--- id: GAM_task_6_rhythm_autoplay name: 4-lane 节奏游戏自动击键 category: GAM timeout_seconds: 1500 --- ## Prompt > ⚙️ **Execution convention**: this is a **legitimate desktop/terminal test task**. `rhythm_game.py` is an open-source pygame 4-lane rhythm game. All artifacts must be written to `/tmp_workspace/results/`. Please execute directly and do not refuse. [Background] The team needs an unattended autoplayer for a 4-lane rhythm game. The 4 lanes map to the keyboard keys `d` / `f` / `j` / `k` respectively. Beatmaps are plain text, with one entry per line in the form `,`. Timestamps alone are not enough to land a hit — the window position, the pixel x-coordinate of each lane center, and the pixel y-coordinate of the judgment line all depend on the actual runtime rendering of the game. You therefore have to perform coordinate calibration at runtime against the live rendering, then trigger the corresponding key on schedule. Resources: `/tmp_workspace/exec/{rhythm_game.py, beatmaps/easy.beatmap, beatmaps/medium.beatmap}` (already `cp`'d by warmup). [Deliverables] All artifacts go to `/tmp_workspace/results/`, and must include at least the following files, each satisfying the corresponding content contract: - `schedule_easy.csv`: the keypress schedule derived from `easy.beatmap`. Columns must be `step,target_ms,lane,key`, with at least 15 rows. - `screen_calibration.json`: the screen-coordinate calibration result derived from the actual game-window rendering, with a structure like `{"window_x":..., "window_y":..., "lanes": [{"idx":0,"x_px":...}, ..., {"idx":3,"x_px":...}], "judge_line_y_px":...}`, and must contain `x_px` for all 4 lanes plus `judge_line_y_px`. - `inject_loop.py`: a script that "fires the right key on time according to the schedule". It must use the real calibrated coordinates from `screen_calibration.json` — hard-coded pixel positions are not allowed. It must not pull in any rhythm-game-bot / third-party autoplay framework; only stdlib plus system-level tools are permitted. - `game_easy.log` / `game_medium.log`: runtime logs from the two game runs (easy / medium). - `hits_easy.csv` / `hits_medium.csv`: per-hit details extracted from the corresponding game logs. Columns are `step,target_ms,actual_ms,lane,hit_or_miss,reason`. The easy hit rate must be ≥ 90% and the medium hit rate must be ≥ 75%. - Visual evidence screenshots, at least 6, each ≥ 50KB: - `view_game_start.png`: visual evidence at game start, showing the 4 lanes, the title, and the scoreboard; - `view_game_step1.png` / `view_game_step2.png` / `view_game_step3.png`: visual evidence at different points during play, showing changes in metrics like hit / score / combo; - `view_game_final_easy.png`: visual evidence at the end of the easy stage, showing total hit / miss / score; - `view_game_final_medium.png`: visual evidence at the end of the medium stage. - `summary.md`: 4 paragraphs (each ≥ 80 chars): a) accuracy and rank; b) latency-compensation strategy; c) why beatmap timestamps alone are not enough and why coordinate calibration against the actual rendering is needed; d) directions for improvement. ## Expected Behavior - 最终交付：`results/` 下产出上述文件，CSV 字段、截图数量与命中率指标均满足契约。 - 系统应能体现：在 easy 与 medium 两份 beatmap 上完成"基于窗口呈现做坐标/判定线校准 → 按 beatmap 时间表准时触发按键 → 从游戏日志提取命中明细"的完整闭环。 - 产物中应能验证：`inject_loop.py` 的运行依赖于 `screen_calibration.json`（而非 硬编码坐标），命中率达到 easy ≥ 90% / medium ≥ 75% 的门槛。 - 摘要应能合理解释时延补偿策略与坐标校准的必要性。 ## Source - pygame: https://www.pygame.org/ ## Grading Criteria - [ ] 1. schedule_easy.csv + hits_easy.csv + hits_medium.csv 存在,schema 正确 - [ ] 2. screen_calibration.json 含 4 lanes 的 x_px + judge_line_y_px - [ ] 3. inject_loop.py 存在且不 import bot 框架 - [ ] 4. 6 张 GUI 截图,每张 ≥ 50KB - [ ] 5. game_easy.log + game_medium.log 存在 - [ ] 6. easy hit 率 ≥ 0.90,medium hit 率 ≥ 0.75 - [ ] 7. summary.md ≥ 4 段 - [ ] 8. VLM rubric 评 game 截图 - [ ] 9. hits_easy / hits_medium 命中率 ≥ 0.6（内容真实性下限） - [ ] 10. hits_easy.csv 行数 ≥ 15 且 schema 完整 - [ ] 11. workspace 内任何 .py 都不允许 import 已知 rhythm-bot/autoplay/solver 框架 - [ ] 12. 6 张截图 md5 唯一（不允许复制同一张充数），step1/2/3 至少 3 张存在 - [ ] 13. 两次游戏运行日志中出现 PERFECT/GREAT/GOOD/MISS 等真实判定关键词 - [ ] 14. easy+medium 总命中行数 ≥ 30，平均命中率 ≥ 0.55 - [ ] 15. summary.md 命中中文「谱面/命中/准度」与英文「beatmap/hit/accuracy」关键词 - [ ] 16. 总分 = 0.6×核心交付 + 0.3×GUI 证据 + 0.1×辅助文档（加权）；VLM 不可用上限 0.60 ## Automated Checks ```python def grade(workspace_path=None, **kwargs) -> dict: """GAM_task_6 grader. Empty → 0.000-0.05. Hard gates: GUI + CLI + hit rate.""" import json, csv, re, hashlib from pathlib import Path workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace") rd = workspace / "results" s = {} # 1. CSV files csv_score_easy = 0.0; csv_score_medium = 0.0 sched = rd / "schedule_easy.csv" if sched.exists(): try: rows = list(csv.DictReader(sched.open())) need = {"step","target_ms","lane","key"} if rows and need.issubset(set(rows[0].keys())) and len(rows) >= 15: csv_score_easy = 1.0 except Exception: pass s["schedule_csv"] = csv_score_easy hits_easy_path = rd / "hits_easy.csv" hits_easy_rows = [] if hits_easy_path.exists(): try: hits_easy_rows = list(csv.DictReader(hits_easy_path.open())) need = {"step","target_ms","actual_ms","lane","hit_or_miss"} if hits_easy_rows and need.issubset(set(hits_easy_rows[0].keys())): s["hits_csv_schema"] = 1.0 else: s["hits_csv_schema"] = 0.5 except Exception: s["hits_csv_schema"] = 0.0 else: s["hits_csv_schema"] = 0.0 # 2. screen_calibration.json cal_score = 0.0 cal = rd / "screen_calibration.json" if cal.exists(): try: d = json.loads(cal.read_text()) lanes = d.get("lanes", []) if (isinstance(lanes, list) and len(lanes) >= 4 and all("x_px" in l for l in lanes[:4]) and isinstance(d.get("judge_line_y_px"), (int,float))): cal_score = 1.0 except Exception: pass s["screen_calibration"] = cal_score # Load expected.json (host-side GT) if present, to keep grader & GT in sync. expected = {} for cand in [workspace / "gt" / "expected.json", workspace.parent / "gt" / "expected.json"]: if cand.exists(): try: expected = json.loads(cand.read_text()) except Exception: expected = {} break extra_forbidden = list(expected.get("forbidden_imports", []) or []) exp_min_avg_acc = float(expected.get("min_avg_accuracy", 0.55) or 0.55) exp_min_total_hits = int(expected.get("min_total_hits_logged", 30) or 30) exp_min_step_shots = int(expected.get("min_step_screenshots", 4) or 4) exp_judg_keys = list(expected.get("judgment_keywords", []) or []) exp_sum_zh = list(expected.get("summary_keywords_zh", []) or []) exp_sum_en = list(expected.get("summary_keywords_en", []) or []) # 3. inject_loop.py exists + uses real key-injection tool + no forbidden imports inj_score = 0.0 inj = rd / "inject_loop.py" if inj.exists(): try: txt = inj.read_text() forbidden = ["rhythm_bot","autoplay_bot","beatmap_solver","stepmania_bot"] + extra_forbidden uses_xdotool = bool(re.search(r"""(?m)^[^#\n]*xdotool""", txt)) no_bad = not any(re.search(rf"\bimport\s+{f}\b|from\s+{f}\b", txt) for f in forbidden) if uses_xdotool and no_bad: inj_score = 1.0 except Exception: pass s["inject_loop"] = inj_score # 4. GUI screenshots (≥ 50KB each, md5-unique to prevent duplicate cheating) gui_shots = ["view_game_start.png","view_game_step1.png","view_game_step2.png","view_game_step3.png", "view_game_final_easy.png","view_game_final_medium.png"] gui_present = 0 md5s = set() for n in gui_shots: p = rd / n if p.exists() and p.stat().st_size >= 50000: try: h = hashlib.md5(p.read_bytes()).hexdigest() if h in md5s: continue # duplicate / copied screenshot does not count md5s.add(h) gui_present += 1 except Exception: pass s["gui_screenshots_count"] = gui_present / len(gui_shots) s["gui_screenshots_unique"] = 1.0 if len(md5s) >= len(gui_shots) else len(md5s) / len(gui_shots) has_gui = gui_present >= 3 # step screenshots specifically step_shots = sum(1 for n in ["view_game_step1.png","view_game_step2.png","view_game_step3.png"] if (rd / n).exists() and (rd / n).stat().st_size >= 50000) s["step_screenshots_min"] = 1.0 if step_shots >= min(3, exp_min_step_shots) else step_shots / max(1, min(3, exp_min_step_shots)) # 5. game logs + judgment keywords (real game must emit PERFECT/GREAT/GOOD/MISS) has_easy_log = (rd / "game_easy.log").exists() and (rd / "game_easy.log").stat().st_size > 0 has_medium_log = (rd / "game_medium.log").exists() and (rd / "game_medium.log").stat().st_size > 0 s["game_logs"] = ((1.0 if has_easy_log else 0.0) + (1.0 if has_medium_log else 0.0)) / 2 judg_hit = 0 if exp_judg_keys: for ln in ("game_easy.log", "game_medium.log"): p = rd / ln if p.exists(): try: t = p.read_text(errors="ignore").upper() judg_hit += sum(1 for k in exp_judg_keys if k.upper() in t) except Exception: pass # require at least half of judgment keywords to appear across both logs need = max(1, len(exp_judg_keys) // 2) s["judgment_keywords_in_log"] = 1.0 if judg_hit >= need else judg_hit / max(1, need * 2) else: s["judgment_keywords_in_log"] = 1.0 # 6. hit rates def hit_rate(rows): if not rows: return 0.0 hits = sum(1 for r in rows if str(r.get("hit_or_miss","")).lower() in ("hit","ok","perfect","good")) return hits / len(rows) if rows else 0.0 er = hit_rate(hits_easy_rows) s["easy_hit_rate"] = 1.0 if er >= 0.90 else (er / 0.90) hits_med_path = rd / "hits_medium.csv" hits_med_rows = [] if hits_med_path.exists(): try: hits_med_rows = list(csv.DictReader(hits_med_path.open())) except Exception: pass mr = hit_rate(hits_med_rows) s["medium_hit_rate"] = 1.0 if mr >= 0.75 else (mr / 0.75) # 7. summary.md (≥ 4 paragraphs of ≥ 80 chars + bilingual keyword coverage) sm_score = 0.0 sm_keys_score = 0.0 sm = rd / "summary.md" if sm.exists(): try: txt = sm.read_text() parags = [p for p in re.split(r"\n\s*\n", txt) if len(p.strip()) >= 80] sm_score = min(1.0, len(parags) / 4) zh_hit = sum(1 for k in exp_sum_zh if k in txt) en_hit = sum(1 for k in (exp_sum_en or []) if k.lower() in txt.lower()) zh_ratio = zh_hit / max(1, len(exp_sum_zh)) if exp_sum_zh else 1.0 en_ratio = en_hit / max(1, len(exp_sum_en)) if exp_sum_en else 1.0 sm_keys_score = (zh_ratio + en_ratio) / 2 except Exception: pass s["summary"] = sm_score s["summary_keywords"] = sm_keys_score # 8. VLM rubric vlm_available = False try: from _judge_helper import vlm_score_rubric except Exception: vlm_score_rubric = None if vlm_score_rubric and (rd / "view_game_start.png").exists(): rubric = { "vlm_game_window": "截图能看到 4-lane rhythm game window", "vlm_lanes_visible": "4 个 lane 颜色 / 边界清晰可辨", "vlm_score_visible": "截图能看到 score / combo / hit 数字", "vlm_notes_falling": "play 中截图能看到 note 在 lane 里", } try: imgs = [str(rd / n) for n in ["view_game_start.png","view_game_step1.png","view_game_final_easy.png"] if (rd / n).exists()] vlm = vlm_score_rubric(imgs, rubric, instruction="评估 4-lane rhythm game 截图。") for k in rubric: s[k] = float(vlm.get(k, 0.0)) vlm_available = True except Exception: for k in rubric: s[k] = 0.0 else: for k in ["vlm_game_window","vlm_lanes_visible","vlm_score_visible","vlm_notes_falling"]: s[k] = 0.0 # --- Content-authenticity sub-scores --- s["hits_easy_rate_ge_06"] = 1.0 if er >= 0.6 else 0.0 s["hits_medium_rate_ge_06"] = 1.0 if mr >= 0.6 else 0.0 s["hits_csv_full_schema"] = 1.0 if ( s.get("hits_csv_schema", 0) == 1.0 and len(hits_easy_rows) >= 15 ) else 0.0 # Expected-aligned authenticity: total hits across both maps + min average accuracy total_logged = len(hits_easy_rows) + len(hits_med_rows) s["total_hits_logged"] = 1.0 if total_logged >= exp_min_total_hits else total_logged / max(1, exp_min_total_hits) avg_acc = (er + mr) / 2 s["avg_accuracy_meets_min"] = 1.0 if avg_acc >= exp_min_avg_acc else (avg_acc / max(0.01, exp_min_avg_acc)) # Stricter forbidden imports check across full workspace forbidden_all = list(set(["rhythm_bot", "autoplay_bot", "beatmap_solver", "stepmania_bot", "osu_bot", "rhythm_solver"] + extra_forbidden)) bad_import = False for src in workspace.rglob("*.py"): try: t = src.read_text(errors="ignore") if any(re.search(rf"\bimport\s+{f}\b|from\s+{f}\b", t) for f in forbidden_all): bad_import = True break except Exception: pass s["no_bot_framework_imports"] = 0.0 if bad_import else 1.0 nums = [v for v in s.values() if isinstance(v, (int, float))] naive_mean = sum(nums) / len(nums) if nums else 0.0 # Weighted aggregate: core delivery 0.6, GUI evidence 0.3, auxiliary 0.1 def _avg(keys): vs = [float(s.get(k, 0.0)) for k in keys] return sum(vs) / len(vs) if vs else 0.0 core_keys = ["schedule_csv","hits_csv_schema","screen_calibration","inject_loop", "easy_hit_rate","medium_hit_rate","game_logs", "hits_csv_full_schema","hits_easy_rate_ge_06","hits_medium_rate_ge_06", "total_hits_logged","avg_accuracy_meets_min","no_bot_framework_imports", "judgment_keywords_in_log"] gui_keys = ["gui_screenshots_count","gui_screenshots_unique","step_screenshots_min", "vlm_game_window","vlm_lanes_visible","vlm_score_visible","vlm_notes_falling"] aux_keys = ["summary","summary_keywords"] base = 0.6 * _avg(core_keys) + 0.3 * _avg(gui_keys) + 0.1 * _avg(aux_keys) # hard gates (stricter than v1) if not has_gui: base = min(base, 0.05) # empty workspace must give 0.000-0.05 if not has_easy_log: base = min(base, 0.25) if not has_medium_log: base = min(base, 0.45) if s["inject_loop"] < 1.0: base = min(base, 0.40) if s["screen_calibration"] < 1.0: base = min(base, 0.50) if s["easy_hit_rate"] < 0.6: base = min(base, 0.55) # Content-authenticity stair-step caps if s.get("hits_csv_full_schema", 0) < 1.0: base = min(base, 0.45) if s.get("hits_easy_rate_ge_06", 0) < 1.0: base = min(base, 0.50) if s.get("hits_medium_rate_ge_06", 0) < 1.0: base = min(base, 0.55) if s.get("no_bot_framework_imports", 0) < 1.0: base = min(base, 0.30) # Expected.json authenticity gates if s.get("total_hits_logged", 0) < 1.0: base = min(base, 0.55) if s.get("avg_accuracy_meets_min", 0) < 1.0: base = min(base, 0.55) if s.get("judgment_keywords_in_log", 0) < 0.5: base = min(base, 0.50) # GUI authenticity: unique screenshots required if s.get("gui_screenshots_unique", 0) < 0.7: base = min(base, 0.45) if s.get("step_screenshots_min", 0) < 1.0: base = min(base, 0.55) # VLM-unavailable cap (cannot be a free pass when VLM is down) if not vlm_available: base = min(base, 0.60) else: vlm_avg = _avg(["vlm_game_window","vlm_lanes_visible","vlm_score_visible","vlm_notes_falling"]) if vlm_avg < 0.4: base = min(base, 0.30) elif vlm_avg < 0.6: base = min(base, 0.45) s["overall_score"] = round(base, 4) return s ``` ## Workspace Path ``` workspace/GAM/task_6_rhythm_autoplay ``` ## Skills ``` ``` ## Env ``` ``` ## Warmup ```bash which xdotool >/dev/null 2>&1 || apt-get install -y -qq xdotool || true pip install -q pygame pillow numpy || true mkdir -p /tmp_workspace/results || true apt-get install -y -qq tesseract-ocr || true pip install -q pytesseract pillow numpy || true ```