autoresearch

name	autoresearch
description	Autonomous codebase improvement loop inspired by Karpathy's autoresearch. USE WHEN user wants to iteratively improve a codebase, run autonomous code improvement, or apply the autoresearch pattern. Individual commands use /autoresearch directly.

Autoresearch

Autonomous codebase improvement loop that converges on measurable improvements through iterative improve-evaluate-iterate cycles.

Quick Start

/autoresearch                              # Interactive discovery mode
/autoresearch src/ --profile quality       # Quality-focused on src/
/autoresearch --profile coverage           # Maximize test coverage
/autoresearch --profile performance        # Optimize performance
/autoresearch --resume                     # Resume a previous run
/autoresearch --dry-run                    # Preview what would be evaluated

How It Works

Autoresearch runs a tight loop inspired by Karpathy's autoresearch pattern:

┌─ DISCOVER ──────────────────────────────┐
│ Analyze codebase → propose constraints  │
│ → interview user via AskUserQuestion    │
│ → lock evaluation commands              │
└─────────────────────────────────────────┘
         ↓
┌─ BASELINE ──────────────────────────────┐
│ Create git branch → run all evaluators  │
│ → capture baseline scores              │
└─────────────────────────────────────────┘
         ↓
┌─ LOOP (until convergence) ──────────────┐
│ Improve → Evaluate → Decide → Track    │
│                                         │
│ Keep if score improves (git commit)     │
│ Revert if score regresses (git reset)   │
│ Stop on diminishing returns             │
└─────────────────────────────────────────┘
         ↓
┌─ REPORT ────────────────────────────────┐
│ Full LLM evaluation → learning report   │
│ → improvement table → convergence data  │
└─────────────────────────────────────────┘

Arguments

Argument	Description	Default