# Validation scope

This public repository contains three skill packages and wholly synthetic course-tool fixtures. It contains no real course media, transcripts, or personal learning records. An example learner profile in a fixture is synthetic, not a record about a real person.

## Reproducible checks

Run from the repository root:

```sh
python3 -m pip install -r requirements-dev.txt
python3 scripts/check_package.py
python3 -m unittest discover -s tests/course-to-notes -v
```

- **Package validation:** skill names and descriptions, UI metadata, invocation policy, bundled licenses, and local Markdown link targets outside fenced examples. This checks packaging, not model behavior.
- **Subtitle parsing:** repeated text, overlapping cues, multiline cues, timestamps, invalid inputs, and refusal to overwrite existing output. Supports UTF-8 SRT and WebVTT only.
- **Frame extraction:** a synthetic red/blue video, requested seek times, output pixel checks, invalid times, missing dependencies, and overwrite protection. Requires FFmpeg/ffprobe; missing tools cause an explicit skip. Requested timestamps are not measured presentation timestamps.
- **Output validation:** coverage gaps, section identifiers, missing assets, learner-evidence references, stale annotations, section/index markers, manual-edit hash warnings, and invalid data. The validator does not modify the course files.

The initial publication was locally checked on macOS with Python 3.13.14 and FFmpeg/ffprobe 9.0.1: all 28 course-tool tests passed with no skips. All three packages also passed the locally available Codex skill-creator structural validator. The repository checker is self-contained apart from its PyYAML development dependency.

The GitHub Actions workflow runs package checks and the same tests on Ubuntu with Python 3.10 and 3.13. Consult the actual [workflow runs](https://github.com/XSY-28/learning-skills/actions) for their status; configuration alone is not a successful run.

## DSH plugin checks

The DSH package was locally checked on macOS with Node.js 24.19.0, DSH skill registry 0.1.7-rc.2, and Cordis 4.0.4. All four Node tests passed:

- Discover, load, and render all three skills through the real DSH registry, including their referenced files.
- Remove skills on plugin disposal and restore them on remount.
- Preserve a same-name runtime customization and restore the bundled default when that customization is removed.
- Pack and extract the plugin into a new directory, load it without a dependency symlink, and run the packaged course helper's CLI help.

The actual DSH CLI also installed the tarball in an isolated temporary profile. Its composed configuration contained the plugin row, and the installed provider loaded all three skills through the real registry. No user profile or model request was used. The runtime adapter imports only Node built-ins; its optional DSH peer documents the tested host API without requiring a second host SDK inside the installed package.

To reproduce the repository checks:

```sh
python3 scripts/build_dsh_catalog.py --check
npm ci --ignore-scripts
npm test
```

The catalog is generated from the shared skill headers; after changing a header, run `python3 scripts/build_dsh_catalog.py` and commit the updated catalog. PyYAML is a development dependency, not a plugin runtime dependency. The package explicitly includes required resource types and excludes Python caches.

These checks establish installation and instruction/resource loading, not a Web UI conversation, reliable model selection, teaching quality, or whole-course processing. The DSH CI job executes the registry and packed-artifact tests; its actual result must be read from the workflow run.

## Not established by this release

These tests do not establish:

- Reliable automatic skill routing in a new client session.
- Model-generated explanation quality, factual correctness, or actual learner mastery.
- Whole-lecture or whole-course coverage and long-running continuation.
- YouTube/Bilibili acquisition, real audio transcription, or reconstruction of unclear audience questions.
- Accurate recognition of a real learner's history.
- Word/PDF/HTML export or Obsidian/Microsoft Word application rendering.
- Cross-client or cross-platform compatibility beyond checks actually reported by CI.

The two instruction-only skills have packaging checks in this release, not a newly executed teaching-effectiveness study. The course skill's [behavioral acceptance cases](../skills/course-to-notes/references/acceptance-cases.md) define useful future evaluation tasks; their presence does not mean they all passed.

## 中文说明

本次开源验证聚焦可复现的脚本与打包：本地 28 项测试全部通过、无跳过，三个技能通过结构检查。合成视频和合成档案只验证对应脚本行为，不能推广为真实课程端到端效果或真实学习成效。

仓库不包含课程原视频、字幕、截图或私人学习记录。自动触发、整讲/整课覆盖、真实 ASR、模糊问答还原、真实学习档案判断、Word/PDF/HTML 导出及 Obsidian/Microsoft Word 应用内显示，均不属于本次发布已验证能力。CI 的实际结果以工作流页面为准。