metadata
id: GAM_task_13_hedgewars_lua_mission_debug
name: Hedgewars 自定义训练 Lua 任务修复 + Frontend+对战双视图通关
category: GAM
timeout_seconds: 5400
Prompt
Background: /tmp_workspace/exec/scripts/CrateCrash.lua is a hand-crafted Hedgewars training
mission. The player should collect 3 health crates on a small Bamboo-themed island to
complete the "Crate Crash Course". The current script has several issues spanning team,
ammo and map configuration that you need to locate yourself.
Your goal is to parse the original Lua mission, record reproduction evidence, produce a
syntactically valid and numerically sane fixed version, and provide image evidence of the
mission list, the bug states, and the completion state.
Please produce the following files, keeping field names and file names exactly as specified:
/tmp_workspace/results/before/CrateCrash.lua: an unmodified copy of the original mission script, for diff bookkeeping./tmp_workspace/results/parse_before.json: JSON object containing the six keysmap,theme,team,hogs,ammo,crates; thecrateslist must have length 3. Recommended:teamincludesname,color_hex,grave,fort,voicepack,flag,hogsentries includename,hat,spawn,ammoincludesamBazooka,amGrenade,amSkip,cratesentries includeindex,x,y./tmp_workspace/results/bugs.json: JSON object containing three items:bug_team_color_invisible(withdetected,color_hex, optionallytheme,rationale),bug_player_no_ammo(withdetected,weapons_with_zero),bug_crate_offscreen(withdetected,crate_index,y, optionallymap_max_y). Each item's fields should document the colour, ammo or coordinate evidence that triggered detection./tmp_workspace/work/scripts/CrateCrash_fixed.lua: the fixed Lua script; it must be syntactically valid, the colour passed toAddTeam(...)must be a legal hex value and not close to0xFFFFFF, the ammo counts foramBazookaandamGrenademust each be at least 5, and theycoordinate of the secondSpawnHealthCratemust lie within[0, 2048]./tmp_workspace/results/view_frontend.png: image evidence showing that, before the fix, the mission appears in the training mission list, with visible text such asTraining,Single,Mission,CrateorCourse./tmp_workspace/results/view_play_bug.png: image evidence showing one of the real bug states in the unfixed mission — e.g. character/name tag invisibility, abnormal Bazooka/ammo state, or insufficient health crates; recognisable text should include one ofTrainee,Rookie,Bazooka,Cratesorammo./tmp_workspace/results/view_frontend_after.png: image evidence showing the training mission list after the fix, with the mission still visible and selectable./tmp_workspace/results/view_play_finish.png: image evidence showing the mission completion state, with result text such asMission complete,All crates collectedorYou wonvisible./tmp_workspace/results/engine.log: a non-empty log containing mission/ammo-related markers such asTrainee,Rookie,ammoorbazooka./tmp_workspace/results/parse_after.json: JSON object with the same schema asparse_before.json, reflecting the structure of the fixed script./tmp_workspace/results/diff.txt: text diff between the original script andCrateCrash_fixed.lua, with between 6 and 200 lines./tmp_workspace/results/report.json: JSON object containingbugs_found,fixes_applied,team_color_after,ammo_after,crates_after,playthrough_finished,screenshots,engine_log_lines;screenshotslists the four image evidence file names.
Expected Behavior
成功时，parse_before.json 和 parse_after.json 都以
map/theme/team/hogs/ammo/crates 描述任务结构，且包含 3 个血包条目；bugs.json
对队伍颜色不可见、玩家关键武器无弹药、血包越界三类问题均给出 detected=true
和可核查字段。/tmp_workspace/work/scripts/CrateCrash_fixed.lua 应语法有效，颜色不接近白色，
amBazooka 与 amGrenade 弹药数达标，第二个血包的 y 坐标回到允许地图范围内。
图像证据应分别显示修复前训练任务列表、修复前问题状态、修复后训练任务列表和任务完成结果；
完成图像应可见 Mission complete、All crates collected 或 You won。engine.log
应含任务或弹药相关标记，diff.txt 行数处于允许范围，report.json 应汇总问题、修复、
修复后颜色/弹药/血包坐标、完成状态、图像文件名和日志行数。
Source
- Hedgewars Lua API reference: https://www.hedgewars.org/kb/LuaAPI
- Reference training mission (real upstream file): https://github.com/hedgewars/hw/blob/master/share/hedgewars/Data/Missions/Training/Basic_Training_-_Bazooka.lua
- Hedgewars
hwengineCLI flags: https://www.hedgewars.org/kb/EngineCommandLineParameters - Forum thread on white-on-snow team colour bug: https://hedgewars.org/forum/threads/team-colour-invisible-on-bamboo
- 对应 benchmark：GAM 视觉×文件双通道 Lua 任务修复 case
Grading Criteria
-
parse_before.jsonschema 完整（map + theme + team + hogs + ammo + crates 六 key 都在，crates 列表长度 == 3）
-
bugs.json三个 bug 都 detected=true 且证据字段非空
-
work/scripts/CrateCrash_fixed.lua存在且luac -p通过
-
- fixed 脚本里
AddTeam(...)第二个实参是合法十六进制颜色，且 与0xFFFFFF的距离 ≥ 0x808080（即不是接近白）
- fixed 脚本里
-
- fixed 脚本里
amBazooka与amGrenade的 ammo count 都 ≥ 5
- fixed 脚本里
-
- fixed 脚本里第 2 个
SpawnHealthCrate调用的 y 坐标在 [0, 2048] 区间内
- fixed 脚本里第 2 个
-
view_frontend.png/view_frontend_after.png都存在且 OCR 命中 "Training"/"Single"/"Mission"/"Crate" 之一
-
view_play_bug.png存在且看起来是真游戏画面（OCR 命中 "Trainee" / "Rookie" / "Bazooka" / "Crates" 之一，或 VLM 判断是真窗口）
-
view_play_finish.pngOCR 命中 "Mission complete" / "All crates collected" / "You won" 之一
-
engine.log存在且非空，含 "Trainee" / "Rookie" / "ammo" 字样
-
diff.txt存在，行数在 [6, 200] 区间
-
- VLM rubric 4 项 ≥ 0.6
Automated Checks
def grade(workspace_path=None, **kwargs) -> dict:
"""Hedgewars Lua training-mission repair + GUI playthrough verifier.
Sub-scores (12) + 3 hard gates. Real signal channels:
- File: regex parse of the Lua script for AddTeam / AddAmmo /
SpawnHealthCrate; numeric checks on team color hex,
ammo counts, crate y coord.
- GUI : 4 screenshots OCR for frontend / engine / finish overlay.
- Run : engine.log captures stderr from hwengine.
No GT leakage: gt/expected.json holds invariant counts only
(e.g. number of bugs, OCR keyword set, map_max_y, diff bounds).
"""
import json, os, re, subprocess
from pathlib import Path
ws = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
rd = ws / "results"
work = ws / "work"
gt_dir = ws / "gt"
expected = {}
if (gt_dir / "expected.json").exists():
try: expected = json.loads((gt_dir/"expected.json").read_text())
except Exception: expected = {}
s = {}
# ---- 1. parse_before.json schema ----
pb = rd / "parse_before.json"
schema_ok = 0
if pb.exists():
try:
d = json.loads(pb.read_text())
need = {"map","theme","team","hogs","ammo","crates"}
if (need.issubset(d.keys())
and isinstance(d.get("crates"), list)
and len(d["crates"]) == 3):
schema_ok = 1
except Exception: pass
s["parse_before_schema"] = float(schema_ok)
# ---- 2. bugs.json all 3 bugs detected with evidence ----
bj = rd / "bugs.json"
bugs_hit = 0
if bj.exists():
try:
b = json.loads(bj.read_text())
color_hex_gt = str(expected.get("broken_team_color_hex", "FFFFFF")).upper()
crate_idx_set = expected.get("broken_crate_index_set", [1, 2])
try: crate_idx_set = list(crate_idx_set)
except Exception: crate_idx_set = [1, 2]
offscreen_y_min = int(expected.get("map_max_landscape_y", 2048)) + 1
for k, must in [
("bug_team_color_invisible", ["color_hex"]),
("bug_player_no_ammo", ["weapons_with_zero"]),
("bug_crate_offscreen", ["crate_index","y"]),
]:
v = b.get(k, {})
ok = bool(v.get("detected")) and all(m in v for m in must)
if ok and k == "bug_team_color_invisible":
ok = color_hex_gt in str(v.get("color_hex","")).upper()
if ok and k == "bug_player_no_ammo":
wz = [str(x).lower() for x in (v.get("weapons_with_zero") or [])]
ok = ("ambazooka" in wz) and ("amgrenade" in wz)
if ok and k == "bug_crate_offscreen":
ok = (int(v.get("crate_index",-1)) in crate_idx_set) and \
int(v.get("y",0)) >= offscreen_y_min
if ok: bugs_hit += 1
except Exception: pass
s["bugs_detected_with_evidence"] = bugs_hit / 3.0
# ---- 3-6. fixed lua parses & invariants ----
fixed_paths = [work/"scripts/CrateCrash_fixed.lua",
ws/"exec/scripts/CrateCrash_fixed.lua"]
fixed = next((p for p in fixed_paths if p.exists()), None)
luac_ok = color_ok = ammo_ok = crate_ok = 0.0
txt = ""
if fixed:
try: txt = fixed.read_text(errors="ignore")
except Exception: txt = ""
try:
r = subprocess.run(["luac","-p",str(fixed)],
capture_output=True, timeout=10)
luac_ok = 1.0 if r.returncode == 0 else 0.0
except Exception:
# fallback: structural sanity
if "function onGameInit" in txt and "AddTeam" in txt:
luac_ok = 0.7
s["fixed_lua_parses"] = luac_ok
if txt:
# Team color: AddTeam(<name>, 0xRRGGBB, ...)
m = re.search(
r"AddTeam\s*\(\s*[^,]+,\s*0x([0-9A-Fa-f]{6})\b", txt)
if m:
hx = int(m.group(1), 16)
white = 0xFFFFFF
# crude colour distance: how far each channel from 0xFF
dr = abs(((hx>>16)&0xFF) - 0xFF)
dg = abs(((hx>>8 )&0xFF) - 0xFF)
db = abs(((hx )&0xFF) - 0xFF)
if (dr + dg + db) >= 0x180: # ≥ ~half-white in L1
near_white = (dr + dg + db) < 0x240
chan_dark = max((hx>>16)&0xFF,(hx>>8)&0xFF,hx&0xFF) < 0xC0
if (not near_white) and chan_dark:
color_ok = 1.0
# Ammo counts for amBazooka and amGrenade
baz = re.search(
r"AddAmmo\s*\(\s*\w+\s*,\s*amBazooka\s*,\s*(\d+)\s*\)", txt)
gre = re.search(
r"AddAmmo\s*\(\s*\w+\s*,\s*amGrenade\s*,\s*(\d+)\s*\)", txt)
if baz and gre and int(baz.group(1)) >= 5 and int(gre.group(1)) >= 5:
ammo_ok = 1.0
# Crates: collect all SpawnHealthCrate(x, y) in order
crates = [(int(a), int(b)) for a,b in re.findall(
r"SpawnHealthCrate\s*\(\s*(-?\d+)\s*,\s*(-?\d+)\s*\)", txt)]
max_y = expected.get("map_max_landscape_y", 2048)
if (len(crates) >= 3
and 0 <= crates[1][1] <= max_y
and 0 <= crates[0][1] <= max_y
and 0 <= crates[2][1] <= max_y):
crate_ok = 1.0
s["fixed_team_color"] = color_ok
s["fixed_ammo_counts"] = ammo_ok
s["fixed_crates_in_map"] = crate_ok
# ---- 7. frontend screenshots present + OCR ----
def _is_real_capture(p):
# Content-based realness check: any non-trivial PNG ≥ 800×600 with
# enough colour diversity counts as real. Drops the previous narrow
# resolution whitelist that rejected legitimate KasmVNC screenshots
# at non-standard sizes (e.g. 1024×768, 1366×768).
try:
from PIL import Image
im = Image.open(p); w, h = im.size
if w < 800 or h < 600:
return False
try:
import collections
px = list(im.convert("RGB").resize((64, 64)).getdata())
return len(collections.Counter(px)) >= 200
except Exception:
return True
except Exception:
return False
shots_fe = ["view_frontend.png","view_frontend_after.png"]
fe_present = sum(1 for n in shots_fe if (rd/n).exists() and _is_real_capture(rd/n))
s["frontend_shots_present"] = fe_present / 2.0
fe_kw = expected.get("frontend_keywords",
["training","mission","crash","course","crate","play","single"])
try:
import pytesseract
from PIL import Image
fe_ocr_hits = 0
for n in shots_fe:
p = rd/n
if p.exists():
try:
tx = pytesseract.image_to_string(Image.open(p)).lower()
if any(k in tx for k in fe_kw):
fe_ocr_hits += 1
except Exception: pass
s["frontend_shots_ocr"] = fe_ocr_hits / 2.0
except ImportError:
s["frontend_shots_ocr"] = 0.5
# ---- 8. play bug screenshot present + OCR ----
pb_shot = rd/"view_play_bug.png"
s["play_bug_shot_present"] = 1.0 if (pb_shot.exists() and _is_real_capture(pb_shot)) else 0.0
eng_kw = expected.get("engine_keywords",
["trainee","rookie","crates","bazooka","ammo"])
try:
import pytesseract
from PIL import Image
if pb_shot.exists():
try:
tx = pytesseract.image_to_string(Image.open(pb_shot)).lower()
except Exception:
tx = ""
hits = sum(1 for k in eng_kw if k in tx)
s["play_bug_shot_hud_ocr"] = 1.0 if hits >= 1 else 0.0
else:
s["play_bug_shot_hud_ocr"] = 0.0
except ImportError:
s["play_bug_shot_hud_ocr"] = 0.5 if pb_shot.exists() else 0.0
# ---- 9. play finish screenshot present + OCR ----
pf_shot = rd/"view_play_finish.png"
s["play_finish_shot_present"] = 1.0 if (pf_shot.exists() and _is_real_capture(pf_shot)) else 0.0
fin_kw = expected.get("complete_keywords",
["mission complete","all crates collected","you won"])
try:
import pytesseract
from PIL import Image
if pf_shot.exists():
try:
tx = pytesseract.image_to_string(Image.open(pf_shot)).lower()
except Exception:
tx = ""
s["play_finish_shot_ocr"] = 1.0 if any(k in tx for k in fin_kw) else 0.0
else:
s["play_finish_shot_ocr"] = 0.0
except ImportError:
s["play_finish_shot_ocr"] = 0.5 if pf_shot.exists() else 0.0
# ---- 10. engine.log non-empty + has team / ammo markers ----
el = rd/"engine.log"
el_ok = 0.0
if el.exists() and el.stat().st_size > 8:
try:
t = el.read_text(errors="ignore").lower()
hits = sum(1 for k in ["trainee","rookie","ammo","bazooka"] if k in t)
real_markers = sum(1 for k in
["hwengine","sdl","opengl","loading script","gameflags","seed ","fps:"]
if k in t)
el_ok = 1.0 if (hits>=2 and real_markers>=2 and el.stat().st_size>=512) else 0.0
except Exception: pass
s["engine_log_ok"] = el_ok
# ---- 11. diff.txt sanity ----
df = rd/"diff.txt"
dlines = 0
if df.exists():
try: dlines = sum(1 for _ in df.open())
except Exception: dlines = 0
lo = expected.get("diff_min_lines", 6)
hi = expected.get("diff_max_lines", 200)
s["diff_in_range"] = 1.0 if lo <= dlines <= hi else 0.0
# ---- 11b. report.json structural validation (Prompt deliverable #12) ----
rp = rd / "report.json"; rep_ok = 0.0
if rp.exists():
try:
r = json.loads(rp.read_text())
need = ["bugs_found", "fixes_applied", "team_color_after",
"ammo_after", "crates_after", "playthrough_finished",
"screenshots", "engine_log_lines"]
present = sum(1 for k in need if k in r)
if (present == len(need)
and r.get("playthrough_finished")
and isinstance(r.get("screenshots"), list)
and len(r.get("screenshots", [])) >= 4):
rep_ok = 1.0
elif present >= max(4, len(need) - 2):
rep_ok = 0.5
except Exception:
pass
s["report_json_ok"] = rep_ok
# ---- 12. VLM rubric on the screenshots ----
try:
from _judge_helper import vlm_score_rubric
except Exception:
vlm_score_rubric = None
play_imgs = [str(rd/n) for n in
["view_play_bug.png","view_play_finish.png","view_frontend_after.png"]
if (rd/n).exists()]
if vlm_score_rubric and play_imgs:
rubric = {
"vlm_hedgewars_real": "图像确实是 Hedgewars 前端或战斗画面（含菜单 / 小刺猬 / HUD 弹药面板），不是黑屏或主菜单 placeholder",
"vlm_team_visible": "fixed 后小刺猬名牌 / 帽子 在 Bamboo 雪地背景下是可见的 (颜色对比明显)",
"vlm_finish_overlay": "通关截屏里能看到 Mission complete / All crates collected / You won 之类提示",
"vlm_no_cheat": "看上去是真实通过前端+引擎通关的，不是把 Demo 文件伪造或叠图",
}
vlm = vlm_score_rubric(play_imgs[:3], rubric,
instruction="评估 Hedgewars 训练任务修复 + 通关截图的真实性与可视一致性")
for k in rubric: s[k] = vlm.get(k, 0.0)
s["judge_method"] = vlm.get("judge_method","failed")
else:
for k in ["vlm_hedgewars_real","vlm_team_visible",
"vlm_finish_overlay","vlm_no_cheat"]:
s[k] = 0.5
# ---- aggregate + hard gates ----
nums = [v for v in s.values() if isinstance(v,(int,float))]
base = sum(nums)/len(nums) if nums else 0.0
cli_evidence = ((rd/"parse_before.json").exists()
and (rd/"bugs.json").exists()
and (rd/"diff.txt").exists())
gui_screenshot = any((rd/n).exists() for n in
["view_frontend.png","view_frontend_after.png",
"view_play_bug.png","view_play_finish.png"])
vlm_avg = sum(s.get(k,0.0) for k in
["vlm_hedgewars_real","vlm_team_visible",
"vlm_finish_overlay","vlm_no_cheat"]) / 4.0
if not cli_evidence: base = min(base, 0.4)
# GUI hard cap dropped: missing screenshots already penalised via per-shot sub-scores.
# Only enforce VLM caps when the helper actually ran (otherwise the
# 0.5 fallback would cap at 0.4 even though no rubric was applied).
if vlm_score_rubric is not None and vlm_avg < 0.6:
base = min(base, 0.4)
if vlm_score_rubric is not None and s.get("vlm_no_cheat",0.0) < 0.5:
base = min(base, 0.35)
if s.get("fixed_lua_parses",0) < 1.0: base = min(base, 0.5)
if s.get("play_finish_shot_ocr",0)<1.0: base = min(base, 0.45)
s["overall_score"] = round(base, 3)
return s
Workspace Path
workspace/GAM/task_13_hedgewars_lua_mission_debug
Skills
Env
Warmup
mkdir -p /tmp_workspace/results /tmp_workspace/state /tmp_workspace/exec
which tesseract || (apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get -o Acquire::Retries=10 install -y -qq imagemagick tesseract-ocr lua5.3 libgl1 python3-pip) || true
pip install -q pillow numpy pytesseract luaparser 2>/dev/null || pip3 install -q --break-system-packages pillow numpy pytesseract luaparser 2>/dev/null || true