lit-summarize

name	lit-summarize
description	Summarize research papers (PDF or arXiv) into structured markdown (problem / method / results / contribution / limitations) and index each summary into the AgentDB `papers/` namespace for later semantic search. Use when the user drops PDFs into `inputs/papers/` or pastes an arXiv URL.

lit-summarize

STATUS: active. research_assistant.lit.extract_pdf_text / fetch_arxiv / parse_metadata are implemented; Claude generates the structured summary in-skill.

When to use

• User mentions a PDF path under inputs/papers/.
• User pastes an arXiv URL or DOI.
• User says "summarize this paper", "what's the key idea of X", "intake these papers".

Workflow

1. Resolve sources — collect every PDF in inputs/papers/ not yet summarized, plus any arXiv URL / DOI the user named.
2. Extract text — call research_assistant.lit.extract_pdf_text (PyMuPDF).
3. Pull canonical BibTeX (when input is a DOI or the PDF has a DOI in metadata) — call research_assistant.lit.publisher_bibtex.fetch_bibtex_from_publisher(doi). It tries CrossRef → doi.org content negotiation → CloakBrowser stealth fallback (last tier needs pip install -e ".[crawl]"). Stash the returned BibTeX string in the AgentDB record's metadata.bibtex field (step 6) so /cite can resolve \cite{slug} from it without a second network round-trip. If no DOI is known, skip — /cite re-fetches on demand.
4. Summarize — Claude produces a structured markdown with sections: Problem, Method, Results, Contribution, Limitations, Related work pointers.
5. Persist — write to outputs/summaries/<slug>.md.
6. Index — store the summary into AgentDB:
   • tool: mcp__claude-flow__memory_store
   • namespace: papers
   • key: paper slug
   • value: the summary markdown (vector-indexed automatically).
   • metadata: {"doi": <doi-or-null>, "arxiv_id": <id-or-null>, "bibtex": <entry-string-or-null>} — when step 3 produced a BibTeX entry, put it in metadata.bibtex so /cite reads it back instead of re-fetching.

Outputs

• outputs/summaries/<slug>.md (one file per paper).
• AgentDB entry in namespace papers/ keyed by slug.

Memory keys touched

• papers/<slug> — write
• papers/* — read (to avoid re-summarizing what's already indexed)

Open enhancements

All previously-tracked items deferred indefinitely. Re-open here when slug-format / GROBID / structured-schema decisions are needed.