ingest

Pass the arXiv URL as the source argument. The skill extracts the arXiv ID, downloads the TeX source via init_discovery.py, enriches metadata from Semantic Scholar, then creates the paper page plus linked concept, claim, and people pages with graph edges.

Provide the local .pdf path as the source. The skill runs a preprocessing pipeline that recovers the title and arXiv ID, fetches original TeX source when available, or builds a synthetic .tex fallback under raw/tmp/ before ingesting.

Yes. Before creating any concept or claim it calls find-similar-concept or find-similar-claim and merges into the top match when the similarity score is 0.80 or higher. Per-paper creation limits cap new entities based on paper importance.

The ingest degrades gracefully: it skips S2 enrichment, defaults the paper importance to 3, skips citation backfill, and notes the provisional importance in the final report. The paper page is still created.

Yes. In INIT MODE the parent /init runs each ingest in an isolated git worktree. Shared files like edges.jsonl, citations.jsonl, index.md, and log.md are written only through append-only tool commands, and slugs are allocated deterministically.

Avoid creating a concept when a similar one scores above 0.40 without a clear technical distinction, when the candidate matches a foundation (link it as derived_from instead), or when the per-paper creation limit is reached. Merge or defer to /check instead.