---

name: research-executor description: "Execute research experiments using TDD methodology. Pops plans from .claude/plans/research_tasks/, implements with smoke/unit/integration tests, and documents results to results.md. Use when: (1) executing the next experiment from the queue, (2) executing a specific plan by number, (3) running an ad-hoc research idea with TDD." user_invocable: true

Research Executor

Execute research experiments using Test-Driven Development (TDD).

Plan Queue

Plans are stored in .claude/plans/research_tasks/plan-*.md.

Pop and Execute

1. List plans: ls .claude/plans/research_tasks/plan-*.md
2. Select plan: Default is plan-1.md, or specify plan-N
3. Move to executed:

   Copymkdir -p .claude/plans/research_tasks/executed
   mv .claude/plans/research_tasks/plan-{N}.md .claude/plans/research_tasks/executed/{YYYY-MM-DD}_{name}.md
   

4. Renumber remaining plans sequentially (plan-2 → plan-1, plan-3 → plan-2, etc.)

If no plans exist, ask user for an ad-hoc research idea.

Execution Workflow

Phase 1: Setup

• Parse hypothesis, variables, success criteria from plan
• Create directory:

  Copyexperiments/{experiment_name}/
  ├── README.md
  ├── notebook.ipynb          # Primary deliverable
  ├── tests/
  │   ├── smoke/
  │   ├── unit/
  │   └── integration/
  ├── src/
  └── results/
      └── figures/            # All generated plots
  

Phase 2: Smoke Tests

• Test data loading, API calls, metric computation
• Implement minimal code to pass
• Run: uv run pytest experiments/{name}/tests/smoke/ -v

Phase 3: Unit Tests

• Test each component (preprocessing, features, model, evaluation)
• TDD: Red → Green → Refactor
• Run: uv run pytest experiments/{name}/tests/unit/ -v

Phase 4: Integration Tests

• Test full pipeline end-to-end
• Run on sampled data first
• Run: uv run pytest experiments/{name}/tests/integration/ -v

Phase 5: Finalize

• Run full experiment
• Create Jupyter notebook deliverable (see below)
• Document results (see Results Documentation)
• Commit changes

Jupyter Notebook Deliverable

Required: Every experiment MUST produce a Jupyter notebook at experiments/{name}/notebook.ipynb.

Notebook Structure

Copy1. Header & Setup
   - Experiment title, date, hypothesis
   - Import statements and configuration

2. Data Loading & Exploration
   - Load experimental data
   - Show sample data, shapes, dtypes
   - Basic statistics (describe(), value_counts())

3. Implementation
   - Core experiment code with explanatory markdown cells
   - Each major step in its own cell for re-runnability

4. Results & Visualizations
   - All figures generated inline (use %matplotlib inline)
   - Statistical tests with p-values, confidence intervals
   - Summary tables of key metrics

5. Conclusions
   - Key findings from the experiment
   - Link to hypothesis: supported/refuted/inconclusive
   - Next steps or follow-up questions

Requirements

• All cells must be executed (outputs saved in notebook)
• Use markdown cells to explain each section
• Save figures both inline AND to experiments/{name}/results/figures/
• Include reproducibility info: random seeds, package versions
• Final cell: print summary statistics and status

Results Documentation

Create results.md in .claude/plans/research_tasks/executed/ alongside the executed plan:

Copy# Experiment: {name}

**Date**: {YYYY-MM-DD}
**Plan**: {original plan path}
**Status**: success | failure | partial

## Hypothesis
{from plan}

## Test Results
- Smoke: X/Y passing
- Unit: X/Y passing
- Integration: X/Y passing

## Findings
{key observations, metrics, conclusions}

## Artifacts
- **Notebook**: `experiments/{name}/notebook.ipynb` (primary deliverable)
- Code: `experiments/{name}/src/`
- Figures: `experiments/{name}/results/figures/`
- Data: `experiments/{name}/results/`

Code Standards

• Type hints for all functions
• Docstrings for public APIs
• Reproducible random seeds (document in README)
• Use uv run ruff check and uv run ruff format before commits