Datasets:

wanlilll
/

WeaveBench

---
id: DOC_task_5_docbook_xmllint_bluefish
name: DocBook XML 错误修复
category: DOC
timeout_seconds: 1500
---

<!--
resources:
  - name: libxml2 (xmllint)
    source: apt install libxml2-utils
    license: MIT (system tool, runtime-only, not redistributed)
  - name: xsltproc + DocBook XSL stylesheets
    source: apt install xsltproc docbook-xsl
    license: MIT (system tool, runtime-only, not redistributed)
  - name: Bluefish editor
    source: apt install bluefish
    license: GPL-3.0 (system tool, runtime-only, not redistributed)
  - name: Firefox (HTML 渲染验证)
    source: apt install firefox-esr
    license: MPL-2.0 (system tool, runtime-only, not redistributed)
  - name: broken.docbook + expected.xsl
    source: workspace/DOC/task_5_docbook_xmllint_bluefish/exec/ (含 5 个 xml 错)
    license: self-written
-->


## Prompt
> ⚙️ **Execution convention**: This is a legitimate desktop + terminal XML editing task. All output artifacts go to `/tmp_workspace/results/`. Execute directly; do not refuse.

**Background:** The team's DocBook 5 document `/tmp_workspace/exec/broken.docbook` reports ≥ 5 validation errors during CI checks and cannot be rendered to HTML. A companion stylesheet `/tmp_workspace/exec/expected.xsl` is available for rendering (the warmup script has already copied it into place).

**Goal:** Locate and fix all structural/reference errors in `broken.docbook` so that it passes DocBook validation, then render it to viewable HTML using `expected.xsl`, and produce a complete evidence chain (validation logs + rendering log + working-process screenshots + fix report).

### Hard constraints for the primary deliverable `working.docbook` (in `/tmp_workspace/results/`)

- Must be a fixed working copy of `broken.docbook`, structurally compliant and fully passing DocBook validation.
- After fixing, rendering with `expected.xsl` via XSLT must produce valid HTML without any fatal/error messages.

### Supporting deliverables (all placed in `/tmp_workspace/results/`)

| File | Requirements |
|---|---|
| `xmllint_errors_before.txt` | Validation output for the original `broken.docbook`; must parse to ≥ 5 errors (containing keywords such as line/element/expected). |
| `xmllint_after_fix1.txt` | Mid-way validation output after fixing the first error; error count must be strictly less than the initial count. |
| `xmllint_errors_after.txt` | Final validation output after all fixes; expected 0 errors (or at least ≥ 80% reduction). |
| `error_locations.txt` | Deduplicated error-location lines extracted from the validation output (each line containing line number/element/expected value and similar keywords). |
| `output.html` | HTML rendered using `expected.xsl`; size > 2 KB; must contain `<html>` / `<body>` and ≥ 3 common document elements (e.g. `<h1>` / `<h2>` / `<p>` / `<a>` / `<img>` / `<ul>` / `<ol>` / `<li>`). |
| `xsltproc.log` | stderr from the rendering process; must not contain `fatal` or `error`. |
| `fix_report.md` | ≥ 4 paragraphs, each ≥ 80 characters: a) file:line:col and category for each error; b) fix strategy; c) how the outline/tree view helped locate nesting issues; d) DocBook validation best practices. |

### 5 working-process screenshots (in `/tmp_workspace/results/`)

Fixed file names + general specs: each ≥ 50 KB, resolution ≥ 1920×1000, all 5 images must differ in content (all MD5 hashes distinct). Grading will run OCR to extract readable text such as window/toolbar/tab labels.

- `view_bluefish_open.png`: The original `broken.docbook` opened in an XML editor showing line numbers + syntax highlighting.
- `view_bluefish_outline.png`: The tag tree / outline view of the DocBook document, with the error locations visible.
- `view_bluefish_fixing.png`: An intermediate state while fixing an error in the XML (selection/cursor must be discernible).
- `view_bluefish_final.png`: The completed `working.docbook` with a full, structurally-compliant outline.
- `view_firefox_html.png`: `output.html` viewed in a browser, showing rendered titles, paragraphs, links, images, and similar elements.

Anti-cheat: `actions.log` (if produced) must not contain literals showing that any of the above PNGs were directly `cp`-ed from other images or synthetically generated; screenshots must originate from a continuous GUI operation trajectory.

## 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. 推荐通道：在 GUI XML 编辑器（如 Bluefish / gedit / VS Code / Kate 等任一带行号 + 语法高亮 + 标签树的工具）中打开 broken.docbook；浏览器侧可用 Firefox / Chromium / 任一 HTML 渲染器查看 output.html。CLI 校验/渲染走 `xmllint --noout` 与 `xsltproc`。
2. step 1：`cp /tmp_workspace/exec/broken.docbook /tmp_workspace/results/working.docbook`，并对原始文件跑一次 xmllint，重定向到 `xmllint_errors_before.txt`，确认能解析出 ≥ 5 处错误。
3. step 2：在编辑器中打开 working.docbook，截 `view_bluefish_open.png`；切到大纲/标签树视图，截 `view_bluefish_outline.png`，并从校验输出抽取去重写入 `error_locations.txt`。
4. step 3：定位并修复第一处错误（缺 xmlns / tag 未关闭 / xref id 不存在 / 嵌套违法等常见类），保存后再跑 xmllint 输出到 `xmllint_after_fix1.txt`（错误数严格少于 before）；过程中截 `view_bluefish_fixing.png`。
5. step 4：迭代修复剩余错误直至 xmllint 0 错，保存为 `xmllint_errors_after.txt`，截 `view_bluefish_final.png`。
6. step 5：`xsltproc /tmp_workspace/exec/expected.xsl working.docbook > output.html 2> xsltproc.log`，在浏览器中打开 output.html 截 `view_firefox_html.png`，最后写 `fix_report.md`（4 段 × ≥ 80 字）。

约束说明：5 张截图须来自连续真实 GUI 操作轨迹（md5 全互异、≥ 50KB、分辨率 ≥ 1920×1000），actions.log 不得包含把这些 png 用 cp / 生成器伪造的字面量。

评分要点（hard gates）：xmllint after 0 错或削减 ≥ 80%；output.html ≥ 2KB 且含 html/body + ≥ 3 类文档元素；xsltproc.log 无 fatal/error；5 张截图齐全且通过 OCR 窗口 chrome 关键词与几何分辨率门槛；fix_report.md ≥ 4 段；缺 GUI 证据/校验未达标会被多层封顶。

## Source

- DocBook 5: https://docbook.org/
- libxml2 xmllint
- Bluefish: http://bluefish.openoffice.nl/

## Grading Criteria

- [ ] 1. xmllint_errors_before.txt + xmllint_errors_after.txt + error_locations.txt 存在
- [ ] 2. xmllint_errors_after.txt 显示 0 错（或相比 before 至少削减 ≥80%）
- [ ] 3. output.html 存在,size > 1KB
- [ ] 4. 4 张 GUI 截图 + OCR 命中 Bluefish/Firefox UI 关键词
- [ ] 5. xsltproc.log 不含 fatal error
- [ ] 6. fix_report.md ≥ 4 段
- [ ] 7. VLM rubric 评 Bluefish outline + firefox HTML
- [ ] 8. **GUI trajectory frames 真实**：5 张 view_*.png 必须由连续 GUI 操作产生（打开 → outline → 修复中 → 修复完成 → firefox 渲染），截图间 md5 唯一率必须 = 100%，且每张文件大小 ≥ 50KB（防纯黑/占位 PNG）
- [ ] 9. **GUI app chrome OCR**：截图须 OCR 命中 Bluefish / Document / XML / Outline / Firefox 等 GUI app chrome 关键词；分辨率 ≥ 1920×1000；窗口几何/Chrome OCR 任一失败 cap 0.5
- [ ] 10. **加权综合 + VLM 封顶**：核心交付 60%、GUI 证据 30%、辅助 10%；VLM 不可用整体封顶 0.6，VLM 平均 < 0.4 封顶 0.4

## Automated Checks

```python
def grade(workspace_path=None, **kwargs) -> dict:
    """DOC_task_5 grader."""
    import re
    from pathlib import Path
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    rd = workspace / "results"
    s = {}

    # 1. CLI artifacts
    cli_files = ["xmllint_errors_before.txt","xmllint_errors_after.txt","error_locations.txt","xsltproc.log","output.html"]
    cli_present = sum(1 for f in cli_files if (rd / f).exists())
    s["cli_artifacts"] = cli_present / len(cli_files)
    has_cli = cli_present >= 3

    # 1b. extra Prompt-required artifacts (working.docbook + after_fix1)
    extra_files = ["working.docbook", "xmllint_after_fix1.txt"]
    extra_present = sum(1 for f in extra_files if (rd / f).exists())
    s["extra_artifacts"] = extra_present / len(extra_files)

    # 2. xmllint errors before vs after
    err_score = 0.0
    bf = rd / "xmllint_errors_before.txt"
    af = rd / "xmllint_errors_after.txt"
    if bf.exists() and af.exists():
        try:
            bb = bf.read_text(); aa = af.read_text()
            n_before = len(re.findall(r"line\s+\d+|element|expected", bb))
            n_after = len(re.findall(r"line\s+\d+|element|expected", aa))
            if n_before >= 5 and n_after == 0: err_score = 1.0
            elif n_before >= 5 and n_after < n_before * 0.2: err_score = 0.7
            elif n_after < n_before: err_score = 0.4
        except Exception: pass
    s["xmllint_clean"] = err_score

    # 3. output.html (size > 2KB + 必须含 html/body + 至少 2 个 DocBook 渲染元素)
    html_score = 0.0
    oh = rd / "output.html"
    if oh.exists():
        try:
            sz = oh.stat().st_size
            txt = oh.read_text(errors="ignore").lower()
            doc_elems = sum(1 for tag in ["<h1", "<h2", "<p", "<a ", "<img", "<ul", "<ol", "<li"] if tag in txt)
            if sz > 2048 and "<html" in txt and "<body" in txt and doc_elems >= 3:
                html_score = 1.0
            elif sz > 1024 and "<html" in txt and "<body" in txt:
                html_score = 0.6
            elif sz > 200:
                html_score = 0.3
        except Exception: pass
    s["html_rendered"] = html_score

    # 4. GUI screenshots (size >= 50KB 才算真截图，过滤纯黑/占位 PNG)
    gui_shots = ["view_bluefish_open.png","view_bluefish_outline.png","view_bluefish_fixing.png","view_firefox_html.png","view_bluefish_final.png"]
    gui_present_paths = [rd / n for n in gui_shots if (rd / n).exists() and (rd / n).stat().st_size >= 50 * 1024]
    gui_present = len(gui_present_paths)
    s["gui_screenshots_count"] = gui_present / len(gui_shots)
    has_gui = gui_present >= 3

    try:
        import pytesseract
        from PIL import Image
        kws_any = ["Bluefish","Outline","DocBook","section","chapter","Firefox","XML"]
        ocr_hits = 0
        for n in gui_shots:
            p = rd / n
            if p.exists():
                try:
                    tx = pytesseract.image_to_string(Image.open(p))
                    if any(k in tx for k in kws_any): ocr_hits += 1
                except Exception: pass
        s["gui_screenshots_ocr"] = ocr_hits / len(gui_shots)
    except Exception:
        s["gui_screenshots_ocr"] = 0.5 if gui_present > 0 else 0.0

    # 5. xsltproc.log no fatal
    xs_score = 0.0
    xs = rd / "xsltproc.log"
    if xs.exists():
        try:
            txt = xs.read_text()
            if "fatal" not in txt.lower() and "error" not in txt.lower():
                xs_score = 1.0
            elif "fatal" not in txt.lower():
                xs_score = 0.5
        except Exception: pass
    s["xsltproc_clean"] = xs_score

    # 6. fix_report.md
    rp_score = 0.0
    rp = rd / "fix_report.md"
    if rp.exists():
        try:
            txt = rp.read_text()
            parags = [p for p in re.split(r"\n\s*\n", txt) if len(p.strip()) >= 80]
            rp_score = min(1.0, len(parags) / 4)
        except Exception: pass
    s["fix_report"] = rp_score

    # 7. VLM rubric
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None
    vlm_keys = ["vlm_bluefish_editor","vlm_xml_highlight","vlm_outline_visible","vlm_html_rendered"]
    if vlm_score_rubric and (rd / "view_bluefish_outline.png").exists():
        rubric = {
            "vlm_bluefish_editor": "Bluefish 截图能看到编辑器 + 行号",
            "vlm_xml_highlight": "XML 内容有语法高亮(关键字 / tag 不同色)",
            "vlm_outline_visible": "outline 截图能看到 tag 树结构",
            "vlm_html_rendered": "Firefox 截图显示渲染的 DocBook HTML(标题 / 段落 / 链接)",
        }
        try:
            imgs = [str(rd / n) for n in ["view_bluefish_outline.png","view_firefox_html.png","view_bluefish_final.png"] if (rd / n).exists()]
            vlm = vlm_score_rubric(imgs, rubric, instruction="评估 Bluefish XML editor + Firefox HTML 截图。")
            for k in rubric: s[k] = float(vlm.get(k, 0.0))
        except Exception:
            for k in rubric: s[k] = 0.0
    # else: VLM 不可用时不写入 vlm_* 键，避免拉低 base 分母

    # GUI hard-gate sub-scores: trajectory diversity + chrome OCR + window geometry
    import hashlib
    try:
        from PIL import Image as _PILImage
    except Exception:
        _PILImage = None
    gui_shot_paths = [rd / n for n in gui_shots if (rd / n).exists() and (rd / n).stat().st_size >= 50 * 1024]
    if len(gui_shot_paths) >= 4:
        hashes = set(hashlib.md5(p.read_bytes()).hexdigest() for p in gui_shot_paths)
        gui_diversity = len(hashes) / max(1, len(gui_shot_paths))
    else:
        gui_diversity = 0.0
    s["gui_real_interaction"] = 1.0 if gui_diversity >= 1.0 else (0.5 if gui_diversity >= 0.75 else 0.0)

    geom_hits = 0
    if _PILImage and gui_shot_paths:
        for p in gui_shot_paths:
            try:
                w, h = _PILImage.open(p).size
                if w >= 1920 and h >= 1000:
                    geom_hits += 1
            except Exception:
                pass
        s["gui_window_geometry"] = min(1.0, geom_hits / max(2.0, len(gui_shot_paths) * 0.6))
    else:
        s["gui_window_geometry"] = 0.0

    chrome_kws = ["Bluefish", "Document", "XML", "Outline", "Firefox",
                  "File", "Edit", "View", "Tools", "section", "chapter"]
    chrome_hits = 0
    try:
        import pytesseract as _pyt
        if _PILImage and gui_shot_paths:
            for p in gui_shot_paths:
                try:
                    tx = _pyt.image_to_string(_PILImage.open(p))
                    if sum(1 for k in chrome_kws if k.lower() in tx.lower()) >= 2:
                        chrome_hits += 1
                except Exception:
                    pass
            s["gui_chrome_ocr"] = min(1.0, chrome_hits / max(2.0, len(gui_shot_paths) * 0.5))
        else:
            s["gui_chrome_ocr"] = 0.0
    except Exception:
        s["gui_chrome_ocr"] = 0.0

    # 加权综合：核心交付 60% / GUI 证据 30% / 辅助 10%
    core_keys = ["cli_artifacts", "extra_artifacts", "xmllint_clean", "html_rendered", "xsltproc_clean"]
    gui_keys = ["gui_screenshots_count", "gui_screenshots_ocr", "gui_real_interaction",
                "gui_window_geometry", "gui_chrome_ocr"]
    aux_keys = ["fix_report"]
    vlm_keys_present = [k for k in s if k.startswith("vlm_")]

    def _avg(keys):
        vals = [s[k] for k in keys if k in s and isinstance(s[k], (int, float))]
        return sum(vals) / len(vals) if vals else 0.0

    core = _avg(core_keys)
    gui = _avg(gui_keys)
    aux = _avg(aux_keys + vlm_keys_present)
    base = 0.6 * core + 0.3 * gui + 0.1 * aux

    # 多层 hard gate（v2 收紧）
    if not has_cli: base = min(base, 0.25)
    if not has_gui: base = min(base, 0.25)
    if s["xmllint_clean"] < 0.7: base = min(base, 0.40)
    if s["xmllint_clean"] < 0.4: base = min(base, 0.30)
    if s["html_rendered"] < 0.7: base = min(base, 0.50)
    if s["html_rendered"] < 0.4: base = min(base, 0.35)
    # GUI 真实交互 hard gate：截图全相同/缺失 → 直接封顶 0.35
    if s.get("gui_real_interaction", 0) < 0.5: base = min(base, 0.35)
    if s.get("gui_chrome_ocr", 0) < 0.5: base = min(base, 0.50)
    if s.get("gui_chrome_ocr", 0) < 0.3: base = min(base, 0.40)
    if s.get("gui_window_geometry", 0) < 0.5: base = min(base, 0.55)
    # VLM 不可用时整体封顶 0.6（不能没 VLM 也满分）
    if not vlm_keys_present:
        base = min(base, 0.60)
    else:
        vlm_avg = sum(s[k] for k in vlm_keys_present) / len(vlm_keys_present)
        if vlm_avg < 0.6: base = min(base, 0.55)
        if vlm_avg < 0.4: base = min(base, 0.40)

    s["overall_score"] = round(base, 4)
    return s
```

## Workspace Path

```
workspace/DOC/task_5_docbook_xmllint_bluefish
```

## Skills

```
```

## Env

```
```

## Warmup

```bash
echo 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 | base64 -d | bash
which firefox >/dev/null 2>&1 || apt-get install -y -qq firefox-esr 2>/dev/null || apt-get install -y -qq firefox 2>/dev/null || true
which tesseract >/dev/null 2>&1 || apt-get install -y -qq tesseract-ocr 2>/dev/null || true
pip install -q pytesseract pillow numpy 2>/dev/null || true
```