Autoloop Controller

Wraps improvement-orchestrator in a persistent loop with convergence detection and cost control. Each iteration runs the full 5-stage pipeline (generate, discriminate, evaluate, execute, gate), then checks five termination conditions before deciding whether to continue. State is persisted to disk after every iteration, so the loop survives crashes and can resume across sessions.

When to Use

• Continuously improve a skill over multiple iterations until scores plateau
• Run overnight improvement (Karpathy autoresearch style) where you start the loop, walk away, and review results next morning
• Schedule periodic improvement cycles via system cron (scheduled mode exits after each run, cron triggers the next)
• Resume a previously interrupted improvement run from saved state
• Apply budget-constrained batch improvement when you want to spend at most $N improving a skill
• Drive a skill from EMERGING to SOLID quality tier through compounding gains across rounds
• Detect and halt on oscillation patterns (keep-reject-keep-reject) that waste resources without convergence
• Compare improvement velocity across skills by examining iteration_log.jsonl outputs

When NOT to Use

• Single-shot improvement -- use improvement-orchestrator directly; the autoloop overhead (state persistence, convergence checks) adds no value for one-off runs
• Only want quality scores -- use improvement-learner; the autoloop controller calls the orchestrator which does more than just scoring