Spaces:
Running
Running
| """Day-1 smoke tests: run 100 episodes end-to-end with scripted agents. | |
| If this passes, the environment is ready for LLM integration on Day 2. | |
| """ | |
| from __future__ import annotations | |
| import pytest | |
| from chakravyuh_env import ChakravyuhEnv, EpisodeOutcome, VictimProfile # noqa: F401 | |
| def test_env_reset_returns_observation(): | |
| env = ChakravyuhEnv() | |
| obs = env.reset(seed=42) | |
| assert obs.agent_role == "scammer" | |
| assert obs.turn == 0 | |
| def test_deterministic_with_seed(): | |
| """Same seed → same outcome. Required for replay-first demo.""" | |
| env1 = ChakravyuhEnv() | |
| env2 = ChakravyuhEnv() | |
| env1.reset(seed=123) | |
| env2.reset(seed=123) | |
| done1, done2 = False, False | |
| outcomes = [] | |
| while not done1: | |
| _, _, done1, info = env1.step() | |
| if done1: | |
| outcomes.append(info["outcome"]) | |
| while not done2: | |
| _, _, done2, info = env2.step() | |
| if done2: | |
| outcomes.append(info["outcome"]) | |
| # Deterministic: same category, same analyzer flag, same detection turn | |
| assert outcomes[0].scam_category == outcomes[1].scam_category | |
| assert outcomes[0].analyzer_flagged == outcomes[1].analyzer_flagged | |
| assert outcomes[0].turns_used == outcomes[1].turns_used | |
| def test_episode_terminates_within_max_turns(): | |
| env = ChakravyuhEnv() | |
| env.reset(seed=7) | |
| done = False | |
| for _ in range(15): # safety cap | |
| if done: | |
| break | |
| _, _, done, _ = env.step() | |
| assert done is True | |
| def test_100_episodes_run_without_crash(): | |
| """Main smoke test — the one that gates Day 2. | |
| Iterates over all three victim profiles to guarantee the full decision | |
| surface (comply / refuse / verify) is covered at least once. | |
| """ | |
| profiles = [VictimProfile.SENIOR, VictimProfile.SEMI_URBAN, VictimProfile.YOUNG_URBAN] | |
| outcomes: list[EpisodeOutcome] = [] | |
| for i in range(100): | |
| profile = profiles[i % 3] | |
| gullibility = {"senior": 1.5, "semi_urban": 1.0, "young_urban": 0.7}[profile.value] | |
| env = ChakravyuhEnv(victim_profile=profile, gullibility=gullibility) | |
| env.reset(seed=1000 + i) | |
| done = False | |
| while not done: | |
| _, _, done, info = env.step() | |
| if done: | |
| outcomes.append(info["outcome"]) | |
| assert len(outcomes) == 100 | |
| flagged = sum(1 for o in outcomes if o.analyzer_flagged) | |
| assert flagged > 0, "Analyzer never flagged in 100 episodes — rules broken" | |
| extracted = sum(1 for o in outcomes if o.money_extracted) | |
| refused = sum(1 for o in outcomes if o.victim_refused) | |
| assert extracted > 0, "No extraction across mixed profiles — victim logic broken" | |
| assert refused > 0, "No refusal across mixed profiles — victim logic broken" | |
| def test_reward_signs_are_sane(): | |
| """Scammer reward > 0 when money extracted; < 0 when refused early.""" | |
| env = ChakravyuhEnv(victim_profile=VictimProfile.SENIOR, gullibility=1.5) | |
| env.reset(seed=42) | |
| done = False | |
| reward = None | |
| while not done: | |
| _, reward, done, _ = env.step() | |
| assert reward is not None | |
| # Reward breakdown must sum correctly | |
| assert isinstance(reward.scammer, float) | |
| assert isinstance(reward.victim, float) | |
| assert isinstance(reward.analyzer, float) | |
| def test_all_scam_categories_reachable(): | |
| """Seed diversity must cover all 5 scam categories.""" | |
| env = ChakravyuhEnv() | |
| seen = set() | |
| for i in range(200): | |
| env.reset(seed=i) | |
| done = False | |
| while not done: | |
| _, _, done, info = env.step() | |
| if done: | |
| seen.add(info["outcome"].scam_category.value) | |
| assert len(seen) >= 3, f"Only {len(seen)} categories seen: {seen}" | |