#!/usr/bin/env python3
"""Laya decision model — simple runner (torch-free).

  pip install coreai-core transformers numpy

  python3 run.py --task guardrail --state "rm -rf /home/user/projects" \
      --question '{"disposition":{"type":"choice","instructions":"...",
                    "criteria":{"allow":"...","block":"..."}}}'
  python3 run.py --task lang_route --state "wake me at nine" \
      --question '{"lang":{"type":"choice","instructions":"...",
                  "criteria":{"en":"...","de":"..."}}}'

Prints one JSON verdict: chain, choice, confidence, probs, acted.
Questions with no "criteria" render as binary noul (yes/no). --unit
gpu|cpu|ne (gpu recommended; unpinned ANE loads can SIGABRT). One process,
one agent — reuse it for many decide() calls in your own code via
CombinedAgent directly (see src/laya_port/combined_agent.py).
"""
import argparse, json, os, sys

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.join(HERE, "src"))


def main():
    ap = argparse.ArgumentParser(description="Laya combined decision model")
    ap.add_argument("--task", required=True,
                    help="triage|lang_route|guardrail|act_escalate|tool_route|"
                         "skill_route|mail_sort|supervise|choose|compact|rerank")
    ap.add_argument("--state", required=True, help="agent state text")
    ap.add_argument("--question", required=True,
                    help='typed question JSON: {"<name>": {type, instructions, criteria}}')
    ap.add_argument("--unit", default="gpu", choices=["gpu", "cpu", "ne"])
    ap.add_argument("--pad", action="store_true",
                    help="pad every call to asset L_max (pays re-specialization "
                         "once at startup; ~16 ms/call warm vs seconds per shape)")
    a = ap.parse_args()

    from laya_port.combined_agent import CombinedAgent
    ag = CombinedAgent(os.path.join(HERE, "laya-combined-f16.aimodel"),
                       os.path.join(HERE, "configs"), unit=a.unit)
    q = json.loads(a.question)
    shape = ag.prov["shape"]["L_max"] if a.pad else None
    print(json.dumps(ag.decide(a.task, a.state, q, pad_to=shape), indent=1))


if __name__ == "__main__":
    main()