research-bootstrap

name	research-bootstrap
description	Orchestrator that routes research work across three external skills — oh-my-research (report-first analysis), academic-research-skills (rigor, verification, paper writing), and autoresearch (iteration/verification loop) — using JSON workflow patterns with per-node backend routing. Triggers on: deep research, literature review, survey, report, research-to-paper, systematic review, bootstrap research, bridge research, 深度研究, 文献综述, 调研报告.
version	0.1.0

Research Bootstrap

A pattern-driven orchestrator that bridges three external research skills. It does not reimplement any of them — it routes each workflow node to whichever backend does that node best, and adapts artifacts across the boundaries.

North star: route each phase of research to the backend that is strongest at it, then pass clean artifacts across the seams.

Backends (all external — referenced, never vendored)

Backend	External skill	Source	Role in bootstrap
omr	oh-my-research	caesar0301/oh-my-research	collect → analyze → think → synth (report-first, disk-state, incremental)
ars	academic-research-skills	Imbad0202/academic-research-skills	verify → review → question → paper (adversarial, citation, methodology rigor)
auto	autoresearch	uditgoenka/autoresearch	modify → verify → keep/discard iteration loop

All three are registered as external skills via nano.yml skills: and invoked by name. The bootstrap skill itself ships only routing + patterns + adapters — no backend content.

License note: ars (academic-research-skills) is CC BY-NC 4.0. The bootstrap references it by name only and never copies its content. Keep it that way.

Dispatch

Parse the invocation in this order:

1. Canonical pattern name — bootstrap --pattern <name> wins over everything.
2. Intent signals — match against references/routing.md, select a pattern.
3. Workspace state — if a pattern is already active in .bootstrap/state.json, resume it.
4. Ambiguous — show the pattern library and ask.

Print a banner on every invocation: [bootstrap] pattern: <name> | backends: omr, ars, auto.

Patterns

Pattern	Route map	Gates	Use when
evidence-deep	collect/analyze/think/synth=omr; verify/review=ars	M→A→T→P→D	Default. Survey/report with adversarial checkpoints
evidence-first	collect/analyze/synth=omr; review=ars	M→A→P→D	Lighter; one review pass
rapid	all=omr	off	Quick brief, no rigor
rigor-first	research/source-verify/synthesis=ars; collect/synth=omr	ARS checkpoints + D	ARS deep-research → OMR long report
paper-pipeline	research+paper+review=ars; ingest/export=omr	ARS mandatory	Research → manuscript → peer review
systematic-review	PRISMA/RoB/meta=ars; report synth=omr	ARS compliance	PRISMA systematic review + meta-analysis
idea-first	idea=omr; question=ars (socratic); analyze/synth=omr	M→A→P→D	Vague idea, guided RQ formulation
stance-first	decide=omr; analyze/synth=omr; adversarial=ars	B→A→P→D	Position/claim paper with devil's-advocate stress test
verify-loop	produce=omr; verify=auto+ars; revise=omr	L (iterate vs advance)	Converge claims before shipping
loop	collect/analyze=omr; gap-detect=ars	L cycles	Ongoing monitoring + drift reconciliation

Pattern files: patterns/*.json. See references/routing.md for intent → pattern selection.

Node → Backend Contract

Node	Backend	Produces
collect	omr	materials/**, docs/index/papers-index.json
analyze	omr	docs/plans/{brief,evidence,judgment}-*.md
think	omr	refined artifact (THINK playbook)
verify	ars	verification/graded sources (source_verification_agent)
review	ars	editorial + ethics + devil's-advocate verdict
question	ars	RQ brief (research_question_agent / socratic_mentor)
synth	omr	docs/<mode>/ long report + DOCX/PDF export
paper	ars	manuscript (academic-paper)
iterate	auto	modify → verify → keep/discard loop

Artifact Adapters (the seams)

Two deterministic adapters translate between OMR and ARS formats:

Adapter	Direction	Transform
scripts/papers_index_to_csljson.py	omr → ars	docs/index/papers-index.json → Material Passport literature_corpus[] (CSL-JSON)
scripts/csljson_to_omr.py	ars → omr	literature_corpus[] → docs/index/papers-index.json + _citation-map.md

Adapters are thin and deterministic (no network, no LLM). Metadata enrichment (authors, year, venue) is expected to come from the OMR analyze step or the _citation-map.md — never invented by the adapter.

State

.bootstrap/state.json — bootstrap-owned, additive. Tracks: active pattern, completed nodes, current backend, last adapter run, per-node outcomes. Each backend keeps its own state (.omr/ for OMR, Material Passport for ARS, autoresearch/ output dirs for auto). The bootstrap reads them; it does not own them.

Safety Invariants

1. Never relicense ARS content — reference by name, never copy into this repo.
2. No silent backend substitution — a node routed to ars stays on ars; switching backends is a pattern change, announced to the user.
3. Adapters never invent metadata — a missing year/authors goes to rejection_log.yaml, not a placeholder (mirrors ARS's own adapter discipline).
4. No un-gated ship — paper and synth publish nodes require their backend's own gates (ARS mandatory checkpoints, OMR Gate D) before any deliverable leaves the workspace.

References

File	Purpose
references/routing.md	Intent detection + pattern selection table
references/backends.md	External skill contracts (what each owns)
references/handoff-omr-ars.md	OMR artifacts → ARS ports
references/handoff-ars-omr.md	ARS artifacts → OMR ports