Data Maskit 0.14.0 herunterladen

更新摘要 / Summary
- 新增「图片内嵌文字脱敏」：截图、文档里的图片、以及 PDF 页面先在本机识别并打码，再发给 AI。
- 开关默认关闭，不开时行为与 0.13.0 完全一致；开了但本机识别服务不可用时，内容会被拦住而不是原样发出去。
- 修复一处文字链路的邮箱漏检：半角冒号后紧跟邮箱地址时不再原样发出去。
- Optional on-device OCR masking for images, pictures embedded in Office files, and PDF pages, before anything is sent.
- Off by default; if it is on and the local OCR service is unavailable, content is blocked instead of being sent unmasked.
- Fixed a text-path email miss: an address typed right after a half-width colon is no longer sent in the clear.
新增
- 新增图片多模态脱敏：代理链路的截图、扩展上传的图片、Office 内嵌图片、PDF 页面都先经本机 OCR 识别，命中的文字在图上物理打码后上行（默认关闭，只走识别与打码，不把图转成文本）。
- Optional multimodal masking: screenshots (proxy path), uploads, Office-embedded images and PDF pages are scanned by a local OCR service and sensitive text is physically redacted (off by default; masking only, never image-to-text).
- PDF 逐页扫描：干净页保留矢量原文，只有含敏感文字的页变成位图；页数/耗时超限时直接阻断并提示拆分。
- PDF support: clean pages keep their vector text, only pages with sensitive text become bitmaps; oversized documents are blocked with a hint to split them.
- OCR 支持指向自建网关的视觉模型（ocr_endpoint + ocr_allow_remote + MASKIT_OCR_MODEL / MASKIT_OCR_API_KEY），提示词与重试按实测调优（标记合规率 50% → 3/3）。
- OCR can point at a self-hosted gateway vision model (ocr_endpoint / ocr_allow_remote / MASKIT_OCR_MODEL / MASKIT_OCR_API_KEY); prompt and retry tuned from measurements (tag compliance 50% → 3/3).
- 设置页新增开关；引擎新增 ocr_endpoint / MASKIT_OCR_ENDPOINT（默认只允许本机地址，自建网关需显式打开 ocr_allow_remote）。
- New settings switch; ocr_endpoint / MASKIT_OCR_ENDPOINT default to loopback only, self-hosted gateways require the explicit ocr_allow_remote opt-in.
- OCR 改为本机 det+rec 引擎（零出网、真检测框），坐标不再依赖模型"报数"；打码改为实体级精遮并用「裁图复认」保证不漏字。
- OCR now uses a local det+rec engine (no egress, real detection boxes) instead of model-reported coordinates; masking is entity-level and self-verified by re-recognition.
- 新增「OCR 组件」按需下载（可选运行时）：主安装包不变，启用时自动下载/校验/自检/回滚。
- New optional OCR runtime downloaded on demand: the main installer stays small, with automatic verify, self-test and rollback.
- OCR 组件下载地址内置官方默认（固定预发布位，不干扰主程序更新），内网/镜像可用环境变量指向自己的清单；配套发布流水线已就位。
- The OCR component now has a built-in official download location (a fixed pre-release, separate from app updates); air-gapped or mirrored setups can point it at their own manifest, and the release pipeline is in place.
- OCR 输出新增「坐标覆盖率」闸门：模型只标了部分行时按「没定位全」处理（阻断或整页遮盖），不拿漏掉的行赌运气。
- New tag-coverage gate: if the OCR model only tags part of the lines, the page is treated as not fully localised and is blocked or fully covered.
修复
- 修复审计在时间预算截断时不留「解析已跳过”痕迹的问题，排障数据不再在这种路径上缺口。
- Fixed audit time-budget truncation not recording the "parse skipped" marker; diagnostic data no longer has a gap on that path.
- 复用表容量提升 4 倍：重度使用（一天几千次请求）下，早前对话的占位符不再因容量淘汰而偶尔还原不了。
- Reuse-table capacity raised 4x: under heavy daily use, placeholders from earlier conversations no longer occasionally fail to restore due to capacity eviction.
- 修复“接管模式”冒烟验收长期误报：接管其实一直生效（连接复用正常），只是子进程日志被缓冲，判据读不到启动行。
- Fixed a long-standing false failure in the takeover smoke test: takeover was working (connection reuse intact), but the child process log line was stuck in stdio buffering.
- 修复高并发图片识别的验收误判：排队等待属于设计内的保护行为，不再计为失败。
- Fixed a concurrency-test false alarm: queue-wait backpressure is by design and no longer counts as a failure.
- 修复邮箱漏检：「邮箱：」半角冒号后紧跟地址时（无空格）此前不脱敏，纯文字请求同样受影响；连接串密码仍由连接串规则保护，不受影响。
- Fixed an email miss: an address typed directly after a half-width colon (no space) was previously not masked, affecting plain text requests too; connection-string passwords remain protected by the connection-string rule.
- 图片打码框增加上下留白，盖住英文降部与高位数字的笔画残边。
- Image redaction boxes now include vertical padding to cover descenders and tall digits.
- 图片打码框增加随行宽缩放的左右留白：中文前缀较窄的字体（如微软雅黑）上，宽行的实体框起点会系统性右偏导致首位数字露边，现已修复。
- Redaction boxes now add width-proportional horizontal padding: on fonts with narrower CJK glyphs (e.g. Microsoft YaHei), entity boxes on wide lines started systematically too far right and exposed the leading digits; fixed.
- 修复超宽单行截图漏检，校正手机照片方向；图片元数据与 PDF 的作者、注释、表单、书签、文本附件里的敏感值现在会被替换，其余内容原样保留；暂不支持的动图会阻断。
- Fixed missed detections on extra-wide screenshots and phone-photo orientation; sensitive values in image metadata and in PDF authors, comments, forms, bookmarks and text attachments are now replaced while everything else is kept; unsupported animations are blocked.
- 修复引擎热重载将最小日志模式重置为详细模式的问题，避免意外保存普通个人信息原文。
- Fixed engine reloads reverting minimal logging to detailed mode, preventing unintended retention of ordinary personal data.

扩展更新提示 / Extension update notice
⚠️ 本次发版扩展有改动：请在 chrome://extensions 重新加载扩展（若提示新增站点权限，点允许）。
⚠️ The browser extension changed in this release: reload it at chrome://extensions (allow the new site permission if prompted).
Full Changelog: https://github.com/xiaYuTian11/maskit/compare/v0.13.0...v0.14.0