scholar-deep-research — From Question to Cited Report

English · 中文 · 📖 Online Docs

Ask an LLM to "write a literature review on X" and you get three failure modes: citations that don't exist, the canonical paper missing because the search ran once and stopped, and no record of why any paper made the cut. scholar-deep-research fixes those by running an 8-phase (Phase 0..7), script-driven workflow with enforced citation anchoring, multi-round saturation gating, and a per-paper audit trail.

Zero LLM calls inside the pipeline. Every script under scripts/ is pure data — search, dedupe, rank, citation-chase, bibliography export. The host LLM orchestrates from outside via JSON envelopes on stdout. That separation buys reproducibility, auditability, and a 343-test smoke suite that runs in ~13 s with no API keys.

Works with Claude Code, Cursor, Codex, OpenCode, OpenClaw / ClawHub, Hermes Agent, pi-mono, and SkillsMP — any agent that supports the Agent Skills format.

See it in action

Three end-to-end runs committed verbatim with full audit trail (state, evidence, report, BibTeX, run notes):

• GLP-1 for Non-Diabetic Obesity — systematic_review, 833 papers, 5 deep reads, 12-citation report on efficacy / safety / discontinuation.
• Mamba vs Transformer for Long Context — comparative_analysis, 1005 papers, 3 deep reads (Mamba, Jamba, xLSTM), 5-axis verdict.
• AAV Capsids for CNS Gene Therapy — grant_background, 681 papers, 4 mechanistic deep reads (LY6A, CAP-B10, single-residue BBB, LRP6).

What's different from "just ask the LLM"

Full feature matrix in docs/COMPARISON.md.

How it works (1-line per phase)

Phase 0  Scope        question decomposition + archetype + state init
Phase 1  Discovery    multi-source search → dedupe → multi-axis saturation check
Phase 2  Triage       ranking → top-N selection → tier triage → optional PDF prefetch
Phase 3  Deep read    parallel agent fan-out (deep tier) + abstract stub (skim tier)
Phase 4  Chasing      citation graph (forward + backward, OpenAlex + S2)
Phase 5  Synthesis    thematic clustering → tension map
Phase 6  Self-critique  14-point adversarial checklist (mandatory)
Phase 7  Report       render archetype template → export bibliography

Full diagram, gate semantics, and state model in docs/ARCHITECTURE.md.

Quick Start

# Any agent
npx skills add Agents365-ai/365-skills -g

# Claude Code only
> /plugin marketplace add Agents365-ai/365-skills
> /plugin install scholar-deep-research

Then pip install -r requirements.txt inside the install dir.

Once installed, just describe what you want:

Run a deep research report on CRISPR base editing for Duchenne muscular dystrophy.

The skill walks the 8 phases automatically and writes the report to reports/<slug>_<YYYYMMDD>.md with a matching .bib. Full install details in docs/INSTALL.md; phase-by-phase walkthrough in docs/WALKTHROUGH.md.

What this doesn't do

• No Google Scholar / Web of Science / Scopus — no public API; cite as "not consulted" in the report appendix if your topic needs them.
• No automatic full-text for paywalled papers — open-access only; gated papers fall back to landing-page abstracts.
• No semantic re-ranking out of the box — relevance is bag-of-words; plug an embedding model into state.papers[*].score_components.relevance if you need it.
• Saturation is novelty-based, not exhaustiveness-based — catches "exploration has stalled", not "every relevant paper found". Use systematic_review + SCHOLAR_SATURATION_NEW_PCT=20 for stricter coverage.

Full list in docs/LIMITATIONS.md.

Documentation

🔗 Related Skills

Pick the right tool for the research workflow you're running:

❤️ Support

If this skill helps you, consider supporting the author:

👤 Author

Agents365-ai

• GitHub: https://github.com/Agents365-ai
• Bilibili: https://space.bilibili.com/441831884

📄 License

MIT