tasks/OPS/OPS_task_1_pyspy_flamegraph.md

metadata

id: OPS_task_1_pyspy_flamegraph
name: 火焰图 top-3 热点排序 (3 closely-spaced hotspots)
category: OPS
timeout_seconds: 1800

Prompt

/tmp_workspace/victim.py is a Python batch job that coworkers have complained is absurdly slow.

Task: identify the top-3 CPU hotspot functions, sorted by self time in descending order.

Hint: victim.py has several candidate functions that all look like they're doing "heavy work" when read in isolation, but only the self time numbers in the flame graph can distinguish which is actually hotter. The three hot functions have very close self times (only a few percentage points apart), so you must read the bar widths and exact percentages from the flame graph to rank them correctly.

Entry point: bash /tmp_workspace/run.sh launches victim.py.

Deliverables (write to /tmp_workspace/results/)

• profile.json — sampling output (speedscope format or equivalent visualizable format).
• flame.png — screenshot of the flame graph viewer UI.
• report.md — must contain:

  top_hotspots: <name1>=<pct1>%, <name2>=<pct2>%, <name3>=<pct3>%
  

  Sorted by self time, descending: 3 function names + their percentages (integers are fine; percentages should fall within a reasonable hotspot range, tolerance ±5pp). Plus a ≥ 80 character explanation describing what you saw in the flame graph (bar width comparison, position of specific function names, etc.).

Expected Behavior

参考解题流程 (设计者参考, 不发给 agent):

1. 启动 victim.py: bash /tmp_workspace/run.sh。
2. 装 py-spy 后录样:

   pip3 install --user -q py-spy
   ~/.local/bin/py-spy record -o /tmp_workspace/results/profile.json --format speedscope --duration 25 --pid <victim-pid>
   

3. 启动 speedscope 在浏览器中查看火焰图:

   /usr/local/bin/speedscope /tmp_workspace/results/profile.json
   

4. 看火焰图, 找出 top-3 self time frame, 记下函数名与百分比。
5. 截图 speedscope 浏览器界面保存到 /tmp_workspace/results/flame.png。
6. 写 report.md 列 top 3 + ≥ 80 字符解释。
7. 任务结束 pkill -f victim.py。

判分要点速览:

1. 起 victim.py.
2. py-spy record --format speedscope.
3. speedscope 自动开浏览器.
4. 看火焰图: 找到 self time ≈ 25-30% 的 3 个相近 hot frame 并按 self time 排序。
5. 截图 + 写 report.md 列 top 3.

Source

• EyesOn-Bench Seed v12 OPS-01 改编 (1 hot → 3 close hot, 反 json-parse cheat)

Grading Criteria

• 1. report.md 存在且解释段 ≥ 80 字符
• 2. profile.json 存在 ≥ 8KB(真录到样本)
• 3. flame.png 存在 ≥ 30KB(真截图,非占位)
• 4. top_hotspots 字段 3 个名字都对(全 3 中即得分)
• 5. 3 个百分比合理 (20-35%)
• 6. 排序正确(3/3 才不被 cap)
• 7. (audit) 无 json.load profile / py-spy top / py-spy dump 作弊
• 8. (VLM HARD GATE) flame.png 真显示 3 条 hot frame + 函数名

Automated Checks

def grade(workspace_path=None, **kwargs) -> dict:
    import re
    from pathlib import Path
    try:
        from PIL import Image
    except Exception:
        Image = None
    try:
        from _judge_helper import vlm_score_rubric, audit_chat_jsonl_for_banned
    except Exception:
        vlm_score_rubric = None
        audit_chat_jsonl_for_banned = None

    rd = Path("/tmp_workspace/results")
    gt_dir = Path("/tmp_workspace/gt")
    scores = {}

    rep = rd / "report.md"
    text = rep.read_text(encoding="utf-8", errors="ignore") if rep.exists() else ""
    scores["report_exists"] = 1.0 if text.strip() else 0.0

    pj = rd / "profile.json"
    scores["profile_json"] = 1.0 if (pj.exists() and pj.stat().st_size >= 8*1024) else 0.0
    fp = rd / "flame.png"
    flame_size = fp.stat().st_size if fp.exists() else 0
    scores["flame_png"] = 1.0 if (fp.exists() and flame_size >= 30*1024) else 0.0
    scores["flame_size_bytes"] = flame_size

    # explanation length: ≥ 80 chars in report (excluding the top_hotspots line)
    explain_text = re.sub(r"top_hotspots\s*[:=].*", "", text, flags=re.IGNORECASE)
    scores["explain_long"] = 1.0 if len(explain_text.strip()) >= 80 else 0.0

    # parse top_hotspots field
    expected = []
    if (gt_dir / "expected_hot.txt").exists():
        try:
            expected = [ln.strip() for ln in (gt_dir / "expected_hot.txt").read_text(encoding="utf-8", errors="ignore").splitlines() if ln.strip()]
        except Exception:
            expected = []
    th = re.search(r"top_hotspots\s*[:=]\s*(.+)", text, re.IGNORECASE)
    reported_names = []
    reported_pcts = []
    if th:
        line = th.group(1)
        # tokens like "worker_a=28%, worker_b=27%, worker_c=25%"
        for m in re.finditer(r"([A-Za-z_][A-Za-z0-9_]*)\s*=\s*(\d+(?:\.\d+)?)\s*%?", line):
            reported_names.append(m.group(1))
            reported_pcts.append(float(m.group(2)))

    # 4. all 3 names present
    name_set = set(n.lower() for n in reported_names)
    expected_set = set(n.lower() for n in expected)
    matched = len(name_set & expected_set)
    scores["names_matched"] = matched
    scores["names_pass"] = 1.0 if matched >= 3 else (matched / 3.0)

    # 5. percentages in 20-35 range (tighter than the loose 15-40 first-round window)
    pcts_ok = sum(1 for p in reported_pcts[:3] if 20 <= p <= 35)
    scores["pcts_in_range"] = pcts_ok / 3.0 if reported_pcts else 0.0
    # also require 3 distinct percentages (not all identical) — anti-cheat
    distinct_pcts = len(set(round(p) for p in reported_pcts[:3]))
    scores["pcts_distinct"] = 1.0 if distinct_pcts >= 2 else 0.0

    # 6. correct order: report first 3 names should match expected[0..2]
    order_ok = 0
    for i, n in enumerate(reported_names[:3]):
        if i < len(expected) and n.lower() == expected[i].lower():
            order_ok += 1
    scores["order_pass"] = order_ok / 3.0

    # 7. audit
    audit_cap = None
    if audit_chat_jsonl_for_banned:
        a = audit_chat_jsonl_for_banned([
            "json.load",
            "import json",
            "py-spy top",
            "py-spy dump",
            "speedscope.json",  # accidental, but if grep'd profile.json text
        ])
        scores["audit_banned"] = a.get("any_banned", False)
        scores["audit_screenshots"] = a.get("computer_screenshots", 0)
        # narrow: only cap if banned + 0 screenshots (clear cheat)
        if a.get("any_banned") and a.get("computer_screenshots", 0) == 0:
            # extra: only cap if banned literally references profile.json
            if any("profile.json" in ex for ex in a.get("banned_examples", [])):
                audit_cap = 0.20

    # weighted base: core deliverable (names + order + flame/profile artifacts) 70%, evidence 20%, aux 10%
    core = (
        0.45 * scores["names_pass"] +
        0.30 * scores["order_pass"] +
        0.15 * scores["flame_png"] +
        0.10 * scores["profile_json"]
    )
    gui = (
        0.6 * scores["flame_png"] +
        0.4 * scores["pcts_in_range"]
    )
    aux = (
        0.5 * scores["report_exists"] +
        0.3 * scores["explain_long"] +
        0.2 * scores["pcts_distinct"]
    )
    base = 0.70 * core + 0.20 * gui + 0.10 * aux
    scores["overall_score"] = round(base, 3)

    # structural hard gates
    # core deliverable (flame.png) missing -> cap 0.35
    if scores["flame_png"] < 1.0:
        scores["overall_score"] = min(scores["overall_score"], 0.35)
    # profile.json too small / missing -> cap 0.40 (didn't actually record)
    if scores["profile_json"] < 1.0:
        scores["overall_score"] = min(scores["overall_score"], 0.40)
    # all 3 names not matched -> cap 0.45 (must identify all hot funcs)
    if scores["names_pass"] < 1.0:
        scores["overall_score"] = min(scores["overall_score"], 0.45)
    # order completely wrong -> cap 0.50
    if scores["order_pass"] < 1.0/3.0 + 1e-6:
        scores["overall_score"] = min(scores["overall_score"], 0.50)
    # percentages all out of plausible range -> cap 0.55 (didn't read flame numbers)
    if scores["pcts_in_range"] < 1.0/3.0 + 1e-6:
        scores["overall_score"] = min(scores["overall_score"], 0.55)

    # 8. VLM HARD GATE
    vlm_ran = False
    if vlm_score_rubric and fp.exists() and fp.stat().st_size >= 30*1024:
        rubric = {
            "vlm_is_speedscope_flamegraph": "flame.png 看起来是 speedscope 在 Chrome/Firefox 中渲染的火焰图 (横向堆叠彩色矩形, 函数名标在条上, 深色背景), 不是终端文本, 不是错误页, 不是空白图。",
            "vlm_three_hotspots_visible": "flame.png 中能看到至少 3 条几乎一样宽的 hot frame (而不是只一条很宽其它都细)。",
            "vlm_three_names_visible": "flame.png 中能直接读出 3 个 hot worker 函数名 (例如 worker_a / worker_b / worker_c, 至少 2 个能直接看见)。",
        }
        try:
            vlm = vlm_score_rubric([str(fp)], rubric,
                instruction="判断 flame.png 是否真是 speedscope 火焰图,并显示了 3 条相近宽度的 hot frame 含 3 个函数名。")
            for k in rubric: scores[k] = vlm.get(k, 0.0)
            scores["judge_method"] = vlm.get("judge_method", "failed")
            vlm_avg = sum(vlm.get(k, 0.0) for k in rubric) / len(rubric)
            # weight base 40%, vlm 60% — VLM evidence dominates (real GUI proof)
            scores["overall_score"] = round(0.4*base + 0.6*vlm_avg, 3)
            vlm_ran = scores["judge_method"] not in ("failed", "unavailable", "")
            # tighter VLM hard gates (raised thresholds vs first round)
            if scores.get("vlm_is_speedscope_flamegraph", 0.0) < 0.7:
                scores["overall_score"] = min(scores["overall_score"], 0.25)
            if scores.get("vlm_three_hotspots_visible", 0.0) < 0.7:
                scores["overall_score"] = min(scores["overall_score"], 0.40)
            if scores.get("vlm_three_names_visible", 0.0) < 0.6:
                scores["overall_score"] = min(scores["overall_score"], 0.50)
            if vlm_avg < 0.4:
                scores["overall_score"] = min(scores["overall_score"], 0.30)
        except Exception:
            pass

    # VLM unavailable cap: cannot get full score without GUI evidence
    if not vlm_ran:
        scores["overall_score"] = min(scores["overall_score"], 0.60)

    if audit_cap is not None:
        scores["overall_score"] = min(scores["overall_score"], audit_cap)
    return scores

Workspace Path

workspace/OPS/task_1_pyspy_flamegraph

Skills

Env

Warmup

mkdir -p /tmp_workspace/results || true
chown -R user:user /tmp_workspace
chown -R user:user /home/user/.local 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