| |
| """Run one authenticated official SWE-bench verification phase in Docker.""" |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| from collections.abc import Mapping, Sequence |
| from dataclasses import dataclass |
| import hashlib |
| import importlib |
| import importlib.util |
| import json |
| import math |
| import os |
| from pathlib import Path, PurePosixPath |
| import re |
| import secrets |
| import shutil |
| import signal |
| import stat |
| import subprocess |
| import sys |
| import time |
| from typing import Any |
|
|
|
|
| PHASES = frozenset({"red", "green", "scored"}) |
| PASSING_STATUSES = frozenset({"PASSED", "XFAIL"}) |
| RED_FAILURE_STATUSES = frozenset({"FAILED", "ERROR"}) |
| KNOWN_STATUSES = PASSING_STATUSES | RED_FAILURE_STATUSES | {"SKIPPED"} |
| REVISION_PATTERN = re.compile(r"^[0-9a-f]{40}$") |
| SHA256_PATTERN = re.compile(r"^[0-9a-f]{64}$") |
| INSTANCE_PATTERN = re.compile(r"^[A-Za-z0-9_.-]{1,200}$") |
| RUN_ID_PATTERN = re.compile(r"^[a-z0-9][a-z0-9-]{0,79}$") |
| CONTAINER_ID_PATTERN = re.compile(r"^[0-9a-f]{12,64}$") |
| IMAGE_ID_PATTERN = re.compile(r"^sha256:[0-9a-f]{64}$") |
| SAFE_PATH_PATTERN = re.compile(r"^[A-Za-z0-9._/-]+$") |
| TEST_PATH_COMPONENTS = frozenset({"test", "tests", "testing", "__tests__"}) |
| CONTAINER_MEMORY_BYTES = 8 * 1024 * 1024 * 1024 |
| CONTAINER_NANO_CPUS = 4_000_000_000 |
| CONTAINER_PIDS_LIMIT = 2048 |
| CONTAINER_SECURITY_OPT = "no-new-privileges:true" |
| RUN_LABEL = "ctx.benchmark.run_id" |
| PROCESS_MARKER = "CTX_SWEBENCH_PROCESS_TOKEN" |
| REQUIRED_ARTIFACTS = ( |
| "report.json", |
| "raw-status.json", |
| "eval.sh", |
| "test_output.txt", |
| "run_instance.log", |
| "patch.diff", |
| ) |
| MAX_EVIDENCE_BYTES = 256 * 1024 * 1024 |
| SCRIPT_PATH = Path(__file__).resolve() |
|
|
|
|
| class SWEbenchVerificationError(RuntimeError): |
| """The official SWE-bench bridge failed closed.""" |
|
|
| def __init__( |
| self, |
| message: str = "SWE-bench verification failed", |
| *, |
| evidence: Mapping[str, Any] | None = None, |
| ) -> None: |
| super().__init__(message) |
| self.evidence = dict(evidence or {}) |
|
|
|
|
| @dataclass(frozen=True) |
| class CommandResult: |
| returncode: int |
| stdout: str |
| stderr: str |
| elapsed: float |
| timed_out: bool = False |
| reaped_descendants: int = 0 |
| residual_descendants: tuple[int, ...] = () |
|
|
|
|
| def _descendant_pids(root_pid: int) -> list[int]: |
| try: |
| result = subprocess.run( |
| ["/bin/ps", "-axo", "pid=,ppid="], |
| check=False, |
| capture_output=True, |
| text=True, |
| timeout=3, |
| ) |
| except (OSError, subprocess.SubprocessError): |
| return [] |
| children: dict[int, list[int]] = {} |
| for line in result.stdout.splitlines(): |
| try: |
| pid_text, parent_text = line.split() |
| children.setdefault(int(parent_text), []).append(int(pid_text)) |
| except (ValueError, TypeError): |
| continue |
| found: list[int] = [] |
| pending = [root_pid] |
| while pending: |
| current = pending.pop() |
| direct = children.get(current, []) |
| found.extend(direct) |
| pending.extend(direct) |
| return found |
|
|
|
|
| def _signal_process_tree(process: subprocess.Popen[str], sig: signal.Signals) -> None: |
| descendants = _descendant_pids(process.pid) |
| try: |
| os.killpg(process.pid, sig) |
| except (ProcessLookupError, PermissionError): |
| pass |
| for pid in reversed(descendants): |
| try: |
| os.kill(pid, sig) |
| except (ProcessLookupError, PermissionError): |
| pass |
|
|
|
|
| def _terminate_process_tree(process: subprocess.Popen[str]) -> tuple[str, str]: |
| _signal_process_tree(process, signal.SIGTERM) |
| try: |
| return process.communicate(timeout=5) |
| except subprocess.TimeoutExpired as first: |
| _signal_process_tree(process, getattr(signal, "SIGKILL", signal.SIGTERM)) |
| try: |
| return process.communicate(timeout=5) |
| except subprocess.TimeoutExpired: |
| for stream in (process.stdout, process.stderr): |
| if stream is not None: |
| stream.close() |
| try: |
| process.wait(timeout=2) |
| except subprocess.TimeoutExpired: |
| pass |
| stdout = first.stdout if isinstance(first.stdout, str) else "" |
| stderr = first.stderr if isinstance(first.stderr, str) else "" |
| return stdout, stderr + "\nprocess tree did not reap within 12 seconds" |
|
|
|
|
| def _marked_process_pids(token: str) -> tuple[set[int], str | None]: |
| try: |
| result = subprocess.run( |
| ["/bin/ps", "eww", "-A", "-o", "pid=", "-o", "args="], |
| check=False, |
| capture_output=True, |
| text=True, |
| timeout=5, |
| ) |
| except (OSError, subprocess.SubprocessError) as exc: |
| return set(), f"process marker scan failed: {exc}" |
| if result.returncode: |
| return set(), f"process marker scan exited {result.returncode}" |
| marker = f"{PROCESS_MARKER}={token}" |
| pids: set[int] = set() |
| for line in result.stdout.splitlines(): |
| fields = line.strip().split(maxsplit=1) |
| if len(fields) != 2 or marker not in fields[1]: |
| continue |
| try: |
| pid = int(fields[0]) |
| except ValueError: |
| continue |
| if pid != os.getpid(): |
| pids.add(pid) |
| return pids, None |
|
|
|
|
| def _cleanup_marked_processes(token: str) -> tuple[int, tuple[int, ...], str | None]: |
| pids, error = _marked_process_pids(token) |
| if error: |
| return 0, (), error |
| signaled = set(pids) |
| for pid in pids: |
| try: |
| os.kill(pid, signal.SIGTERM) |
| except (ProcessLookupError, PermissionError): |
| pass |
| deadline = time.monotonic() + 2 |
| while pids and time.monotonic() < deadline: |
| time.sleep(0.05) |
| pids, error = _marked_process_pids(token) |
| if error: |
| return len(signaled), (), error |
| signaled.update(pids) |
| for pid in pids: |
| try: |
| os.kill(pid, getattr(signal, "SIGKILL", signal.SIGTERM)) |
| except (ProcessLookupError, PermissionError): |
| pass |
| if pids: |
| time.sleep(0.05) |
| residual, error = _marked_process_pids(token) |
| return len(signaled), tuple(sorted(residual)), error |
|
|
|
|
| def _run_process( |
| argv: list[str], |
| *, |
| cwd: Path, |
| env: dict[str, str] | None = None, |
| timeout: float = 600, |
| input_text: str | None = None, |
| contain_descendants: bool = False, |
| ) -> CommandResult: |
| started = time.perf_counter() |
| process_token = secrets.token_hex(16) if contain_descendants else None |
| child_env = env |
| if process_token is not None: |
| child_env = dict(os.environ if env is None else env) |
| child_env[PROCESS_MARKER] = process_token |
| process = subprocess.Popen( |
| argv, |
| cwd=cwd, |
| env=child_env, |
| text=True, |
| stdin=subprocess.PIPE if input_text is not None else subprocess.DEVNULL, |
| stdout=subprocess.PIPE, |
| stderr=subprocess.PIPE, |
| start_new_session=True, |
| ) |
| timed_out = False |
| reaped = 0 |
| residual: tuple[int, ...] = () |
| cleanup_error: str | None = None |
|
|
| def cleanup_marked_processes() -> None: |
| nonlocal reaped, residual, cleanup_error |
| if process_token is None: |
| return |
| count, remaining, error = _cleanup_marked_processes(process_token) |
| reaped += count |
| residual = remaining |
| cleanup_error = cleanup_error or error |
|
|
| try: |
| stdout, stderr = process.communicate(input=input_text, timeout=timeout) |
| except subprocess.TimeoutExpired: |
| timed_out = True |
| cleanup_marked_processes() |
| stdout, stderr = _terminate_process_tree(process) |
| except BaseException: |
| cleanup_marked_processes() |
| _terminate_process_tree(process) |
| raise |
| finally: |
| cleanup_marked_processes() |
| returncode = 124 if timed_out else (process.returncode or 0) |
| if cleanup_error or residual: |
| returncode = returncode or 125 |
| detail = cleanup_error or f"residual descendants: {list(residual)}" |
| stderr = f"{stderr}\nprocess containment failed: {detail}".lstrip() |
| return CommandResult( |
| returncode, |
| stdout, |
| stderr, |
| time.perf_counter() - started, |
| timed_out=timed_out, |
| reaped_descendants=reaped, |
| residual_descendants=residual, |
| ) |
|
|
|
|
| def _canonical_bytes(value: Any) -> bytes: |
| return json.dumps( |
| value, |
| ensure_ascii=False, |
| separators=(",", ":"), |
| sort_keys=True, |
| ).encode("utf-8") |
|
|
|
|
| def _sha256(data: bytes) -> str: |
| return hashlib.sha256(data).hexdigest() |
|
|
|
|
| def _remaining(deadline: float) -> float: |
| remaining = deadline - time.monotonic() |
| if remaining <= 0: |
| raise SWEbenchVerificationError("SWE-bench verification deadline expired") |
| return remaining |
|
|
|
|
| def _command_evidence(argv: Sequence[str]) -> dict[str, Any]: |
| preimage = {"argv": list(argv)} |
| encoded = _canonical_bytes(preimage) |
| return { |
| "argc": len(argv), |
| "bytes": len(encoded), |
| "executable": Path(argv[0]).name if argv else None, |
| "sha256": _sha256(encoded), |
| } |
|
|
|
|
| def _process_evidence(result: Any | None, *, error_type: str | None = None) -> dict[str, Any]: |
| if result is None: |
| preimage: dict[str, Any] = { |
| "returncode": None, |
| "stdout": "", |
| "stderr": "", |
| "elapsed": 0.0, |
| "timed_out": False, |
| "reaped_descendants": 0, |
| "residual_descendants": [], |
| "error_type": error_type, |
| } |
| else: |
| preimage = { |
| "returncode": int(result.returncode), |
| "stdout": str(result.stdout), |
| "stderr": str(result.stderr), |
| "elapsed": float(result.elapsed), |
| "timed_out": bool(result.timed_out), |
| "reaped_descendants": int(result.reaped_descendants), |
| "residual_descendants": list(result.residual_descendants), |
| "error_type": error_type, |
| } |
| encoded = _canonical_bytes(preimage) |
| stdout = str(preimage.pop("stdout")) |
| stderr = str(preimage.pop("stderr")) |
| residual_descendants = list(preimage.pop("residual_descendants")) |
| return { |
| **preimage, |
| "bytes": len(encoded), |
| "residual_descendant_count": len(residual_descendants), |
| "residual_descendants_sha256": _sha256(_canonical_bytes(residual_descendants)), |
| "sha256": _sha256(encoded), |
| "stderr_bytes": len(stderr.encode("utf-8")), |
| "stderr_sha256": _sha256(stderr.encode("utf-8")), |
| "stdout_bytes": len(stdout.encode("utf-8")), |
| "stdout_sha256": _sha256(stdout.encode("utf-8")), |
| } |
|
|
|
|
| def _regular_file(path: Path, *, label: str) -> Path: |
| if not path.is_absolute(): |
| raise SWEbenchVerificationError(f"{label} must be absolute") |
| try: |
| resolved = path.resolve(strict=True) |
| mode = resolved.stat().st_mode |
| except OSError as exc: |
| raise SWEbenchVerificationError(f"{label} is unavailable") from exc |
| if not stat.S_ISREG(mode): |
| raise SWEbenchVerificationError(f"{label} must be a regular file") |
| return resolved |
|
|
|
|
| def _executable(path: Path, *, label: str) -> Path: |
| if not path.is_absolute(): |
| raise SWEbenchVerificationError(f"{label} must be absolute") |
| candidate = Path(os.path.abspath(path)) |
| _regular_file(candidate, label=label) |
| if not os.access(candidate, os.X_OK): |
| raise SWEbenchVerificationError(f"{label} must be executable") |
| return candidate |
|
|
|
|
| def _checkout(path: Path) -> Path: |
| if not path.is_absolute(): |
| raise SWEbenchVerificationError("SWE-bench checkout must be absolute") |
| try: |
| resolved = path.resolve(strict=True) |
| except OSError as exc: |
| raise SWEbenchVerificationError("SWE-bench checkout is unavailable") from exc |
| if not resolved.is_dir(): |
| raise SWEbenchVerificationError("SWE-bench checkout must be a directory") |
| return resolved |
|
|
|
|
| def _unix_docker_host(value: str) -> tuple[str, Path]: |
| if not value.startswith("unix://") or any(character in value for character in "?#\0"): |
| raise SWEbenchVerificationError("DOCKER_HOST must name a unix socket") |
| raw_path = value.removeprefix("unix://") |
| path = Path(raw_path) |
| if not path.is_absolute(): |
| raise SWEbenchVerificationError("DOCKER_HOST socket path must be absolute") |
| try: |
| resolved = path.resolve(strict=True) |
| mode = resolved.stat().st_mode |
| except OSError as exc: |
| raise SWEbenchVerificationError("DOCKER_HOST socket is unavailable") from exc |
| if not stat.S_ISSOCK(mode): |
| raise SWEbenchVerificationError("DOCKER_HOST must name a unix socket") |
| return f"unix://{resolved}", resolved |
|
|
|
|
| def _private_write_json(path: Path, value: Mapping[str, Any]) -> None: |
| data = _canonical_bytes(value) |
| descriptor = os.open( |
| path, |
| os.O_WRONLY | os.O_CREAT | os.O_TRUNC | getattr(os, "O_NOFOLLOW", 0), |
| 0o600, |
| ) |
| try: |
| with os.fdopen(descriptor, "wb") as handle: |
| handle.write(data) |
| handle.flush() |
| os.fsync(handle.fileno()) |
| except BaseException: |
| try: |
| os.close(descriptor) |
| except OSError: |
| pass |
| raise |
|
|
|
|
| def _safe_relative(path: Path, root: Path) -> str: |
| try: |
| resolved = path.resolve(strict=True) |
| relative = resolved.relative_to(root.resolve(strict=True)) |
| except (OSError, ValueError) as exc: |
| raise SWEbenchVerificationError("evidence path escaped the private run") from exc |
| if path.is_symlink(): |
| raise SWEbenchVerificationError("evidence artifacts must not be symlinks") |
| return relative.as_posix() |
|
|
|
|
| def _artifact_evidence(path: Path, root: Path) -> dict[str, Any]: |
| relative = path.relative_to(root).as_posix() |
| if not path.exists(): |
| return {"bytes": 0, "name": path.name, "present": False, "sha256": None} |
| _safe_relative(path, root) |
| mode = path.stat().st_mode |
| if not stat.S_ISREG(mode): |
| raise SWEbenchVerificationError("evidence artifact must be a regular file") |
| size = path.stat().st_size |
| if size > MAX_EVIDENCE_BYTES: |
| raise SWEbenchVerificationError("evidence artifact exceeds the retained-size limit") |
| data = path.read_bytes() |
| return { |
| "bytes": len(data), |
| "name": Path(relative).name, |
| "present": True, |
| "sha256": _sha256(data), |
| } |
|
|
|
|
| def _json_file(path: Path, root: Path) -> dict[str, Any]: |
| evidence = _artifact_evidence(path, root) |
| if evidence["present"] is not True: |
| raise SWEbenchVerificationError("required JSON evidence is unavailable") |
| try: |
| value = json.loads(path.read_text(encoding="utf-8")) |
| except (OSError, UnicodeDecodeError, json.JSONDecodeError) as exc: |
| raise SWEbenchVerificationError("required JSON evidence is invalid") from exc |
| if not isinstance(value, dict): |
| raise SWEbenchVerificationError("required JSON evidence must be an object") |
| return value |
|
|
|
|
| def _safe_production_python_path(value: str) -> bool: |
| if ( |
| not value |
| or not SAFE_PATH_PATTERN.fullmatch(value) |
| or "\\" in value |
| or value.startswith("/") |
| ): |
| return False |
| pure = PurePosixPath(value) |
| if str(pure) != value or ".." in pure.parts or pure.suffix != ".py": |
| return False |
| lowered = tuple(part.lower() for part in pure.parts) |
| filename = lowered[-1] |
| return not ( |
| any(part in TEST_PATH_COMPONENTS for part in lowered[:-1]) |
| or filename == "test.py" |
| or filename.startswith("test_") |
| or filename.endswith("_test.py") |
| ) |
|
|
|
|
| def select_mode_target( |
| allowed_paths: Sequence[str], |
| tracked_modes: Mapping[str, str], |
| ) -> str: |
| """Choose a deterministic tracked 100644 Python production path.""" |
| candidates = sorted( |
| { |
| path |
| for path in allowed_paths |
| if isinstance(path, str) |
| and _safe_production_python_path(path) |
| and tracked_modes.get(path) == "100644" |
| } |
| ) |
| if not candidates: |
| raise SWEbenchVerificationError("red control has no safe mode-only target") |
| return candidates[0] |
|
|
|
|
| def mode_only_patch(path: str) -> str: |
| """Build the inert 100644-to-100755 Git mode patch used by red controls.""" |
| if not _safe_production_python_path(path): |
| raise SWEbenchVerificationError("mode-only target is not a production Python path") |
| return f"diff --git a/{path} b/{path}\nold mode 100644\nnew mode 100755\n" |
|
|
|
|
| def _string_list(value: Any, *, allow_empty: bool) -> list[str]: |
| if not isinstance(value, list) or not all(isinstance(item, str) and item for item in value): |
| raise SWEbenchVerificationError("test selector lists are invalid") |
| items = list(value) |
| if not allow_empty and not items: |
| raise SWEbenchVerificationError("FAIL_TO_PASS must not be empty") |
| if len(items) != len(set(items)): |
| raise SWEbenchVerificationError("test selector lists contain duplicates") |
| return items |
|
|
|
|
| def _report_group(entry: Mapping[str, Any], expected: Sequence[str]) -> dict[str, list[str]]: |
| success = _string_list(entry.get("success"), allow_empty=True) |
| failure = _string_list(entry.get("failure"), allow_empty=True) |
| if set(success) & set(failure) or set(success) | set(failure) != set(expected): |
| raise SWEbenchVerificationError("official report selector identity is inconsistent") |
| return {"success": success, "failure": failure} |
|
|
|
|
| def _normalize_raw_statuses( |
| statuses: Mapping[str, str], |
| expected: set[str], |
| ) -> tuple[dict[str, str], dict[str, int]]: |
| observations: dict[str, list[str]] = {selector: [] for selector in expected} |
| exact_selectors: set[str] = set() |
| for raw_selector, raw_status in statuses.items(): |
| if not isinstance(raw_selector, str) or not isinstance(raw_status, str): |
| raise SWEbenchVerificationError("raw test statuses are malformed") |
| matches = [selector for selector in expected if raw_selector == selector] |
| if matches: |
| exact_selectors.add(matches[0]) |
| if not matches: |
| matches = [ |
| selector |
| for selector in expected |
| if "/" in raw_selector |
| and raw_selector.endswith(f":{selector}") |
| and SAFE_PATH_PATTERN.fullmatch(raw_selector.replace(":", "/")) |
| ] |
| if len(matches) != 1: |
| raise SWEbenchVerificationError("raw test selector identity is not exact") |
| observations[matches[0]].append(raw_status) |
| if exact_selectors != expected or any( |
| not values or len(set(values)) != 1 for values in observations.values() |
| ): |
| raise SWEbenchVerificationError("raw test selector occurrences conflict or are missing") |
| return ( |
| {selector: values[0] for selector, values in observations.items()}, |
| {selector: len(values) for selector, values in observations.items()}, |
| ) |
|
|
|
|
| def validate_outcome( |
| *, |
| phase: str, |
| fail_to_pass: Sequence[str], |
| pass_to_pass: Sequence[str], |
| statuses: Mapping[str, str], |
| report_entry: Mapping[str, Any], |
| official_resolution: str, |
| ) -> dict[str, Any]: |
| """Validate exact raw and grouped official SWE-bench outcomes.""" |
| if phase not in PHASES: |
| raise SWEbenchVerificationError("unsupported SWE-bench verification phase") |
| f2p = _string_list(list(fail_to_pass), allow_empty=False) |
| p2p = _string_list(list(pass_to_pass), allow_empty=True) |
| if set(f2p) & set(p2p): |
| raise SWEbenchVerificationError("FAIL_TO_PASS and PASS_TO_PASS overlap") |
| expected = set(f2p) | set(p2p) |
| if not isinstance(statuses, Mapping): |
| raise SWEbenchVerificationError("raw test selector identity is not exact") |
| raw, occurrences = _normalize_raw_statuses(statuses, expected) |
| if any(status not in KNOWN_STATUSES or status == "SKIPPED" for status in raw.values()): |
| raise SWEbenchVerificationError("raw test status is missing, skipped, or unsupported") |
|
|
| red = phase == "red" |
| f2p_allowed = RED_FAILURE_STATUSES if red else PASSING_STATUSES |
| if any(raw[selector] not in f2p_allowed for selector in f2p): |
| raise SWEbenchVerificationError("FAIL_TO_PASS statuses violate the phase contract") |
| if any(raw[selector] not in PASSING_STATUSES for selector in p2p): |
| raise SWEbenchVerificationError("PASS_TO_PASS statuses violate the phase contract") |
|
|
| for key, expected_value in ( |
| ("patch_is_None", False), |
| ("patch_exists", True), |
| ("patch_successfully_applied", True), |
| ("resolved", not red), |
| ): |
| if report_entry.get(key) is not expected_value: |
| raise SWEbenchVerificationError("official report patch or resolution state is invalid") |
| expected_resolution = "RESOLVED_NO" if red else "RESOLVED_FULL" |
| if official_resolution != expected_resolution: |
| raise SWEbenchVerificationError("official resolution does not match the phase") |
|
|
| tests_status = report_entry.get("tests_status") |
| if not isinstance(tests_status, Mapping): |
| raise SWEbenchVerificationError("official report omitted test status groups") |
| f2p_report = tests_status.get("FAIL_TO_PASS") |
| p2p_report = tests_status.get("PASS_TO_PASS") |
| if not isinstance(f2p_report, Mapping) or not isinstance(p2p_report, Mapping): |
| raise SWEbenchVerificationError("official report omitted required status groups") |
| normalized_f2p = _report_group(f2p_report, f2p) |
| normalized_p2p = _report_group(p2p_report, p2p) |
| expected_f2p_success = [] if red else f2p |
| expected_f2p_failure = f2p if red else [] |
| if ( |
| set(normalized_f2p["success"]) != set(expected_f2p_success) |
| or set(normalized_f2p["failure"]) != set(expected_f2p_failure) |
| or set(normalized_p2p["success"]) != set(p2p) |
| or normalized_p2p["failure"] |
| ): |
| raise SWEbenchVerificationError("official grouped statuses violate the phase contract") |
| return { |
| "exact_selector_identity": True, |
| "fail_to_pass_count": len(f2p), |
| "phase": phase, |
| "pass_to_pass_count": len(p2p), |
| "exact_selector_keys_present": True, |
| "parsed_status_key_count": sum(occurrences.values()), |
| "resolution": official_resolution, |
| "resolved": not red, |
| "status_counts": { |
| status: sum(1 for value in raw.values() if value == status) |
| for status in sorted(set(raw.values())) |
| }, |
| } |
|
|
|
|
| def _auth_environment() -> dict[str, str]: |
| return { |
| "GIT_CONFIG_GLOBAL": os.devnull, |
| "GIT_CONFIG_NOSYSTEM": "1", |
| "GIT_OPTIONAL_LOCKS": "0", |
| "LANG": "C", |
| "LC_ALL": "C", |
| "PATH": os.defpath, |
| "PYTHONDONTWRITEBYTECODE": "1", |
| } |
|
|
|
|
| def _freeze_tree(root: Path) -> None: |
| for path in sorted(root.rglob("*"), key=lambda item: len(item.parts), reverse=True): |
| if path.is_symlink(): |
| raise SWEbenchVerificationError("authenticated input contains a symlink") |
| if path.is_dir(): |
| os.chmod(path, 0o500) |
| elif path.is_file(): |
| executable = bool(path.stat().st_mode & 0o111) |
| os.chmod(path, 0o500 if executable else 0o400) |
| else: |
| raise SWEbenchVerificationError("authenticated input has an unsupported file type") |
| os.chmod(root, 0o500) |
|
|
|
|
| def _snapshot_file( |
| source: Path, |
| *, |
| inputs: Path, |
| label: str, |
| expected_sha256: str, |
| ) -> Path: |
| if not SHA256_PATTERN.fullmatch(expected_sha256): |
| raise SWEbenchVerificationError(f"{label} snapshot identity is invalid") |
| data = _regular_file(source, label=label).read_bytes() |
| if _sha256(data) != expected_sha256: |
| raise SWEbenchVerificationError(f"{label} identity drifted") |
| inputs.mkdir(mode=0o700, parents=True, exist_ok=True) |
| destination = inputs / f"{label.lower().replace(' ', '-')}-{expected_sha256}{source.suffix}" |
| descriptor = os.open( |
| destination, |
| os.O_WRONLY | os.O_CREAT | os.O_EXCL | getattr(os, "O_NOFOLLOW", 0), |
| 0o400, |
| ) |
| try: |
| with os.fdopen(descriptor, "wb") as handle: |
| descriptor = -1 |
| handle.write(data) |
| handle.flush() |
| os.fsync(handle.fileno()) |
| finally: |
| if descriptor >= 0: |
| os.close(descriptor) |
| os.chmod(destination, 0o400) |
| if _sha256(destination.read_bytes()) != expected_sha256: |
| raise SWEbenchVerificationError(f"{label} snapshot drifted") |
| return destination |
|
|
|
|
| def worker_environment( |
| *, |
| checkout: Path, |
| private_cwd: Path, |
| docker_host: str, |
| ) -> dict[str, str]: |
| """Return the complete, secret-free worker environment.""" |
| return { |
| "DOCKER_HOST": docker_host, |
| "GIT_CONFIG_GLOBAL": os.devnull, |
| "GIT_CONFIG_NOSYSTEM": "1", |
| "HF_DATASETS_OFFLINE": "1", |
| "HF_HUB_OFFLINE": "1", |
| "HOME": str(private_cwd / "home"), |
| "LANG": "C.UTF-8", |
| "LC_ALL": "C.UTF-8", |
| "NO_COLOR": "1", |
| "PATH": os.defpath, |
| "PYTHONDONTWRITEBYTECODE": "1", |
| "PYTHONHASHSEED": "0", |
| "PYTHONNOUSERSITE": "1", |
| "PYTHONPATH": str(checkout), |
| "PYTHONUTF8": "1", |
| "TMPDIR": str(private_cwd / "tmp"), |
| "TZ": "UTC", |
| } |
|
|
|
|
| def _git_executable() -> Path: |
| candidate = shutil.which("git", path=os.defpath) |
| if not candidate: |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| return _executable(Path(candidate), label="Git executable") |
|
|
|
|
| def _materialize_authenticated_harness( |
| *, |
| source_checkout: Path, |
| destination: Path, |
| git: Path, |
| expected_revision: str, |
| expected_run_evaluation_sha256: str, |
| deadline: float, |
| ) -> dict[str, Any]: |
| if not REVISION_PATTERN.fullmatch(expected_revision) or not SHA256_PATTERN.fullmatch( |
| expected_run_evaluation_sha256 |
| ): |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| clone_command = [ |
| str(git), |
| "-c", |
| "core.hooksPath=/dev/null", |
| "clone", |
| "--quiet", |
| "--no-checkout", |
| "--no-hardlinks", |
| "--", |
| str(source_checkout), |
| str(destination), |
| ] |
| clone = _run_process( |
| clone_command, |
| cwd=destination.parent, |
| env=_auth_environment(), |
| timeout=min(120.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if clone.returncode or clone.timed_out or clone.residual_descendants: |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| checkout_command = [ |
| str(git), |
| "-c", |
| "core.hooksPath=/dev/null", |
| "-C", |
| str(destination), |
| "checkout", |
| "--quiet", |
| "--detach", |
| expected_revision, |
| ] |
| checked_out = _run_process( |
| checkout_command, |
| cwd=destination, |
| env=_auth_environment(), |
| timeout=min(60.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if checked_out.returncode or checked_out.timed_out or checked_out.residual_descendants: |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| revision_command = [ |
| str(git), |
| "-C", |
| str(destination), |
| "rev-parse", |
| "--verify", |
| "HEAD^{commit}", |
| ] |
| revision = _run_process( |
| revision_command, |
| cwd=destination, |
| env=_auth_environment(), |
| timeout=min(15.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if revision.returncode or revision.stdout.strip() != expected_revision: |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| status_command = [ |
| str(git), |
| "-C", |
| str(destination), |
| "status", |
| "--porcelain=v1", |
| "--untracked-files=all", |
| "--ignored=matching", |
| ] |
| status_result = _run_process( |
| status_command, |
| cwd=destination, |
| env=_auth_environment(), |
| timeout=min(15.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| run_evaluation = _regular_file( |
| destination / "swebench/harness/run_evaluation.py", |
| label="run_evaluation.py", |
| ) |
| observed_sha256 = _sha256(run_evaluation.read_bytes()) |
| if ( |
| revision.returncode |
| or revision.stdout.strip() != expected_revision |
| or status_result.returncode |
| or status_result.stdout.strip() |
| or observed_sha256 != expected_run_evaluation_sha256 |
| ): |
| raise SWEbenchVerificationError("SWE-bench harness authentication failed") |
| source_manifest = _package_manifest(destination / "swebench") |
| _freeze_tree(destination) |
| return { |
| "commands": [ |
| { |
| "command": _command_evidence(clone_command), |
| "process": _process_evidence(clone), |
| }, |
| { |
| "command": _command_evidence(checkout_command), |
| "process": _process_evidence(checked_out), |
| }, |
| { |
| "command": _command_evidence(revision_command), |
| "process": _process_evidence(revision), |
| }, |
| { |
| "command": _command_evidence(status_command), |
| "process": _process_evidence(status_result), |
| }, |
| ], |
| "git_revision": expected_revision, |
| "run_evaluation_sha256": observed_sha256, |
| "source_file_count": source_manifest["file_count"], |
| "source_sha256": source_manifest["sha256"], |
| } |
|
|
|
|
| def _validate_scored_patch( |
| *, |
| patch: str, |
| allowed_paths: Sequence[str], |
| git: Path, |
| cwd: Path, |
| deadline: float, |
| ) -> tuple[list[str], dict[str, Any]]: |
| encoded = patch.encode("utf-8") |
| if not patch.strip() or len(encoded) > 1_000_000: |
| raise SWEbenchVerificationError("scored patch is empty or exceeds the size limit") |
| command = [str(git), "apply", "--numstat", "-z", "--no-unsafe-paths", "-"] |
| result = _run_process( |
| command, |
| cwd=cwd, |
| env=_auth_environment(), |
| input_text=patch, |
| timeout=min(30.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if result.returncode or result.timed_out or result.residual_descendants: |
| raise SWEbenchVerificationError("scored patch metadata is invalid") |
| summary_command = [str(git), "apply", "--summary", "--no-unsafe-paths", "-"] |
| summary = _run_process( |
| summary_command, |
| cwd=cwd, |
| env=_auth_environment(), |
| input_text=patch, |
| timeout=min(30.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if ( |
| summary.returncode |
| or summary.timed_out |
| or summary.residual_descendants |
| or any( |
| line.lstrip().startswith(("rename ", "copy ")) for line in summary.stdout.splitlines() |
| ) |
| ): |
| raise SWEbenchVerificationError("scored patch renames or copies repository paths") |
| paths: list[str] = [] |
| for row in result.stdout.split("\0"): |
| if not row: |
| continue |
| fields = row.split("\t", 2) |
| if len(fields) != 3 or not fields[2]: |
| raise SWEbenchVerificationError("scored patch path metadata is malformed") |
| paths.append(fields[2]) |
| allowed = set(allowed_paths) |
| if not paths or any(path not in allowed for path in paths): |
| raise SWEbenchVerificationError("scored patch changes a disallowed path") |
| return sorted(set(paths)), { |
| "bytes": len(encoded), |
| "command": _command_evidence(command), |
| "path_count": len(set(paths)), |
| "paths_sha256": _sha256(_canonical_bytes(sorted(set(paths)))), |
| "process": _process_evidence(result), |
| "sha256": _sha256(encoded), |
| "summary_command": _command_evidence(summary_command), |
| "summary_process": _process_evidence(summary), |
| } |
|
|
|
|
| def _python_environment_identity( |
| *, |
| python: Path, |
| expected_sha256: str, |
| cwd: Path, |
| deadline: float, |
| ) -> dict[str, Any]: |
| if not SHA256_PATTERN.fullmatch(expected_sha256): |
| raise SWEbenchVerificationError("SWE-bench Python environment identity is invalid") |
| command = [ |
| str(python), |
| "-m", |
| "pip", |
| "--disable-pip-version-check", |
| "freeze", |
| "--all", |
| ] |
| environment = { |
| **_auth_environment(), |
| "HOME": str(cwd / "home"), |
| "PIP_CONFIG_FILE": os.devnull, |
| "PYTHONNOUSERSITE": "1", |
| } |
| result = _run_process( |
| command, |
| cwd=cwd, |
| env=environment, |
| timeout=min(60.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| lines = sorted(line.strip() for line in result.stdout.splitlines() if line.strip()) |
| observed = _sha256(_canonical_bytes(lines)) |
| if ( |
| result.returncode |
| or result.timed_out |
| or result.residual_descendants |
| or observed != expected_sha256 |
| ): |
| raise SWEbenchVerificationError("SWE-bench Python environment identity drifted") |
| return { |
| "command": _command_evidence(command), |
| "distribution_count": len(lines), |
| "process": _process_evidence(result), |
| "sha256": observed, |
| } |
|
|
|
|
| def _package_manifest(root: Path) -> dict[str, Any]: |
| resolved = root.resolve(strict=True) |
| if not resolved.is_dir(): |
| raise SWEbenchVerificationError("Python package root is invalid") |
| files: list[dict[str, Any]] = [] |
| for path in sorted(resolved.rglob("*")): |
| if path.is_symlink(): |
| raise SWEbenchVerificationError("Python package contains a symlink") |
| if not path.is_file() or path.suffix not in {".py", ".pyi"}: |
| continue |
| data = path.read_bytes() |
| files.append( |
| { |
| "bytes": len(data), |
| "path": path.relative_to(resolved).as_posix(), |
| "sha256": _sha256(data), |
| } |
| ) |
| if not files: |
| raise SWEbenchVerificationError("Python package contains no source files") |
| return { |
| "file_count": len(files), |
| "sha256": _sha256(_canonical_bytes(files)), |
| } |
|
|
|
|
| def _package_root_for_name(package: str) -> Path: |
| if not re.fullmatch(r"[A-Za-z_][A-Za-z0-9_]*", package): |
| raise SWEbenchVerificationError("Python package name is invalid") |
| spec = importlib.util.find_spec(package) |
| if spec is None or not isinstance(spec.origin, str): |
| raise SWEbenchVerificationError("Python package is unavailable") |
| root = Path(spec.origin).resolve(strict=True).parent |
| try: |
| root.relative_to(Path(sys.prefix).resolve(strict=True)) |
| except ValueError as exc: |
| raise SWEbenchVerificationError("Python package escaped the pinned environment") from exc |
| return root |
|
|
|
|
| def _package_manifest_for_name(package: str) -> dict[str, Any]: |
| return _package_manifest(_package_root_for_name(package)) |
|
|
|
|
| def _package_snapshot_for_name(package: str, destination: Path) -> dict[str, Any]: |
| if not destination.is_absolute() or destination.exists() or destination.is_symlink(): |
| raise SWEbenchVerificationError("Python package snapshot destination is invalid") |
| source = _package_root_for_name(package) |
| shutil.copytree(source, destination, symlinks=False) |
| manifest = _package_manifest(destination) |
| _freeze_tree(destination) |
| return manifest |
|
|
|
|
| def _python_package_snapshot( |
| *, |
| python: Path, |
| bridge: Path, |
| package: str, |
| expected_sha256: str, |
| destination: Path, |
| cwd: Path, |
| deadline: float, |
| ) -> tuple[Path, dict[str, Any]]: |
| if not SHA256_PATTERN.fullmatch(expected_sha256): |
| raise SWEbenchVerificationError("Python package identity is invalid") |
| command = [ |
| str(python), |
| str(bridge), |
| "package-snapshot", |
| "--package", |
| package, |
| "--destination", |
| str(destination), |
| ] |
| result = _run_process( |
| command, |
| cwd=cwd, |
| env={ |
| **_auth_environment(), |
| "HOME": str(cwd / "home"), |
| "PYTHONNOUSERSITE": "1", |
| }, |
| timeout=min(30.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| try: |
| manifest = json.loads(result.stdout) |
| except json.JSONDecodeError as exc: |
| raise SWEbenchVerificationError("Python package identity probe failed") from exc |
| if ( |
| result.returncode |
| or result.timed_out |
| or result.residual_descendants |
| or not isinstance(manifest, dict) |
| or manifest.get("sha256") != expected_sha256 |
| or not isinstance(manifest.get("file_count"), int) |
| or manifest["file_count"] < 1 |
| or not destination.is_dir() |
| or _package_manifest(destination) != manifest |
| ): |
| raise SWEbenchVerificationError("Python package identity drifted") |
| return ( |
| destination, |
| { |
| "command": _command_evidence(command), |
| "file_count": manifest["file_count"], |
| "process": _process_evidence(result), |
| "sha256": expected_sha256, |
| }, |
| ) |
|
|
|
|
| def _new_run_id(phase: str) -> str: |
| return f"ctx-sb-{phase}-{secrets.token_hex(10)}" |
|
|
|
|
| def _model_name(phase: str) -> str: |
| return f"ctx-swebench-{phase}" |
|
|
|
|
| def _prepare_private_cwd(path: Path) -> Path: |
| if not path.is_absolute(): |
| raise SWEbenchVerificationError("private worker CWD must be absolute") |
| resolved = path.resolve(strict=False) |
| if resolved.exists(): |
| raise SWEbenchVerificationError("private worker CWD must not already exist") |
| resolved.mkdir(mode=0o700, parents=True) |
| os.chmod(resolved, 0o700) |
| for name in ("home", "tmp"): |
| child = resolved / name |
| child.mkdir(mode=0o700) |
| os.chmod(child, 0o700) |
| return resolved |
|
|
|
|
| def _docker_ids(stdout: str) -> list[str]: |
| ids = [line.strip() for line in stdout.splitlines() if line.strip()] |
| if any(not CONTAINER_ID_PATTERN.fullmatch(item) for item in ids): |
| raise SWEbenchVerificationError("Docker cleanup returned invalid container identifiers") |
| return sorted(set(ids)) |
|
|
|
|
| def _docker_query( |
| *, |
| docker_cli: Path, |
| docker_host: str, |
| arguments: Sequence[str], |
| cwd: Path, |
| env: Mapping[str, str], |
| deadline: float, |
| ) -> tuple[list[str], dict[str, Any]]: |
| command = [str(docker_cli), "--host", docker_host, *arguments] |
| result = _run_process( |
| command, |
| cwd=cwd, |
| env=dict(env), |
| timeout=min(15.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| if result.returncode or result.timed_out or result.residual_descendants: |
| raise SWEbenchVerificationError("Docker authentication or inventory failed") |
| values = sorted({line.strip() for line in result.stdout.splitlines() if line.strip()}) |
| return values, { |
| "command": _command_evidence(command), |
| "count": len(values), |
| "process": _process_evidence(result), |
| "sha256": _sha256(_canonical_bytes(values)), |
| } |
|
|
|
|
| def _docker_identity( |
| *, |
| docker_cli: Path, |
| docker_host: str, |
| expected_daemon_id: str, |
| expected_server_version: str, |
| cwd: Path, |
| env: Mapping[str, str], |
| deadline: float, |
| ) -> dict[str, Any]: |
| if ( |
| not expected_daemon_id |
| or len(expected_daemon_id) > 200 |
| or not expected_server_version |
| or len(expected_server_version) > 100 |
| ): |
| raise SWEbenchVerificationError("Docker daemon identity is invalid") |
| daemon_ids, daemon_evidence = _docker_query( |
| docker_cli=docker_cli, |
| docker_host=docker_host, |
| arguments=("info", "--format", "{{.ID}}"), |
| cwd=cwd, |
| env=env, |
| deadline=deadline, |
| ) |
| versions, version_evidence = _docker_query( |
| docker_cli=docker_cli, |
| docker_host=docker_host, |
| arguments=("version", "--format", "{{.Server.Version}}"), |
| cwd=cwd, |
| env=env, |
| deadline=deadline, |
| ) |
| if daemon_ids != [expected_daemon_id] or versions != [expected_server_version]: |
| raise SWEbenchVerificationError("Docker daemon identity drifted") |
| return { |
| "daemon_id_sha256": _sha256(expected_daemon_id.encode()), |
| "info": daemon_evidence, |
| "server_version": expected_server_version, |
| "version": version_evidence, |
| } |
|
|
|
|
| def _docker_inventory( |
| *, |
| docker_cli: Path, |
| docker_host: str, |
| cwd: Path, |
| env: Mapping[str, str], |
| deadline: float, |
| ) -> tuple[dict[str, frozenset[str]], dict[str, Any]]: |
| commands = { |
| "containers": ("ps", "--all", "--quiet", "--no-trunc"), |
| "images": ("image", "ls", "--all", "--no-trunc", "--quiet"), |
| "networks": ("network", "ls", "--quiet", "--no-trunc"), |
| "volumes": ("volume", "ls", "--quiet"), |
| } |
| values: dict[str, frozenset[str]] = {} |
| evidence: dict[str, Any] = {} |
| for name, arguments in commands.items(): |
| observed, record = _docker_query( |
| docker_cli=docker_cli, |
| docker_host=docker_host, |
| arguments=arguments, |
| cwd=cwd, |
| env=env, |
| deadline=deadline, |
| ) |
| values[name] = frozenset(observed) |
| evidence[name] = record |
| return values, evidence |
|
|
|
|
| def _cleanup_containers( |
| *, |
| docker_cli: Path, |
| docker_host: str, |
| run_id: str, |
| cwd: Path, |
| env: Mapping[str, str], |
| ) -> dict[str, Any]: |
| records: list[dict[str, Any]] = [] |
| errors: list[str] = [] |
| residual: list[str] = [] |
| deadline = time.monotonic() + 30 |
|
|
| def execute(argv: list[str]) -> Any | None: |
| try: |
| result = _run_process( |
| argv, |
| cwd=cwd, |
| env=dict(env), |
| timeout=min(10.0, _remaining(deadline)), |
| contain_descendants=True, |
| ) |
| except BaseException as exc: |
| records.append( |
| { |
| "command": _command_evidence(argv), |
| "process": _process_evidence(None, error_type=type(exc).__name__), |
| } |
| ) |
| errors.append("cleanup process failed") |
| return None |
| records.append( |
| { |
| "command": _command_evidence(argv), |
| "process": _process_evidence(result), |
| } |
| ) |
| if result.returncode or result.timed_out or result.residual_descendants: |
| errors.append("cleanup command failed") |
| return result |
|
|
| for _attempt in range(3): |
| list_command = [ |
| str(docker_cli), |
| "--host", |
| docker_host, |
| "ps", |
| "--all", |
| "--quiet", |
| "--filter", |
| f"label={RUN_LABEL}={run_id}", |
| ] |
| listed = execute(list_command) |
| if listed is None or listed.returncode: |
| break |
| try: |
| residual = _docker_ids(str(listed.stdout)) |
| except SWEbenchVerificationError: |
| errors.append("cleanup list was invalid") |
| break |
| if not residual: |
| break |
| removed = execute( |
| [ |
| str(docker_cli), |
| "--host", |
| docker_host, |
| "rm", |
| "--force", |
| *residual, |
| ] |
| ) |
| if removed is None: |
| break |
|
|
| final_command = [ |
| str(docker_cli), |
| "--host", |
| docker_host, |
| "ps", |
| "--all", |
| "--quiet", |
| "--filter", |
| f"label={RUN_LABEL}={run_id}", |
| ] |
| final = execute(final_command) |
| verification_complete = final is not None and final.returncode == 0 |
| if final is not None and final.returncode == 0: |
| try: |
| residual = _docker_ids(str(final.stdout)) |
| except SWEbenchVerificationError: |
| residual = [] |
| errors.append("final cleanup list was invalid") |
| else: |
| errors.append("final cleanup verification failed") |
| return { |
| "commands": records, |
| "errors": errors, |
| "ok": not errors and not residual and verification_complete, |
| "residual_container_ids": residual, |
| "verification_complete": verification_complete, |
| } |
|
|
|
|
| def _container_policy_attestation( |
| container: Any, |
| *, |
| run_id: str, |
| expected_image_id: str, |
| ) -> dict[str, Any]: |
| container.reload() |
| attrs = container.attrs |
| if not isinstance(attrs, Mapping): |
| raise SWEbenchVerificationError("Docker container inspection is invalid") |
| host = attrs.get("HostConfig") |
| config = attrs.get("Config") |
| mounts = attrs.get("Mounts") |
| if ( |
| not isinstance(host, Mapping) |
| or not isinstance(config, Mapping) |
| or not isinstance(mounts, list) |
| ): |
| raise SWEbenchVerificationError("Docker container inspection is incomplete") |
| labels = config.get("Labels") |
| security_opt = host.get("SecurityOpt") or [] |
| cap_drop = host.get("CapDrop") or [] |
| if ( |
| not isinstance(labels, Mapping) |
| or labels.get(RUN_LABEL) != run_id |
| or attrs.get("Image") != expected_image_id |
| or host.get("NetworkMode") != "none" |
| or host.get("Memory") != CONTAINER_MEMORY_BYTES |
| or host.get("NanoCpus") != CONTAINER_NANO_CPUS |
| or host.get("PidsLimit") != CONTAINER_PIDS_LIMIT |
| or host.get("Privileged") is not False |
| or host.get("PidMode") not in (None, "") |
| or host.get("IpcMode") not in (None, "", "private") |
| or host.get("UTSMode") not in (None, "") |
| or host.get("UsernsMode") not in (None, "") |
| or host.get("CgroupnsMode") != "private" |
| or host.get("Binds") not in (None, []) |
| or mounts |
| or cap_drop != ["ALL"] |
| or CONTAINER_SECURITY_OPT not in security_opt |
| ): |
| raise SWEbenchVerificationError("Docker container policy was not enforced") |
| return { |
| "cap_drop": ["ALL"], |
| "cgroupns_mode": "private", |
| "image_id": expected_image_id, |
| "memory_bytes": CONTAINER_MEMORY_BYTES, |
| "mount_count": 0, |
| "nano_cpus": CONTAINER_NANO_CPUS, |
| "network_mode": "none", |
| "pids_limit": CONTAINER_PIDS_LIMIT, |
| "privileged": False, |
| "security_opt": CONTAINER_SECURITY_OPT, |
| } |
|
|
|
|
| class _HardenedContainerCollection: |
| def __init__( |
| self, |
| collection: Any, |
| *, |
| run_id: str, |
| expected_image_id: str, |
| attestations: list[dict[str, Any]], |
| ) -> None: |
| self._collection = collection |
| self._run_id = run_id |
| self._expected_image_id = expected_image_id |
| self._attestations = attestations |
|
|
| def __getattr__(self, name: str) -> Any: |
| return getattr(self._collection, name) |
|
|
| def create(self, *args: Any, **kwargs: Any) -> Any: |
| for field in ("binds", "devices", "device_requests", "mounts", "tmpfs", "volumes"): |
| if kwargs.get(field): |
| raise SWEbenchVerificationError("Docker host resources are forbidden") |
| for field in ("cgroupns", "ipc_mode", "pid_mode", "uts_mode", "userns_mode"): |
| if kwargs.get(field): |
| raise SWEbenchVerificationError("Docker host namespaces are forbidden") |
| if kwargs.get("privileged") is True or kwargs.get("cap_add"): |
| raise SWEbenchVerificationError("Docker privilege escalation is forbidden") |
| labels = kwargs.get("labels") or {} |
| if not isinstance(labels, Mapping) or ( |
| RUN_LABEL in labels and labels.get(RUN_LABEL) != self._run_id |
| ): |
| raise SWEbenchVerificationError("Docker container labels are invalid") |
| kwargs.update( |
| { |
| "cap_drop": ["ALL"], |
| "cgroupns": "private", |
| "init": True, |
| "labels": {**dict(labels), RUN_LABEL: self._run_id}, |
| "mem_limit": CONTAINER_MEMORY_BYTES, |
| "nano_cpus": CONTAINER_NANO_CPUS, |
| "network_mode": "none", |
| "pids_limit": CONTAINER_PIDS_LIMIT, |
| "privileged": False, |
| "security_opt": [CONTAINER_SECURITY_OPT], |
| } |
| ) |
| if args: |
| args = (self._expected_image_id, *args[1:]) |
| kwargs.pop("image", None) |
| else: |
| kwargs["image"] = self._expected_image_id |
| container = self._collection.create(*args, **kwargs) |
| self._attestations.append( |
| _container_policy_attestation( |
| container, |
| run_id=self._run_id, |
| expected_image_id=self._expected_image_id, |
| ) |
| ) |
| return container |
|
|
|
|
| class _HardenedDockerClient: |
| def __init__( |
| self, |
| client: Any, |
| *, |
| run_id: str, |
| expected_image_id: str, |
| attestations: list[dict[str, Any]], |
| ) -> None: |
| self._client = client |
| self.containers = _HardenedContainerCollection( |
| client.containers, |
| run_id=run_id, |
| expected_image_id=expected_image_id, |
| attestations=attestations, |
| ) |
|
|
| def __getattr__(self, name: str) -> Any: |
| return getattr(self._client, name) |
|
|
|
|
| def _raw_document( |
| value: Mapping[str, Any], *, phase: str, instance_id: str |
| ) -> tuple[ |
| list[str], |
| list[str], |
| dict[str, str], |
| str, |
| ]: |
| if ( |
| value.get("schema_version") != 1 |
| or value.get("phase") != phase |
| or value.get("instance_id") != instance_id |
| ): |
| raise SWEbenchVerificationError("raw status evidence identity is invalid") |
| expected = value.get("expected") |
| statuses = value.get("statuses") |
| resolution = value.get("official_resolution") |
| if ( |
| not isinstance(expected, Mapping) |
| or not isinstance(statuses, Mapping) |
| or not isinstance(resolution, str) |
| ): |
| raise SWEbenchVerificationError("raw status evidence is incomplete") |
| f2p = _string_list(expected.get("FAIL_TO_PASS"), allow_empty=False) |
| p2p = _string_list(expected.get("PASS_TO_PASS"), allow_empty=True) |
| normalized_statuses: dict[str, str] = {} |
| for key, status_value in statuses.items(): |
| if not isinstance(key, str) or not isinstance(status_value, str): |
| raise SWEbenchVerificationError("raw status evidence is malformed") |
| normalized_statuses[key] = status_value |
| return f2p, p2p, normalized_statuses, resolution |
|
|
|
|
| def _validate_parent_artifacts( |
| *, |
| phase: str, |
| instance_id: str, |
| run_id: str, |
| model_name: str, |
| allowed_paths: Sequence[str], |
| expected_image_id: str | None, |
| model_patch: str | None, |
| work_dir: Path, |
| log_dir: Path, |
| ) -> dict[str, Any]: |
| worker = _json_file(work_dir / "worker-result.json", work_dir) |
| if ( |
| worker.get("ok") is not True |
| or worker.get("phase") != phase |
| or worker.get("run_id") != run_id |
| or worker.get("model_name") != model_name |
| or worker.get("log_dir") != log_dir.relative_to(work_dir).as_posix() |
| ): |
| raise SWEbenchVerificationError("worker result identity is invalid") |
| image_id = worker.get("image_id") |
| container_policy = worker.get("container_policy") |
| if ( |
| not isinstance(image_id, str) |
| or not IMAGE_ID_PATTERN.fullmatch(image_id) |
| or (expected_image_id is not None and image_id != expected_image_id) |
| or not isinstance(container_policy, list) |
| or not container_policy |
| or not all( |
| item |
| == { |
| "cap_drop": ["ALL"], |
| "cgroupns_mode": "private", |
| "image_id": image_id, |
| "memory_bytes": CONTAINER_MEMORY_BYTES, |
| "mount_count": 0, |
| "nano_cpus": CONTAINER_NANO_CPUS, |
| "network_mode": "none", |
| "pids_limit": CONTAINER_PIDS_LIMIT, |
| "privileged": False, |
| "security_opt": CONTAINER_SECURITY_OPT, |
| } |
| for item in container_policy |
| ) |
| ): |
| raise SWEbenchVerificationError("worker runtime attestation is invalid") |
| report = _json_file(log_dir / "report.json", work_dir) |
| raw = _json_file(log_dir / "raw-status.json", work_dir) |
| if set(report) != {instance_id} or not isinstance(report.get(instance_id), Mapping): |
| raise SWEbenchVerificationError("official report instance identity is invalid") |
| f2p, p2p, statuses, resolution = _raw_document( |
| raw, |
| phase=phase, |
| instance_id=instance_id, |
| ) |
| outcome = validate_outcome( |
| phase=phase, |
| fail_to_pass=f2p, |
| pass_to_pass=p2p, |
| statuses=statuses, |
| report_entry=report[instance_id], |
| official_resolution=resolution, |
| ) |
| patch = (log_dir / "patch.diff").read_text(encoding="utf-8") |
| if phase == "red": |
| matching_targets = [ |
| path |
| for path in allowed_paths |
| if _safe_production_python_path(path) and patch == mode_only_patch(path) |
| ] |
| if len(matching_targets) != 1: |
| raise SWEbenchVerificationError("red mode-only patch evidence is invalid") |
| outcome["mode_target_sha256"] = _sha256(matching_targets[0].encode()) |
| elif phase == "scored" and patch != model_patch: |
| raise SWEbenchVerificationError("scored patch evidence changed across the worker boundary") |
| outcome["container_policy_count"] = len(container_policy) |
| outcome["image_id"] = image_id |
| return outcome |
|
|
|
|
| def verify_swebench( |
| *, |
| phase: str, |
| dataset_path: Path, |
| instance_id: str, |
| allowed_paths: Sequence[str], |
| swebench_checkout: Path, |
| swebench_python: Path, |
| expected_revision: str, |
| expected_run_evaluation_sha256: str, |
| expected_bridge_sha256: str, |
| expected_dataset_sha256: str, |
| expected_python_sha256: str, |
| expected_python_environment_sha256: str, |
| expected_docker_package_sha256: str, |
| docker_cli: Path, |
| expected_docker_cli_sha256: str, |
| docker_host: str, |
| expected_docker_daemon_id: str, |
| expected_docker_server_version: str, |
| work_dir: Path, |
| timeout: float = 900, |
| model_patch: str | None = None, |
| namespace: str | None = None, |
| expected_image_id: str | None = None, |
| allow_image_pull: bool = False, |
| ) -> dict[str, Any]: |
| """Verify one red, gold-green, or scored prediction with official SWE-bench.""" |
| if phase not in PHASES: |
| raise SWEbenchVerificationError("unsupported SWE-bench verification phase") |
| if not INSTANCE_PATTERN.fullmatch(instance_id): |
| raise SWEbenchVerificationError("SWE-bench instance identifier is invalid") |
| if isinstance(timeout, bool) or not math.isfinite(timeout) or timeout <= 0: |
| raise SWEbenchVerificationError("SWE-bench timeout must be positive and finite") |
| if phase in {"red", "green"} and model_patch is not None: |
| raise SWEbenchVerificationError("red and green phases select their own control patches") |
| if phase == "scored" and (not isinstance(model_patch, str) or not model_patch.strip()): |
| raise SWEbenchVerificationError("scored phase requires a non-empty prediction patch") |
| if namespace != "swebench": |
| raise SWEbenchVerificationError("official SWE-bench namespace is required") |
| if not isinstance(allow_image_pull, bool): |
| raise SWEbenchVerificationError("image acquisition policy is invalid") |
| if allow_image_pull and phase != "red": |
| raise SWEbenchVerificationError("only the acquisition red control may pull an image") |
| if phase in {"green", "scored"} and ( |
| expected_image_id is None or not IMAGE_ID_PATTERN.fullmatch(expected_image_id) |
| ): |
| raise SWEbenchVerificationError("green and scored phases require a pinned image ID") |
| if expected_image_id is not None and not IMAGE_ID_PATTERN.fullmatch(expected_image_id): |
| raise SWEbenchVerificationError("expected image ID is invalid") |
| normalized_allowed = list(allowed_paths) |
| if ( |
| not normalized_allowed |
| or len(normalized_allowed) != len(set(normalized_allowed)) |
| or not all( |
| isinstance(path, str) and _safe_production_python_path(path) |
| for path in normalized_allowed |
| ) |
| ): |
| raise SWEbenchVerificationError("allowed production paths are invalid") |
| if phase == "red" and not any( |
| _safe_production_python_path(path) for path in normalized_allowed |
| ): |
| raise SWEbenchVerificationError("red control has no safe mode-only target") |
|
|
| deadline = time.monotonic() + timeout |
| source_checkout = _checkout(swebench_checkout) |
| python = _executable(swebench_python, label="SWE-bench Python") |
| docker = _executable(docker_cli, label="Docker CLI") |
| dataset = _regular_file(dataset_path, label="SWE-bench dataset") |
| for expected, observed, label in ( |
| (expected_bridge_sha256, _sha256(SCRIPT_PATH.read_bytes()), "bridge"), |
| (expected_dataset_sha256, _sha256(dataset.read_bytes()), "dataset"), |
| (expected_python_sha256, _sha256(python.read_bytes()), "Python"), |
| (expected_docker_cli_sha256, _sha256(docker.read_bytes()), "Docker CLI"), |
| ): |
| if not SHA256_PATTERN.fullmatch(expected) or observed != expected: |
| raise SWEbenchVerificationError(f"SWE-bench {label} identity drifted") |
| authenticated_host, _socket_path = _unix_docker_host(docker_host) |
| private_cwd = _prepare_private_cwd(work_dir) |
| inputs = private_cwd / "inputs" |
| inputs.mkdir(mode=0o700) |
| bridge_snapshot = _snapshot_file( |
| SCRIPT_PATH, |
| inputs=inputs, |
| label="bridge", |
| expected_sha256=expected_bridge_sha256, |
| ) |
| dataset_snapshot = _snapshot_file( |
| dataset, |
| inputs=inputs, |
| label="dataset", |
| expected_sha256=expected_dataset_sha256, |
| ) |
| python_environment = _python_environment_identity( |
| python=python, |
| expected_sha256=expected_python_environment_sha256, |
| cwd=private_cwd, |
| deadline=deadline, |
| ) |
| docker_package_path, docker_package = _python_package_snapshot( |
| python=python, |
| bridge=bridge_snapshot, |
| package="docker", |
| expected_sha256=expected_docker_package_sha256, |
| destination=(inputs / f"docker-package-{expected_docker_package_sha256}" / "docker"), |
| cwd=private_cwd, |
| deadline=deadline, |
| ) |
| git = _git_executable() |
| checkout = inputs / f"swebench-{expected_revision}" |
| authentication = _materialize_authenticated_harness( |
| source_checkout=source_checkout, |
| destination=checkout, |
| git=git, |
| expected_revision=expected_revision, |
| expected_run_evaluation_sha256=expected_run_evaluation_sha256, |
| deadline=deadline, |
| ) |
| os.chmod(inputs, 0o500) |
| environment = worker_environment( |
| checkout=checkout, |
| private_cwd=private_cwd, |
| docker_host=authenticated_host, |
| ) |
| run_id = _new_run_id(phase) |
| if not RUN_ID_PATTERN.fullmatch(run_id): |
| raise SWEbenchVerificationError("generated SWE-bench run ID is invalid") |
| model_name = _model_name(phase) |
| docker_identity = _docker_identity( |
| docker_cli=docker, |
| docker_host=authenticated_host, |
| expected_daemon_id=expected_docker_daemon_id, |
| expected_server_version=expected_docker_server_version, |
| cwd=private_cwd, |
| env=environment, |
| deadline=deadline, |
| ) |
| baseline_inventory, baseline_inventory_evidence = _docker_inventory( |
| docker_cli=docker, |
| docker_host=authenticated_host, |
| cwd=private_cwd, |
| env=environment, |
| deadline=deadline, |
| ) |
| patch_paths: list[str] = [] |
| patch_evidence: dict[str, Any] | None = None |
| if phase == "scored": |
| assert model_patch is not None |
| patch_paths, patch_evidence = _validate_scored_patch( |
| patch=model_patch, |
| allowed_paths=normalized_allowed, |
| git=git, |
| cwd=private_cwd, |
| deadline=deadline, |
| ) |
| request = { |
| "allow_image_pull": allow_image_pull, |
| "allowed_paths": normalized_allowed, |
| "dataset_path": str(dataset_snapshot), |
| "docker_package_path": str(docker_package_path), |
| "expected_docker_daemon_id": expected_docker_daemon_id, |
| "expected_docker_package_sha256": expected_docker_package_sha256, |
| "expected_docker_server_version": expected_docker_server_version, |
| "expected_harness_source_file_count": authentication["source_file_count"], |
| "expected_harness_source_sha256": authentication["source_sha256"], |
| "expected_image_id": expected_image_id, |
| "expected_bridge_sha256": expected_bridge_sha256, |
| "expected_dataset_sha256": expected_dataset_sha256, |
| "expected_revision": expected_revision, |
| "expected_run_evaluation_sha256": expected_run_evaluation_sha256, |
| "git_cli": str(git), |
| "harness_timeout_seconds": max(1, int(_remaining(deadline))), |
| "instance_id": instance_id, |
| "model_name": model_name, |
| "model_patch": model_patch, |
| "model_patch_sha256": _sha256(model_patch.encode()) if model_patch is not None else None, |
| "namespace": namespace, |
| "patch_paths": patch_paths, |
| "phase": phase, |
| "run_id": run_id, |
| "schema_version": 1, |
| "swebench_checkout": str(checkout), |
| } |
| request_path = private_cwd / "worker-request.json" |
| result_path = private_cwd / "worker-result.json" |
| _private_write_json(request_path, request) |
| command = [ |
| str(python), |
| str(bridge_snapshot), |
| "worker", |
| "--request", |
| str(request_path), |
| "--result", |
| str(result_path), |
| ] |
|
|
| process: Any | None = None |
| process_error_type: str | None = None |
| final_inventory: dict[str, frozenset[str]] = {} |
| final_inventory_evidence: dict[str, Any] = {} |
| inventory_error_type: str | None = None |
| try: |
| process = _run_process( |
| command, |
| cwd=private_cwd, |
| env=environment, |
| timeout=_remaining(deadline), |
| contain_descendants=True, |
| ) |
| except BaseException as exc: |
| process_error_type = type(exc).__name__ |
| finally: |
| cleanup = _cleanup_containers( |
| docker_cli=docker, |
| docker_host=authenticated_host, |
| run_id=run_id, |
| cwd=private_cwd, |
| env=environment, |
| ) |
| try: |
| final_inventory, final_inventory_evidence = _docker_inventory( |
| docker_cli=docker, |
| docker_host=authenticated_host, |
| cwd=private_cwd, |
| env=environment, |
| deadline=time.monotonic() + 60, |
| ) |
| except BaseException as exc: |
| inventory_error_type = type(exc).__name__ |
|
|
| process_preimage = { |
| "command": command, |
| "environment": environment, |
| "error_type": process_error_type, |
| "process": { |
| "elapsed": float(process.elapsed) if process is not None else 0.0, |
| "residual_descendants": ( |
| list(process.residual_descendants) if process is not None else [] |
| ), |
| "returncode": int(process.returncode) if process is not None else None, |
| "stderr": str(process.stderr) if process is not None else "", |
| "stdout": str(process.stdout) if process is not None else "", |
| "timed_out": bool(process.timed_out) if process is not None else False, |
| }, |
| } |
| parent_preimage_path = private_cwd / "parent-process.json" |
| _private_write_json(parent_preimage_path, process_preimage) |
|
|
| log_dir = private_cwd / "logs/run_evaluation" / run_id / model_name / instance_id |
| artifact_paths = {name: log_dir / name for name in REQUIRED_ARTIFACTS} |
| artifact_paths["worker-request.json"] = request_path |
| artifact_paths["worker-result.json"] = result_path |
| artifact_paths["parent-process.json"] = parent_preimage_path |
| if phase == "red": |
| artifact_paths["mode-probe.log"] = private_cwd / "mode-probe.log" |
| artifacts: dict[str, Any] = {} |
| artifact_error_type: str | None = None |
| try: |
| artifacts = { |
| name: _artifact_evidence(path, private_cwd) for name, path in artifact_paths.items() |
| } |
| except SWEbenchVerificationError as exc: |
| artifact_error_type = type(exc).__name__ |
|
|
| evidence: dict[str, Any] = { |
| "artifacts": artifacts, |
| "authentication": authentication, |
| "cleanup": cleanup, |
| "command": _command_evidence(command), |
| "docker_identity": docker_identity, |
| "docker_package": docker_package, |
| "environment": { |
| "keys": sorted(environment), |
| "sha256": _sha256(_canonical_bytes(environment)), |
| }, |
| "inventory": { |
| "baseline": baseline_inventory_evidence, |
| "final": final_inventory_evidence, |
| }, |
| "input_snapshots": { |
| "bridge_sha256": _sha256(bridge_snapshot.read_bytes()), |
| "dataset_sha256": _sha256(dataset_snapshot.read_bytes()), |
| "docker_package": _package_manifest(docker_package_path), |
| "harness_source": _package_manifest(checkout / "swebench"), |
| }, |
| "model_name": model_name, |
| "patch": patch_evidence, |
| "phase": phase, |
| "process": _process_evidence(process, error_type=process_error_type), |
| "python_environment": python_environment, |
| "run_id": run_id, |
| "schema_version": 1, |
| } |
| audit_path = private_cwd / "verification-evidence.json" |
| audit = { |
| "cleanup": { |
| "command_count": len(cleanup["commands"]), |
| "error_count": len(cleanup["errors"]), |
| "ok": cleanup["ok"], |
| "residual_container_count": len(cleanup["residual_container_ids"]), |
| "verification_complete": cleanup["verification_complete"], |
| }, |
| "inventory": { |
| stage: { |
| name: { |
| "count": record.get("count"), |
| "sha256": record.get("sha256"), |
| } |
| for name, record in records.items() |
| } |
| for stage, records in evidence["inventory"].items() |
| }, |
| "phase": phase, |
| "process": evidence["process"], |
| "run_id": run_id, |
| "schema_version": 1, |
| } |
| try: |
| _private_write_json(audit_path, audit) |
| artifacts[audit_path.name] = _artifact_evidence(audit_path, private_cwd) |
| except (OSError, SWEbenchVerificationError) as exc: |
| evidence["audit_error_type"] = type(exc).__name__ |
| raise SWEbenchVerificationError(evidence=evidence) from exc |
| if artifact_error_type: |
| evidence["artifact_error_type"] = artifact_error_type |
| raise SWEbenchVerificationError(evidence=evidence) |
| if not cleanup["ok"]: |
| raise SWEbenchVerificationError( |
| "SWE-bench Docker containment failed", |
| evidence=evidence, |
| ) |
| if evidence["input_snapshots"] != { |
| "bridge_sha256": expected_bridge_sha256, |
| "dataset_sha256": expected_dataset_sha256, |
| "docker_package": { |
| "file_count": docker_package["file_count"], |
| "sha256": expected_docker_package_sha256, |
| }, |
| "harness_source": { |
| "file_count": authentication["source_file_count"], |
| "sha256": authentication["source_sha256"], |
| }, |
| }: |
| raise SWEbenchVerificationError( |
| "SWE-bench authenticated inputs drifted", |
| evidence=evidence, |
| ) |
| containment_drift = ( |
| inventory_error_type is not None |
| or not final_inventory |
| or baseline_inventory.get("containers") != final_inventory.get("containers") |
| or baseline_inventory.get("networks") != final_inventory.get("networks") |
| or baseline_inventory.get("volumes") != final_inventory.get("volumes") |
| or ( |
| not allow_image_pull |
| and baseline_inventory.get("images") != final_inventory.get("images") |
| ) |
| or ( |
| allow_image_pull |
| and not baseline_inventory.get("images", frozenset()).issubset( |
| final_inventory.get("images", frozenset()) |
| ) |
| ) |
| ) |
| if containment_drift: |
| evidence["inventory_error_type"] = inventory_error_type |
| raise SWEbenchVerificationError( |
| "SWE-bench Docker inventory drifted", |
| evidence=evidence, |
| ) |
| if ( |
| process is None |
| or process.returncode |
| or process.timed_out |
| or process.residual_descendants |
| or process_error_type |
| ): |
| raise SWEbenchVerificationError(evidence=evidence) |
| if any(artifacts[name]["present"] is not True for name in REQUIRED_ARTIFACTS): |
| raise SWEbenchVerificationError(evidence=evidence) |
| try: |
| evidence["validation"] = _validate_parent_artifacts( |
| phase=phase, |
| instance_id=instance_id, |
| run_id=run_id, |
| model_name=model_name, |
| allowed_paths=normalized_allowed, |
| expected_image_id=expected_image_id, |
| model_patch=model_patch, |
| work_dir=private_cwd, |
| log_dir=log_dir, |
| ) |
| if allow_image_pull: |
| added_images = final_inventory["images"] - baseline_inventory["images"] |
| if added_images not in ( |
| frozenset(), |
| frozenset({str(evidence["validation"]["image_id"])}), |
| ): |
| raise SWEbenchVerificationError("official image pull changed extra images") |
| except SWEbenchVerificationError as exc: |
| raise SWEbenchVerificationError(evidence=evidence) from exc |
| return evidence |
|
|
|
|
| def _worker_authenticate(request: Mapping[str, Any]) -> Path: |
| checkout = _checkout(Path(str(request["swebench_checkout"]))) |
| expected_revision = str(request["expected_revision"]) |
| expected_sha256 = str(request["expected_run_evaluation_sha256"]) |
| expected_bridge_sha256 = str(request["expected_bridge_sha256"]) |
| expected_source_sha256 = str(request.get("expected_harness_source_sha256") or "") |
| expected_source_file_count = request.get("expected_harness_source_file_count") |
| git = _executable(Path(str(request["git_cli"])), label="Git executable") |
| run_evaluation = _regular_file( |
| checkout / "swebench/harness/run_evaluation.py", |
| label="run_evaluation.py", |
| ) |
| if ( |
| not REVISION_PATTERN.fullmatch(expected_revision) |
| or not SHA256_PATTERN.fullmatch(expected_sha256) |
| or not SHA256_PATTERN.fullmatch(expected_bridge_sha256) |
| or _sha256(SCRIPT_PATH.read_bytes()) != expected_bridge_sha256 |
| or _sha256(run_evaluation.read_bytes()) != expected_sha256 |
| or not SHA256_PATTERN.fullmatch(expected_source_sha256) |
| or not isinstance(expected_source_file_count, int) |
| or isinstance(expected_source_file_count, bool) |
| or expected_source_file_count < 1 |
| ): |
| raise SWEbenchVerificationError("SWE-bench worker authentication failed") |
| revision = subprocess.run( |
| [str(git), "-C", str(checkout), "rev-parse", "--verify", "HEAD^{commit}"], |
| cwd=checkout, |
| env=_auth_environment(), |
| check=False, |
| capture_output=True, |
| text=True, |
| timeout=15, |
| ) |
| dirty = subprocess.run( |
| [ |
| str(git), |
| "-C", |
| str(checkout), |
| "status", |
| "--porcelain=v1", |
| "--untracked-files=all", |
| "--ignored=matching", |
| ], |
| cwd=checkout, |
| env=_auth_environment(), |
| check=False, |
| capture_output=True, |
| text=True, |
| timeout=15, |
| ) |
| if ( |
| revision.returncode |
| or revision.stdout.strip() != expected_revision |
| or dirty.returncode |
| or dirty.stdout.strip() |
| ): |
| raise SWEbenchVerificationError("SWE-bench worker authentication failed") |
| source_manifest = _package_manifest(checkout / "swebench") |
| if source_manifest != { |
| "file_count": expected_source_file_count, |
| "sha256": expected_source_sha256, |
| }: |
| raise SWEbenchVerificationError("SWE-bench worker source snapshot drifted") |
| return run_evaluation |
|
|
|
|
| def _require_authenticated_module(module: Any, checkout: Path) -> None: |
| raw_path = getattr(module, "__file__", None) |
| if not isinstance(raw_path, str): |
| raise SWEbenchVerificationError("imported SWE-bench module has no source path") |
| try: |
| Path(raw_path).resolve(strict=True).relative_to(checkout.resolve(strict=True)) |
| except (OSError, ValueError) as exc: |
| raise SWEbenchVerificationError( |
| "imported SWE-bench module escaped the authenticated checkout" |
| ) from exc |
|
|
|
|
| def _tracked_mode_map( |
| *, |
| spec: Any, |
| client: Any, |
| allowed_paths: Sequence[str], |
| run_id: str, |
| docker_build: Any, |
| docker_utils: Any, |
| ) -> dict[str, str]: |
| docker_logging = importlib.import_module("swebench.harness.docker_build") |
| logger = docker_logging.setup_logger(f"{run_id}-mode-probe", Path("mode-probe.log")) |
| container = None |
| tracked: dict[str, str] = {} |
| try: |
| container = docker_build.build_container( |
| spec, |
| client, |
| f"{run_id}-mode-probe", |
| logger, |
| False, |
| False, |
| ) |
| container.start() |
| for candidate in sorted(set(allowed_paths)): |
| if not _safe_production_python_path(candidate): |
| continue |
| result = container.exec_run( |
| ["git", "ls-files", "--stage", "--", candidate], |
| workdir="/testbed", |
| user="root", |
| ) |
| if result.exit_code: |
| continue |
| output = result.output.decode("utf-8", errors="strict").strip() |
| lines = output.splitlines() |
| if len(lines) != 1 or "\t" not in lines[0]: |
| continue |
| metadata, observed_path = lines[0].split("\t", 1) |
| fields = metadata.split() |
| if ( |
| len(fields) == 3 |
| and fields[0] in {"100644", "100755"} |
| and fields[2] == "0" |
| and observed_path == candidate |
| ): |
| tracked[candidate] = fields[0] |
| finally: |
| docker_utils.cleanup_container(client, container, logger) |
| docker_logging.close_logger(logger) |
| return tracked |
|
|
|
|
| def _run_official_instance(run_evaluation: Any, *args: Any) -> Any: |
| previous_umask = os.umask(0o022) |
| try: |
| return run_evaluation.run_instance(*args) |
| finally: |
| os.umask(previous_umask) |
|
|
|
|
| def _worker_run(request: Mapping[str, Any], result_path: Path) -> None: |
| os.umask(0o077) |
| run_evaluation_path = _worker_authenticate(request) |
| checkout = Path(str(request["swebench_checkout"])).resolve(strict=True) |
| expected_docker_package_sha256 = request.get("expected_docker_package_sha256") |
| docker_package_path = Path(str(request.get("docker_package_path") or "")) |
| if ( |
| not isinstance(expected_docker_package_sha256, str) |
| or not SHA256_PATTERN.fullmatch(expected_docker_package_sha256) |
| or not docker_package_path.is_absolute() |
| or docker_package_path.is_symlink() |
| or not docker_package_path.is_dir() |
| or docker_package_path.name != "docker" |
| or _package_manifest(docker_package_path)["sha256"] != expected_docker_package_sha256 |
| ): |
| raise SWEbenchVerificationError("worker Docker package identity drifted") |
| sys.path.insert(0, str(docker_package_path.parent)) |
| importlib.invalidate_caches() |
| docker_module = importlib.import_module("docker") |
| try: |
| Path(str(docker_module.__file__)).resolve(strict=True).relative_to( |
| docker_package_path.resolve(strict=True) |
| ) |
| except ValueError as exc: |
| raise SWEbenchVerificationError( |
| "worker Docker package escaped the authenticated snapshot" |
| ) from exc |
| if _package_manifest(docker_package_path)["sha256"] != expected_docker_package_sha256: |
| raise SWEbenchVerificationError("worker Docker package changed during import") |
| run_evaluation = importlib.import_module("swebench.harness.run_evaluation") |
| if Path(str(run_evaluation.__file__)).resolve() != run_evaluation_path: |
| raise SWEbenchVerificationError("imported SWE-bench harness path is not authenticated") |
| constants = importlib.import_module("swebench.harness.constants") |
| docker_build = importlib.import_module("swebench.harness.docker_build") |
| docker_utils = importlib.import_module("swebench.harness.docker_utils") |
| grading = importlib.import_module("swebench.harness.grading") |
| test_spec_module = importlib.import_module("swebench.harness.test_spec.test_spec") |
| harness_utils = importlib.import_module("swebench.harness.utils") |
| for module in ( |
| run_evaluation, |
| constants, |
| docker_build, |
| docker_utils, |
| grading, |
| test_spec_module, |
| harness_utils, |
| ): |
| _require_authenticated_module(module, checkout) |
|
|
| phase = str(request["phase"]) |
| instance_id = str(request["instance_id"]) |
| run_id = str(request["run_id"]) |
| model_name = str(request["model_name"]) |
| if ( |
| request.get("schema_version") != 1 |
| or phase not in PHASES |
| or not INSTANCE_PATTERN.fullmatch(instance_id) |
| or not RUN_ID_PATTERN.fullmatch(run_id) |
| or model_name != _model_name(phase) |
| ): |
| raise SWEbenchVerificationError("worker request identity is invalid") |
| dataset_path = _regular_file( |
| Path(str(request["dataset_path"])), |
| label="SWE-bench dataset", |
| ) |
| expected_dataset_sha256 = request.get("expected_dataset_sha256") |
| if ( |
| not isinstance(expected_dataset_sha256, str) |
| or not SHA256_PATTERN.fullmatch(expected_dataset_sha256) |
| or _sha256(dataset_path.read_bytes()) != expected_dataset_sha256 |
| ): |
| raise SWEbenchVerificationError("worker dataset identity drifted") |
| rows = harness_utils.load_swebench_dataset(str(dataset_path), "test", [instance_id]) |
| if len(rows) != 1 or rows[0].get("instance_id") != instance_id: |
| raise SWEbenchVerificationError("worker dataset selection is not exact") |
| row = rows[0] |
| namespace = request.get("namespace") |
| if namespace != "swebench": |
| raise SWEbenchVerificationError("worker official namespace is invalid") |
| spec = test_spec_module.make_test_spec( |
| row, |
| namespace=namespace, |
| base_image_tag="latest", |
| env_image_tag="latest", |
| instance_image_tag="latest", |
| arch="x86_64", |
| ) |
| client = docker_module.from_env() |
| try: |
| info = client.info() |
| version = client.version() |
| if ( |
| not isinstance(info, Mapping) |
| or info.get("ID") != request.get("expected_docker_daemon_id") |
| or not isinstance(version, Mapping) |
| or version.get("Version") != request.get("expected_docker_server_version") |
| ): |
| raise SWEbenchVerificationError("worker Docker daemon identity drifted") |
| allow_image_pull = request.get("allow_image_pull") |
| expected_image_id = request.get("expected_image_id") |
| if not isinstance(allow_image_pull, bool) or ( |
| expected_image_id is not None |
| and ( |
| not isinstance(expected_image_id, str) |
| or not IMAGE_ID_PATTERN.fullmatch(expected_image_id) |
| ) |
| ): |
| raise SWEbenchVerificationError("worker image policy is invalid") |
| try: |
| initial_image = client.images.get(spec.instance_image_key) |
| initial_image_id: str | None = str(initial_image.id) |
| except docker_module.errors.ImageNotFound: |
| initial_image_id = None |
| if initial_image_id is None and allow_image_pull: |
| client.images.pull(spec.instance_image_key) |
| initial_image = client.images.get(spec.instance_image_key) |
| initial_image_id = str(initial_image.id) |
| if not allow_image_pull and initial_image_id is None: |
| raise SWEbenchVerificationError("pinned instance image is unavailable") |
| if initial_image_id is None or not IMAGE_ID_PATTERN.fullmatch(initial_image_id): |
| raise SWEbenchVerificationError("instance image identity is invalid") |
| if expected_image_id is not None and initial_image_id != expected_image_id: |
| raise SWEbenchVerificationError("pinned instance image drifted") |
| pinned_image_id = initial_image_id |
| container_policy: list[dict[str, Any]] = [] |
| hardened_client = _HardenedDockerClient( |
| client, |
| run_id=run_id, |
| expected_image_id=pinned_image_id, |
| attestations=container_policy, |
| ) |
| allowed_paths_value = request.get("allowed_paths") |
| if not isinstance(allowed_paths_value, list) or not all( |
| isinstance(path, str) for path in allowed_paths_value |
| ): |
| raise SWEbenchVerificationError("worker allowed paths are invalid") |
| allowed_paths = list(allowed_paths_value) |
| mode_target: str | None = None |
| patch: str |
| if phase == "red": |
| tracked = _tracked_mode_map( |
| spec=spec, |
| client=hardened_client, |
| allowed_paths=allowed_paths, |
| run_id=run_id, |
| docker_build=docker_build, |
| docker_utils=docker_utils, |
| ) |
| mode_target = select_mode_target(allowed_paths, tracked) |
| patch = mode_only_patch(mode_target) |
| elif phase == "green": |
| gold_patch = row.get("patch") |
| if not isinstance(gold_patch, str) or not gold_patch.strip(): |
| raise SWEbenchVerificationError("gold patch is unavailable") |
| patch = gold_patch |
| else: |
| prediction_patch = request.get("model_patch") |
| if not isinstance(prediction_patch, str) or not prediction_patch.strip(): |
| raise SWEbenchVerificationError("scored prediction patch is unavailable") |
| if _sha256(prediction_patch.encode()) != request.get("model_patch_sha256"): |
| raise SWEbenchVerificationError("scored prediction patch identity drifted") |
| patch_paths = request.get("patch_paths") |
| if ( |
| not isinstance(patch_paths, list) |
| or not patch_paths |
| or not all(isinstance(path, str) and path in allowed_paths for path in patch_paths) |
| ): |
| raise SWEbenchVerificationError("scored prediction paths are invalid") |
| patch = prediction_patch |
|
|
| timeout_value = request.get("harness_timeout_seconds") |
| if ( |
| isinstance(timeout_value, bool) |
| or not isinstance(timeout_value, int) |
| or timeout_value <= 0 |
| ): |
| raise SWEbenchVerificationError("worker harness timeout is invalid") |
| prediction = { |
| constants.KEY_INSTANCE_ID: instance_id, |
| constants.KEY_MODEL: model_name, |
| constants.KEY_PREDICTION: patch, |
| } |
| |
| |
| result = _run_official_instance( |
| run_evaluation, |
| spec, |
| prediction, |
| False, |
| False, |
| |
| |
| hardened_client, |
| run_id, |
| timeout_value, |
| False, |
| ) |
| if result.get("completed") is not True: |
| raise SWEbenchVerificationError("official SWE-bench instance did not complete") |
| log_dir = constants.RUN_EVALUATION_LOG_DIR / run_id / model_name / instance_id |
| report_path = log_dir / constants.LOG_REPORT |
| test_output_path = log_dir / constants.LOG_TEST_OUTPUT |
| report = json.loads(report_path.read_text(encoding="utf-8")) |
| if set(report) != {instance_id} or not isinstance(report.get(instance_id), Mapping): |
| raise SWEbenchVerificationError("official report identity is invalid") |
| statuses, found = grading.get_logs_eval(spec, str(test_output_path)) |
| if not found or not isinstance(statuses, dict): |
| raise SWEbenchVerificationError("official raw test statuses are unavailable") |
| tests_status = report[instance_id].get("tests_status") |
| if not isinstance(tests_status, Mapping): |
| raise SWEbenchVerificationError("official grouped test statuses are unavailable") |
| resolution = grading.get_resolution_status(tests_status) |
| raw_status = { |
| "expected": { |
| "FAIL_TO_PASS": list(spec.FAIL_TO_PASS), |
| "PASS_TO_PASS": list(spec.PASS_TO_PASS), |
| }, |
| "instance_id": instance_id, |
| "official_resolution": resolution, |
| "phase": phase, |
| "schema_version": 1, |
| "statuses": statuses, |
| } |
| _private_write_json(log_dir / "raw-status.json", raw_status) |
| validate_outcome( |
| phase=phase, |
| fail_to_pass=spec.FAIL_TO_PASS, |
| pass_to_pass=spec.PASS_TO_PASS, |
| statuses=statuses, |
| report_entry=report[instance_id], |
| official_resolution=resolution, |
| ) |
| _worker_authenticate(request) |
| if ( |
| _sha256(dataset_path.read_bytes()) != expected_dataset_sha256 |
| or _package_manifest(docker_package_path)["sha256"] != expected_docker_package_sha256 |
| or _sha256(SCRIPT_PATH.read_bytes()) != request.get("expected_bridge_sha256") |
| ): |
| raise SWEbenchVerificationError("authenticated worker inputs changed during use") |
| final_image = client.images.get(spec.instance_image_key) |
| image_id = str(final_image.id) |
| if image_id != pinned_image_id or ( |
| expected_image_id is not None and image_id != expected_image_id |
| ): |
| raise SWEbenchVerificationError("instance image identity drifted") |
| if not container_policy: |
| raise SWEbenchVerificationError("no evaluation container policy was attested") |
| worker_result: dict[str, Any] = { |
| "container_policy": container_policy, |
| "image_id": image_id, |
| "log_dir": log_dir.as_posix(), |
| "model_name": model_name, |
| "ok": True, |
| "phase": phase, |
| "run_id": run_id, |
| } |
| if mode_target is not None: |
| worker_result["mode_target"] = mode_target |
| _private_write_json(result_path, worker_result) |
| finally: |
| client.close() |
|
|
|
|
| def _worker_main(request_path: Path, result_path: Path) -> int: |
| os.umask(0o077) |
| try: |
| request_value = json.loads(request_path.read_text(encoding="utf-8")) |
| if not isinstance(request_value, dict): |
| raise SWEbenchVerificationError("worker request must be an object") |
| _worker_run(request_value, result_path) |
| except Exception as exc: |
| try: |
| _private_write_json( |
| result_path, |
| { |
| "error_type": type(exc).__name__, |
| "ok": False, |
| "schema_version": 1, |
| }, |
| ) |
| except OSError: |
| pass |
| print("official SWE-bench worker failed", file=sys.stderr) |
| return 2 |
| return 0 |
|
|
|
|
| def main(argv: Sequence[str] | None = None) -> int: |
| parser = argparse.ArgumentParser(description=__doc__) |
| subparsers = parser.add_subparsers(dest="command", required=True) |
| worker = subparsers.add_parser("worker") |
| worker.add_argument("--request", type=Path, required=True) |
| worker.add_argument("--result", type=Path, required=True) |
| package_manifest = subparsers.add_parser("package-manifest") |
| package_manifest.add_argument("--package", required=True) |
| package_snapshot = subparsers.add_parser("package-snapshot") |
| package_snapshot.add_argument("--package", required=True) |
| package_snapshot.add_argument("--destination", type=Path, required=True) |
| args = parser.parse_args(argv) |
| if args.command == "worker": |
| return _worker_main(args.request, args.result) |
| if args.command == "package-manifest": |
| print(json.dumps(_package_manifest_for_name(args.package), sort_keys=True)) |
| return 0 |
| if args.command == "package-snapshot": |
| print( |
| json.dumps( |
| _package_snapshot_for_name(args.package, args.destination), |
| sort_keys=True, |
| ) |
| ) |
| return 0 |
| parser.error("unsupported command") |
| return 2 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|