Autoloop Controller#

Wraps improvement-orchestrator in a persistent loop with convergence detection and cost control.

When to Use#

• Continuously improve a skill over multiple iterations (overnight runs)
• Schedule periodic improvement cycles via cron
• Run a fixed number of improvement iterations with automatic plateau detection

When NOT to Use#

• Single-shot improvement → use improvement-orchestrator
• Only want scores → use improvement-learner
• Only want baseline data → use benchmark-store

CLI#

bash

python3 scripts/autoloop.py \
  --target /path/to/skill \        # REQUIRED: skill directory to improve
  --state-root /path/to/state \    # REQUIRED: persistent state directory
  --max-iterations 5 \             # default 5: total iterations before stop
  --max-cost 50.0 \                # default 50.0: cost cap in USD
  --plateau-window 3 \             # default 3: consecutive no-improvement rounds → stop
  --cooldown-minutes 30 \          # default 30: delay between iterations in continuous mode
  --mode single-run \              # single-run | continuous | scheduled
  --dry-run                        # simulate without calling orchestrator

ParamDefaultWhen to change--max-iterations5Raise to 20 for overnight runs; lower to 2 for quick tests--max-cost$50.0Lower to $5 for testing; raise for production overnight runs--plateau-window3Raise to 5 if improvements are slow but steady--cooldown-minutes30Lower to 5 for rapid iteration; raise to 60 for rate-limited APIs--dry-runfalseUse to test loop logic without running real orchestrator

3 Modes#

ModeBehaviorExit Conditionsingle-runLoop through iterations, exit when termination hitmax_iterations / cost_cap / plateau / oscillationcontinuousLoop with cooldown_minutes sleep between iterationsSame + Ctrl+C saves state gracefullyscheduledRun exactly 1 iteration then exitCron triggers next run; state persists in autoloop_state.json

Termination (4 conditions, OR logic)#

1. max_iterations reached
2. cost_cap exceeded (rough estimate: ~$0.10/min of LLM time)
3. Score plateau: weighted_score no improvement for plateau_window consecutive rounds
4. Oscillation: keep-reject alternating pattern detected over 4-round window

Cross-Session State#

State persists in {state-root}/autoloop_state.json (dataclass → JSON):

json

{"schema_version": "1.0", "target": "/path/to/skill",
 "iterations_completed": 3, "max_iterations": 5,
 "total_cost_usd": 1.23, "max_cost_usd": 50.0,
 "score_history": [{"iteration": 1, "weighted_score": 0.72, "decision": "keep"}, ...],
 "plateau_counter": 1, "status": "running"}

Resume a stopped run: just re-run the same command — AutoloopState.load() picks up from where it left off.

Output Artifacts#

• handoffs/iteration-N.md — per-iteration handoff with Decided/Rejected/Scores/Remaining sections for cross-iteration context survival
• iteration_log.jsonl — one JSON line per iteration: {"iteration": 1, "decision": "keep", "weighted_score": 0.72, "cost_usd": 0.15, "candidate_id": "cand-01-docs"}
• autoloop_state.json — full serialized state for cross-session resume

Run 5 iterations with cost cap: $ python3 scripts/autoloop.py --target /path/to/skill --state-root ./state --max-iterations 5 --max-cost 10.0 → --- Iteration 1/5 --- → Decision: keep Weighted score: 0.7200 Cumulative cost: $0.1500 → --- Iteration 2/5 --- → Decision: keep Weighted score: 0.7800 Cumulative cost: $0.3100 → --- Iteration 3/5 --- → Decision: keep Weighted score: 0.7800 Cumulative cost: $0.4600 → Stopped: plateau detected (no improvement in last 3 iterations) → State saved: ./state/autoloop_state.json

Running continuous mode without checking cost cap: $ python3 scripts/autoloop.py --target /path/to/skill --state-root ./state --mode continuous --max-cost 50.0 → Default $50 cap applies. For overnight runs, explicitly set --max-cost. → Always verify cost cap BEFORE launching continuous mode.

Error Handling#

• Orchestrator subprocess failure: state.status = "error", stderr saved to last_failure_trace, loop exits
• KeyboardInterrupt (Ctrl+C): state saved as status=completed, graceful exit
• Unhandled exceptions: traceback saved to last_failure_trace, status=error, exit code 1

Related Skills#

• improvement-orchestrator: Single pipeline run (called as subprocess each iteration)
• improvement-learner: Provides dimension scores consumed by _load_latest_scores() for weighted_score computation
• benchmark-store: Pareto front data for convergence tracking