"""The turn loop (SPEC.md §1, §5).

This is the game — not the model. A turn is:

    build_prompt -> converse (model, with §4 retry/fallback) -> apply_action
    (deterministic) -> check_win (mechanical) -> maybe-offer a concept to learn.

Kept free of Gradio so the whole loop is testable against StubBrain at zero GPU
(tests/test_loop_stub.py). app.py is thin wiring over these functions.

Ledger additions are GATED on player confirmation (SPEC §5): a candidate_concept
becomes ``session.pending_candidate`` and only enters the ledger when the player
confirms. That keeps the "it learned!" beat deliberate and screenshot-worthy.
"""

from __future__ import annotations

from dataclasses import dataclass
from typing import Optional

from . import ledger as ledger_mod
from .brain import Brain
from .challenges import CHALLENGES
from .models import GameSession
from .parsing import ParsedResponse, converse
from .prompt import build_prompt
from .world import apply_action, check_win, initial_world


@dataclass
class TurnResult:
    response: ParsedResponse
    log_line: str
    won: bool                       # the challenge win-predicate is now satisfied
    newly_won: bool                 # ...and it wasn't before this turn
    learn_offer: Optional[dict]     # a candidate concept awaiting confirmation
    reapplied: tuple[str, ...] = () # concept ids whose times_applied just bumped


def new_session() -> GameSession:
    """Fresh per-visitor state. Lives in one ``gr.State`` (SPEC §0)."""
    return GameSession(
        ledger=ledger_mod.seed_ledger(),
        world=initial_world(),
        challenge_index=0,
        turn=0,
        history=[],
        pending_candidate=None,
        won_current=False,
    )


def current_challenge(session: GameSession):
    return CHALLENGES[session.challenge_index]


def run_turn(session: GameSession, utterance: str, brain: Brain) -> TurnResult:
    """One full turn. Mutates ``session`` in place and returns a TurnResult.

    The GPU (when LocalBrain is used) is held ONLY inside ``converse`` -> the
    decorated ``respond``. Everything here — state mutation, win check, learning
    — runs on CPU, outside the GPU function (SPEC §6)."""
    session.turn += 1
    challenge = current_challenge(session)

    session.history.append({"who": "player", "text": utterance, "gap": None, "kind": None})

    prompt = build_prompt(session.ledger, session.world, challenge, utterance)
    response = converse(brain, prompt, session.world)

    # Deterministic application — the model only ever *proposed* the action.
    log_line = apply_action(session.world, response.action)

    session.history.append(
        {"who": "alien", "text": response.utterance, "gap": response.gap,
         "action": log_line, "kind": None}
    )

    was_won = session.won_current
    won = check_win(session.world, challenge)
    session.won_current = won
    newly_won = won and not was_won

    # --- times_applied: a learned concept transferring to a new situation (§5) --
    reapplied: list[str] = []
    if newly_won and challenge.teaches is None and challenge.relies_on:
        # Light up the concepts the player ACTUALLY taught — "it drew on what you
        # gave it". We can't match challenge.relies_on by id: the real model names
        # concepts freely (the stub used the canonical ids, which masked this).
        for c in session.ledger:
            if c.taught_on_turn != 0 and not c.via_generalization:
                c.times_applied += 1
                reapplied.append(c.id)

    # --- learning gate: offer the candidate, don't auto-add (§5) ---------------
    learn_offer: Optional[dict] = None
    candidate = response.candidate_concept
    if candidate and candidate.get("id"):
        if not ledger_mod.has(session.ledger, candidate["id"]):
            session.pending_candidate = {**candidate, "_player_phrase": utterance}
            learn_offer = candidate

    if newly_won:
        session.history.append(
            {"who": "system", "text": f"✦ {challenge.title} — understood.",
             "gap": None, "kind": "win"}
        )

    return TurnResult(
        response=response,
        log_line=log_line,
        won=won,
        newly_won=newly_won,
        learn_offer=learn_offer,
        reapplied=tuple(reapplied),
    )


def confirm_candidate(session: GameSession) -> Optional[str]:
    """Player pressed "Yes, it learned that". Move the pending candidate into the
    ledger. Returns the new concept's label, or None if nothing pending."""
    cand = session.pending_candidate
    if not cand:
        return None
    phrase = cand.get("_player_phrase", "")
    payload = {k: v for k, v in cand.items() if not k.startswith("_")}
    concept = ledger_mod.add_concept(session.ledger, payload, session.turn, phrase)
    session.pending_candidate = None
    if concept is None:
        return None
    ch = current_challenge(session)
    concept.via_generalization = ch.teaches is None and bool(ch.relies_on)
    if concept.via_generalization and not concept.built_from:
        # Lineage = the concepts the player ACTUALLY taught before this one (their
        # real ids). The model names concepts freely, so the challenge's canonical
        # relies_on ids won't be in the ledger — using them produced phantom
        # "from hidden_info + gift" lineage pointing at concepts that don't exist.
        concept.built_from = tuple(
            c.id for c in session.ledger
            if c.taught_on_turn != 0 and not c.via_generalization and c.id != concept.id
        )
    session.history.append(
        {"who": "system", "text": f"✦ learned: {concept.label}", "gap": None, "kind": "learn"}
    )
    return concept.label


def reject_candidate(session: GameSession) -> None:
    """Player declined the candidate. Drop it; nothing enters the ledger."""
    session.pending_candidate = None


def advance_challenge(session: GameSession) -> bool:
    """Move to the next challenge, resetting the world so its win-predicate is
    well-posed. Returns False if the arc is complete."""
    if session.challenge_index >= len(CHALLENGES) - 1:
        return False
    session.challenge_index += 1
    session.world = initial_world()
    session.won_current = False
    session.pending_candidate = None
    nxt = current_challenge(session)
    session.history.append(
        {"who": "system", "text": f"— {nxt.title} —", "gap": None, "kind": "challenge"}
    )
    return True


def is_arc_complete(session: GameSession) -> bool:
    return session.challenge_index >= len(CHALLENGES) - 1 and session.won_current