minimal-agent 0.8.2 → 0.10.0

package/skills/ui-design/scripts/extract_source_assets_test.py

@@ -0,0 +1,256 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+extract_source_assets_test.py —— extract_source_assets.py 的单测（内造 pptx/docx fixture）。
+
+锁定：
+  - pptx 抽出 slide 文字（含标题作 section）+ picture，manifest 记对 aspect_ratio/orient/section。
+  - docx 抽出 heading 分节 + 内嵌图，图关联到最近 heading。
+  - 横图判 landscape、竖图判 portrait。
+  - 无源文档 → has_source:false，不报错。
+依赖：python-pptx / python-docx / Pillow（已就绪）。
+"""
+
+import json
+import os
+import sys
+import tempfile
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import extract_source_assets as ex  # noqa: E402
+
+
+def _make_png(path, w, h, color=(80, 120, 200)):
+    from PIL import Image
+    Image.new("RGB", (w, h), color).save(path)
+
+
+def _make_pptx(path, png_landscape):
+    from pptx import Presentation
+    from pptx.util import Inches
+    prs = Presentation()
+    slide = prs.slides.add_slide(prs.slide_layouts[5])  # Title Only（含标题占位符）
+    slide.shapes.title.text = "产品概览"
+    tb = slide.shapes.add_textbox(Inches(1), Inches(5), Inches(8), Inches(1))
+    tb.text_frame.text = "这是产品概览页的说明文字"
+    slide.shapes.add_picture(png_landscape, Inches(1), Inches(1), Inches(4))
+    prs.save(path)
+
+
+def _make_docx(path, png_portrait):
+    from docx import Document
+    from docx.shared import Inches
+    doc = Document()
+    doc.add_heading("第一章 背景", level=1)
+    doc.add_paragraph("这是第一章的正文段落，描述背景。")
+    doc.add_picture(png_portrait, width=Inches(2))
+    doc.save(path)
+
+
+class TestExtractPptx(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.mkdtemp(prefix="extract-pptx-")
+        self.png = os.path.join(self.tmp, "land.png")
+        _make_png(self.png, 1600, 900)               # 横图
+        self.doc = os.path.join(self.tmp, "brief.pptx")
+        _make_pptx(self.doc, self.png)
+        self.out = os.path.join(self.tmp, "build")
+
+    def test_pptx_text_image_manifest(self):
+        m = ex.extract(self.doc, self.out)
+        self.assertTrue(m["has_source"])
+        self.assertIn("brief.pptx", m["sources"])
+        # source.md 含标题
+        with open(os.path.join(self.out, "source.md"), encoding="utf-8") as f:
+            md = f.read()
+        self.assertIn("产品概览", md)
+        # 图片落盘 + manifest 正确
+        self.assertEqual(len(m["images"]), 1)
+        img = m["images"][0]
+        self.assertTrue(os.path.isfile(os.path.join(self.out, "assets", img["file"])))
+        self.assertEqual((img["w"], img["h"]), (1600, 900))
+        self.assertEqual(img["orient"], "landscape")
+        self.assertGreater(img["aspect_ratio"], 1.5)
+        self.assertEqual(img["section"], "产品概览")
+        self.assertIn("说明文字", img["near_text"])
+
+
+class TestExtractDocx(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.mkdtemp(prefix="extract-docx-")
+        self.png = os.path.join(self.tmp, "port.png")
+        _make_png(self.png, 600, 1000)               # 竖图
+        self.doc = os.path.join(self.tmp, "report.docx")
+        _make_docx(self.doc, self.png)
+        self.out = os.path.join(self.tmp, "build")
+
+    def test_docx_heading_image_assoc(self):
+        m = ex.extract(self.doc, self.out)
+        self.assertTrue(m["has_source"])
+        with open(os.path.join(self.out, "source.md"), encoding="utf-8") as f:
+            md = f.read()
+        self.assertIn("第一章 背景", md)
+        self.assertEqual(len(m["images"]), 1)
+        img = m["images"][0]
+        self.assertEqual((img["w"], img["h"]), (600, 1000))
+        self.assertEqual(img["orient"], "portrait")
+        self.assertLess(img["aspect_ratio"], 0.8)
+        self.assertEqual(img["section"], "第一章 背景")
+
+    def test_docx_word_media_fallback_without_duplicate(self):
+        import zipfile
+
+        extra = os.path.join(self.tmp, "extra.png")
+        _make_png(extra, 1200, 500)
+        with open(extra, "rb") as f:
+            blob = f.read()
+        with zipfile.ZipFile(self.doc, "a") as z:
+            z.writestr("word/media/orphan_extra.png", blob)
+
+        m = ex.extract(self.doc, self.out)
+        self.assertTrue(m["has_source"])
+        self.assertEqual(len(m["images"]), 2)
+        sizes = {(im["w"], im["h"]) for im in m["images"]}
+        self.assertIn((600, 1000), sizes)
+        self.assertIn((1200, 500), sizes)
+
+
+class TestExtractLegacyDoc(unittest.TestCase):
+    def test_doc_converts_then_uses_docx_extractor(self):
+        tmp = tempfile.mkdtemp(prefix="extract-doc-")
+        png = os.path.join(tmp, "port.png")
+        _make_png(png, 600, 1000)
+        converted = os.path.join(tmp, "converted.docx")
+        _make_docx(converted, png)
+        legacy = os.path.join(tmp, "legacy.doc")
+        with open(legacy, "wb") as f:
+            f.write(b"fake old word; conversion is mocked")
+        out = os.path.join(tmp, "build")
+
+        old = ex._convert_doc_to_docx
+        try:
+            ex._convert_doc_to_docx = lambda path, verbose: converted
+            m = ex.extract(legacy, out)
+        finally:
+            ex._convert_doc_to_docx = old
+
+        self.assertTrue(m["has_source"])
+        self.assertIn("legacy.doc", m["sources"])
+        self.assertEqual(len(m["images"]), 1)
+        with open(os.path.join(out, "source.md"), encoding="utf-8") as f:
+            md = f.read()
+        self.assertIn("老 Word `.doc` 已通过 LibreOffice 转换", md)
+        self.assertIn("第一章 背景", md)
+
+    def test_doc_conversion_error_does_not_claim_success(self):
+        tmp = tempfile.mkdtemp(prefix="extract-doc-fail-")
+        legacy = os.path.join(tmp, "bad.doc")
+        with open(legacy, "wb") as f:
+            f.write(b"bad")
+        out = os.path.join(tmp, "build")
+
+        old = ex._convert_doc_to_docx
+        try:
+            ex._convert_doc_to_docx = lambda path, verbose: (_ for _ in ()).throw(RuntimeError("LibreOffice missing"))
+            m = ex.extract(legacy, out)
+        finally:
+            ex._convert_doc_to_docx = old
+
+        self.assertFalse(m["has_source"])
+        self.assertEqual(m["images"], [])
+
+
+class TestScanAndEmpty(unittest.TestCase):
+    def test_scan_dir_finds_docs_and_standalone_images(self):
+        # 混合输入：a.pptx + b.docx + 两张独立图片，全部应被识别为 source。
+        tmp = tempfile.mkdtemp(prefix="extract-scan-")
+        land = os.path.join(tmp, "l.png")
+        port = os.path.join(tmp, "p.png")
+        _make_png(land, 1600, 900)
+        _make_png(port, 600, 1000)
+        _make_pptx(os.path.join(tmp, "a.pptx"), land)
+        _make_docx(os.path.join(tmp, "b.docx"), port)
+        out = os.path.join(tmp, "build")
+        m = ex.extract(tmp, out, scan=True)
+        self.assertTrue(m["has_source"])
+        # 4 个 source：pptx + docx + 2 张独立图片（独立图片现也是合法输入）
+        self.assertEqual(len(m["sources"]), 4)
+        # 4 张图：pptx 内嵌 1 + docx 内嵌 1 + 独立 2
+        self.assertEqual(len(m["images"]), 4)
+        # src 编号连续不撞（src_1..src_4）
+        files = {im["file"] for im in m["images"]}
+        self.assertEqual(len(files), 4)
+        # 独立图片以文件名为 section
+        sections = {im["section"] for im in m["images"]}
+        self.assertTrue("l" in sections and "p" in sections)
+
+    def test_no_source_graceful(self):
+        tmp = tempfile.mkdtemp(prefix="extract-empty-")
+        out = os.path.join(tmp, "build")
+        m = ex.extract(tmp, out, scan=True)
+        self.assertFalse(m["has_source"])
+        self.assertEqual(m["images"], [])
+        # manifest 仍落盘（下游 Read 不炸）
+        with open(os.path.join(out, "source_assets.json"), encoding="utf-8") as f:
+            self.assertFalse(json.load(f)["has_source"])
+
+
+class TestNewInputTypes(unittest.TestCase):
+    """csv / 独立图片 / xlsx 内嵌图（免 openpyxl）——本轮新增的输入类型。"""
+
+    def test_standalone_image_single_file(self):
+        tmp = tempfile.mkdtemp(prefix="extract-img-")
+        img = os.path.join(tmp, "封面图.png")
+        _make_png(img, 1920, 1080)
+        out = os.path.join(tmp, "build")
+        m = ex.extract(img, out)  # 直接传单个图片文件
+        self.assertTrue(m["has_source"])
+        self.assertEqual(len(m["images"]), 1)
+        one = m["images"][0]
+        self.assertEqual((one["w"], one["h"]), (1920, 1080))
+        self.assertEqual(one["orient"], "landscape")
+        self.assertEqual(one["section"], "封面图")  # section = 文件名（无扩展）
+        self.assertTrue(os.path.isfile(os.path.join(out, "assets", one["file"])))
+
+    def test_csv_to_markdown_table(self):
+        tmp = tempfile.mkdtemp(prefix="extract-csv-")
+        csv = os.path.join(tmp, "data.csv")
+        with open(csv, "w", encoding="utf-8") as f:
+            f.write("产品,销量,占比\n相机,1200,35%\n镜头,800,24%\n")
+        out = os.path.join(tmp, "build")
+        m = ex.extract(csv, out)
+        self.assertTrue(m["has_source"])
+        self.assertEqual(len(m["images"]), 0)  # csv 无图
+        with open(os.path.join(out, "source.md"), encoding="utf-8") as f:
+            md = f.read()
+        self.assertIn("| 产品 | 销量 | 占比 |", md)  # markdown 表头
+        self.assertIn("相机", md)
+
+    def test_xlsx_embedded_image_via_zipfile(self):
+        # openpyxl 未必安装：手工造一个含 xl/media 的 .xlsx（OOXML=zip），
+        # 验证「免 openpyxl 也能抽内嵌图」的 zipfile 路径。
+        import zipfile as zf
+        from PIL import Image
+
+        tmp = tempfile.mkdtemp(prefix="extract-xlsx-")
+        png_bytes_path = os.path.join(tmp, "_tmp.png")
+        Image.new("RGB", (800, 600), (10, 20, 30)).save(png_bytes_path)
+        with open(png_bytes_path, "rb") as f:
+            png_bytes = f.read()
+        xlsx = os.path.join(tmp, "报表.xlsx")
+        with zf.ZipFile(xlsx, "w") as z:
+            z.writestr("[Content_Types].xml", "<Types/>")  # 占位，足够 zipfile 读取
+            z.writestr("xl/media/image1.png", png_bytes)
+        out = os.path.join(tmp, "build")
+        m = ex.extract(xlsx, out)
+        self.assertTrue(m["has_source"])
+        self.assertEqual(len(m["images"]), 1)  # zipfile 抽到内嵌图（无需 openpyxl）
+        one = m["images"][0]
+        self.assertEqual((one["w"], one["h"]), (800, 600))
+        self.assertEqual(one["section"], "工作簿")
+        self.assertTrue(os.path.isfile(os.path.join(out, "assets", one["file"])))
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

package/skills/ui-design/scripts/generated_image_review_collect.py

@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""Collect generated-image semantic reviews and make a retry report.
+
+The workflow writes one JSON review per generated image using a multimodal LLM.
+This helper aggregates those results and creates a retry-degraded compatible
+report for GenerateImage, so only semantically failed images are retried.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from typing import Any
+import re
+
+
+BAKED_TEXT_RE = re.compile(
+    r"(?:文字|数字|编号|序号|标签|题注|水印|logo|Logo|字母|1\s*/\s*2\s*/\s*3\s*/\s*4|"
+    r"\b(?:text|word|words|number|numbers|digit|digits|letter|letters|label|caption|logo|watermark)\b)",
+    re.I,
+)
+
+
+def _load_json(path: Path, default: Any) -> Any:
+    if not path.exists():
+        return default
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        text = path.read_text(encoding="utf-8", errors="replace")
+        # Some LLMs wrap JSON in prose. Try a coarse extraction.
+        start = text.find("{")
+        end = text.rfind("}")
+        if start >= 0 and end > start:
+            try:
+                return json.loads(text[start : end + 1])
+            except Exception:
+                pass
+    return default
+
+
+def _is_pass(review: dict[str, Any]) -> bool:
+    value = review.get("pass")
+    if isinstance(value, bool):
+        return value
+    verdict = str(review.get("verdict") or review.get("status") or "").lower()
+    if verdict in ("pass", "ok", "passed"):
+        return True
+    if verdict in ("fail", "failed", "reject"):
+        return False
+    score = review.get("score")
+    try:
+        return float(score) >= 7.0
+    except Exception:
+        return False
+
+
+def _text_of(value: Any) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, str):
+        return value
+    if isinstance(value, (int, float, bool)):
+        return str(value)
+    if isinstance(value, list):
+        return " ".join(_text_of(v) for v in value)
+    if isinstance(value, dict):
+        return " ".join(_text_of(v) for v in value.values())
+    return str(value)
+
+
+def _is_faithful_or_dense(item: dict[str, Any] | None) -> bool:
+    if not isinstance(item, dict):
+        return False
+    text = _text_of(
+        {
+            "mode": item.get("mode"),
+            "role": item.get("role"),
+            "target_layer": item.get("target_layer"),
+            "enhance_style": item.get("enhance_style"),
+            "enhance_intent": item.get("enhance_intent"),
+            "image_role": item.get("image_role"),
+        }
+    ).lower()
+    return any(term in text for term in ("dense", "faithful", "document", "info"))
+
+
+def _is_background_like(file_name: str, item: dict[str, Any] | None) -> bool:
+    name = str(file_name or "").lower()
+    if "_fg." in name:
+        return False
+    if name.startswith("ai_p"):
+        return True
+    if not isinstance(item, dict):
+        return False
+    role_text = _text_of(
+        {
+            "role": item.get("role"),
+            "target_layer": item.get("target_layer"),
+            "visual_role": item.get("visual_role"),
+            "enhance_style": item.get("enhance_style"),
+            "enhance_intent": item.get("enhance_intent"),
+        }
+    ).lower()
+    return "background" in role_text or "bg" in role_text
+
+
+def _has_baked_text_issue(review: dict[str, Any], file_name: str, item: dict[str, Any] | None) -> bool:
+    if _is_faithful_or_dense(item) or not _is_background_like(file_name, item):
+        return False
+    scan = _text_of(
+        {
+            "reason": review.get("reason"),
+            "issues": review.get("issues"),
+            "blocking_issues": review.get("blocking_issues"),
+            "aesthetic_issues": review.get("aesthetic_issues"),
+            "description": review.get("description"),
+            "notes": review.get("notes"),
+        }
+    )
+    return bool(BAKED_TEXT_RE.search(scan))
+
+
+def collect(build: Path) -> dict[str, Any]:
+    build.mkdir(parents=True, exist_ok=True)
+    review_dir = build / "generated-image-reviews"
+    plan = _load_json(build / "img_plan.json", [])
+    by_file = {str(item.get("file")): item for item in plan if isinstance(item, dict) and item.get("file")}
+    reviews = []
+    failed = []
+    if review_dir.exists():
+        for path in sorted(review_dir.glob("*.json")):
+            data = _load_json(path, {})
+            file_name = data.get("file") or path.stem
+            if not str(file_name).lower().endswith((".png", ".jpg", ".jpeg", ".webp")):
+                # path.stem loses extension for files like ai_p1.png.json. Restore common case.
+                for candidate in by_file:
+                    if path.name == f"{candidate}.json":
+                        file_name = candidate
+                        break
+            item = by_file.get(str(file_name))
+            passed = _is_pass(data if isinstance(data, dict) else {})
+            baked_text_issue = False
+            if isinstance(data, dict) and _has_baked_text_issue(data, str(file_name), item):
+                passed = False
+                baked_text_issue = True
+            entry = {
+                "file": file_name,
+                "pass": passed,
+                "score": data.get("score") if isinstance(data, dict) else None,
+                "reason": data.get("reason") or data.get("issues") if isinstance(data, dict) else "invalid review json",
+            }
+            if baked_text_issue:
+                entry["reason"] = f"{entry['reason']}；背景/hero 图含不可编辑文字、数字、Logo 或标签，必须重试"
+                entry["hard_error"] = "baked_text_in_background"
+            reviews.append(entry)
+            if not passed and str(file_name) in by_file:
+                failed.append(
+                    {
+                        "file": str(file_name),
+                        "reason": entry["reason"] or "semantic_review_failed",
+                        "mode": by_file[str(file_name)].get("mode", "text-to-image"),
+                    }
+                )
+
+    report = {
+        "ok": len(failed) == 0,
+        "reviewed": len(reviews),
+        "failed_count": len(failed),
+        "reviews": reviews,
+        "failed": failed,
+    }
+    (build / "generated_image_review.json").write_text(
+        json.dumps(report, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    retry_report = {
+        "ok": len(failed) == 0,
+        "qualityOk": len(failed) == 0,
+        "degraded": failed,
+        "hardErrors": [],
+    }
+    (build / "generated_image_semantic_retry_report.json").write_text(
+        json.dumps(retry_report, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    lines = ["# 生成图语义复核", "", f"- reviewed: {len(reviews)}", f"- failed: {len(failed)}", ""]
+    for item in failed:
+        lines.append(f"- {item['file']}: {item.get('reason')}")
+    (build / "generated_image_review.md").write_text("\n".join(lines).rstrip() + "\n", encoding="utf-8")
+    print("GENERATED_IMAGE_REVIEW_OK" if not failed else "GENERATED_IMAGE_REVIEW_RETRY")
+    return report
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Collect generated image semantic reviews.")
+    parser.add_argument("--build", default="./build")
+    args = parser.parse_args()
+    collect(Path(args.build))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/skills/ui-design/scripts/generated_image_review_collect_test.py

@@ -0,0 +1,91 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import json
+import os
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+from generated_image_review_collect import collect  # noqa: E402
+
+
+class GeneratedImageReviewCollectTest(unittest.TestCase):
+    def test_background_baked_numbers_turn_pass_into_fail(self):
+        with tempfile.TemporaryDirectory() as d:
+            build = Path(d)
+            (build / "generated-image-reviews").mkdir()
+            (build / "img_plan.json").write_text(
+                json.dumps(
+                    [
+                        {
+                            "file": "ai_p2.png",
+                            "mode": "text-to-image",
+                            "prompt": "clean background, no text",
+                        }
+                    ],
+                    ensure_ascii=False,
+                ),
+                encoding="utf-8",
+            )
+            (build / "generated-image-reviews" / "ai_p2.png.json").write_text(
+                json.dumps(
+                    {
+                        "file": "ai_p2.png",
+                        "pass": True,
+                        "score": 7,
+                        "reason": "画面可用，但背景里有 1/2/3/4 数字，可能和前景章节序号重复。",
+                    },
+                    ensure_ascii=False,
+                ),
+                encoding="utf-8",
+            )
+
+            report = collect(build)
+            self.assertFalse(report["ok"])
+            self.assertEqual(report["failed_count"], 1)
+            retry = json.loads((build / "generated_image_semantic_retry_report.json").read_text(encoding="utf-8"))
+            self.assertFalse(retry["qualityOk"])
+            self.assertEqual(retry["degraded"][0]["file"], "ai_p2.png")
+
+    def test_faithful_document_can_keep_text(self):
+        with tempfile.TemporaryDirectory() as d:
+            build = Path(d)
+            (build / "generated-image-reviews").mkdir()
+            (build / "img_plan.json").write_text(
+                json.dumps(
+                    [
+                        {
+                            "file": "enhanced_src_5.png",
+                            "mode": "image-to-image",
+                            "enhance_style": "dense",
+                            "prompt": "faithfully restore screenshot",
+                        }
+                    ],
+                    ensure_ascii=False,
+                ),
+                encoding="utf-8",
+            )
+            (build / "generated-image-reviews" / "enhanced_src_5.png.json").write_text(
+                json.dumps(
+                    {
+                        "file": "enhanced_src_5.png",
+                        "pass": True,
+                        "score": 8,
+                        "reason": "信息截图保真增强，保留原有文字和数字。",
+                    },
+                    ensure_ascii=False,
+                ),
+                encoding="utf-8",
+            )
+
+            report = collect(build)
+            self.assertTrue(report["ok"])
+            self.assertEqual(report["failed_count"], 0)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)