tests/runtime/test_auto_regressive.py

"""The agent plays auto-regressively: each turn's observation carries the model's
OWN prior moves this run AND the handover memory every turn (not just turn 1,
not just the current grid) — so a stateless agent can maintain its trajectory."""
from __future__ import annotations

import proteus.game.scenarios  # noqa: F401
from proteus.game.agents.human import HumanAgent
from proteus.game.engine.difficulty import Difficulty
from proteus.game.runtime.session import SessionRunner


def _human(actions):
    feed = iter(actions)
    return HumanAgent(input_fn=lambda _p: next(feed), output_fn=lambda _t: None)


def _move_log(observation: str) -> str:
    """The 'your moves so far' segment of an observation (empty if absent)."""
    if "Your moves so far" not in observation:
        return ""
    return observation.split("Your moves so far", 1)[1].split("Now:", 1)[0]


def test_memory_shown_every_turn_and_prior_actions_are_auto_regressive():
    actions = ["up", "left", "down", "right"]
    runner = SessionRunner(
        "template", _human(actions), difficulty=Difficulty.EASY, seed=42,
        play_turns=len(actions), use_probe=False,
    )
    trace = runner.run()
    # template attaches a persona memory by default -> it is present on EVERY
    # turn now (previously only turn 1).
    assert all("MEMORY" in t.observation for t in trace.turns)
    # Turn 1: nothing chosen yet.
    assert _move_log(trace.turns[0].observation) == ""
    # Turn 2: the turn-1 move is fed back.
    assert "up" in _move_log(trace.turns[1].observation)
    # Turn 3: the running line of play accumulates (auto-regressive).
    log3 = _move_log(trace.turns[2].observation)
    assert "up" in log3 and "left" in log3
    # Turn 4: all three prior moves.
    log4 = _move_log(trace.turns[3].observation)
    assert log4.count(",") == 2  # three actions -> two separators