tasks/DSK/DSK_task_1_dup_triage.md

metadata

id: DSK_task_1_dup_triage
name: Nautilus 文件冲突批量解决
category: DSK
timeout_seconds: 1200

Prompt

Copy /tmp_workspace/source/ (contains 135 files, of which 45 filenames collide with destination) into /tmp_workspace/dest/ (which already has 95 files, 45 of which will collide with source by name). The file manager will pop up a conflict dialog for each duplicate filename during the copy.

Task

Handle conflicts one by one according to the following policy (4 decision classes total):

• If source and dest files have mtime difference ≤ 1 day and identical size → skip (skip);
• If filenames are identical but contents are in entirely different domains (e.g. meeting notes on one side, recipe on the other) → keep both (keep_both, the new file gets a (copy) suffix);
• If the dst file has a newer mtime but content is obviously truncated (placeholder) → skip and preserve the older but more complete dst (skip_older);
• Otherwise → replace and preserve the old version (replace + old file gets a .bak.<unix_ts> suffix, i.e. use rename to first rename the old dest file and then overwrite).

Specific scale:

• source/ contains 135 files, dest/ contains 95, with 45 pairs of name collisions.
• Of these, 10 pairs are "same name, different domain" — identical filenames but entirely different contents (meeting notes on one side, recipe on the other), and must be keep_both.
• Of these, 5 pairs are "newer mtime but worse content" — dst is a placeholder (newer mtime), src is 30 days old but more complete; these must be skip_older (i.e. preserve the older but better dst and discard src's fake "newer" version). Looking at mtime alone will misclassify all of them.

Deliverables (write to /tmp_workspace/results/)

• conflict_log.csv: columns filename,decision(skip|replace|keep_both|skip_older),reason.
• final_inventory.csv: every file under the dest directory + size + mtime + sha256.
• conflict_<n>.png: screenshot of the dialog / operation for each conflict decision, at least 15 images.
• Final /tmp_workspace/dest/ file count ≥ 150 and no original important dest file is lost (the list of important files is in keep_list.txt).
• /tmp_workspace/dest/ must contain ≥ 10 newly added files with the .bak.<unix_ts> suffix (this format is used when renaming old dest files).

Expected Behavior

参考解题流程 (设计者参考, 不发给 agent):

1. 用 Nautilus 打开两个窗口 (source / dest), 全选 source 的 135 个文件, 拖拽到 dest。
2. 对弹出的 "Conflict" 对话框, 按以下策略逐项处理 (共 4 类 decision):
   • 如果 source 与 dest 文件 mtime 差 ≤ 1 天且 size 相同 → 跳过 (skip);
   • 如果文件名相同但内容属于完全不同领域 (例如一边是 meeting notes, 一边是 recipe) → 保留双方 (keep_both, Nautilus 会把新文件加 (copy) 后缀);
   • 如果 dst 文件 mtime 较新但内容明显残缺 (placeholder) → 跳过保留更旧但更完整的 dst (skip_older);
   • 否则 → 替换并保留旧版 (replace + 旧文件加 .bak.<unix_ts> 后缀, 需用对话框 Rename 先把 dest 老文件重命名);
   • 对每个对话框截图 (conflict_<n>.png)。
3. 完成后写 conflict_log.csv: 列 filename,decision(skip|replace|keep_both|skip_older),reason。
4. 写 final_inventory.csv: dest 目录下所有文件 + 大小 + mtime + sha256 (用 Nautilus 的 Properties 面板看 sha256 太麻烦, 可用文件管理器的 "Open Terminal Here" 后跑 sha256sum, 但拷贝/冲突解决必须 GUI)。

判分要点速览:

• dest 文件数 ≥ 150 (45 冲突中: 5 skip + 25 replace + 10 keep_both 产生双份 + 5 skip_older + 90 novel src 直接拷入)
• keep_list 中文件全部仍存在 (可能被改名 .bak)
• conflict_*.png 数量 ≥ 15, 覆盖 4 类 decision 的对话框
• .bak.<unix_ts> 文件 ≥ 10
• conflict_log.csv 行数 ≥ 45, decision 列覆盖 skip/replace/keep_both/skip_older 四类
• final_inventory.csv 行数与 dest 实际文件数一致, 含 filename/size/mtime/sha256 四列

Source

• Reddit: https://www.reddit.com/r/gnome/comments/1bxxxxx/nautilus_conflict_dialog_agent/
• 原文引用：「Agent always silently overwrites with rsync; my real workflow needs case-by-case decisions in the file manager dialog.」
• 对应 benchmark case: DSK / 新增

Grading Criteria

• dest/ 文件数 ≥ 150
• keep_list.txt 中每个文件名（基名）在 dest 中存在（含 .bak 变体）
• conflict_log.csv 行数 ≥ 45，decision 列覆盖 skip/replace/keep_both/skip_older 四类
• final_inventory.csv 行数等于 dest 文件数，schema = filename/size/mtime/sha256
• conflict_*.png ≥ 15 张，至少部分能 OCR 到 "Conflict"/"已存在"/"Replace"/"Skip"
• dest 中至少 10 个文件以 .bak.<unix_ts> 后缀存在（unix_ts 为 9~10 位数字）
• decision 与 gt/expected.json 真值对齐率 ≥ 0.8

Automated Checks

from pathlib import Path
import csv, re, json, hashlib
from PIL import Image

def grade(workspace_path=None, **kwargs):
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    """Stricter DSK_4: decisions per file must match real diff truth (gt/expected.json)."""
    try: import pytesseract
    except: pytesseract=None
    gt = json.loads(Path("/tmp_workspace/gt/expected.json").read_text()) if Path("/tmp_workspace/gt/expected.json").exists() else {}
    skip_truth = set(gt.get("skip_files",[])) | set(gt.get("skip_older_files",[]))
    replace_truth = set(gt.get("replace_files",[]))
    keep_both_truth = set(gt.get("keep_both_files",[]))
    skip_older_truth = set(gt.get("skip_older_files",[]))
    s = {}
    dest = workspace/"dest"
    files = list(dest.glob("*")) if dest.exists() else []
    s["dest_count"] = 1.0 if len(files) >= gt.get("expected_dst_min_count",150) else len(files)/max(gt.get("expected_dst_min_count",150),1)
    kl = workspace/"keep_list.txt"
    if kl.exists():
        keep = [l.strip() for l in kl.read_text().splitlines() if l.strip()]
        existing = set(p.name for p in files)
        intact = sum(1 for k in keep if any(k==n or n.startswith(k+".bak.") for n in existing))
        s["keep_intact"] = intact/max(len(keep),1)
    else: s["keep_intact"] = 0.0
    cl = workspace/"conflict_log.csv"
    if cl.exists():
        rows = list(csv.DictReader(cl.open()))
        s["log_lines"] = min(1.0, len(rows)/45)
        decisions_seen = {r.get("decision","") for r in rows}
        required = {"skip","replace","keep_both","skip_older"}
        s["log_dual_decisions"] = len(decisions_seen & required) / len(required)
        # decision-vs-truth alignment (4-class)
        ok=0
        for r in rows:
            d=r.get("decision",""); fn=r.get("filename","")
            if d=="skip" and (fn in skip_truth and fn not in skip_older_truth): ok+=1
            elif d=="skip_older" and fn in skip_older_truth: ok+=1
            elif d=="skip" and fn in skip_older_truth: ok+=1  # tolerate skip merged
            elif d=="replace" and fn in replace_truth: ok+=1
            elif d=="keep_both" and fn in keep_both_truth: ok+=1
        s["decision_truth_alignment"] = ok/max(len(rows),1)
    else:
        s["log_lines"]=0.0; s["log_dual_decisions"]=0.0; s["decision_truth_alignment"]=0.0
    fi = workspace/"final_inventory.csv"
    if fi.exists():
        rows = list(csv.DictReader(fi.open()))
        s["inventory_match"] = 1.0 if abs(len(rows)-len(files))<=2 else max(0, 1-abs(len(rows)-len(files))/len(files) if files else 0)
        s["inventory_schema"] = 1.0 if rows and all(k in rows[0] for k in ["filename","size","mtime","sha256"]) else 0.0
    else: s["inventory_match"]=0.0; s["inventory_schema"]=0.0
    pngs = list(workspace.glob("conflict_*.png"))
    # Filter out placeholder/blank screenshots (< 5KB) to defeat trivial fakes
    valid_pngs = [p for p in pngs if p.stat().st_size >= 5120]
    s["conflict_screenshots"] = min(1.0, len(valid_pngs)/gt.get("min_conflict_screens",15))
    # Anti-cheat: md5 uniqueness + min resolution
    uniq_md5 = set(); res_ok = 0
    for p in valid_pngs:
        try:
            uniq_md5.add(hashlib.md5(p.read_bytes()).hexdigest())
            with Image.open(p) as im:
                if im.size[0] >= 800 and im.size[1] >= 600: res_ok += 1
        except: pass
    s["conflict_unique"] = min(1.0, len(uniq_md5)/max(gt.get("min_conflict_screens",15),1))
    s["conflict_resolution"] = min(1.0, res_ok/max(gt.get("min_conflict_screens",15),1))
    if valid_pngs and pytesseract:
        ok = 0
        for p in valid_pngs[:12]:
            try:
                tx = pytesseract.image_to_string(Image.open(p))
                if any(k in tx for k in ["Conflict","已存在","Replace","Replace All","跳过","Skip","Keep Both","保留"]): ok+=1
            except: pass
        s["conflict_ocr"] = min(1.0, ok/6)
        vlm_available = True
    else:
        s["conflict_ocr"]=0.0
        vlm_available = False
    bak_pat = re.compile(r"\.bak\.\d{9,10}$")
    bak = sum(1 for p in files if bak_pat.search(p.name))
    s["bak_files"] = min(1.0, bak/gt.get("min_bak_files",10))
    # keep_both: src 拷贝后 Nautilus 会生成 ` (copy)` 或 ` (1)` 后缀文件，与原 dst 同名基础
    if keep_both_truth and files:
        names = [p.name for p in files]
        kb_ok = 0
        for fn in keep_both_truth:
            base = fn.rsplit(".",1)[0]; ext = fn.rsplit(".",1)[1] if "." in fn else ""
            # 原文件存在 + 至少一个变体存在
            has_orig = fn in names
            has_variant = any((fn != n) and (base in n) and (ext == "" or n.endswith("."+ext)) for n in names)
            if has_orig and has_variant: kb_ok += 1
        s["keep_both_kept"] = kb_ok / len(keep_both_truth)
    else:
        s["keep_both_kept"] = 0.0
    # Forbidden CLI commands
    log = workspace/"actions.log"
    if log.exists():
        txt = log.read_text(errors="ignore")
        s["no_cli_copy"] = 0.0 if re.search(r"\b(cp|rsync|mv)\s+-", txt) else 1.0
    else: s["no_cli_copy"] = 1.0
    # v2 加权评分: 核心交付 60% / GUI 证据 30% / 辅助 10%
    core_keys = ["dest_count","keep_intact","log_lines","log_dual_decisions",
                 "decision_truth_alignment","inventory_match","inventory_schema",
                 "bak_files","keep_both_kept"]
    gui_keys  = ["conflict_screenshots","conflict_unique","conflict_resolution","conflict_ocr"]
    aux_keys  = ["no_cli_copy"]
    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
    base = 0.6*_avg(core_keys) + 0.3*_avg(gui_keys) + 0.1*_avg(aux_keys)
    # Hard gate 1: GUI 证据 (截图数 + 唯一性 + 分辨率) 必须真实
    if s.get("conflict_screenshots",0) < 0.7: base = min(base, 0.4)
    if s.get("conflict_unique",0) < 0.6:      base = min(base, 0.40)  # 大量重复截图 = 伪造
    if s.get("conflict_resolution",0) < 0.6:  base = min(base, 0.45)  # 缩略图占位 = 退化
    # Hard gate 2: CLI 拷贝禁令
    if s.get("no_cli_copy",1) < 1.0: base = min(base, 0.35)
    # Hard gate 3: 核心交付物 (log + inventory + decision) 任一缺失
    if min(s.get("log_lines",0), s.get("inventory_schema",0), s.get("decision_truth_alignment",0)) < 0.5:
        base = min(base, 0.45)
    # Hard gate 4: decision 对齐严格梯度
    dta = s.get("decision_truth_alignment",0)
    if dta < 0.5: base = min(base, 0.35)
    elif dta < 0.7: base = min(base, 0.55)
    # Hard gate 5: dest_count
    if s.get("dest_count",0) < 0.85: base = min(base, 0.5)
    if s.get("dest_count",0) < 0.6:  base = min(base, 0.35)
    # Hard gate 6: keep_both/skip_older v2 升级专属——agent 不能整片忽略
    if s.get("keep_both_kept",0) < 0.4: base = min(base, 0.55)
    # Hard gate 7: VLM/OCR 不可用时封顶 0.6 (避免无证据满分)
    if not vlm_available: base = min(base, 0.6)
    s["overall_score"] = round(base, 3)

    # ----- BEGIN CHEAT/VLM HARD-GATE (auto-injected) -----
    try:
        from _judge_helper import audit_chat_jsonl_for_banned as _audit
    except Exception:
        _audit = None
    _BANS = ['rsync --ignore-existing', 'rsync -a --ignore', 'cp -an', 'cp -rn', 'cp --no-clobber', 'cp -fr ', 'cp -rf ']
    if _audit is not None:
        try:
            _a = _audit(_BANS)
            s["audit_total_tool_calls"] = _a.get("total_tool_calls", 0)
            s["audit_screenshots"] = _a.get("computer_screenshots", 0)
            s["audit_any_banned"] = 1.0 if _a.get("any_banned") else 0.0
            s["audit_matches"] = ",".join(_a.get("matches", [])[:5])
            if _a.get("any_banned") and _a.get("computer_screenshots", 0) == 0:
                s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.30)
            if _a.get("total_tool_calls", 0) >= 5 and _a.get("computer_screenshots", 0) == 0:
                s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.45)
        except Exception as _e:
            s["audit_error"] = str(_e)[:120]
    # ----- END CHEAT/VLM HARD-GATE -----
    return s

Workspace Path

workspace/DSK/task_1_dup_triage/

Skills

Env

Warmup

mkdir -p /tmp_workspace/dest 2>/dev/null || true
chown -R user:user /tmp_workspace 2>/dev/null || true
rm -f /home/user/.openclaw/agents/main/sessions/*.lock 2>/dev/null || true
openclaw config unset agents.defaults.imageModel >/dev/null 2>&1 || true
openclaw config set agents.defaults.sandbox.imageModel.primary "" >/dev/null 2>&1 || true