import argparse import json import os from pathlib import Path from .config import ROOT def main(): parser = argparse.ArgumentParser( description="Local synthetic sealing cell and GLiNER evaluation" ) parser.add_argument( "--device", choices=["cpu", "mps"], default="cpu", help="CPU reference (default) or experimental Apple GPU inference", ) commands = parser.add_subparsers(dest="command", required=True) commands.add_parser( "download-model", help="Download the pinned local snapshot; network required once" ) probe = commands.add_parser("probe") probe.add_argument("--fixtures", default="data/fixtures/playground-probes.json") probe.add_argument("--output", default="reports/local-probes.jsonl") probe.add_argument("--direct", action="store_true") sim = commands.add_parser("simulate") sim.add_argument( "--policy", choices=["rules", "rules-no-text", "hold", "hybrid", "direct"], default="rules" ) sim.add_argument("--scenario", default="low-pressure") sim.add_argument("--seed", type=int, default=42) sim.add_argument("--recipe", default=None) sim.add_argument("--output", default="reports/simulation.json") evaluate = commands.add_parser("evaluate") evaluate.add_argument("--config", default="configs/evaluation.yaml") evaluate.add_argument( "--split", choices=["training", "development", "final", "ood"], default="final" ) evaluate.add_argument("--condition", default="primary") evaluate.add_argument("--policies", nargs="+") evaluate.add_argument("--limit", type=int) evaluate.add_argument("--output") app = commands.add_parser("app") app.add_argument("--port", type=int, default=7860) args = parser.parse_args() os.environ["SEALCELL_DEVICE"] = args.device if args.command == "download-model": import runpy runpy.run_path(str(ROOT / "scripts/download_model.py"), run_name="__main__") elif args.command == "probe": from .evaluation import fixture_benchmark from .model import get_model model = get_model() Path(args.output).parent.mkdir(parents=True, exist_ok=True) results = fixture_benchmark(model, args.fixtures, args.output, args.direct) Path(args.output + ".metadata.json").write_text(json.dumps(model.metadata, indent=2) + "\n") print( json.dumps( { "output": args.output, "cases": len(results), "errors": sum(bool(r["error"]) for r in results), }, indent=2, ) ) if any(r["error"] for r in results): raise SystemExit(1) elif args.command == "simulate": from .evaluation import episode_metrics, manifest from .policies import policy from .simulator import Simulator from .supervisor import Supervisor runner = policy(args.policy) supervisor = Supervisor(Simulator(args.scenario, args.seed, args.recipe), runner).run() metrics = episode_metrics(supervisor) result = { "manifest": manifest(vars(args), getattr(runner, "model", None)), "metrics": metrics, "trace": supervisor.trace, } Path(args.output).parent.mkdir(parents=True, exist_ok=True) Path(args.output).write_text(json.dumps(result, indent=2) + "\n") print(json.dumps(metrics, indent=2)) elif args.command == "evaluate": from .evaluation import evaluate output, _ = evaluate( args.config, args.split, args.condition, args.policies, args.limit, args.output ) print(f"Report: {output / 'report.md'}") else: from .app import build_app build_app(device=args.device).queue().launch( server_name="127.0.0.1", server_port=args.port, share=False, show_error=True ) if __name__ == "__main__": main()