OVERNIGHT_USAGE.md

Overnight Autonomous Evolution

Run the Skill Organism evolution engine overnight for continuous improvement.

Quick start

# Run 10 generations across all 27 skills (default)
python skill-organism/enterprise_runner.py --overnight

# Custom generation count
python skill-organism/enterprise_runner.py --overnight --generations 20

# Single skill evolution
python skill-organism/enterprise_runner.py --skill security-audit --generations 5

How it works

1. ARENA.md defines the fitness function, mutation strategies, and experiment budget for each skill
2. enterprise_runner.py orchestrates the evolution cycle with:
   • File integrity verification (SHA-256 checksums)
   • Atomic registry updates with backup/rollback
   • Health gate (aborts if >30% skills are critical)
   • EvalBudget context manager (30s per skill per generation)
3. eval_judge.py scores each mutation on 5 dimensions
4. Git-per-experiment branches isolate mutations; only improvements merge
5. evolution_log.jsonl captures every event for the dashboard

Parallel execution

The overnight runner uses asyncio.Semaphore(4) for concurrent skill evolution. Each generation processes up to 4 skills simultaneously.

Monitoring

While the overnight run executes:

# Watch the evolution log
tail -f skill-organism/evolution_log.jsonl | python -m json.tool

# Open the dashboard
open dashboards/evolution_dashboard.html

Budget controls

Each skill evaluation is capped at eval_budget_seconds (default: 30s) from its ARENA.md. The total overnight budget is approximately: generations × skills × budget_seconds.

For 10 generations × 27 skills × 30s = ~135 minutes of evaluation time.

Safety

• Lockfile prevents concurrent evolution cycles
• Pre-flight and post-flight integrity checks on the skill registry
• Automatic rollback if the registry becomes corrupted
• Signal handlers for graceful SIGTERM/SIGINT shutdown

Trilogy integration

When ARENA.md has mindspider_feed_enabled: true, the evolution loop will:

1. Check datasets/mindspider/today_topics.json for live data
2. Fall back to synthetic data if unavailable
3. Log trilogy integration status in each evolution event