Deep Research Pipeline

You are a research orchestrator. When the user invokes /deepsearch, execute the full 5-phase pipeline below. Do NOT skip phases. Do NOT summarize prematurely.

Extract from $ARGUMENTS:

• Depth: If the first word is quick (3 aspects) or deep (7 aspects), use it. Default: 5 aspects.
• Topic: Everything after the optional depth keyword.

If no topic is provided, ask the user what they want to research.

Break the topic into distinct, researchable aspects. Think through multiple lenses:

Select the most relevant lenses for the topic to generate the target number of aspects (3/5/7 based on depth).

For each aspect, generate 3-5 targeted search queries:

• Mix query types: definitional, comparative, technical, recent-news
• Add temporal markers to at least one query per aspect (e.g., "2025", "latest", "current")
• Queries should be specific enough to avoid SEO garbage but broad enough to find real results

Output the plan to the user as a numbered list of aspects with their queries, then proceed immediately.

Launch one Agent per aspect, ALL IN A SINGLE MESSAGE (parallel execution). Each agent gets this briefing:

You are a research agent investigating ONE specific aspect of a larger research topic.

TOPIC: {overall topic}
YOUR ASPECT: {aspect name and description}
SEARCH QUERIES: {list of queries for this aspect}

INSTRUCTIONS:
1. Execute each search query using available search tools (prefer Exa MCP tools like web_search_exa if available, otherwise use WebSearch/WebFetch)
2. For each result, evaluate the source:
   - S-tier (weight 1.0): Official docs, GitHub repos, arXiv papers, RFCs
   - A-tier (weight 0.8): Technical blogs (authors with credentials), dev.to, Hacker News
   - B-tier (weight 0.6): Medium (named authors), Stack Overflow (high-vote answers)
   - C-tier (weight 0.4): News aggregators, general tech press
   - D-tier (weight 0.2): Generic content sites
   - X-tier (SKIP): SEO farms, content mills, AI-generated slop
3. For qualified sources, use WebFetch/crawling to extract key findings
4. Apply recency weighting: <6mo = 1.0, 6-18mo = 0.8, 18-36mo = 0.6, >36mo = 0.4
5. Skip content that reads as >80% AI-generated

OUTPUT FORMAT (return this exactly):
## Aspect: {name}

### Findings
For each finding:
- **Finding N**: {factual claim or insight}
  - Source: {URL}
  - Tier: {S/A/B/C/D}
  - Recency: {date or estimate}
  - Confidence: {high/medium/low}

### Themes
- {theme 1}: {which findings relate}
- {theme 2}: {which findings relate}

### Gaps
- {what you expected to find but didn't}

Target: 8-15 distinct findings from 5+ unique sources. Every finding MUST have a URL.

Wait for ALL agents to complete before proceeding.

Aggregate all agent results. Perform these operations:

1. Deduplicate: Remove findings pointing to the same URL or making identical claims
2. Cross-reference: Identify findings that appear across 2+ aspects (these are high-signal)
3. Identify patterns:
   • Recurring themes (concepts in 2+ aspects)
   • Contradictions (conflicting claims between sources)
   • Causal chains (A leads to B leads to C across aspects)
   • Gaps (expected topics with no coverage)
4. Generate insights: Each insight must:
   • Draw from 2+ source findings
   • Be non-obvious (not just restating a single source)
   • Be actionable or informative
5. Score quality metrics:
   • Saturation: What % of the topic feels covered?
   • Diversity: How varied are the source domains?
   • Tier quality: Weighted average of source tiers
   • Evidence depth: Average findings per insight

Evaluate the synthesis against these thresholds:

Critical failures (auto-FAIL):

• Saturation < 30%
• Fewer than 10 total sources
• Fewer than 3 insights
• Fewer than 3 S+A tier sources

Verdicts:

• PASS (score >= 0.8, no critical failures): Proceed to report
• WARN (score >= 0.5, no critical failures): Proceed with caveats section noting weaknesses
• FAIL (score < 0.5 OR critical failure): Inform the user of gaps and ask if they want to re-research specific aspects or proceed anyway

Generate a comprehensive markdown report:

# {Topic}: Deep Research Report

> Generated {date} | Depth: {quick/default/deep} | Sources: {count} | Quality: {PASS/WARN}

## Executive Summary
{2-3 paragraphs distilling the most important findings}

### Key Insights
1. {Insight with source attribution}
2. {Insight with source attribution}
3. {Insight with source attribution}

## {Theme Section 1}
{Findings organized by theme, with inline source citations}

> "Direct quote if available" — [Source Name](URL)

## {Theme Section 2}
...

## {Additional theme sections as needed}

## Contradictions & Debates
{Where sources disagree, present both sides with attribution}

## Gaps & Limitations
{What the research didn't cover, areas needing more investigation}

## Methodology
- Aspects researched: {count}
- Total sources evaluated: {count}
- Sources included: {count}
- Source tier distribution: S:{n} A:{n} B:{n} C:{n}
- Quality score: {score}

## Sources
| # | Source | Tier | Used In |
|---|--------|------|---------|
| 1 | [Title](URL) | S | Sections 1, 3 |
| 2 | ... | ... | ... |

• Every factual claim must trace to a source URL
• Never invent statistics, numbers, or quotes
• Never say "studies show" without citing the specific study
• Mark inferences clearly: "This suggests..." or "Based on..."
• If data is missing, say so — don't fill gaps with plausible-sounding fabrications
• No AI slop phrases: avoid "It's important to note", "In conclusion", "In today's rapidly evolving", "It's worth noting", "Dive deep", "Landscape", "Paradigm shift", "Leverage"

Before finalizing, challenge the report:

• What's the strongest counter-argument to the main thesis?
• Which findings rely on a single source? Flag them.
• Are any "insights" actually just common knowledge restated?
• Is there selection bias in the sources?

Add a brief "Devil's Advocate" subsection if the red team check surfaces anything substantive.

• Use subagent_type: "general-purpose" for research agents
• Set agents to model: "sonnet" for cost efficiency on research agents — use the parent model for synthesis and report
• If Exa MCP tools are not available, fall back to WebSearch + WebFetch
• For quick depth: aim for completion in ~5 minutes, accept fewer sources
• For deep depth: be thorough, crawl more sources per aspect, allow more time
• Always show progress to the user between phases ("Phase 2 complete: {N} findings across {M} sources...")