Scholar megasearch

scholar-megasearch

Integrates every academic search MCP/skill in this environment into one fan-out → merge → synthesize pipeline. Each subagent owns one source bucket and searches in parallel; results are merged into a single deduplicated, provenance-tracked, ranked corpus. Prefer this over single-source searches whenever breadth matters.

Works in both Claude Code and Codex. Use the host's native tool discovery when MCP schemas are deferred: Claude Code may expose ToolSearch; Codex may expose tool_search. Use the skill directory from the loaded skill path for bundled scripts. Default install locations:

• Claude Code: ~/.claude/skills/scholar-megasearch, venv ~/.claude/skill_venv.
• Codex: ~/.agents/skills/scholar-megasearch, venv ${CODEX_HOME:-~/.codex}/skill_venv.

Core engines expected when fully installed:

• MCP servers: arxiv-mcp-server, asta, and paper-search-mcp.
• Local fallbacks: scripts/search_local.py {arxiv|semanticscholar|ddg} and scripts/resilient_search.py, plus scripts/fetch_pdfs.py.

Source buckets A-G:

• A arXiv; B Semantic Scholar via Ai2 Asta; C Crossref + OpenAlex.
• D PubMed/PMC/bioRxiv/medRxiv/Europe PMC; E DOAJ/CORE/BASE/OpenAIRE/Zenodo/Unpaywall/HAL.
• F DBLP/IACR/CiteSeerX/SSRN; G web, GitHub, grey literature, and page scraping.

For the full source list and which tools live in each bucket, read references/sources.md. For the orchestration templates and the record schema, read references/orchestration.md.

Workflow

1. Frame the query + pick a depth level

Restate the topic in one line. If it is underspecified (e.g. "find papers on neural networks"), do a mini survey before fanning out. Ask for exactly:

• Field: e.g. cs-ml, biomed, physics, chem-materials, crypto-security, econ-social-law, math, or interdisciplinary.
• Goal: survey, systematic, newest, seminal, implementation, or pdf-corpus.
• Depth: numeric 1–5 only.

If the user already gave these, do not ask. Otherwise ask once, then continue. For terminal planning or repeatable runs, generate the same plan with:

python3 <skill-dir>/scripts/plan_run.py "<topic>" --field cs-ml --goal survey --depth 3

Depth sets the facet count, bucket count, per-source hit cap, and how many waves run. An explicit depth=N / LN / bare 1–5 in the request wins; otherwise use the user's numeric mini-survey answer; otherwise default L2 only when the user asks not to be asked.

2. Decompose into facets + route to buckets

• Facets (count set by the depth level, 3–8): synonyms, sub-aspects, method vs. phenomenon, key authors, and at least one each of a broad and a narrow phrasing. For topics with strong non-English literature, add a localized query in the relevant language for Bucket G.
• Buckets (count set by the depth level, 4–7): pick from the domain→bucket routing table in references/sources.md based on the topic's field. Default for unknown/interdisciplinary: A, B, C, D, E, G.

3. Set up the run directory

Create ./literature_search/<slug>_<YYYY-MM-DD>/raw/ under the current working directory (slug = short kebab of the topic; use today's date). All artifacts go here.

4. Fan out the searchers (wave 1)

• If the user opted into workflows ("workflow" keyword / ultracode): run the Workflow script in references/orchestration.md, passing {topic, facets, buckets, cap} as args (cap = the level's hits/subquery). Then write each returned raw[i] to raw/<bucket>.json.
• Otherwise: spawn one Agent per bucket in a single message (concurrent), each writing its own raw/<bucket>.json. Use the Agent prompt skeleton in references/orchestration.md.

Every searcher returns records in the schema (title, authors, year, doi, arxiv_id, pdf_url, url, citations, abstract, source, query) and does NOT dedupe. This is wave 1; L3+ add further waves after the first merge — see ## Depth levels.

5. Merge into one corpus

Dedupes by DOI → arXiv-id → normalized title, merges duplicates (keeping the richest fields + max citations), then ranks with the five-layer scorer described below. Pass the goal and topic so the relevance/weight profile is aligned with the mini survey:

python3 <skill-dir>/scripts/merge_corpus.py \
  ./literature_search/<slug>_<date>/raw \
  -o ./literature_search/<slug>_<date>/corpus.json \
  --md ./literature_search/<slug>_<date>/corpus.md \
  --goal <goal> --topic "<topic>"

corpus.md is the human-readable digest. Use --min-sources 2 to keep only papers corroborated by ≥2 databases (high-precision shortlist). Use --ranking classic only when reproducing old runs.

6. Synthesize

Read corpus.json and write summary.md in the run dir:

• Headline count (unique papers, sources hit, year span).
• Top ~15–25 papers grouped by sub-theme, each with a one-line "why it matters".
• Number every paper by its corpus.json rank (1-based) shown as [#NN]. That same number is the NN_ prefix of the acquired pdfs/NN_*.pdf and the rank/i in pdfs/manifest.json, so a reader jumps from a summary [#NN] straight to its file.
• Seminal/most-cited works, recent frontier (last 2 yrs), and notable gaps.
• Cite by DOI/arXiv id. Report honestly what was searched and any source that failed — no fabricated entries (see memory feedback_honest_writing).

7. Acquire original PDFs

Pull the original PDFs for the depth level's count — L1 → top 10, L2 → 30, L3 → 50, L4 → 100, L5 → all (every paper in the corpus). --top all (or 0) takes the whole corpus; files are saved as NN_<slug>.pdf by corpus.json rank, matching the [#NN] in summary.md:

python3 <skill-dir>/scripts/fetch_pdfs.py \
  ./literature_search/<slug>_<date>/corpus.json \
  -o ./literature_search/<slug>_<date>/pdfs \
  --email [email protected] --top 30

This auto-acquires via the free/legal routes — known open-access pdf_url, arXiv direct, then Unpaywall OA API — verifying each file is a real PDF, and writes pdfs/manifest.json. Papers with no free route are flagged "status": "needs_mcp". For those, fetch via the session MCP download tools (paper-search-mcp. download_with_fallback, source-specific download_*, or download_scihub) — a standalone script cannot reach MCP. To read extracted full text afterward, use the read_*_paper MCP tools or pdfplumber/pymupdf from the installed host venv. See references/sources.md for the full acquisition tool list.

Depth levels (L1–L5)

One knob: breadth (facets × buckets × hits) and recursion (extra waves) scale together. Pick one per run — explicit depth=N / LN / bare 1–5 wins; else infer from phrasing; else default L2. Clamp out-of-range to 1–5, and state the level you ran at.

Lvl	facets	buckets	hits/subq	waves	PDFs	output
L1 Quick	3	4	15	wave 1 only	top 10	corpus
L2 Standard (default)	5	5	25	wave 1 only	top 30	corpus
L3 Deep	6	6	30	+ citation-snowball	top 50	corpus
L4 Exhaustive	8	7 (all)	40	+ snowball + 1 completeness-critic pass	top 100	corpus + ≥2 shortlist
L5 Total / Exhaustive	8	7 (all)	40	+ snowball + critic loop-until-dry	all	corpus + ≥2 shortlist

Phrasing → level when not explicit: quick·first look·taste → L1 · (no signal) → L2 · deep·snowball·trace citations → L3 · systematic review·comprehensive·thorough → L4 · exhaustive·every source·all of them·to the end → L5. Equivalent phrases in other languages map the same way. Higher levels spawn more subagents and cost more tokens (L5 is bounded only by the token budget, not a fixed wave count).

Five-layer ranking

merge_corpus.py ranks each merged paper with five orthogonal layers. Each layer is normalized to 0–1 and written to rank_layers; the weighted total is score.

1. Provenance: independent source agreement (sources_count), not citation-based.
2. Impact: citation count plus age-normalized citation velocity.
3. Recency: publication-year frontier signal, independent of citations.
4. Access/completeness: DOI/arXiv id, PDF URL, abstract, authors, venue/year/url.
5. Relevance: overlap between topic/query terms and title/abstract/venue/query text.

Goal-specific weights stay intentionally separate: systematic emphasizes provenance, seminal emphasizes impact, newest emphasizes recency, implementation emphasizes relevance/access, and pdf-corpus emphasizes access. survey is balanced.

Waves — each is a fan-out followed by a merge_corpus.py pass into the same corpus; applies to both the Workflow and Agent paths:

1. Wave 1 (all levels): buckets searchers, each running the facets subqueries, ~hits per subquery.
2. Citation-snowball (L3+): take the top ~10 DOIs/arXiv ids from the corpus so far and fan out one wave that expands their forward (cited-by) + backward (references) neighbours. Workflow: re-run the script with seeds:[...]. Agent: tell each searcher to run asta get_citations / arXiv citation_graph / OpenAlex cited-by on the seeds.
3. Completeness-critic (L4+): a critic agent reads corpus.md and names the missing subtopics / seminal authors; those become new facets for one more wave-1-style fan-out.
4. Loop-until-dry (L5): repeat the critic → facets → fan-out → merge cycle until two consecutive critic passes surface nothing new (or, under Workflow, budget.remaining() runs low).

L4/L5 also re-run the merge with --min-sources 2 → corpus_shortlist.json (papers corroborated by ≥2 databases) alongside the full corpus.json.

Fallback when MCP is unavailable

If MCP servers are down/headless, searchers use scripts/search_local.py {arxiv| semanticscholar|ddg|kisti} "query" with the installed host venv Python. arXiv may rate-limit (HTTP 429) under heavy fan-out — stagger or lean on Asta/OpenAlex. The Asta (Semantic Scholar) MCP is remote and needs no key (a key only raises rate limits) — in headless/cron runs just ensure network access, or fall back to search_local.py semanticscholar. Never let a host-specific scholar gateway be a bucket's only tool (absent in headless runs). kisti 소스는 국내(KCI/보고서/특허) 전용이며 KISTI_CLIENT_ID/KISTI_AUTH_KEY/KISTI_MAC 환경변수가 있어야 동작한다(references/kisti.md). 자격증명이 없으면 이 버킷을 건너뛴다.

For failure-recovery runs, use the resilient local ladder instead of aborting:

python3 <skill-dir>/scripts/resilient_search.py "<query>" \
  --sources arxiv,semanticscholar,ddg -n 20 \
  -o ./literature_search/<slug>_<date>/raw/local_recovery.json \
  --status ./literature_search/<slug>_<date>/raw/local_recovery.status.json

Each searcher should follow the same policy: preferred MCP → alternate MCP in the bucket → local resilient fallback where applicable → record the failed source in a status file and continue with partial results. Do not fail the whole run because one source fails.