deep-research

/interdeep:deep-research

Orchestrate a deep research session: decompose a query, search multiple sources, extract and evaluate content, synthesize findings, and persist the result.

Protocol

Execute these five phases in order. Adapt depth based on the mode (see Depth Modes below).

Phase 1: Orient

1. Use the research-planner agent to decompose the query into sub-queries with source routing.
2. Review the planner output: query_type, sub_queries (each with query, sources, priority), depth_recommendation, rationale.
3. If the user did not specify a depth mode, use the planner's depth_recommendation.

Phase 2: Search

For each sub-query, dispatch searches to the routed sources:

• Web search — call interject_search (interject plugin) or web_search_exa (interflux/exa plugin):

  mcp tool: interject_search
  args: { "query": "<sub_query>", "source": "exa", "max_results": 10 }
  

• Academic — call interject_search with "source": "arxiv".
• Knowledge base — call interknow_qmd__search or interknow_qmd__vector_search (interknow plugin) for local knowledge.
• Cached results — check intercache if available to avoid redundant fetches.

Collect URLs and snippets from all search results.

Phase 3: Extract

For each URL returned in Phase 2:

1. Call extract_content (interdeep MCP tool) for single URLs:

   mcp tool: extract_content
   args: { "url": "<url>", "include_metadata": true }
   

2. For batches (5+ URLs), use extract_batch:

   mcp tool: extract_batch
   args: { "urls": ["<url1>", "<url2>", ...], "max_concurrent": 5 }
   

3. Use the source-evaluator agent on each extraction result to score relevance and credibility.
4. Filter to sources where include_in_report is true.

Phase 4: Synthesize

1. If interlens is available, call detect_thinking_gaps to identify blind spots:

   mcp tool: detect_thinking_gaps
   args: { "context": "<research summary so far>" }
   

2. If gaps are found and depth mode is deep, run additional searches targeting the gaps.
3. Pass evaluated findings to the report-compiler agent.
4. Alternatively, call compile_report (interdeep MCP tool) for a structured markdown report:

   mcp tool: compile_report
   args: {
     "title": "<report title>",
     "query": "<original query>",
     "findings": [...],
     "sources": [...]
   }
   

5. If intersynth is available, use it for additional synthesis passes.

Phase 5: Persist

1. Save to filesystem — always write the report to a research/ directory in the current working directory:
   • Create the research/ directory if it doesn't exist.
   • Filename format: research/YYYY-MM-DD-<slug>.md where <slug> is a short kebab-case summary of the query (max 60 chars).
   • If a file with the same name already exists, append -2, -3, etc.
   • The saved file should include the full report with all sections from the Output Contract.
2. If interknow is available, also store the report for future retrieval:

   mcp tool: interknow_qmd__search
   

   (Check for duplicates before storing.)
3. Present the final report to the user, noting the saved file path.
4. Call research_status to confirm extraction capabilities used:

   mcp tool: research_status
   args: {}
   

Depth Modes

Mode	Sub-queries	Sources per query	Extract limit	Thinking gaps
quick	1-2	1-2	5 URLs	skip
balanced	3-5	2-3	15 URLs	if available
deep	5-10	all available	30+ URLs	required

Default: balanced unless the planner recommends otherwise.

Companion Plugins

Plugin	Role	Required?
interject	Search providers (Exa, arXiv, HN, GitHub)	recommended
interflux/exa	Alternative web search via Exa	optional
interknow	Knowledge base storage and retrieval	optional
intercache	Result caching to avoid redundant fetches	optional
interlens	Thinking gap detection and lens analysis	optional
intersynth	Advanced synthesis passes	optional

interdeep works standalone (extract + compile) but reaches full capability with companion plugins.

Output Contract

The skill produces a markdown report with these sections:

1. Executive Summary — 2-3 sentence overview
2. Key Findings — organized by theme, each with confidence level
3. Detailed Analysis — full exploration of each theme
4. Recommendations — actionable next steps
5. Sources — bibliography with relevance annotations
6. Methodology — depth mode used, sources searched, extraction stats

The report is returned as a single markdown document in the chat.