WebVTT export; bilingual
+  Markdown/CSV/TMX; document exports for PO, XLIFF, XLSX, and PPTX;
+  current-document JSON; structure-preserving JSON; DOCX source-package export.
+- CSV column mapping by header name or 1-based number.
+- SRT/VTT cue timing/settings/identifier metadata; semantic inline tags retained;
+  invalid ranges rejected; untimed cues follow the previous cue end + 2 s.
+- PO contexts, plural forms, translator comments, and obsolete `#~` entries.
+- XLIFF inline placeholders converted to `{n}` tokens and expanded back on
+  export; confirmation rejects dropped tokens with `PLACEHOLDER_MISMATCH`.
+- DOCX export patches translated text into the original package, preserving
+  paragraph/run properties, styles, numbering, images, headers, tables, and
+  layout.
+- PPTX export patches translated body text into the retained source package,
+  preserving slides, masters, shapes, images, and layout.
+- XLSX two-column Source/Target workbook import/export.
+- JSON original-template preservation through `formatMeta.jsonTemplate`.
+
+#### API and security
+
+- Same-origin JSON bridge at `GET /cait/api/health` and
+  `POST /cait/api/rpc`.
+- `ctx.connection.requestRejection` browser/Origin check.
+- 32 MB body limit (`413 BODY_TOO_LARGE`), JSON content-type check, and
+  allow-listed handlers.
+- `clearTM` / `clearGlossary` require `confirm: true` (`409
+  CONFIRMATION_REQUIRED` otherwise).
+- No API key in the browser, snapshot, RPC response, or logs.
+
+### Changed
+
+- Renamed the UI fallback from “mock” to **术语替换（离线） / Glossary
+  substitution (offline)** while keeping internal ids compatible.
+- Replaced CAIT-managed OpenAI-compatible HTTP calls with DSH provider routes.
+- Split the 1,718-line `lib/store.js` into a `CatStore` facade plus
+  `libraries.js`, `tm.js`, `glossary.js`, `persistence.js`,
+  `import-export.js`, `store-utils.js`, and `errors.js`.
+- Split the Web UI into `client/helpers.js`, `rpc.js`, `styles.js`,
+  `components.js`, `locales.js`, and `entry.js`; `client.js` is now generated by
+  `scripts/build-client.mjs`.
+- JSON export was renamed from a purported “project backup” to
+  `*.document.json` (current document only); structured JSON export is a
+  separate `json-tree` format.
+- Subtitle export now renumbers cues in file order and maintains a monotonic
+  cursor for untimed cues instead of overlapping `index * 2s`.
+- Removed the never-incremented `usageCount` TM metric.
+- `cat_submit_translation` now returns `tm.count` as the mounted count and
+  `tm.total` as the all-library count.
+- Open documents persist `openDocIds` and `activeDocId`; settings persist, but
+  drafts, cursor, and library-modal UI state do not.
+- API key and endpoint fields were removed from settings; `publicSettings()`
+  exposes only route/model/temperature.
+- README rewritten for GitHub; this changelog added.
+
+### Fixed
+
+- Mount checkbox was disabled for every unmounted library when only one was
+  mounted; unmounted libraries are now selectable and only the last mounted one
+  is protected.
+- `loadDemo()` TM entries were not persisted; demo state now survives reload.
+- Save errors were swallowed and confirmations could report success without a
+  disk write; strict confirm + rollback + banner now surface the failure.
+- Agent-driven document switches/confirmations were invisible in an open
+  workbench; revision polling and stale/draft handling fixed this.
+- DOCX table paragraphs were flattened into the body text despite the docs; table
+  blocks are now excluded during import.
+- TMX inline `bpt/ept/ph` markup was silently dropped; imports now return a
+  `TMX_INLINE_MARKUP_LOST` warning with a count.
+- Glossary language matching was case-sensitive in hits but case-insensitive in
+  the tool; both paths now canonicalize.
+- XLIFF placeholders, PO plurals/comments/obsolete blocks, and subtitle semantic
+  tags were dropped in earlier adapters; all are preserved now.
+- Subtitle zero/negative-duration cues were silently rewritten; they now fail
+  with an explicit parse error.
+- Duplicate imports always created a new document; identical content now reuses
+  the existing document.
+- Closed documents could not be cleaned up; the recent-documents dialog can now
+  delete one or all closed documents.
+- `cat_lookup_glossary` parameter text still claimed case-insensitive matching
+  after case-sensitive support was added; corrected.
+- PPTX export no longer silently emits an unformatted deck when the source
+  package is missing; it returns `PPTX_SOURCE_REQUIRED`.
+- JSON `formatMeta.jsonTemplate`, XLSX metadata, and PPTX source sidecars were
+  not preserved through reload in the first integration pass; fixed and tested.
+- Save-label text was invisible in the settings modal because the generic button
+  reset overrode `.cait-btn-primary`; fixed and covered by the mount test.
+
+### Security
+
+- Removed the CAIT-managed `baseUrl` and API key path, eliminating the SSRF /
+  keyless-regex / remote-error-body findings from the audit.
+- MT failures expose only `MT_PROVIDER_FAILED`; provider/model/code/status stay in
+  the Host log.
+- All RPC calls retain the DSH Origin/browser-trust check; clear methods require
+  explicit confirmation.
+- Snapshot excludes API keys by design; MT credentials remain inside DSH.
+
+### Tests
+
+- 13 suites: `test-core`, `test-imports`, `test-formats`,
+  `test-second-tier`, `test-docx-export`, `test-api`, `test-plugin`,
+  `test-persistence`, `test-documents`, `test-libraries`, `test-mounts`,
+  `test-ui-scale`, and `test-mt`.
+- Coverage includes TM grading and boundaries, multi-library mounts,
+  mounted-only lookup, unmounted-target rejection, clear scope/confirmation,
+  corruption fallback, v1 migration, strict persistence rollback, demo reload,
+  document tabs/dedupe/delete, DOCX/XLSX/PPTX source preservation, CSV mapping,
+  SRT/VTT timelines, PO plurals/comments/obsolete, XLIFF 1.2/2.0 placeholders,
+  second-tier round-trips, API 413/403/404/409/415, and UI helper logic.
+- Fixtures: `sample.docx`, `sample-table.docx`, `sample-format.docx` with
+  `make-*.py` regeneration scripts.
+
+### Documentation
+
+- Replaced the long development README with a GitHub-oriented README: badges,
+  quick start, feature overview, format matrix, install paths, configuration,
+  tools, architecture, development, security, and limits.
+- Added this changelog and set `package.json` version to `1.0.0`.
+- Added a complete Simplified Chinese README (`README.zh-CN.md`) and a language
+  switcher; the English README now contains English-only prose and UI labels.
+
+### Known limitations
+
+- No PDF, EPUB, or image/OCR import yet; no PDF layout-preserving export.
+- DOCX: tables, headers/footers, footnotes, fields, alt text, and tracked
+  deletions are not imported; mixed-run translation inherits the first run style.
+- PPTX: body text only; notes, charts, SmartArt, and fields are not imported.
+- XLSX: two-column table only; styles, formulas, merged cells, images, and extra
+  sheets are not preserved.
+- HTML/RTF export does not round-trip original styling; JSON only translates
+  string values.
+- TM matching is lexical, not semantic.
+- Library rename/delete and per-tab mount state are not part of v1.0.0.
+- Persistence is one snapshot with last-write-wins semantics, not a multi-user
+  journal.
+- No browser pixel/click test is included; the generated client is validated by
+  `node --check`, helper tests, and plugin slot activation.
+
+## Development history
+
+| Date | Milestone |
+| --- | --- |
+| 2026-11-03 | Plugin skeleton, Host/Client split, CAT store, segmentation, TM, glossary, three tools, workbench UI, TXT/MD/DOCX/TMX import/export, mock MT. |
+| 2026-11-03 | Persistence layer (`dataDir`, v1 snapshot, atomic writes, corrupt fallback), `tmId`/`glossaryId` reservations, reload/corruption tests. |
+| 2026-11-03 | Graded `matchType` (`exact`/`near_exact`/three fuzzy bands) with locale labels and boundary tests. |
+| 2026-11-03 | Four additional agent tools (`cat_lookup_glossary`, `cat_list_documents`, `cat_switch_document`, `cat_suggest_mt`). |
+| 2026-11-03 | Multi-library TM/glossary mounts, `targetTmId`, v1→v2 migration, library modal, mounted-only semantics. |
+| 2026-11-03 | Read-only audit and hardening: mount checkbox, MT route reuse, clear confirmation, mounted write boundaries, strict persistence, revision polling, language/metadata fixes, store/client split. |
+| 2026-11-04 | Document tabs, close/reopen, duplicate fingerprint, recent-document cleanup. |
+| 2026-11-04 | Source-format-preserving DOCX export with sidecar storage and fixtures. |
+| 2026-11-04 | First-tier exchange formats: CSV mapping, SRT, WebVTT, PO, XLIFF 1.2, metadata persistence, timeline fixes. |
+| 2026-11-04 | Second-tier formats: HTML, JSON, XLSX, PPTX, RTF; binary upload/export; source-preserving PPTX. |
+| 2026-11-04 | Responsive UI density 1×–1.8×; button contrast fix; v1.0.0 release. |
+| 2026-11-06 | EPUB 3 import/export with source-package patching, shared HTML/XHTML scanner, import preprocessing stage, CSV/RTF hardening, PDF out of scope; v1.0.5 release. |