research-orchestrator
Route substantive academic and scientific research goals through the current Agentic Research Workflow source of truth. Use for research planning, literature reviews, paper or method analysis, data-an
SKILL.mdskills/research-orchestrator/SKILL.md
GitHubResearch Orchestrator
Use this Skill as a thin cross-surface router. Retrieve current rules from GitHub, load only task-relevant context, select a specialist workflow or Skill, and report concise QC. Do not duplicate the repository manual here.
Establish the workflow source
Use Lzy599775/agent-auto-sci-skills as the authoritative workflow repository.
- Use
mainfor released work. Use an unmerged candidate only when the user explicitly selects that reviewed ref; report it as a candidate. - Retrieve
docs/version.md,AGENTS.md,ARCHITECTURE.md, anddocs/workflows/index.mdat the start of a substantive workflow. - Report the workflow version and ref used.
Read references/source-retrieval.md for GitHub retrieval, fallback, and version-drift behavior.
Route the goal
Infer one primary route from the intended decision or deliverable:
- papers, literature, DOI, citations, Zotero, or synthesis -> literature;
- reproduce, replicate, reimplement, or benchmark -> paper reproduction;
- methodology / research design, causal identification, estimands, bias, sampling/measurement validity, inference assumptions, or model appropriateness -> methodology, using
agent-auto-sci-methodology; - cleaning data, fitting models, executing statistics/regression, or evaluating prediction performance -> data analysis;
- drafting, rewriting, or restructuring manuscript sections -> manuscript writing;
- whether a citation supports a statement -> citation-claim audit;
- plot, map, table, chart, or figure -> figure workflow;
- novelty, feasibility, research topic, or new idea -> research idea evaluation.
Do not require the user to name the workflow. Load the selected file under docs/workflows/, then only the standards and specialist Skills required to complete the task.
Select by the scientific decision or deliverable, not prompt keywords. Critiquing Methods
validity uses methodology; rewriting its prose uses manuscript writing. Reproducing a published
result uses paper reproduction; novelty/feasibility uses research idea evaluation, with
methodology as needed. Artifact identity checks alone use literature plus the artifact protocol.
A standalone resource preflight uses data analysis in preflight_only mode without starting
computation. The canonical router's ROUTING_DECISION_CONTRACT_V1.json records these choices.
There is no default discipline, dataset, or paper. Domain specialists are optional and loaded only for the current task. Research questions, methods, data, results, figures, papers/Zotero collections, and manuscript/exemplar state stay in the corresponding private project. Resolve identity before private-state access; an unidentified task never inherits a prior project. Ordinary method consultation and micro-edits do not read manuscript exemplar state or initialize Master20/Core3.
Automatic companion routes
- Online paper search automatically invokes
PAPER_ARTIFACT_AVAILABILITY_PROTOCOL_V1for user-facing, qualified, shortlisted, evidence-table, exemplar, method-comparison, and synthesis papers. Raw monitor discovery does not trigger deep search; qualified candidates get Level 1, and reading/exemplar/reproduction records get Level 2. Network failure is not a negative finding. - Substantive manuscript writing automatically invokes the isolated per-manuscript workflow under the unique
agent-auto-sci-scicommowner. Read that manuscript's authoritative private state and versions; never infer from memory or reuse another manuscript's corpus/artifact status. Micro-edits skip heavy initialization.
Execute with evidence discipline
Apply these minimal fallback invariants on every surface:
- prioritize evidence integrity, reproducibility, and traceability;
- prefer original and authoritative sources;
- never invent citations, identifiers, source text, methods, values, results, or availability;
- separate source-reported fact, derived result, and inference;
- keep claim strength at or below evidence strength;
- preserve explicit unknown, unavailable, unverified, and conflicting states;
- treat Zotero as bibliographic/PDF source of truth when available;
- treat Research Vault records as derived knowledge, not original evidence;
- require explicit approval for destructive, publishing, submission, external-message, merge, force-push, private-data, or destructive Zotero actions.
Read references/routing-and-qc.md before finalizing the route or reporting completion.
Use specialists
Prefer an installed specialist Skill when its scope matches. In particular, use literature for evidence-traceable literature search, extraction, synthesis, Zotero work, and citation-support audits. Specialist Skills implement procedures; this Skill owns orchestration and cross-surface source retrieval.
Handle unavailable GitHub context
If the repository cannot be retrieved:
- State
WORKFLOW_SOURCE_FALLBACK_USED. - Identify the intended ref and files that could not be verified.
- Apply only the minimal invariants in this Skill.
- Avoid claiming the fallback is current or complete.
- Continue low-risk work when valid; stop only at genuine access, legal, destructive, or material scientific-choice boundaries.
Return
Report the selected route, workflow repository/ref/version, repository files loaded, specialist Skills used, evidence and uncertainty state, artifacts created, QC results, fallback status, and unresolved blockers.