EdgeTranslate 4.3

EdgeTranslate 4.3

AI 페이지 번역 엔진 전면 재구축 (vNext)

OpenAI · Google AI Studio · OpenAI-호환 백엔드를 쓰는 LLM 기반 전체 페이지 번역을 하나의 통합 엔진으로 새로 만들었습니다.

• 빠른 속도 — 짧은 문서 ~6–8초, 초대형 위키 문서 ~22초 (gpt-5.6-luna, reasoning off 기준 실측). 즉시 최대 동시성 + 스트래글러 헤징으로 꼬리 지연 제거.
• 높은 품질 — 블록 단위 + 인라인 플레이스홀더 wire로 어순이 다른 언어쌍의 문장 파편화 문제 해결. URL·식별자·코드는 로컬 마스킹으로 보존.
• 안정성 — 원본 DOM에 CAS 커밋(링크·포커스·이벤트 보존), 유한한 복구 재시도, 정직한 미완료 보고.
• 코드 태그 처리 개선 — <code>/<kbd>/<samp> 안의 자연어(가나·한글·CJK 등)는 번역하고 실제 ASCII 코드는 보존.

옵션

• AI 번역 모델 목록을 provider API에서 자동 갱신 (수동 업데이트 불필요).

UI

• 번역 진행바가 0%에서 정상 시작하며 단조 증가. 헤더에서 개발자용 토큰/요청 카운터 제거.

---

설치: 아래 edge_translate_chrome.zip을 받아 압축 해제 후, chrome://extensions에서 개발자 모드 → "압축해제된 확장 프로그램을 로드"로 폴더를 선택하세요.

v4.2 — Smarter Page Translation & Caption Reliability

v4.2 - Smarter Page Translation & Caption Reliability

Highlights

• Reworked page translation around a compact LLM translation IR so models translate language content instead of regenerating page HTML.
• Added DOM coverage verification and completion sweeps to reduce cases where translation appears complete while source-language content remains.
• Improved page root selection for articles, headings, comments, discussions, and late-loaded content while continuing to avoid site chrome.
• Added Chrome built-in/on-device page translation routing with support for compact [[n]] segment markers.
• Optimized first-run settings defaults for a calmer UX: intentional double-click translation is enabled, aggressive automatic translation stays off, and result panels start cleaner.

Captions & Reliability

• Improved YouTube realtime caption merging so fragment captions are replaced more cleanly by stabilized captions.
• Scoped Google realtime captions to the current visible line and collapsed duplicate or repeated multiline caption output.
• Added safeguards against source echoes, hallucinated subtitle samples, stale page translation responses, and invalid DOM translation output.
• Hardened runtime channel messaging against malformed messages and synchronous service failures.

Documentation & Tests

• Added docs/LLM_PAGE_TRANSLATION_SPEC.md to document the LLM page translation architecture and safety model.
• Added unit coverage for settings defaults and channel hardening.
• Expanded page translation and caption tests for coverage scans, partial retries, marker parsing, comment roots, on-device routing, and stale-output prevention.

Validation

• Passed 202 targeted EdgeTranslate unit tests.
• Passed 14 translators local quality tests.
• Built and packaged successfully with:
  npm run build -w @edge_translate/translators && npm run pack:chrome -w edge_translate
• Passed git diff --cached --check before the release commit.

Download

• Chrome extension package: edge_translate_chrome.zip
• SHA-256: cee111ae3669db8c4d6b7e647ec758a1ec0f0a08cbb958f7f5226bd1d81853ed

v4.1 — Optimized AI Translation Pipelines

What's New in v4.1

🚀 On-Device Pipeline (Gemini Nano) — Complete Overhaul

• Single-pass streaming — Removed the redundant 2nd-pass refinement logic that added latency without meaningful quality improvement. Translations now stream in a single pass for significantly faster response times.
• Smart text chunking — Long inputs are intelligently split at paragraph → line → sentence boundaries (800-char optimal window) to maintain formatting and translation quality.
• Proper noun protection — System prompt now explicitly forbids hallucination and improper substitution of proper nouns, brands, and titles (e.g. 「はじまりのピカチュウ」 no longer gets replaced with random movie titles).
• Post-translation safety net — Known false-friend corrections (e.g. 国勢調査 → 인구총조사, not 국세조사) applied as a final pass.
• Passthrough punctuation — Decorative symbols (※, →, ★, ●, etc.) are preserved through translation via placeholder substitution.

⚡ AI Studio Pipeline (Gemini 2.5 Flash Lite) — Streamlined

• Prompt optimization — Reduced prompt token count by ~50% while maintaining translation quality. Removed verbose script rules, redundant instructions, and unnecessary role declarations.
• Generation config — Removed redundant candidateCount and topK parameters; thinkingBudget: 0 for maximum throughput.
• Script constraints — Simplified from 22-language verbose rules to 9-language concise constraints.

🐛 Bug Fixes

• Fixed mainMeaning field not receiving post-translation rules (e.g. 국세조사 correction only applied to translatedText, not mainMeaning).
• Fixed originalText in chunked translations incorrectly set to the full input text instead of the current chunk.
• Fixed parseInt calls missing radix parameter (ESLint warnings).
• Fixed unnecessary arguments passed to buildGeminiNanoSystemPrompt() in the on-device bridge.

🧹 Code Cleanup

• Removed ~250 lines of dead code across both pipelines: getPrimaryScript, getTargetLanguageOutputRule, findSoftBreak, splitLongParagraphAtSentenceEnds, splitTextIntoParagraphChunks, needsRefinement, and unused constants.
• Extracted smartSplitText() as a shared utility function.
• Unified prompt structure between chrome_builtin_translate.js and on_device_bridge.js.

📖 Full Changelog

v4.0.3...v4.1

EdgeTranslate 4.0.3

EdgeTranslate 4.0.3

This patch release tightens the 4.x on-device translation path and refreshes the extension polish prepared for store submission.

On-Device AI Translation

• Strengthened Chrome Gemini Nano prompts so translation tasks explicitly translate ordinary text instead of copying source sentences unchanged.
• Added lightweight output validation that blocks likely source-copy results without issuing a second Gemini Nano prompt, preserving speed and thermal behavior.
• Kept the existing no-second-prompt cleanup for common untranslated Japanese/Chinese fragments such as mixed-script security terms.
• Added regression coverage to ensure copied source output is rejected after only one on-device prompt.

UI and Localization Polish

• Updated the extension version to 4.0.3.
• Updated visible copyright years from 2025 to 2026.
• Improved native select spacing in the popup and settings UI while keeping browser-native dropdown behavior.
• Refined floating translation panel provider sizing so the provider pill does not stretch awkwardly on wide panels.
• Audited supported locale files, removed stale endpoint-related strings, and preserved product/provider proper nouns such as Edge Translate, Google, Bing, DeepL, Chrome, Gemini Nano, and Google AI Studio.

Packaging

• Rebuilt Chrome and Firefox packages with manifest version 4.0.3.
• Included Chrome Web Store package: edge_translate_chrome.zip.
• Included Firefox package: edge_translate_firefox.xpi.

Validation

• Passed npm test -w edge_translate -- --runInBand.
• Passed npm run build:chrome -w edge_translate.
• Passed npm run pack:chrome -w edge_translate.
• Passed npm run pack:firefox -w edge_translate.

EdgeTranslate 4.0.2

EdgeTranslate 4.0.2

This patch release focuses on PDF routing performance, floating translation panel polish, and refreshed public-facing documentation for the 4.x line.

PDF Viewer and PDF Hijacking

• Reduced PDF detection overhead on ordinary web pages by avoiding unnecessary cache writes for non-PDF main-frame responses.
• Preserved automatic PDF hijacking for normal PDF URLs, local file PDFs, and header-confirmed PDFs whose URLs do not end in .pdf.
• Added coverage to ensure header-confirmed PDFs still open in the built-in pdf.js viewer.
• Kept the guard that prevents unnecessary network probing on regular Chrome Web Store pages.

Translation Panel UI

• Fixed a floating-panel layout issue where the settings, pin, and close buttons could overflow when the panel width was narrow.
• Improved the translator selector so long provider names truncate cleanly instead of pushing action buttons outside the panel.
• Slightly compacted header action buttons while keeping the Material-inspired visual language.

Documentation and Store Copy

• Rewrote the main README for the 4.x extension direction.
• Refreshed localized README files for Simplified Chinese, Traditional Chinese, Japanese, Korean, and Russian.
• Added Chrome Web Store listing copy in English, Korean, Simplified Chinese, and Traditional Chinese.
• Added README screenshots showing selection translation, pinned side-panel mode, the PDF viewer, and PDF selection translation.

Validation

• Passed npm test -w edge_translate -- --runInBand.
• Passed npm run build:chrome.
• Built and verified edge_translate_chrome.zip and edge_translate_firefox.xpi with manifest version 4.0.2.

EdgeTranslate 4.0.1

EdgeTranslate 4.0.1

EdgeTranslate 4.0.1 is a focused stability release for the 4.0 line. It fixes the most important regressions found after the large Material UI and PDF viewer update, especially around PDF translation, Chrome Gemini Nano behavior, translation panel readability, and extension-side error reporting.

Highlights

• Fixed AI translation from the extension PDF viewer.
• Fixed Chrome LanguageModel API requests that failed when the target language was not one of Chrome's attestation languages.
• Removed the slow Gemini Nano second-pass repair prompt that could make translations feel dramatically slower.
• Added lightweight cleanup for common untranslated fragments such as Japanese 不正/ログイン leftovers in AI translation output.
• Restored a more practical translation panel layout by removing pronunciation text display while keeping TTS playback.
• Fixed PDF hijacking and PDF thumbnail/sidebar behavior for local, blob, and remote PDF sources.
• Stopped PDF detection from probing ordinary non-PDF pages such as Chrome Web Store developer console URLs, avoiding noisy CORS errors.

Chrome Gemini Nano and AI Translation

• Always supplies expectedOutputs to Chrome's LanguageModel.create() requests so newer Chrome versions no longer report: No output language was specified in a LanguageModel API request.
• Uses a safe fallback output-language hint for targets outside Chrome's currently supported attestation set while the prompt still asks for the user's actual target language.
• Skips the page-bridge route for extension-hosted PDF viewer tabs and routes Gemini Nano requests through the extension/offscreen path instead.
• Improves malformed JSON handling so raw JSON fragments and removed pronunciation fields do not leak into the translation panel.
• Keeps normal AI translation concurrency at the tuned value while avoiding the expensive extra Gemini Nano repair call.
• Adds target-language prompt guidance for multiple language families so AI output is more likely to stay in the requested writing system.

Translation Panel UX

• Removed visible pronunciation text from translation results because it created noisy, oversized cards and awkward expand/collapse behavior.
• Kept speaker/TTS actions available.
• Refined result truncation and drawer behavior so the main translation remains readable while secondary content stays less intrusive.
• Improved dark-mode handling and layout consistency for translation result cards, drawer blocks, and action icons.
• Fixed object-shaped extension errors so the UI shows meaningful error messages instead of [object Object].

PDF Viewer Fixes

• Fixed PDF drag/drop and extension PDF hijacking paths that could open the viewer shell without loading the document contents.
• Fixed blob URL handling and repaired previously double-encoded PDF viewer file parameters.
• Improved remote PDF loading behavior for links such as arXiv PDF URLs.
• Fixed PDF thumbnail sidebar mouse scrolling.
• Preserved PDF viewer UI refinements from 4.0 while addressing regressions in content loading and navigation.

Error Handling and Detection

• Prevented the PDF detection fallback from running network probes against unrelated non-PDF pages.
• This removes noisy CORS errors from pages like https://chrome.google.com/webstore/devconsole/....
• Improved channel error formatting for background/content/offscreen messaging failures.

Validation

• npm test -w edge_translate -- --runInBand
• npm run build:chrome
• npm run pack:chrome -w edge_translate
• npm run pack:firefox -w edge_translate
• Verified packaged manifests report version 4.0.1.

Release Assets

• edge_translate_chrome.zip
• edge_translate_firefox.xpi

EdgeTranslate 4.0

EdgeTranslate 4.0

EdgeTranslate 4.0 is a major update focused on a more polished extension experience, a deeply refreshed PDF viewer, Chrome built-in AI translation experiments, better dark mode support, and smoother interactions across the UI.

Release Highlights

• Updated the extension version to 4.0.
• Reworked the PDF viewer with a much stronger Material Design-inspired visual language.
• Added broad dark mode support across the PDF viewer, translation panel, popup, and settings page.
• Refined the translation UI with cleaner cards, menus, spacing, scrollbars, and animations.
• Introduced Chrome built-in AI translation support for regular text and word translation flows.
• Removed Gemini Nano page translation because the current on-device model path is not practical enough for full-page translation speed and thermals.
• Packaged both Chrome and Firefox release artifacts.

Chrome Built-in AI Translation

• Added support for Chrome's built-in on-device AI model path for normal translation requests.
• Tuned AI translation concurrency to improve response time for regular translation flows.
• Replaced the old “Local Translation” wording with clearer AI translation terminology.
• Improved word translation behavior so translation, detailed meaning, definitions, and examples can be handled through the AI provider path when configured that way.
• Removed the Gemini Nano full-page translation option from page translation settings after real-world testing showed it was too slow for practical use.

PDF Viewer Overhaul

• Upgraded the bundled PDF.js viewer to the v5.7.284 generation.
• Fixed dragged PDF/blob loading so extension-hosted blob URLs no longer fail with unexpected server response errors.
• Redesigned the top toolbar controls, page counter, zoom controls, annotation group, print/download group, and overflow menu with a cleaner Material-style layout.
• Reworked the page counter into a more unified control with a selected-page inner chip treatment.
• Improved the thumbnail sidebar header alignment, spacing, and visual hierarchy.
• Fixed multiple icon clipping, stray divider, and vertical alignment issues in the PDF toolbar and menus.
• Refined the “More” menu row layout so icons and labels align consistently.
• Redesigned the document properties dialog, including its header, icon alignment, information layout, and close button styling.
• Improved theme menus and PDF viewer dark mode so toolbar, sidebar, popovers, scrollbars, and document background feel cohesive.
• Kept the PDF UI compact while still improving touch targets, grouping, spacing, and visual clarity.

Translation Panel and Popup UI

• Refreshed the translation panel, result cards, provider dropdown, loading states, error states, drawer blocks, and notification surfaces.
• Fixed visual issues where scrollbars and borders appeared too close to the translation window edge.
• Improved icon sizing and alignment in translation cards and action buttons.
• Added smoother transitions for panel open/close, dropdowns, drawers, hover states, and interactive controls.
• Added dark mode styling for the translation panel and popup.

Settings Page Refresh

• Cleaned up settings page layout, spacing, form controls, menus, and section hierarchy.
• Fixed awkward wrapping in settings labels such as translation button position controls.
• Removed obsolete page translation settings for the removed Gemini Nano page translator path.
• Added dark mode support to the settings UI.
• Reduced awkward horizontal overflow in detailed settings menus.

Performance and Practicality

• Increased normal AI translation concurrency to 4.
• Kept full-page translation on practical engines instead of forcing the current on-device model path where it causes unacceptable latency and heat.
• Tuned prompt and request behavior around the AI translation path for faster regular translation use.
• Preserved Google/other page translation paths where they remain much faster and more suitable for whole-page workloads.

Quality and Tests

• Added and updated tests around page translation behavior, translation manager behavior, Chrome built-in translation handling, banner control, PDF viewer controls, and PDF file loading.
• Verified Chrome packaging with npm run pack:chrome.
• Verified Firefox packaging with npm run pack:firefox.
• Release assets are attached for both browsers.

Release Assets

• edge_translate_chrome.zip
• edge_translate_firefox.xpi

Notes

This release intentionally removes Gemini Nano full-page translation from the user-facing page translation options. Gemini Nano remains useful for smaller translation and word-assistance tasks, but the current browser on-device model path is not fast enough for a good full-page translation experience yet.

EdgeTranslate v3.0.2

EdgeTranslate v3.0.2

Changes

• Tune Local Translate concurrency for the local/VPS translation provider.
• Reduce local translation request concurrency to 2 to match the measured best server-side setting.
• Keep Local Translate endpoint/API key user-configurable and disabled by default.

Verification

• npm test
• npm run pack:chrome -w edge_translate
• npm run pack:firefox -w edge_translate
• Packaged Chrome and Firefox manifests verified as 3.0.2.

EdgeTranslate v3.0.1

EdgeTranslate v3.0.1

Patch release focused on Chrome Web Store readiness and configurable translation improvements.

Highlights

• Added a user-configurable Local API translator provider.
  • Disabled by default.
  • No hardcoded private endpoint or API key.
  • Users can configure endpoint and API key from Options.
• Improved translation request performance.
  • Uses sender tab id fast path when available.
  • Preserves Google detected source language to avoid redundant detection.
• Fixed local/dragged PDF loading in the bundled PDF viewer.
  • Local file:// PDFs are preloaded before PDF.js origin validation.
• Reduced unused extension permissions for Chrome Web Store submission.
  • Removed unused cookies, notifications, and declarativeNetRequest declarations.
  • Removed unused DNR rules.json packaging.
• Removed Google API key-like literals from the current source tree.

Validation

• Full test suite passed.
• Chrome package built and manifest version verified as 3.0.1.
• Firefox XPI built and manifest version verified as 3.0.1.
• Firefox web-ext lint completed with 0 errors.

Edge Translate v3.0.0

Changelog

Browser release packages

• Bumped the extension version to 3.0.0.
• Added an installable Firefox .xpi release artifact.
• Validated the Firefox build with web-ext lint and confirmed errors 0.
• Kept Chrome release output as a packaged .zip artifact.

Firefox compatibility

• Fixed the generated Firefox MV3 manifest for installable releases.
• Replaced legacy applications with browser_specific_settings.gecko.
• Converted Firefox background declaration from service_worker to background.scripts.
• Removed Chrome-only manifest entries from Firefox builds, including declarative net request/page-translate commands.
• Raised Firefox minimum version metadata to the MV3-compatible target.

PDF viewer fixes

• Preserved the extension PDF viewer hijack behavior, including local file:// PDFs.
• Fixed local/downloaded PDF blank loading by avoiding blob preloading for file: and blob: URLs.
• Added an "Open original PDF" escape hatch with a bypass marker to avoid redirect loops.

Build, tests, and CI

• Made Chrome the default development build target while keeping explicit Firefox/Safari targets.
• Added Firefox package validation to CI.
• Fixed clean-checkout CI by building workspace dependencies before tests.
• Refactored cache, page translate, TTS voice selection, and Bing parser helpers into smaller modules with regression tests.

Release assets

• edge_translate_chrome_v3.0.0.zip — Chrome/Chromium package
• edge_translate_firefox_v3.0.0.xpi — Firefox install package

Verification

• npm test passed.
• Chrome package built successfully with manifest version 3.0.0.
• Firefox package built successfully with manifest version 3.0.0.
• web-ext lint completed with errors 0 for the Firefox build.