Releases: coolplayagent/into-markdown
Release list
Into Markdown 0.0.8 [UNSIGNED]
本次覆盖更新（2026-09-18）
- 文档与转录转换统一串行，导入队列按顺序推进。
- 修复批量转换中数据库空间记账误判导致队列停止的问题。
- 工作台新增本地文件选择及图片／文件粘贴入口，先暂存再提交。
- 更新内嵌前端、四平台 Core、Agent Skill 与完整审计附件。
代码：cc3d5bc。原生交互与本地专项验证边界见验收记录。此前下载过 0.0.8 的用户请重新下载对应平台文件。
0.0.8 更新
本次重点整理 Web 文件任务从上传、转换到结果预览的完整流程，保留现有页面布局和转换输出契约。
- 文档与转录共用两个上传槽位；文件独立推进，支持单项取消、重试和批量取消。
- 持久化上传接收凭据，切换页面持续上传，刷新恢复服务端已接收任务。未完成上传可重新选择文件继续提交。
- 合并轻量状态查询，区分上传、接收准备、排队、转换和结果发布，完成后主动打开结果。
- 结果窗口先显示框架，再分别读取预览与详情；预览、渲染和缓存均有大小上限。
- 历史搜索筛选与批次导航支持服务端分页；缩短磁盘 I/O 持锁时间，清理终态事件缓存与过期接收记录。
下载的 macOS 0.0.8 正式包通过混合 PPTX/XLSX、切页与刷新恢复、102 文件批次验收。重复打开已发布结果的预览 P95 为 69.5 ms，百任务状态查询 P95 为 7.7 ms；这是固定本机、带填充仓库样本的测试结果。用户原始复杂文件与长期内存压力尚未覆盖。四平台正式产物通过发布流程中的原生验收。
下载
选择与你的系统对应的 Core 压缩包：
- Windows x86_64：
into-md-windows-x86_64.zip - macOS Apple Silicon：
into-md-macos-arm64.zip - Linux x86_64：
into-md-linux-x86_64.zip - Linux ARM64：
into-md-linux-arm64.zip
Core 内置 OCR。音视频转录按需下载对应平台的 official.media.whisper-*.imp。Agent 集成使用 into-markdown-skill.zip；审计归档包含产物与验收证据。
本次沿用未做操作系统发布者签名的分发方式（[UNSIGNED]），系统可能显示信任提示。
Conversion stability update — 2026-09-18
This replacement includes automatic recovery when local file receipt or progress reads temporarily fail, keeps accepted local copies running after a lost response, and removes repeated full-history scans during task updates. The embedded Web UI is updated.
Source revision: 72535b14ed144f0119aad501afea099d9a93077e (PR #374). Version remains 0.0.8. Download the package again and fully exit the previous program before replacing it.
Into Markdown 0.0.7 [UNSIGNED]
Download one Core file for your computer and run it directly:
- Windows x86_64: into-md-windows-x86_64.zip
- macOS Apple Silicon: into-md-macos-arm64.zip
- Linux x86_64: into-md-linux-x86_64.zip
- Linux ARM64: into-md-linux-arm64.zip
OCR is already included in Core. Speech is optional; download the matching official.media.whisper .imp only if needed. Agent users download into-markdown-skill.zip. The audit ZIP is for auditors. GitHub provides source archives automatically. These builds are installable but may show an operating-system trust warning.
0.0.7 conversion and OCR validation
- Default Web and CLI behavior uses
best-effortwith OCRauto; Web resource overrides are submitted only when edited by the user. - The final arXiv retest contains 52 conversions across 26 frozen papers (OCR auto/off), all successful. The earlier fixed 100-paper, 200-conversion run is retained as separate evidence.
- The user-supplied multi-format corpus contains 2,243 readable inputs, all delivered; 42 encrypted, unreadable inputs are reported separately.
- Independent OCR evaluation recognized text in 110 of 114 text-bearing image frames (96.49%). Four missed text-bearing frames remain in the denominator; 14 independently reviewed no-text frames are excluded. All 51 outputs recognized by 0.0.6 retained their text.
- The final arXiv auto run attempted and completed OCR for all 1,123 image sources, with zero failures or skips. User-corpus raw OCR diagnostics and independently verified invisible-image exclusions remain available in the evidence.
- Source-content comparisons against the official 0.0.6 binary and Web/CLI parity checks are recorded in the acceptance evidence.
Changes: #367, #368, #369, #370, #371.
Downloaded macOS 0.0.7 artifacts passed the focused installed acceptance: native and legacy Office conversion with real OCR, the 15-case PDF recovery regression, four mask-image conversions with content checks, four Web/CLI comparisons, and speech-plugin installation plus transcription. Missing-runtime best-effort recovery and strict failure are now tested by the same assertions on all four release platforms. No long corpus rerun was required for the final test-only fixes.
Into Markdown 0.0.6 [UNSIGNED]
Download one Core file for your computer and run it directly:
- Windows x86_64: into-md-windows-x86_64.zip
- macOS Apple Silicon: into-md-macos-arm64.zip
- Linux x86_64: into-md-linux-x86_64.zip
- Linux ARM64: into-md-linux-arm64.zip
OCR is already included in Core. Speech is optional; download the matching official.media.whisper .imp only if needed. Agent users download into-markdown-skill.zip. The audit ZIP is for auditors. GitHub provides source archives automatically. These builds are installable but may show an operating-system trust warning.
Into Markdown 0.0.5 [UNSIGNED]
0.0.5 更新
- 新增 Draw.io 页面、分组和连接关系的语义转换。
- 改进 PDF 链接容错与大文档预算，完善 PPTX 结构预算及 ZIP/EPUB/RAR 兼容性。
- 统一格式准入，修复 HTML IR；改善 Markdown 样式、资产路径和空备注输出。
- 完善 OCR 共享内存预算与逐图片可选识别降级。发版前的真实语料测试修复了两处过度保守的预估：按实际 OCR 节点执行硬上限（#351），并按串行图片的临时工作空间峰值预留内存（#352）；所有保留输出仍累计，节点和内存硬上限不提高。
- Agent Skill 版本与产品同步为 0.0.5。
完整真实语料对比
同一台 Windows 主机、2,285 个输入、1,979,990,949 字节（1.844 GiB），比较已发布 0.0.4、最终 0.0.5、MarkItDown 0.1.7 与 Firecrawl AnyDoc 0.2.4。完整方法、质量评分、逐文件结果与复现脚本。
| 工具 / 模式 | 非空 | 失败 | 空 | 全量耗时（秒） | 峰值 RSS（MiB） |
|---|---|---|---|---|---|
| 0.0.4 OCR off | 1,825 | 449 | 11 | 528.9 | 1,525.0 |
| 0.0.5 OCR off | 1,827 | 447 | 11 | 509.9 | 1,531.6 |
| 0.0.4 OCR auto | 1,818 | 456 | 11 | 2,635.1 | 1,520.5 |
| 0.0.5 OCR auto | 1,819 | 455 | 11 | 3,325.8 | 1,543.0 |
| AnyDoc 0.2.4（本地，无 hosted OCR） | 1,512 | 568 | 205 | 163.6 | 12,292.4 |
| MarkItDown 0.1.7（本地，无云服务） | 1,469 | 682 | 134 | 1,644.7 | 12,377.1 |
非空不等于正文完整、OCR 正确或语音转写成功；其中包括降级或仅元数据输出。语料含损坏、加密和资源边界样本；工具覆盖范围不同，不能仅凭总耗时认定谁全面更快。超时及超过 12 GiB 后的终止和补跑均保留。质量采用独立参考的文本/结构代理、相同输入交集和 OCR 人工抽检，详见报告。
全量测试保留各版本默认逻辑内存预算：本机 0.0.4 为 4 GiB，0.0.5 无压力时约 15.9 GiB，不能解释为相同资源上限的微基准。报告另含固定 4 GiB 的代表文件复测；逻辑预算不等于实际 RSS。
原候选中 18 个基线成功输入因 documentNodes 预检失败，修复后的最终全量复测恢复了其中 17 个非空结果；原候选数据不计入上表。
时限边界：《绿野仙踪》EPUB 的 155 张图片已进入实际 OCR，但在全量扫描的 120 秒测试预算内超时；0.0.4 当时跳过了这部分 OCR。两个版本另以相同 600 秒时限复测，0.0.5 状态为 success，耗时 148.2 秒。补测不替换上表中的超时结果。
兼容性提示：未知文件扩展名不再自动按 XML 接受。包含 KML 的 ZIP 会报告成员级不支持并保留可转换成员；需转换 KML 时可提取后显式使用 --format xml。
下载与信任说明
按平台下载一个 Core ZIP，解压后直接运行。OCR 已包含在 Core 中：
- Windows x86_64：
into-md-windows-x86_64.zip - macOS Apple Silicon：
into-md-macos-arm64.zip - Linux x86_64：
into-md-linux-x86_64.zip - Linux ARM64：
into-md-linux-arm64.zip
语音转写按需下载对应平台的 official.media.whisper-*.imp。Agent 用户下载 into-markdown-skill.zip。审计、许可、来源、SBOM 与安装验收证据收录于 into-markdown-0.0.5-audit.zip。共 10 个正式下载，不另附受限制的原始语料或转换正文。
UNSIGNED：本次未提供操作系统发行者签名/公证，可能出现 Unknown Publisher、SmartScreen 或 macOS 信任提示；不要把内部完整性校验当作操作系统已验证发行者。
源码提交：527fa5281c127a93b3d712afcc403540703182e0。四平台正式构建与原生验收。
Windows Core ZIP SHA-256：a9b1e979dcaaa692eb1876d97ad6b9c2e4adb422df343b327185fa5de96d6b35。
Into Markdown 0.0.4 [UNSIGNED]
Changes
- Broader best-effort conversion for Office, legacy Office, OpenDocument, RTF, EPUB and Outlook messages, with explicit diagnostics for omitted content.
- Bounded spreadsheet paging and output transactions, shared automatic memory budgets, and lower observed memory use on the fixed regression corpus.
- Better preservation of spreadsheet display values, supported BIFF formulas, merged tables and document text order.
- PDF runtime packaging and discovery fixes, plus bounded OCR detection for scanned pages while retaining original-resolution recognition crops.
- Correct embedded-image OCR coordinate frames and preserve the original JPEG asset when harmless trailing container data is omitted from its OCR input.
- Release PDF OCR pixels after each page and reuse bounded recognition contributions for repeated pages. Stream large typed batch reports through accounted temporary storage without discarding diagnostics.
- Preserve native PDF content when optional page OCR cannot fit its preflight budget or an image placement cannot provide reliable page coordinates. Avoid node-ID collisions when OCR is merged with previously recovered tables and other retained layout blocks.
Conversions still enforce integrity and resource limits. Best-effort output is not a guarantee of lossless layout: unsupported objects and formulas are diagnosed, and legacy DOC/PPT/RTF/XLSB compatibility gaps remain. OCR quality varies with source resolution, language, small text and page layout.
Downloads
Download one Core file for your computer and run it directly:
- Windows x86_64:
into-md-windows-x86_64.zip - macOS Apple Silicon:
into-md-macos-arm64.zip - Linux x86_64:
into-md-linux-x86_64.zip - Linux ARM64:
into-md-linux-arm64.zip
OCR is already included in Core. Speech is optional; download the matching official.media.whisper .imp only if needed. Agent users download into-markdown-skill.zip. The audit ZIP is for auditors. GitHub provides source archives automatically.
These builds are unsigned and may show an operating-system trust warning.
Into Markdown 0.0.3 [UNSIGNED]
Download one Core file for your computer and run it directly:
- Windows x86_64: into-md-windows-x86_64.zip
- macOS Apple Silicon: into-md-macos-arm64.zip
- Linux x86_64: into-md-linux-x86_64.zip
- Linux ARM64: into-md-linux-arm64.zip
OCR is already included in Core. Speech is optional; download the matching official.media.whisper .imp only if needed. Agent users download into-markdown-skill.zip. The audit ZIP is for auditors. GitHub provides source archives automatically. These builds are installable but may show an operating-system trust warning.
Into Markdown 0.0.2 [UNSIGNED]
Download exactly one Core package for your computer:
- Windows x86_64: into-md-windows-x86_64-core.zip
- macOS Apple Silicon: into-md-macos-arm64-core.dmg
- Linux x86_64: into-md-linux-x86_64-core.tar.gz
- Linux ARM64: into-md-linux-arm64-core.tar.gz
OCR is built into Core and is ready after installation. Speech transcription is optional; only then download the official.media.whisper .imp with the same platform suffix. The single release-evidence ZIP contains checksums and audit reports and is not needed by ordinary users. This unsigned build may show an operating-system trust warning.
Into Markdown 0.0.1 [UNSIGNED]
普通用户下载指南
只看与你电脑平台对应的一行。仅使用基础文档转换时下载 Core；需要完整功能时，再下载同一行的 OCR 和 语音 两个 .imp。
Windows 64 位（大多数 Windows 电脑）
macOS Apple 芯片（M1/M2/M3/M4/M5）
Linux x86_64（Intel/AMD）
Linux ARM64
每个运行时文件旁的 .sha256 用于下载校验。其余 SBOM、来源、NOTICE、平台审计、installed-smoke 和 release-set 已统一收进 release-evidence.zip，普通用户不需要下载。
签名说明
本次 OS 分发包可安装但未做外部代码签名。请优先核对对应 .sha256；Windows 可能显示 Unknown publisher/SmartScreen，macOS 可能需要在系统设置中选择“仍要打开”。OCR 和语音 .imp 保留内部 Ed25519 manifest 签名及固定 SHA-256。
发布内容只有 Core、OCR 和语音，不包含 LibreOffice 插件。
Into Markdown runtime assets
Build-time pinned runtime assets for deterministic Into Markdown builds. This pre-release is not an end-user product download; use a numbered release for application binaries.
FFmpeg 8.1.2 provenance:
- Audited build commit:
af8cca88c8997f2dec7c6f030e2fdbf7ebc731c2 - Build workflow: https://github.com/coolplayagent/into-markdown/actions/runs/33191296093
- Source: https://ffmpeg.org/releases/ffmpeg-8.1.2.tar.xz
- Source SHA-256:
464beb5e7bf0c311e68b45ae2f04e9cc2af88851abb4082231742a74d97b524c - Signature SHA-256:
0a0963fccd70597838073f3e31b20f4a4d8cc2b5e577472c9a5a1f22624246f8 - Upstream signing fingerprint:
FCF986EA15E6E293A5644F10B4322F04D67658D8
Pinned runtime ZIP SHA-256:
- macOS ARM64:
1c14a5eb88dbff1f1462f7588a7ba967163782a67edae8b5e242d18a737e9030 - Linux ARM64:
1023427cef602bfc144e73602867596bca0fa9440baccae3733a199b879915c3 - Windows x86_64:
a7f3c0049d46f2b8557eb081423d78bd2569eaa57028b3e1314a8782abc96c1e - Linux x86_64:
dd3bd7320098824dc2a5aafbb16009d54a3776e96da84c27b6d70e56c0bb64e0