Research

Session: ${CODEX_SESSION_ID}

Research Topic: $ARGUMENTS

Research Skill

Research Workflows

Standard Query

1. THINK → Sequential Thinking decomposition
2. RESOLVE → Context7 resolve-library-id
3. DOCUMENT → Context7 query-docs (5000-10000 tokens)
4. SUPPLEMENT → Exa code context search
5. SYNTHESIZE → Structured answer with sources

Complex Investigation

1. DEEP THINK → Multi-hypothesis Sequential Thinking
2. DEEP RESEARCH → Exa deep researcher (45s-2min)
3. MONITOR → Check status until completed
4. VALIDATE → Cross-check Context7 official sources
5. REPORT → Comprehensive solution

Technology Trends

1. WEB SCAN → Exa search latest developments
2. CODE PATTERNS → Exa code context for practices
3. ECOSYSTEM → Company research for key players
4. ANALYSIS → Sequential Thinking for implications
5. RECOMMENDATIONS → Actionable insights

Context7 Usage

// Step 1: Resolve library ID
mcp__context7__resolve-library-id({
  libraryName: "next.js",
  query: "App Router server actions"
})

// Step 2: Query docs
mcp__context7__query-docs({
  libraryId: "/vercel/next.js",
  query: "server actions authentication"
})

Best Practices:

• Always resolve-library-id BEFORE query-docs
• Specify topic parameter to focus retrieval
• Start with 5000 tokens, increase to 10000 if needed
• Handle variations: "nextjs" vs "/vercel/next.js"

Exa Search Types

// Code context search
mcp__exa__get_code_context_exa({
  query: "Next.js 16 server actions authentication",
  tokensNum: 5000
})

// Web search
mcp__exa__web_search_exa({
  query: "React 2025 best practices",
  type: "auto",
  numResults: 5
})

Exa Deep Research

Reserve for investigations requiring >30min manual effort.

// Start research
const { taskId } = await mcp__exa__deep_researcher_start({
  instructions: "Compare authentication solutions for Node.js",
  model: "exa-research-pro" // or "exa-research" for faster
})

// Poll until complete
mcp__exa__deep_researcher_check({ taskId })

Models:

• exa-research: Standard depth (15-45s)
• exa-research-pro: Complex topics (45s-2min)

Sequential Thinking

mcp__sequential-thinking__sequentialthinking({
  thought: "Analyzing authentication approaches",
  thoughtNumber: 1,
  totalThoughts: 5,
  nextThoughtNeeded: true,
  // Optional for revisions:
  isRevision: false,
  revisesThought: null,
  branchId: null,
  branchFromThought: null,
  needsMoreThoughts: false
})

Best Practices:

• Start with realistic totalThoughts, adjust dynamically
• Use isRevision: true to reconsider hypotheses
• Create branches (branchId) for alternatives
• Set needsMoreThoughts: true if incomplete

Multi-Source Synthesis

Parallelization:

• Run resolve-library-id + web_search_exa simultaneously
• Launch multiple Exa searches concurrently
• Execute Context7 docs + Exa code search in parallel

Source Priority:

1. Official documentation (Context7)
2. Recent tutorials (Exa, <6 months)
3. Older content (with version verification)

Response Format

## 🔍 Research: [Topic]

### Methodology
- Sequential Thinking: [N thoughts, M revisions]
- Context7: [Library@version consulted]
- Exa: [Search types performed]

### Key Findings
1. **[Finding 1]** (Source: [URL])
   - Technical details
   - Code examples

2. **[Finding 2]** (Source: [URL])

### Recommendations
- [Action 1]: [Why + How]
- [Action 2]: [Why + How]

### Sources
- Context7: [Exact library IDs]
- Exa: [X results analyzed]
- Deep Research: [Task ID if used]

Error Handling

Context7 Failures:

• Verify library name spelling
• Try different formats ("/org/project" vs "project-name")
• Fallback to Exa code context

Exa Timeouts:

• Reduce numResults
• Simplify query
• Switch type: "deep" → type: "fast"

Sequential Thinking Blocks:

• Revise with isRevision: true
• Increase totalThoughts
• Create new branch

Forbidden Behaviors

• ❌ Guess library IDs without resolve-library-id
• ❌ Start deep researcher without checking completion
• ❌ Mix opinions with facts without distinction
• ❌ Provide code without version verification
• ❌ Ignore WebFetch redirects
• ❌ Recommend without citing sources
• ❌ Skip Sequential Thinking for multi-step problems