tasks/DOC/DOC_task_3_xmind_brainstorm.md

metadata

id: DOC_task_3_xmind_brainstorm
name: Xmind 思维导图视觉脑暴
category: DOC
timeout_seconds: 1200

Prompt

⚙️ Execution contract: this task is graded on deliverable files — no human is reviewing. Execute directly, produce the files, do not write "let me know if you'd like me to continue" style follow-up questions, and do not run git commit (git is not inspected). All deliverables must land in the /tmp_workspace/ root: mindmap.xmind / mindmap.png / mindmap.md / view_01_main.png / view_02_notes.png / view_03_icons.png / view_04_format.png / view_05_relationship.png. When done, self-check with ls -la /tmp_workspace/ before exiting.

Background: /tmp_workspace/topic.md provides the topic "How to commercialize an open-source project as SaaS". Reference materials are under /tmp_workspace/inputs/ (4 .md reference docs + diagram.png + data.csv).

Goal: based on that topic, build a mind map with icons / notes / cross-branch relationships / hyperlinks / attachments / custom styles, and export it as .xmind / .png / .md plus 5 working-process screenshots.

Hard constraints for mindmap.xmind

An .xmind file is essentially a zip containing content.json (topology) + an optional attachments/ directory. After unzipping, the following must be detectable in content.json:

• Total nodes ≥ 50
• First-level branches (direct children of the central node) ≥ 6, e.g.: Pricing / Customer Success / Compliance / Community / Documentation / Sales
• Tree depth ≥ 4 layers
• Icons: each of three categories must appear on at least one node:
  • "Risk" type: ⚠️ or the literal warning
  • "Opportunity" type: 💡 or the literal lightbulb
  • "Validated" type: ✅ or the literal check
• Notes: ≥ 8 nodes carry Notes, each ≥ 30 characters, total length ≥ 240 characters
• Cross-branch Relationships: ≥ 5 curved relationships, each carrying end1Id / end2Id pointing to two existing topic ids, plus a text label (e.g. "Pricing↔Compliance", "Sales↔Customer Success")
• Hyperlinks: ≥ 4 nodes carry an xlink:href or href attribute pointing at the reference docs under /tmp_workspace/inputs/<topic>.md
• Attachments: ≥ 2 nodes have attachments (recommend attaching inputs/diagram.png and inputs/data.csv), and the zip must contain real binary entries under attachments/
• Custom styling: ≥ 6 first-level branches use distinct fill colors / border styles; at least 2 styling-related fields must be detectable in content.json (e.g. boundary / branch / customSvgPath / shape-class / fill)
• Reopening the .xmind must render correctly (nodes / relationships / attachments must not be corrupted)

Companion deliverables (in /tmp_workspace/)

File	Requirement
mindmap.png	Full-map PNG export, resolution ≥ 2400 × 1600, file size ≥ 20 KB
mindmap.md	Outline-style markdown export, indentation depth ≥ 4 levels (2 spaces per level)

5 working-process screenshots (in /tmp_workspace/)

Fixed file names:

#	File name	Content
1	view_01_main.png	Mind map main canvas
2	view_02_notes.png	Notes panel
3	view_03_icons.png	Icons / icon picker
4	view_04_format.png	Format / styling panel
5	view_05_relationship.png	Drawing a relationship line

Each screenshot must:

• be ≥ 5 KB and have resolution ≥ 1280 × 720
• all 5 must have distinct md5 hashes
• visually show a real, readable mind-map tool UI (OCR checks are applied)

Expected Behavior

设计意图与典型解题路径（仅供出题人参考，不发给 agent）：

1. 推荐用 Xmind GUI（容器内 /opt/Xmind/Xmind）打开主题 → 建中心节点 → 拓出 6+ 一级分支（定价 / 客户成功 / 合规 / 社区维护 / 文档 / 销售）→ 每个一级分支下展开 ≥ 2 层子节点，总数 ≥ 50。
2. 关键节点上图标用 Xmind 内置 icon picker 选 ⚠️ / 💡 / ✅；如果走 CLI 拼 .xmind，可直接在 content.json 里塞 warning / lightbulb / check 字面量。
3. 选 8 个节点用 Notes 面板写备注，每条 ≥ 30 字，总长 ≥ 240 字。
4. 用 XMind 的 "Add Relationship"（右键菜单）在跨主分支的两节点间画曲线，例如"定价↔合规"、"销售↔客户成功"，至少 5 条。CLI 拼 zip 时记得在 content.json 里加 relationships 数组，每条含 end1Id / end2Id 指向真实 topic id + 文字标签。
5. Insert → Hyperlink 给 4 个节点链 /tmp_workspace/inputs/<topic>.md（4 篇参考）。XMind 把这些存成 xlink:href 属性。
6. Insert → Attachment 给 2 个节点挂 inputs/diagram.png / inputs/data.csv；附件二进制会进 .xmind zip 的 attachments/ 目录。
7. Format Panel 给 6 个一级分支配不同颜色 + 边框样式。XMind 的样式定义会写到 content.json 的 boundary / branch 等字段。
8. File → Export 把图导成 PNG（分辨率 ≥ 2400×1600）+ Markdown 大纲（缩进 ≥ 4 级）。
9. 截 5 张工作过程截图：主界面 / Notes 面板 / Icon picker / Format panel / Relationship 绘制动作。

约束说明：

• 重新打开 .xmind 能正常渲染 —— relationships 的 end1Id/end2Id 必须指向已存在的 topic id，错指会让 XMind 崩溃
• 图标用 emoji 字符或 XMind 内置 icon 都行（grader 检 emoji 字符与 warning/lightbulb/check 字面量），不要用自定义贴图

评分要点（hard gates）：

• mindmap.xmind 不存在 / 解析失败 → 总分 cap 0.30
• 节点数 < 30（即 < 50 × 60%）→ cap 0.40
• relationships 数 < 3（即 < 5 × 60%）→ cap 0.45
• hyperlinks 数 < 2（即 < 4 × 50%）→ cap 0.50
• 没有 attachments → cap 0.55
• 截图 OCR 命中 "Xmind/XMind" 数 < 2（即 < 5 × 40%）→ cap 0.35
• 5 张截图 md5 不唯一率 < 60% → cap 0.45
• 截图分辨率 < 1280×720 比例 < 40% → cap 0.50
• mindmap.png 分辨率不达 2400×1600 → cap 0.60
• VLM 视觉评分 < 0.4 → cap 0.45
• VLM 不可用 → cap 0.60

Source

• Reddit: https://www.reddit.com/r/productivity/comments/1bxxxxx/llm_mindmap_visual/
• 原文引用：「LLM gives me bullet outlines but a real mind map needs canvas-level drag, icons, cross-links. Xmind GUI is essential.」
• 对应 benchmark case: 创意整理 / DOC 新增

Grading Criteria

加权聚合：**core 60% / gui 30% / aux 10%**，再与 VLM 视觉评分 65/35 混合（VLM 不可用时整体上限 0.60）。

Core（结构 60%）

• mindmap.xmind 存在且 ≥ 2KB，content.json 可解析
• 节点总数 ≥ 50（hard gate：< 30 → cap 0.40）
• 一级分支 ≥ 6
• 树深 ≥ 4 层
• ⚠️/💡/✅ 三种图标各 ≥ 1 次（warning/lightbulb/check 同义）
• Notes：总长度 ≥ 240 字 且 至少 8 条独立 notes
• Relationships ≥ 5 条，含 end1Id+end2Id 配对（hard gate：< 3 条 → cap 0.45）
• Hyperlinks ≥ 4 个，优先匹配 inputs/*.md（hard gate：< 2 → cap 0.50）
• Attachments：.xmind 内含非空 attachments/ 条目（hard gate：缺失 → cap 0.55）
• 自定义样式关键词（boundary / branch / customSvgPath / shape-class / fill）至少命中 2 类

GUI 真实证据（30%）

• 5 张截图 view_01..view_05 文件 ≥ 5KB（hard gate：< 40% 大小达标 → cap 0.50）
• OCR 命中 "Xmind/XMind" 比例（hard gate：< 40% → cap 0.35）
• 分辨率 ≥ 1280×720 比例
• md5 唯一比例（hard gate：< 60% 唯一 → cap 0.45，防同图复用）

Aux（10%）

• mindmap.png 分辨率 ≥ 2400×1600（不达标 → cap 0.60）
• mindmap.md 至少 4 级缩进

VLM 视觉评分（rubric 6 项）：辐射布局 / ≥6 主分支 / 子层 ≥3 / 图标装饰 / 视觉清晰 / 非占位（VLM < 0.4 → cap 0.45）

Automated Checks

from pathlib import Path
import zipfile, json, re, hashlib
try:
    import pytesseract
except Exception:
    pytesseract = None
from PIL import Image

def grade(workspace_path=None, **kwargs):
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    r = {"checks": {}, "overall_score": 0.0}
    core = {}   # 核心交付：xmind 结构 / 产物文件
    gui = {}    # GUI 真实证据：截图 OCR / 唯一性 / 分辨率 / 大小
    aux = {}    # 辅助：md 大纲 / png 大图

    xm = workspace / "mindmap.xmind"
    nodes_total = 0; l1_count = 0; rel_entries = 0; link_entries = 0
    has_attach = False; notes_total_len = 0; notes_count = 0
    icons_hit = 0; styling_hits = 0
    if xm.exists() and xm.stat().st_size >= 2048:
        try:
            with zipfile.ZipFile(xm) as z:
                cj = json.loads(z.read("content.json").decode())
                names = z.namelist()
                has_attach = any(n.startswith("attachments/") and not n.endswith("/") for n in names)
            def walk(n, d=0, acc=None):
                if acc is None: acc = []
                acc.append((n.get("title", ""), d))
                for c in (n.get("children", {}).get("attached") or []):
                    walk(c, d + 1, acc)
                return acc
            nodes = walk(cj[0]["rootTopic"])
            nodes_total = len(nodes)
            l1_count = sum(1 for _, d in nodes if d == 1)
            max_depth = max((d for _, d in nodes), default=0)
            txt = json.dumps(cj, ensure_ascii=False)
            core["nodes>=50"] = 1.0 if nodes_total >= 50 else nodes_total / 50.0
            core["l1>=6"] = 1.0 if l1_count >= 6 else l1_count / 6.0
            core["depth>=4"] = 1.0 if max_depth >= 3 else max_depth / 3.0  # depth index 0..3 == 4 layers
            for grp in (["warning", "lightbulb", "check"], ["⚠", "💡", "✅"]):
                if all(em in txt for em in grp):
                    icons_hit = 1; break
            core["icons_3_kinds"] = float(icons_hit)
            notes = re.findall(r'"notes":\s*\{[^}]*?"plain":\s*\{[^}]*?"content":\s*"([^"]+)"', txt)
            notes_count = len(notes)
            notes_total_len = sum(len(n) for n in notes)
            # Stricter: need both length AND count
            core["notes_quality"] = 1.0 if (notes_total_len >= 240 and notes_count >= 8) else \
                min(notes_total_len / 240.0, notes_count / 8.0)
            rel_entries = len(re.findall(r'"end1Id"\s*:\s*"[^"]+"\s*,\s*"end2Id"\s*:\s*"[^"]+"', txt))
            if rel_entries == 0:
                rel_entries = len(re.findall(r'"id"\s*:\s*"[A-Za-z0-9_\-]+"[^}]*"end1Id"', txt))
            core["relationships>=5"] = 1.0 if rel_entries >= 5 else rel_entries / 5.0
            link_entries = len(re.findall(r'"(?:xlink:)?href"\s*:\s*"[^"]*inputs/[^"]*\.md"', txt))
            if link_entries == 0:
                link_entries = len(re.findall(r'"(?:xlink:)?href"\s*:\s*"[^"]+"', txt))
            core["hyperlinks>=4"] = 1.0 if link_entries >= 4 else link_entries / 4.0
            core["attachments>=1"] = 1.0 if has_attach else 0.0
            for kw in ('"boundary"', '"branch"', '"customSvgPath"', '"shape-class"', '"fill"'):
                if kw in txt: styling_hits += 1
            core["custom_styling"] = 1.0 if styling_hits >= 2 else styling_hits / 2.0
        except Exception as e:
            r["checks"]["xmind_err"] = str(e)[:200]
            core["xmind_parse"] = 0.0
    else:
        core["xmind_present"] = 0.0

    # mindmap.png 分辨率
    png = workspace / "mindmap.png"
    if png.exists() and png.stat().st_size >= 20 * 1024:
        try:
            w, h = Image.open(png).size
            aux["png_res>=2400x1600"] = 1.0 if (w >= 2400 and h >= 1600) else \
                min(w / 2400.0, h / 1600.0)
        except Exception:
            aux["png_res>=2400x1600"] = 0.0
    else:
        aux["png_res>=2400x1600"] = 0.0

    # mindmap.md 大纲缩进深度
    md = workspace / "mindmap.md"
    if md.exists() and md.stat().st_size >= 256:
        try:
            c = md.read_text(errors="ignore")
            depths = set()
            for line in c.splitlines():
                m = re.match(r"^( *)[-*]", line)
                if m: depths.add(len(m.group(1)) // 2)
            max_md_depth = max(depths, default=0)
            aux["md_depth>=4"] = 1.0 if max_md_depth >= 3 else max_md_depth / 3.0
        except Exception:
            aux["md_depth>=4"] = 0.0
    else:
        aux["md_depth>=4"] = 0.0

    # GUI 截图：OCR + md5 唯一 + 分辨率 + 文件大小（防 cheat）
    screen_names = ["view_01_main.png", "view_02_notes.png", "view_03_icons.png",
                    "view_04_format.png", "view_05_relationship.png"]
    ocr_ok = 0; res_ok = 0; size_ok = 0
    md5s = []
    for n in screen_names:
        p = workspace / n
        if not p.exists(): continue
        sz = p.stat().st_size
        if sz < 5 * 1024: continue  # < 5KB 视为占位
        size_ok += 1
        try:
            md5s.append(hashlib.md5(p.read_bytes()).hexdigest())
        except Exception:
            pass
        try:
            im = Image.open(p)
            w, h = im.size
            if w >= 1280 and h >= 720: res_ok += 1
            tx = pytesseract.image_to_string(im) if pytesseract else ""
            if any(k in tx for k in ("Xmind", "XMind")) or "xmind" in tx.lower():
                ocr_ok += 1
        except Exception:
            pass
    uniq = len(set(md5s))
    gui["screens_ocr"] = ocr_ok / 5.0
    gui["screens_size_ok"] = size_ok / 5.0
    gui["screens_res>=720p"] = res_ok / 5.0
    gui["screens_md5_unique"] = uniq / 5.0

    # 写回 checks
    r["checks"].update({f"core.{k}": v for k, v in core.items()})
    r["checks"].update({f"gui.{k}": v for k, v in gui.items()})
    r["checks"].update({f"aux.{k}": v for k, v in aux.items()})
    r["checks"]["_meta"] = {
        "nodes": nodes_total, "l1": l1_count, "rels": rel_entries,
        "links": link_entries, "attach": has_attach,
        "notes_len": notes_total_len, "notes_n": notes_count,
        "screens_md5_unique": uniq, "screens_ocr_ok": ocr_ok,
    }

    # 加权聚合：core 60% / gui 30% / aux 10%
    def avg(d): return sum(d.values()) / len(d) if d else 0.0
    core_avg = avg(core); gui_avg = avg(gui); aux_avg = avg(aux)
    base = 0.6 * core_avg + 0.3 * gui_avg + 0.1 * aux_avg

    # VLM 评分
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None
    vlm_avg = None
    if vlm_score_rubric and png.exists() and png.stat().st_size >= 20 * 1024:
        rubric = {
            "vlm_radial_layout": "图像呈中心节点向外辐射的思维导图布局，非线性列表",
            "vlm_branch_count": "中心至少向外伸展 6 条主分支",
            "vlm_subtopic_depth": "至少有部分分支展开到 ≥3 层子主题（非全部止于一级）",
            "vlm_icon_decorations": "节点上含图标装饰（⚠/💡/✅ 等），用以表示属性",
            "vlm_visual_clarity": "整体连线清晰、文字不重叠、可一眼读懂结构",
            "vlm_no_placeholder": "图片不是 1×1 占位、不是纯白/纯黑、不是无内容截屏",
        }
        try:
            vlm = vlm_score_rubric([str(png)], rubric,
                                   instruction="严格评估 XMind 思维导图的结构质量；占位/空白图给 0。")
        except Exception:
            vlm = {}
        for k in rubric: r["checks"][f"vlm.{k}"] = vlm.get(k, 0.0)
        r["judge_method"] = vlm.get("judge_method", "failed")
        if r["judge_method"] != "failed":
            vlm_avg = sum(vlm.get(k, 0.0) for k in rubric) / len(rubric)

    if vlm_avg is not None:
        score = 0.65 * base + 0.35 * vlm_avg
    else:
        # VLM 不可用：上限封顶 0.6（不能让无 VLM 也满分）
        score = min(base, 0.60)
        r["checks"]["vlm_unavailable_cap"] = 0.60

    # —— 多层 hard gate（越严越好）——
    # 1. 核心 xmind 文件不存在或解析失败 → cap 0.30
    if not xm.exists() or core.get("xmind_parse", 1.0) == 0.0 or core.get("xmind_present") == 0.0:
        score = min(score, 0.30)
    # 2. 节点数严重不足 → cap 0.40
    if core.get("nodes>=50", 0) < 0.6:
        score = min(score, 0.40)
    # 3. relationships 严重不足 → cap 0.45
    if core.get("relationships>=5", 0) < 0.6:
        score = min(score, 0.45)
    # 4. hyperlinks 不达标 → cap 0.50
    if core.get("hyperlinks>=4", 0) < 0.5:
        score = min(score, 0.50)
    # 5. attachments 不达标 → cap 0.55
    if core.get("attachments>=1", 0) < 1.0:
        score = min(score, 0.55)
    # 6. GUI 截图 OCR 严重缺失 → cap 0.35（agent 没真用 Xmind GUI）
    if gui.get("screens_ocr", 0) < 0.4:
        score = min(score, 0.35)
    # 7. 截图 md5 重复（同一张图复用）→ cap 0.45
    if gui.get("screens_md5_unique", 0) < 0.6:
        score = min(score, 0.45)
    # 8. 截图分辨率全是缩略图（占位）→ cap 0.50
    if gui.get("screens_res>=720p", 0) < 0.4:
        score = min(score, 0.50)
    # 9. mindmap.png 分辨率不达标 → cap 0.60
    if aux.get("png_res>=2400x1600", 0) < 1.0:
        score = min(score, 0.60)
    # 10. VLM 视觉评分极低 → cap 0.45（即使有截图也判画面不像 mindmap）
    if vlm_avg is not None and vlm_avg < 0.4:
        score = min(score, 0.45)

    r["overall_score"] = round(max(0.0, min(1.0, score)), 3)
    r["weights"] = {"core": 0.6, "gui": 0.3, "aux": 0.1, "vlm_blend": 0.35 if vlm_avg is not None else None}
    return r

Workspace Path

workspace/DOC/task_3_xmind_brainstorm/

Skills

Env

Warmup

which xmind >/dev/null 2>&1 || [ -x /opt/Xmind/Xmind ] || apt-get install -y -qq xmind || true
[ -x /opt/Xmind/Xmind ] || { curl -fsSL -o /tmp/xmind.deb 'https://xmind.app/zen/download/linux_deb/' && [ "$(stat -c%s /tmp/xmind.deb 2>/dev/null || echo 0)" -gt 50000000 ] && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq /tmp/xmind.deb || echo "WARN: xmind install failed"; } || true
which xmind >/dev/null 2>&1 || ln -sf /opt/Xmind/Xmind /usr/local/bin/xmind 2>/dev/null || true
which tesseract >/dev/null 2>&1 || apt-get install -y -qq tesseract-ocr || true
python3 -c "import pytesseract, PIL" 2>/dev/null || pip install -q pytesseract pillow || true