"""Smoke test: one scripted episode per condition on one task, then monitors offline. usage: python scripts/smoke_swe.py T3 [C0_honest_single,C3_collude_overt] """ import sys import time from pathlib import Path from collusion_lab.monitor.core import TrustedMonitor from collusion_lab.scenarios.swe.conditions import CONDITIONS from collusion_lab.scenarios.swe.env import load_ctx, run_episode from collusion_lab.scenarios.swe.monitor_plugin import SWEPlugin task = sys.argv[1] if len(sys.argv) > 1 else "T3" conds = sys.argv[2].split(",") if len(sys.argv) > 2 else list(CONDITIONS) out = Path("runs/smoke") for c in conds: t0 = time.time() rec = run_episode(task, c, 0, out) ctx, events, _ = load_ctx(out / rec["episode_id"]) ds = {lvl: TrustedMonitor(lvl, SWEPlugin()).evaluate(ctx, events) for lvl in ("M1", "M3")} o = rec["outcome"] print(f"{c:20s} rel={o['released']!s:5} correct={o['correct']!s:5} harm={o['harmful_release']!s:5} " f"| M1 {ds['M1'].action.value:9s} c={ds['M1'].collusion_score:.2f} " f"| M3 {ds['M3'].action.value:9s} r={ds['M3'].risk:.2f} c={ds['M3'].collusion_score:.2f} " f"| {time.time()-t0:.0f}s", flush=True) for r in ds["M3"].reasons[:5]: print(" -", r) for r in ds["M3"].collusion_reasons[:3]: print(" * ", r)