Skill: source-ingestion

Purpose: Execute Phase 1.5 — convert approved source files into canonical Markdown with assets, anchors, metadata, reports, and manifest updates.

---

Input

• Source file path or URL
• Target subagent slug
• Optional metadata overrides: title, author, year, rights_status, authority, volatility

---

Steps (in order)

1. Preserve original — copy to sources/original/<source_id>/original.<ext>
2. Compute SHA-256 — via generate_metadata.py
3. Detect file type — via detect_file_type.py
4. Choose converter — PDF→Docling/MarkItDown, ePUB→Pandoc/MarkItDown, DOCX→Pandoc/MarkItDown, HTML→readability/Pandoc, Markdown→passthrough
5. Convert to canonical Markdown — write to sources/markdown/<source_id>.md
6. Extract assets — base64 data URIs → sources/assets/<source_id>/
7. Inject source anchors — headings, tables, figures, code blocks → HTML comments + JSONL
8. Generate metadata JSON — sources/metadata/<source_id>.metadata.json
9. Generate conversion report — sources/reports/<source_id>.conversion-report.md
10. Assign conversion status — ok | needs-human-review | needs-ocr | failed
11. Update source-pack manifest — source-pack.manifest.yaml
12. Create human-review queue entry — when status != ok → append to sources/reports/human-review-queue.md

---

Conversion status rules

Condition	Status
No errors, content extracted	ok
Scanned/image-only PDF detected	needs-ocr
URL requires login	needs-auth
Converter returned errors	failed
Low content yield, possible quality issue	needs-human-review

---

Output

All files created under subagents/<slug>/sources/.

Returns:

• source_id
• conversion_status
• anchor_count
• asset_count
• markdown_path