meta-deep-research-execute

Full multi-model deep research protocol. This skill is executed by an Opus subagent dispatched from /meta-deep-research. It reads its input from deep_research_prompt.md in the research folder, orchestrates all workers, and writes the final summary.

Do not invoke this skill directly. It is triggered by /meta-deep-research.

For detailed phase instructions, read references/protocol-detail.md.

Source Counting Target

Deep research target: 1000+ sources scanned.

Every connector subagent and external CLI worker must track sources. The orchestrator aggregates counts after Phase 2 and again after any addendum cycles. The final summary header reports the total.

Inputs

The dispatching skill provides:

• The NNN identifier for this deep research run (e.g., 003D)
• Path to deep_research_prompt.md (in artifacts/research/ or passed directly)

Read the prompt file first. It contains: the research question, sub-questions, scope, project context, and output configuration.

File Naming Convention

All intermediate files use 3-4 word kebab-case names that describe the content:

• websocket-scaling-limits.md (not codex-worker-1.md)
• rag-accuracy-benchmarks.md (not sonnet-connector-3.md)
• dynamo-failure-stories.md (not gemini-contradictions.md)

The filename should tell you what's inside without opening it.

Phase Overview

Phase 1: Decomposition

Validate and refine the sub-questions from the dispatcher. Classify each by evidence type (Academic, Technical, Market, Reasoning). Assign each sub-question to 2-3 model families ensuring cross-coverage. Write the dispatch table to the artifact DB:

source artifacts/db.sh
db_upsert 'meta-deep-research-execute' 'dispatch-table' '{NNN}D' "$CONTENT"

Phase 2: Parallel Research Fan-Out

Launch all 4 tracks simultaneously:

• Track A — Opus Deep Reasoning (2-3 subagents): Extended thinking on "Reasoning" sub-questions. Full MCP connector access. Cite every claim.
• Track B — Sonnet Connector Sweep (7-10 subagents): One per MCP connector (Consensus, Scholar Gateway, PubMed, Synapse, Clinical Trials, Context7, GitHub, MS Learn, HuggingFace, WebSearch). Uses research-connector agent with multi-query protocol (3-5 queries per topic). Each subagent writes to: research-connector / findings / {NNN}D/{descriptive-name} in the DB.
• Track C — Codex Technical Validation (up to 4 workers): Workers 1-3 do primary technical research. Worker 4 runs devil's advocate (find evidence AGAINST conventional wisdom). Slot 5 reserved for Phase 2.5/3.
• Track D — Gemini Web Grounding (2 instances): Instance 1 does broad Google Search research + case studies. Instance 2 hunts contradictions. Phase 2.5/3 Gemini runs AFTER Track D completes (sequential).

All workers include a Source Tally in their output. After completion, aggregate all tallies and store:

source artifacts/db.sh
db_upsert 'meta-deep-research-execute' 'source-tally' '{NNN}D' "$TALLY_CONTENT"

Phase 2.5: Coverage Expansion (MANDATORY)

This phase ALWAYS runs. It is not optional or conditional. Initial research inevitably surfaces topics and alternatives not in the original prompt.

Step 1 — Coverage Debate: Opus + Codex reviewers run in parallel. Gemini reviewer runs AFTER Track D completes (respects 2-session limit). Each reads the original prompt, dispatch table, all findings, and the source tally. They identify:

• Thin evidence areas needing reinforcement
• Emergent topics that surfaced during research
• Missing well-known approaches/tools
• Underperforming connectors
• Source count gaps vs. 1000+ target

Output stored in artifact DB:

• meta-deep-research-execute / coverage-review / {NNN}D/claude
• meta-deep-research-execute / coverage-review / {NNN}D/gemini
• meta-deep-research-execute / coverage-review / {NNN}D/codex

Step 2 — Addendum Creation: A fresh Opus subagent (not a prior participant) synthesizes the three reviews into an addendum stored at meta-deep-research-execute / addendum / {NNN}D in the DB. Content: reinforcement targets, new sub-questions from emergent topics, missed alternatives, source count plan, and worker allocation.

Step 3 — Addendum Research Cycle: Execute the addendum using Phase 2 patterns. Update dispatch table with [ADDENDUM] tags. Name output files with -addendum suffix. Re-aggregate source tallies. Max 1 addendum cycle to prevent infinite loops.

Phase 3: Cross-Model Debate (3 Rounds)

All debate outputs go to the artifact DB under meta-deep-research-execute / debate / {NNN}D/{file-stem}. The debate covers ALL findings (original + addendum).

• Round 1 — Present: Each model family (Claude/Codex/Gemini) compiles its findings into a position paper. Per sub-question: claim, evidence, confidence, gaps. DB label: {NNN}D/position-{model}
• Round 2 — Challenge: Each model reads the OTHER two positions and attacks them (insufficient evidence, wrong details, contradictions, hallucinations). DB label: {NNN}D/challenge-{model}
• Round 3 — Respond + Converge: Each model responds to challenges with CONCEDE / REBUT / ESCALATE per claim. DB label: {NNN}D/response-{model}

Phase 4: Convergence Scoring

Read all 9 debate files. Score every major claim using extended thinking.

Outcome	Confidence	Action
3/3 agree	VERIFIED	Report as fact
2/3 agree, 1 conceded	HIGH	Report with concession note
2/3 agree, 1 rebutted	CONTESTED	Report both sides
All 3 differ	UNCERTAIN	Open Questions section
No model could rebut challenge	DEBUNKED	Debate trail only
All 3 escalated	UNRESOLVED	Gaps section + follow-up

Source quality weighting: academic papers > official docs > engineering blogs

forums > LLM inference. Recent sources (2025-2026) weighted higher. First-hand experience weighted higher than theory.

Store results in the artifact DB:

source artifacts/db.sh
db_upsert 'meta-deep-research-execute' 'convergence-scoring' '{NNN}D' "$SCORING_CONTENT"

Phase 5: Write Summary

Write to artifacts/research/summary/{NNN}D-{topic-slug}.md (300-500 lines). This is a final output file — create artifacts/research/summary/ if it doesn't exist. Intermediate artifacts (dispatch table, findings, coverage reviews, addendum, debate, convergence scoring) are all in the artifact DB. Structure:

• Header block (models, connectors, source counts, claim counts)
• Executive Summary (10-15 highest-confidence bullets)
• Confidence Map (table of all sub-questions with confidence + agreement)
• Detailed Findings (per sub-question: confidence, agreement, finding, evidence, debate trail)
• Addendum Findings (emergent topics and their impact)
• Contested Findings (both sides presented)
• Open Questions (UNCERTAIN/UNRESOLVED with follow-up suggestions)
• Debunked Claims (hallucinations caught by debate)
• Source Index (academic, docs, web, code — with tally by track)
• Methodology (worker allocation, debate structure, addendum rationale)

Phase 6: Report Completion

Report back to the dispatcher with ONLY:

• Summary file path
• Source tally: {N} queries | {N} scanned | {N} cited
• Claim counts: verified / high / contested / debunked
• Whether addendum cycle ran (and what it added)
• Any CONTESTED findings needing human judgment (one line each)

Homelab Tools memory sync (MANDATORY): Before returning, store the deep research summary in Qdrant. Per cross-cutting rule 7:

• Use mcp__claude_ai_Homelab_Tools__memory_call with tool: 'store_memory'.
• Content: executive summary + source tally + claim counts + contested findings (condensed to ~500 words).
• Tags: research, deep-research, {NNN}D, {project-name}.
• Search first to avoid duplicating a recent entry for the same NNN.

Error Handling

• Gemini unavailable: Try Copilot as fallback for Track D (same 2-slot limit, same prompt patterns). If both fail, use WebSearch. Debate becomes 2-model.
• Codex unavailable: Redistribute to Sonnet. Debate becomes 2-model.
• Both Gemini+Copilot unavailable: WebSearch for Track D. Debate becomes 2-model.
• All CLIs unavailable: Claude only + self-consistency (3 Sonnet subagents). Note "single-model" in methodology.
• Subagent/debate failure: Note gap, mark affected claims UNCERTAIN. Proceed with available responses.
• Coverage review failure: Proceed with available reviews. If ALL fail, write minimal addendum targeting source count gaps only.
• Addendum cycle: Exactly 1 cycle, mandatory. Remaining gaps noted in summary, no looping.

Cost Awareness

~17 workers + mandatory coverage expansion + addendum workers + 3 debate rounds. Up to 5 Opus subagents, 8-10+ Sonnet subagents, 4 Codex workers (max 5 concurrent), 2 Gemini instances (max 2 concurrent), plus 9 debate exchanges. Gemini and Codex phases run sequentially when needed to respect hard concurrency limits. Reserve for decisions where being wrong costs more than the research.

---

Before completing, read and follow ../references/cross-cutting-rules.md.