Spaces:
Running
Running
File size: 9,672 Bytes
b4ac377 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""Test that PythonCodeActEnv properly computes rewards via transform pipeline."""
import os
import sys
from pathlib import Path
import pytest
# Add the project root and src to the path
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))
sys.path.insert(0, str(Path(__file__).parent.parent.parent / "src"))
# Skip entire module if smolagents is not installed (optional dependency)
pytest.importorskip("smolagents", reason="smolagents is not installed")
from envs.coding_env.models import CodeAction
from envs.coding_env.server.python_codeact_env import PythonCodeActEnv
# ============================================================================
# Fixtures
# ============================================================================
@pytest.fixture
def env():
"""Provides a fresh PythonCodeActEnv for each test."""
environment = PythonCodeActEnv()
environment.reset()
return environment
@pytest.fixture
def env_with_variable(env):
"""Environment with a variable already defined."""
env.step(CodeAction(code="test_var = 42"))
return env
# ============================================================================
# Parametrized Tests - Reward Computation
# ============================================================================
@pytest.mark.parametrize(
"code,expected_reward,expected_exit_code,description",
[
# Safe + concise code
("x = 5", 0.1, 0, "safe + concise"),
("print('Hello')", 0.1, 0, "safe + concise print"),
("y = 10 + 5", 0.1, 0, "safe + concise calculation"),
# Safe + verbose code (>100 chars, no concise bonus)
("x = " + " + ".join(str(i) for i in range(50)), 0.0, 0, "safe + verbose"),
# Dangerous + concise (-1.0 safety + 0.1 concise = -0.9)
# NOTE: These actually fail at execution, so exit_code=1
("import os", -0.9, 1, "dangerous + concise"),
("eval('1+1')", -0.9, 1, "dangerous eval"),
("exec('x=1')", -0.9, 1, "dangerous exec"),
("with open('f.txt') as f: pass", -0.9, 1, "dangerous open"),
# Dangerous + verbose (-1.0 safety, no concise bonus)
("import os\n" + "x = 1\n" * 50, -1.0, 1, "dangerous + verbose"),
# Syntax error + concise (0.0 safe - 0.2 syntax + 0.1 concise = -0.1)
("print('unclosed", -0.1, 1, "syntax error + concise"),
# Syntax error + verbose (0.0 safe - 0.2 syntax = -0.2)
(
"x = " + " + ".join(str(i) for i in range(50)) + "\nprint('unclosed",
-0.2,
1,
"syntax error + verbose",
),
],
ids=lambda x: (
x if isinstance(x, str) and len(x) < 20 else None
), # Use description for test IDs
)
def test_reward_computation(
env, code, expected_reward, expected_exit_code, description
):
"""Test reward computation for various code patterns.
Parametrized test covering:
- Safe code (concise and verbose)
- Dangerous patterns (import os, eval, exec, open)
- Syntax errors
- Combinations of safety and quality transforms
Uses pytest.approx() for all float comparisons since rewards are computed
via floating point addition in the transform pipeline (transforms.py line 101).
"""
action = CodeAction(code=code)
obs = env.step(action)
assert obs.reward == pytest.approx(expected_reward, rel=1e-9), (
f"{description}: expected reward {expected_reward}, got {obs.reward}"
)
assert obs.exit_code == expected_exit_code, (
f"{description}: expected exit_code {expected_exit_code}, got {obs.exit_code}"
)
# ============================================================================
# Metadata Tests
# ============================================================================
def test_metadata_contains_last_code(env):
"""Test that step() includes executed code in observation metadata.
This is CRITICAL for the transform pipeline to evaluate code and assign rewards.
Without metadata["last_code"], transforms cannot access the code and rewards
will always be None.
"""
code = "print('Hello, World!')"
action = CodeAction(code=code)
obs = env.step(action)
assert "last_code" in obs.metadata, (
"metadata must contain 'last_code' for transform pipeline to evaluate code"
)
assert obs.metadata["last_code"] == code, (
f"metadata['last_code'] should be '{code}', got '{obs.metadata.get('last_code')}'"
)
@pytest.mark.parametrize(
"code,should_have_violation",
[
("import os", True),
("eval('1+1')", True),
("open('file.txt')", True),
("print('safe')", False),
("x = 1 + 2", False),
],
)
def test_metadata_safety_violations(env, code, should_have_violation):
"""Test that metadata correctly tracks safety violations."""
action = CodeAction(code=code)
obs = env.step(action)
assert "last_code" in obs.metadata
assert obs.metadata["last_code"] == code
if should_have_violation:
assert "safety_violation" in obs.metadata, (
f"Code '{code}' should have safety_violation in metadata"
)
else:
assert "safety_violation" not in obs.metadata, (
f"Code '{code}' should NOT have safety_violation in metadata"
)
# ============================================================================
# Consistency and State Tests
# ============================================================================
def test_reward_not_none_for_safe_code(env):
"""Test that safe code always receives a non-None reward."""
action = CodeAction(code="print('Hello')")
obs = env.step(action)
assert obs.reward is not None, "Safe code should receive a reward (not None)"
assert obs.exit_code == 0, "Safe code should execute successfully"
def test_reward_consistency_across_steps(env):
"""Test that rewards are computed consistently across multiple steps."""
for i in range(5):
action = CodeAction(code=f"x = {i}")
obs = env.step(action)
assert obs.reward is not None, f"Step {i}: Reward should not be None"
assert obs.reward == pytest.approx(0.1, rel=1e-9), (
f"Step {i}: Should get consistent 0.1 reward, got {obs.reward}"
)
def test_reset_preserves_transform_functionality(env):
"""Test that reset() doesn't break reward computation."""
# First episode
action1 = CodeAction(code="x = 1")
obs1 = env.step(action1)
assert obs1.reward == pytest.approx(0.1, rel=1e-9)
# Reset and start new episode
env.reset()
action2 = CodeAction(code="y = 2")
obs2 = env.step(action2)
assert obs2.reward == pytest.approx(0.1, rel=1e-9), (
"Reward computation should work after reset"
)
# ============================================================================
# Fixture Composition Tests
# ============================================================================
def test_using_composed_fixture(env_with_variable):
"""Test using an environment that builds on base fixture."""
action = CodeAction(code="print(test_var)")
obs = env_with_variable.step(action)
assert obs.exit_code == 0
assert "42" in obs.stdout
assert obs.reward == pytest.approx(0.1, rel=1e-9)
@pytest.mark.parametrize(
"code,expected_output",
[
("print(test_var)", "42"),
("print(test_var * 2)", "84"),
("print(test_var + 8)", "50"),
],
)
def test_fixture_with_parametrization(env_with_variable, code, expected_output):
"""Test combining fixtures with parametrization."""
action = CodeAction(code=code)
obs = env_with_variable.step(action)
assert obs.exit_code == 0
assert expected_output in obs.stdout
assert obs.reward == pytest.approx(0.1, rel=1e-9)
# ============================================================================
# Edge Cases and Special Patterns
# ============================================================================
@pytest.mark.parametrize(
"dangerous_pattern",
[
"import os",
"import subprocess",
"eval('x')",
"exec('x=1')",
"__import__('os')",
"open('file.txt')",
],
)
def test_all_dangerous_patterns_detected(env, dangerous_pattern):
"""Test that all dangerous patterns are correctly detected and penalized."""
action = CodeAction(code=dangerous_pattern)
obs = env.step(action)
# Concise dangerous code gets -0.9 (-1.0 safety + 0.1 concise)
assert obs.reward == pytest.approx(-0.9, rel=1e-9), (
f"Pattern '{dangerous_pattern}' should get -0.9 reward, got {obs.reward}"
)
assert "safety_violation" in obs.metadata
def test_multiline_code_with_mixed_patterns(env):
"""Test code with both safe and dangerous patterns (dangerous wins)."""
code = """
x = 5
y = 10
import os
z = x + y
"""
action = CodeAction(code=code)
obs = env.step(action)
# Should be flagged as dangerous even with safe code mixed in
assert obs.reward < 0, "Code with dangerous import should have negative reward"
assert "safety_violation" in obs.metadata
|