science

Research Skill

Orchestrate parallel scientist agents for comprehensive research workflows with optional AUTO mode for fully autonomous execution.

Overview

Research is a multi-stage workflow that decomposes complex research goals into parallel investigations:

1. Decomposition - Break research goal into independent stages/hypotheses
2. Execution - Run parallel scientist agents on each stage
3. Verification - Cross-validate findings, check consistency
4. Synthesis - Aggregate results into comprehensive report

Usage Examples

/science <goal>                    # Standard research with user checkpoints
/science AUTO: <goal>              # Fully autonomous until complete
/science status                    # Check current research session status
/science resume                    # Resume interrupted research session
/science list                      # List all research sessions
/science report <session-id>       # Generate report for session

Quick Examples

/science What are the performance characteristics of different sorting algorithms?
/science AUTO: Analyze authentication patterns in this codebase
/science How does the error handling work across the API layer?

Research Protocol

Stage Decomposition Pattern

When given a research goal, decompose into 3-7 independent stages:

## Research Decomposition

**Goal:** <original research goal>

### Stage 1: <stage-name>
- **Focus:** What this stage investigates
- **Hypothesis:** Expected finding (if applicable)
- **Scope:** Files/areas to examine
- **Tier:** LOW | MEDIUM | HIGH

### Stage 2: <stage-name>
...

Parallel Scientist Invocation

Fire independent stages in parallel via Task tool:

// Stage 1 - Simple data gathering
Task(subagent_type="scientist", model="haiku", prompt="[RESEARCH_STAGE:1] Investigate...")

// Stage 2 - Standard analysis
Task(subagent_type="scientist", model="sonnet", prompt="[RESEARCH_STAGE:2] Analyze...")

// Stage 3 - Complex reasoning
Task(subagent_type="scientist", model="opus", prompt="[RESEARCH_STAGE:3] Deep analysis of...")

Smart Model Routing

CRITICAL: Always pass model parameter explicitly!

Task Complexity	Agent	Model	Use For
Data gathering	scientist (model=haiku)	haiku	File enumeration, pattern counting, simple lookups
Standard analysis	scientist	sonnet	Code analysis, pattern detection, documentation review
Complex reasoning	scientist	opus	Architecture analysis, cross-cutting concerns, hypothesis validation

Routing Decision Guide

Research Task	Tier	Example Prompt
"Count occurrences of X"	LOW	"Count all usages of useState hook"
"Find all files matching Y"	LOW	"List all test files in the project"
"Analyze pattern Z"	MEDIUM	"Analyze error handling patterns in API routes"
"Document how W works"	MEDIUM	"Document the authentication flow"
"Explain why X happens"	HIGH	"Explain why race conditions occur in the cache layer"
"Compare approaches A vs B"	HIGH	"Compare Redux vs Context for state management here"

Verification Loop

After parallel execution completes, verify findings:

// Cross-validation stage
Task(subagent_type="scientist", model="sonnet", prompt="
[RESEARCH_VERIFICATION]
Cross-validate these findings for consistency:

Stage 1 findings: <summary>
Stage 2 findings: <summary>
Stage 3 findings: <summary>

Check for:
1. Contradictions between stages
2. Missing connections
3. Gaps in coverage
4. Evidence quality

Output: [VERIFIED] or [CONFLICTS:<list>]
")

AUTO Mode

AUTO mode runs the complete research workflow autonomously with loop control.

Loop Control Protocol

[RESEARCH + AUTO - ITERATION {{ITERATION}}/{{MAX}}]

Your previous attempt did not output the completion promise. Continue working.

Current state: {{STATE}}
Completed stages: {{COMPLETED_STAGES}}
Pending stages: {{PENDING_STAGES}}

Promise Tags

Tag	Meaning	When to Use
[PROMISE:RESEARCH_COMPLETE]	Research finished successfully	All stages done, verified, report generated
[PROMISE:RESEARCH_BLOCKED]	Cannot proceed	Missing data, access issues, circular dependency

AUTO Mode Rules

1. Max Iterations: 10 (configurable)
2. Continue until: Promise tag emitted OR max iterations
3. State tracking: Persist after each stage completion
4. Cancellation: /cancel or "stop", "cancel"

AUTO Mode Example

/science AUTO: Comprehensive security analysis of the authentication system

[Decomposition]
- Stage 1 (LOW): Enumerate auth-related files
- Stage 2 (MEDIUM): Analyze token handling
- Stage 3 (MEDIUM): Review session management
- Stage 4 (HIGH): Identify vulnerability patterns
- Stage 5 (MEDIUM): Document security controls

[Execution - Parallel]
Firing stages 1-3 in parallel...
Firing stages 4-5 after dependencies complete...

[Verification]
Cross-validating findings...

[Synthesis]
Generating report...

[PROMISE:RESEARCH_COMPLETE]

Parallel Execution Patterns

Independent Dataset Analysis (Parallel)

When stages analyze different data sources:

// All fire simultaneously
Task(subagent_type="scientist", model="haiku", prompt="[STAGE:1] Analyze src/api/...")
Task(subagent_type="scientist", model="haiku", prompt="[STAGE:2] Analyze src/utils/...")
Task(subagent_type="scientist", model="haiku", prompt="[STAGE:3] Analyze src/components/...")

Hypothesis Battery (Parallel)

When testing multiple hypotheses:

// Test hypotheses simultaneously
Task(subagent_type="scientist", model="sonnet", prompt="[HYPOTHESIS:A] Test if caching improves...")
Task(subagent_type="scientist", model="sonnet", prompt="[HYPOTHESIS:B] Test if batching reduces...")
Task(subagent_type="scientist", model="sonnet", prompt="[HYPOTHESIS:C] Test if lazy loading helps...")

Cross-Validation (Sequential)

When verification depends on all findings:

// Wait for all parallel stages
[stages complete]

// Then sequential verification
Task(subagent_type="scientist", model="opus", prompt="
[CROSS_VALIDATION]
Validate consistency across all findings:
- Finding 1: ...
- Finding 2: ...
- Finding 3: ...
")

Concurrency Limit

Maximum 20 concurrent scientist agents to prevent resource exhaustion.

If more than 20 stages, batch them:

Batch 1: Stages 1-5 (parallel)
[wait for completion]
Batch 2: Stages 6-7 (parallel)

Session Management

Directory Structure

.omc/research/{session-id}/
  state.json              # Session state and progress
  stages/
    stage-1.md            # Stage 1 findings
    stage-2.md            # Stage 2 findings
    ...
  findings/
    raw/                  # Raw findings from scientists
    verified/             # Post-verification findings
  figures/
    figure-1.png          # Generated visualizations
    ...
  report.md               # Final synthesized report

State File Format

{
  "id": "research-20240115-abc123",
  "goal": "Original research goal",
  "status": "in_progress | complete | blocked | cancelled",
  "mode": "standard | auto",
  "iteration": 3,
  "maxIterations": 10,
  "stages": [
    {
      "id": 1,
      "name": "Stage name",
      "tier": "LOW | MEDIUM | HIGH",
      "status": "pending | running | complete | failed",
      "startedAt": "ISO timestamp",
      "completedAt": "ISO timestamp",
      "findingsFile": "stages/stage-1.md"
    }
  ],
  "verification": {
    "status": "pending | passed | failed",
    "conflicts": [],
    "completedAt": "ISO timestamp"
  },
  "createdAt": "ISO timestamp",
  "updatedAt": "ISO timestamp"
}

Session Commands

Command	Action
/science status	Show current session progress
/science resume	Resume most recent interrupted session
/science resume <session-id>	Resume specific session
/science list	List all sessions with status
/science report <session-id>	Generate/regenerate report
/science cancel	Cancel current session (preserves state)

Tag Extraction

Scientists use structured tags for findings. Extract them with these patterns:

Finding Tags

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

[EVIDENCE:<finding-id>]
- File: <path>
- Lines: <range>
- Content: <relevant code/text>
[/EVIDENCE]

[CONFIDENCE:<level>] # HIGH | MEDIUM | LOW
<reasoning for confidence level>

Extraction Regex Patterns

// Finding extraction
const findingPattern = /\[FINDING:(\w+)\]\s*(.*?)\n([\s\S]*?)\[\/FINDING\]/g;

// Evidence extraction
const evidencePattern = /\[EVIDENCE:(\w+)\]([\s\S]*?)\[\/EVIDENCE\]/g;

// Confidence extraction
const confidencePattern = /\[CONFIDENCE:(HIGH|MEDIUM|LOW)\]\s*(.*)/g;

// Stage completion
const stageCompletePattern = /\[STAGE_COMPLETE:(\d+)\]/;

// Verification result
const verificationPattern = /\[(VERIFIED|CONFLICTS):?(.*?)\]/;

Evidence Window

When extracting evidence, include context window:

[EVIDENCE:F1]
- File: /src/auth/login.ts
- Lines: 45-52 (context: 40-57)
- Content:
  ```typescript
  // Lines 45-52 with 5 lines context above/below

[/EVIDENCE]

### Quality Validation

Findings must meet quality threshold:

| Quality Check | Requirement |
|---------------|-------------|
| Evidence present | At least 1 [EVIDENCE] per [FINDING] |
| Confidence stated | Each finding has [CONFIDENCE] |
| Source cited | File paths are absolute and valid |
| Reproducible | Another agent could verify |

## Report Generation

### Report Template

```markdown
# Research Report: {{GOAL}}

**Session ID:** {{SESSION_ID}}
**Date:** {{DATE}}
**Status:** {{STATUS}}

## Executive Summary

{{2-3 paragraph summary of key findings}}

## Methodology

### Research Stages

| Stage | Focus | Tier | Status |
|-------|-------|------|--------|
{{STAGES_TABLE}}

### Approach

{{Description of decomposition rationale and execution strategy}}

## Key Findings

### Finding 1: {{TITLE}}

**Confidence:** {{HIGH|MEDIUM|LOW}}

{{Detailed finding with evidence}}

#### Evidence

{{Embedded evidence blocks}}

### Finding 2: {{TITLE}}
...

## Visualizations

{{FIGURES}}

## Cross-Validation Results

{{Verification summary, any conflicts resolved}}

## Limitations

- {{Limitation 1}}
- {{Limitation 2}}
- {{Areas not covered and why}}

## Recommendations

1. {{Actionable recommendation}}
2. {{Actionable recommendation}}

## Appendix

### Raw Data

{{Links to raw findings files}}

### Session State

{{Link to state.json}}

Figure Embedding Protocol

Scientists generate visualizations using this marker:

[FIGURE:path/to/figure.png]
Caption: Description of what the figure shows
Alt: Accessibility description
[/FIGURE]

Report generator embeds figures:

## Visualizations

![Figure 1: Description](figures/figure-1.png)
*Caption: Description of what the figure shows*

![Figure 2: Description](figures/figure-2.png)
*Caption: Description of what the figure shows*

Figure Types

Type	Use For	Generated By
Architecture diagram	System structure	scientist
Flow chart	Process flows	scientist
Dependency graph	Module relationships	scientist
Timeline	Sequence of events	scientist
Comparison table	A vs B analysis	scientist

Configuration

Optional settings in .claude/settings.json:

{
  "omc": {
    "research": {
      "maxIterations": 10,
      "maxConcurrentScientists": 5,
      "defaultTier": "MEDIUM",
      "autoVerify": true,
      "generateFigures": true,
      "evidenceContextLines": 5
    }
  }
}

Cancellation

/cancel

Or say: "stop research", "cancel research", "abort"

Progress is preserved in .omc/research/{session-id}/ for resume.

Troubleshooting

Stuck in verification loop?

• Check for conflicting findings between stages
• Review state.json for specific conflicts
• May need to re-run specific stages with different approach

Scientists returning low-quality findings?

• Check tier assignment - complex analysis needs HIGH tier
• Ensure prompts include clear scope and expected output format
• Review if research goal is too broad

AUTO mode exhausted iterations?

• Review state to see where it's stuck
• Check if goal is achievable with available data
• Consider breaking into smaller research sessions

Missing figures in report?

• Verify figures/ directory exists
• Check [FIGURE:] tags in findings
• Ensure paths are relative to session directory