cardbuddy 0.13.3 → 0.14.1

package/dist/vendored/skills/card-skill-trainer/scripts/build_dataset.py

@@ -57,33 +57,103 @@ def split_for(case_id: str) -> str:
 
 
 def discover_sources(root: Path):
+    # --src 直接指向一份 MeaXure 导出（自带 index.html）时，它本身就是唯一 source。
+    # 此前必须指到父目录才能被发现，会把同级所有设计稿一起卷进来（训练常只跑一两块画板）。
+    if (root / "index.html").exists():
+        yield root
+        return
     for path in sorted(root.iterdir(), key=lambda p: p.name):
         if path.is_dir() and (path / "index.html").exists():
             yield path
 
 
-def parse_source(source: Path, raw_root: Path, force: bool, sample_size: int = 0) -> bool:
+def parse_artboard_selector(value: Optional[str]):
+    """--artboards 拆成「1 起序号」与「画板名关键词」两类，与门禁清单同口径。"""
+    indices: list = []
+    keywords: list = []
+    for raw in (value or "").split(","):
+        token = raw.strip()
+        if not token:
+            continue
+        if token.isdigit():
+            index = int(token)
+            if index >= 1 and index not in indices:
+                indices.append(index)
+        elif token not in keywords:
+            keywords.append(token)
+    return indices, keywords
+
+
+def matches_artboard_selector(canvas: dict, indices: list, keywords: list) -> bool:
+    if not indices and not keywords:
+        return True
+    order = canvas.get("order")
+    if indices and isinstance(order, int) and order in indices:
+        return True
+    name = str(canvas.get("name") or "")
+    return bool(keywords) and any(keyword in name for keyword in keywords)
+
+
+def scoped_canvases_ready(target: Path, indices: list, keywords: list) -> bool:
+    """增量场景：所选画板已全部 generated 且 spec.json 落盘，才允许跳过重解析。"""
+    if not indices and not keywords:
+        return (target / "manifest.json").exists()
+    manifest_path = target / "manifest.json"
+    if not manifest_path.exists():
+        return False
+    try:
+        manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return False
+    picked = [
+        canvas for canvas in manifest.get("canvases", [])
+        if isinstance(canvas, dict) and matches_artboard_selector(canvas, indices, keywords)
+    ]
+    if not picked:
+        return False
+    return all(
+        canvas.get("status") == "generated"
+        and (target / str(canvas.get("name") or "") / "spec.json").exists()
+        for canvas in picked
+    )
+
+
+def parse_source(source: Path, raw_root: Path, force: bool, sample_size: int = 0,
+                 indices: Optional[list] = None, keywords: Optional[list] = None) -> bool:
     target = raw_root / safe_name(source.name)
-    if (target / "manifest.json").exists() and not force:
+    indices = indices or []
+    keywords = keywords or []
+    if not force and scoped_canvases_ready(target, indices, keywords):
         print(f"[skip] {source.name}")
         return True
     target.mkdir(parents=True, exist_ok=True)
-    command = [
-        sys.executable,
-        str(PARSER),
-        "--src",
-        str(source),
-        "--out",
-        str(target),
-        "--build",
-        # 无人值守训练 = 显式批量授权（rules.md §13.3 例外）：直接过 parse_meaxure 的多画板选择门禁
-        "--scope-all",
-    ]
-    if sample_size:
-        command.extend(["--sample", str(sample_size)])
-    print(f"[parse] {source.name}")
-    result = subprocess.run(command)
-    return result.returncode == 0 and (target / "manifest.json").exists()
+    # parse_meaxure 的 --only 优先于 --indices，混用时拆成两次局部解析：
+    # 局部运行只重解析选中画板，未选中的状态从旧 manifest 继承，两次结果自然合并。
+    scopes: list = []
+    if indices:
+        # 门禁清单是 1 起，--indices 是 0 起
+        scopes.append(["--indices", ",".join(str(index - 1) for index in sorted(indices))])
+    if keywords:
+        scopes.append(["--only", ",".join(keywords)])
+    if not scopes:
+        scopes.append(["--sample", str(sample_size)] if sample_size else [])
+    for scope in scopes:
+        command = [
+            sys.executable,
+            str(PARSER),
+            "--src",
+            str(source),
+            "--out",
+            str(target),
+            "--build",
+            # 无人值守训练 = 显式批量授权（rules.md §13.3 例外）：直接过 parse_meaxure 的多画板选择门禁
+            "--scope-all",
+        ] + scope
+        print(f"[parse] {source.name}{' ' + ' '.join(scope) if scope else ''}")
+        result = subprocess.run(command)
+        if result.returncode != 0:
+            return False
+    return (target / "manifest.json").exists()
 
 
 def first_image(spec_dir: Path, card: dict) -> Optional[Path]:
@@ -287,6 +357,11 @@ def main():
     parser.add_argument("--max-cases", type=int, default=160)
     parser.add_argument("--max-sources", type=int, default=0)
     parser.add_argument("--artboards-per-source", type=int, default=0)
+    parser.add_argument(
+        "--artboards", default=None,
+        help="只构建指定画板：1 起序号或画板名关键词，逗号分隔（如 4,14 / 金额极限,保险保障）。"
+             "序号转 parse_meaxure 的 0 起 --indices，关键词走 --only；未选画板不进数据集。",
+    )
     parser.add_argument("--force", action="store_true")
     parser.add_argument("--duplicate-threshold", type=int, default=4)
     parser.add_argument("--no-artboard-dedup", action="store_true")
@@ -297,6 +372,10 @@ def main():
     raw_root = out_root / "raw"
     raw_root.mkdir(parents=True, exist_ok=True)
 
+    sel_indices, sel_keywords = parse_artboard_selector(args.artboards)
+    if sel_indices or sel_keywords:
+        print(f"[scope] 只构建画板：indices={sel_indices or '[]'} keywords={sel_keywords or '[]'}")
+
     sources = list(discover_sources(source_root))
     if args.max_sources:
         sources = sources[: args.max_sources]
@@ -311,7 +390,8 @@ def main():
     eligibleRecords = []
     generatedArtboards = 0
     for source in sources:
-        ok = parse_source(source, raw_root, args.force, args.artboards_per_source)
+        ok = parse_source(source, raw_root, args.force, args.artboards_per_source,
+                          sel_indices, sel_keywords)
         if not ok:
             failures.append({"source": source.name, "error": "parse-failed"})
             continue
@@ -324,6 +404,9 @@ def main():
             spec_path = spec_dir / "spec.json"
             if status == "skipped" or not spec_path.exists():
                 continue
+            # 旧 raw 目录可能来自一次全量解析：按本次 --artboards 收口，避免把未选画板混进数据集
+            if not matches_artboard_selector(canvas, sel_indices, sel_keywords):
+                continue
             generatedArtboards += 1
             if canvas.get("trainingEligible") is False:
                 skippedPopups.append({"source": source.name, "canvas": name, "reason": "manifest-training-ineligible"})
@@ -442,6 +525,7 @@ def main():
             "parser": str(PARSER.relative_to(REPO_ROOT)),
             "duplicateThreshold": args.duplicate_threshold,
             "artboardsPerSource": args.artboards_per_source,
+            "artboards": args.artboards or "",
             "maxCases": args.max_cases,
             "noArtboardDedup": bool(args.no_artboard_dedup),
         },