minimal-agent 0.8.2 → 0.10.0

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

@@ -0,0 +1,301 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""Bounded visual quality gate for PPT HTML decks.
+
+Hard engineering checks still fail. Product-detail visual issues are repairable
+for at most N rounds; after that this script writes a residual report and lets
+the workflow continue, so the workflow does not loop forever during peak model
+traffic.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import subprocess
+from pathlib import Path
+
+
+def _run(cmd: list[str], cwd: Path) -> tuple[int, str]:
+    # 显式 UTF-8 解码子进程输出：中文 Windows 上 text=True 默认走 locale=GBK，
+    # 解码子脚本的 UTF-8 中文日志会崩（UnicodeDecodeError）。errors="replace" 再兜一层，
+    # 保证直接 CLI / 测试 / 非经 bash.ts 的宿主也不炸（bash.ts 已注入 PYTHONUTF8 兜底）。
+    proc = subprocess.run(
+        cmd,
+        cwd=str(cwd),
+        text=True,
+        encoding="utf-8",
+        errors="replace",
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+    )
+    return proc.returncode, proc.stdout
+
+
+def _load_json(path: Path, default):
+    if not path.exists():
+        return default
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return default
+
+
+def _errors_first(text: str, limit: int = 12000) -> str:
+    """截断前先把 ERROR/FAILED/【…】缺陷行提到段首（各组内保持原顺序）。
+
+    治真机事故（2026-07-02）：layout_balance 输出超 12000 字符时旧版 `text[-12000:]` 尾截，
+    **前几页的 ERROR 被裁掉**、修复 agent 只看得到后半段 WARNING → 第 9 页标题遮挡/越界没人修。
+    现在缺陷行永远在头部，截断只牺牲低价值 WARNING 尾巴。
+    """
+    if not text:
+        return ""
+    hot, cold = [], []
+    for line in text.splitlines():
+        is_hot = line.lstrip().startswith(("ERROR", "- 【", "【")) or "FAILED" in line
+        (hot if is_hot else cold).append(line)
+    return "\n".join(hot + cold)[:limit]
+
+
+def _write_report(build: Path, title: str, sections: list[tuple[str, str]]) -> None:
+    lines = [f"# {title}", ""]
+    for name, text in sections:
+        lines.append(f"## {name}")
+        lines.append("")
+        lines.append("```text")
+        lines.append(_errors_first(text))
+        lines.append("```")
+        lines.append("")
+    (build / "visual_quality_report.md").write_text("\n".join(lines), encoding="utf-8")
+
+
+def _generated_review_log(build: Path) -> tuple[bool, str]:
+    report = _load_json(build / "generated_image_review.json", {})
+    if not isinstance(report, dict) or not report:
+        return True, (
+            "generated_image_review.json 不存在或为空；跳过生成图语义硬门"
+            "（阶段7c 已按 2026-07-02 裁决屏蔽 LLM 图像复核，此为预期状态、非 bug；"
+            "生成图带字风险由 image_prompt_rewrite 背景层去内容化 + image_plan_lint 确定性拦截）。"
+        )
+    ok = bool(report.get("ok", False))
+    failed = report.get("failed") if isinstance(report.get("failed"), list) else []
+    reviews = report.get("reviews") if isinstance(report.get("reviews"), list) else []
+    lines = [
+        f"ok={ok}",
+        f"reviewed={report.get('reviewed', len(reviews))}",
+        f"failed_count={report.get('failed_count', len(failed))}",
+    ]
+    for item in failed[:20]:
+        if isinstance(item, dict):
+            lines.append(f"- {item.get('file')}: {item.get('reason')}")
+    for item in reviews[:20]:
+        if isinstance(item, dict) and item.get("hard_error"):
+            lines.append(f"- {item.get('file')}: {item.get('hard_error')} {item.get('reason')}")
+            ok = False
+    return ok, "\n".join(lines)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Run bounded visual quality checks.")
+    parser.add_argument("--build", default="./build")
+    parser.add_argument("--max-rounds", type=int, default=3)
+    parser.add_argument("--strict-product-detail", action="store_true")
+    args = parser.parse_args()
+
+    root = Path.cwd()
+    build = Path(args.build)
+    build.mkdir(parents=True, exist_ok=True)
+    state_path = build / "visual_quality_state.json"
+    state = _load_json(state_path, {"round": 0})
+    current_round = int(state.get("round") or 0) + 1
+    state["round"] = current_round
+    state_path.write_text(json.dumps(state, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    ok_path = build / "visual_quality_ok"
+    residual_path = build / "visual_quality_residual.md"
+    ok_path.unlink(missing_ok=True)
+
+    # lint 只扫 img_plan prompt（不带 --outline）：ppt 链路里 t2i prompt 已由
+    # image_prompt_rewrite.py 确定性净化（共享 lint 词表），design.visual.note 到不了生图模型，
+    # 扫 note 只剩误伤面（页 note/标题含「架构/对比图」即 halt 的假阳事故）。
+    # 若未来移除 image_prompt_rewrite，需恢复 --outline ./build/design.json。
+    lint_cmd = [
+        "python3",
+        "skills/ui-design/scripts/image_plan_lint.py",
+        str(build / "img_plan.json"),
+        "-v",
+    ]
+    source_cmd = [
+        "python3",
+        "skills/ui-design/scripts/source_usage_check.py",
+        "--build",
+        str(build),
+        "--require-existing-derived",
+    ]
+    style_cmd = [
+        "python3",
+        "skills/ui-design/scripts/deck_visual_style_check.py",
+        str(build / "deck.html"),
+        "--design",
+        str(build / "design.json"),
+        "--image-report",
+        str(build / "image_generation_report.json"),
+        "--source-plan",
+        str(build / "source_layout_plan.json"),
+        "-v",
+    ]
+    if args.strict_product_detail:
+        style_cmd.append("--strict-product-detail")
+
+    # 版式规划执行检查（2026-07-03 真机第4页事故）：contain 类客户图被 build 违规铺成整页背景。
+    # Edit 可修 → 归**可修档**驱动修复轮（source_usage_check 主模式在 hard 档，违规不能混进去 halt）。
+    intent_cmd = [
+        "python3",
+        "skills/ui-design/scripts/source_usage_check.py",
+        "--build",
+        str(build),
+        "--layout-intent-only",
+    ]
+
+    lint_status, lint_log = _run(lint_cmd, root)
+    source_status, source_log = _run(source_cmd, root)
+    intent_status, intent_log = _run(intent_cmd, root)
+    style_status, style_log = _run(style_cmd, root)
+    generated_review_ok, generated_review_log = _generated_review_log(build)
+
+    # 几何构图门（2026-07-02 接线升级：catastrophic-only → 全量 + --ppt-strict）：
+    # 旧口径只判「内容不可见灾难」——第 9 页标题被卡遮挡 / 内容超出下边缘这类**测到了却不驱动修复**
+    # （全量只在导出阶段 || true 报告态），41 个缺陷带残留放行是真机事故的直接原因。
+    # 现在全量跑 + --ppt-strict（重叠去噪防追鬼 / 越界升 ERROR / 大空卡升 ERROR，见 layout_balance.py），
+    # 归入**可修档**（visual_failed）驱动 repair；「内容不可见」零容忍仍由导出阶段 catastrophic 硬门兜底。
+    export_dir = build / "export"
+    geometry_status = 0
+    if (export_dir / "layout-report.json").exists() or (export_dir / "export-manifest.json").exists():
+        geometry_cmd = [
+            "python3",
+            "skills/ui-design/scripts/layout_balance.py",
+            str(export_dir),
+            "--ppt-strict",
+            "--no-occupancy",
+        ]
+        # source_layout_plan.json 只在有客户图版式规划时才存在（webppt 无此文件）——传入路径即可，
+        # layout_balance.py 自行判空跳过，额外反查 is_dense_info 客户图渲染宽度是否达标（WARNING）。
+        source_layout_plan = build / "source_layout_plan.json"
+        if source_layout_plan.exists():
+            geometry_cmd += ["--source-layout-plan", str(source_layout_plan)]
+        geometry_status, geometry_log = _run(geometry_cmd, root)
+    else:
+        geometry_log = (
+            "layout-report.json 不存在（应由本阶段起头的 measure-only 产出）；跳过几何复核"
+            "——不阻断，但意味着本轮未做几何检查（如 chromium 缺失导致 measure 失败）。"
+        )
+
+    # 像素级文字可读性门（2026-07-02 新增）：CSS 对比度在「文字压背景图」页物理上测不准
+    # （resolveBg 只拿得到页面底色）——用 measure-only --with-bg-shots 落的背景截图采真实像素。
+    # 缺 PIL/缺截图时脚本响亮 SKIP 且 exit 0（fail-open），不阻断。归可修档。
+    readability_status, readability_log = _run(
+        [
+            "python3",
+            "skills/ui-design/scripts/text_readability_pixel.py",
+            str(export_dir),
+        ],
+        root,
+    )
+
+    # SVG 图表一致性门（2026-07-02 新增）：柱高与百分比标签颠倒 / 标签装不下（slide11/19 事故）。
+    # 复杂图 fail-open 跳过。归可修档。
+    chart_status, chart_log = _run(
+        [
+            "python3",
+            "skills/ui-design/scripts/chart_sanity_check.py",
+            str(build / "deck.html"),
+        ],
+        root,
+    )
+
+    # 用户裁决（2026-07-02）：构图/视觉验证门**只吃确定性脚本信号**，LLM 图像识别裁决屏蔽——
+    # generated_image_review（多模态语义复核）不再进 hard_failed（ppt.yaml 阶段14-16 已同步 when:"false"）。
+    # 复核结果仍记录进报告（信息性，供人工参考/未来恢复）；恢复 = 把 `or not generated_review_ok` 加回。
+    hard_failed = (
+        lint_status != 0
+        or source_status != 0
+    )
+    visual_failed = (
+        style_status != 0
+        or intent_status != 0
+        or geometry_status != 0
+        or readability_status != 0
+        or chart_status != 0
+    )
+
+    report = {
+        "round": current_round,
+        "max_rounds": args.max_rounds,
+        "hard_ok": not hard_failed,
+        "visual_ok": not visual_failed,
+        "ok": not hard_failed and not visual_failed,
+        "accepted_with_residual": False,
+        "lint_status": lint_status,
+        "source_status": source_status,
+        "intent_status": intent_status,
+        "style_status": style_status,
+        "geometry_status": geometry_status,
+        "readability_status": readability_status,
+        "chart_status": chart_status,
+        "generated_image_review_ok": generated_review_ok,
+    }
+    (build / "visual_quality_report.json").write_text(
+        json.dumps(report, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    _write_report(
+        build,
+        f"视觉质量门第 {current_round}/{args.max_rounds} 轮",
+        [
+            ("image_plan_lint", lint_log),
+            ("generated_image_review", generated_review_log),
+            ("source_usage_check", source_log),
+            ("source_layout_intent (版式规划执行)", intent_log),
+            ("deck_visual_style_check", style_log),
+            ("layout_balance (几何·全量 --ppt-strict)", geometry_log),
+            ("text_readability_pixel (像素级文字压图可读性)", readability_log),
+            ("chart_sanity_check (SVG 图表一致性)", chart_log),
+        ],
+    )
+
+    if hard_failed:
+        print("VISUAL_QUALITY_HARD_FAILED")
+        return 1
+    if not visual_failed:
+        ok_path.touch()
+        residual_path.unlink(missing_ok=True)
+        print("VISUAL_QUALITY_OK")
+        return 0
+    if current_round >= args.max_rounds:
+        # 达轮数**带残留放行**（落实 docstring「lets the workflow continue」与「逐页有界修复后放行」裁决）：
+        # 视觉档（style/geometry）问题记录进 residual 后 touch ok 放行导出——fast（max_rounds=1）没有
+        # 修复步，不放行 = 无人能修的确定性死端；best 修满轮数后同理不再空转。
+        # 硬档（lint/source/生成图复核）不走这里（上面已 exit 1）；「内容不可见」灾难仍由
+        # stage_export 的 layout_balance --catastrophic-only 硬门兜底，绝不因放行而出货不可见 deck。
+        report["accepted_with_residual"] = True
+        (build / "visual_quality_report.json").write_text(
+            json.dumps(report, ensure_ascii=False, indent=2),
+            encoding="utf-8",
+        )
+        residual_path.write_text(
+            "# 视觉质量残留问题（带残留放行）\n\n"
+            f"已达到最多 {args.max_rounds} 轮视觉修复，带残留放行导出。以下问题供人工复核参考；\n"
+            "「内容不可见」灾难仍由导出阶段的 catastrophic 几何硬门兜底拦截。\n\n"
+            + (build / "visual_quality_report.md").read_text(encoding="utf-8"),
+            encoding="utf-8",
+        )
+        ok_path.touch()
+        print("VISUAL_QUALITY_ACCEPTED_WITH_RESIDUAL")
+        return 0
+
+    print("VISUAL_QUALITY_NEEDS_REPAIR")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

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

@@ -0,0 +1,233 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+
+HERE = Path(__file__).resolve().parent
+SCRIPT = HERE / "visual_quality_gate.py"
+ROOT = HERE.parents[2]
+
+
+def _write(path: Path, text: str) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(text, encoding="utf-8")
+
+
+class VisualQualityGateTest(unittest.TestCase):
+    def test_visual_residual_accepted_after_max_rounds(self):
+        # 达 --max-rounds 且仅视觉档失败 → **带残留放行**：ok 与 residual 并存、
+        # accepted_with_residual=true、打印 ACCEPTED_WITH_RESIDUAL。
+        # （原断言「残留不产 ok」已知情反转：fast 无修复步，阻断=无人能修的确定性死端；
+        #  docstring 与「有界修复后放行」裁决都是放行语义。硬档 halt 见下一测试，不受影响。）
+        with tempfile.TemporaryDirectory() as d:
+            build = Path(d)
+            _write(
+                build / "deck.html",
+                """
+                <html><body>
+                <section class="slide fullbleed" data-screen-label="01 测试">
+                  <div class="edit-bg"><img class="bg-img" src="assets/ai_p1.png"></div>
+                  <div class="edit-content"><h1>只有文字</h1></div>
+                </section>
+                </body></html>
+                """,
+            )
+            _write(
+                build / "design.json",
+                json.dumps(
+                    {
+                        "pages": [
+                            {
+                                "page": 1,
+                                "composition_style": "scenario-fullbleed",
+                                "page_grammar": "hero-scene",
+                                "max_visible_words": 80,
+                                "supporting_visual_layer": "前景产品特写",
+                            }
+                        ]
+                    },
+                    ensure_ascii=False,
+                ),
+            )
+            _write(build / "img_plan.json", "[]")
+            _write(build / "source_assets.json", json.dumps({"images": []}, ensure_ascii=False))
+            _write(build / "source_layout_plan.json", json.dumps({"enhance_entries": []}, ensure_ascii=False))
+            _write(build / "image_generation_report.json", json.dumps({"degraded": []}, ensure_ascii=False))
+
+            res = subprocess.run(
+                [sys.executable, str(SCRIPT), "--build", str(build), "--max-rounds", "1", "--strict-product-detail"],
+                cwd=str(ROOT),
+                text=True,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                check=False,
+            )
+            self.assertEqual(res.returncode, 0, res.stdout)
+            self.assertIn("VISUAL_QUALITY_ACCEPTED_WITH_RESIDUAL", res.stdout)
+            self.assertTrue((build / "visual_quality_residual.md").exists())
+            self.assertTrue((build / "visual_quality_ok").exists())
+            report = json.loads((build / "visual_quality_report.json").read_text(encoding="utf-8"))
+            self.assertTrue(report.get("accepted_with_residual"), report)
+
+    def test_generated_image_review_failure_is_informational_only(self):
+        # 知情反转（2026-07-02 用户裁决）：构图/视觉验证门只吃**确定性脚本**信号，
+        # LLM 图像识别复核（generated_image_review.json）已屏蔽出 hard_failed——
+        # 即便复核 ok=false 也**不**硬 halt，只记录进报告（generated_image_review_ok=false）。
+        # 原断言「复核 fail → VISUAL_QUALITY_HARD_FAILED」是被反转的旧行为。
+        with tempfile.TemporaryDirectory() as d:
+            build = Path(d)
+            _write(
+                build / "deck.html",
+                """
+                <html><body>
+                <section class="slide fullbleed" data-screen-label="01 测试">
+                  <div class="edit-bg"><img class="bg-img" src="assets/ai_p1.png"></div>
+                  <div class="edit-content"><h1>全过程可追溯</h1><p>一眼看见关键链路。</p><img class="product-visual" src="assets/hero.png"></div>
+                </section>
+                </body></html>
+                """,
+            )
+            _write(
+                build / "design.json",
+                json.dumps(
+                    {
+                        "pages": [
+                            {
+                                "page": 1,
+                                "composition_style": "scenario-fullbleed",
+                                "page_grammar": "hero-scene",
+                                "max_visible_words": 80,
+                                "supporting_visual_layer": "前景产品特写",
+                            }
+                        ]
+                    },
+                    ensure_ascii=False,
+                ),
+            )
+            _write(build / "img_plan.json", "[]")
+            _write(build / "source_assets.json", json.dumps({"images": []}, ensure_ascii=False))
+            _write(build / "source_layout_plan.json", json.dumps({"enhance_entries": []}, ensure_ascii=False))
+            _write(build / "image_generation_report.json", json.dumps({"degraded": []}, ensure_ascii=False))
+            _write(
+                build / "generated_image_review.json",
+                json.dumps(
+                    {
+                        "ok": False,
+                        "failed_count": 1,
+                        "failed": [{"file": "ai_p1.png", "reason": "背景含数字"}],
+                    },
+                    ensure_ascii=False,
+                ),
+            )
+
+            res = subprocess.run(
+                [sys.executable, str(SCRIPT), "--build", str(build), "--max-rounds", "1", "--strict-product-detail"],
+                cwd=str(ROOT),
+                text=True,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                check=False,
+            )
+            self.assertEqual(res.returncode, 0, res.stdout)
+            self.assertNotIn("VISUAL_QUALITY_HARD_FAILED", res.stdout)
+            report = json.loads((build / "visual_quality_report.json").read_text(encoding="utf-8"))
+            self.assertTrue(report.get("hard_ok"), report)  # 硬档只由确定性脚本决定
+            self.assertFalse(report.get("generated_image_review_ok"), report)  # 复核结果仍被记录（信息性）
+            self.assertTrue((build / "visual_quality_ok").exists())  # 放行（直接 OK 或带残留放行皆产 ok）
+
+    def test_catastrophic_geometry_is_repairable_not_hard(self):
+        # F2 回归锁：有 layout-report.json 且几何 catastrophic（空页无视觉）时——
+        # 几何走**可修档**(visual_failed)而非 hard_failed：returncode==0（非 HARD_FAILED）、
+        # report.geometry_status!=0。改前（几何算 hard）此用例会 return 1。
+        # 达轮数后带残留放行（ok 会存在）——「内容不可见不出货」由 stage_export 的
+        # layout_balance --catastrophic-only 硬门保证（该门行为不动），不靠此处阻断。
+        with tempfile.TemporaryDirectory() as d:
+            build = Path(d)
+            _write(
+                build / "deck.html",
+                """
+                <html><body>
+                <section class="slide" data-screen-label="01 空页"><div class="edit-content"></div></section>
+                </body></html>
+                """,
+            )
+            _write(
+                build / "design.json",
+                json.dumps({"pages": [{"page": 1, "composition_style": "scenario-fullbleed",
+                                        "page_grammar": "hero-scene", "max_visible_words": 80}]},
+                           ensure_ascii=False),
+            )
+            _write(build / "img_plan.json", "[]")
+            _write(build / "source_assets.json", json.dumps({"images": []}, ensure_ascii=False))
+            _write(build / "source_layout_plan.json", json.dumps({"enhance_entries": []}, ensure_ascii=False))
+            _write(build / "image_generation_report.json", json.dumps({"degraded": []}, ensure_ascii=False))
+            # 灾难级几何：content_boxes==0 且无任何真实视觉（无底图/无前景切图）→ catastrophic ERROR。
+            empty_balance = {
+                "grid": {"cols": 16, "rows": 8}, "contentBoxes": 0,
+                "hasBgImage": False, "hasForegroundVisual": False, "coverage": 0.0,
+                "halfOccupancy": {"top": 0, "bottom": 0, "left": 0, "right": 0},
+                "bbox": {"x": 0, "y": 0, "w": 1920, "h": 1080}, "centroid": {"x": 960, "y": 540},
+                "overlaps": [], "cards": [], "overflowRuns": [], "smallText": [],
+                "bakedText": [], "offCanvas": [], "lowContrast": [],
+            }
+            _write(
+                build / "export" / "layout-report.json",
+                json.dumps({"mode": "hybrid", "width": 1920, "height": 1080,
+                            "slides": [{"index": 1, "label": "01 空页", "balance": empty_balance}]},
+                           ensure_ascii=False),
+            )
+
+            res = subprocess.run(
+                [sys.executable, str(SCRIPT), "--build", str(build), "--max-rounds", "1", "--strict-product-detail"],
+                cwd=str(ROOT),
+                text=True,
+                encoding="utf-8",
+                stdout=subprocess.PIPE,
+                stderr=subprocess.STDOUT,
+                check=False,
+            )
+            self.assertEqual(res.returncode, 0, res.stdout)  # 几何不再 hard-fail
+            self.assertNotIn("VISUAL_QUALITY_HARD_FAILED", res.stdout)
+            report = json.loads((build / "visual_quality_report.json").read_text(encoding="utf-8"))
+            self.assertNotEqual(report.get("geometry_status"), 0, report)  # 几何确被检出
+            # max-rounds=1 → 本轮即达轮数：带残留放行（ok+residual 并存，出货兜底在导出 catastrophic 硬门）
+            self.assertTrue(report.get("accepted_with_residual"), report)
+            self.assertTrue((build / "visual_quality_ok").exists())
+            self.assertTrue((build / "visual_quality_residual.md").exists())
+
+
+class ErrorsFirstTest(unittest.TestCase):
+    """报告截断前 ERROR 行提头（2026-07-02：layout_balance 超 12000 字符尾截把第 9 页 ERROR 裁掉的事故）。"""
+
+    def test_errors_moved_to_head_before_truncation(self):
+        sys.path.insert(0, str(HERE))
+        from visual_quality_gate import _errors_first
+        text = chr(10).join(
+            [f"WARNING: 噪声 {i}" for i in range(3)]
+            + ["  - 【内容相互叠加】09 页 · 标题被卡压", "ERROR: 硬缺陷一条"]
+            + [f"WARNING: 噪声尾 {i}" for i in range(3)]
+        )
+        out = _errors_first(text)
+        lines = out.splitlines()
+        self.assertIn("【内容相互叠加】", lines[0])
+        self.assertTrue(lines[1].startswith("ERROR"))
+        # WARNING 全保留（无截断场景不丢行）
+        self.assertEqual(len(lines), 8)
+
+    def test_truncation_sacrifices_tail_warnings_not_errors(self):
+        from visual_quality_gate import _errors_first
+        text = chr(10).join(["ERROR: 关键缺陷"] + ["WARNING: " + "x" * 100 for _ in range(300)])
+        out = _errors_first(text, limit=2000)
+        self.assertTrue(out.startswith("ERROR: 关键缺陷"))
+        self.assertLessEqual(len(out), 2000)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

package/skills/ui-design/tests/e2e/prompts/deck.md

@@ -0,0 +1,19 @@
+# Deck 端到端测试题
+
+使用 `ui-design` 技能，为 **舒适视觉 第三季度港口韧性简报** 创建一个 5 页 HTML deck。
+
+受众：港口管理层和计划负责人。
+
+设计约束：
+- 视觉上像可编辑的 PPT 风格管理层简报，但实现方式是 HTML。
+- 使用固定 16:9 舞台，并实现 viewport scaling。
+- slide 必须使用 1 起始 `data-screen-label`，例如 `01 标题`。
+- 使用 `localStorage` 持久化当前 slide。
+- 提供上一页/下一页控件和键盘导航。
+- 视觉系统要与 Web 页面测试相关，但节奏要更适合演示文稿。
+- 避免密集段落页、填充结论、通用渐变背景和素材库感装饰。
+- 交付一个能通过 `scripts/check_ui_artifact.py` 和 `scripts/check_language_policy.py` 的单文件 HTML。
+- 所有用户可见文案必须使用简体中文，品牌名 舒适视觉 可以保留英文。
+
+期望输出路径：
+`minimal-agent/skills/ui-design/tests/e2e/runs/iteration-1/deck/outputs/deck.html`

package/skills/ui-design/tests/e2e/prompts/webpage.md

@@ -0,0 +1,17 @@
+# Web 页面端到端测试题
+
+使用 `ui-design` 技能，为 **舒适视觉** 创建一个高保真 Web 页面。舒适视觉 是面向区域港口管理机构的气候风险运营产品。
+
+受众：运营总监。他们需要在每周计划会之前理解泊位中断风险、排放暴露和天气导致的容量变化。
+
+设计约束：
+- 首屏必须像真实运营产品，而不是泛泛 SaaS 落地页。
+- 使用克制但有辨识度的视觉系统：海事中性色、信号色、精确字体、高密度但可读的信息，以及一个令人记住的视觉母题。
+- 内容要真实可信：船舶队列、天气窗口、泊位状态、排放暴露和计划动作。
+- 至少包含一个使用 `ui-design` edit-mode 协议的可调设计维度。
+- 避免填充指标、紫色渐变、通用图标网格、表情符号和装饰性卡片堆叠。
+- 交付一个能通过 `scripts/check_ui_artifact.py` 和 `scripts/check_language_policy.py` 的单文件 HTML。
+- 所有用户可见文案必须使用简体中文，品牌名 舒适视觉 可以保留英文。
+
+期望输出路径：
+`minimal-agent/skills/ui-design/tests/e2e/runs/iteration-1/webpage/outputs/index.html`

package/skills/ui-design/tests/e2e/render_check.mjs

@@ -0,0 +1,101 @@
+#!/usr/bin/env node
+import { mkdir, writeFile } from 'node:fs/promises';
+import { resolve } from 'node:path';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+import { chromium } from 'playwright-core';
+
+const skillRoot = resolve(fileURLToPath(new URL('../..', import.meta.url)));
+const runRoot = resolve(skillRoot, 'tests/e2e/runs/iteration-1');
+
+const targets = {
+  webpage: {
+    html: resolve(runRoot, 'webpage/outputs/index.html'),
+    shots: resolve(runRoot, 'webpage/screenshots'),
+    checks: resolve(runRoot, 'webpage/checks/browser-check.json'),
+  },
+  deck: {
+    html: resolve(runRoot, 'deck/outputs/deck.html'),
+    shots: resolve(runRoot, 'deck/screenshots'),
+    checks: resolve(runRoot, 'deck/checks/browser-check.json'),
+  },
+};
+
+async function inspectPage(page) {
+  return page.evaluate(() => ({
+    title: document.title,
+    screenLabels: Array.from(document.querySelectorAll('[data-screen-label]')).map((node) =>
+      node.getAttribute('data-screen-label'),
+    ),
+    editModeBlocks: (document.documentElement.outerHTML.match(/EDITMODE-BEGIN/g) || []).length,
+    hasEditModeAvailable: document.documentElement.outerHTML.includes('__edit_mode_available'),
+    bodyOverflowX: getComputedStyle(document.body).overflowX,
+    documentWidth: document.documentElement.scrollWidth,
+    viewportWidth: window.innerWidth,
+  }));
+}
+
+async function main() {
+  const browser = await chromium.launch({ headless: true });
+  const results = {};
+
+  try {
+    const page = await browser.newPage({ viewport: { width: 1440, height: 980 } });
+    const webpageConsole = [];
+    page.on('console', (msg) => webpageConsole.push({ type: msg.type(), text: msg.text() }));
+    page.on('pageerror', (error) => webpageConsole.push({ type: 'pageerror', text: error.message }));
+    await page.goto(pathToFileURL(targets.webpage.html).href, { waitUntil: 'load' });
+    await mkdir(targets.webpage.shots, { recursive: true });
+    await page.screenshot({ path: resolve(targets.webpage.shots, 'desktop.png'), fullPage: true });
+    const webpageDesktop = await inspectPage(page);
+    await page.setViewportSize({ width: 390, height: 844 });
+    await page.reload({ waitUntil: 'load' });
+    await page.screenshot({ path: resolve(targets.webpage.shots, 'mobile.png'), fullPage: true });
+    const webpageMobile = await inspectPage(page);
+    results.webpage = {
+      console: webpageConsole,
+      desktop: webpageDesktop,
+      mobile: webpageMobile,
+      pass: webpageConsole.filter((item) => item.type === 'error' || item.type === 'pageerror').length === 0,
+    };
+
+    const deckPage = await browser.newPage({ viewport: { width: 1440, height: 900 } });
+    const deckConsole = [];
+    deckPage.on('console', (msg) => deckConsole.push({ type: msg.type(), text: msg.text() }));
+    deckPage.on('pageerror', (error) => deckConsole.push({ type: 'pageerror', text: error.message }));
+    await deckPage.goto(pathToFileURL(targets.deck.html).href, { waitUntil: 'load' });
+    await mkdir(targets.deck.shots, { recursive: true });
+    await deckPage.screenshot({ path: resolve(targets.deck.shots, 'slide-01.png'), fullPage: false });
+    const deckSlideOne = await inspectPage(deckPage);
+    await deckPage.getByRole('button', { name: '下一页' }).click();
+    await deckPage.screenshot({ path: resolve(targets.deck.shots, 'slide-02.png'), fullPage: false });
+    const deckSlideTwo = await deckPage.evaluate(() => ({
+      activeSlide: localStorage.getItem('comfort-vision-active-slide'),
+      counter: document.getElementById('slide-counter')?.textContent,
+      visibleLabel: Array.from(document.querySelectorAll('.slide'))
+        .find((node) => !node.hidden)
+        ?.getAttribute('data-screen-label'),
+    }));
+    results.deck = {
+      console: deckConsole,
+      slideOne: deckSlideOne,
+      slideTwo: deckSlideTwo,
+      pass:
+        deckConsole.filter((item) => item.type === 'error' || item.type === 'pageerror').length === 0 &&
+        deckSlideTwo.activeSlide === '1' &&
+        deckSlideTwo.visibleLabel === '02 运营态势',
+    };
+  } finally {
+    await browser.close();
+  }
+
+  await mkdir(resolve(runRoot, 'webpage/checks'), { recursive: true });
+  await mkdir(resolve(runRoot, 'deck/checks'), { recursive: true });
+  await writeFile(targets.webpage.checks, JSON.stringify(results.webpage, null, 2));
+  await writeFile(targets.deck.checks, JSON.stringify(results.deck, null, 2));
+  console.log(JSON.stringify({ webpage: results.webpage.pass, deck: results.deck.pass }, null, 2));
+}
+
+main().catch((error) => {
+  console.error(error);
+  process.exit(1);
+});