DS-STAR: Multi-Agent Data Science Framework

Overview

DS-STAR (Data Science - Structured Thought and Action) is an intelligent multi-agent framework adapted for Claude Code that automates data science workflows through specialized AI agents. Originally designed for Google's Gemini models, this skill adapts DS-STAR to leverage Claude's model family (Haiku, Sonnet, Opus) for cost-efficient, high-quality data analysis.

What It Does

DS-STAR orchestrates seven specialized agents that work collaboratively to:

1. •Analyze data files (structured/unstructured)
2. •Plan analysis strategies iteratively
3. •Generate Python code to execute plans
4. •Execute code with automatic debugging
5. •Verify results against the query
6. •Route refinement decisions
7. •Finalize formatted outputs

Key Innovation: Multi-model optimization - route exploration tasks to Haiku, complex reasoning to Sonnet, and critical analysis to Opus for optimal cost/performance balance.

When to Use This Skill

• •Exploratory Data Analysis (EDA): Automated insights from CSV, JSON, or text files
• •Data Science Questions: Answer factual questions from data programmatically
• •Iterative Analysis: Refine analyses through multiple rounds of verification
• •Cost-Optimized Workflows: Use cheaper models for simple tasks, powerful models for complex reasoning
• •Reproducible Research: All artifacts (prompts, code, results) are saved for audit trails

Performance Characteristics

Estimated Cost (typical analysis):

• •Traditional (all Sonnet): ~$0.50-1.00
• •DS-STAR (mixed models): ~$0.20-0.40
• •Savings: 50-60%

How It Works

Multi-Agent Pipeline

┌─────────────────────────────────────────────────────────┐
│  1. ANALYZER (Haiku)                                    │
│     Input: Data files                                   │
│     Output: Data descriptions, schemas, summaries       │
└─────────────────────────────────────────────────────────┘
                         ↓
┌─────────────────────────────────────────────────────────┐
│  2. PLANNER (Sonnet)                                    │
│     Input: Query + Data descriptions                    │
│     Output: Initial analysis step                       │
└─────────────────────────────────────────────────────────┘
                         ↓
┌─────────────────────────────────────────────────────────┐
│  3. CODER (Sonnet)                                      │
│     Input: Plan + Data descriptions                     │
│     Output: Python code to execute plan                 │
└─────────────────────────────────────────────────────────┘
                         ↓
┌─────────────────────────────────────────────────────────┐
│  4. DEBUGGER (Sonnet) [If needed]                       │
│     Input: Code + Error                                 │
│     Output: Fixed code                                  │
└─────────────────────────────────────────────────────────┘
                         ↓
┌─────────────────────────────────────────────────────────┐
│  5. VERIFIER (Sonnet)                                   │
│     Input: Code + Results + Query                       │
│     Output: "Sufficient" or "Needs refinement"          │
└─────────────────────────────────────────────────────────┘
                         ↓
           ┌─────────────┴─────────────┐
           │    Sufficient?            │
           └─────────────┬─────────────┘
                 YES │   │ NO
                     │   └──────────────┐
                     │                  ↓
                     │    ┌──────────────────────────────┐
                     │    │  6. ROUTER (Haiku)           │
                     │    │     Decide: Fix step or Add  │
                     │    └──────────────────────────────┘
                     │                  │
                     │                  ↓
                     │         [Loop back to PLANNER]
                     │
                     ↓
┌─────────────────────────────────────────────────────────┐
│  7. FINALYZER (Sonnet)                                  │
│     Input: Final code + Results + Query                 │
│     Output: Formatted answer (JSON/text)                │
└─────────────────────────────────────────────────────────┘

Iterative Refinement

DS-STAR uses up to 5 refinement rounds (configurable):

• •Each round can add a new analysis step or fix a previous step
• •Automatic debugging retries failed code
• •Verification gates ensure quality before finalization
• •All artifacts saved in runs/<run_id>/ for reproducibility

Installation

Quick Start

# Add the marketplace (if not already added)
/plugin marketplace add token-eater/skills-marketplace

# Install DS-STAR skill
/plugin install ds-star

Manual Installation

# Clone the marketplace
git clone https://github.com/token-eater/skills-marketplace.git
cd skills-marketplace

# The skill is in skills/ds-star/

Usage

Basic Usage

# In Claude Code, invoke the skill
/skill ds-star

# The skill will prompt you for:
# 1. Data files (CSV, JSON, TXT, etc.)
# 2. Your analysis query
# 3. Model configuration (optional)

Example Queries

1. •

   Basic Statistics:

   code

   Query: "What is the average age and gender distribution in this dataset?"
   Data: customers.csv
   

2. •

   Time Series Analysis:

   code

   Query: "What are the monthly sales trends over the past year?"
   Data: sales_2024.csv
   

3. •

   Data Quality:

   code

   Query: "How many missing values are in each column, and what percentage of the total?"
   Data: survey_responses.csv
   

4. •

   Correlation Analysis:

   code

   Query: "Which features have the strongest correlation with customer churn?"
   Data: customer_features.csv
   

Advanced Configuration

Configure which models to use for each agent:

# In your query, you can specify:
agent_models:
  ANALYZER: haiku      # Fast data inspection
  PLANNER: sonnet      # Strategic planning
  CODER: sonnet        # Code generation
  VERIFIER: sonnet     # Result validation
  ROUTER: haiku        # Simple routing decisions
  DEBUGGER: opus       # Complex debugging (if needed)
  FINALYZER: sonnet    # Output formatting

Or use presets:

• •Fast (70% cost savings): All Haiku except Coder/Debugger
• •Balanced (50% cost savings): Haiku for simple, Sonnet for complex
• •Quality (Highest accuracy): All Sonnet or Opus

Output Structure

Every run creates a structured artifact directory:

runs/<run_id>/
├── steps/
│   ├── 001_analyzer/
│   │   ├── prompt.md       # Agent prompt
│   │   ├── code.py         # Generated code
│   │   ├── result.txt      # Execution output
│   │   └── metadata.json   # Step metadata
│   ├── 002_planner_init/
│   ├── 003_coder/
│   └── ...
├── exec_env/               # Executed scripts
├── logs/
│   ├── pipeline.log        # Structured logs
│   └── execution.log       # Code execution logs
├── final_output/
│   └── result.json         # Final answer
└── pipeline_state.json     # Resume state

Resume Support

DS-STAR supports resuming interrupted runs:

# Resume from a previous run ID
/skill ds-star --resume 20241123_143022_a1b2c3

Multi-Model Optimization

Cost Comparison (50K token analysis)

Model Selection Guidelines

Use Haiku for:

• •Data file inspection (Analyzer)
• •Simple routing decisions (Router)
• •Basic extraction tasks

Use Sonnet for:

• •Planning strategies (Planner)
• •Code generation (Coder)
• •Result verification (Verifier)
• •Output formatting (Finalyzer)

Use Opus for:

• •Complex debugging scenarios
• •Critical analysis requiring highest accuracy
• •Novel/ambiguous data formats

Examples

Example 1: Iris Dataset Analysis

Data: iris.csv (150 rows, 5 columns)

Query: "What is the average petal length for each species, and which species has the highest variance?"

Result:

{
  "final_answer": {
    "averages": {
      "setosa": 1.46,
      "versicolor": 4.26,
      "virginica": 5.55
    },
    "highest_variance": "virginica",
    "variance_value": 0.304
  }
}

Cost: $0.18 (Balanced configuration) Time: ~25 seconds Steps: 3 (Analyze → Plan → Code → Verify → Finalize)

Example 2: Sales Data Time Series

Data: sales_2024.csv (365 rows, daily sales)

Query: "Identify the top 3 months by total sales and calculate month-over-month growth rates."

Result:

{
  "final_answer": {
    "top_3_months": [
      {"month": "December", "total_sales": 125000, "growth": "+15%"},
      {"month": "November", "total_sales": 108000, "growth": "+8%"},
      {"month": "July", "total_sales": 95000, "growth": "+12%"}
    ],
    "average_growth": "+8.3%"
  }
}

Cost: $0.22 (Balanced configuration) Time: ~30 seconds Steps: 4 (with 1 refinement round)

Architecture

Claude Provider Implementation

The skill includes a custom ClaudeProvider that integrates with the Claude Agent SDK:

class ClaudeProvider(ModelProvider):
    """Provider for Claude models via Agent SDK."""

    def __init__(self, model_name: str = "sonnet"):
        self.model_name = model_name  # haiku, sonnet, or opus

    def generate_content(self, prompt: str) -> str:
        # Use SDK's subagent system for context efficiency
        # Automatically handles model routing
        # Returns generated content

Integration with Claude Code

• •Context Efficient: Uses subagents to avoid context bloat
• •Parallel Execution: Independent agents can run concurrently
• •Artifact Persistence: All outputs saved to project directory
• •Resume Support: Can pause/resume long analyses

Troubleshooting

Common Issues

1. "Missing data files" error

• •Ensure files are in the data/ directory
• •Use absolute paths or relative to project root
• •Check file permissions

2. Code execution timeout

• •Increase execution_timeout in config (default: 60s)
• •Simplify the query or split into multiple queries
• •Check for infinite loops in generated code

3. "API key not found"

• •Claude Agent SDK handles auth automatically
• •No API keys needed for built-in Claude models
• •External models (OpenAI, Gemini) require env vars

4. Verification always fails

• •Query may be too ambiguous - make it more specific
• •Increase max_refinement_rounds (default: 5)
• •Check intermediate results in runs/<id>/steps/

Debug Mode

Enable detailed logging:

# Run with interactive mode to pause at each step
/skill ds-star --interactive

# Or inspect artifacts manually
ls runs/<run_id>/steps/
cat runs/<run_id>/logs/pipeline.log

Configuration

Config Options

# config.yaml (optional, in project root)
run_id: "my_experiment"           # Custom run ID
model_name: "sonnet"               # Default model
interactive: false                 # Pause between steps
max_refinement_rounds: 5           # Max iterations
execution_timeout: 60              # Code timeout (seconds)
preserve_artifacts: true           # Save all outputs
runs_dir: "runs"                   # Artifact directory
data_dir: "data"                   # Data file directory

# Agent-specific models
agent_models:
  ANALYZER: haiku
  PLANNER: sonnet
  CODER: sonnet
  VERIFIER: sonnet
  ROUTER: haiku
  DEBUGGER: opus
  FINALYZER: sonnet

Comparison with Original

Best Practices

Query Design

✅ Good Queries:

• •"Calculate the correlation between X and Y columns"
• •"What percentage of users have more than 3 purchases?"
• •"Identify outliers in the price column using IQR method"

❌ Avoid:

• •"Tell me about this data" (too vague)
• •"Do everything" (use multiple specific queries)
• •Queries requiring external data not in files

Data Preparation

✅ Best Formats:

• •CSV with headers
• •JSON with consistent structure
• •Clean text files (UTF-8)

❌ Problematic:

• •Excel files with multiple sheets (convert to CSV first)
• •PDFs (extract text first)
• •Binary formats (images, etc.)

Model Selection

• •Start with Balanced configuration (50% cost savings)
• •Use Fast for large datasets (>100K rows)
• •Use Quality (Opus) only for novel/complex analyses
• •Override specific agents as needed (e.g., Opus for Debugger only)

Limitations

• •Python-only execution: Generated code is Python (pandas, numpy, etc.)
• •Local execution: Code runs locally (be cautious with untrusted data)
• •No external APIs: Agents can't call external services (by design)
• •Single-file outputs: Finalyzer produces one result file
• •No interactive visualizations: Text/JSON outputs only (no plots saved)

Contributing

Extending the Framework

Want to add new agents or capabilities?

1. •Add new agent type in scripts/dsstar.py
2. •Create prompt template in scripts/prompts.py
3. •Wire up in pipeline in run_pipeline() method
4. •Test with sample data
5. •Submit PR with examples

Ideas for Extensions

• •Visualization Agent: Generate matplotlib/seaborn plots
• •Export Agent: Save results to various formats (Excel, PDF, etc.)
• •Schema Agent: Automatically detect and validate data schemas
• •Anomaly Agent: Specialized outlier detection
• •ML Agent: Train simple models (regression, classification)

License

MIT License - Adapted from DS-Star by Jules Lescx.

Original DS-STAR based on Google Research paper.

Support & Resources

• •📖 Original Paper: DS-STAR: Domain Adaptive Data Science via Large Language Models
• •💾 Original Repository: github.com/JulesLscx/DS-Star
• •🐛 Report Issues: Skills Marketplace Issues
• •💬 Discussions: Skills Marketplace Discussions

Ready to automate your data science workflows? Install DS-STAR and let multi-agent AI handle the heavy lifting!