| name | research-orchestration |
| description | Agents should invoke this skill for broad multi-claim research projects needing planning, parallel investigation, source merging, gap closure, citation audit, and final synthesis when narrower research skills are insufficient. |
Research orchestration
Single workflow entry point for end-to-end investigations. Detailed rules live in workspace-researcher/AGENTS.md (task class, parallel investigation, Deep Research Protocol, citation audit); this skill is the ordered checklist for executing them.
When to use
- Broad question with several independent sub-questions or claim clusters.
- Mixed sources (web + docs + papers) and one final report or contract output.
- User or delegating agent asked for deep research without pointing at a single comparison or single paper.
Prefer tech-deep-dive, competitor-analysis, or paper-summarizer when the task clearly matches one of those shapes.
Pipeline
- Classify depth — Quick / standard / deep (
AGENTS.md task-class table). Honor explicit quick / standard / deep or /deep-research quick|standard|max when the user names them. Set expected tool budget.
- Plan sub-questions — List facets and claims to verify; note which are high-stakes (citation audit mandatory). Prefer internal → official → web (
AGENTS.md data source order) when gathering evidence.
- Parallel pass — Batch
web_search / web_fetch per independent facet; optional subagents with tight scopes and structured handoff.
- Merge — One synthesis pass; resolve contradictions; prefer primary sources.
- Gap closure — List unresolved material gaps; targeted follow-up searches or explicit “blocked” notes.
- Citation audit — For mandatory cases: every key finding → source row or
unsupported / inferential (AGENTS.md).
- Deliver — Default report shape, contract JSON, or consumer-specific packet from
AGENTS.md; include Limitations and Research trace when depth is standard/deep or requested (APPEND_SYSTEM.md).
- Log — Append a row to
workspace-researcher/MEMORY.md → Research History (topic, task class, tool bucket, major gaps, notable sources).
Scout scripts → pipeline step
Run from the installed skill directory, or adjust paths to absolute locations. Policy defaults to ./scripts/policy.json next to the bundled scripts.
| Step | Script | When |
|---|
| After depth + topic fixed | ./scripts/scout_query_plan.py | Emit ordered queries (Q001…) and plan_hash for the research trace |
| After URL collection | ./scripts/scout_normalize_sources.py | Dedupe and sort sources before the final table |
| Long / multi-session runs | ./scripts/scout_evidence_bundle.py | Freeze fetch records (JSONL → evidence_bundle.json + content_sha256) |
| Before high-stakes delivery | ./scripts/scout_citation_audit.py | Gate: every key_findings[] row has source_ids or unsupported / inferential |
Details and exit codes: ./scripts/README.md.
Handoff schema (optional, for subagents)
## Scope
[One sentence]
## Findings
- [Bullet + URL]
## Gaps
- [What could not be verified]
## Sources touched
- [Title](URL) — quality flag
Scout skill — Orchestrated research pipeline · 2026-04-10
