orchestrator

name	orchestrator
description	High-level orchestrator for iterative review planning and hypothesis evolution. Builds the current autor planning package, triggers targeted acquisition, maintains corpus layers, and stops before writing.

Orchestrator: Iterative Research Planning

Use this skill for large or iterative review planning, especially when the user wants coverage checking, framework evolution, targeted acquisition, trial mapping, or independent verification before writing.

Do not draft manuscript prose.

Canonical Outputs

The orchestrator must leave the workspace in the current planning format:

workspace/<name>/references.bib
workspace/<name>/reference-map.json
workspace/<name>/review-plan.md
workspace/<name>/evidence-ledger.md
workspace/<name>/table-figure-plan.md
workspace/<name>/acquisition-log.md if acquisition, trials, or citation-network work occurred
workspace/<name>/sidecars/

references.bib citation keys are canonical across all downstream artifacts.

Compatibility exports (paper-classification.md, section-evidence.md, table-plan.md, execution-tasks.md, corpus-ledger.md, query-matrix.md) are optional derived files only.

Core Rules

1. Treat the global library as orientation, not proof of completeness.
2. For full reviews, build a universe corpus before pruning to the working and core analytical corpora.
3. Use the user framework as a first-class input. Audit every unit as keep, merge, split, rename, reposition, defer, or drop.
4. Use Records-backed autodownload for justified external acquisition. Do not use Playwright fallback.
5. Ingest full PDFs before using new papers for major planning claims.
6. Maintain citation-key identity through references.bib and reference-map.json.
7. Keep trial records as a parallel evidence layer.
8. Stop at an approved plan; manuscript drafting is outside this orchestrator and must be chosen explicitly by the user.

For full-review coverage-first work, read references/full-review-coverage-mode.md.

Workflow

1. Diagnose

• inspect workspace/<name>/papers.json
• inspect existing canonical planning files
• run autor ws status <name> --papers and autor ws export-evidence <name> for a machine-readable completeness check
• run local workspace search and L2/L3 reading for orientation; preserve L3 mode (explicit_section, inferred_synthesis, etc.) when it affects evidence confidence
• identify current corpus layers, missing branches, and unresolved references If L3 coverage is incomplete, use workspace-scoped enrichment (autor enrich-l3 --workspace <name> --only-missing or MCP workspace_enrich_l3) before planning claims that depend on paper-level conclusions.

2. Normalize identity

• export or repair references.bib
• create or update reference-map.json
• ensure retained papers have stable citation keys
• keep UUIDs, dir names, PMIDs, DOIs, and full-text status in reference-map.json Use autor ws plan-package <name> for a fresh identity/evidence scaffold when these files are missing, then revise the generated plan into the current review framework.

3. Calibrate coverage

For full reviews, use landscape tools when available and/or external metadata retrieval to estimate:

• field size
• recent growth
• review saturation
• seminal papers
• branch structure
• identifier normalization problems

Build a query matrix when external coverage checking is performed, but store it inside acquisition-log.md unless a tool requires a sidecar.

4. Acquire and ingest when needed

Use autodownload only when a real gap or full-review coverage requirement justifies it.

Record:

• query branch
• candidate identifiers
• include/exclude decisions
• download status
• ingest status
• workspace attachment status

5. Re-plan

Run the planning logic after acquisition:

• update framework audit
• update section cards
• update corpus layers
• update section evidence
• update table and figure plan
• document evidence-thin and true research-gap areas

6. Trials and citation network

Use trials when clinically relevant. Save trial outputs under workspace/<name>/trials/ and map retained trial IDs to sections.

If citation-network mapping is used, save raw graph data under sidecars/ and summarize lineages or conceptual bridges in acquisition-log.md or review-plan.md.

7. Verify

Before approval, verify:

• every retained citation key exists in references.bib
• every retained key appears in reference-map.json
• every working/core paper maps to at least one section
• every major section has evidence support
• every table or figure has a source evidence set
• excluded and unresolved records have reasons
• external completeness judgments are not based only on local library density

Use an independent verification subagent when the task is large enough and subagents are available.

Final Output

Return the same handoff block used by plan:

FRAME_STATUS: APPROVED | REPLAN_REQUIRED | BLOCKED
WORKSPACE: <name>
LANGUAGE: <language>
ARTICLE_TYPE: <type>
STYLE_BASELINE: <style>
REVIEW_SCOPE_PROFILE: <scope>
TRIALS_STATUS: COMPLETED | NOT_APPLICABLE | BLOCKED
CITATION_NETWORK_STATUS: COMPLETED | NOT_APPLICABLE | BLOCKED
CORE_HYPOTHESIS: <1-2 sentences>
REFERENCE_POLICY: references.bib citation keys are canonical
CORPUS_COUNTS:
- universe: <n>
- working: <n>
- core: <n>
- excluded: <n>
- unresolved: <n>
HANDOFF_FILES:
- workspace/<name>/references.bib
- workspace/<name>/reference-map.json
- workspace/<name>/review-plan.md
- workspace/<name>/evidence-ledger.md
- workspace/<name>/table-figure-plan.md
- workspace/<name>/acquisition-log.md if applicable
OPEN_GAPS:
- <NONE or explicit list>
NEXT_ACTION: ready_for_manuscript_drafting | run_targeted_acquisition | return_to_user