---

name: research description: Use this skill for multi-stage, evidence-based research with parallel investigation, source ranking, verification, and synthesis. argument-hint: version: 0.3.0 source: fork checksum: 40502175b1337bd8c4d30d0a3d8acc4097bdbaa442314ce8b76aa478f8a11ffb updated_at: 2026-03-12T18:57:03+08:00 intent: research layer: research

Research Skill

Codex invocation: use $research ... or research: ...

Orchestrate parallel scientist agents for comprehensive, citation-backed research.

Capabilities

• Decompose broad questions into independent research stages.
• Run parallel investigations with bounded concurrency.
• Rank and filter sources by authority and recency.
• Cross-validate findings and resolve contradictions.
• Generate reproducible report artifacts with evidence trails.

Input Requirements

• goal (required): research question or decision to support.
• scope (required): domain/system/time window.
• deliverable (required): summary, comparison, recommendation, or deep report.
• constraints (optional): budget, risk tolerance, compliance context.
• deadline (optional): timebox and required depth.

How to Use

$research Compare OAuth token rotation strategies for Node backends
$research AUTO: Analyze auth failure patterns in this codebase
$research status
$research resume
$research list
$research report <session-id>

Native Subagent Protocol (Codex)

Codex supports native subagents. Delegate with spawn_agent, coordinate with send_input, collect via wait_agent, and clean up with close_agent.

Minimal orchestration pattern:

spawn_agent -> send_input (optional) -> wait_agent -> close_agent

Fallback: if delegation is unavailable, execute stages sequentially in one thread.

Research Operating Rules

1. Retrieval over memory: prefer repository facts and authoritative external sources.
2. Primary sources first: official docs, standards, papers, vendor references.
3. Date awareness: for unstable topics, verify publication/update dates.
4. Evidence minimum: each finding must include at least one source and one artifact.
5. Contradiction handling: unresolved conflicts must be surfaced, not hidden.

Source Quality Rubric

Priority order:

1. Official documentation/specifications
2. Peer-reviewed papers / standards bodies
3. Maintainer-authored technical references
4. High-quality practitioner analysis (only when primary sources are insufficient)

For each citation, capture:

• URL
• source type
• publication/update date
• relevance note

Workflow

Stage 1: Decomposition

Break goal into 3-7 independent stages:

## Research Decomposition

Goal: <original goal>

### Stage N: <name>
- Focus: <what this stage investigates>
- Scope: <files/systems/sources>
- Expected output: <artifact>
- Tier: LOW | MEDIUM | HIGH

Stage 2: Parallel Execution

Run independent stages in parallel (max 5 concurrent scientists).

Tier guidance:

• LOW: enumeration/counting/simple extraction
• MEDIUM: pattern analysis/comparison/synthesis
• HIGH: tradeoff reasoning/architecture/security implications

If stages > 5, execute in batches.

Stage 3: Verification

Cross-check for:

1. Contradictions between stage findings
2. Missing dependencies or blind spots
3. Weak evidence or outdated sources
4. Overstated confidence

Verification output must be one of:

• [VERIFIED]
• [CONFLICTS:<list>]

Stage 4: Synthesis

Produce final report with:

• executive summary
• key findings with confidence
• options/tradeoffs
• recommendations
• limitations and open questions

AUTO Mode

AUTO runs decomposition -> execution -> verification -> synthesis continuously.

Loop control:

• maxIterations: 10 (default)
• stop on completion promise or blocking promise
• persist state after each stage transition

Promise tags:

• [PROMISE:RESEARCH_COMPLETE]
• [PROMISE:RESEARCH_BLOCKED]

Session State

Directory:

.omc/research/{session-id}/
  state.json
  stages/stage-*.md
  findings/raw/
  findings/verified/
  report.md

state.json minimum contract:

{
  "id": "research-<timestamp>-<id>",
  "goal": "",
  "status": "pending|in_progress|complete|blocked|cancelled",
  "mode": "standard|auto",
  "iteration": 1,
  "maxIterations": 10,
  "stages": [],
  "verification": {
    "status": "pending|passed|failed",
    "conflicts": []
  },
  "updatedAt": "ISO-8601"
}

Structured Evidence Tags

Use structured tags in stage outputs:

[FINDING:<id>] <title>
<analysis>
[/FINDING]

[EVIDENCE:<finding-id>]
- Source: <url>
- File: <path or n/a>
- Date: <YYYY-MM-DD>
- Note: <why this evidence supports finding>
[/EVIDENCE]

[CONFIDENCE:HIGH|MEDIUM|LOW]
<confidence rationale>

Quality gates:

• every [FINDING] must include [EVIDENCE]
• every finding must include [CONFIDENCE]
• unsupported claims must be downgraded or removed

Report Template

# Research Report: <goal>

## Executive Summary
<2-3 paragraphs>

## Methodology
- decomposition strategy
- source selection criteria
- verification approach

## Findings
### Finding 1: <title>
Confidence: <HIGH|MEDIUM|LOW>
Evidence: <linked citations>

## Conflicts and Resolutions
<resolved/unresolved contradictions>

## Recommendations
1. <action>
2. <action>

## Limitations
- <gap>
- <gap>

Session Commands

• $research status: show current session state.
• $research resume [session-id]: continue interrupted research.
• $research list: list saved sessions.
• $research report <session-id>: regenerate report from saved stage files.
• $research cancel: stop and persist current state.

Cancellation and Resume

Cancellation keeps .omc/research/{session-id}/ intact.

Resume loads state.json and continues pending stages only.

Failure Handling

Stop and emit [PROMISE:RESEARCH_BLOCKED] when:

• required data is inaccessible
• source quality is below minimum and cannot be improved
• contradictions cannot be resolved within iteration budget

Always include blocker details and recommended next action.