Skip to content

Releases

v0.2.2

Choose a tag to compare

@github-actions github-actions released this 31 Oct 13:02

0.2.2 validation

Primary workflow: HTML → translated HTML. No browser or PDF renderer is used for HTML jobs.

Local validation: TypeScript typecheck and build pass; 15 Node tests pass; Python suite has 53 passing tests and 3 skipped optional real-PDF tests (56 total).

HTML coverage includes formula-node preservation (MathML), references, appendix, translated captions and table cells, inline formatting, embedded local images, input immutability, placeholder loss rejection, script removal, escaped translation markup, unsupported-prose detection, source staleness and URL/path restrictions. SDK integration exercises prepare/read/submit/export/status using the installed 0.2.1-rc.3 SDK. GitHub release workflow additionally tests 0.2.2-alpha.2; its result must be checked separately.

These are synthetic and interface tests, not full-paper translation-quality acceptance. DDPM full HTML translation has not been validated. Structural checks do not guarantee semantic accuracy, browser visual fidelity, or universal website compatibility.

Legacy PDF extraction remains experimental. The previously supplied DDPM PDF still has unresolved boundary issues; this release does not claim to fix general PDF formula detection. Optional real-PDF tests require DDPM_FIXTURE_PATH and do not bundle the user's paper.

Environment regression tests: launch the real worker with Python -S to simulate missing site-packages, verify all missing packages and the actual executable are returned before any job directory is created; verify PowerShell/POSIX interpreter-path quoting, nonexistent interpreters and nonzero interpreter exits. Windows Store alias behavior is diagnosed by launch/exit/protocol errors; no Windows machine has been exercised in this local validation.

Encoding regression: real worker subprocesses run under forced cp936, cp932, cp949 and cp1251 pipe encodings with PYTHONUTF8=0. Tests preserve Chinese, Japanese, Korean, Cyrillic and emoji through prepare/read/submit/export/status, retain formula nodes, verify output files at exact Unicode paths, and preserve Unicode dependency-error fields. Bridge tests reject invalid UTF-8 inside valid-looking JSON, verify the subprocess encoding, and preserve multibyte output across writes. These checks passed locally on Linux. A new Windows CI job executes the worker encoding and environment tests; its result is separate from local validation.

v0.2.1

Choose a tag to compare

@github-actions github-actions released this 31 Oct 12:31

0.2.1 validation

Primary workflow: HTML → translated HTML. No browser or PDF renderer is used for HTML jobs.

Local validation: TypeScript typecheck and build pass; 12 Node tests pass; Python suite has 51 passing tests and 3 skipped optional real-PDF tests (54 total).

HTML coverage includes formula-node preservation (MathML), references, appendix, translated captions and table cells, inline formatting, embedded local images, input immutability, placeholder loss rejection, script removal, escaped translation markup, unsupported-prose detection, source staleness and URL/path restrictions. SDK integration exercises prepare/read/submit/export/status using the installed 0.2.1-rc.3 SDK. GitHub release workflow additionally tests 0.2.2-alpha.2; its result must be checked separately.

These are synthetic and interface tests, not full-paper translation-quality acceptance. DDPM full HTML translation has not been validated. Structural checks do not guarantee semantic accuracy, browser visual fidelity, or universal website compatibility.

Legacy PDF extraction remains experimental. The previously supplied DDPM PDF still has unresolved boundary issues; this release does not claim to fix general PDF formula detection. Optional real-PDF tests require DDPM_FIXTURE_PATH and do not bundle the user's paper.

Environment regression tests: launch the real worker with Python -S to simulate missing site-packages, verify all missing packages and the actual executable are returned before any job directory is created; verify PowerShell/POSIX interpreter-path quoting, nonexistent interpreters and nonzero interpreter exits. Windows Store alias behavior is diagnosed by launch/exit/protocol errors; no Windows machine has been exercised in this local validation.

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 29 Oct 14:03

0.2.0 validation

Primary workflow: HTML → translated HTML. No browser or PDF renderer is used for HTML jobs.

Local validation: TypeScript typecheck and build pass; 11 Node tests pass; Python suite has 49 passing tests and 3 skipped optional real-PDF tests (52 total).

HTML coverage includes formula-node preservation (MathML), references, appendix, translated captions and table cells, inline formatting, embedded local images, input immutability, placeholder loss rejection, script removal, escaped translation markup, unsupported-prose detection, source staleness and URL/path restrictions. SDK integration exercises prepare/read/submit/export/status using the installed 0.2.1-rc.3 SDK. GitHub release workflow additionally tests 0.2.2-alpha.2; its result must be checked separately.

These are synthetic and interface tests, not full-paper translation-quality acceptance. DDPM full HTML translation has not been validated. Structural checks do not guarantee semantic accuracy, browser visual fidelity, or universal website compatibility.

Legacy PDF extraction remains experimental. The previously supplied DDPM PDF still has unresolved boundary issues; this release does not claim to fix general PDF formula detection. Optional real-PDF tests require DDPM_FIXTURE_PATH and do not bundle the user's paper.

v0.1.1

Choose a tag to compare

@github-actions github-actions released this 29 Oct 07:29

0.1.1 兼容性补充(2026-10-29)

  • 目标:支持用户的 DSH Desktop 0.2.1-rc.3;原 0.1.0 的 peer 范围错误地排除了它。
  • 开发依赖切换为 dsh-tools 0.2.1-rc.3 和其要求的 Cordis ~4.0.5;peer 范围声明 ^0.2.1-rc.3 || ^0.2.2-alpha.2。
  • 原 SDK 将注册器导出为 ToolRuntime,新 SDK 导出为 ToolsService;仅测试适配该命名差异,生产入口使用的 defineTool / ctx.tools.register 无需修改。
  • 0.2.1-rc.3 隔离环境:构建、类型检查、10 项 Node 测试和 21 项 Python 测试通过。使用真实目标 SDK 注册五工具并执行完整离线闭环。
  • GitHub CI 增加 0.2.1-rc.3 / 0.2.2-alpha.2 两组 SDK 验证,发布任务须等两组成功。
  • 测试运行在 Linux,尚未验证用户 Windows Desktop 安装、实际模型会话或完整真实论文;不能将 SDK 测试等同于客户端完整验收。
  • GitHub Release 提供 0.1.1 安装包;npm 0.1.1 需要维护者另行发布,不能用 README 或市场索引修改替代 npm 发版。

验证报告

验证日期:2026-10-28。版本:0.1.0;解析器版本:1.0.1。

环境与结果

Node.js 24.22.1、Python 3.12.14、PyMuPDF 1.26.6;Harness SDK 使用 package-lock.json 锁定的依赖。

检查 结果
npm run build 通过,生成 lib/
npm run typecheck 通过
npm test 10 项通过:8 项进程桥接、2 项真实 SDK 注册测试
python3 -m unittest discover -s tests -p 'test_*.py' 21 项通过
python3 examples/generate_demo.py 合成英文论文到中文 PDF 闭环通过

测试输入均为代码生成的人工合成 PDF;没有使用真实研究论文,测试数量不代表真实论文样本数。

已验证行为

  • 通过真实 Cordis Context / ToolsService 注册五个工具,并执行 prepare → 分批 read → submit → export → status,校验 SDK 输入输出 schema 与渲染。接口测试提交英文原文,只验证协议,不作为翻译质量证据。
  • 人工编写中文译文的单页合成论文包含标题、摘要、正文、引用、矢量图及图注和参考文献。导出可重新打开,逐页可渲染,中文可提取,无未恢复占位符。已目视检查 examples/preview.png,未见重叠、缺字、裁切或丢图。
  • 样例图表及图注区域保持 320 × 125.74 PDF 点,源/目标渲染平均绝对通道差为 0;实际数据见 examples/validation.json。源文件哈希不变。
  • 回归覆盖长中文排版增页、图表外原文不残留、行内公式/引用恢复、缺译拒绝、整批占位符校验原子性、重复与过期提交、中断恢复、Markdown 关闭、输入路径授权边界、页数和大小上限。
  • 双栏参考文献按左栏后右栏排序,跨栏参考文献标题作为区段边界;作者/机构等前置信息保留。正文 URL/DOI 不再自动归为元数据。
  • 同一段落多条脚注保留顺序与关联;脚注正文、引用和标记分别保护。缺失或歧义锚点停止处理。
  • 扫描及混合扫描合成页返回停止状态与准确页码,停止任务禁止读块、提交和导出。
  • 进程桥接覆盖依赖缺失、取消、超时、请求和响应上限、无效 JSON,路径中的 shell 字符不执行。

支持边界与尚未完成的验收

这是保守的规则解析首版,不是任意论文通用的布局模型。测试验证了选定合成版式,不证明未知复杂布局都能可靠识别;OCR、复杂数学边界、多栏浮动框、表格/图注边界不明确时须停止。跨页段落保留独立块,不猜测合并。

尚未在用户实际 Harness 安装及模型会话中验收,也未使用完整真实论文完成模型翻译。真实论文的无框表格、跨页表格、复杂公式、整页图与更多字体仍需要样本验收。离线 SDK 闭环不能替代这些项目。

目标语言测试覆盖中文 zh 与英文 en;其他字体支持语言未做完整验收,阿拉伯语等复杂塑形语言会明确拒绝。结构和字符覆盖校验不能证明译文语义准确;中文样例仅抽查了人工译文,未测模型翻译准确率。

复现

按 README 安装 Python 依赖和开发依赖后,依次执行上表命令。examples/generate_demo.py 重新生成源样例、中文输出、原文 Markdown、验证 JSON 和预览。测试产生的临时任务会自动清理。源码、构建结果、测试、配置及本报告随安装包提供。

v0.1.0

Choose a tag to compare

@github-actions github-actions released this 28 Oct 14:53

验证报告

验证日期:2026-10-28。版本:0.1.0;解析器版本:1.0.1。

环境与结果

Node.js 24.22.1、Python 3.12.14、PyMuPDF 1.26.6;Harness SDK 使用 package-lock.json 锁定的依赖。

检查 结果
npm run build 通过,生成 lib/
npm run typecheck 通过
npm test 10 项通过:8 项进程桥接、2 项真实 SDK 注册测试
python3 -m unittest discover -s tests -p 'test_*.py' 21 项通过
python3 examples/generate_demo.py 合成英文论文到中文 PDF 闭环通过

测试输入均为代码生成的人工合成 PDF;没有使用真实研究论文,测试数量不代表真实论文样本数。

已验证行为

  • 通过真实 Cordis Context / ToolsService 注册五个工具,并执行 prepare → 分批 read → submit → export → status,校验 SDK 输入输出 schema 与渲染。接口测试提交英文原文,只验证协议,不作为翻译质量证据。
  • 人工编写中文译文的单页合成论文包含标题、摘要、正文、引用、矢量图及图注和参考文献。导出可重新打开,逐页可渲染,中文可提取,无未恢复占位符。已目视检查 examples/preview.png,未见重叠、缺字、裁切或丢图。
  • 样例图表及图注区域保持 320 × 125.74 PDF 点,源/目标渲染平均绝对通道差为 0;实际数据见 examples/validation.json。源文件哈希不变。
  • 回归覆盖长中文排版增页、图表外原文不残留、行内公式/引用恢复、缺译拒绝、整批占位符校验原子性、重复与过期提交、中断恢复、Markdown 关闭、输入路径授权边界、页数和大小上限。
  • 双栏参考文献按左栏后右栏排序,跨栏参考文献标题作为区段边界;作者/机构等前置信息保留。正文 URL/DOI 不再自动归为元数据。
  • 同一段落多条脚注保留顺序与关联;脚注正文、引用和标记分别保护。缺失或歧义锚点停止处理。
  • 扫描及混合扫描合成页返回停止状态与准确页码,停止任务禁止读块、提交和导出。
  • 进程桥接覆盖依赖缺失、取消、超时、请求和响应上限、无效 JSON,路径中的 shell 字符不执行。

支持边界与尚未完成的验收

这是保守的规则解析首版,不是任意论文通用的布局模型。测试验证了选定合成版式,不证明未知复杂布局都能可靠识别;OCR、复杂数学边界、多栏浮动框、表格/图注边界不明确时须停止。跨页段落保留独立块,不猜测合并。

尚未在用户实际 Harness 安装及模型会话中验收,也未使用完整真实论文完成模型翻译。真实论文的无框表格、跨页表格、复杂公式、整页图与更多字体仍需要样本验收。离线 SDK 闭环不能替代这些项目。

目标语言测试覆盖中文 zh 与英文 en;其他字体支持语言未做完整验收,阿拉伯语等复杂塑形语言会明确拒绝。结构和字符覆盖校验不能证明译文语义准确;中文样例仅抽查了人工译文,未测模型翻译准确率。

复现

按 README 安装 Python 依赖和开发依赖后,依次执行上表命令。examples/generate_demo.py 重新生成源样例、中文输出、原文 Markdown、验证 JSON 和预览。测试产生的临时任务会自动清理。源码、构建结果、测试、配置及本报告随安装包提供。