QudadFlow Translator 0.9.2

QudadFlow Translator 0.9.2

• 连接程序改为按需引导：只有主动选择 ChatGPT 计划且未检测到可用连接程序时，才显示下载操作。兼容的已安装版本直接复用。
• 下载权限为可选权限，仅在“下载并继续授权”中请求，跟踪本次下载。下载中断时提供重试及手动下载，连接程序就绪前不会打开账号授权。
• 用户打开下载文件并确认安装后，返回设置页自动检测并继续官方授权。停止等待、切换方案或关闭设置页会停止自动继续。浏览器不会静默执行程序，安装器当前未签名。
• 弹窗不再等待系统平台检查或注入整套网页翻译脚本，只读取所选文字；网站权限和本地设置并行读取。
• 后台唤醒复用已持久注册的快捷键脚本，不再扫描、注入所有已打开标签页。
• 合并相同的在途翻译请求；MyMemory 相同片段可跨网页和文档请求复用。保留有界缓存，按接口、凭据、模型、语言与上下文隔离。连接测试和 ChatGPT 计划结果不使用译文缓存。
• 本地 Translator 模型短时复用，最多保留两个空闲实例，30 秒后释放。语言方向不可用不会拖累其他已准备好的语言；手动准备模型后立即清除该方向的不可用记录。
• 沉浸式翻译分发前复用同一轮的元素样式检查，减少深层 DOM 重复计算，继续按原文节点定位和当前视区优先派发。
• 修复 iframe 重载后残留原文节点被再次派发，以及 PDF 连续阅读中布局调整产生的滚动事件覆盖新阅读位置的问题。
• 本机连接保留 30 秒供连续请求复用；Windows 凭据保护使用编译入口，兼容旧 DPAPI 文件，且每次检查磁盘变化。另一浏览器的退出、令牌轮换、额度暂停仍生效。
• 扩展版本号及显示版本统一为 0.9.2。

性能数据、测试范围和复现方法见 性能报告。ChatGPT 计划资格、安装要求和已知限制见 登录说明。发布与审核状态见 三平台记录。

验证：518 项主回归与 75 项专项检查通过。模拟提供方测试不代表真实服务商的速度、额度或翻译质量。Edge／Chrome 商店送审状态单独记录。

QudadFlow Translator v0.9.1

QudadFlow Translator 0.9.1

Windows x64 商店版首次打开时会询问是否下载可选的 ChatGPT 本地连接程序。用户同意后扩展打开本页附带的 EXE 下载地址；下载完成后仍须用户自行打开文件。安装器校验内置文件并自动注册当前 Windows 用户的 Edge／Chrome 本机宿主。侧载开发版显示针对当前扩展 ID 的命令。安装器尚未签名，ZIP 手动安装包仍可选用。默认免 Key 机器翻译无需连接程序。

官方开源计划接入文档明确列出符合资格的 Plus／Pro；Business／Team 成员通过此路径使用自己额度的资格尚未确认。授权仍在 OpenAI 官方页面完成，达到限额时不会自动切到付费 API。

本地验证：完整回归 513／513，隔离 Edge 的 ChatGPT 界面 22／22（商店 ID 分支为模拟）、Native Host 29／29、安装器校验与负面样本 4／4。安装器的真实图形安装过程、Windows 安全提示和商店安装态下载尚未覆盖。运行版 ZIP 已逐文件审计，内含 271 个扩展文件，不含 EXE 或 Node.js。附件 SHA256SUMS-v0.9.1.txt 列出全部发布文件的校验值。

详细更新说明 · 安装与限制 · 三平台状态

English: This update adds an optional, consent initiated Windows x64 installer for the ChatGPT plan preview. The user still opens the downloaded unsigned EXE. Default key free machine translation works without it. See the linked docs and checksums for details and limitations.

QudadFlow Translator 0.9.0

QudadFlow Translator 0.9.0

发布日期：2026-10-29。此版本的商店提交、审核和公开上架进度分别记录在 三平台发布记录。

翻译与设置

• 无 AI 配置时默认使用免 Key 机器翻译：优先尝试已准备好的浏览器本地 Translator 模型，否则立即使用 MyMemory。首次下载本地模型需要用户在同一设置分区点击“下载并准备本地模型”；浏览器不支持时该按钮停用。后台 Worker 不调用本地模型，翻译在扩展隐藏页面完成；仅在该页面的 localStorage 缓存语言方向可用性，不保存翻译文字。MyMemory 受服务商额度和保留条款限制，匿名用量可能按共享网络计数，不承诺每位用户独享。可自愿填写联系邮箱使用服务商公布的较高额度，或配置自己的服务商密钥。
• MyMemory 请求按段落及 500 UTF-8 字节切分，单任务最多并行 3 个片段、全部任务最多并行 4 个请求；译文仍按原顺序展示。短期内重复成功片段复用内存缓存。速度取决于浏览器模型、网络和服务商，本版未给出真实站点的通用加速比例。
• 沉浸式译文定位改进链接节点附近的插入位置，并继承竖排书写方向；保留原文与链接。布局不保证适配所有网页，遇到错位可恢复原文。
• “已保存方案”、接口类型、上游模型、连接测试和保存设置集中在“API 方案与模型”分区。账号模式与自备 API 均可从上游模型列表手动选择并填入；不会擅自替用户选模型。

ChatGPT 官方登录预览

• 用户主动选择后，通过本机 Native Messaging 程序按 OpenAI 官方 Sign in with ChatGPT 开源流程授权。用户各自授权；令牌留在 Windows DPAPI 加密文件中，扩展只保存账号引用和模型。推理请求使用官方 Responses API，store:false、stream:true，必须收到 response.completed 才判定成功。不会读取 ChatGPT 网页 Cookie 或自动转到付费 API Key。
• Windows x64 用户可从本版 GitHub Release 下载独立连接程序 ZIP，解压后双击安装。它自带经过 Node.js 官方 SHA256 校验的运行时，按当前用户注册 Edge 和 Chrome 商店版 ID，无需预装 Node.js 或管理员权限。商店扩展本身无法静默安装本机程序；升级后首次打开弹窗只提供一次可关闭的安装引导，默认机器翻译不依赖此程序。此连接程序尚未签名。
• 个人 Pro 账号已在本地真实读取 8 个模型，并完成一次固定短句推理。其他账号、真实图片质量、长期稳定性和 Chrome 原生桥接未完成实机验证。OpenAI 开源接入文档目前只明确列出符合资格的 Plus / Pro 可使用计划额度。Business / Team 成员经此路径使用自己的额度尚未获得官方确认；组织策略拒绝不等于套餐资格结论。界面在授权前告知，并保留错误码与 Request ID 便于排查。
• 授权码过期或失效时可重新发起授权；首次兑换失败保留本机加密的应用注册信息。连接测试修复了上游 SSE 缺少 Content-Type 时的误报。账户、模型和 API 方案均由用户明确选择。具体安装、注销和组织策略说明见 ChatGPT 登录说明。

验证范围

本版回归采用隔离 Edge、生成页面和模拟上游响应；真实翻译速度、准确率及不同账号的计划额度无法由合成测试证明。最终构建的逐文件校验与本轮测试结果见 发布记录。历史 0.8.16 及各开发版的测试记录保留在各自版本文件中。

完整本地回归 21 份报告共 513 项通过、0 项失败，覆盖默认设置、可关闭的 Windows 安装引导、沉浸式定位、文档与论文阅读、模型切换及打包审计。隔离 Chromium 中另有两次真实 MyMemory 固定短句请求得到中文译文；单次样本不能代表普遍速度。Chromium 151 报告英译中本地模型可下载，但在全新测试配置中等待 180 秒仍未完成下载，因此尚未验证真实本地模型译文；该场景会立即使用 MyMemory 回退。

QudadFlow Translator 0.8.16

QudadFlow Translator 0.8.16

October 23, 2026. All-platform update authorized by the owner. GitHub publication, store submission and store approval are tracked separately in the publication record.

Roomier figure workbenches

• 展开工作台 opens the existing paper or PDF/Word/TXT/Markdown workbench in a dedicated view that fills the browser viewport. 返回阅读 or Escape restores the reading layout. Selection, search, drafts, results and pending AI requests stay in the same workspace.
• The expanded overview uses larger figure previews and a fixed toolbar for search, selection and batch actions on wide screens. Native webpage tables join the overview while expanded and return to the source column afterward. Optional webpage supplementation stays folded under 核对漏项与补图.
• Questions and settings occupy the left column; answers and conversation history have a larger reading column. Expanded answers and figure results use the main scroll area instead of small capped result boxes. Narrow screens stack the columns.
• AI candidate review uses larger previews. 查看原页 returns to the PDF without discarding the candidates. Cropping reveals the original page first, then returns to the expanded workbench when a crop is completed. Escape cancels the crop and stays in the reader.
• The installed version remains visible at the bottom in both layouts. Compact document mode still preserves reading space.

中文说明

在 Edge 扩展管理中重新加载本地扩展，然后重新打开论文窗口或文档阅读器，确认底部 v0.8.16。点击工作台标题右侧的 展开工作台，即可单独查看总览、解读与问答、AI 候选核对。点击 返回阅读 或按 Esc 返回原布局。

宽屏下图表预览更大，搜索、全选、批量载入及提问操作集中在总览上方；问答设置和长篇结果分栏展示。展开与返回不会重新提取图表或发送 AI 请求；已选择的图表、问题草稿、分析结果和进行中的请求继续保留。

QQ 交流群：1048123205 · 支持邮箱：lcf1249288484@outlook.com。

Validation scope

87 behavior checks passed, zero failures: 8 expanded-workbench checks, 29 PDF AI extraction/review checks, 25 web companion checks and 25 visual supplementation checks. The package also passes manifest, 266 runtime-file, 34 script syntax and local dependency validation. See verification.

Validation uses isolated Microsoft Edge profiles, generated papers/images and synthetic API responses. Reports and preview screenshots accompany this release. They check layout, state retention, keyboard navigation, source-page access and existing analysis behavior; they do not measure a real model's accuracy or speed.

QudadFlow Translator 0.8.15

QudadFlow Translator 0.8.15

October 23, 2026. Publication authorized by the owner. GitHub release, store submission and store approval are tracked separately in the publication record.

Upstream models in every reading view

The popup, PDF/Word/TXT/Markdown reader and paper/figure workbench automatically read the active API's model list when no cached catalog exists. Refresh or cancel inline and select a shared model. Dropdowns display model IDs only. No first result is selected automatically; the existing model survives a failed or empty read. API connections are created and switched only in Settings, with 新增 API 配置并使用 preserving the previous connection.

Discovery uses the saved endpoint, headers, credentials and host grant; it sends no page text. Simultaneous views share one request per connection, and lists persist within a bounded local cache. Connection changes cancel stale discovery. Some gateways do not expose a model route; enter their documented ID in Settings. Model names do not guarantee image support.

PDF reading and evidence extraction since 0.8.8

• Switch between page-by-page and continuous reading. Keep original PDF page proportions, chart positions and a full-text fallback for translations that cannot fit.
• Extract named figures and tables from the current page or the whole document using captions, drawing coordinates, continuation evidence and ordered figure legends. Reconstruct supported native text tables as copyable Markdown/TSV. Decorative images and repeated headers are excluded.
• Find AI 补充识别 inside 图表总览. Explicitly send chosen page images/text to an image-capable model, then review crops in the same compact workbench. Coordinate formats are normalized; valid neighbors survive invalid candidates. One corrective request can repair unusable page output; actual API errors stop requests.
• Share identity/overlap deduplication between local and AI extraction. Repeated or shifted table crops merge; uncertain replacements require review. Continued tables on different pages remain separate.
• Use select-all and batch figure questions. Extraction, overview, answers and candidate review share a collapsible tabbed workbench; the PDF expanded body stays within 34% of the viewport. Manual tools and AI page options begin folded.
• The first-use current-site permission reminder disappears after consent and returns only when the actual grant is absent. Protected browser/store pages omit it.

中文使用说明

更新后重新打开阅读页，确认底部 v0.8.15。在设置中保存当前 API；阅读界面只切换该 API 支持的模型，模型下拉框不附带“能力未知”等标签。

PDF 工具栏可切换逐页／连续阅读。底部图表工作台支持本页和全文普通提取；漏提取时进入 图表总览 → AI 补充识别，输入页码后识别并核对候选，确认后才加入总览。可全选图表结合正文提问，图片表格可主动调用 AI 转为可复制文本。

支持邮箱：lcf1249288484@outlook.com · QQ 交流群：1048123205。

Validation and limits

0.8.15 passed 121 checks (110 behavior + 11 build/package), with zero failures. The bare-ID refinement reran the same 27 model-selection checks successfully; these are not added to the total. 0.8.14 passed 241 checks, including ordinary extraction and manually annotated simulated AI responses for three supplied PDFs. Earlier results are historical, not newly rerun counts. See verification.

Tests use isolated Edge profiles, generated fixtures and synthetic providers/credentials. They establish software behavior, not live-model accuracy, speed or cost. Extraction is heuristic; scanned body replacement needs external OCR. AI table transcriptions, numbers and generated Python/R code require checking. Code is not executed. File/page/image resource limits remain. There is no layout-preserving translated PDF/DOCX export. Only the final runtime ZIP is for store upload; private sample PDFs and credentials are excluded.

QudadFlow Translator 0.8.8

QudadFlow Translator 0.8.8 — Figures, full tables and version visibility

October 22, 2026. Download the versioned installation and source packages from the GitHub release. Edge 0.8.8 is In review, with 0.8.7 still Live. Chrome remains deferred. Actual submission/approval status is recorded in STORE_SUBMISSION_0.8.8.md.

Figures and linked tables

Springer thumbnail, reading-companion and original-image variants now share a canonical image identity. Repeated size/format variants merge while distinct panels and identity-bearing query parameters remain separate. Extraction keeps the fuller caption and excludes small download/decorative controls. Items are not merged solely because they have the same figure number.

Explicit article capture discovers eligible same-origin numbered full-table links. The research window automatically reads those HTML pages using the original article's existing site session, parses them locally without executing embedded resources, and adds the complete table to copy/export controls and paper Q&A material. This needs no AI request or new host permissions. Failed tables retain the source link and retry action; navigating the source page invalidates its binding. Ordinary citations, downloads, supplements and foreign-site links are excluded.

Popup, Settings, document reader and research window show QudadFlow Translator · v0.8.8 at the bottom, obtained from the running extension manifest.

Verification

147 targeted checks passed, zero failures: 136 behavior checks and 11 build/package checks. Saved complete public DOM from the requested Springer article produces five figures and one linked table; its saved table page preserves all 54 rows (including the header) and five columns. Simulated source-bound downloads cover errors, redirects, invalid sources, stale documents, duplicate saves, retry, export and complete paper Q&A material. Responsive layouts and version footers were verified.

Tests use isolated Edge profiles, saved public DOM and simulated APIs. They do not establish live publisher download availability, real AI accuracy or speed. Live publisher availability may depend on the source session and site access; it was not established by the fixture tests. Earlier 0.8.7 full-suite results remain historical; current Chrome integration is not verified. See VERIFICATION.md.

中文更新说明与升级检查

本版修复 Springer 图表因缩略图、侧栏和原图版本而重复展示的问题；同编号的不同分图仍保留。完整表格链接自动尝试读取，成功后可复制或导出 Markdown/TSV，并随正文参与论文问答，无需 AI 转写。读取失败保留原表和重试入口。

1. 在 Edge 扩展管理页找到现有的 QudadFlow Translator 解压安装，点击重新加载，确认 0.8.8。保持原安装目录，沿用当前浏览器设置。
2. 刷新原论文网页。旧论文窗口保留旧捕获，请关闭旧窗口并从原网页重新打开论文分析。
3. 在给定 Springer 论文中，检查总览为 Figure 1–5 各一项；等待读取 Table 1，核对 54 行（含表头）、5 列，再试用复制/TSV 下载。
4. 可继续测试其他论文，核对独立分图未被误合并。需要结合图像提问时，仍先批量载入原图并勾选。
5. 在弹出面板、设置、文档阅读器或论文窗口底部确认实际运行版本。只复制文件不代表浏览器已启用新版。

本次发布 GitHub 并已提交 Edge 更新，0.8.8 状态为审核中，商店公开版仍为 0.8.7；Chrome 暂缓。GitHub 下载可用与商店审核通过分别记录。翻译和 AI 请求继续等待完成、实际调用错误或手动取消；20 秒超时仅用于表格 HTML 等资源下载。

Support: lcf1249288484@outlook.com · QQ community: 1048123205.

QudadFlow Translator 0.8.7

QudadFlow Translator 0.8.7 brings the cumulative 0.8.x reading, paper and build improvements to the published 0.7.1 baseline.

• Editable/recordable webpage shortcut, default Shift+A, with save, validation, reset, disable and safe recording cancellation.
• Custom immersive text colors update existing translations without another API call. Viewport-first translation supplements dynamic text and skips confidently detected target-language text.
• Floating paper Q&A; caption-aware figures, image tables and native tables; select all, batch high-resolution originals, Markdown/TSV table export and explicit vision transcription.
• PDF/DOCX/TXT/Markdown reading, editable Skills, optional exportable cross-page notebook memory, upstream model discovery and downloadable Python/R drafts. Code is never executed.
• Deterministic runtime/source packaging, dependency validation and exact ZIP byte auditing. Updated English/Chinese guides, descriptions and six store screenshots.

Validation: 342 checks passed, zero failures: 331 behavior checks and 11 build/package checks. Synthetic credentials, generated documents and local simulated APIs were used. Real-provider speed/accuracy and current Chrome integration are not established.

Edge: 0.8.7 submitted and confirmed In review on October 22, 2026; public store remains 0.7.1 pending approval. Chrome: deferred for this release at the owner's request.

Download qudadflow-translator-v0.8.7.zip for unpacked installation; the source ZIP includes development files. For an existing unpacked installation, replace files in the same folder, click Reload at edge://extensions, confirm 0.8.7 and refresh webpages. Store installations update after approval. SHA256 checksums and an entry-by-entry audit accompany the ZIPs.

中文：默认 Shift+A，支持手动保存或录制；修正取消录制，保留自定义颜色。加入论文图表全选、批量原图、可复制表格及浮窗问答等累计改进。Edge 已提交审核，Chrome 暂缓。解压版请在原目录更新后重新加载并确认版本。

Full release notes · Verification · Platform status · 中文指南

Support: lcf1249288484@outlook.com · QQ community: 1048123205.

QudadFlow Translator 0.7.1

QudadFlow Translator 0.7.1

Settings can now read the model list from your OpenAI-compatible provider. Search IDs, filter reported text/image capabilities, inspect returned metadata and explicitly choose a model. Missing metadata stays unknown; the extension does not guess vision support, speed, cost or quality from names, select the first entry, or overwrite your existing model.

The responsive settings layout adds section navigation, a clear unsaved-change state and accessible save/test controls. Connection tests use the current draft and a fixed short sentence, bypass the cache and report elapsed time without saving or enabling the draft. Tests do not validate image input. Model discovery requests support cancellation; changing the connection invalidates previous discovery and verification results. API-host grants may persist independently of saving.

Model discovery uses the same-origin models route, with a 15-second deadline, 2 MB response limit and at most 1,000 unique entries. Lists remain only in the settings tab. Error text hides configured credentials. Root-domain endpoint construction now produces valid model and translation routes.

The popup, Settings, English/Chinese READMEs and website include QQ community: 1048123205. Search the group number in QQ to join. English and Chinese privacy policies describe model discovery and draft testing. Chrome Web Store submission text is included alongside the Edge update materials; the runtime ZIP is shared by both targets.

Validation: 191 Edge checks passed, zero failures: 22 model/settings, 87 webpage/provider, 34 document, 13 configuration, 17 layout/figure and 18 research checks. Tests use isolated profiles, synthetic keys and local fixtures. An installed-Chrome attempt timed out waiting for an extension service worker; Chrome behavior and store acceptance remain unverified. Live-provider speed, vision quality and scientific reproduction are not established by these tests.

中文说明

设置页新增上游模型读取、ID 搜索、上游已声明的文本/图片能力筛选及元数据详情。必须明确选择模型，不自动选第一项、不覆盖已有模型；缺失信息显示未知，不凭名字猜读图、速度或价格。

设置界面改为分区导航，增加未保存状态、保存/测试操作栏及窄屏布局。连接测试用当前草稿发送固定短句，显示耗时，绕过缓存，不保存或启用草稿，也不验证图片能力。模型读取可取消，连接更改后旧的模型读取和验证结果失效。API 域名授权可独立保留。

模型列表仅保存在当前设置页内存；接口同源、15 秒超时、响应上限 2 MB、最多 1000 个去重模型。错误提示隐藏配置密钥。修复仅填写根域名时模型/翻译路径出现双斜杠的问题。

产品和 GitHub 中英文资料加入 QQ 交流群：1048123205，在 QQ 搜索群号加入。同步准备 Edge 更新与 Chrome 商店资料，并补充双语隐私说明。

Edge 191 项检查通过、0 失败。独立 Chrome 检查未能加载扩展服务工作线程，不能宣称 Chrome 已验收或已上架。模拟接口结果不代表真实服务商速度、质量或图表精确复现。详情见 验证记录。

Published on October 21, 2026: GitHub 0.7.1 is available. Edge Add-ons confirms 0.7.0 Live and 0.7.1 In review; submission evidence. Microsoft estimates seven business days for review; 0.7.1 approval is pending.

2026 年 10 月 21 日：GitHub 0.7.1 已发布；Edge 0.7.0 已上线，0.7.1 已提交审核。微软页面预计七个工作日，0.7.1 尚待批准。

QudadFlow Translator 0.7.0

QudadFlow Translator 0.7.0

Edge Add-ons update submitted October 21, 2026: In review. The store currently serves 0.5.0 Live; 0.7.0 approval is pending.

Paper research now works across webpages and PDF, Word DOCX, TXT and Markdown, alongside immersive and selection translation.

• Extract webpage article text, figure captions and images into an editable research workspace. Visible SVG/canvas charts support viewport crops; remote images load only after an explicit action and source permission.
• Analyze research questions, methods, statistics and evidence using three built-in Skills or your own imported/edited prompt templates. Stream, stop and export results as Markdown.
• Opt into cross-page memory, isolated by notebook. IndexedDB stores deduplicated bounded text records; Settings provides retention/context limits, search, deletion and JSON/Markdown export. Memory defaults off and does not affect ordinary translation.
• Recognize figures and generate separate downloadable Python and R drafts with data requirements. The extension never executes code. Original data and independent verification are needed for scientific reproduction.
• Preserve a figure's extraction-page attribution when navigating a PDF. Report memory-save failures without losing the completed analysis.

This release also includes 0.6.0's PDF original-position replacement, Word browser layout and optional vision workflow. Traditional translation and AI both support basic reading; enhanced paper/figure analysis requires an AI profile. Scanned-body OCR and layout-preserving PDF/DOCX export remain outside this release.

Limits: paper input 60,000 characters; 20 custom Skills; memory 300 records / two million content characters, with 90-day default retention; per request at most six memory records and 6,000 characters by default. Web captures use a separate four-entry, one-hour local cache with up to 16 MB image bytes each. See the bilingual privacy policy for transmitted sources, image capture, exports and deletion.

Validation: 169 isolated checks passed, zero failures (87 webpage/provider, 34 document, 13 configuration, 17 layout/figures, 18 research). Generated fixtures and simulated APIs were used; real-provider performance, model accuracy and generated-code execution were not tested.

Edge 125+ required. Reload an unpacked installation after replacing its files and refresh open pages. Store status is recorded in STORE_SUBMISSION_0.7.0.md; GitHub release availability does not imply Microsoft approval.

中文说明

0.7.0 新增网页论文图表提取，与 PDF、DOCX、TXT、Markdown 共用论文分析、Skill、跨页记忆和 R/Python 复现草稿流程。内置三种分析模板，可自行编辑和导入/导出；记忆默认关闭，开启后按笔记本隔离，设置中可搜索、删除和导出。分析结果流式显示，支持停止和 Markdown 下载。

图表识别后可分别生成 .py/.R 草稿及说明，不自动执行；缺失数据需用户提供，估读值需核对，不能仅凭图片保证精确复现。PDF 翻页后仍保留图表原始来源页码。本版同时包含 0.6.0 的文档原位翻译与底部图文 AI。169 项隔离测试全部通过，不代表真实服务商速度或模型质量；商店是否已通过审核以提交记录为准。

QudadFlow Translator 0.6.0 Preview

QudadFlow Translator 0.6.0

Development preview. Not submitted to Microsoft Edge Add-ons. Version 0.5.0 was last confirmed In review on October 21, 2026; this preview does not replace that submission.

Traditional MT and AI both provide the base immersive and selection translation features. This release adds original-position document translation and optional vision AI tools.

• PDF text replacement keeps page proportions, raster images and vector charts. Original-layout comparison and paragraph comparison remain available. Long/unsupported text keeps its source and opens its full translation on click.
• Word DOCX renders tables, images, margins and pages in the browser, with measured text overlays. TXT/literal Markdown keep reading-paragraph positions. Word fidelity can differ from Microsoft Word; use PDF for strict layout requirements.
• The bottom AI workflow extracts local pictures, allows PDF chart cropping or whole-page capture, previews/downloads images, and translates their text with chart explanations. A separate saved vision AI profile works alongside traditional body translation. Only explicit analysis sends selected images; cancel and configuration changes stop active work.
• Body export remains bilingual TXT. No layout-preserving PDF/DOCX export or built-in scanned-page OCR replacement. Chart AI provides text and explanations, not redrawn charts.

Limits: 25 MB per document, 300 pages/reading sections, one million extracted text characters. Images: longest edge 1,600 pixels, at most 4 MB encoded each and 20 images/24 MB total. Document and image state stays in tab memory. Images bypass the text cache; downloaded files persist. See the updated privacy policy.

All parsing/rendering libraries are bundled: PDF.js 6.4.299, fflate 0.8.3, docx-preview 0.4.1 and JSZip 3.10.2. Requires Chromium-based Edge 125+. Reload an existing unpacked installation after replacing its runtime files.

Verification: 151 checks passed, zero failures, in isolated Edge with generated documents, synthetic keys and local APIs: 87 webpage/provider + 34 document + 13 configuration + 17 layout/figure checks. Pixel checks verify unchanged chart/image regions and retained overflow originals. Real provider accounts, private settings, latency and image-analysis quality were not tested.

中文说明

传统接口与 AI 均提供基础沉浸式翻译及划词翻译。0.6.0 新增 PDF 原位文字替换，保持页面比例和图表位置；Word 在浏览器中保留表格、图片和页边距，复杂排版可能与 Microsoft Word 有差异。支持沉浸原位、原版左右对照、段落对照；长译文放不下时保留原文，点击查看完整译文。

底部 AI 图文工作流可本地提取图片、框选 PDF 图表或提取整页，预览后点击分析，才向所选视觉 AI 发送图片并生成文字翻译和图表解释。图表 AI 可单独选择已保存方案，正文继续用传统接口。支持下载图片、停止分析与配置切换取消。

正文仍导出双语 TXT，不导出版式保真的译后 PDF/DOCX。扫描正文原位替换需先外部 OCR；AI 阅读页图不自动替换扫描文字。151 项模拟接口检查通过，不代表真实 API 速度或翻译质量。此版本为开发预览，尚未提交商店；更新侧载目录后须在 Edge 重新加载扩展。