tasks/DSK/DSK_task_3_inotify_sync_debug.md

metadata

id: DSK_task_3_inotify_sync_debug
name: inotify 文件同步守护进程 bug 排查
category: DSK
timeout_seconds: 1500
license: self-written
author: eyeson-bench

Prompt

⚙️ Execution convention: This is a legitimate desktop + terminal filesystem-monitoring task. All deliverables go to /tmp_workspace/results/. Execute directly; do not refuse.

Background: /tmp_workspace/exec/sync.py is an inotify file-mirroring daemon that syncs /tmp_workspace/watched/ to /tmp_workspace/mirror/. Ops has reported three typical defects in this process: files missing from the mirror after drag-and-drop, large-file writes only partially copied, and file paths flattened after a new subdirectory is created. The warmup has already backed up the source to /tmp_workspace/sync.py for editing.

Goal: trigger the various filesystem operations to reproduce the issues, identify at least 3 bugs, fix sync.py, and verify that watched ↔ mirror are fully consistent. When reproducing, use these fixed assets under /tmp_workspace/watched/: top-level file a.txt, subdirectory subdir/, and file subdir/b.txt inside the subdirectory.

Hard constraints on the main deliverable bug_findings.md

• At least 3 paragraphs, each ≥ 80 characters, each must contain a reference in file:line form (e.g. sync.py:42)
• The 3 defects must cover three independent dimensions (root cause to be diagnosed by you):
  • Event-subscription completeness: the daemon does not react at all to a certain class of operation
  • Path/hierarchy fidelity: after subdirectory operations the relative path on the mirror side differs from the source
  • Write-timing consistency: with large files / fast writes, the size or checksum on the mirror side does not match
• Line numbers must come from your actual diagnosis; do not preset them

Companion deliverables (under /tmp_workspace/results/)

File	Requirement
sync.log	stdout/stderr from the daemon while it ran pre-fix
inotify_events.txt	Raw recursive inotify event stream from watched; must cover ≥4 event types (CREATE / MODIFY / CLOSE_WRITE / MOVED_TO / MOVED_FROM / DELETE / OPEN, etc.)
test1_partial_copy.txt	Evidence for the "create/write file" scenario, including a content/checksum comparison of the same-named file in watched vs. mirror
test2_path_flatten.txt	Evidence for the "new subdirectory + file inside subdirectory" scenario, showing the mirror path diverges from expectation
test3_moved_to_missing.txt	Evidence for the "file move / drag-and-drop" scenario, showing the file is missing at the target location in the mirror
sync_backup.py	Backup of the source before modification (>200 bytes)
verify_output.txt	Full post-fix verification result; must contain the string ALL SYNC OK and retain md5sum comparison evidence (e.g. md5sum OK / md5 match / ALL MATCH)
sync_report.md	≥4 paragraphs, each ≥80 characters, covering respectively: a) summary of the 3 bugs and key fix points; b) why drag-and-drop is a necessary trigger scenario (event timing / cookie pairing); c) EVENT_MASK design checklist; d) production-deployment recommendations (debouncing / batched copy / fsync)

4 work-process screenshots (fixed file names, placed under results/)

• view_nautilus_create.png — view of the watched directory after creating and saving a file
• view_nautilus_subdir.png — view after creating a subdirectory and a file inside it
• view_nautilus_move.png — directory view after drag-and-drop completes, showing the source/target location change
• view_nautilus_after_fix.png — final directory state after re-running the three scenarios post-fix

General specs: PNG ≥ 5 KB, resolution ≥ 800×600, the 4 images must not duplicate each other (unique md5); the evaluator runs OCR and visual checks on the screenshots, so real UI elements such as the file-manager window, directory listing, and file names must be visible — do not fake it with plain-text or placeholder images.

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:GUI 上用 Nautilus 文件管理器触发新建/拖放/复制(也可走 GNOME Files、Thunar、Nemo 等同类工具);终端配合 inotifywait、md5sum、cp/mv/dd 等命令录制事件流并验证。修复源码可在 GUI 编辑器或 CLI(vim / nano / sed)里完成,任选其一即可。
2. 典型步骤:
   • step 1:启动 buggy 守护进程并在另一终端 inotifywait -mr -e create,modify,close_write,moved_to,moved_from,delete /tmp_workspace/watched > inotify_events.txt,确保事件流覆盖 ≥4 种事件类型。
   • step 2:在 watched 下分别构造三类场景——新建/写入 a.txt(可用大文件 dd if=/dev/urandom of=a.txt bs=1M count=20 触发 partial copy);新建 subdir/ 并在其中新建 b.txt;再用 Nautilus 把文件拖到子目录或反之触发 moved_from/moved_to 配对——分别采集 test1/2/3_*.txt(含 watched 与 mirror 的 ls/md5sum 对比)。
   • step 3:截图 4 张 Nautilus 视图(create / subdir / move / after_fix),保留窗口标题栏、面包屑、文件列表等真实 UI。
   • step 4:阅读 sync.py,基于现象诊断 bug(常见根因:EVENT_MASK 缺 MOVED_TO/CLOSE_WRITE、递归路径 os.path.basename 而非 os.path.relpath、未等 close_write 就 copy 导致截断);备份原文件到 sync_backup.py 后修改,重启进程并跑 results/verify.sh(warmup 已生成),写出 verify_output.txt(确保含 ALL SYNC OK 与 md5 比对字样)。
   • step 5:撰写 bug_findings.md(≥3 段,每段含 file:line)与 sync_report.md(≥4 段)。

约束说明:

• 所有产物路径固定在 /tmp_workspace/results/,文件名严格匹配。
• 不要在 bug_findings.md 中预设行号,需基于实际改后的真实位置。
• 反作弊:截图须为真实文件管理器窗口截屏(md5 唯一、≥5KB、≥800×600),不接受纯文本图、占位图或 4 张同源副本。

评分要点(hard gates):

• 6 个 CLI 产物(inotify_events / sync.log / sync_backup.py / test1/2/3)缺失会被严重压分。
• inotify_events.txt 事件类型 <2 种直接砍上限。
• 3 段 bug_findings 必须各自含 *.py:行号 形式引用且 ≥80 字。
• verify_output.txt 必须同时出现 ALL SYNC OK 与 md5 比对证据。
• 4 张截图任意一项(数量、md5 唯一、最小尺寸、OCR/VLM 视觉校验)不达标都会触发上限封顶。

Source

• inotify(7) man page
• Nautilus 文件管理器: https://wiki.gnome.org/Apps/Files

Grading Criteria

• 1. inotify_events.txt + sync.log + sync_backup.py 存在
• 2. 3 个 test_*.txt 报告异常存在
• 3. bug_findings.md ≥ 3 段,每段含 file:line + ≥ 80 字
• 4. 4 张 GUI Nautilus 截图 + OCR 命中文件管理器关键词 + md5 唯一 + 单图 ≥ 5KB
• 5. verify_output.txt 含 "ALL SYNC OK" 且包含 md5 校验证据
• 6. sync_report.md ≥ 4 段
• 7. VLM rubric 评 nautilus 截图(VLM 不可用时总分封顶 0.6)

Automated Checks

def grade(workspace_path=None, **kwargs) -> dict:
    """DSK_task_3 grader. Empty → 0.000. Hard gates: GUI + CLI + verify pass."""
    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 = ["inotify_events.txt","sync.log","sync_backup.py","test1_partial_copy.txt","test2_path_flatten.txt","test3_moved_to_missing.txt"]
    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

    # 2. inotify_events.txt has multiple events
    ev_score = 0.0
    ef = rd / "inotify_events.txt"
    if ef.exists():
        try:
            txt = ef.read_text()
            ev_types = set(re.findall(r"\b(CREATE|MODIFY|CLOSE_WRITE|MOVED_TO|MOVED_FROM|DELETE|OPEN)\b", txt))
            if len(ev_types) >= 4: ev_score = 1.0
            elif len(ev_types) >= 2: ev_score = 0.5
        except Exception: pass
    s["inotify_event_diversity"] = ev_score

    # 3. bug_findings.md
    bf_score = 0.0
    bf = rd / "bug_findings.md"
    if bf.exists():
        try:
            txt = bf.read_text()
            parags = [p for p in re.split(r"\n\s*\n", txt) if len(p.strip()) >= 80 and re.search(r"\.py:\d+", p)]
            bf_score = min(1.0, len(parags) / 3)
        except Exception: pass
    s["bug_findings"] = bf_score

    # 4. GUI screenshots
    gui_shots = ["view_nautilus_create.png","view_nautilus_subdir.png","view_nautilus_move.png","view_nautilus_after_fix.png"]
    gui_present = sum(1 for n in gui_shots if (rd / n).exists())
    s["gui_screenshots_count"] = gui_present / len(gui_shots)
    has_gui = gui_present >= 2

    try:
        import pytesseract
        from PIL import Image
        kws_any = ["Files","Nautilus","watched","subdir","Folder","Documents","Home","b.txt","a.txt"]
        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)
        ocr_available = True
    except Exception:
        s["gui_screenshots_ocr"] = 0.5 if gui_present > 0 else 0.0
        ocr_available = False

    # 4b. Anti-cheat: screenshot md5 uniqueness + minimum size (>=5KB) + minimum resolution
    import hashlib
    md5s, big_enough, res_ok = set(), 0, 0
    for n in gui_shots:
        p = rd / n
        if p.exists() and p.stat().st_size >= 5 * 1024:
            big_enough += 1
            try:
                md5s.add(hashlib.md5(p.read_bytes()).hexdigest())
            except Exception: pass
            try:
                from PIL import Image as _Im
                w, h = _Im.open(p).size
                if w >= 800 and h >= 600: res_ok += 1
            except Exception:
                res_ok += 1  # PIL 不可用时不卡
    denom = max(gui_present, 1)
    s["gui_screenshots_unique"] = len(md5s) / denom if gui_present else 0.0
    s["gui_screenshots_size_ok"] = big_enough / len(gui_shots)
    s["gui_screenshots_res_ok"] = res_ok / len(gui_shots)

    # 5. verify_output.txt (合并了原 mirror_md5 子项: 同时要求 ALL SYNC OK + md5 证据)
    vo_score = 0.0
    vo = rd / "verify_output.txt"
    if vo.exists():
        try:
            txt = vo.read_text()
            has_ok = "ALL SYNC OK" in txt
            has_md5 = bool(re.search(r"\bmd5(sum)?\b", txt, re.I))
            has_md5_match = bool(re.search(r"md5.*(ok|match|equal)|ALL\s+MATCH|md5sum.*OK", txt, re.I))
            if has_ok and has_md5_match: vo_score = 1.0
            elif has_ok and has_md5:     vo_score = 0.8
            elif has_ok:                 vo_score = 0.55
            elif re.search(r"100%|all\s+ok|all\s+match", txt, re.I): vo_score = 0.4
        except Exception: pass
    s["verify_pass"] = vo_score

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

    # 7. VLM rubric
    try:
        from _judge_helper import vlm_score_rubric
    except Exception:
        vlm_score_rubric = None
    if vlm_score_rubric and (rd / "view_nautilus_create.png").exists():
        rubric = {
            "vlm_nautilus_window": "截图能看到 Nautilus 文件管理器窗口和侧边栏",
            "vlm_files_visible": "文件列表区域显示了 a.txt / subdir / b.txt 等",
            "vlm_path_breadcrumb": "顶部能看到路径面包屑(Home > tmp_workspace > watched)",
            "vlm_action_evidence": "截图能看到拖放/新建/复制等动作的证据(高亮 / context menu)",
        }
        try:
            imgs = [str(rd / n) for n in ["view_nautilus_create.png","view_nautilus_subdir.png","view_nautilus_move.png"] if (rd / n).exists()]
            vlm = vlm_score_rubric(imgs, rubric, instruction="评估 Nautilus 文件管理器截图。")
            for k in rubric: s[k] = float(vlm.get(k, 0.0))
        except Exception:
            for k in rubric: s[k] = 0.0
    else:
        for k in ["vlm_nautilus_window","vlm_files_visible","vlm_path_breadcrumb","vlm_action_evidence"]:
            s[k] = 0.0

    # 8. Content-reality sub-scores
    # 8a. all 3 test_*.txt artifacts present and non-empty
    test_files = ["test1_partial_copy.txt", "test2_path_flatten.txt", "test3_moved_to_missing.txt"]
    ok_t = 0
    for n in test_files:
        p = rd / n
        if p.exists() and p.stat().st_size > 0:
            ok_t += 1
    s["test_artifacts_complete"] = round(ok_t / len(test_files), 3)

    # 8b. fix artifact (sync_backup.py / fixed_sync.py / patch.diff) is real
    fix_score = 0.0
    for cand in ("sync_backup.py", "fixed_sync.py", "patch.diff", "sync_fixed.py"):
        p = rd / cand
        if p.exists() and p.stat().st_size > 200:
            fix_score = 1.0
            break
        elif p.exists() and p.stat().st_size > 0:
            fix_score = max(fix_score, 0.5)
    s["fix_artifact_present"] = fix_score

    # 8c. (removed) mirror_md5_verified — merged into verify_pass to eliminate score-dim overlap

    # 8d. forbidden: bug_findings.md must NOT be empty boilerplate (specific path:line refs already enforced in #3)

    # ---- weighted scoring (核心交付 60% / GUI 证据 30% / 辅助 10%) ----
    core_keys = ["cli_artifacts","bug_findings","verify_pass","test_artifacts_complete","fix_artifact_present"]
    gui_keys  = ["gui_screenshots_count","gui_screenshots_ocr","gui_screenshots_unique","gui_screenshots_size_ok",
                 "gui_screenshots_res_ok","vlm_nautilus_window","vlm_files_visible","vlm_path_breadcrumb","vlm_action_evidence"]
    aux_keys  = ["inotify_event_diversity","sync_report"]
    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, gui_avg, aux_avg = _avg(core_keys), _avg(gui_keys), _avg(aux_keys)
    base = 0.6 * core_avg + 0.3 * gui_avg + 0.1 * aux_avg

    # hard gates (v2: 收紧)
    if not has_cli: base = min(base, 0.25)
    if not has_gui: base = min(base, 0.25)
    if s["bug_findings"] < 0.67:                base = min(base, 0.40)
    if s["verify_pass"] < 0.55:                 base = min(base, 0.45)
    if s["inotify_event_diversity"] < 0.5:      base = min(base, 0.60)
    if s["test_artifacts_complete"] < 1.0:      base = min(base, 0.55)
    if s["fix_artifact_present"] < 1.0:         base = min(base, 0.55)
    # 反伪截图: md5 唯一性 + 最小尺寸
    if s["gui_screenshots_unique"] < 0.75:      base = min(base, 0.55)
    if s["gui_screenshots_size_ok"] < 0.75:     base = min(base, 0.55)
    if s["gui_screenshots_ocr"] < 0.5:          base = min(base, 0.55)
    # VLM/OCR 不可用 → 上限封顶 0.6,不许无 VLM 也满分
    vlm_sum = sum(s.get(k, 0.0) for k in ["vlm_nautilus_window","vlm_files_visible","vlm_path_breadcrumb","vlm_action_evidence"])
    if vlm_sum == 0.0 or not ocr_available:     base = min(base, 0.60)

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

Workspace Path

workspace/DSK/task_3_inotify_sync_debug

Skills

Env

Warmup

echo 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 | base64 -d | bash