"""GameCrafter v5 capture -> training index (v4 format + layout remix summary).  usage: export_v4.py cache RAW OUT BUDGET | index RAW OUT"""
import sys, os, json, glob
import numpy as np
MODE, RAW, OUT = sys.argv[1], os.path.abspath(sys.argv[2]), os.path.abspath(sys.argv[3])
FPS, BLOCK, CLIP = 16, 4, 81
CLASSES = {0: 'sky/background', 1: 'terrain', 2: 'paving', 3: 'architecture', 4: 'rock', 5: 'vegetation', 6: 'prop', 10: 'player', 11: 'boss'}
os.makedirs(OUT, exist_ok=True)
R = os.path.join(RAW, 'raw')
N = len(glob.glob(os.path.join(R, '*.json')))
if MODE == 'cache':
    import time; t0 = time.time(); budget = float(sys.argv[4]) if len(sys.argv) > 4 else 150
    C = os.path.join(OUT, '_cache'); os.makedirs(C, exist_ok=True)
    for k in range((N + 999) // 1000):
        path = os.path.join(C, f'{k:04d}.json')
        if os.path.exists(path): continue
        if time.time() - t0 > budget: break
        chunk = [json.load(open(os.path.join(R, f'{f:06d}.json'))) for f in range(k * 1000, min(N, (k + 1) * 1000))]
        json.dump(chunk, open(path + '.tmp', 'w')); os.replace(path + '.tmp', path)
    left = [k for k in range((N + 999) // 1000) if not os.path.exists(os.path.join(C, f'{k:04d}.json'))]
    print('cache chunks remaining', len(left)); sys.exit(0)

acts = np.zeros((N, 5), np.uint8); eps = np.zeros(N, np.int32); cams = np.zeros(N, np.int32); invalid = 0
modes, states, rows, ev_frames = {}, {}, [], []
models = {'player': {}, 'boss': {}}; scales = {'player': [], 'boss': []}
C = os.path.join(OUT, '_cache'); cached = {}
def get(f):
    k = f // 1000
    if k not in cached:
        cp = os.path.join(C, f'{k:04d}.json'); cached.clear(); cached[k] = json.load(open(cp)) if os.path.exists(cp) else None
    return cached[k][f - k * 1000] if cached[k] is not None else json.load(open(os.path.join(R, f'{f:06d}.json')))
appearance = None
with open(os.path.join(OUT, 'frames.jsonl'), 'w') as fj:
    for f in range(N):
        j = get(f); a = j['action']; acts[f] = [a['w'], a['a'], a['s'], a['d'], a['attack']]
        eps[f] = j['episode']; cams[f] = j.get('camera_variant', 0); invalid += j.get('label_invalid_pixels', 0)
        appearance = j.get('appearance', appearance)
        modes[a['bot_mode']] = modes.get(a['bot_mode'], 0) + 1
        pl = next(x for x in j['actors'] if x['role'] == 'player'); bs = next(x for x in j['actors'] if x['role'] == 'boss')
        states[pl['action']] = states.get(pl['action'], 0) + 1
        ch = {c['role']: c for c in j.get('characters', [])}
        if f % 4 == 0:
            for role in ('player', 'boss'):
                if role in ch: models[role][ch[role]['model']] = models[role].get(ch[role]['model'], 0) + 1; scales[role].append(ch[role]['scale'])
        for e in j.get('events', []): ev_frames.append(dict(e, frame=f))
        rows.append(f"{f},{j['time_seconds']:.4f},{a['w']},{a['a']},{a['s']},{a['d']},{a['attack']},{j['episode']},{cams[f]},{a['bot_mode']},{a.get('fight_sub','')},{pl['action']},{pl['health']:.0f},{bs['action']},{bs['health']:.0f},"
                    f"{ch.get('player',{}).get('model','')},{ch.get('player',{}).get('scale',1):.3f},{ch.get('boss',{}).get('model','')},{ch.get('boss',{}).get('scale',1):.3f},{len(j.get('events',[]))}")
        keep = ['frame_id', 'time_seconds', 'episode', 'camera_variant', 'mirrored_world', 'personality', 'action', 'camera_world_cm', 'camera_pitch_yaw_roll', 'ortho_width_cm', 'world_to_clip', 'label_invalid_pixels', 'appearance', 'characters', 'events']
        fj.write(json.dumps({k: j[k] for k in keep if k in j} | {'actors': [{k: x[k] for k in ['role', 'action', 'health', 'position_cm', 'velocity_cm_s']} for x in j['actors']]}, separators=(',', ':')) + '\n')
np.save(os.path.join(OUT, 'actions.npy'), acts); np.save(os.path.join(OUT, 'episodes.npy'), eps)
nb = N // BLOCK; blocks = np.zeros((nb, 5), np.uint8)
for b in range(nb):
    seg = acts[b * BLOCK:(b + 1) * BLOCK]; blocks[b, :4] = seg[0, :4]; blocks[b, 4] = seg[:, 4].max()
np.save(os.path.join(OUT, 'actions_block_0.25s.npy'), blocks)
within_change = int(sum((acts[b * BLOCK:(b + 1) * BLOCK, :4] != acts[b * BLOCK, :4]).any() for b in range(nb)))
with open(os.path.join(OUT, 'actions.csv'), 'w') as fc:
    fc.write('frame,time_s,w,a,s,d,attack,episode,camera_variant,bot_mode,fight_sub,player_state,player_hp,boss_state,boss_hp,player_model,player_scale,boss_model,boss_scale,n_events\n' + '\n'.join(rows) + '\n')
# events: the json of frame f lists the events that first become visible at frame f
events = [e for e in ev_frames if e.get('type') in ('delete', 'replace', 'spawn')]
json.dump(ev_frames, open(os.path.join(OUT, 'events.json'), 'w'), indent=0)
clips = []
starts = [0] + [f for f in range(1, N) if eps[f] != eps[f - 1] or cams[f] != cams[f - 1]] + [N]
for s0, s1 in zip(starts[:-1], starts[1:]):
    s = ((s0 + BLOCK - 1) // BLOCK) * BLOCK
    while s + CLIP <= s1:
        inside = [{'offset': e['frame'] - s, 'type': e['type'], 'removed': e.get('removed', {}).get('class'), 'added': e.get('added', {}).get('class')} for e in events if s < e['frame'] < s + CLIP]
        clips.append({'start': s, 'end': s + CLIP, 'episode': int(eps[s]), 'camera_variant': int(cams[s]), 'events': inside}); s += CLIP - 1
json.dump(clips, open(os.path.join(OUT, 'clips_81.json'), 'w'))
res = json.load(open(os.path.join(RAW, 'capture_result.json'))) if os.path.exists(os.path.join(RAW, 'capture_result.json')) else {}
seg_len = np.diff([i for i in range(1, N) if (acts[i, :4] != acts[i - 1, :4]).any()] or [0])
ev_counts = {}
for e in events: ev_counts[e['type']] = ev_counts.get(e['type'], 0) + 1
ev_classes = {}
for e in events:
    for side in ('removed', 'added'):
        if side in e: k = f"{side}:{e[side]['class']}"; ev_classes[k] = ev_classes.get(k, 0) + 1
q = lambda v: [round(float(x), 3) for x in np.percentile(v, [0, 10, 50, 90, 100])] if len(v) else []
meta = {'run': os.path.basename(RAW), 'frames': N, 'fps': FPS, 'duration_min': round(N / FPS / 60, 2), 'resolution': [1280, 704], 'camera': 'isometric orthographic 2.5D',
        'appearance': appearance, 'appearance_detail': res.get('appearance'), 'layout_seed': res.get('layout_seed'), 'mirrored_world': res.get('mirrored_world'),
        'mirrored_y': res.get('mirrored_y'), 'layout_remix': {k: v for k, v in (res.get('layout_remix') or {}).items() if k != 'items'},
        'cross_scene': {k: v for k, v in (res.get('cross_scene') or {}).items() if k not in ('items', 'style_map')},
        'files_per_frame': {'rgb': '%06d.jpg', 'semantic_ids': '%06d.ids.png (uint8 class id, lossless)', 'whitebox': '%06d.whitebox.jpg', 'meta': '%06d.json'},
        'action_layout': ['w', 'a', 's', 'd', 'attack'],
        'semantic_classes': {str(k): v for k, v in CLASSES.items()}, 'episodes': int(eps.max() + 1), 'camera_variants': int(len(set(cams.tolist()))), 'clips_81': len(clips),
        'clips_with_events': sum(1 for c in clips if c['events']), 'events': ev_counts, 'event_classes': ev_classes,
        'label_invalid_pixels_total': int(invalid), 'blocks_with_key_change_inside': within_change,
        'key_segment_frames': {'median': float(np.median(seg_len)) if len(seg_len) else 0, 'p10': float(np.percentile(seg_len, 10)) if len(seg_len) else 0, 'min': int(seg_len.min()) if len(seg_len) else 0},
        'action_frequency': {k: round(float(acts[:, i].mean()), 4) for i, k in enumerate('wasd')} | {'attack_per_block': round(float(blocks[:, 4].mean()), 4), 'no_key': round(float((acts[:, :4].sum(1) == 0).mean()), 4)},
        'bot_modes': {k: round(v / N, 4) for k, v in sorted(modes.items(), key=lambda x: -x[1])}, 'player_states': {k: round(v / N, 4) for k, v in sorted(states.items(), key=lambda x: -x[1])},
        'character_models': {r: {k: round(v / max(1, sum(models[r].values())), 3) for k, v in models[r].items()} for r in models},
        'character_scale_percentiles': {r: q(scales[r]) for r in scales},
        'capture_result': res}
json.dump(meta, open(os.path.join(OUT, 'meta.json'), 'w'), indent=1, ensure_ascii=False)
print(json.dumps({k: meta[k] for k in ['run', 'frames', 'appearance', 'episodes', 'clips_81', 'clips_with_events', 'events', 'label_invalid_pixels_total', 'blocks_with_key_change_inside', 'action_frequency', 'bot_modes', 'character_models', 'character_scale_percentiles']}, ensure_ascii=False))