File size: 4,441 Bytes
ae9e249 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 | """Fuzz every tool of every task with hostile arguments; no call may raise, and every reward must be in [0, 1].
Usage: .venv/bin/python tests/fuzz_tools.py [--seeds 4] [--rounds 60]
"""
import argparse
import math
import random
import sys
import time
import traceback
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from diligence.environment import TASK_IDS, DiligenceEnvironment # noqa: E402
from diligence.models import ToolAction # noqa: E402
STRINGS = ["", " ", "\n", "None", "null", "-1", "0", "1e309", "nan", "../../etc/passwd", "/etc/passwd", "./", ".",
"x" * 20000, "\x00", "é漢字🙂", "'; DROP TABLE invoices; --", "{{}}", "%s%s%n", "\\", "e1", "E1",
"2026-10-05", "2026-02-30", "25:99", "09:00", "9:00", "09:00:00", "-09:00", "README.md", "DEPLOYMENT.md",
"Dana Whitfield", "CWE-89", "cwe-89 sql injection", "SELECT 1", "print(1)", "import os; os._exit(0)"]
INTS = [0, 1, -1, 2, 7, 13, 99, 10 ** 18, -(10 ** 18), 2 ** 63]
FLOATS = [0.0, -0.0, 1.5, float("inf"), float("-inf"), float("nan"), 1e308]
def value(rng, depth=0):
kind = rng.random()
if kind < 0.35:
return rng.choice(STRINGS)
if kind < 0.5:
return rng.choice(INTS)
if kind < 0.6:
return rng.choice(FLOATS)
if kind < 0.65:
return rng.choice([True, False, None])
if depth > 2:
return rng.choice(STRINGS)
if kind < 0.82:
return [value(rng, depth + 1) for _ in range(rng.randint(0, 4))]
keys = STRINGS[:8] + ["answer", "value", "file", "line", "cwe", "total_usd", "top_customer", "date", "start"]
return {str(rng.choice(keys)): value(rng, depth + 1) for _ in range(rng.randint(0, 4))}
def args_for(rng, tool):
"""Mostly well-typed arguments (so the tool body runs), sometimes extra, missing or mistyped ones."""
out = {}
for name, typ in tool.args.items():
if rng.random() < 0.08:
continue
v = value(rng)
if rng.random() < 0.8:
for _ in range(50):
if isinstance(v, typ) and not (typ in (int, float) and isinstance(v, bool)):
break
v = value(rng)
out[name] = v
if rng.random() < 0.05:
out[rng.choice(STRINGS[:5])] = value(rng)
return out
def check(obs, where):
assert isinstance(obs.output, str), where
assert isinstance(obs.done, bool), where
r = obs.reward
assert r is None or (isinstance(r, (int, float)) and math.isfinite(r) and 0 <= r <= 1), (where, r)
assert len(obs.output) <= 6100, (where, len(obs.output))
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--seeds", type=int, default=4)
ap.add_argument("--rounds", type=int, default=60)
args = ap.parse_args()
rng = random.Random(12345)
calls = failures = 0
slowest = (0.0, "")
for task_id in TASK_IDS:
for seed in range(args.seeds):
env = DiligenceEnvironment()
for _ in range(args.rounds):
if env._done:
check(env.reset(task_id=task_id, seed=seed), task_id)
tool = rng.choice(list(env._tools.values()))
if tool.name == "python" and rng.random() < 0.7:
continue # python is slow; it gets its own adversarial tests
action = {"tool": tool.name, "args": args_for(rng, tool)}
if rng.random() < 0.03:
action["tool"] = rng.choice(STRINGS[:6]) or "x"
where = f"{task_id} seed {seed} {str(action)[:300]}"
t0 = time.monotonic()
try:
check(env.step(ToolAction(**action)), where)
except Exception: # noqa: BLE001
failures += 1
print("FAIL", where)
traceback.print_exc(limit=3)
took = time.monotonic() - t0
slowest = max(slowest, (took, where[:160]))
calls += 1
# stepping after the episode is over must be harmless too
env._done = True
check(env.step(ToolAction(tool="submit", args={})), task_id)
print(f"{task_id}: done ({calls} calls so far, {failures} failures)", flush=True)
print(f"{calls} calls, {failures} failures; slowest {slowest[0]:.1f}s: {slowest[1]}")
sys.exit(1 if failures else 0)
if __name__ == "__main__":
main()
|