tasks/GAM/GAM_task_2_mines_visual.md

metadata

id: GAM_task_2_mines_visual
name: gnome-mines 30x16 解题录步
category: GAM
timeout_seconds: 1500

Prompt

Start a game of Big preset in gnome-mines (30×16, 99 mines, safe cells = 480 - 99 = 381). Play the game through the GUI and record your decision process.

Task

1. Reveal as many non-mine cells as possible, target ≥ 380 (i.e. as close as possible to all 381 revealed).
2. Flag at least 95 mines (right-click to place a flag).
3. Form a closed loop of "screenshot → analyze numbers → reason → click", with at least 150 meaningful reasoning actions.
4. If you hit a mine and the game ends, you may restart, up to 3 times total.

Deliverables (write to the /tmp_workspace/ root, not to results/ or any subdirectory)

• /tmp_workspace/clicks.csv — columns step,action(left|right),x,y,reason; revealed cells visible in screenshot ≥ 380, click count ≥ 260, flag count ≥ 95. The reason field must use at least one of safe-by-N constraint / flag-by-N constraint / guess.
• /tmp_workspace/final_board.png — final screenshot of the whole board.
• /tmp_workspace/summary.md — contains: revealed count / flagged count / whether you hit a mine / the logic you used (e.g. constraint propagation from 211/121 number patterns).

Expected Behavior

参考解题流程 (设计者参考, 不发给 agent):

1. 启动 gnome-mines, 新建一局 **Big (30×16, 99 颗雷)**。
2. 通过界面点击揭开格子并标记雷, 目标:
   • 至少揭开 ≥ 380 个非雷格子 (总安全格 = 480 - 99 = 381)。
   • 标记至少 95 颗雷 (右键插旗)。
   • 至少 150 次 有意义的推理动作。
3. 如果踩雷游戏结束, 最多 3 次 重开尝试。
4. 记录每一步的点击坐标到 clicks.csv (列 step,action(left|right),x,y,reason)。
5. 最终截图整张棋盘 final_board.png, 并写 summary.md, 含: 揭开数 / 标记数 / 是否爆雷 / 用到的推理逻辑 (数字 211/121 约束传播等)。
6. 全部产物写到 /tmp_workspace/ 根目录 (绝对路径: /tmp_workspace/clicks.csv、/tmp_workspace/final_board.png、/tmp_workspace/summary.md), 不要写到 /tmp_workspace/results/ 或子目录。

判分要点速览:

• Agent 能识别 gnome-mines 棋盘格状态 (数字、未揭开、旗)。
• 形成"截图 → 分析 → 点击"循环, 至少进行 200 次有意义的操作。
• 在含约束传播的位置准确推理 (不会在已确定 0 邻雷的格子标旗)。

设计版本备注 (v2 加难, additive):

• 局面是 gnome-mines Big preset (30×16, 99 雷, 381 安全格)。
• 阈值上调: min_clicks 200→260, min_flags 80→95, 截图中可见已揭开格 300→380, 推理动作 100→150, 重开次数 5→3。

Source

• Reddit: https://www.reddit.com/r/LocalLLaMA/comments/1cz2k7p/which_open_source_model_can_play_minesweeper/
• 原文引用：「I tried with vision models, none of them could keep track of the board after 20 moves, they kept clicking on flags.」
• 对应 benchmark case: 视觉循环推理 / GAM 新增

Grading Criteria

• clicks.csv 行数 ≥ 260 且列齐全（v2 加难阈值）
• 至少 95 行 action=right（标旗，v2 加难阈值）
• final_board.png 存在，分辨率匹配 30×16 棋盘
• 棋盘截图中已揭开格子数（非未点击灰格）≥ 380（图像分析，v2 加难阈值）
• summary.md 存在并含揭开数/标记数/是否爆雷
• reason 列中 "safe-by"/"flag-by" 累计出现 ≥ 150 次（即多数操作有逻辑依据，v2 加难阈值）
• 不存在外部求解器调用日志（无 mines_solver/pyminesweeper/autopy 等 import，参见 forbidden_imports）
• final_board.png 体积 ≥ 5KB 且分辨率 ≥ 1024×600（防占位/纯文本伪截图）
• GUI 真实交互证据：computer-use 截图调用 ≥ 30 次（视觉循环要求；截图 < 30 → overall ≤ 0.55；< 10 → overall ≤ 0.40）
• 评分采用加权（core 60% / gui 30% / aux 10%）；revealed<70%、爆解器、缺关键产物均触发分段封顶

Automated Checks

import csv, re, json, hashlib
from pathlib import Path
from PIL import Image
import numpy as np

def grade(workspace_path=None, **kwargs):
    workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    """Stricter GAM_2: require min_revealed_cells from board image, no solver libs, reasoned >=100."""
    _gt_path = workspace.parent / "gt" / "expected.json"
    if not _gt_path.exists():
        _gt_path = Path("/tmp_workspace/gt/expected.json")
    gt = json.loads(_gt_path.read_text()) if _gt_path.exists() else {}
    s = {}
    cf = workspace/"clicks.csv"
    rows = []
    if cf.exists():
        rows = list(csv.DictReader(cf.open()))
    s["clicks_count"] = min(1.0, len(rows)/gt.get("min_clicks",200))
    s["clicks_schema"] = 1.0 if rows and all(k in rows[0] for k in ["step","action","x","y","reason"]) else 0.0
    flags = sum(1 for x in rows if x.get("action")=="right")
    s["flags_right_click"] = min(1.0, flags/gt.get("min_flags_right_click",80))
    reasoned = sum(1 for x in rows if re.search(r"(safe|flag)-by-?\w*", x.get("reason","")))
    s["reasoned_actions"] = min(1.0, reasoned/gt.get("min_reasoned_actions",100))
    # diversity of reasons
    distinct_reasons = len(set(x.get("reason","").split("-")[0] for x in rows))
    s["reason_diversity"] = min(1.0, distinct_reasons/3)
    fb = workspace/"final_board.png"
    fb_size = fb.stat().st_size if fb.exists() else 0
    s["final_board_exists"] = 1.0 if (fb.exists() and fb_size >= 5120) else 0.0
    s["final_board_md5"] = hashlib.md5(fb.read_bytes()).hexdigest()[:10] if fb.exists() else ""
    s["final_board_resolution_ok"] = 0.0
    if fb.exists() and fb_size >= 5120:
        try:
            im = Image.open(fb).convert("L"); a = np.array(im)
            h,w = a.shape
            s["final_board_resolution_ok"] = 1.0 if (w >= 1024 and h >= 600) else 0.0
            s["board_nontrivial"] = 1.0 if a.std()>40 else a.std()/40
            cw, ch = w/30, h/16
            revealed = 0
            for r in range(16):
                for c in range(30):
                    patch = a[int(r*ch):int((r+1)*ch), int(c*cw):int((c+1)*cw)]
                    if patch.size and patch.std()>15:
                        revealed += 1
            s["revealed_cells_raw"] = revealed
            s["revealed_cells"] = min(1.0, revealed/gt.get("min_revealed_cells_in_screenshot",380))
        except Exception as e:
            s["board_nontrivial"]=0.0; s["revealed_cells"]=0.0
    else:
        s["board_nontrivial"]=0.0; s["revealed_cells"]=0.0
    sm = workspace/"summary.md"
    if sm.exists():
        c = sm.read_text(errors="ignore").lower()
        zh_ok = all(k in c for k in gt.get("summary_required_keywords_zh",["揭开","标记"]))
        en_ok = all(k in c for k in gt.get("summary_required_keywords_en",["revealed","flagg"]))
        s["summary_complete"] = 1.0 if (zh_ok or en_ok) else 0.0
    else:
        s["summary_complete"]=0.0
    forbidden = gt.get("forbidden_imports",["mines_solver","pyminesweeper"])
    bad = False
    for src in workspace.rglob("*.py"):
        try:
            t = src.read_text(errors="ignore")
            if any(re.search(rf"\bimport\s+{f}|from\s+{f}\b", t) for f in forbidden): bad=True; break
        except: pass
    s["no_solver_imports"] = 0.0 if bad else 1.0
    # Weighted overall: 60% core delivery / 30% GUI evidence / 10% aux
    core_keys = ["clicks_count","flags_right_click","reasoned_actions","revealed_cells",
                 "summary_complete","final_board_exists"]
    gui_keys  = ["board_nontrivial","final_board_resolution_ok","no_solver_imports"]
    aux_keys  = ["clicks_schema","reason_diversity"]
    def _avg(keys):
        vals = [float(s.get(k,0.0)) for k in keys]
        return sum(vals)/len(vals) if vals else 0.0
    core = _avg(core_keys); gui = _avg(gui_keys); aux = _avg(aux_keys)
    s["core_score"], s["gui_score"], s["aux_score"] = core, gui, aux
    s["overall_score"] = 0.6*core + 0.3*gui + 0.1*aux
    # Multi-tier hard gates
    if s.get("final_board_exists",0) < 1 or s.get("summary_complete",0) < 1:
        s["overall_score"] = min(float(s["overall_score"]), 0.40)
    if s.get("revealed_cells",0) < 0.7:
        s["overall_score"] = min(float(s["overall_score"]), 0.45)
    if s.get("revealed_cells",0) < 0.4:
        s["overall_score"] = min(float(s["overall_score"]), 0.30)
    if s.get("reasoned_actions",0) < 0.6:
        s["overall_score"] = min(float(s["overall_score"]), 0.50)
    if s.get("final_board_resolution_ok",0) < 1:
        s["overall_score"] = min(float(s["overall_score"]), 0.55)
    if s.get("no_solver_imports",1) < 1:
        s["overall_score"] = min(float(s["overall_score"]), 0.30)

    # ----- BEGIN CHEAT/VLM HARD-GATE (auto-injected) -----
    try:
        from _judge_helper import audit_chat_jsonl_for_banned as _audit
    except Exception:
        _audit = None
    _BANS = ['gnome-mines --solve', 'dconf read', '/dev/shm/mines']
    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"):
                s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.25)
            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)), 0.40)
            # GUI 真实交互证据：截图数下限（visual loop 至少 ~30 次截图）
            if _a.get("computer_screenshots", 0) < 30:
                s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.55)
            if _a.get("computer_screenshots", 0) < 10:
                s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.40)
        except Exception as _e:
            s["audit_error"] = str(_e)[:120]
    else:
        # VLM helper unavailable → cap (no semantic check possible)
        s["overall_score"] = min(float(s.get("overall_score", 1.0)), 0.60)
    # ----- END CHEAT/VLM HARD-GATE -----
    return s

Workspace Path

workspace/GAM/task_2_mines_visual/

Skills

Env

Warmup

which gnome-mines >/dev/null 2>&1 || (apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq gnome-mines)
pip install -q pillow numpy
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