tests/runtime/test_integration_golden.py

import json

from proteus.game.runtime import SessionRunner, SessionTrace
from proteus.providers import FakeProvider
from proteus.game.agents import VanillaAgent

# Self-captured deterministic snapshot for template (seed=42, EASY, 8 turns,
# the agent always answers "up"). template has no ToM divergence, so the model's
# "up" matches the optimal escape and motive-reading accuracy is perfect.
EXPECTED_MRA = 100.0


def test_full_session_serializes_to_jsonl_and_reloads():
    agent = VanillaAgent(FakeProvider(responses=["ACTION: up"], model_name="fake-1"))
    trace = SessionRunner(
        "template", agent, seed=42, play_turns=8, use_probe=True,
    ).run()

    # Serialize the whole session as one JSON line, reload, and verify.
    line = trace.model_dump_json()
    reloaded = SessionTrace.model_validate_json(line)
    # Full round-trip fidelity: every field survives serialization unchanged.
    assert reloaded.model_dump() == trace.model_dump()
    assert reloaded.model == "fake-1"
    # Deterministic self-captured snapshot (regression guard): the agent always
    # answers "up", so the first committed action is "up".
    assert reloaded.turns[0].motive_action == "up"
    # Concrete metric anchor (regression guard, not a vacuous >= 0 check).
    assert reloaded.metrics["motive_reading_accuracy"] == EXPECTED_MRA

    # Per-turn JSONL is also valid line-by-line.
    for t in trace.turns:
        assert json.loads(t.model_dump_json())["turn_idx"] == t.turn_idx