Spaces:
Sleeping
Sleeping
File size: 50,170 Bytes
ee933ab | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 | """V3 FlakeForge Environment — Unified step loop.
Key changes from V2:
- No hypothesis gating — agent outputs think+patch directly
- No judge calls — reward is fully deterministic
- Deep flakiness signals injected into observation
- Free-form patch application via search/replace
- Single-step flow: observe → generate → patch → run → reward
"""
from __future__ import annotations
import ast
import math
import os
import re
import runpy
import traceback
import uuid
import warnings
from collections import Counter
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
warnings.filterwarnings("ignore", category=SyntaxWarning)
from openenv.core.env_server.interfaces import Environment
try:
from models import (
FlakeForgeAction,
FlakeForgeObservation,
FlakeForgeState,
RunRecord,
PatchRecord,
RewardBreakdown,
failure_mode_entropy,
)
from server.state import EpisodeState
from server.deep_flakiness import (
build_deep_observation_signals,
extract_failure_frontier,
)
from server.patch_applier import restore_repo_files, write_validated_sources
from server.patch_validator import PatchValidator
from server.reward import compute_verifiable_reward
from server.oracle_engine import verify_structured_think
from server.causal_graph import CrossRepoGraphBuilder
try:
from server.tools import build_agent_targeting_hints
except ImportError:
def build_agent_targeting_hints(**_kwargs: Any) -> List[str]:
return []
from server.docker_runner import DockerTestRunner
except ImportError:
try:
from ..models import (
FlakeForgeAction,
FlakeForgeObservation,
FlakeForgeState,
RunRecord,
PatchRecord,
RewardBreakdown,
failure_mode_entropy,
)
from ..server.state import EpisodeState
from ..server.deep_flakiness import (
build_deep_observation_signals,
extract_failure_frontier,
)
from ..server.patch_applier import restore_repo_files, write_validated_sources
from ..server.patch_validator import PatchValidator
from ..server.reward import compute_verifiable_reward
from ..server.oracle_engine import verify_structured_think
from ..server.causal_graph import CrossRepoGraphBuilder
try:
from ..server.tools import build_agent_targeting_hints
except ImportError:
def build_agent_targeting_hints(**_kwargs: Any) -> List[str]:
return []
from ..server.docker_runner import DockerTestRunner
except (ImportError, ValueError):
from FlakeForge.models import (
FlakeForgeAction,
FlakeForgeObservation,
FlakeForgeState,
RunRecord,
PatchRecord,
RewardBreakdown,
failure_mode_entropy,
)
from FlakeForge.server.state import EpisodeState
from FlakeForge.server.deep_flakiness import (
build_deep_observation_signals,
extract_failure_frontier,
)
from FlakeForge.server.patch_applier import restore_repo_files, write_validated_sources
from FlakeForge.server.patch_validator import PatchValidator
from FlakeForge.server.reward import compute_verifiable_reward
from FlakeForge.server.oracle_engine import verify_structured_think
from FlakeForge.server.causal_graph import CrossRepoGraphBuilder
try:
from FlakeForge.server.tools import build_agent_targeting_hints
except ImportError:
def build_agent_targeting_hints(**_kwargs: Any) -> List[str]:
return []
from FlakeForge.server.docker_runner import DockerTestRunner
try:
from utils.logger import get_logger
except ImportError:
try:
from ..utils.logger import get_logger
except ImportError:
import logging
get_logger = lambda n, **kw: logging.getLogger(n)
logger = get_logger(__name__)
_INFRA_ERROR_TYPES = {
"ImportError",
"ModuleNotFoundError",
"SyntaxError",
"IndentationError",
"FileNotFoundError",
}
class FlakeForgeEnvironment(Environment[FlakeForgeAction, FlakeForgeObservation, FlakeForgeState]):
"""V3 RL environment for flaky test repair.
Unified step loop:
1. Build observation with deep flakiness signals
2. Agent generates <think> + <patch> in one forward pass
3. Apply search/replace patch atomically
4. Run test suite to verify fix
5. Compute 6-signal verifiable reward
6. Return observation with updated signals
"""
def __init__(
self,
repo_path: Optional[str] = None,
test_identifier: Optional[str] = None,
max_steps: int = 8,
num_runs: int = 10,
runner: Optional[Any] = None,
chaos_runner: Optional[Any] = None,
) -> None:
default_repo = os.environ.get("FF_REPO_PATH", str(Path("test_repos") / "timing_race_minimal"))
default_test = os.environ.get("FF_TEST_ID", "tests/test_flaky.py::test_fetch_should_complete")
self.repo_path = Path(repo_path or default_repo)
self.test_identifier = test_identifier or default_test
self.max_steps = max_steps
self.num_runs = num_runs
# Default to DockerTestRunner if no runner provided
self.runner = runner or DockerTestRunner(str(self.repo_path))
self.chaos_runner = chaos_runner
self._episode_state: Optional[EpisodeState] = None
self._openenv_state: Optional[FlakeForgeState] = None
# First successful reset() captures a full .py tree snapshot; every later reset() restores
# it so consecutive episodes (and GRPO group rollouts that reset each time) do not stack patches.
self._pristine_file_snapshots: Optional[Dict[str, str]] = None
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> FlakeForgeObservation:
"""Initialize a new episode."""
del seed
# Allow remote clients to configure env at reset-time.
if "repo_path" in kwargs and kwargs["repo_path"]:
new_repo = Path(str(kwargs["repo_path"]))
if self.repo_path.resolve() != new_repo.resolve():
self._pristine_file_snapshots = None # different tree — re-baseline on next line
self.repo_path = new_repo
if "test_identifier" in kwargs and kwargs["test_identifier"]:
self.test_identifier = str(kwargs["test_identifier"])
if "max_steps" in kwargs and kwargs["max_steps"] is not None:
self.max_steps = int(kwargs["max_steps"])
if "num_runs" in kwargs and kwargs["num_runs"] is not None:
self.num_runs = int(kwargs["num_runs"])
episode_id = episode_id or str(uuid.uuid4())[:8]
logger.info("[ENV] RESET episode=%s test=%s", episode_id, self.test_identifier)
if self._pristine_file_snapshots:
try:
restore_repo_files(self.repo_path, self._pristine_file_snapshots)
logger.info(
"[ENV] Restored %d .py file(s) from pristine snapshot",
len(self._pristine_file_snapshots),
)
except Exception as exc:
logger.warning("[ENV] Pristine restore failed (non-fatal): %s", exc)
self._reset_demo_repo_if_present()
# Read source files
test_source, source_under_test = self._read_sources()
file_tree = self._build_file_tree()
# Three-stage gate: Sanity → Determinism → Flakiness.
# skip_preflight=True reuses cached baseline from the previous reset
# (used during GRPO rollouts where flakiness was already confirmed).
skip_preflight = bool(kwargs.get("skip_preflight", False))
if skip_preflight and self._episode_state is not None:
baseline_runs = list(self._episode_state.run_history)
baseline_pass_rate = self._episode_state.baseline_pass_rate
baseline_entropy = self._episode_state.baseline_entropy
preflight = {
"runs": baseline_runs,
"pass_rate": baseline_pass_rate,
"failure_entropy": baseline_entropy,
"env_type": self._episode_state.env_type or "flaky",
"should_train": True,
"summary": dict(self._episode_state.preflight_result)
if self._episode_state.preflight_result else {},
}
else:
preflight = self._preflight_gate(
quick_runs=int(kwargs.get("preflight_quick_runs", 5)),
confirm_runs=int(kwargs.get("preflight_confirm_runs", 10)),
drop_deterministic_bugs=bool(kwargs.get("drop_deterministic_bugs", True)),
)
baseline_runs = preflight["runs"]
baseline_pass_rate = preflight["pass_rate"]
baseline_entropy = preflight["failure_entropy"]
# Extract failing stack trace
failing_trace = ""
last_error_type = None
for r in baseline_runs:
if not r.passed:
failing_trace = r.stderr_excerpt or r.error_message or ""
last_error_type = r.error_type
break
# Build deep flakiness signals (AST-based, <5ms)
deep_signals = build_deep_observation_signals(self.repo_path)
# Extract causal frontier from stack trace
failure_frontier, call_chain, boundary_crossings = extract_failure_frontier(
failing_trace, self.repo_path
)
# Check order dependency (reverse run)
order_dep = self._check_order_dependency(baseline_pass_rate)
# Check infrastructure sensitivity (chaos run)
infra_sensitive = self._check_infrastructure_sensitivity(baseline_pass_rate)
# Build causal graph
causal_graph_data, causal_hints = self._build_causal_graph(test_source)
# Build additional file-targeting hints from stack trace/imports/deep signals.
targeting_hints = build_agent_targeting_hints(
repo_path=str(self.repo_path),
test_identifier=self.test_identifier,
failing_stack_trace=failing_trace,
source_under_test=source_under_test,
causal_frontier=failure_frontier,
deep_signals=deep_signals,
max_hints=8,
)
merged_hints = list(dict.fromkeys([*causal_hints, *targeting_hints]))[:10]
# Initialize state
self._episode_state = EpisodeState(
episode_id=episode_id,
test_identifier=self.test_identifier,
repo_path=str(self.repo_path),
max_steps=self.max_steps,
original_test_source=test_source,
original_source_under_test=source_under_test,
current_test_source=test_source,
current_source_under_test=source_under_test,
run_history=baseline_runs,
baseline_pass_rate=baseline_pass_rate,
current_pass_rate=baseline_pass_rate,
baseline_entropy=baseline_entropy,
env_type=preflight["env_type"],
should_train=preflight["should_train"],
preflight_result=preflight["summary"],
failing_stack_trace=failing_trace,
last_error_type=last_error_type,
failure_frontier=failure_frontier,
call_chain_to_frontier=call_chain,
boundary_crossings=boundary_crossings,
order_dependency_detected=order_dep,
infrastructure_sensitive=infra_sensitive,
causal_graph=causal_graph_data,
causal_hints=merged_hints,
file_tree=file_tree,
**deep_signals,
)
observation = self._build_observation()
observation.reward = 0.0
observation.done = not preflight["should_train"]
self._openenv_state = FlakeForgeState(
episode_id=episode_id,
step_count=0,
done=not preflight["should_train"],
current_pass_rate=baseline_pass_rate,
baseline_pass_rate=baseline_pass_rate,
env_type=preflight["env_type"],
should_train=preflight["should_train"],
)
if not preflight["should_train"]:
self._episode_state.done = True
self._episode_state.last_done_reason = preflight["summary"]["reason"]
observation.done = True
observation.done_reason = self._episode_state.last_done_reason
if self._pristine_file_snapshots is None:
self._pristine_file_snapshots = dict(self._collect_sources())
logger.info(
"[ENV] Recorded pristine snapshot of %d .py file(s) for future resets",
len(self._pristine_file_snapshots),
)
return observation
def step(
self,
action: FlakeForgeAction,
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> FlakeForgeObservation:
"""Execute one step of the unified agent loop."""
del timeout_s, kwargs
if self._episode_state is None:
raise RuntimeError("Environment not initialized. Call reset() first.")
if self._episode_state.done and not self._episode_state.should_train:
observation = self._build_observation()
observation.done = True
observation.done_reason = self._episode_state.last_done_reason or "preflight_rejected"
return observation
self._episode_state.step_count += 1
logger.info(
"[ENV] STEP %d/%d category=%s",
self._episode_state.step_count,
self.max_steps,
action.predicted_category,
)
# --- 1. Patch validation → apply (disk unchanged if invalid) ---
# Snapshot sources *before* any write: oracle + PatchValidator simulation.
pre_sources: Dict[str, str] = {}
if action.patch_text.strip():
pre_sources = self._collect_sources()
patch_result: Dict[str, Any] = {
"success": False,
"error": "empty_patch",
"files_modified": [],
"lines_changed": 0,
"diff": "",
"rejected_by_validator": False,
"validation_errors": [],
"validation_warnings": [],
"validation_score": None,
}
rollback_snapshots: Dict[str, str] = {}
if action.patch_text.strip():
validator = PatchValidator()
validation = validator.validate(
action.patch_text,
repo_path=self.repo_path,
pre_sources=pre_sources or None,
claims=(
action.structured_think.claims
if action.structured_think is not None and action.structured_think.claims
else None
),
default_target=self._resolve_default_target(),
failure_frontier=self._episode_state.failure_frontier,
call_chain=self._episode_state.call_chain_to_frontier,
)
rollback_snapshots = dict(
validation.simulate_result.get("original_sources")
or validation.simulate_result.get("rollback_snapshots")
or {}
)
if not validation.is_valid:
first_error = validation.errors[0] if validation.errors else "patch_validation_failed"
patch_result = {
"success": False,
"error": first_error,
"validation_errors": list(validation.errors),
"validation_warnings": list(validation.warnings),
"validation_score": validation.score,
"files_modified": [],
"lines_changed": 0,
"diff": "",
"noop": False,
"protected_file": False,
"fuzzy_applied": False,
"rejected_by_validator": True,
}
logger.warning(
"[ENV] PATCH VALIDATION FAILED errors=%s warnings=%s",
validation.errors,
validation.warnings,
)
else:
sim = validation.simulate_result
try:
write_validated_sources(
self.repo_path,
dict(sim.get("modified_sources") or {}),
)
patch_result = {
"success": True,
"files_modified": list(sim.get("files_modified") or []),
"lines_changed": int(sim.get("lines_changed") or 0),
"hunks_applied": int(sim.get("hunks_applied") or 0),
"diff": sim.get("diff") or "",
"error": None,
"noop": bool(sim.get("noop", False)),
"protected_file": bool(sim.get("protected_file", False)),
"fuzzy_applied": bool(sim.get("fuzzy_applied", False)),
}
except Exception as exc:
if rollback_snapshots:
restore_repo_files(self.repo_path, rollback_snapshots)
patch_result = {
"success": False,
"files_modified": [],
"lines_changed": 0,
"hunks_applied": 0,
"diff": "",
"error": f"validated_write_failed: {exc}",
"noop": False,
"protected_file": False,
"fuzzy_applied": False,
"rolled_back": True,
}
patch_result["validation_errors"] = []
patch_result["validation_warnings"] = list(validation.warnings)
patch_result["validation_score"] = validation.score
patch_result["rejected_by_validator"] = False
logger.info(
"[ENV] PATCH success=%s files=%s lines=%d error=%s validation_score=%s",
patch_result["success"],
patch_result.get("files_modified", []),
patch_result.get("lines_changed", 0),
patch_result.get("error"),
validation.score,
)
# --- 2. Syntax check (sanity after apply); rollback if broken ---
syntax_error = None
if patch_result["success"]:
syntax_error = self._check_syntax(patch_result.get("files_modified", []))
patch_result["syntax_error"] = syntax_error
if syntax_error:
logger.warning("[ENV] SYNTAX ERROR after apply (rolling back): %s", syntax_error)
if rollback_snapshots:
try:
restore_repo_files(self.repo_path, rollback_snapshots)
except Exception as exc:
logger.error("[ENV] Rollback failed: %s", exc)
patch_result["success"] = False
patch_result["error"] = "syntax_error_after_apply"
patch_result["rolled_back"] = True
# --- 3. Run tests ---
post_runs: List[RunRecord] = []
post_run_dicts: List[Dict[str, Any]] = []
if patch_result["success"] and not syntax_error:
post_runs = self._run_tests(self.num_runs)
post_run_dicts = [
{"passed": r.passed, "error_type": r.error_type, "duration_ms": r.duration_ms}
for r in post_runs
]
post_pass_rate = sum(1 for r in post_runs if r.passed) / max(len(post_runs), 1)
else:
post_pass_rate = self._episode_state.current_pass_rate
# --- 4. Regression check ---
regression_detected = post_pass_rate < self._episode_state.baseline_pass_rate - 0.1
if patch_result["success"] and not syntax_error and self.runner is not None:
try:
if hasattr(self.runner, "check_regressions"):
regression_detected = regression_detected or bool(
self.runner.check_regressions(self.test_identifier)
)
except Exception as exc:
logger.debug("[ENV] Regression check failed: %s", exc)
patch_result["regression_detected"] = regression_detected
# --- 5. Compute reward ---
pre_entropy = failure_mode_entropy(self._episode_state.run_history[-self.num_runs:])
observation = self._build_observation() # Build before reward for causal proximity
# Oracle: verify structured claims against pre/post patch sources.
oracle_score: Optional[float] = None
if action.structured_think is not None and action.structured_think.claims:
post_sources = self._collect_sources() # current disk = post-patch
try:
annotated_think, oracle_score = verify_structured_think(
action.structured_think,
pre_sources=pre_sources,
post_sources=post_sources,
patch_hunks=(
action.structured_patch.hunks
if action.structured_patch is not None
else ()
),
)
action = action.model_copy(update={"structured_think": annotated_think})
logger.info("[ENV] ORACLE score=%.3f claims=%d", oracle_score, len(annotated_think.claims))
except Exception as exc:
logger.warning("[ENV] Oracle verification failed (non-fatal): %s", exc)
reward_breakdown = compute_verifiable_reward(
action=action,
observation=observation,
patch_result=patch_result,
post_run_results=post_run_dicts,
baseline_pass_rate=self._episode_state.baseline_pass_rate,
pre_entropy=pre_entropy,
oracle_score=oracle_score,
regression_detected=regression_detected,
think_history=self._episode_state.step_think_history,
)
# --- 6. Update state ---
self._episode_state.current_pass_rate = post_pass_rate
self._episode_state.run_history.extend(post_runs)
self._episode_state.last_think_text = action.think_text
self._episode_state.last_patch_text = action.patch_text
self._episode_state.last_reward = reward_breakdown.total_reward
self._episode_state.last_reward_breakdown = reward_breakdown.to_dict()
self._episode_state.last_patch_result = patch_result
# Build and store per-step think summary for diversity tracking.
think_summary = self._build_think_summary(
action=action,
oracle_score=oracle_score,
pass_rate_after=post_pass_rate,
reward=reward_breakdown.total_reward,
)
self._episode_state.step_think_history.append(think_summary)
if patch_result["success"]:
self._episode_state.patches_applied.append(PatchRecord(
patch_text=action.patch_text,
target_files=patch_result.get("files_modified", []),
lines_changed=patch_result.get("lines_changed", 0),
pass_rate_after=post_pass_rate,
applied_successfully=True,
))
self._episode_state.total_diff_lines += patch_result.get("lines_changed", 0)
# Check for regression
if regression_detected:
self._episode_state.regression_detected = True
# Re-read modified sources
self._episode_state.current_test_source, self._episode_state.current_source_under_test = self._read_sources()
# Determine if episode is terminal
done = (
post_pass_rate >= 1.0 # Full stability achieved
or self._episode_state.step_count >= self.max_steps
)
self._episode_state.done = done
self._episode_state.last_done_reason = self._done_reason(post_pass_rate, done)
# Build final observation with updated signals
final_observation = self._build_observation()
final_observation.reward = reward_breakdown.total_reward
final_observation.done = done
final_observation.patch_result = patch_result
final_observation.done_reason = self._episode_state.last_done_reason
self._openenv_state = FlakeForgeState(
episode_id=self._episode_state.episode_id,
step_count=self._episode_state.step_count,
done=done,
current_pass_rate=post_pass_rate,
baseline_pass_rate=self._episode_state.baseline_pass_rate,
regression_detected=self._episode_state.regression_detected,
env_type=self._episode_state.env_type,
should_train=self._episode_state.should_train,
)
logger.info(
"[ENV] REWARD total=%.4f breakdown=%s done=%s pass_rate=%.2f->%.2f",
reward_breakdown.total_reward,
{k: round(v, 3) for k, v in reward_breakdown.to_dict().items()},
done,
self._episode_state.baseline_pass_rate,
post_pass_rate,
)
return final_observation
def _build_observation(self) -> FlakeForgeObservation:
"""Build a V3 observation from current state."""
if self._episode_state is None:
raise RuntimeError("State not initialized")
# Duration fingerprint
durations = [r.duration_ms for r in self._episode_state.run_history[-self.num_runs:]]
dur_mean = sum(durations) / max(len(durations), 1) if durations else 0
dur_std = (
math.sqrt(sum((d - dur_mean) ** 2 for d in durations) / max(len(durations), 1))
if durations else 0
)
return FlakeForgeObservation(
episode_id=self._episode_state.episode_id,
test_identifier=self._episode_state.test_identifier,
step=self._episode_state.step_count,
steps_remaining=self._episode_state.steps_remaining,
test_function_source=self._episode_state.current_test_source,
source_under_test=self._episode_state.current_source_under_test,
relevant_imports=self._extract_imports(self._episode_state.current_test_source),
file_tree=self._episode_state.file_tree,
run_history=self._episode_state.run_history[-20:],
current_pass_rate=self._episode_state.current_pass_rate,
baseline_pass_rate=self._episode_state.baseline_pass_rate,
env_type=self._episode_state.env_type,
should_train=self._episode_state.should_train,
preflight_result=dict(self._episode_state.preflight_result),
patches_applied=self._episode_state.patches_applied,
total_diff_lines=self._episode_state.total_diff_lines,
# V3 deep signals
module_cache_violations=self._episode_state.module_cache_violations,
fixture_scope_risks=self._episode_state.fixture_scope_risks,
mock_residue_sites=self._episode_state.mock_residue_sites,
import_side_effect_files=self._episode_state.import_side_effect_files,
async_contamination_alive=self._episode_state.async_contamination_alive,
# Causal frontier
failure_frontier=self._episode_state.failure_frontier,
call_chain_to_frontier=self._episode_state.call_chain_to_frontier,
boundary_crossings=self._episode_state.boundary_crossings,
# iDFlakies
order_dependency_detected=self._episode_state.order_dependency_detected,
infrastructure_sensitive=self._episode_state.infrastructure_sensitive,
# Causal graph
causal_graph=self._episode_state.causal_graph,
causal_hints=self._episode_state.causal_hints,
# Failure analysis
failing_stack_trace=self._episode_state.failing_stack_trace,
duration_fingerprint={"mean": dur_mean, "std": dur_std},
# Episode context
last_think_text=self._episode_state.last_think_text,
last_patch_text=self._episode_state.last_patch_text,
last_reward=self._episode_state.last_reward,
reward_breakdown=self._episode_state.last_reward_breakdown,
patch_result=self._episode_state.last_patch_result,
done_reason=self._episode_state.last_done_reason,
reward=self._episode_state.last_reward,
think_history=list(self._episode_state.step_think_history),
)
def _build_think_summary(
self,
action: FlakeForgeAction,
oracle_score: Optional[float],
pass_rate_after: float,
reward: float,
) -> Dict[str, Any]:
"""Build a compact think summary dict for history tracking."""
categories: List[str] = []
entities: List[str] = []
reason_signatures: List[str] = []
if action.structured_think and action.structured_think.claims:
for claim in action.structured_think.claims:
categories.append(claim.category)
if claim.entity:
entities.append(claim.entity)
if claim.reason:
reason_signatures.append(claim.reason[:35].lower().strip())
elif action.predicted_category:
categories = [action.predicted_category]
return {
"step": self._episode_state.step_count,
"categories": categories,
"entities": entities,
"reason_signatures": reason_signatures,
"oracle_score": round(oracle_score, 3) if oracle_score is not None else None,
"pass_rate_after": round(pass_rate_after, 3),
"reward": round(reward, 4),
}
def _preflight_gate(
self,
*,
quick_runs: int = 10,
confirm_runs: int = 20,
drop_deterministic_bugs: bool = True,
) -> Dict[str, Any]:
"""Classify environment before training: sanity → determinism → flakiness.
Pass rate alone is not enough: 0/N can be a deterministic bug, infra
breakage, or a hard flaky case where success was not observed yet.
This gate separates those cases using error consistency/entropy.
"""
quick_runs = max(1, quick_runs)
confirm_runs = max(1, confirm_runs)
sanity = self._run_tests(1)
runs: List[RunRecord] = list(sanity)
sanity_record = sanity[0] if sanity else RunRecord(
passed=False,
duration_ms=0,
error_type="RunnerError",
error_message="runner returned no result",
)
if self._is_infra_failure(sanity_record):
return self._preflight_result(
runs=runs,
env_type="infra_broken",
should_train=False,
reason="preflight_infra_broken",
stage="sanity",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
# Stage 2: cheap deterministic baseline. We already ran one sanity pass.
remaining_quick = max(quick_runs - len(runs), 0)
if remaining_quick:
runs.extend(self._run_tests(remaining_quick))
quick_passes = sum(1 for r in runs[:quick_runs] if r.passed)
if quick_passes == quick_runs:
return self._preflight_result(
runs=runs[:quick_runs],
env_type="stable",
should_train=False,
reason="preflight_stable_pass",
stage="determinism",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
if 0 < quick_passes < quick_runs:
return self._preflight_result(
runs=runs[:quick_runs],
env_type="flaky",
should_train=True,
reason="preflight_mixed_pass_fail",
stage="determinism",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
# Stage 3: 0/N quick passes. Do not drop yet; confirm failure type.
runs.extend(self._run_tests(confirm_runs))
confirm_window = runs[quick_runs:quick_runs + confirm_runs]
confirm_passes = sum(1 for r in confirm_window if r.passed)
if confirm_passes > 0:
return self._preflight_result(
runs=runs,
env_type="flaky",
should_train=True,
reason="preflight_late_success_after_zero_quick_passes",
stage="flakiness_confirm",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
failure_keys = self._failure_keys(runs)
unique_failures = set(failure_keys)
if len(unique_failures) <= 1:
return self._preflight_result(
runs=runs,
env_type="deterministic_bug",
should_train=not drop_deterministic_bugs,
reason=(
"preflight_deterministic_bug_dropped"
if drop_deterministic_bugs
else "preflight_deterministic_bug_labeled"
),
stage="flakiness_confirm",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
return self._preflight_result(
runs=runs,
env_type="deterministic_bug",
should_train=not drop_deterministic_bugs,
reason=(
"preflight_deterministic_bug_dropped"
if drop_deterministic_bugs
else "preflight_deterministic_multi_error_labeled"
),
stage="flakiness_confirm",
quick_runs=quick_runs,
confirm_runs=confirm_runs,
)
def _preflight_result(
self,
*,
runs: List[RunRecord],
env_type: str,
should_train: bool,
reason: str,
stage: str,
quick_runs: int,
confirm_runs: int,
) -> Dict[str, Any]:
pass_count = sum(1 for r in runs if r.passed)
pass_rate = pass_count / max(len(runs), 1)
failure_keys = self._failure_keys(runs)
error_distribution = dict(Counter(failure_keys))
unique_failure_types = len(error_distribution)
entropy = failure_mode_entropy(runs)
summary = {
"env_type": env_type,
"should_train": should_train,
"reason": reason,
"stage": stage,
"runs": len(runs),
"passes": pass_count,
"pass_rate": round(pass_rate, 4),
"quick_runs": quick_runs,
"confirm_runs": confirm_runs,
"unique_failure_types": unique_failure_types,
"failure_entropy": entropy,
"error_distribution": error_distribution,
}
logger.info("[ENV] PREFLIGHT %s", summary)
return {
"runs": runs,
"pass_rate": pass_rate,
"failure_entropy": entropy,
"env_type": env_type,
"should_train": should_train,
"summary": summary,
}
_TIMING_RE = re.compile(r"\b\d+\.\d+s\b")
def _failure_keys(self, runs: List[RunRecord]) -> List[str]:
keys: List[str] = []
for r in runs:
if r.passed:
continue
error_type = r.error_type or "UnknownError"
message = (r.error_message or r.stderr_excerpt or "").strip()
msg_sig = message[:80] if message else ""
# Strip wall-clock durations ("0.17s", "1.23s") from the signature
# so that the same deterministic failure with slightly different
# timing doesn't inflate failure-entropy into a false "flaky" verdict.
msg_sig = self._TIMING_RE.sub("Xs", msg_sig)
keys.append(f"{error_type}:{msg_sig}")
return keys
def _is_infra_failure(self, run: RunRecord) -> bool:
if run.passed:
return False
error_type = run.error_type or ""
message = f"{run.error_message or ''}\n{run.stderr_excerpt or ''}"
if error_type in _INFRA_ERROR_TYPES:
return True
infra_needles = (
"ImportError",
"ModuleNotFoundError",
"SyntaxError",
"IndentationError",
"ERROR collecting",
"collected 0 items",
"fixture",
"pytest timed out",
)
return any(needle in message for needle in infra_needles)
def _run_tests(self, n: int) -> List[RunRecord]:
"""Run the target test n times, collecting results."""
if self.runner is None:
logger.warning("[ENV] No runner configured — returning synthetic runs")
return self._synthetic_runs(n)
try:
if hasattr(self.runner, "run_test_n_times"):
batch = self.runner.run_test_n_times(self.test_identifier, n)
if isinstance(batch, list) and batch:
return batch
except Exception as exc:
logger.debug("[ENV] run_test_n_times failed, falling back to loop: %s", exc)
results: List[RunRecord] = []
for _ in range(n):
try:
result = self.runner.run_test(self.test_identifier)
results.append(result)
except Exception as exc:
results.append(RunRecord(
passed=False,
duration_ms=0,
error_type=type(exc).__name__,
error_message=str(exc)[:200],
stderr_excerpt=None,
))
return results
def _synthetic_runs(self, n: int) -> List[RunRecord]:
"""Generate synthetic run results for development/testing."""
import random
from server.docker_runner import RunRecord
results = []
for _ in range(n):
passed = random.random() > 0.5
results.append(RunRecord(
passed=passed,
duration_ms=random.randint(10, 500),
error_type=None if passed else "TimeoutError",
error_message=None if passed else "Operation timed out",
stderr_excerpt=None,
))
return results
def _read_sources(self) -> Tuple[str, str]:
"""Read test and source-under-test files."""
test_source = ""
source_under_test = ""
# Find test file
test_parts = self.test_identifier.split("::")
test_file_hint = test_parts[0] if test_parts else ""
if test_file_hint:
test_path = self.repo_path / test_file_hint
if test_path.exists():
try:
test_source = test_path.read_text(encoding="utf-8", errors="ignore")[:8000]
except Exception:
pass
# Try to find source under test from imports
if test_source:
for candidate in self._source_candidates_from_test(test_source, test_file_hint):
if candidate.exists() and candidate.is_file():
try:
source_under_test = candidate.read_text(encoding="utf-8", errors="ignore")[:8000]
break
except Exception:
pass
return test_source, source_under_test
def _source_candidates_from_test(self, test_source: str, test_file_hint: str) -> List[Path]:
"""Resolve likely source files from imports in the target test."""
candidates: List[Path] = []
test_dir = (self.repo_path / test_file_hint).parent if test_file_hint else self.repo_path
def add_module_candidates(module: str) -> None:
if not module:
return
parts = module.replace(".", "/")
candidates.extend([
self.repo_path / f"{parts}.py",
self.repo_path / parts / "__init__.py",
self.repo_path / "src" / f"{parts}.py",
self.repo_path / "src" / parts / "__init__.py",
])
try:
tree = ast.parse(test_source)
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for alias in node.names:
add_module_candidates(alias.name)
elif isinstance(node, ast.ImportFrom):
base_module = node.module or ""
if node.level:
base_dir = test_dir
for _ in range(max(node.level - 1, 0)):
base_dir = base_dir.parent
if base_module:
candidates.extend([
base_dir / f"{base_module.replace('.', '/')}.py",
base_dir / base_module.replace(".", "/") / "__init__.py",
])
for alias in node.names:
if alias.name != "*":
candidates.append(base_dir / f"{alias.name}.py")
else:
add_module_candidates(base_module)
for alias in node.names:
if alias.name != "*":
add_module_candidates(f"{base_module}.{alias.name}" if base_module else alias.name)
except Exception:
for imp in self._extract_imports(test_source):
add_module_candidates(imp)
seen = set()
unique_candidates: List[Path] = []
for candidate in candidates:
try:
resolved = candidate.resolve()
except Exception:
resolved = candidate
if resolved in seen:
continue
seen.add(resolved)
unique_candidates.append(candidate)
return unique_candidates
def _build_file_tree(self) -> List[str]:
"""Build a compact file tree of the repo."""
tree: List[str] = []
skip = {"__pycache__", ".git", "node_modules", "venv", ".venv", ".pytest_cache", ".tox"}
for root, dirs, files in os.walk(self.repo_path):
dirs[:] = [d for d in dirs if d not in skip]
rel = os.path.relpath(root, self.repo_path)
depth = rel.count(os.sep)
if depth > 3:
continue
for f in files:
if f.endswith(".py"):
tree.append(os.path.join(rel, f).replace("\\", "/"))
return sorted(tree)[:50]
def _extract_imports(self, source: str) -> List[str]:
"""Extract import statements from source."""
imports = []
try:
tree = ast.parse(source)
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for alias in node.names:
imports.append(alias.name)
elif isinstance(node, ast.ImportFrom) and node.module:
imports.append(node.module)
except Exception:
pass
return imports
def _check_syntax(self, files: List[str]) -> Optional[str]:
"""Check that modified files have valid Python syntax."""
for f in files:
path = self.repo_path / f
if not path.exists() or path.suffix != ".py":
continue
try:
source = path.read_text(encoding="utf-8", errors="ignore")
ast.parse(source)
except SyntaxError as exc:
return f"{path.name}:{exc.lineno}: {exc.msg}"
return None
def _check_order_dependency(self, baseline_rate: float) -> bool:
"""Check for order dependency by running in reverse order."""
if self.runner is None:
return False
try:
if hasattr(self.runner, "run_reversed"):
result = self.runner.run_reversed(self.test_identifier)
reverse_rate = result.get("pass_rate", baseline_rate)
return abs(reverse_rate - baseline_rate) > 0.15
except Exception:
pass
return False
def _check_infrastructure_sensitivity(self, baseline_rate: float) -> bool:
"""Check if the test is sensitive to infrastructure pressure."""
if self.chaos_runner is None:
return False
try:
result = self.chaos_runner.run_single(self.test_identifier)
chaos_rate = 1.0 if result.get("passed", False) else 0.0
return abs(chaos_rate - baseline_rate) > 0.2
except Exception:
pass
return False
def _build_causal_graph(self, test_source: str) -> Tuple[Optional[Dict], List[str]]:
"""Build causal graph for the test."""
del test_source
try:
test_file, _, test_func = self.test_identifier.partition("::")
entry_file = str(self.repo_path / test_file)
entry_function = test_func or ""
if not entry_function:
return None, []
builder = CrossRepoGraphBuilder(str(self.repo_path), max_depth=3)
graph = builder.build(entry_file=entry_file, entry_function=entry_function)
graph_dict = graph.to_observation_dict()
hints = list(graph_dict.get("boundary_warnings", []))[:5]
return graph_dict, hints
except Exception as exc:
logger.debug("[ENV] Causal graph construction failed: %s", exc)
return None, []
def _resolve_default_target(self) -> str:
"""Resolve the default target file for patches."""
test_parts = self.test_identifier.split("::")
return test_parts[0] if test_parts else ""
def _collect_sources(self) -> Dict[str, str]:
"""Collect current on-disk source texts keyed by path relative to repo_path.
Used to feed the oracle engine with pre/post source snapshots.
This is called *before* a patch is applied for pre-sources and *after*
for post-sources, so it just reads the current disk state.
"""
sources: Dict[str, str] = {}
if self._episode_state is None:
return sources
skip = {"__pycache__", ".git", "node_modules", "venv", ".venv", ".pytest_cache"}
for root, dirs, files in os.walk(self.repo_path):
dirs[:] = [d for d in dirs if d not in skip]
for f in files:
if not f.endswith(".py"):
continue
full = Path(root) / f
rel = str(full.relative_to(self.repo_path)).replace("\\", "/")
try:
sources[rel] = full.read_text(encoding="utf-8", errors="ignore")
except Exception:
pass
return sources
def _done_reason(self, pass_rate: float, done: bool) -> str:
if not done:
return "in_progress"
if pass_rate >= 1.0:
return "fully_stable"
if self._episode_state and self._episode_state.regression_detected:
return "regression_detected"
if self._episode_state and self._episode_state.step_count >= self.max_steps:
return "max_steps_reached"
return "unknown"
def _reset_demo_repo_if_present(self) -> None:
"""Restore bundled demo repos before each episode when they provide a reset script."""
if os.environ.get("FF_SKIP_DEMO_RESET", "0") == "1":
return
reset_script = self.repo_path / "reset_demo.py"
if not reset_script.exists():
return
try:
runpy.run_path(str(reset_script), run_name="__main__")
except Exception as exc:
logger.warning("[ENV] Demo reset script failed: %s", exc)
@property
def state(self) -> FlakeForgeState:
if self._openenv_state is not None:
return self._openenv_state
if self._episode_state is not None:
return FlakeForgeState(
episode_id=self._episode_state.episode_id,
step_count=self._episode_state.step_count,
done=self._episode_state.done,
current_pass_rate=self._episode_state.current_pass_rate,
baseline_pass_rate=self._episode_state.baseline_pass_rate,
regression_detected=self._episode_state.regression_detected,
env_type=self._episode_state.env_type,
should_train=self._episode_state.should_train,
)
return FlakeForgeState(
episode_id="",
step_count=0,
done=False,
current_pass_rate=0.0,
baseline_pass_rate=0.0,
regression_detected=False,
env_type="unknown",
should_train=True,
)
# ── Factory for OpenEnv ──────────────────────────────────────────────────────
def create_flakeforge_environment(
repo_path: str,
test_identifier: str,
**kwargs: Any,
) -> FlakeForgeEnvironment:
"""Create a FlakeForge V3 environment instance."""
return FlakeForgeEnvironment(
repo_path=repo_path,
test_identifier=test_identifier,
**kwargs,
) |