sovereign-research (Deep Research Swarm & CEO Edition)

(opencode - Skill) The March 2026 gold standard for deep, multi-source research. Fuses a bounded multi-agent swarm with the A2A-SIN-Research pipeline, enabling Perplexity-killer answers, live authenticated multi-platform crawling (via A2A-SIN-Authenticator), and persistent Drive artifacts. Inspired by Gemini Deep Research workflows.

Triggers

MANDATORY USE: Keywords "research", "find out", "what is the best", "compare", "state of the art", "deep dive", "investigate", "analyze topic", "deep research".

Architecture

This skill replaces linear, single-agent searches with an Aggressive Multi-Agent Swarm:

User Query
    |
    v
[Phase 1: Swarm Fan-Out (Parallel)]
    ├─> Librarian Task A: Google Search & Webfetch (General Context)
    ├─> Librarian Task B: Exa Search (Category filtered: papers/company)
    ├─> Explore Task C: GitHub/Code search (grep_app_searchGitHub)
    └─> Explore Task D: Authenticated Portal Search (X, Discord, Custom via A2A-SIN-Authenticator)
    |
    v
[Phase 2: Source Gather & Validate]
    Wait for tasks (background_output) -> deduplicate, score authority, date-check
    |
    v
[Phase 3: Deep Research (Optional / Complex Queries)]
    sin.research.run (A2A-SIN-Research) -> NotebookLM swarms, NIM analysis, Drive persistence
    |
    v
[Phase 4: Review & Synthesis]
    Artistry Task: Critical Review (Gaps, counter-arguments, missing sources)
    Output: Structured cited answer with confidence scores.

Protocol

Step 1: Deep Research Swarm Fan-Out (Max 5-10 Minutes)

Launch parallel background tasks using the task tool to gather maximum context without blocking.

// Example: Launching parallel subagents
task({
  subagent_type: "librarian",
  run_in_background: true,
  load_skills: [],
  description: "Exa-based academic/company research",
  prompt: "Run websearch_web_search_exa for [Query]. Deep read the top 3 URLs using webfetch. Return structured findings."
});

task({
  subagent_type: "explore",
  run_in_background: true,
  load_skills: [],
  description: "Codebase & GitHub research",
  prompt: "Use grep_app_searchGitHub to find real implementations of [Query]. Return code patterns."
});

For Authenticated Portals (e.g., hundreds of platforms like X, Discord, Reddit, custom enterprise tools): Request A2A-SIN-Authenticator to handle mass-platform login and cookie injection via webauto-nodriver-mcp or curl_cffi, allowing the explore agents to read auth-gated data safely.

Step 2: Gather & Source Validation

Check background_output(task_id=...) for all launched agents. Do not proceed until you have gathered all findings.

Validate the sources:

• Score authority (Official Docs > Blogs > Forums).
• Flag stale content (>6 months old).
• Ensure explicit traceability: every claim MUST link directly to its source URL.

Step 3: A2A Deep Research (For Highly Complex Queries)

Only invoke if the query spans 10+ sources, requires video/image parsing, or requires persistent artifacts.

# Plan first
sin-research run-action '{"action":"sin.research.plan","query":"...","targetDepth":"deep","includeAuthSources":true}'

# Then execute
sin-research run-action '{"action":"sin.research.run","query":"...","targetDepth":"deep","confirm":true}'

This triggers the native A2A backend to spin up NotebookLM instances, process multimodal data (NVIDIA NIM), and persist to Google Drive.

Step 4: Verification & Synthesis

Launch a final Artistry task to review the synthesized findings:

task({
  subagent_type: "artistry",
  run_in_background: false,
  load_skills: [],
  description: "Critical review of research",
  prompt: "Review these findings. Are there gaps? False claims? Uncited sources? Provide concrete rewrites."
})

Return the final structured answer to the user:

## Answer
<Direct answer to the query>

## Key Findings
1. <Finding with [Source 1] citation>
2. <Finding with [Source 2] citation>

## Sources
| # | Source | Authority | Date | URL |
|---|--------|-----------|------|-----|
| 1 | Official Docs | High | 2026-03 | https://... |

## Confidence: <HIGH/MEDIUM/LOW>
## Research Depth: <SWARM/DEEP>

Anti-Patterns

• NEVER return search results without launching parallel background agents first. Deep research requires volume.
• NEVER cite a source without verifying its date (stale = warning).
• NEVER run sin.research.run for simple factual queries (waste of NotebookLM quota).
• NEVER autonomously switch technologies during research. Technology decisions are EXCLUSIVELY the user's prerogative.
• PERMANENTLY BANNED technologies: Camoufox, Playwright, Puppeteer, Selenium, any Firefox-based browser automation. The ONLY approved browser automation stack is nodriver + Chrome profile, webauto-nodriver-mcp, CDP, and curl_cffi. Violation = immediate ban from the project.

Integration Points

• A2A-SIN-Authenticator: Controls mass-platform auth orchestration. Provides authenticated handles/sessions to webauto-nodriver-mcp/curl_cffi so research agents can bypass auth gates.
• A2A-SIN-Research (sin.research.*): Full deep-research pipeline with NVIDIA NIM, NotebookLM, Drive persistence.
• google_search / websearch_web_search_exa: Live crawling.
• grep_app_searchGitHub: Real code pattern search across 1M+ public GitHub repos.

Related Skills

• See Skill Catalog for related skills