"""Lightweight H* compatible panorama interface, no Habitat dependency.

The panorama is private environment state. Only rendered perspective images and
public instructions/pose should be sent to a policy. Angles are degrees.
"""
import math, re
import cv2
import numpy as np

HFOV=100.0
SYSTEM = ('You perform visual search from a fixed position. You see only the current '
          'perspective image. Rotate to gather evidence, then center the requested object '
          'or face the requested traversable path. Take one action per turn: '
          '<answer>rotate(delta_yaw,delta_pitch)</answer> or '
          '<answer>submit(current_yaw,current_pitch)</answer>. Use integer degrees. '
          'Positive yaw turns right; positive pitch looks up. Submit the CURRENT pose. '
          'Do not infer that an unseen target is already visible.')

def circular_delta(target, current):
    return (target-current+180)%360-180

def perspective(pano, yaw, pitch, height=360, fov=HFOV, crosshair=True):
    """Same spherical convention as Hstar: center longitude zero at W/2."""
    width=height//9*16
    focal=width/(2*math.tan(math.radians(fov)/2))
    v,u=np.indices((height,width),dtype=np.float32)
    rays=np.stack(((u-width/2)/focal,(v-height/2)/focal,np.ones_like(u)),axis=-1)
    rays/=np.linalg.norm(rays,axis=-1,keepdims=True)
    a,b=math.radians(yaw),math.radians(pitch)
    x,y,z=rays[...,0],rays[...,1],rays[...,2]
    yp=y*math.cos(b)-z*math.sin(b);zp=y*math.sin(b)+z*math.cos(b)
    xp=x*math.cos(a)+zp*math.sin(a);zp=-x*math.sin(a)+zp*math.cos(a)
    longitude=np.arctan2(xp,zp);latitude=np.arcsin(np.clip(yp,-1,1))
    ph,pw=pano.shape[:2]
    mx=((longitude/(2*np.pi)+.5)*pw).astype(np.float32)
    my=((latitude/np.pi+.5)*ph).astype(np.float32)
    # wrap only longitude; avoid pole-to-pole interpolation across vertical edge.
    result=cv2.remap(pano,mx%pw,np.clip(my,0,ph-1),cv2.INTER_LINEAR,borderMode=cv2.BORDER_WRAP)
    if crosshair:
        cx,cy=width//2,height//2
        cv2.line(result,(cx-8,cy),(cx+8,cy),(0,255,0),1)
        cv2.line(result,(cx,cy-8),(cx,cy+8),(0,255,0),1)
    return result

def bbox_angles(bbox,width,height):
    x1,y1,x2,y2=map(float,bbox)
    if not (0<=x1<x2<width and 0<=y1<y2<height):
        raise ValueError('invalid or seam-spanning bounding box')
    # Source bbox upper endpoints are inclusive pixel centers.
    yaw=[(x1/width*360-180)%360,(x2/width*360-180)%360]
    pitch=[90-y2/height*180,90-y1/height*180]
    center=[((x1+x2)/2/width*360-180)%360,sum(pitch)/2]
    return yaw,pitch,center

def bbox_visible_fraction(bbox,shape,yaw,pitch=0,fov=HFOV,aspect=16/9):
    """Spherical-area-weighted bbox proxy, NOT a true instance-mask visibility."""
    h,w=shape[:2];x1,y1,x2,y2=bbox
    xs=np.linspace(x1,x2,31);ys=np.linspace(y1,y2,21)
    lon,lat=np.meshgrid(xs/w*2*np.pi-np.pi,ys/h*np.pi-np.pi/2)
    dx=np.cos(lat)*np.sin(lon);dy=np.sin(lat);dz=np.cos(lat)*np.cos(lon)
    a,b=math.radians(yaw),math.radians(pitch)
    cx=dx*math.cos(a)-dz*math.sin(a);cz=dx*math.sin(a)+dz*math.cos(a)
    cy=dy*math.cos(b)+cz*math.sin(b);cz=-dy*math.sin(b)+cz*math.cos(b)
    tanx=math.tan(math.radians(fov)/2);tany=tanx/aspect
    inside=(cz>0)&(np.abs(cx)<=cz*tanx)&(np.abs(cy)<=cz*tany)
    weights=np.cos(lat)
    return float(weights[inside].sum()/weights.sum())

def in_interval(value,interval,circular=False,tolerance=0):
    lo,hi=interval
    if circular:
        span=(hi-lo)%360;center=(lo+span/2)%360
        return abs(circular_delta(value,center))<=max(span/2,tolerance)+1e-6
    return abs(value-(lo+hi)/2)<=max((hi-lo)/2,tolerance)+1e-6

class SearchEnv:
    def __init__(self,pano,annotation,start_yaw=0,height=360):
        self.pano=pano;self.annotation=annotation;self.yaw=float(start_yaw)%360
        self.pitch=0.;self.height=height;self.done=False;self.steps=0
    def observe(self):
        return perspective(self.pano,self.yaw,self.pitch,self.height)
    def public_text(self):
        return ('<image>\nTask: '+self.annotation['task']+
                f'\nCurrent direction (yaw,pitch): ({self.yaw:g},{self.pitch:g}). Decide your next action.')
    def step(self,text):
        if self.done:raise ValueError('episode already ended')
        match=re.fullmatch(r'\s*<answer>(rotate|submit)\((-?\d+),\s*(-?\d+)\)</answer>\s*',text)
        if not match:raise ValueError('one integer rotate or submit action is required')
        op,a,b=match[1],int(match[2]),int(match[3]);self.steps+=1
        if op=='rotate':
            self.yaw=(self.yaw+a)%360;self.pitch=max(-90,min(90,self.pitch+b))
            return {'done':False,'pose':[self.yaw,self.pitch]}
        if abs(circular_delta(a,self.yaw))>.51 or abs(b-self.pitch)>.51:
            raise ValueError('submit must name the actual current pose; teleport submission rejected')
        if 'yaw' not in self.annotation:
            self.done=True
            return {'done':True,'success':None,'evaluated':False,'pose':[self.yaw,self.pitch]}
        is_hos='pitch' in self.annotation
        success=in_interval(self.yaw,self.annotation['yaw'],True,30 if is_hos else 10)
        if is_hos:success=success and in_interval(self.pitch,self.annotation['pitch'],False,20)
        self.done=True
        return {'done':True,'success':bool(success),'pose':[self.yaw,self.pitch]}