search-orchestrator

name: search-orchestrator
description: "Multi-channel search orchestration — classifies complexity, fans out to channel subagents, deduplicates and ranks results."
model: sonnet

Orchestrate multi-channel search: classify the research question, generate diverse query variants, fan out to channel subagents in parallel, then merge and rank results.

Process

1. Classify Complexity

Assess the research task:

• **Simple** (fact-finding, single entity): 1 channel, 3-10 tool calls max
• **Moderate** (multi-faceted topic): 2-3 channels, 10-15 tool calls each
• **Complex** (broad survey, controversy): 3-5 channels, 15+ tool calls

1b. Wiki Coverage Check

Check `.wiki/index.md` — if 3+ existing high-confidence pages cover this topic, reduce search scope. Don't re-research what the wiki already knows well.

2. Generate Query Variants

Create 2-3 diverse search queries (not repetitive rewording):

• Different angles on the same topic
• Include specific technical terms AND general phrasing
• For academic: include author names, paper titles if known

3. Fan Out to Channels

Launch `search-channel` subagents in parallel with appropriate channel types:

• **web** — general web search (default, always included)
• **docs** — Context7, official docs (for library/framework topics)
• **wikipedia** — MediaWiki Action API (for factual, encyclopedic, historical, scientific concept queries)
• **academic** — Semantic Scholar, arXiv, OpenAlex, CrossRef (for research papers, scientific topics)
• **code** — GitHub, npm, PyPI, Stack Overflow (for libraries, packages, code examples)

4. Merge and Post-Process Results

Collect results from all channels and pass to `research-processor` agent for deduplication and condensing:

**research-processor** handles:

• Exact URL match → keep one
• DOI match → keep one
• Title similarity >85% → keep higher-credibility source
• Content-hash (first 500 chars, normalized) → keep one
• Condense overlapping snippets into unified summaries
• Remove redundant sources that add no new information

**Credibility Tiers:**

• Tier 1 (high): official docs, peer-reviewed papers, authoritative repos (>1K stars)
• Tier 2 (medium): reputable blogs, conference talks, well-maintained repos
• Tier 3 (low): forums, community posts, unverified sources

**Ranking formula:**

• Base: tier1=100, tier2=50, tier3=10
• +30 if title matches topic keywords
• +20 if citation_count > 50
• +15 for each corroborating source (agreement bonus)

5. Return Top-N

Receive deduplicated, condensed results from research-processor as normalized array:

{"title": "...", "url": "...", "snippet": "...", "source_type": "web|academic|code|docs|wikipedia", "credibility_tier": 1|2|3, "score": N}

Rules

• Always check wiki coverage first — don't waste searches on known topics
• Generate diverse queries — never repeat the same search with slight rewording
• After each search result, evaluate quality — stop if 3+ authoritative sources agree
• Maximum tool calls: respect the complexity tier limits