Search Plan

Build a source discovery plan and execute it.

Boundary:

• Do not write sources.jsonl, evidence-matrix.jsonl, source notes, final reports, QA reports, or review reports.
• Search Strategist output is a candidate handoff only. Source Reader is responsible for fetching content and registering retrievable sources.
• If you use WebFetch to sanity-check a candidate, label it as fetch_status in the handoff. Do not treat it as a registered source.

Inputs:

• Research contract.
• Scope brief.
• Existing sources and evidence gaps.

Procedure:

1. Generate query families

   • Core concept query.
   • Synonym query.
   • Primary-source query.
   • Critical/limitations query.
   • Recent query when freshness matters.
   • Academic query when papers matter.

2. Select channels

   • Use WebSearch with categories: "science" for academic/technical discovery when useful.
   • Use WebSearch with site for official docs, standards, or specific domains.
   • Use arXiv MCP mcp__arxiv__search_arxiv for academic paper discovery.
   • Use arXiv MCP mcp__arxiv__get_paper_details for candidate metadata or content previews.
   • Use arXiv MCP mcp__arxiv__search_and_summarize when a query needs a compact paper set overview.
   • If the contract requires official, regulatory, standard, or authoritative medical sources and WebSearch returns no results for that source class, use documented shell/curl fallback against the public domain itself, such as curl -L https://www.fda.gov/... or a public site search endpoint. Record the prior WebSearch error or empty result and mark the discovery channel as fallback.

3. Execute in rounds

   • Round 1: broad discovery, 5-10 results.
   • Round 2: targeted source types/domains.
   • Round 3: conflicts, critiques, negative evidence.
   • Stop when the contract coverage is met. Do not keep searching after you have enough credible candidates for every required source class.
   • For interactive use, prefer a bounded handoff: normally 12-20 candidates, at least the requested minimum plus 2-4 backups. More than 25 candidates usually slows Source Reader without improving quality.

4. Score candidates

   • Relevance: high/medium/low.
   • Credibility: primary/peer-reviewed/official/secondary/unknown.
   • Freshness.
   • Likely use: definition, empirical evidence, benchmark, method, limitation, context.

5. Return candidate list

   • Do not claim snippets as evidence.
   • Mark which sources Source Reader should fetch first.
   • Mark fetch_status: not_fetched|fetched_preview|fetch_failed if known.
   • Mark evidence_allowed: no for every candidate, because only Source Reader can promote a candidate to evidence after reading the source.
   • Include a coverage_summary that explicitly states whether every required source class is covered. If a required class is missing, do not hand off as complete; run targeted recovery or fallback first.

Output format:

## Search Plan and Candidate Sources

### Query Families
- F1: ...

### Searches Run
| Round | Tool | Query | Category/Site | Result count | Notes |
|---|---|---|---|---|---|

### Candidate Sources
| Candidate ID | Title | URL/arXiv ID | Type | Relevance | Credibility | Read Priority | Fetch Status | Evidence Allowed | Reason |
|---|---|---|---|---|---|---|---|---|---|

### Coverage Gaps
- ...

Recovery examples:

• If SearxNG general engines time out, retry with categories: "science" or site: restrictions.
• If all web search fails, use arXiv MCP for academic topics and report WebSearch errors exactly.
• If official/regulatory/medical authority searches are empty but required, use shell/curl fallback on the public official site and record the exact failed WebSearch query and fallback reason.
• If a query family produces many irrelevant arXiv hits, stop broad querying and switch to known exact titles, author names, or site/domain-restricted queries.