Research Team Spawn Recipe

Team Composition

Role	Model	Tools	Purpose
Lead (you)	inherit	All	Decompose task, assign sources, synthesize results
Researcher(s)	Haiku	Read, Glob, Grep, Bash, Write, Edit + agent-fetch	Fetch + analyze + write findings
Reviewer	Sonnet	Read, Glob, Grep	Cross-reference, identify contradictions, quality check

Spawn Protocol

1. Decompose the research topic into 2-4 independent source groups
2. Spawn researchers (one per source group) as background agents:

   For each source group, spawn a researcher agent with:
   - Specific URLs or search terms to investigate
   - Target output file: docs/research/trail/deep-reviews/<topic>.md
   - Provenance records: docs/research/trail/findings/finding-<slug>.yaml
   

3. Spawn reviewer after researchers complete:
   • Cross-reference all findings for contradictions
   • Identify gaps requiring follow-up
   • Quality-check source citations
4. Synthesize results into a summary

File Ownership (No Conflicts)

Owner	Writes To	Cannot Touch
Researcher A	findings/finding-<topic>-a-*.yaml	Other researchers' files
Researcher B	findings/finding-<topic>-b-*.yaml	Other researchers' files
Reviewer	(read-only)	Everything
Lead	deep-reviews/<topic>.md, source-catalog.md	findings/

Quality Gate

• Every URL fetched must be logged in source-catalog.md
• Every finding must have a YAML provenance record
• Contradictions between sources must be flagged
• Final synthesis must cite specific findings

Source Discovery Protocol

• Log EVERY URL encountered (HIGH/MEDIUM/LOW/SKIP classification)
• Never silently discard a URL
• Use npx agent-fetch "<url>" --json for full content (never WebFetch)