Spaces:
Sleeping
Sleeping
File size: 19,148 Bytes
1b35d41 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 | # 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.
"""
Code Migration Environment β full OpenEnv-compatible RL environment.
"""
from __future__ import annotations
import atexit
import math
import os
import re
import tempfile
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..dataset_loader import DatasetLoader, Task
from ..docker_sandbox import DockerSandbox
from ..models import CodeMigrationAction, CodeMigrationObservation
from ..prompt import MIGRATION_SYSTEM_PROMPT
from ..repo_manager import RepoManager
from ..tool_executor import ToolExecutor
except ImportError:
from dataset_loader import DatasetLoader, Task
from docker_sandbox import DockerSandbox
from models import CodeMigrationAction, CodeMigrationObservation
from prompt import MIGRATION_SYSTEM_PROMPT
from repo_manager import RepoManager
from tool_executor import ToolExecutor
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Test log parsing β extract pass/fail counts for continuous reward
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# pytest: "5 passed, 2 failed in 0.12s"
_RE_PYTEST_PASSED = re.compile(r"(\d+)\s+passed")
_RE_PYTEST_FAILED = re.compile(r"(\d+)\s+failed")
_RE_PYTEST_ERROR = re.compile(r"(\d+)\s+error")
# unittest: "Ran 7 tests" + "OK" or "FAILED (failures=2, errors=1)"
_RE_UNITTEST_RAN = re.compile(r"Ran\s+(\d+)\s+tests?")
_RE_UNITTEST_FAIL = re.compile(r"failures=(\d+)")
_RE_UNITTEST_ERR = re.compile(r"errors=(\d+)")
def _parse_test_pass_rate(log_text: str) -> float | None:
"""Extract the fraction of tests passing from a test log.
Returns a float in [0.0, 1.0] or None if we can't parse.
"""
if not log_text:
return None
# Try pytest format first
passed_m = _RE_PYTEST_PASSED.search(log_text)
failed_m = _RE_PYTEST_FAILED.search(log_text)
error_m = _RE_PYTEST_ERROR.search(log_text)
if passed_m:
passed = int(passed_m.group(1))
failed = int(failed_m.group(1)) if failed_m else 0
errors = int(error_m.group(1)) if error_m else 0
total = passed + failed + errors
if total > 0:
return passed / total
# Try unittest format
ran_m = _RE_UNITTEST_RAN.search(log_text)
if ran_m:
total = int(ran_m.group(1))
if total == 0:
return None
fail_m = _RE_UNITTEST_FAIL.search(log_text)
err_m = _RE_UNITTEST_ERR.search(log_text)
failures = int(fail_m.group(1)) if fail_m else 0
errors = int(err_m.group(1)) if err_m else 0
# If "OK" appears and no failures/errors, all passed
if "OK" in log_text and failures == 0 and errors == 0:
return 1.0
passed = total - failures - errors
return max(0.0, passed / total)
return None
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Reward function
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_reward(
*,
tool_name: str,
result_output: str,
result_patch: str | None,
test_exit_code: int | None,
test_log: str | None,
prev_pass_rate: float | None,
curr_pass_rate: float | None,
step_count: int,
max_steps: int,
is_limit_hit: bool,
) -> float:
"""Compute reward for a single step.
Design:
- Intermediate steps: ALWAYS positive (encourage exploration)
- Success (tests pass): large positive, with efficiency bonus
- Terminal failure (hit limit without solving): large negative
Intermediate reward scale (per step):
0.10 successful edit
0.08 test run that improved pass rate
0.05 test run (no improvement but informative)
0.04 found useful search matches
0.03 viewed file / gathered info
0.02 any other valid action
0.01 minimum (even failed actions get a tiny positive)
Terminal rewards:
+1.0 to +2.0 tests pass (higher = fewer steps used)
-1.0 hit step/test limit without passing
"""
# ββ Terminal: hit limits without solving ββ
if is_limit_hit:
return -3.0
# ββ Terminal: tests pass ββ
if tool_name == "execute_tests" and test_exit_code == 0:
# Big positive: 5.0 at step 1, down to 3.0 at max_steps
efficiency = 5.0 - 2.0 * (step_count / max(max_steps, 1))
return max(3.0, efficiency)
# ββ Intermediate: execute_tests (didn't pass) ββ
if tool_name == "execute_tests":
if curr_pass_rate is not None and prev_pass_rate is not None:
delta = curr_pass_rate - prev_pass_rate
if delta > 0:
# Improved pass rate β good signal
return 0.05 + delta * 0.3 # 0.05 to ~0.35
else:
# No improvement or regression β still positive but small
return 0.02
return 0.03 # ran tests, can't parse rate β still informative
# ββ Intermediate: successful edit ββ
if tool_name in ("edit_file", "replace_all_in_file"):
if result_patch:
return 0.10 # applied a real change
return 0.01 # edit refused but still a valid action
# ββ Intermediate: search found results ββ
if tool_name in ("search_file", "search_dir"):
if "match" in result_output.lower() and "no match" not in result_output.lower():
return 0.04 # found something useful
return 0.01 # searched, found nothing β still exploring
# ββ Intermediate: information gathering ββ
if tool_name in ("view_file", "view_last_log", "search_last_log"):
return 0.03
if tool_name == "list_dir":
return 0.02
if tool_name == "revert_last":
return 0.02
# ββ Fallback: any valid action ββ
return 0.01
class CodeMigrationEnvironment(Environment):
"""OpenEnv environment for Python code-migration tasks."""
SUPPORTS_CONCURRENT_SESSIONS: bool = False
def __init__(
self,
dataset_path: str | None = None,
max_steps: int = 200,
max_test_executions: int = 10,
container_timeout: int = 600,
container_memory_limit: str = "16g",
difficulty_filter: str | None = None,
) -> None:
self._loader = DatasetLoader(dataset_path)
if difficulty_filter:
tasks = self._loader.filter_by_difficulty(difficulty_filter)
self._loader._tasks = tasks
self._max_steps = max_steps
self._max_test_executions = max_test_executions
self._repo_manager = RepoManager()
self._sandbox = DockerSandbox(
timeout=container_timeout, memory_limit=container_memory_limit
)
self._tool_executor = ToolExecutor()
# Episode state
self._current_task: Task | None = None
self._workspace_dir: str | None = None
self._image_name: str | None = None
self._step_count: int = 0
self._patch_history: list[tuple[str, str]] = []
self._last_log_path: str | None = None
self._last_test_exit_code: int | None = None
self._last_pass_rate: float | None = None # continuous test pass rate
self._num_test_executions: int = 0
self._done: bool = False
self._task_index: int = 0
self._state = State(episode_id=str(uuid4()), step_count=0)
atexit.register(self._atexit_cleanup)
# ------------------------------------------------------------------
# reset
# ------------------------------------------------------------------
def reset(
self,
*,
task_index: int | None = None,
repo_name: str | None = None,
) -> CodeMigrationObservation:
"""Prepare a fresh workspace for the next (or specified) task."""
self._cleanup_episode()
# Select task
try:
if repo_name is not None:
task = self._loader.get_by_repo_name(repo_name)
if task is None:
return CodeMigrationObservation(
tool_output=f"No task found for repo_name: {repo_name}",
done=True,
)
elif task_index is not None:
task = self._loader[task_index]
else:
task = self._loader[self._task_index]
self._task_index = (self._task_index + 1) % len(self._loader)
self._current_task = task
except Exception as e:
return CodeMigrationObservation(
tool_output=f"Failed to select task: {e}", done=True,
)
# Setup workspace
try:
self._workspace_dir = self._repo_manager.setup_workspace(task)
except Exception as e:
return CodeMigrationObservation(
tool_output=f"Failed to setup workspace: {e}", done=True,
)
# Derive image name
escaped_name = task.repo_name.replace("/", "__").lower()
self._image_name = escaped_name + "_new"
# Create temp log file
tmp = tempfile.NamedTemporaryFile(mode="w", delete=False, suffix=".log")
self._last_log_path = tmp.name
tmp.close()
# Reset episode state
self._step_count = 0
self._patch_history = []
self._last_test_exit_code = None
self._last_pass_rate = None
self._num_test_executions = 0
self._done = False
self._state = State(episode_id=str(uuid4()), step_count=0)
# Initial test run
try:
test_result = self._sandbox.run_tests(self._image_name, self._workspace_dir)
self._num_test_executions += 1
if test_result.full_log is not None:
with open(self._last_log_path, "w", newline="") as f:
f.write(test_result.full_log)
self._last_test_exit_code = test_result.exit_code
self._last_pass_rate = _parse_test_pass_rate(test_result.full_log or "")
initial_test_output = (
"Test execution completed. Here is the test log.\n\n"
f"<test_log>\n{test_result.truncated_log}\n</test_log>"
)
except Exception as e:
initial_test_output = f"Initial test execution failed: {e}"
# Build system prompt
system_prompt = MIGRATION_SYSTEM_PROMPT.strip().format(
python_version=task.migration_target_version,
dependency_versions=task.dependency_versions,
)
combined_output = system_prompt + "\n\n" + initial_test_output
return CodeMigrationObservation(
tool_output=combined_output,
reward=0.0,
done=False,
metadata=self._build_metadata("reset"),
)
# ------------------------------------------------------------------
# step
# ------------------------------------------------------------------
def step(self, action: CodeMigrationAction) -> CodeMigrationObservation: # type: ignore[override]
"""Execute one tool call and return the observation."""
if self._done:
return CodeMigrationObservation(
tool_output="Episode is already done. Call reset() to start a new episode.",
reward=0.0, done=True,
metadata=self._build_metadata(action.tool_name),
)
self._step_count += 1
self._state.step_count = self._step_count
# Check step limit
if self._step_count > self._max_steps:
self._done = True
reward = compute_reward(
tool_name=action.tool_name, result_output="", result_patch=None,
test_exit_code=None, test_log=None,
prev_pass_rate=self._last_pass_rate, curr_pass_rate=self._last_pass_rate,
step_count=self._step_count, max_steps=self._max_steps,
is_limit_hit=True,
)
return CodeMigrationObservation(
tool_output=f"Step limit ({self._max_steps}) reached.",
reward=reward, done=True,
metadata=self._build_metadata(action.tool_name),
)
# Check test execution limit
if action.tool_name == "execute_tests":
if self._num_test_executions >= self._max_test_executions:
self._done = True
reward = compute_reward(
tool_name=action.tool_name, result_output="", result_patch=None,
test_exit_code=None, test_log=None,
prev_pass_rate=self._last_pass_rate, curr_pass_rate=self._last_pass_rate,
step_count=self._step_count, max_steps=self._max_steps,
is_limit_hit=True,
)
return CodeMigrationObservation(
tool_output=f"Test execution limit ({self._max_test_executions}) reached.",
reward=reward, done=True,
metadata=self._build_metadata(action.tool_name),
)
# Determine last_patch for revert
last_patch = self._patch_history[-1] if self._patch_history else None
# Dispatch tool
test_files = [
tf.strip()
for tf in (self._current_task.test_files or "").split(",")
if tf.strip()
]
result = self._tool_executor.execute(
tool_name=action.tool_name,
tool_args=action.tool_args,
host_repo_dir=self._workspace_dir,
repo_name=self._current_task.repo_name,
test_files=test_files,
image_name=self._image_name,
last_log_path=self._last_log_path,
last_patch=last_patch,
sandbox=self._sandbox,
)
# Track patches
if action.tool_name in ("edit_file", "replace_all_in_file") and result.patch:
file_path = action.tool_args.get("file_path", "")
self._patch_history.append((file_path, result.patch))
self._patch_history = self._patch_history[-5:]
# Handle revert
if action.tool_name == "revert_last" and last_patch is not None:
if "succeeded" in result.output.lower():
if self._patch_history:
self._patch_history.pop()
# Handle execute_tests β update state and parse pass rate
curr_pass_rate = self._last_pass_rate
if action.tool_name == "execute_tests":
self._num_test_executions += 1
if result.full_log is not None and self._last_log_path:
with open(self._last_log_path, "w", newline="") as f:
f.write(result.full_log)
self._last_test_exit_code = result.exit_code
curr_pass_rate = _parse_test_pass_rate(result.full_log or "")
# Compute continuous reward
reward = compute_reward(
tool_name=action.tool_name,
result_output=result.output,
result_patch=result.patch,
test_exit_code=result.exit_code if action.tool_name == "execute_tests" else None,
test_log=result.full_log if action.tool_name == "execute_tests" else None,
prev_pass_rate=self._last_pass_rate,
curr_pass_rate=curr_pass_rate,
step_count=self._step_count,
max_steps=self._max_steps,
is_limit_hit=False,
)
# Update pass rate after reward computation
if action.tool_name == "execute_tests" and curr_pass_rate is not None:
self._last_pass_rate = curr_pass_rate
# Check if tests passed
if action.tool_name == "execute_tests" and result.exit_code == 0:
self._done = True
# Check step limit
if self._step_count >= self._max_steps:
self._done = True
metadata = self._build_metadata(action.tool_name)
metadata["pass_rate"] = curr_pass_rate
metadata["prev_pass_rate"] = self._last_pass_rate
return CodeMigrationObservation(
tool_output=result.output,
reward=round(reward, 4),
done=self._done,
metadata=metadata,
)
# ------------------------------------------------------------------
# state
# ------------------------------------------------------------------
@property
def state(self) -> State:
meta: dict = {}
if self._current_task:
meta.update({
"repo_name": self._current_task.repo_name,
"difficulty": self._current_task.difficulty,
"test_type": self._current_task.test_type,
"test_count": self._current_task.test_count,
"num_test_executions": self._num_test_executions,
"last_test_exit_code": self._last_test_exit_code,
"last_pass_rate": self._last_pass_rate,
"migration_target_version": self._current_task.migration_target_version,
"reproduction_target_version": self._current_task.reproduction_target_version,
})
self._state.metadata = meta
return self._state
# ------------------------------------------------------------------
# helpers
# ------------------------------------------------------------------
def _build_metadata(self, tool_name: str) -> dict:
meta: dict = {
"step_count": self._step_count,
"tool_name": tool_name,
"last_test_exit_code": self._last_test_exit_code,
"num_test_executions": self._num_test_executions,
}
if self._current_task:
meta["repo_name"] = self._current_task.repo_name
meta["difficulty"] = self._current_task.difficulty
return meta
def _cleanup_episode(self) -> None:
if self._workspace_dir:
self._repo_manager.cleanup(self._workspace_dir)
self._workspace_dir = None
if self._last_log_path and os.path.exists(self._last_log_path):
try:
os.remove(self._last_log_path)
except Exception:
pass
self._last_log_path = None
def _atexit_cleanup(self) -> None:
try:
self._cleanup_episode()
except Exception:
pass
|