import sys, os, time
sys.path.insert(0, '/home/paulh/projects/donkeycar-rl-autoresearch/agent')
from multitrack_runner import wrap_env, log, _send_exit_scene
from donkeycar_sb3_runner import ThrottleClampWrapper
from reward_wrapper import SpeedRewardWrapper
from multitrack_runner import StuckTerminationWrapper
from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import DummyVecEnv, VecTransposeImage
from stable_baselines3.common.callbacks import BaseCallback
import gymnasium as gym, numpy as np

LR           = 0.000725
TOTAL_STEPS  = 90000
THROTTLE_MIN = 0.5           # raised from 0.2
SAVE_PATH    = '/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/exp3-mountain-throttle05/model'
os.makedirs(os.path.dirname(SAVE_PATH), exist_ok=True)

def make_env_high_throttle(env_id):
    raw = gym.make(env_id)
    env = ThrottleClampWrapper(raw, throttle_min=THROTTLE_MIN)
    env = StuckTerminationWrapper(env, stuck_steps=80, min_displacement=0.5)
    env = SpeedRewardWrapper(env, speed_scale=0.1)
    return env

def switch_to(current_id, next_id, name, throttle_min=THROTTLE_MIN):
    log(f'  → Switching to {name} (throttle_min={throttle_min})...')
    tmp = gym.make(current_id); time.sleep(2)
    _send_exit_scene(tmp, verbose=False); tmp.close(); time.sleep(5)
    raw = gym.make(next_id)
    env = VecTransposeImage(DummyVecEnv(
        [lambda e=raw, t=throttle_min: (
            lambda ee: (ThrottleClampWrapper(ee, throttle_min=t),
                        StuckTerminationWrapper(ThrottleClampWrapper(ee, throttle_min=t), 80, 0.5),
                        SpeedRewardWrapper(StuckTerminationWrapper(ThrottleClampWrapper(ee, throttle_min=t), 80, 0.5), 0.1)
                       )[-1]
        )(e)]
    ))
    log(f'  Connected to {name}'); return env

def simple_switch(current_id, next_id, name):
    log(f'  → Switching to {name}...')
    tmp = gym.make(current_id); time.sleep(2)
    _send_exit_scene(tmp, verbose=False); tmp.close(); time.sleep(5)
    raw = gym.make(next_id)
    env = VecTransposeImage(DummyVecEnv([lambda e=raw: make_env_high_throttle(next_id)]))
    log(f'  Connected to {name}'); return env

class ProgressCB(BaseCallback):
    def __init__(self, total): super().__init__(verbose=0); self._last=0; self._total=total
    def _on_step(self):
        if self.num_timesteps - self._last >= 10000:
            log(f'    step {self.num_timesteps:,}/{self._total:,}')
            self._last = self.num_timesteps
        return True

log('='*60)
log(f'Exp 3: mountain_track ONLY, throttle_min={THROTTLE_MIN}, lr={LR}')
log('Hypothesis: higher min throttle gets car over the hill,')
log('            allowing it to learn the full track.')
log('='*60)

# Connect - sim is on mountain_track from last test
log('Connecting to mountain_track...')
tmp = gym.make('donkey-mountain-track-v0'); time.sleep(2)
_send_exit_scene(tmp, verbose=False); tmp.close(); time.sleep(5)

raw = gym.make('donkey-mountain-track-v0')
train_env = VecTransposeImage(DummyVecEnv([lambda: make_env_high_throttle('donkey-mountain-track-v0')]))
log('Connected. Training...')

model = PPO('CnnPolicy', train_env, learning_rate=LR, verbose=1, device='cpu')
model.learn(total_timesteps=TOTAL_STEPS, callback=ProgressCB(TOTAL_STEPS),
            reset_num_timesteps=True)
model.save(SAVE_PATH)
log(f'Saved: {SAVE_PATH}.zip')
train_env.close(); time.sleep(3)

# Eval on all 4 tracks
def eval_track(current_id, track_id, track_name, model_path, n=3, max_steps=2000):
    log(f'\n--- EVAL: {track_name} ({n} episodes) ---')
    tmp = gym.make(current_id); time.sleep(2)
    _send_exit_scene(tmp, verbose=False); tmp.close(); time.sleep(5)
    raw = gym.make(track_id)
    ev = VecTransposeImage(DummyVecEnv([lambda e=raw: make_env_high_throttle(track_id)]))
    m = PPO.load(model_path, env=ev, device='cpu')
    for ep in range(1, n+1):
        obs = ev.reset(); total, steps, done = 0.0, 0, False
        while not done and steps < max_steps:
            action, _ = m.predict(obs, deterministic=True)
            result = ev.step(action)
            if len(result)==5: obs,r,t,tr,info=result; done=bool(t[0] or tr[0])
            else: obs,r,d,info=result; done=bool(d[0])
            total+=float(r[0]); steps+=1
        status='✅ FULL' if steps>=max_steps else f'❌ crash@{steps}'
        log(f'  ep{ep}: {total:.0f} reward / {steps} steps — {status}')
        time.sleep(1)
    ev.close(); time.sleep(3)
    return track_id

current = 'donkey-mountain-track-v0'
current = eval_track(current, 'donkey-mountain-track-v0', 'mountain_track (training track)', SAVE_PATH)
current = eval_track(current, 'donkey-generated-track-v0', 'generated_track (zero-shot)', SAVE_PATH)
current = eval_track(current, 'donkey-minimonaco-track-v0', 'mini_monaco (zero-shot)', SAVE_PATH)
current = eval_track(current, 'donkey-generated-roads-v0', 'generated_road (zero-shot)', SAVE_PATH)

log('\n=== Exp 3 COMPLETE ===')