"""实测 deepseek-flash 的 grounding：能否给出可用的「放大区域」归一化坐标。

流程：截一张真实图框整体图 → 问 VLM「哪些区域需要放大」→ 解析它给的归一化坐标
→ 映射成 CAD 世界坐标（WCS）→ 按坐标放大截图，供目视验证 grounding 是否可用。

用法：
    python verify_grounding.py [--frames 1] [--out reports/_probe]

前置：CAD 在线（127.0.0.1:8080 监听）、DEEPSEEK_API_KEY 可用。
"""

import argparse
import base64
import json
import time
from pathlib import Path

from dwg_review import config
from dwg_review.findings import extract_json_block
from dwg_review.llm_client import LlmClient, downsample_image_b64
from dwg_review.mcp_client import result_images
from dwg_review.phases import planning
from dwg_review.context import ReviewContext

_PROMPT = (
    "这是某工程图纸中一个图框的截图。请判断：图中哪些区域**细节不清、需要放大**才能看清？\n"
    "用**归一化坐标**给出想放大的区域：x0,y0,x1,y1 取值 0~1，"
    "图片左上角为 (0,0)、右下角为 (1,1)。\n"
    "最多给 3 个区域，按重要性排序。\n"
    '只输出 JSON，不要解释：{"regions":[{"bbox":[x0,y0,x1,y1],"reason":"..."}]}'
)


def _save(b64, path):
    path.write_bytes(base64.b64decode(b64))
    return path


def _norm_to_wcs(win, x0, y0, x1, y1):
    """归一化坐标（左上原点）→ WCS window。"""
    dx = win["max_x"] - win["min_x"]
    dy = win["max_y"] - win["min_y"]
    return {
        "min_x": win["min_x"] + x0 * dx,
        "max_x": win["min_x"] + x1 * dx,
        "min_y": win["max_y"] - y1 * dy,  # 图片 y 向下，WCS y 向上
        "max_y": win["max_y"] - y0 * dy,
    }


def main(argv=None):
    ap = argparse.ArgumentParser()
    ap.add_argument("--frames", type=int, default=1)
    ap.add_argument("--out", default="reports/_probe")
    args = ap.parse_args(argv)

    ctx = ReviewContext.create()
    plan = planning.plan(ctx)
    frames = plan.get("frames", [])[: args.frames]
    if not frames:
        print("未识别到图框")
        return 1

    out_dir = Path(config.PROJECT_ROOT) / args.out
    out_dir.mkdir(parents=True, exist_ok=True)

    frame = frames[0]
    win = {
        k: float(v)
        for k, v in zip(("min_x", "min_y", "max_x", "max_y"), frame["bbox"])
    }
    print("图框 window:", win)

    ctx.tools.call("zoom_to_window", dict(win, margin=0.02))
    time.sleep(config.RENDER_WAIT)
    shot = ctx.tools.call("get_viewport_screenshot", {})
    imgs = result_images(shot)
    if not imgs:
        print("无截图")
        return 1
    b64, mime = imgs[0]
    nb, nm = downsample_image_b64(b64, mime)
    print("整体图:", _save(b64, out_dir / "overall.png"))

    llm = LlmClient()
    print("模型:", llm.model)
    messages = [
        {"role": "system", "content": "你是资深工程图纸审图助手。"},
        {
            "role": "user",
            "content": [
                {"type": "text", "text": _PROMPT},
                LlmClient.image_content(nb, nm),
            ],
        },
    ]
    resp = llm.chat(messages)
    text = resp.choices[0].message.content or ""
    print("=== VLM 原始响应 ===")
    print(text)

    block = extract_json_block(text)
    if not block:
        print("[X] 未解析出 JSON")
        return 1
    data = json.loads(block)
    regions = data.get("regions") or []
    print(f"=== 解析到 {len(regions)} 个区域 ===")

    for i, r in enumerate(regions, 1):
        bbox = r.get("bbox")
        if not (isinstance(bbox, list) and len(bbox) == 4):
            print(f"  区域 {i}: bbox 非法 {bbox}")
            continue
        x0, y0, x1, y1 = (float(v) for v in bbox)
        sub = _norm_to_wcs(win, x0, y0, x1, y1)
        print(f"  区域 {i}: bbox={bbox} reason={r.get('reason')!r}")
        print(f"    -> WCS {sub}")
        ctx.tools.call("zoom_to_window", dict(sub, margin=0.05))
        time.sleep(config.RENDER_WAIT)
        s2 = ctx.tools.call("get_viewport_screenshot", {})
        im2 = result_images(s2)
        if im2:
            print("    -> 已保存", _save(im2[0][0], out_dir / f"region_{i}.png"))

    print("\n截图目录:", out_dir)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())