88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 1 | #!/usr/bin/env python3
|
| 2 | """The full task loop, end to end β the product's signature motion.
|
| 3 |
|
| 4 | oak: mount -> edit -> commit -> push -> desc -> end, then REMOUNT for a
|
| 5 | follow-up task (the warm-cache story) and repeat edit/commit/push/end.
|
| 6 | git: worktree add -> branch edit -> add/commit -> push -> worktree remove,
|
| 7 | then a second worktree for the follow-up. The task description lives in
|
| 8 | the commit message git already paid for; oak pays a separate `desc`
|
| 9 | call. That asymmetry is the measurement, not a harness gap
|
| 10 | (docs/command-semantics.md): this is agent-default-track, whole-task
|
| 11 | evidence, never a step-by-step equivalence claim.
|
| 12 |
|
| 13 | Transport honesty: oak rows always cross the network to a real oak server.
|
| 14 | git runs against GitHub (--git-remote github, transport-comparable) or a
|
| 15 | local bare repo (--git-remote local, the core-equivalent track). Rows carry
|
| 16 | remote_transport/remote_server; latency comparisons across transports are
|
| 17 | forbidden β cross-subject evidence at unlike transports is tool calls and
|
| 18 | tokens only.
|
| 19 |
|
| 20 | Setup:
|
| 21 | OAK_BENCH_SYNC_REPO=oak/bench-sync-tmp \\
|
| 22 | [email protected]:oakdotspace/bench-task-loop-tmp.git \\
|
| 23 | python3 scripts/task_loop.py --runs 3
|
| 24 |
|
| 25 | Unconfigured remotes emit skip rows (returncode 77), never silent absence.
|
| 26 | """
|
| 27 |
|
| 28 | from __future__ import annotations
|
| 29 |
|
| 30 | import argparse
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 31 | import hashlib
|
| 32 | import json
|
| 33 | import os
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 34 | import platform
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 35 | import random
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 36 | import re
|
| 37 | import shutil
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 38 | import socket
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 39 | import sqlite3
|
| 40 | import stat
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 41 | import subprocess
|
| 42 | import tempfile
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 43 | import threading
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 44 | import time
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 45 | import urllib.error
|
| 46 | import urllib.parse
|
| 47 | import urllib.request
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 48 | from datetime import datetime, timezone
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 49 | from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 50 | from pathlib import Path
|
| 51 | from typing import Any
|
| 52 |
|
| 53 | from oakbench import remotes as oakbench_remotes
|
| 54 | from oakbench import tokens as oakbench_tokens
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 55 | from oakbench.byteproxy import CountingProxy
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 56 | from oakbench.diskprobe import bounded_tree_usage
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 57 | from oakbench.environment import ENV_ISOLATION_VERSION, base_env as oakbench_base_env
|
| 58 | from oakbench.environment import command_display
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 59 | from oakbench.execution import run_timed, stop_owned_process
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 60 | from oakbench.results import ResultsStore
|
| 61 | from oakbench.rows import row_returncode
|
| 62 | from oakbench.runlock import measurement_lock
|
| 63 | from oakbench.subjects import (
|
| 64 | DEFAULT_OAK_REPO,
|
| 65 | Subject,
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 66 | binary_sha256,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 67 | load_subjects,
|
| 68 | source_metadata,
|
| 69 | subject_details,
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 70 | subject_version,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 71 | subject_versions,
|
| 72 | )
|
| 73 |
|
| 74 | ROOT = Path(__file__).resolve().parents[1]
|
| 75 | DEFAULT_WORKDIR = Path(tempfile.gettempdir()) / "oak-task-loop"
|
| 76 | SKIP_RETURNCODE = 77
|
| 77 | VIRTUAL_BRANCH_RE = re.compile(r"virtual branch ([^\s)]+)")
|
| 78 | PHASES = ("task", "followup")
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 79 | TASK_LOOP_SCENARIO = "task_loop"
|
| 80 | AGENT_TASK_SCENARIO = "agent_task_baseline_v1"
|
| 81 | AGENT_TASK_OAK_LOCAL_SCENARIO = "agent_task_oak_local_v1"
|
| 82 | INTERRUPTED_PUBLICATION_SCENARIO = "interrupted_publication_after_request_v1"
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 83 | LEGACY_SELECTED_MATERIALIZATION_SCENARIO = "legacy_selected_clone_materialization_v1"
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 84 | INTERRUPTED_PUBLICATION_REQUEST_TIMEOUT_S = 2.0
|
| 85 | INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S = 30.0
|
| 86 | INTERRUPTED_PUBLICATION_MAX_REQUEST_BYTES = 64 * 1024 * 1024
|
| 87 | LOCAL_OAK_SERVER_READY_TIMEOUT_S = 10.0
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 88 |
|
| 89 |
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 90 | def legacy_selected_receipt_evidence(receipt: dict[str, Any]) -> dict[str, Any]:
|
| 91 | result = receipt.get("result") if isinstance(receipt.get("result"), dict) else {}
|
| 92 | observed = receipt.get("observed") if isinstance(receipt.get("observed"), dict) else {}
|
| 93 | phases = observed.get("pull_phases") if isinstance(observed.get("pull_phases"), list) else []
|
| 94 | phase_names = [phase.get("name") for phase in phases if isinstance(phase, dict)]
|
| 95 | materialization = (
|
| 96 | observed.get("materialization")
|
| 97 | if isinstance(observed.get("materialization"), dict)
|
| 98 | else {}
|
| 99 | )
|
| 100 | return {
|
| 101 | "legacy_route_observed": phase_names == ["initial", "legacy_selected_branch"],
|
| 102 | "pull_phase_names": phase_names,
|
| 103 | "pull_phases": phases,
|
| 104 | "materialization_passes_observed": materialization.get("passes_observed"),
|
| 105 | "written_paths": materialization.get("written_paths"),
|
| 106 | "snapshot_bound": result.get("snapshot_bound"),
|
| 107 | "head": result.get("head"),
|
| 108 | "manifest": result.get("manifest"),
|
| 109 | }
|
| 110 |
|
| 111 |
|
| 112 | def legacy_selected_receipt_contract(evidence: dict[str, Any]) -> dict[str, Any]:
|
| 113 | phases = []
|
| 114 | for phase in evidence.get("pull_phases", []):
|
| 115 | if not isinstance(phase, dict):
|
| 116 | continue
|
| 117 | phases.append(
|
| 118 | {
|
| 119 | key: value
|
| 120 | for key, value in phase.items()
|
| 121 | if key not in {"materialization_passes"}
|
| 122 | }
|
| 123 | )
|
| 124 | return {
|
| 125 | "legacy_route_observed": evidence.get("legacy_route_observed"),
|
| 126 | "pull_phases": phases,
|
| 127 | "snapshot_bound": evidence.get("snapshot_bound"),
|
| 128 | "head": evidence.get("head"),
|
| 129 | "manifest": evidence.get("manifest"),
|
| 130 | "written_paths": evidence.get("written_paths"),
|
| 131 | }
|
| 132 |
|
| 133 |
|
| 134 | def normalized_cache_contract(snapshot: dict[str, Any] | None) -> list[dict[str, Any]] | None:
|
| 135 | if not isinstance(snapshot, dict) or not isinstance(snapshot.get("rows"), list):
|
| 136 | return None
|
| 137 | return [
|
| 138 | {key: row.get(key) for key in ("path", "size", "blob_hash")}
|
| 139 | for row in snapshot["rows"]
|
| 140 | if isinstance(row, dict)
|
| 141 | ]
|
| 142 |
|
| 143 |
|
| 144 | def normalized_status(value: dict[str, Any] | list[Any] | None) -> dict[str, Any] | None:
|
| 145 | if not isinstance(value, dict):
|
| 146 | return None
|
| 147 | return {
|
| 148 | key: value.get(key)
|
| 149 | for key in ("branch", "parent", "branch_status", "head", "working_changes", "branch_changes")
|
| 150 | }
|
| 151 |
|
| 152 |
|
| 153 | def materialization_timing_controls_met(
|
| 154 | metadata: dict[str, Any], runs: int, randomize_order: bool, controls_requested: bool
|
| 155 | ) -> bool:
|
| 156 | return bool(
|
| 157 | controls_requested
|
| 158 | and metadata.get("measurement_lock") == "held"
|
| 159 | and randomize_order
|
| 160 | and runs >= 10
|
| 161 | )
|
| 162 |
|
| 163 |
|
| 164 | def isolated_oak_fixture_env(task_config_root: Path) -> dict[str, str]:
|
| 165 | task_config_root.mkdir(mode=0o700, parents=True, exist_ok=True)
|
| 166 | retained = {
|
| 167 | "PATH",
|
| 168 | "TMPDIR",
|
| 169 | "TMP",
|
| 170 | "TEMP",
|
| 171 | "HOME",
|
| 172 | "USER",
|
| 173 | "USERNAME",
|
| 174 | "SHELL",
|
| 175 | "SYSTEMROOT",
|
| 176 | "WINDIR",
|
| 177 | "COMSPEC",
|
| 178 | "PATHEXT",
|
| 179 | "SSL_CERT_FILE",
|
| 180 | "SSL_CERT_DIR",
|
| 181 | }
|
| 182 | env = {key: value for key, value in os.environ.items() if key in retained}
|
| 183 | env.update(
|
| 184 | {
|
| 185 | "GIT_AUTHOR_NAME": "Oak Materialization Bench",
|
| 186 | "GIT_AUTHOR_EMAIL": "[email protected]",
|
| 187 | "GIT_COMMITTER_NAME": "Oak Materialization Bench",
|
| 188 | "GIT_COMMITTER_EMAIL": "[email protected]",
|
| 189 | "GIT_CONFIG_GLOBAL": os.devnull,
|
| 190 | "GIT_CONFIG_NOSYSTEM": "1",
|
| 191 | "OAK_AUTHOR": "oak-materialization-bench",
|
| 192 | "OAK_NO_UPDATE_CHECK": "1",
|
| 193 | "OAK_BENCH_ENV_ISOLATION_VERSION": ENV_ISOLATION_VERSION,
|
| 194 | "NO_COLOR": "1",
|
| 195 | "CLICOLOR": "0",
|
| 196 | "LC_ALL": "C",
|
| 197 | "LANG": "C",
|
| 198 | "TZ": "UTC",
|
| 199 | "GIT_AUTHOR_DATE": "2026-01-01T00:00:00Z",
|
| 200 | "GIT_COMMITTER_DATE": "2026-01-01T00:00:00Z",
|
| 201 | }
|
| 202 | )
|
| 203 | return env
|
| 204 |
|
| 205 |
|
| 206 | def normalized_worktree_snapshot(root: Path) -> list[dict[str, Any]]:
|
| 207 | rows: list[dict[str, Any]] = []
|
| 208 | for path in sorted(root.rglob("*")):
|
| 209 | relative = path.relative_to(root)
|
| 210 | if relative.parts and relative.parts[0] in {".oak", ".git"}:
|
| 211 | continue
|
| 212 | info = path.lstat()
|
| 213 | if stat.S_ISDIR(info.st_mode):
|
| 214 | continue
|
| 215 | row: dict[str, Any] = {
|
| 216 | "path": relative.as_posix(),
|
| 217 | "mode": stat.S_IMODE(info.st_mode),
|
| 218 | }
|
| 219 | if stat.S_ISLNK(info.st_mode):
|
| 220 | row.update({"kind": "symlink", "target": os.readlink(path)})
|
| 221 | else:
|
| 222 | row.update({"kind": "file", "sha256": hashlib.sha256(path.read_bytes()).hexdigest()})
|
| 223 | rows.append(row)
|
| 224 | return rows
|
| 225 |
|
| 226 |
|
| 227 | def json_command(
|
| 228 | command: list[str], cwd: Path, env: dict[str, str]
|
| 229 | ) -> dict[str, Any] | list[Any] | None:
|
| 230 | result = run_bounded_harness_command(command, cwd, env=env)
|
| 231 | if result.returncode != 0:
|
| 232 | return None
|
| 233 | try:
|
| 234 | value = json.loads(result.stdout)
|
| 235 | except (TypeError, ValueError):
|
| 236 | return None
|
| 237 | return value if isinstance(value, (dict, list)) else None
|
| 238 |
|
| 239 |
|
| 240 | def normalized_history(value: dict[str, Any] | list[Any] | None) -> list[dict[str, Any]] | None:
|
| 241 | if not isinstance(value, list):
|
| 242 | return None
|
| 243 | rows: list[dict[str, Any]] = []
|
| 244 | for item in value:
|
| 245 | if not isinstance(item, dict):
|
| 246 | return None
|
| 247 | rows.append(
|
| 248 | {
|
| 249 | key: item.get(key)
|
| 250 | for key in (
|
| 251 | "hash",
|
| 252 | "timestamp",
|
| 253 | "branch",
|
| 254 | "description_or_subject",
|
| 255 | "files_changed",
|
| 256 | )
|
| 257 | }
|
| 258 | )
|
| 259 | return rows
|
| 260 |
|
| 261 |
|
| 262 | def normalized_oak_artifacts(root: Path) -> list[str] | None:
|
| 263 | oak_dir = root / ".oak"
|
| 264 | if not oak_dir.is_dir():
|
| 265 | return None
|
| 266 | return sorted(
|
| 267 | path.relative_to(oak_dir).as_posix() + ("/" if path.is_dir() else "")
|
| 268 | for path in oak_dir.rglob("*")
|
| 269 | )
|
| 270 |
|
| 271 |
|
| 272 | def stat_cache_snapshot(root: Path) -> dict[str, Any] | None:
|
| 273 | db = root / ".oak/oak.db"
|
| 274 | if not db.is_file():
|
| 275 | return {"rows": None, "all_metadata_valid": False, "error": "oak_db_missing"}
|
| 276 | try:
|
| 277 | connection = sqlite3.connect(str(db))
|
| 278 | connection.execute("PRAGMA query_only = ON")
|
| 279 | raw_rows = connection.execute(
|
| 280 | "SELECT path, mtime_ns, ctime_ns, size, blob_hash FROM stat_cache ORDER BY path"
|
| 281 | ).fetchall()
|
| 282 | except sqlite3.Error as error:
|
| 283 | return {"rows": None, "all_metadata_valid": False, "error": str(error)}
|
| 284 | finally:
|
| 285 | if "connection" in locals():
|
| 286 | connection.close()
|
| 287 | rows: list[dict[str, Any]] = []
|
| 288 | valid = True
|
| 289 | for path, mtime_ns, ctime_ns, size, blob_hash in raw_rows:
|
| 290 | file_path = root / path
|
| 291 | try:
|
| 292 | info = file_path.stat()
|
| 293 | except OSError:
|
| 294 | valid = False
|
| 295 | continue
|
| 296 | expected_ctime = info.st_ctime_ns if os.name == "posix" else 0
|
| 297 | metadata_valid = (
|
| 298 | int(size) == info.st_size
|
| 299 | and int(mtime_ns) == info.st_mtime_ns
|
| 300 | and int(ctime_ns) == expected_ctime
|
| 301 | )
|
| 302 | valid = valid and metadata_valid
|
| 303 | rows.append(
|
| 304 | {
|
| 305 | "path": str(path),
|
| 306 | "size": int(size),
|
| 307 | "blob_hash": str(blob_hash),
|
| 308 | "metadata_valid": metadata_valid,
|
| 309 | }
|
| 310 | )
|
| 311 | return {"rows": rows, "all_metadata_valid": valid}
|
| 312 |
|
| 313 |
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 314 | def base_env() -> dict[str, str]:
|
| 315 | return oakbench_base_env(
|
| 316 | author_name="Oak Task Loop",
|
| 317 | author_email="[email protected]",
|
| 318 | oak_author="oak-task-loop",
|
| 319 | )
|
| 320 |
|
| 321 |
|
| 322 | def parse_args() -> argparse.Namespace:
|
| 323 | parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
|
| 324 | parser.add_argument("--subjects", help="Comma-separated subject names from config/subjects.toml")
|
| 325 | parser.add_argument("--config", type=Path, default=ROOT / "config" / "subjects.toml")
|
| 326 | parser.add_argument("--git-bin", type=Path)
|
| 327 | parser.add_argument("--oak-installed-bin", type=Path)
|
| 328 | parser.add_argument("--oak-local-bin", type=Path)
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 329 | parser.add_argument(
|
| 330 | "--oak-serve-bin",
|
| 331 | type=Path,
|
| 332 | help="Immutable Oak binary used only as the fixed loopback server and fixture writer.",
|
| 333 | )
|
| 334 | parser.add_argument(
|
| 335 | "--oak-serve-source",
|
| 336 | type=Path,
|
| 337 | help="Read-only source checkout that produced --oak-serve-bin.",
|
| 338 | )
|
| 339 | parser.add_argument(
|
| 340 | "--oak-serve-source-head",
|
| 341 | help="Exact 64-hex Oak source head that produced --oak-serve-bin.",
|
| 342 | )
|
| 343 | parser.add_argument(
|
| 344 | "--materialization-baseline-subject",
|
| 345 | help="Named control subject required by the legacy materialization scenario.",
|
| 346 | )
|
| 347 | parser.add_argument(
|
| 348 | "--materialization-timing-controls-met",
|
| 349 | action="store_true",
|
| 350 | help=(
|
| 351 | "Record that harness-verifiable controls are met only with a held measurement lock, "
|
| 352 | "randomized subject order, and at least 10 runs. This is not a speedup claim: "
|
| 353 | "separate A/A noise and externally verified quiet-host evidence remain required."
|
| 354 | ),
|
| 355 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 356 | parser.add_argument("--runs", type=int, default=3)
|
| 357 | parser.add_argument(
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 358 | "--scenarios",
|
| 359 | default=TASK_LOOP_SCENARIO,
|
| 360 | help=(
|
| 361 | "Comma-separated scenarios: task_loop, agent_task_baseline_v1, "
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 362 | "agent_task_oak_local_v1, legacy_selected_clone_materialization_v1, "
|
| 363 | "and/or interrupted_publication_after_request_v1. "
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 364 | "The interruption "
|
| 365 | "scenario is a loopback-only Oak fixture and never contacts a remote service."
|
| 366 | ),
|
| 367 | )
|
| 368 | parser.add_argument(
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 369 | "--git-remote",
|
| 370 | choices=("github", "local"),
|
| 371 | default="github",
|
| 372 | help=(
|
| 373 | "github: agent-default track, real push to GIT_BENCH_REMOTE "
|
| 374 | "(transport-comparable to oak). local: core-equivalent track, "
|
| 375 | "local bare remote (no network in git's loop)."
|
| 376 | ),
|
| 377 | )
|
| 378 | parser.add_argument("--workdir", type=Path, default=DEFAULT_WORKDIR)
|
| 379 | parser.add_argument("--results", type=Path, default=ROOT / "results" / "task-loop")
|
| 380 | parser.add_argument("--oak-repo", type=Path, default=DEFAULT_OAK_REPO)
|
| 381 | parser.add_argument("--admitted-output-chars", type=int, default=20_000)
|
| 382 | parser.add_argument("--keep-workdirs", action="store_true")
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 383 | parser.add_argument(
|
| 384 | "--randomize-subject-order",
|
| 385 | action="store_true",
|
| 386 | help="Shuffle subject order independently in each run using --random-seed.",
|
| 387 | )
|
| 388 | parser.add_argument("--random-seed", type=int, default=20260908)
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 389 | return parser.parse_args()
|
| 390 |
|
| 391 |
|
| 392 | def run_untimed(command: list[str], cwd: Path | None, timeout: float = 300.0) -> subprocess.CompletedProcess[str]:
|
| 393 | return subprocess.run(
|
| 394 | command,
|
| 395 | cwd=cwd,
|
| 396 | env=base_env(),
|
| 397 | text=True,
|
| 398 | stdout=subprocess.PIPE,
|
| 399 | stderr=subprocess.PIPE,
|
| 400 | timeout=timeout,
|
| 401 | check=False,
|
| 402 | )
|
| 403 |
|
| 404 |
|
| 405 | def step_row(
|
| 406 | subject: Subject,
|
| 407 | scenario: str,
|
| 408 | operation: str,
|
| 409 | command: list[str],
|
| 410 | cwd: Path,
|
| 411 | run_index: int,
|
| 412 | metadata: dict[str, Any],
|
| 413 | admitted_output_chars: int,
|
| 414 | *,
|
| 415 | kind: str = "vcs",
|
| 416 | expected_returncodes: tuple[int, ...] = (0,),
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 417 | preserve_stdout: bool = False,
|
| 418 | preserve_stderr: bool = False,
|
| 419 | env: dict[str, str] | None = None,
|
| 420 | timeout_seconds: float | None = None,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 421 | ) -> dict[str, Any]:
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 422 | capture = run_timed(
|
| 423 | command,
|
| 424 | cwd,
|
| 425 | env or base_env(),
|
| 426 | admitted_output_chars,
|
| 427 | timeout_seconds=timeout_seconds,
|
| 428 | )
|
| 429 | expected = not capture.timed_out and capture.returncode in expected_returncodes
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 430 | is_vcs = 1 if kind == "vcs" else 0
|
| 431 | row = {
|
| 432 | **metadata,
|
| 433 | "subject": subject.name,
|
| 434 | "subject_kind": subject.kind,
|
| 435 | "subject_label": subject.label,
|
| 436 | "scenario": scenario,
|
| 437 | "run": run_index,
|
| 438 | "operation": operation,
|
| 439 | "step_kind": kind,
|
| 440 | "elapsed_ms": round(capture.elapsed_ms, 3),
|
| 441 | "returncode": 0 if expected else (capture.returncode or 1),
|
| 442 | "process_returncode": capture.returncode,
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 443 | "timed_out": capture.timed_out,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 444 | "expected_returncodes": list(expected_returncodes),
|
| 445 | "command": command,
|
| 446 | "tool_call_count": 1,
|
| 447 | "terminal_tool_call_count": 1,
|
| 448 | "vcs_tool_call_count": is_vcs,
|
| 449 | **oakbench_tokens.interaction_token_fields(
|
| 450 | command_display(command),
|
| 451 | capture.stdout_text,
|
| 452 | capture.stderr_text,
|
| 453 | capture.stdout_bytes,
|
| 454 | capture.stderr_bytes,
|
| 455 | capture.stdout_truncated,
|
| 456 | capture.stderr_truncated,
|
| 457 | ),
|
| 458 | }
|
| 459 | if not expected:
|
| 460 | row["stderr"] = capture.stderr_text[-4000:]
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 461 | if preserve_stdout:
|
| 462 | row["stdout"] = capture.stdout_text
|
| 463 | if preserve_stderr:
|
| 464 | row["stderr"] = capture.stderr_text
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 465 | return row
|
| 466 |
|
| 467 |
|
| 468 | def skip_row(
|
| 469 | subject: Subject,
|
| 470 | scenario: str,
|
| 471 | run_index: int,
|
| 472 | metadata: dict[str, Any],
|
| 473 | reason: str,
|
| 474 | ) -> dict[str, Any]:
|
| 475 | return {
|
| 476 | **metadata,
|
| 477 | "subject": subject.name,
|
| 478 | "subject_kind": subject.kind,
|
| 479 | "subject_label": subject.label,
|
| 480 | "scenario": scenario,
|
| 481 | "run": run_index,
|
| 482 | "operation": "task_loop.skipped",
|
| 483 | "step_kind": "summary",
|
| 484 | "elapsed_ms": 0.0,
|
| 485 | "returncode": SKIP_RETURNCODE,
|
| 486 | "command": [],
|
| 487 | "skipped": True,
|
| 488 | "skip_reason": reason,
|
| 489 | "tool_call_count": 0,
|
| 490 | "terminal_tool_call_count": 0,
|
| 491 | "vcs_tool_call_count": 0,
|
| 492 | }
|
| 493 |
|
| 494 |
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 495 | def unmeasured_step_row(
|
| 496 | subject: Subject,
|
| 497 | scenario: str,
|
| 498 | operation: str,
|
| 499 | run_index: int,
|
| 500 | metadata: dict[str, Any],
|
| 501 | reason: str,
|
| 502 | ) -> dict[str, Any]:
|
| 503 | row = skip_row(subject, scenario, run_index, metadata, reason)
|
| 504 | row["operation"] = operation
|
| 505 | return row
|
| 506 |
|
| 507 |
|
| 508 | def harness_step_row(
|
| 509 | subject: Subject,
|
| 510 | scenario: str,
|
| 511 | operation: str,
|
| 512 | run_index: int,
|
| 513 | metadata: dict[str, Any],
|
| 514 | elapsed_ms: float,
|
| 515 | *,
|
| 516 | returncode: int = 0,
|
| 517 | harness_command_count: int = 0,
|
| 518 | ) -> dict[str, Any]:
|
| 519 | return {
|
| 520 | **metadata,
|
| 521 | "subject": subject.name,
|
| 522 | "subject_kind": subject.kind,
|
| 523 | "subject_label": subject.label,
|
| 524 | "scenario": scenario,
|
| 525 | "run": run_index,
|
| 526 | "operation": operation,
|
| 527 | "step_kind": "harness",
|
| 528 | "elapsed_ms": round(elapsed_ms, 3),
|
| 529 | "returncode": returncode,
|
| 530 | "command": [],
|
| 531 | "tool_call_count": 0,
|
| 532 | "terminal_tool_call_count": 0,
|
| 533 | "vcs_tool_call_count": 0,
|
| 534 | "harness_command_count": harness_command_count,
|
| 535 | "stdout_bytes": 0,
|
| 536 | "stderr_bytes": 0,
|
| 537 | "retry_count": 0,
|
| 538 | "human_action_count": 0,
|
| 539 | }
|
| 540 |
|
| 541 |
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 542 | def phase_total_row(
|
| 543 | rows: list[dict[str, Any]],
|
| 544 | subject: Subject,
|
| 545 | scenario: str,
|
| 546 | run_index: int,
|
| 547 | metadata: dict[str, Any],
|
| 548 | operation: str,
|
| 549 | ) -> dict[str, Any]:
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 550 | failures = [row for row in rows if row_returncode(row) not in (0, SKIP_RETURNCODE)]
|
| 551 | unmeasured = [
|
| 552 | str(row["operation"])
|
| 553 | for row in rows
|
| 554 | if row_returncode(row) == SKIP_RETURNCODE
|
| 555 | ]
|
| 556 | phase_elapsed_ms: dict[str, float] = {}
|
| 557 | for row in rows:
|
| 558 | phase = str(row.get("operation", "unknown")).split(".", 1)[0]
|
| 559 | phase_elapsed_ms[phase] = phase_elapsed_ms.get(phase, 0.0) + float(
|
| 560 | row.get("elapsed_ms", 0.0)
|
| 561 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 562 | return {
|
| 563 | **metadata,
|
| 564 | "subject": subject.name,
|
| 565 | "subject_kind": subject.kind,
|
| 566 | "subject_label": subject.label,
|
| 567 | "scenario": scenario,
|
| 568 | "run": run_index,
|
| 569 | "operation": operation,
|
| 570 | "step_kind": "summary",
|
| 571 | "elapsed_ms": round(sum(float(row.get("elapsed_ms", 0)) for row in rows), 3),
|
| 572 | "returncode": 1 if failures else 0,
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 573 | "coverage_complete": not unmeasured,
|
| 574 | "unmeasured_operations": unmeasured,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 575 | "command": [row["command"] for row in rows],
|
| 576 | "tool_call_count": sum(oakbench_tokens.int_or_zero(row.get("tool_call_count")) for row in rows),
|
| 577 | "terminal_tool_call_count": sum(
|
| 578 | oakbench_tokens.int_or_zero(row.get("terminal_tool_call_count")) for row in rows
|
| 579 | ),
|
| 580 | "vcs_tool_call_count": sum(oakbench_tokens.int_or_zero(row.get("vcs_tool_call_count")) for row in rows),
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 581 | "agent_tool_call_count": sum(
|
| 582 | oakbench_tokens.int_or_zero(row.get("tool_call_count")) for row in rows
|
| 583 | ),
|
| 584 | "harness_command_count": sum(
|
| 585 | oakbench_tokens.int_or_zero(row.get("harness_command_count")) for row in rows
|
| 586 | ),
|
| 587 | "retry_count": sum(oakbench_tokens.int_or_zero(row.get("retry_count")) for row in rows),
|
| 588 | "human_action_count": sum(
|
| 589 | oakbench_tokens.int_or_zero(row.get("human_action_count")) for row in rows
|
| 590 | ),
|
| 591 | "phase_elapsed_ms": {
|
| 592 | phase: round(elapsed, 3) for phase, elapsed in sorted(phase_elapsed_ms.items())
|
| 593 | },
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 594 | **oakbench_tokens.summed_token_fields(
|
| 595 | rows, "sum_of_task_loop_steps_command_plus_admitted_output_chars_div_4"
|
| 596 | ),
|
| 597 | "summarized_operations": [row["operation"] for row in rows],
|
| 598 | }
|
| 599 |
|
| 600 |
|
| 601 | def has_lane_failures(rows: list[dict[str, Any]]) -> bool:
|
| 602 | return any(row_returncode(row) not in (0, SKIP_RETURNCODE) for row in rows)
|
| 603 |
|
| 604 |
|
| 605 | def enforce_oracle_result(total: dict[str, Any], oracle_ok: bool | None) -> dict[str, Any]:
|
| 606 | if oracle_ok is False:
|
| 607 | total["returncode"] = 1
|
| 608 | total["stderr"] = "oracle_payload_missing_from_remote_branch"
|
| 609 | return total
|
| 610 |
|
| 611 |
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 612 | def interrupted_publication_fields(
|
| 613 | *,
|
| 614 | request_count: int,
|
| 615 | request_bytes: int | None,
|
| 616 | declared_request_bytes: int | None,
|
| 617 | request_framing_valid: bool,
|
| 618 | request_body_timed_out: bool,
|
| 619 | local_head_before: str | None,
|
| 620 | local_head_after: str | None,
|
| 621 | status_after: dict[str, Any] | None,
|
| 622 | worktree_payload_before: str | None,
|
| 623 | worktree_payload_after: str | None,
|
| 624 | cli_observation: dict[str, str],
|
| 625 | ) -> dict[str, Any]:
|
| 626 | """Evidence fields for a single interrupted mutation attempt.
|
| 627 |
|
| 628 | The fixture's receipt of a full request is ground truth for the harness,
|
| 629 | not a server durability receipt available to the CLI. A disconnect after
|
| 630 | that point is therefore unknown to the client and must never imply replay.
|
| 631 | """
|
| 632 | request_complete = bool(
|
| 633 | request_count == 1
|
| 634 | and request_framing_valid
|
| 635 | and declared_request_bytes is not None
|
| 636 | and request_bytes == declared_request_bytes
|
| 637 | )
|
| 638 | if request_count == 0:
|
| 639 | delivery_state = "no_request_observed"
|
| 640 | fixture_outcome = "not_sent_observed"
|
| 641 | elif not request_framing_valid:
|
| 642 | delivery_state = "invalid_request_framing"
|
| 643 | fixture_outcome = "indeterminate_after_invalid_request_disconnect"
|
| 644 | elif request_complete:
|
| 645 | delivery_state = "complete_request_observed"
|
| 646 | fixture_outcome = "unknown_after_complete_request_disconnect"
|
| 647 | else:
|
| 648 | delivery_state = "partial_request_observed"
|
| 649 | fixture_outcome = "indeterminate_after_partial_request_disconnect"
|
| 650 | payload_preserved = bool(
|
| 651 | worktree_payload_before
|
| 652 | and worktree_payload_before == worktree_payload_after
|
| 653 | )
|
| 654 | unpublished_state_preserved = bool(status_after and status_after.get("needs_push"))
|
| 655 | return {
|
| 656 | "publication_outcome": cli_observation["reported_outcome"],
|
| 657 | "cli_observation": cli_observation["observation"],
|
| 658 | "fixture_delivery_state": delivery_state,
|
| 659 | "fixture_outcome_knowledge": fixture_outcome,
|
| 660 | "fixture_received_complete_request_before_disconnect": request_complete,
|
| 661 | "fixture_request_framing_valid": request_framing_valid,
|
| 662 | "fixture_request_body_timed_out": request_body_timed_out,
|
| 663 | "fixture_declared_request_bytes": declared_request_bytes,
|
| 664 | "mutation_requests_observed": request_count,
|
| 665 | "blind_replay_requests": max(0, request_count - 1),
|
| 666 | "network_request_bytes_observed": request_bytes,
|
| 667 | "local_head_preserved": bool(local_head_before) and local_head_before == local_head_after,
|
| 668 | "worktree_payload_preserved": payload_preserved,
|
| 669 | "unpublished_state_preserved": unpublished_state_preserved,
|
| 670 | "local_work_preserved": payload_preserved and unpublished_state_preserved,
|
| 671 | "working_tree_dirty_after": None if status_after is None else status_after.get("dirty"),
|
| 672 | }
|
| 673 |
|
| 674 |
|
| 675 | def cli_publication_observation(row: dict[str, Any], *, json_requested: bool) -> dict[str, str]:
|
| 676 | """Classify only what the CLI reported, never what the fault fixture knows."""
|
| 677 | if row.get("timed_out"):
|
| 678 | return {"reported_outcome": "unreported", "observation": "process_timeout"}
|
| 679 | stdout = str(row.get("stdout") or "")
|
| 680 | stderr = str(row.get("stderr") or "")
|
| 681 | value = _parse_json_object(stdout) if json_requested else None
|
| 682 | if value is not None and row.get("process_returncode") == 0 and value.get("published") is True:
|
| 683 | return {"reported_outcome": "published", "observation": "json_success_receipt"}
|
| 684 | if value is not None and isinstance(value.get("error"), dict):
|
| 685 | message = str(value["error"].get("message") or "").lower()
|
| 686 | if any(marker in message for marker in ("unconfirmed", "unknown", "may have")):
|
| 687 | return {"reported_outcome": "unknown", "observation": "json_error_unknown"}
|
| 688 | return {
|
| 689 | "reported_outcome": "unreported",
|
| 690 | "observation": "json_error_without_publication_outcome",
|
| 691 | }
|
| 692 | message = f"{stdout}\n{stderr}".lower()
|
| 693 | if any(marker in message for marker in ("unconfirmed", "unknown", "may have")):
|
| 694 | return {"reported_outcome": "unknown", "observation": "text_error_unknown"}
|
| 695 | return {"reported_outcome": "unreported", "observation": "text_error_without_publication_outcome"}
|
| 696 |
|
| 697 |
|
| 698 | def file_sha256(path: Path) -> str | None:
|
| 699 | try:
|
| 700 | return hashlib.sha256(path.read_bytes()).hexdigest()
|
| 701 | except OSError:
|
| 702 | return None
|
| 703 |
|
| 704 |
|
| 705 | class _DisconnectAfterPushHandler(BaseHTTPRequestHandler):
|
| 706 | """Minimal loopback fixture for a complete request with a lost reply."""
|
| 707 |
|
| 708 | server_version = "oakbench-interruption/1"
|
| 709 |
|
| 710 | def log_message(self, _format: str, *_args: object) -> None:
|
| 711 | return
|
| 712 |
|
| 713 | def _json(self, status: int, payload: dict[str, Any]) -> None:
|
| 714 | encoded = json.dumps(payload).encode()
|
| 715 | self.send_response(status)
|
| 716 | self.send_header("Content-Type", "application/json")
|
| 717 | self.send_header("Content-Length", str(len(encoded)))
|
| 718 | self.end_headers()
|
| 719 | self.wfile.write(encoded)
|
| 720 |
|
| 721 | def do_GET(self) -> None: # noqa: N802 - BaseHTTPRequestHandler API
|
| 722 | self._json(404, {"error": "absent synthetic repo"})
|
| 723 |
|
| 724 | def do_POST(self) -> None: # noqa: N802 - BaseHTTPRequestHandler API
|
| 725 | declared_size: int | None = None
|
| 726 | framing_valid = False
|
| 727 | body = b""
|
| 728 | body_timed_out = False
|
| 729 | raw_sizes = self.headers.get_all("Content-Length", failobj=[])
|
| 730 | try:
|
| 731 | raw_size = raw_sizes[0].strip(" \t") if len(raw_sizes) == 1 else ""
|
| 732 | parsed_size = int(raw_size) if raw_size.isascii() and raw_size.isdecimal() else -1
|
| 733 | framing_valid = (
|
| 734 | 0 <= parsed_size <= INTERRUPTED_PUBLICATION_MAX_REQUEST_BYTES
|
| 735 | and self.headers.get("Transfer-Encoding") is None
|
| 736 | )
|
| 737 | if framing_valid:
|
| 738 | declared_size = parsed_size
|
| 739 | self.connection.settimeout(INTERRUPTED_PUBLICATION_REQUEST_TIMEOUT_S)
|
| 740 | while len(body) < parsed_size:
|
| 741 | try:
|
| 742 | chunk = self.rfile.read1(parsed_size - len(body))
|
| 743 | except (socket.timeout, TimeoutError):
|
| 744 | body_timed_out = True
|
| 745 | break
|
| 746 | except OSError:
|
| 747 | break
|
| 748 | if not chunk:
|
| 749 | break
|
| 750 | body += chunk
|
| 751 | except ValueError:
|
| 752 | framing_valid = False
|
| 753 | server = self.server
|
| 754 | if self.path == "/api/repos":
|
| 755 | self._json(200, {"name": "repo"})
|
| 756 | return
|
| 757 | if self.path == "/api/oak/repo/push":
|
| 758 | with server.receipt_lock: # type: ignore[attr-defined]
|
| 759 | server.push_request_count += 1 # type: ignore[attr-defined]
|
| 760 | server.push_request_bytes += len(body) # type: ignore[attr-defined]
|
| 761 | server.push_declared_request_bytes = declared_size # type: ignore[attr-defined]
|
| 762 | server.push_request_framing_valid = framing_valid # type: ignore[attr-defined]
|
| 763 | server.push_request_body_timed_out = body_timed_out # type: ignore[attr-defined]
|
| 764 | server.request_complete_before_disconnect = ( # type: ignore[attr-defined]
|
| 765 | framing_valid and len(body) == declared_size
|
| 766 | )
|
| 767 | self.close_connection = True
|
| 768 | try:
|
| 769 | self.connection.shutdown(socket.SHUT_RDWR)
|
| 770 | except OSError:
|
| 771 | pass
|
| 772 | self.connection.close()
|
| 773 | return
|
| 774 | self._json(404, {"error": "unexpected synthetic path"})
|
| 775 |
|
| 776 |
|
| 777 | def start_interrupted_publication_server() -> tuple[ThreadingHTTPServer, threading.Thread]:
|
| 778 | server = ThreadingHTTPServer(("127.0.0.1", 0), _DisconnectAfterPushHandler)
|
| 779 | server.receipt_lock = threading.Lock() # type: ignore[attr-defined]
|
| 780 | server.push_request_count = 0 # type: ignore[attr-defined]
|
| 781 | server.push_request_bytes = 0 # type: ignore[attr-defined]
|
| 782 | server.push_declared_request_bytes = None # type: ignore[attr-defined]
|
| 783 | server.push_request_framing_valid = False # type: ignore[attr-defined]
|
| 784 | server.push_request_body_timed_out = False # type: ignore[attr-defined]
|
| 785 | server.request_complete_before_disconnect = False # type: ignore[attr-defined]
|
| 786 | thread = threading.Thread(target=server.serve_forever, name="oakbench-interrupted-push", daemon=True)
|
| 787 | thread.start()
|
| 788 | return server, thread
|
| 789 |
|
| 790 |
|
| 791 | def _parse_json_object(text: str | None) -> dict[str, Any] | None:
|
| 792 | try:
|
| 793 | value = json.loads(text or "")
|
| 794 | except (TypeError, json.JSONDecodeError):
|
| 795 | return None
|
| 796 | return value if isinstance(value, dict) else None
|
| 797 |
|
| 798 |
|
| 799 | def run_interrupted_publication(
|
| 800 | subject: Subject,
|
| 801 | run_root: Path,
|
| 802 | run_index: int,
|
| 803 | metadata: dict[str, Any],
|
| 804 | admitted: int,
|
| 805 | ) -> list[dict[str, Any]]:
|
| 806 | """Run one real Oak push against a loopback server that drops its reply.
|
| 807 |
|
| 808 | The fixture observes request delivery but deliberately has no durable
|
| 809 | remote state. That makes the client's outcome unknown and lets the lane
|
| 810 | prove work preservation and absence of blind replay without production
|
| 811 | mutations or a fake success claim.
|
| 812 | """
|
| 813 | scenario = INTERRUPTED_PUBLICATION_SCENARIO
|
| 814 | metadata = {
|
| 815 | **metadata,
|
| 816 | "benchmark_track": "failure-evidence",
|
| 817 | "remote_transport": "loopback_fault_fixture",
|
| 818 | "remote_server": "127.0.0.1",
|
| 819 | }
|
| 820 | if subject.kind != "oak":
|
| 821 | return [
|
| 822 | skip_row(
|
| 823 | subject,
|
| 824 | scenario,
|
| 825 | run_index,
|
| 826 | metadata,
|
| 827 | "interrupted publication fixture currently exercises the actual Oak HTTP client only",
|
| 828 | )
|
| 829 | ]
|
| 830 |
|
| 831 | workspace = run_root / f"interrupted-{subject.name}-r{run_index}"
|
| 832 | workspace.mkdir(parents=True)
|
| 833 | rows: list[dict[str, Any]] = []
|
| 834 | vcs = str(subject.bin)
|
| 835 | rows.append(
|
| 836 | step_row(subject, scenario, "setup.init", [vcs, "init", "."], workspace, run_index, metadata, admitted)
|
| 837 | )
|
| 838 | if row_returncode(rows[-1]) != 0:
|
| 839 | return rows
|
| 840 | rows.append(
|
| 841 | edit_step(
|
| 842 | subject,
|
| 843 | scenario,
|
| 844 | "setup.edit",
|
| 845 | workspace,
|
| 846 | f"interrupted publication run={run_index}",
|
| 847 | run_index,
|
| 848 | metadata,
|
| 849 | admitted,
|
| 850 | )
|
| 851 | )
|
| 852 | rows.append(
|
| 853 | step_row(subject, scenario, "setup.checkpoint", [vcs, "commit"], workspace, run_index, metadata, admitted)
|
| 854 | )
|
| 855 | if row_returncode(rows[-1]) != 0:
|
| 856 | return rows
|
| 857 |
|
| 858 | before = run_untimed([vcs, "hash"], workspace)
|
| 859 | head_before = before.stdout.strip() if before.returncode == 0 else None
|
| 860 | payload_path = workspace / "task-notes.md"
|
| 861 | payload_before = file_sha256(payload_path)
|
| 862 | before_usage = bounded_tree_usage(workspace)
|
| 863 | push_help = run_untimed([vcs, "push", "--help"], workspace)
|
| 864 | push_supports_json = "--json" in (push_help.stdout + push_help.stderr)
|
| 865 | server, thread = start_interrupted_publication_server()
|
| 866 | host, port = server.server_address[:2]
|
| 867 | remote = f"http://{host}:{port}"
|
| 868 | fixture_env = base_env()
|
| 869 | fixture_env.pop("OAK_API_KEY", None)
|
| 870 | fixture_env.pop("OAK_REMOTE", None)
|
| 871 | fixture_env["OAK_NO_UPDATE_CHECK"] = "1"
|
| 872 | try:
|
| 873 | push_command = [vcs, "push", "--repo", "oak/repo", "--remote", remote]
|
| 874 | if push_supports_json:
|
| 875 | push_command.append("--json")
|
| 876 | publication_attempt = step_row(
|
| 877 | subject,
|
| 878 | scenario,
|
| 879 | "publication.attempt",
|
| 880 | push_command,
|
| 881 | workspace,
|
| 882 | run_index,
|
| 883 | metadata,
|
| 884 | admitted,
|
| 885 | expected_returncodes=tuple(range(1, 128)),
|
| 886 | preserve_stdout=True,
|
| 887 | preserve_stderr=True,
|
| 888 | env=fixture_env,
|
| 889 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 890 | )
|
| 891 | rows.append(publication_attempt)
|
| 892 | finally:
|
| 893 | server.shutdown()
|
| 894 | server.server_close()
|
| 895 | thread.join(timeout=2.0)
|
| 896 |
|
| 897 | state_row = step_row(
|
| 898 | subject,
|
| 899 | scenario,
|
| 900 | "reconcile.local_state",
|
| 901 | [vcs, "agent", "state", "--json", "--compact"],
|
| 902 | workspace,
|
| 903 | run_index,
|
| 904 | metadata,
|
| 905 | admitted,
|
| 906 | preserve_stdout=True,
|
| 907 | env=fixture_env,
|
| 908 | )
|
| 909 | rows.append(state_row)
|
| 910 | after = run_untimed([vcs, "hash"], workspace)
|
| 911 | head_after = after.stdout.strip() if after.returncode == 0 else None
|
| 912 | payload_after = file_sha256(payload_path)
|
| 913 | status_after = _parse_json_object(state_row.get("stdout"))
|
| 914 | with server.receipt_lock: # type: ignore[attr-defined]
|
| 915 | request_count = int(server.push_request_count) # type: ignore[attr-defined]
|
| 916 | request_bytes = int(server.push_request_bytes) # type: ignore[attr-defined]
|
| 917 | declared_request_bytes = server.push_declared_request_bytes # type: ignore[attr-defined]
|
| 918 | request_framing_valid = bool(server.push_request_framing_valid) # type: ignore[attr-defined]
|
| 919 | request_body_timed_out = bool(server.push_request_body_timed_out) # type: ignore[attr-defined]
|
| 920 | cli_observation = cli_publication_observation(
|
| 921 | publication_attempt,
|
| 922 | json_requested=push_supports_json,
|
| 923 | )
|
| 924 | evidence = interrupted_publication_fields(
|
| 925 | request_count=request_count,
|
| 926 | request_bytes=request_bytes if request_count else None,
|
| 927 | declared_request_bytes=declared_request_bytes,
|
| 928 | request_framing_valid=request_framing_valid,
|
| 929 | request_body_timed_out=request_body_timed_out,
|
| 930 | local_head_before=head_before,
|
| 931 | local_head_after=head_after,
|
| 932 | status_after=status_after,
|
| 933 | worktree_payload_before=payload_before,
|
| 934 | worktree_payload_after=payload_after,
|
| 935 | cli_observation=cli_observation,
|
| 936 | )
|
| 937 | evidence["workspace_usage_before"] = before_usage
|
| 938 | evidence["workspace_usage_after"] = bounded_tree_usage(workspace)
|
| 939 | evidence["remote_transport"] = "loopback_fault_fixture"
|
| 940 | evidence["remote_server"] = str(host)
|
| 941 | evidence["human_action_count"] = 0
|
| 942 | evidence["compatibility_probe_calls"] = 1
|
| 943 | evidence["harness_probe_calls"] = 3
|
| 944 | evidence["push_json_supported"] = push_supports_json
|
| 945 | total = phase_total_row(rows, subject, scenario, run_index, metadata, "interrupted_publication.total")
|
| 946 | evidence["harness_command_count"] = total["harness_command_count"] + 3
|
| 947 | total.update(evidence)
|
| 948 | if not (
|
| 949 | evidence["fixture_received_complete_request_before_disconnect"]
|
| 950 | and request_count == 1
|
| 951 | and evidence["local_head_preserved"]
|
| 952 | and evidence["local_work_preserved"]
|
| 953 | and not publication_attempt["timed_out"]
|
| 954 | and row_returncode(rows[-1]) == 0
|
| 955 | ):
|
| 956 | total["returncode"] = 1
|
| 957 | total["stderr"] = "interrupted_publication_evidence_incomplete"
|
| 958 | rows.append(total)
|
| 959 | return rows
|
| 960 |
|
| 961 |
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 962 | def edit_step(
|
| 963 | subject: Subject,
|
| 964 | scenario: str,
|
| 965 | operation: str,
|
| 966 | workspace: Path,
|
| 967 | payload: str,
|
| 968 | run_index: int,
|
| 969 | metadata: dict[str, Any],
|
| 970 | admitted: int,
|
| 971 | ) -> dict[str, Any]:
|
| 972 | command = ["/bin/zsh", "-c", f"echo {payload!r} >> task-notes.md"]
|
| 973 | return step_row(
|
| 974 | subject, scenario, operation, command, workspace, run_index, metadata, admitted, kind="edit"
|
| 975 | )
|
| 976 |
|
| 977 |
|
| 978 | def oak_virtual_branch(subject: Subject, mount_dir: Path) -> str | None:
|
| 979 | proc = run_untimed([str(subject.bin), "status"], mount_dir)
|
| 980 | match = VIRTUAL_BRANCH_RE.search(proc.stdout + proc.stderr)
|
| 981 | return match.group(1) if match else None
|
| 982 |
|
| 983 |
|
| 984 | def run_oak_loop(
|
| 985 | subject: Subject,
|
| 986 | remote: oakbench_remotes.RemoteResolution,
|
| 987 | run_root: Path,
|
| 988 | run_index: int,
|
| 989 | metadata: dict[str, Any],
|
| 990 | admitted: int,
|
| 991 | ) -> list[dict[str, Any]]:
|
| 992 | scenario = "task_loop"
|
| 993 | meta = {**metadata, **remote.row_fields()}
|
| 994 | rows: list[dict[str, Any]] = []
|
| 995 | branches: dict[str, str | None] = {}
|
| 996 | payloads: dict[str, str] = {}
|
| 997 | mounts: list[Path] = []
|
| 998 | vcs = str(subject.bin)
|
| 999 | try:
|
| 1000 | for phase in PHASES:
|
| 1001 | workspace = run_root / f"tl-{subject.name}-r{run_index}-{phase}"
|
| 1002 | payload = f"payload {phase} run={run_index} bench={metadata['bench_id']}"
|
| 1003 | payloads[phase] = payload
|
| 1004 | acquire = step_row(
|
| 1005 | subject, scenario, f"{phase}.acquire", [vcs, "mount", remote.repo or "", str(workspace)],
|
| 1006 | run_root, run_index, meta, admitted,
|
| 1007 | )
|
| 1008 | rows.append(acquire)
|
| 1009 | if int(acquire["returncode"]) != 0:
|
| 1010 | return rows
|
| 1011 | mounts.append(workspace)
|
| 1012 | branches[phase] = oak_virtual_branch(subject, workspace)
|
| 1013 | rows.append(edit_step(subject, scenario, f"{phase}.edit", workspace, payload, run_index, meta, admitted))
|
| 1014 | rows.append(step_row(subject, scenario, f"{phase}.snapshot", [vcs, "commit", "--no-verify"], workspace, run_index, meta, admitted))
|
| 1015 | rows.append(step_row(subject, scenario, f"{phase}.publish", [vcs, "push"], workspace, run_index, meta, admitted))
|
| 1016 | if phase == "task":
|
| 1017 | rows.append(
|
| 1018 | step_row(
|
| 1019 | subject, scenario, "task.describe",
|
| 1020 | [vcs, "desc", f"bench task loop {metadata['bench_id']} r{run_index}"],
|
| 1021 | workspace, run_index, meta, admitted,
|
| 1022 | )
|
| 1023 | )
|
| 1024 | end = step_row(
|
| 1025 | subject, scenario, f"{phase}.end", [vcs, "mount", "end", str(workspace)],
|
| 1026 | run_root, run_index, meta, admitted,
|
| 1027 | )
|
| 1028 | rows.append(end)
|
| 1029 | if int(end["returncode"]) == 0 and workspace in mounts:
|
| 1030 | mounts.remove(workspace)
|
| 1031 |
|
| 1032 | # Untimed oracle: each phase's payload must be on its pushed branch.
|
| 1033 | verifier = run_root / f"tl-verify-{subject.name}-r{run_index}"
|
| 1034 | clone = run_untimed([vcs, "clone", remote.repo or "", str(verifier)], run_root)
|
| 1035 | oracle_ok: bool | None = None
|
| 1036 | if clone.returncode == 0:
|
| 1037 | oracle_ok = True
|
| 1038 | for phase in PHASES:
|
| 1039 | branch = branches.get(phase)
|
| 1040 | if not branch:
|
| 1041 | oracle_ok = None
|
| 1042 | break
|
| 1043 | switch = run_untimed([vcs, "switch", branch], verifier)
|
| 1044 | content = (verifier / "task-notes.md")
|
| 1045 | seen = content.read_text() if content.exists() else ""
|
| 1046 | if switch.returncode != 0 or payloads[phase] not in seen:
|
| 1047 | oracle_ok = False
|
| 1048 | total = {
|
| 1049 | **phase_total_row(rows, subject, scenario, run_index, meta, "task_loop.total"),
|
| 1050 | "oracle_payload_on_remote_branch": oracle_ok,
|
| 1051 | "workspace_branches": branches,
|
| 1052 | "measurement_source": "verifier_clone_switch_grep; null oracle means branch name unparsed",
|
| 1053 | }
|
| 1054 | rows.append(enforce_oracle_result(total, oracle_ok))
|
| 1055 | return rows
|
| 1056 | finally:
|
| 1057 | for mount_dir in mounts:
|
| 1058 | run_untimed([vcs, "mount", "end", str(mount_dir), "-f"], run_root)
|
| 1059 |
|
| 1060 |
|
| 1061 | def ensure_git_remote_seeded(subject: Subject, remote: oakbench_remotes.RemoteResolution, scratch: Path) -> str | None:
|
| 1062 | """Untimed: make sure the git remote has a main with a seed commit."""
|
| 1063 | vcs = str(subject.bin)
|
| 1064 | probe = run_untimed([vcs, "ls-remote", str(remote.repo), "refs/heads/main"], None)
|
| 1065 | if probe.returncode == 0 and probe.stdout.strip():
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 1066 | if remote.transport == oakbench_remotes.TRANSPORT_LOCAL_FILE:
|
| 1067 | head = run_untimed([vcs, "symbolic-ref", "HEAD", "refs/heads/main"], Path(str(remote.repo)))
|
| 1068 | if head.returncode != 0:
|
| 1069 | return f"git remote HEAD setup failed: {head.stderr.strip()[:200]}"
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 1070 | return None
|
| 1071 | seed = scratch / "git-seed"
|
| 1072 | if seed.exists():
|
| 1073 | shutil.rmtree(seed)
|
| 1074 | seed.mkdir(parents=True)
|
| 1075 | (seed / "README.md").write_text("Task-loop benchmark repo (disposable).\n")
|
| 1076 | (seed / "task-notes.md").write_text("# task notes\n")
|
| 1077 | for command in (
|
| 1078 | [vcs, "init", "-q", "-b", "main"],
|
| 1079 | [vcs, "add", "."],
|
| 1080 | [vcs, "commit", "-q", "-m", "seed"],
|
| 1081 | [vcs, "push", "-q", str(remote.repo), "main:main"],
|
| 1082 | ):
|
| 1083 | proc = run_untimed(command, seed)
|
| 1084 | if proc.returncode != 0:
|
| 1085 | return f"git remote seed `{' '.join(command[1:3])}` failed: {proc.stderr.strip()[:200]}"
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 1086 | if remote.transport == oakbench_remotes.TRANSPORT_LOCAL_FILE:
|
| 1087 | head = run_untimed([vcs, "symbolic-ref", "HEAD", "refs/heads/main"], Path(str(remote.repo)))
|
| 1088 | if head.returncode != 0:
|
| 1089 | return f"git remote HEAD setup failed: {head.stderr.strip()[:200]}"
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 1090 | return None
|
| 1091 |
|
| 1092 |
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 1093 | def run_git_agent_task_baseline(
|
| 1094 | subject: Subject,
|
| 1095 | remote: oakbench_remotes.RemoteResolution,
|
| 1096 | run_root: Path,
|
| 1097 | run_index: int,
|
| 1098 | metadata: dict[str, Any],
|
| 1099 | admitted: int,
|
| 1100 | remote_create_ms: float = 0.0,
|
| 1101 | ) -> list[dict[str, Any]]:
|
| 1102 | """Measure a cold clone and a warm isolated worktree on one local remote.
|
| 1103 |
|
| 1104 | This is the core-equivalent, local-file cell. Hosted CI is deliberately a
|
| 1105 | skip row: a local commit oracle is not relabelled as provider CI.
|
| 1106 | """
|
| 1107 | scenario = AGENT_TASK_SCENARIO
|
| 1108 | meta = {
|
| 1109 | **metadata,
|
| 1110 | **remote.row_fields(),
|
| 1111 | "benchmark_track": (
|
| 1112 | "core-equivalent"
|
| 1113 | if remote.transport == oakbench_remotes.TRANSPORT_LOCAL_FILE
|
| 1114 | else "agent-default"
|
| 1115 | ),
|
| 1116 | }
|
| 1117 | vcs = str(subject.bin)
|
| 1118 | rows: list[dict[str, Any]] = [
|
| 1119 | harness_step_row(
|
| 1120 | subject,
|
| 1121 | scenario,
|
| 1122 | "setup.remote_container",
|
| 1123 | run_index,
|
| 1124 | meta,
|
| 1125 | remote_create_ms,
|
| 1126 | harness_command_count=1,
|
| 1127 | )
|
| 1128 | ]
|
| 1129 | setup_start = time.perf_counter()
|
| 1130 | seed_error = ensure_git_remote_seeded(subject, remote, run_root)
|
| 1131 | rows.append(
|
| 1132 | harness_step_row(
|
| 1133 | subject,
|
| 1134 | scenario,
|
| 1135 | "setup.remote_seed",
|
| 1136 | run_index,
|
| 1137 | meta,
|
| 1138 | (time.perf_counter() - setup_start) * 1000,
|
| 1139 | returncode=1 if seed_error else 0,
|
| 1140 | harness_command_count=6,
|
| 1141 | )
|
| 1142 | )
|
| 1143 | if seed_error:
|
| 1144 | rows[-1]["stderr"] = seed_error
|
| 1145 | return rows
|
| 1146 |
|
| 1147 | base = run_root / f"agent-task-{subject.name}-r{run_index}-cold"
|
| 1148 | cold_branch = oakbench_remotes.disposable_branch(
|
| 1149 | str(metadata["bench_id"]), "agent-task-cold", subject.name, run_index
|
| 1150 | )
|
| 1151 | warm_branch = oakbench_remotes.disposable_branch(
|
| 1152 | str(metadata["bench_id"]), "agent-task-warm", subject.name, run_index
|
| 1153 | )
|
| 1154 | cold_payload = f"cold task run={run_index} bench={metadata['bench_id']}"
|
| 1155 | warm_payload = f"warm task run={run_index} bench={metadata['bench_id']}"
|
| 1156 |
|
| 1157 | rows.append(
|
| 1158 | step_row(
|
| 1159 | subject,
|
| 1160 | scenario,
|
| 1161 | "cold.acquire",
|
| 1162 | [vcs, "clone", "-q", str(remote.repo), str(base)],
|
| 1163 | run_root,
|
| 1164 | run_index,
|
| 1165 | meta,
|
| 1166 | admitted,
|
| 1167 | )
|
| 1168 | )
|
| 1169 | if row_returncode(rows[-1]) != 0:
|
| 1170 | return rows
|
| 1171 | rows.append(
|
| 1172 | step_row(
|
| 1173 | subject, scenario, "cold.branch", [vcs, "switch", "-q", "-c", cold_branch],
|
| 1174 | base, run_index, meta, admitted,
|
| 1175 | )
|
| 1176 | )
|
| 1177 | rows.append(
|
| 1178 | step_row(
|
| 1179 | subject, scenario, "cold.inspect", [vcs, "status", "--porcelain=v1"],
|
| 1180 | base, run_index, meta, admitted,
|
| 1181 | )
|
| 1182 | )
|
| 1183 | rows.append(edit_step(subject, scenario, "cold.edit", base, cold_payload, run_index, meta, admitted))
|
| 1184 | rows.append(step_row(subject, scenario, "cold.checkpoint.add", [vcs, "add", "."], base, run_index, meta, admitted))
|
| 1185 | rows.append(
|
| 1186 | step_row(
|
| 1187 | subject, scenario, "cold.checkpoint.commit",
|
| 1188 | [vcs, "commit", "-q", "-m", f"agent task cold {metadata['bench_id']} r{run_index}"],
|
| 1189 | base, run_index, meta, admitted,
|
| 1190 | )
|
| 1191 | )
|
| 1192 | rows.append(
|
| 1193 | step_row(
|
| 1194 | subject, scenario, "cold.publish", [vcs, "push", "-q", "-u", "origin", cold_branch],
|
| 1195 | base, run_index, meta, admitted,
|
| 1196 | )
|
| 1197 | )
|
| 1198 | rows.append(
|
| 1199 | unmeasured_step_row(
|
| 1200 | subject, scenario, "cold.ci.observe", run_index, meta,
|
| 1201 | "local-file transport has no hosted CI provider; exact CI observation is unmeasured",
|
| 1202 | )
|
| 1203 | )
|
| 1204 | rows.append(
|
| 1205 | step_row(
|
| 1206 | subject, scenario, "cold.review_handoff",
|
| 1207 | [vcs, "show", "--format=%H%n%s", "--stat", "--oneline", "HEAD"],
|
| 1208 | base, run_index, meta, admitted,
|
| 1209 | )
|
| 1210 | )
|
| 1211 |
|
| 1212 | warm = run_root / f"agent-task-{subject.name}-r{run_index}-warm"
|
| 1213 | rows.append(
|
| 1214 | step_row(
|
| 1215 | subject, scenario, "warm.acquire",
|
| 1216 | [vcs, "worktree", "add", "-q", str(warm), "-b", warm_branch, "origin/main"],
|
| 1217 | base, run_index, meta, admitted,
|
| 1218 | )
|
| 1219 | )
|
| 1220 | if row_returncode(rows[-1]) == 0:
|
| 1221 | rows.append(
|
| 1222 | step_row(
|
| 1223 | subject, scenario, "warm.inspect", [vcs, "status", "--porcelain=v1"],
|
| 1224 | warm, run_index, meta, admitted,
|
| 1225 | )
|
| 1226 | )
|
| 1227 | rows.append(edit_step(subject, scenario, "warm.edit", warm, warm_payload, run_index, meta, admitted))
|
| 1228 | rows.append(step_row(subject, scenario, "warm.checkpoint.add", [vcs, "add", "."], warm, run_index, meta, admitted))
|
| 1229 | rows.append(
|
| 1230 | step_row(
|
| 1231 | subject, scenario, "warm.checkpoint.commit",
|
| 1232 | [vcs, "commit", "-q", "-m", f"agent task warm {metadata['bench_id']} r{run_index}"],
|
| 1233 | warm, run_index, meta, admitted,
|
| 1234 | )
|
| 1235 | )
|
| 1236 | rows.append(
|
| 1237 | step_row(
|
| 1238 | subject, scenario, "warm.publish", [vcs, "push", "-q", "-u", "origin", warm_branch],
|
| 1239 | warm, run_index, meta, admitted,
|
| 1240 | )
|
| 1241 | )
|
| 1242 | rows.append(
|
| 1243 | unmeasured_step_row(
|
| 1244 | subject, scenario, "warm.ci.observe", run_index, meta,
|
| 1245 | "local-file transport has no hosted CI provider; exact CI observation is unmeasured",
|
| 1246 | )
|
| 1247 | )
|
| 1248 | rows.append(
|
| 1249 | step_row(
|
| 1250 | subject, scenario, "warm.review_handoff",
|
| 1251 | [vcs, "show", "--format=%H%n%s", "--stat", "--oneline", "HEAD"],
|
| 1252 | warm, run_index, meta, admitted,
|
| 1253 | )
|
| 1254 | )
|
| 1255 |
|
| 1256 | usage_before_cleanup = {
|
| 1257 | "cold": bounded_tree_usage(base),
|
| 1258 | "warm": bounded_tree_usage(warm),
|
| 1259 | }
|
| 1260 | oracle_ok = True
|
| 1261 | for branch, payload in ((cold_branch, cold_payload), (warm_branch, warm_payload)):
|
| 1262 | show = run_untimed([vcs, "--git-dir", str(remote.repo), "show", f"{branch}:task-notes.md"], run_root)
|
| 1263 | if show.returncode != 0 or payload not in show.stdout:
|
| 1264 | oracle_ok = False
|
| 1265 |
|
| 1266 | if warm.exists():
|
| 1267 | rows.append(
|
| 1268 | step_row(
|
| 1269 | subject, scenario, "cleanup.warm",
|
| 1270 | [vcs, "worktree", "remove", str(warm)], base, run_index, meta, admitted,
|
| 1271 | )
|
| 1272 | )
|
| 1273 | cleanup_start = time.perf_counter()
|
| 1274 | shutil.rmtree(base, ignore_errors=True)
|
| 1275 | rows.append(
|
| 1276 | harness_step_row(
|
| 1277 | subject, scenario, "cleanup.cold", run_index, meta,
|
| 1278 | (time.perf_counter() - cleanup_start) * 1000,
|
| 1279 | )
|
| 1280 | )
|
| 1281 |
|
| 1282 | total = phase_total_row(rows, subject, scenario, run_index, meta, "agent_task.total")
|
| 1283 | total.update(
|
| 1284 | {
|
| 1285 | "oracle_payload_on_remote_branch": oracle_ok,
|
| 1286 | "acquisition_models": {
|
| 1287 | "cold": "fresh_full_clone",
|
| 1288 | "warm": "shared_object_worktree",
|
| 1289 | },
|
| 1290 | "repository_cache_state": {
|
| 1291 | "cold": "fresh_destination_no_shared_object_store",
|
| 1292 | "warm": "explicit_shared_object_store",
|
| 1293 | },
|
| 1294 | "os_cache_state": "unknown",
|
| 1295 | "os_cache_state_reason": "page_cache_not_purged_or_instrumented",
|
| 1296 | "workspace_usage_before_cleanup": usage_before_cleanup,
|
| 1297 | "network_bytes": None,
|
| 1298 | "network_bytes_reason": "local_file_transport_not_observed_by_byte_proxy",
|
| 1299 | "partial_clone_comparator": None,
|
| 1300 | "partial_clone_reason": "local file transport does not establish a network partial-clone comparison",
|
| 1301 | "human_action_count": 0,
|
| 1302 | "parallel_evidence_lane": "parallel_contention.py/workspace_per_task",
|
| 1303 | "oracle_probe_calls": 2,
|
| 1304 | "harness_command_count": total["harness_command_count"] + 2,
|
| 1305 | }
|
| 1306 | )
|
| 1307 | if not oracle_ok:
|
| 1308 | total["returncode"] = 1
|
| 1309 | total["stderr"] = "agent_task_remote_payload_oracle_failed"
|
| 1310 | rows.append(total)
|
| 1311 | return rows
|
| 1312 |
|
| 1313 |
|
| 1314 | def reserve_loopback_port() -> int:
|
| 1315 | with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as listener:
|
| 1316 | listener.bind(("127.0.0.1", 0))
|
| 1317 | return int(listener.getsockname()[1])
|
| 1318 |
|
| 1319 |
|
| 1320 | def wait_for_loopback_port(
|
| 1321 | process: subprocess.Popen[Any], port: int, timeout_s: float = LOCAL_OAK_SERVER_READY_TIMEOUT_S
|
| 1322 | ) -> bool:
|
| 1323 | deadline = time.monotonic() + timeout_s
|
| 1324 | while time.monotonic() < deadline:
|
| 1325 | if process.poll() is not None:
|
| 1326 | return False
|
| 1327 | try:
|
| 1328 | with socket.create_connection(("127.0.0.1", port), timeout=0.2):
|
| 1329 | return True
|
| 1330 | except OSError:
|
| 1331 | time.sleep(0.05)
|
| 1332 | return False
|
| 1333 |
|
| 1334 |
|
| 1335 | def harness_command_row(
|
| 1336 | subject: Subject,
|
| 1337 | scenario: str,
|
| 1338 | operation: str,
|
| 1339 | command: list[str],
|
| 1340 | cwd: Path,
|
| 1341 | run_index: int,
|
| 1342 | metadata: dict[str, Any],
|
| 1343 | ) -> tuple[dict[str, Any], subprocess.CompletedProcess[str]]:
|
| 1344 | started = time.perf_counter()
|
| 1345 | result = run_bounded_harness_command(command, cwd)
|
| 1346 | row = harness_step_row(
|
| 1347 | subject,
|
| 1348 | scenario,
|
| 1349 | operation,
|
| 1350 | run_index,
|
| 1351 | metadata,
|
| 1352 | (time.perf_counter() - started) * 1000,
|
| 1353 | returncode=0 if result.returncode == 0 else result.returncode,
|
| 1354 | harness_command_count=1,
|
| 1355 | )
|
| 1356 | if result.returncode != 0:
|
| 1357 | row["stderr"] = str(result.stderr)[-4000:]
|
| 1358 | return row, result
|
| 1359 |
|
| 1360 |
|
| 1361 | def run_bounded_harness_command(
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 1362 | command: list[str], cwd: Path, *, env: dict[str, str] | None = None
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 1363 | ) -> subprocess.CompletedProcess[str]:
|
| 1364 | capture = run_timed(
|
| 1365 | command,
|
| 1366 | cwd,
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 1367 | env or base_env(),
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 1368 | admitted_output_chars=4_000,
|
| 1369 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 1370 | )
|
| 1371 | return subprocess.CompletedProcess(
|
| 1372 | command,
|
| 1373 | capture.returncode,
|
| 1374 | capture.stdout_text,
|
| 1375 | capture.stderr_text,
|
| 1376 | )
|
| 1377 |
|
| 1378 |
|
| 1379 | def remote_branch_head(remote: str, repo_spec: str, branch: str) -> str | None:
|
| 1380 | owner, repo = repo_spec.split("/", 1)
|
| 1381 | branch_segment = urllib.parse.quote(branch, safe="")
|
| 1382 | url = f"{remote}/api/{owner}/{repo}/branches/{branch_segment}"
|
| 1383 | try:
|
| 1384 | with urllib.request.urlopen(url, timeout=2.0) as response:
|
| 1385 | value = json.loads(response.read())
|
| 1386 | except (OSError, ValueError, urllib.error.URLError):
|
| 1387 | return None
|
| 1388 | head = value.get("head") if isinstance(value, dict) else None
|
| 1389 | return head if isinstance(head, str) else None
|
| 1390 |
|
| 1391 |
|
| 1392 | def run_oak_local_agent_task_baseline(
|
| 1393 | subject: Subject,
|
| 1394 | run_root: Path,
|
| 1395 | run_index: int,
|
| 1396 | metadata: dict[str, Any],
|
| 1397 | admitted: int,
|
| 1398 | ) -> list[dict[str, Any]]:
|
| 1399 | """Exercise two fresh Oak checkouts against an owned loopback Oak server."""
|
| 1400 | run_root = run_root.resolve()
|
| 1401 | scenario = AGENT_TASK_OAK_LOCAL_SCENARIO
|
| 1402 | meta = {
|
| 1403 | **metadata,
|
| 1404 | "benchmark_track": "oak-local-loopback",
|
| 1405 | "remote_transport": "loopback_http",
|
| 1406 | "remote_server": "127.0.0.1",
|
| 1407 | }
|
| 1408 | rows: list[dict[str, Any]] = []
|
| 1409 | vcs = str(subject.bin)
|
| 1410 | server_root = run_root / f"oak-local-server-{subject.name}-r{run_index}"
|
| 1411 | seed = run_root / f"oak-local-seed-{subject.name}-r{run_index}"
|
| 1412 | cold = run_root / f"oak-local-cold-{subject.name}-r{run_index}"
|
| 1413 | followup = run_root / f"oak-local-followup-{subject.name}-r{run_index}"
|
| 1414 | verifier = run_root / f"oak-local-verifier-{subject.name}-r{run_index}"
|
| 1415 | repo_spec = f"qa/task-loop-{subject.name}-r{run_index}"
|
| 1416 | port = reserve_loopback_port()
|
| 1417 | remote = f"http://127.0.0.1:{port}"
|
| 1418 | server_root.mkdir(parents=True)
|
| 1419 | server_started = time.perf_counter()
|
| 1420 | server: subprocess.Popen[Any] | None = None
|
| 1421 | server_start_error: str | None = None
|
| 1422 | try:
|
| 1423 | server = subprocess.Popen(
|
| 1424 | [vcs, "serve", "--dir", str(server_root), "--host", "127.0.0.1", "--port", str(port)],
|
| 1425 | cwd=run_root,
|
| 1426 | env=base_env(),
|
| 1427 | stdout=subprocess.DEVNULL,
|
| 1428 | stderr=subprocess.DEVNULL,
|
| 1429 | start_new_session=os.name == "posix",
|
| 1430 | )
|
| 1431 | except OSError as error:
|
| 1432 | server_start_error = type(error).__name__
|
| 1433 | ready = server is not None and wait_for_loopback_port(server, port)
|
| 1434 | rows.append(
|
| 1435 | harness_step_row(
|
| 1436 | subject,
|
| 1437 | scenario,
|
| 1438 | "setup.server",
|
| 1439 | run_index,
|
| 1440 | meta,
|
| 1441 | (time.perf_counter() - server_started) * 1000,
|
| 1442 | returncode=0 if ready else 1,
|
| 1443 | harness_command_count=1,
|
| 1444 | )
|
| 1445 | )
|
| 1446 | branch: str | None = None
|
| 1447 | expected_head: str | None = None
|
| 1448 | expected_payload = "# task notes\n"
|
| 1449 | failure: str | None = None if ready else (
|
| 1450 | f"local_oak_server_start_failed:{server_start_error}"
|
| 1451 | if server_start_error else "local_oak_server_not_ready"
|
| 1452 | )
|
| 1453 | try:
|
| 1454 | if failure is None:
|
| 1455 | seed.mkdir(parents=True)
|
| 1456 | (seed / "task-notes.md").write_text(expected_payload)
|
| 1457 | for operation, command in (
|
| 1458 | ("setup.seed_init", [vcs, "init", "."]),
|
| 1459 | ("setup.seed_commit", [vcs, "commit"]),
|
| 1460 | ("setup.seed_push", [vcs, "push", "--repo", repo_spec, "--remote", remote]),
|
| 1461 | ):
|
| 1462 | row, result = harness_command_row(
|
| 1463 | subject, scenario, operation, command, seed, run_index, meta
|
| 1464 | )
|
| 1465 | rows.append(row)
|
| 1466 | if result.returncode != 0:
|
| 1467 | failure = f"{operation}_failed"
|
| 1468 | break
|
| 1469 |
|
| 1470 | def agent_step(operation: str, command: list[str], cwd: Path, *, preserve: bool = False) -> dict[str, Any]:
|
| 1471 | row = step_row(
|
| 1472 | subject,
|
| 1473 | scenario,
|
| 1474 | operation,
|
| 1475 | command,
|
| 1476 | cwd,
|
| 1477 | run_index,
|
| 1478 | meta,
|
| 1479 | admitted,
|
| 1480 | preserve_stdout=preserve,
|
| 1481 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 1482 | )
|
| 1483 | rows.append(row)
|
| 1484 | return row
|
| 1485 |
|
| 1486 | if failure is None:
|
| 1487 | row = agent_step(
|
| 1488 | "cold.acquire",
|
| 1489 | [vcs, "clone", repo_spec, str(cold), "--remote", remote, "--allow-legacy-scope"],
|
| 1490 | run_root,
|
| 1491 | )
|
| 1492 | if row_returncode(row) != 0:
|
| 1493 | failure = "cold.acquire_failed"
|
| 1494 | if failure is None:
|
| 1495 | status = agent_step("cold.inspect", [vcs, "status", "--json"], cold, preserve=True)
|
| 1496 | status_value = _parse_json_object(status.get("stdout"))
|
| 1497 | branch_value = status_value.get("branch") if status_value else None
|
| 1498 | branch = branch_value if isinstance(branch_value, str) else None
|
| 1499 | if not branch:
|
| 1500 | failure = "cold.branch_identity_missing"
|
| 1501 | if failure is None:
|
| 1502 | cold_payload = f"cold task run={run_index} bench={metadata['bench_id']}\n"
|
| 1503 | expected_payload += cold_payload
|
| 1504 | rows.append(edit_step(subject, scenario, "cold.edit", cold, cold_payload.strip(), run_index, meta, admitted))
|
| 1505 | if row_returncode(agent_step("cold.checkpoint", [vcs, "commit"], cold)) != 0:
|
| 1506 | failure = "cold.checkpoint_failed"
|
| 1507 | elif row_returncode(
|
| 1508 | agent_step("cold.publish", [vcs, "push", "--repo", repo_spec, "--remote", remote], cold)
|
| 1509 | ) != 0:
|
| 1510 | failure = "cold.publish_failed"
|
| 1511 | rows.append(
|
| 1512 | unmeasured_step_row(
|
| 1513 | subject, scenario, "cold.ci.observe", run_index, meta,
|
| 1514 | "owned local oak serve has no configured CI provider",
|
| 1515 | )
|
| 1516 | )
|
| 1517 | rows.append(
|
| 1518 | unmeasured_step_row(
|
| 1519 | subject, scenario, "cold.review_handoff", run_index, meta,
|
| 1520 | "owned local oak serve does not provide hosted review semantics",
|
| 1521 | )
|
| 1522 | )
|
| 1523 | if failure is None and branch is not None:
|
| 1524 | row = agent_step(
|
| 1525 | "followup.acquire",
|
| 1526 | [
|
| 1527 | vcs, "clone", repo_spec, str(followup), "--remote", remote,
|
| 1528 | "--branch", branch, "--allow-legacy-scope",
|
| 1529 | ],
|
| 1530 | run_root,
|
| 1531 | )
|
| 1532 | if row_returncode(row) != 0:
|
| 1533 | failure = "followup.acquire_failed"
|
| 1534 | if failure is None:
|
| 1535 | agent_step("followup.inspect", [vcs, "status", "--json"], followup)
|
| 1536 | followup_payload = f"followup task run={run_index} bench={metadata['bench_id']}\n"
|
| 1537 | expected_payload += followup_payload
|
| 1538 | rows.append(
|
| 1539 | edit_step(
|
| 1540 | subject, scenario, "followup.edit", followup,
|
| 1541 | followup_payload.strip(), run_index, meta, admitted,
|
| 1542 | )
|
| 1543 | )
|
| 1544 | if row_returncode(agent_step("followup.checkpoint", [vcs, "commit"], followup)) != 0:
|
| 1545 | failure = "followup.checkpoint_failed"
|
| 1546 | else:
|
| 1547 | head_result = run_bounded_harness_command([vcs, "hash"], followup)
|
| 1548 | expected_head = head_result.stdout.strip() if head_result.returncode == 0 else None
|
| 1549 | if row_returncode(
|
| 1550 | agent_step(
|
| 1551 | "followup.publish",
|
| 1552 | [vcs, "push", "--repo", repo_spec, "--remote", remote],
|
| 1553 | followup,
|
| 1554 | )
|
| 1555 | ) != 0:
|
| 1556 | failure = "followup.publish_failed"
|
| 1557 | rows.append(
|
| 1558 | unmeasured_step_row(
|
| 1559 | subject, scenario, "followup.ci.observe", run_index, meta,
|
| 1560 | "owned local oak serve has no configured CI provider",
|
| 1561 | )
|
| 1562 | )
|
| 1563 | rows.append(
|
| 1564 | unmeasured_step_row(
|
| 1565 | subject, scenario, "followup.review_handoff", run_index, meta,
|
| 1566 | "owned local oak serve does not provide hosted review semantics",
|
| 1567 | )
|
| 1568 | )
|
| 1569 |
|
| 1570 | payload_ok = False
|
| 1571 | verifier_head: str | None = None
|
| 1572 | advertised_head: str | None = None
|
| 1573 | if failure is None and branch is not None:
|
| 1574 | clone = run_bounded_harness_command(
|
| 1575 | [
|
| 1576 | vcs, "clone", repo_spec, str(verifier), "--remote", remote,
|
| 1577 | "--branch", branch, "--allow-legacy-scope",
|
| 1578 | ],
|
| 1579 | run_root,
|
| 1580 | )
|
| 1581 | if clone.returncode == 0:
|
| 1582 | payload_path = verifier / "task-notes.md"
|
| 1583 | payload_ok = payload_path.exists() and payload_path.read_text() == expected_payload
|
| 1584 | verifier_result = run_bounded_harness_command([vcs, "hash"], verifier)
|
| 1585 | verifier_head = verifier_result.stdout.strip() if verifier_result.returncode == 0 else None
|
| 1586 | advertised_head = remote_branch_head(remote, repo_spec, branch)
|
| 1587 | else:
|
| 1588 | failure = "verifier_clone_failed"
|
| 1589 | head_ok = bool(expected_head and expected_head == verifier_head == advertised_head)
|
| 1590 | if not payload_ok:
|
| 1591 | failure = failure or "local_oak_payload_oracle_failed"
|
| 1592 | if not head_ok:
|
| 1593 | failure = failure or "local_oak_head_oracle_failed"
|
| 1594 | finally:
|
| 1595 | if server is not None:
|
| 1596 | stop_owned_process(server)
|
| 1597 | server_stopped = server is None or server.poll() is not None
|
| 1598 | cleanup_started = time.perf_counter()
|
| 1599 | for path in (seed, cold, followup, verifier, server_root):
|
| 1600 | shutil.rmtree(path, ignore_errors=True)
|
| 1601 | rows.append(
|
| 1602 | harness_step_row(
|
| 1603 | subject,
|
| 1604 | scenario,
|
| 1605 | "cleanup.local_server_and_checkouts",
|
| 1606 | run_index,
|
| 1607 | meta,
|
| 1608 | (time.perf_counter() - cleanup_started) * 1000,
|
| 1609 | )
|
| 1610 | )
|
| 1611 |
|
| 1612 | total = phase_total_row(rows, subject, scenario, run_index, meta, "agent_task_oak_local.total")
|
| 1613 | total.update(
|
| 1614 | {
|
| 1615 | "oracle_payload_on_remote_branch": payload_ok and head_ok,
|
| 1616 | "oracle_payload_matches": payload_ok,
|
| 1617 | "oracle_branch_head_matches": head_ok,
|
| 1618 | "server_process_stopped": server_stopped,
|
| 1619 | "acquisition_models": {
|
| 1620 | "cold": "fresh_full_clone",
|
| 1621 | "followup": "second_fresh_full_clone",
|
| 1622 | },
|
| 1623 | "repository_cache_state": {
|
| 1624 | "cold": "fresh_destination_no_shared_object_store",
|
| 1625 | "followup": "no_shared_object_store",
|
| 1626 | },
|
| 1627 | "os_cache_state": "unknown",
|
| 1628 | "os_cache_state_reason": "page_cache_not_purged_or_instrumented",
|
| 1629 | "network_bytes": None,
|
| 1630 | "network_bytes_reason": "loopback transport not observed by byte proxy",
|
| 1631 | "hosted_ci_configured": False,
|
| 1632 | "oracle_probe_calls": 4,
|
| 1633 | # The expected-head hash plus verifier clone/hash/HTTP head probe
|
| 1634 | # are oracle/setup commands, not agent-facing tool calls.
|
| 1635 | "harness_command_count": total["harness_command_count"] + 4,
|
| 1636 | }
|
| 1637 | )
|
| 1638 | if failure or not server_stopped:
|
| 1639 | total["returncode"] = 1
|
| 1640 | total["stderr"] = failure or "local_oak_server_cleanup_failed"
|
| 1641 | rows.append(total)
|
| 1642 | return rows
|
| 1643 |
|
| 1644 |
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 1645 | def run_legacy_selected_materialization_campaign(
|
| 1646 | subjects: list[Subject],
|
| 1647 | server_bin: Path | None,
|
| 1648 | server_source: Path | None,
|
| 1649 | server_source_head: str | None,
|
| 1650 | baseline_subject_name: str | None,
|
| 1651 | run_root: Path,
|
| 1652 | runs: int,
|
| 1653 | metadata: dict[str, Any],
|
| 1654 | admitted: int,
|
| 1655 | randomize_order: bool,
|
| 1656 | random_seed: int,
|
| 1657 | timing_controls_requested: bool,
|
| 1658 | ) -> list[dict[str, Any]]:
|
| 1659 | """Measure one legacy selected-branch clone against one fixed Serve fixture."""
|
| 1660 | scenario = LEGACY_SELECTED_MATERIALIZATION_SCENARIO
|
| 1661 | oak_subjects = [subject for subject in subjects if subject.kind == "oak"]
|
| 1662 | if not oak_subjects:
|
| 1663 | return []
|
| 1664 | if (
|
| 1665 | server_bin is None
|
| 1666 | or not server_bin.is_absolute()
|
| 1667 | or not server_bin.is_file()
|
| 1668 | or not os.access(server_bin, os.X_OK)
|
| 1669 | ):
|
| 1670 | return [
|
| 1671 | unmeasured_step_row(
|
| 1672 | subject,
|
| 1673 | scenario,
|
| 1674 | "scenario.skip",
|
| 1675 | 0,
|
| 1676 | metadata,
|
| 1677 | "--oak-serve-bin must name an absolute executable used only as the fixed Serve fixture",
|
| 1678 | )
|
| 1679 | for subject in oak_subjects
|
| 1680 | ]
|
| 1681 | if server_source_head is None or re.fullmatch(r"[0-9a-f]{64}", server_source_head) is None:
|
| 1682 | return [
|
| 1683 | unmeasured_step_row(
|
| 1684 | subject,
|
| 1685 | scenario,
|
| 1686 | "scenario.skip",
|
| 1687 | 0,
|
| 1688 | metadata,
|
| 1689 | "the fixed Serve binary's exact 64-hex --oak-serve-source-head is required",
|
| 1690 | )
|
| 1691 | for subject in oak_subjects
|
| 1692 | ]
|
| 1693 | baseline_subject = next(
|
| 1694 | (subject for subject in oak_subjects if subject.name == baseline_subject_name), None
|
| 1695 | )
|
| 1696 | if baseline_subject is None:
|
| 1697 | return [
|
| 1698 | unmeasured_step_row(
|
| 1699 | subject,
|
| 1700 | scenario,
|
| 1701 | "scenario.skip",
|
| 1702 | 0,
|
| 1703 | metadata,
|
| 1704 | "--materialization-baseline-subject must name one enabled Oak subject",
|
| 1705 | )
|
| 1706 | for subject in oak_subjects
|
| 1707 | ]
|
| 1708 | git_bin = shutil.which("git")
|
| 1709 | if git_bin is None:
|
| 1710 | return [
|
| 1711 | unmeasured_step_row(
|
| 1712 | subject,
|
| 1713 | scenario,
|
| 1714 | "scenario.skip",
|
| 1715 | 0,
|
| 1716 | metadata,
|
| 1717 | "git is required only to create the deterministic main-history fixture",
|
| 1718 | )
|
| 1719 | for subject in oak_subjects
|
| 1720 | ]
|
| 1721 |
|
| 1722 | server_bin = server_bin.resolve()
|
| 1723 | resolved_server_source: Path | None = None
|
| 1724 | if server_source is not None:
|
| 1725 | if not server_source.is_absolute() or not server_source.is_dir():
|
| 1726 | return [
|
| 1727 | unmeasured_step_row(
|
| 1728 | subject,
|
| 1729 | scenario,
|
| 1730 | "scenario.skip",
|
| 1731 | 0,
|
| 1732 | metadata,
|
| 1733 | "--oak-serve-source must name an absolute read-only source checkout",
|
| 1734 | )
|
| 1735 | for subject in oak_subjects
|
| 1736 | ]
|
| 1737 | resolved_server_source = server_source.resolve()
|
| 1738 | source_head_result = run_bounded_harness_command(
|
| 1739 | [str(server_bin), "hash"],
|
| 1740 | resolved_server_source,
|
| 1741 | env=isolated_oak_fixture_env(run_root / "source-check"),
|
| 1742 | )
|
| 1743 | if source_head_result.returncode != 0 or source_head_result.stdout.strip() != server_source_head:
|
| 1744 | return [
|
| 1745 | unmeasured_step_row(
|
| 1746 | subject,
|
| 1747 | scenario,
|
| 1748 | "scenario.skip",
|
| 1749 | 0,
|
| 1750 | metadata,
|
| 1751 | "fixed Serve source checkout does not match --oak-serve-source-head",
|
| 1752 | )
|
| 1753 | for subject in oak_subjects
|
| 1754 | ]
|
| 1755 | baseline_binary_sha256 = binary_sha256(baseline_subject.bin)
|
| 1756 | timing_controls_met = materialization_timing_controls_met(
|
| 1757 | metadata, runs, randomize_order, timing_controls_requested
|
| 1758 | )
|
| 1759 | fixture_root = run_root / "legacy-selected-materialization-fixture"
|
| 1760 | server_root = fixture_root / "server"
|
| 1761 | git_seed = fixture_root / "git-seed.git"
|
| 1762 | seed = fixture_root / "seed"
|
| 1763 | task_config_root = fixture_root / "config"
|
| 1764 | for path in (server_root, git_seed, task_config_root):
|
| 1765 | path.mkdir(parents=True, exist_ok=True)
|
| 1766 | fixture_env = isolated_oak_fixture_env(task_config_root)
|
| 1767 | fixture_meta = {
|
| 1768 | **metadata,
|
| 1769 | "benchmark_track": "oak-local-loopback-materialization",
|
| 1770 | "remote_transport": "loopback_http",
|
| 1771 | "remote_server": "127.0.0.1",
|
| 1772 | "serve_binary": str(server_bin),
|
| 1773 | "serve_binary_sha256": binary_sha256(server_bin),
|
| 1774 | "serve_source": str(resolved_server_source) if resolved_server_source else None,
|
| 1775 | "serve_source_checkout_verification": (
|
| 1776 | "exact_head_verified" if resolved_server_source else "unavailable_binary_hash_and_head_supplied"
|
| 1777 | ),
|
| 1778 | "serve_source_head": server_source_head,
|
| 1779 | "serve_version": subject_version(Subject("serve_fixture", "oak", "Serve fixture", server_bin)),
|
| 1780 | "baseline_subject": baseline_subject.name,
|
| 1781 | "baseline_binary_sha256": baseline_binary_sha256,
|
| 1782 | "fixture_id": "legacy-selected-materialization-192x16k-v1",
|
| 1783 | "fixture_regular_files": 192,
|
| 1784 | "fixture_regular_file_bytes": 16 * 1024,
|
| 1785 | "fixture_empty_files": 1,
|
| 1786 | "fixture_executable_files": 8,
|
| 1787 | "fixture_selected_only_files": 1,
|
| 1788 | "os_cache_state": "warm_or_unknown_after_required_fixture_probe",
|
| 1789 | "materialization_timing_controls_met": timing_controls_met,
|
| 1790 | "materialization_timing_controls_reason": (
|
| 1791 | "harness_controls_met_aa_and_quiet_host_evidence_still_required"
|
| 1792 | if timing_controls_met
|
| 1793 | else "correctness_only_or_harness_controls_not_met"
|
| 1794 | ),
|
| 1795 | }
|
| 1796 | rows: list[dict[str, Any]] = []
|
| 1797 | setup_subject = baseline_subject
|
| 1798 |
|
| 1799 | def setup_command(operation: str, command: list[str], cwd: Path) -> bool:
|
| 1800 | started = time.perf_counter()
|
| 1801 | result = run_bounded_harness_command(command, cwd, env=fixture_env)
|
| 1802 | rows.append(
|
| 1803 | harness_step_row(
|
| 1804 | setup_subject,
|
| 1805 | scenario,
|
| 1806 | operation,
|
| 1807 | 0,
|
| 1808 | fixture_meta,
|
| 1809 | (time.perf_counter() - started) * 1000,
|
| 1810 | returncode=0 if result.returncode == 0 else result.returncode,
|
| 1811 | harness_command_count=1,
|
| 1812 | )
|
| 1813 | )
|
| 1814 | if result.returncode != 0:
|
| 1815 | rows[-1]["stderr"] = result.stderr[-4000:]
|
| 1816 | return result.returncode == 0
|
| 1817 |
|
| 1818 | server: subprocess.Popen[Any] | None = None
|
| 1819 | proxy: CountingProxy | None = None
|
| 1820 | failure: str | None = None
|
| 1821 | expected_artifacts: list[str] | None = None
|
| 1822 | baseline_cache: dict[str, Any] | None = None
|
| 1823 | baseline_receipt_contract: dict[str, Any] | None = None
|
| 1824 | baseline_network_contract: dict[str, Any] | None = None
|
| 1825 | expected_status_contract: dict[str, Any] | None = None
|
| 1826 | selected_branch = "selected-materialization"
|
| 1827 | repo_spec = "qa/materialization-bench"
|
| 1828 | try:
|
| 1829 | for index in range(192):
|
| 1830 | directory = git_seed / f"group-{index % 12:02d}"
|
| 1831 | directory.mkdir(parents=True, exist_ok=True)
|
| 1832 | digest = hashlib.sha256(f"oak-materialization-{index}".encode()).digest()
|
| 1833 | content = (digest * ((16 * 1024 + len(digest) - 1) // len(digest)))[: 16 * 1024]
|
| 1834 | path = directory / f"file-{index:03d}.bin"
|
| 1835 | path.write_bytes(content)
|
| 1836 | if index < 8:
|
| 1837 | path.chmod(0o755)
|
| 1838 | (git_seed / "empty.txt").write_bytes(b"")
|
| 1839 | for operation, command in (
|
| 1840 | ("setup.git_init", [git_bin, "init", "-q", "-b", "main"]),
|
| 1841 | ("setup.git_add", [git_bin, "add", "."]),
|
| 1842 | ("setup.git_commit", [git_bin, "commit", "-q", "-m", "fixture main"]),
|
| 1843 | ):
|
| 1844 | if not setup_command(operation, command, git_seed):
|
| 1845 | failure = f"{operation}_failed"
|
| 1846 | break
|
| 1847 |
|
| 1848 | port = reserve_loopback_port()
|
| 1849 | direct_remote = f"http://127.0.0.1:{port}"
|
| 1850 | if failure is None:
|
| 1851 | server = subprocess.Popen(
|
| 1852 | [
|
| 1853 | str(server_bin),
|
| 1854 | "serve",
|
| 1855 | "--dir",
|
| 1856 | str(server_root),
|
| 1857 | "--host",
|
| 1858 | "127.0.0.1",
|
| 1859 | "--port",
|
| 1860 | str(port),
|
| 1861 | ],
|
| 1862 | cwd=fixture_root,
|
| 1863 | env=fixture_env,
|
| 1864 | stdout=subprocess.DEVNULL,
|
| 1865 | stderr=subprocess.DEVNULL,
|
| 1866 | start_new_session=os.name == "posix",
|
| 1867 | )
|
| 1868 | if not wait_for_loopback_port(server, port):
|
| 1869 | failure = "setup.server_not_ready"
|
| 1870 | if failure is None:
|
| 1871 | for operation, command, cwd in (
|
| 1872 | ("setup.oak_import", [str(server_bin), "clone", str(git_seed), str(seed)], fixture_root),
|
| 1873 | ("setup.branch", [str(server_bin), "switch", "-c", selected_branch], seed),
|
| 1874 | ):
|
| 1875 | if not setup_command(operation, command, cwd):
|
| 1876 | failure = f"{operation}_failed"
|
| 1877 | break
|
| 1878 | if failure is None:
|
| 1879 | (seed / "selected.txt").write_text("selected branch\n")
|
| 1880 | for operation, command in (
|
| 1881 | ("setup.oak_commit", [str(server_bin), "commit"]),
|
| 1882 | (
|
| 1883 | "setup.oak_push",
|
| 1884 | [str(server_bin), "push", "--repo", repo_spec, "--remote", direct_remote],
|
| 1885 | ),
|
| 1886 | ):
|
| 1887 | if not setup_command(operation, command, seed):
|
| 1888 | failure = f"{operation}_failed"
|
| 1889 | break
|
| 1890 |
|
| 1891 | expected_head_result = (
|
| 1892 | run_bounded_harness_command([str(server_bin), "hash"], seed, env=fixture_env)
|
| 1893 | if failure is None
|
| 1894 | else None
|
| 1895 | )
|
| 1896 | expected_head = (
|
| 1897 | expected_head_result.stdout.strip()
|
| 1898 | if expected_head_result is not None and expected_head_result.returncode == 0
|
| 1899 | else None
|
| 1900 | )
|
| 1901 | expected_history = (
|
| 1902 | normalized_history(json_command([str(server_bin), "log", "--json"], seed, fixture_env))
|
| 1903 | if failure is None
|
| 1904 | else None
|
| 1905 | )
|
| 1906 | expected_worktree = normalized_worktree_snapshot(seed) if failure is None else None
|
| 1907 | if not expected_head or expected_history is None or expected_worktree is None:
|
| 1908 | failure = failure or "setup.fixture_oracle_failed"
|
| 1909 |
|
| 1910 | if failure is None:
|
| 1911 | proxy = CountingProxy("127.0.0.1", port).start()
|
| 1912 | measured_remote = f"http://127.0.0.1:{proxy.address[1]}"
|
| 1913 | probe = fixture_root / "baseline-probe"
|
| 1914 | probe_env = isolated_oak_fixture_env(fixture_root / "baseline-probe-config")
|
| 1915 | proxy.reset_counts()
|
| 1916 | result = run_bounded_harness_command(
|
| 1917 | [
|
| 1918 | str(baseline_subject.bin),
|
| 1919 | "clone",
|
| 1920 | "--json",
|
| 1921 | "--allow-legacy-scope",
|
| 1922 | "--remote",
|
| 1923 | measured_remote,
|
| 1924 | "--branch",
|
| 1925 | selected_branch,
|
| 1926 | repo_spec,
|
| 1927 | str(probe),
|
| 1928 | ],
|
| 1929 | fixture_root,
|
| 1930 | env=probe_env,
|
| 1931 | )
|
| 1932 | receipt = _parse_json_object(result.stdout)
|
| 1933 | evidence = legacy_selected_receipt_evidence(receipt or {})
|
| 1934 | baseline_proxy_stats = proxy.snapshot()
|
| 1935 | baseline_cache = stat_cache_snapshot(probe)
|
| 1936 | expected_artifacts = normalized_oak_artifacts(probe)
|
| 1937 | baseline_receipt_contract = legacy_selected_receipt_contract(evidence)
|
| 1938 | baseline_network_contract = {
|
| 1939 | "connections": baseline_proxy_stats.connections,
|
| 1940 | "receipt": baseline_receipt_contract,
|
| 1941 | }
|
| 1942 | expected_status_contract = normalized_status(
|
| 1943 | json_command([str(baseline_subject.bin), "status", "--json"], probe, probe_env)
|
| 1944 | )
|
| 1945 | probe_checks = {
|
| 1946 | "returncode": result.returncode == 0,
|
| 1947 | "legacy_route": evidence["legacy_route_observed"],
|
| 1948 | "passes_two": evidence["materialization_passes_observed"] == 2,
|
| 1949 | "head": evidence["head"] == expected_head,
|
| 1950 | "manifest": bool(evidence["manifest"]),
|
| 1951 | "artifacts": expected_artifacts is not None,
|
| 1952 | "cache": normalized_cache_contract(baseline_cache) is not None,
|
| 1953 | "status": expected_status_contract is not None,
|
| 1954 | "network": baseline_proxy_stats.connection_errors == 0,
|
| 1955 | }
|
| 1956 | if not all(probe_checks.values()):
|
| 1957 | failed_checks = ",".join(key for key, passed in probe_checks.items() if not passed)
|
| 1958 | failure = f"baseline_probe_failed:{failed_checks}"
|
| 1959 | if isinstance(baseline_cache, dict) and baseline_cache.get("error"):
|
| 1960 | failure += f":{baseline_cache['error']}"
|
| 1961 | shutil.rmtree(probe, ignore_errors=True)
|
| 1962 |
|
| 1963 | if failure is None and proxy is not None:
|
| 1964 | measured_remote = f"http://127.0.0.1:{proxy.address[1]}"
|
| 1965 | order_rng = random.Random(random_seed)
|
| 1966 | for run_index in range(runs):
|
| 1967 | run_subjects = list(oak_subjects)
|
| 1968 | if randomize_order:
|
| 1969 | order_rng.shuffle(run_subjects)
|
| 1970 | for subject in run_subjects:
|
| 1971 | destination = run_root / f"materialization-{subject.name}-r{run_index}"
|
| 1972 | subject_env = isolated_oak_fixture_env(
|
| 1973 | run_root / f"materialization-config-{subject.name}-r{run_index}"
|
| 1974 | )
|
| 1975 | proxy.reset_counts()
|
| 1976 | row = step_row(
|
| 1977 | subject,
|
| 1978 | scenario,
|
| 1979 | "legacy_selected.acquire",
|
| 1980 | [
|
| 1981 | str(subject.bin),
|
| 1982 | "clone",
|
| 1983 | "--json",
|
| 1984 | "--allow-legacy-scope",
|
| 1985 | "--remote",
|
| 1986 | measured_remote,
|
| 1987 | "--branch",
|
| 1988 | selected_branch,
|
| 1989 | repo_spec,
|
| 1990 | str(destination),
|
| 1991 | ],
|
| 1992 | fixture_root,
|
| 1993 | run_index,
|
| 1994 | {
|
| 1995 | **fixture_meta,
|
| 1996 | "run_subject_order": [item.name for item in run_subjects],
|
| 1997 | },
|
| 1998 | admitted,
|
| 1999 | preserve_stdout=True,
|
| 2000 | env=subject_env,
|
| 2001 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 2002 | )
|
| 2003 | receipt = _parse_json_object(row.get("stdout"))
|
| 2004 | evidence = legacy_selected_receipt_evidence(receipt or {})
|
| 2005 | proxy_stats = proxy.snapshot()
|
| 2006 | cache_before_status = stat_cache_snapshot(destination)
|
| 2007 | artifacts = normalized_oak_artifacts(destination)
|
| 2008 | trial_meta = {
|
| 2009 | **fixture_meta,
|
| 2010 | "run_subject_order": [item.name for item in run_subjects],
|
| 2011 | }
|
| 2012 | status_row = step_row(
|
| 2013 | subject,
|
| 2014 | scenario,
|
| 2015 | "legacy_selected.first_status",
|
| 2016 | [str(subject.bin), "status", "--json"],
|
| 2017 | destination,
|
| 2018 | run_index,
|
| 2019 | trial_meta,
|
| 2020 | admitted,
|
| 2021 | preserve_stdout=True,
|
| 2022 | env=subject_env,
|
| 2023 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 2024 | )
|
| 2025 | status = normalized_status(_parse_json_object(status_row.get("stdout")) or {})
|
| 2026 | commit_row = step_row(
|
| 2027 | subject,
|
| 2028 | scenario,
|
| 2029 | "legacy_selected.first_noop_commit",
|
| 2030 | [str(subject.bin), "commit"],
|
| 2031 | destination,
|
| 2032 | run_index,
|
| 2033 | trial_meta,
|
| 2034 | admitted,
|
| 2035 | env=subject_env,
|
| 2036 | timeout_seconds=INTERRUPTED_PUBLICATION_PROCESS_TIMEOUT_S,
|
| 2037 | )
|
| 2038 | history = normalized_history(
|
| 2039 | json_command([str(subject.bin), "log", "--json"], destination, subject_env)
|
| 2040 | )
|
| 2041 | head_result = run_bounded_harness_command(
|
| 2042 | [str(subject.bin), "hash"], destination, env=subject_env
|
| 2043 | )
|
| 2044 | local_head = head_result.stdout.strip() if head_result.returncode == 0 else None
|
| 2045 | worktree = normalized_worktree_snapshot(destination) if destination.exists() else None
|
| 2046 | advertised_head = remote_branch_head(direct_remote, repo_spec, selected_branch)
|
| 2047 | subject_sha256 = binary_sha256(subject.bin)
|
| 2048 | expected_passes = 2 if subject_sha256 == baseline_binary_sha256 else 1
|
| 2049 | receipt_contract = legacy_selected_receipt_contract(evidence)
|
| 2050 | network_contract = {
|
| 2051 | "connections": proxy_stats.connections,
|
| 2052 | "receipt": receipt_contract,
|
| 2053 | }
|
| 2054 | cache_contract = normalized_cache_contract(cache_before_status)
|
| 2055 | baseline_cache_contract = normalized_cache_contract(baseline_cache)
|
| 2056 | expected_cache_valid = subject_sha256 != baseline_binary_sha256
|
| 2057 | cache_valid = bool(
|
| 2058 | isinstance(cache_before_status, dict)
|
| 2059 | and cache_before_status.get("all_metadata_valid") is expected_cache_valid
|
| 2060 | )
|
| 2061 | status_ok = status == expected_status_contract
|
| 2062 | artifacts_ok = artifacts == expected_artifacts
|
| 2063 | network_ok = network_contract == baseline_network_contract
|
| 2064 | oracle_ok = bool(
|
| 2065 | row_returncode(row) == 0
|
| 2066 | and row_returncode(status_row) == 0
|
| 2067 | and row_returncode(commit_row) == 0
|
| 2068 | and evidence["legacy_route_observed"]
|
| 2069 | and evidence["snapshot_bound"] is False
|
| 2070 | and evidence["materialization_passes_observed"] == expected_passes
|
| 2071 | and evidence["head"] == expected_head == local_head == advertised_head
|
| 2072 | and evidence["manifest"] == baseline_receipt_contract["manifest"]
|
| 2073 | and evidence["written_paths"] == len(expected_worktree)
|
| 2074 | and receipt_contract == baseline_receipt_contract
|
| 2075 | and history == expected_history
|
| 2076 | and worktree == expected_worktree
|
| 2077 | and status_ok
|
| 2078 | and artifacts_ok
|
| 2079 | and cache_contract == baseline_cache_contract
|
| 2080 | and cache_valid
|
| 2081 | and network_ok
|
| 2082 | and proxy_stats.connection_errors == 0
|
| 2083 | )
|
| 2084 | row.update(
|
| 2085 | {
|
| 2086 | **evidence,
|
| 2087 | "network_connections_observed": proxy_stats.connections,
|
| 2088 | "network_client_to_server_bytes_observed": proxy_stats.client_to_upstream_bytes,
|
| 2089 | "network_server_to_client_bytes_observed": proxy_stats.upstream_to_client_bytes,
|
| 2090 | "network_connection_errors": proxy_stats.connection_errors,
|
| 2091 | "normalized_network_work": network_contract,
|
| 2092 | "oracle_network_work_matches_baseline": network_ok,
|
| 2093 | "oracle_head_matches": expected_head == local_head == advertised_head,
|
| 2094 | "oracle_manifest_matches": evidence["manifest"]
|
| 2095 | == baseline_receipt_contract["manifest"],
|
| 2096 | "oracle_receipt_contract_matches": receipt_contract
|
| 2097 | == baseline_receipt_contract,
|
| 2098 | "oracle_history_matches": history == expected_history,
|
| 2099 | "oracle_worktree_bytes_and_modes_match": worktree == expected_worktree,
|
| 2100 | "oracle_clean_status": status_ok,
|
| 2101 | "oracle_local_artifact_names_match": artifacts_ok,
|
| 2102 | "oracle_stat_cache_contract_matches": cache_contract
|
| 2103 | == baseline_cache_contract,
|
| 2104 | "stat_cache_metadata_valid_before_status": (
|
| 2105 | cache_before_status.get("all_metadata_valid")
|
| 2106 | if isinstance(cache_before_status, dict)
|
| 2107 | else None
|
| 2108 | ),
|
| 2109 | "stat_cache_rows_before_status": (
|
| 2110 | len(cache_before_status.get("rows", []))
|
| 2111 | if isinstance(cache_before_status, dict)
|
| 2112 | and isinstance(cache_before_status.get("rows"), list)
|
| 2113 | else None
|
| 2114 | ),
|
| 2115 | "stat_cache_expected_valid_before_status": expected_cache_valid,
|
| 2116 | "oracle_clone_residue_absent": artifacts_ok,
|
| 2117 | "oracle_passed": oracle_ok,
|
| 2118 | "repository_cache_state": "fresh_destination_no_shared_object_store",
|
| 2119 | }
|
| 2120 | )
|
| 2121 | for followup_row in (status_row, commit_row):
|
| 2122 | followup_row.update(
|
| 2123 | {
|
| 2124 | "materialization_timing_controls_met": timing_controls_met,
|
| 2125 | "stat_cache_expected_valid_before_first_status": expected_cache_valid,
|
| 2126 | "oracle_passed": oracle_ok,
|
| 2127 | }
|
| 2128 | )
|
| 2129 | if not oracle_ok:
|
| 2130 | row["returncode"] = 1
|
| 2131 | row["stderr"] = "legacy_selected_materialization_oracle_failed"
|
| 2132 | status_row["returncode"] = 1
|
| 2133 | commit_row["returncode"] = 1
|
| 2134 | total = phase_total_row(
|
| 2135 | [row, status_row, commit_row],
|
| 2136 | subject,
|
| 2137 | scenario,
|
| 2138 | run_index,
|
| 2139 | trial_meta,
|
| 2140 | "legacy_selected_clone_materialization.total",
|
| 2141 | )
|
| 2142 | total.update(
|
| 2143 | {
|
| 2144 | key: row[key]
|
| 2145 | for key in (
|
| 2146 | "materialization_passes_observed",
|
| 2147 | "network_connections_observed",
|
| 2148 | "network_client_to_server_bytes_observed",
|
| 2149 | "network_server_to_client_bytes_observed",
|
| 2150 | "oracle_passed",
|
| 2151 | "oracle_head_matches",
|
| 2152 | "oracle_manifest_matches",
|
| 2153 | "oracle_receipt_contract_matches",
|
| 2154 | "oracle_history_matches",
|
| 2155 | "oracle_worktree_bytes_and_modes_match",
|
| 2156 | "oracle_clean_status",
|
| 2157 | "oracle_local_artifact_names_match",
|
| 2158 | "oracle_stat_cache_contract_matches",
|
| 2159 | "stat_cache_metadata_valid_before_status",
|
| 2160 | "stat_cache_rows_before_status",
|
| 2161 | "oracle_clone_residue_absent",
|
| 2162 | "oracle_network_work_matches_baseline",
|
| 2163 | "serve_binary_sha256",
|
| 2164 | "serve_source_head",
|
| 2165 | "fixture_id",
|
| 2166 | "os_cache_state",
|
| 2167 | "materialization_timing_controls_met",
|
| 2168 | )
|
| 2169 | }
|
| 2170 | )
|
| 2171 | total.update(
|
| 2172 | {
|
| 2173 | "clone_elapsed_ms": row.get("elapsed_ms"),
|
| 2174 | "first_status_elapsed_ms": status_row.get("elapsed_ms"),
|
| 2175 | "first_noop_commit_elapsed_ms": commit_row.get("elapsed_ms"),
|
| 2176 | }
|
| 2177 | )
|
| 2178 | rows.extend([row, status_row, commit_row, total])
|
| 2179 | shutil.rmtree(destination, ignore_errors=True)
|
| 2180 | finally:
|
| 2181 | if proxy is not None:
|
| 2182 | proxy.stop()
|
| 2183 | if server is not None:
|
| 2184 | stop_owned_process(server)
|
| 2185 | shutil.rmtree(fixture_root, ignore_errors=True)
|
| 2186 |
|
| 2187 | if failure is not None:
|
| 2188 | rows.append(
|
| 2189 | harness_step_row(
|
| 2190 | setup_subject,
|
| 2191 | scenario,
|
| 2192 | "scenario.fixture_failure",
|
| 2193 | 0,
|
| 2194 | fixture_meta,
|
| 2195 | 0.0,
|
| 2196 | returncode=1,
|
| 2197 | )
|
| 2198 | )
|
| 2199 | rows[-1]["stderr"] = failure
|
| 2200 | return rows
|
| 2201 |
|
| 2202 |
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2203 | def run_git_loop(
|
| 2204 | subject: Subject,
|
| 2205 | remote: oakbench_remotes.RemoteResolution,
|
| 2206 | run_root: Path,
|
| 2207 | run_index: int,
|
| 2208 | metadata: dict[str, Any],
|
| 2209 | admitted: int,
|
| 2210 | ) -> list[dict[str, Any]]:
|
| 2211 | scenario = "task_loop"
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2212 | meta = {
|
| 2213 | **metadata,
|
| 2214 | **remote.row_fields(),
|
| 2215 | "benchmark_track": (
|
| 2216 | "core-equivalent"
|
| 2217 | if remote.transport == oakbench_remotes.TRANSPORT_LOCAL_FILE
|
| 2218 | else "agent-default"
|
| 2219 | ),
|
| 2220 | }
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2221 | vcs = str(subject.bin)
|
| 2222 | rows: list[dict[str, Any]] = []
|
| 2223 |
|
| 2224 | seed_error = ensure_git_remote_seeded(subject, remote, run_root)
|
| 2225 | if seed_error:
|
| 2226 | return [skip_row(subject, scenario, run_index, meta, seed_error)]
|
| 2227 |
|
| 2228 | # Untimed: the base clone is the long-lived checkout worktrees hang off;
|
| 2229 | # an agent fleet pays it once, not per task.
|
| 2230 | base = run_root / f"tl-git-base-r{run_index}"
|
| 2231 | clone = run_untimed([vcs, "clone", "-q", str(remote.repo), str(base)], run_root)
|
| 2232 | if clone.returncode != 0:
|
| 2233 | return [skip_row(subject, scenario, run_index, meta, f"base clone failed: {clone.stderr.strip()[:200]}")]
|
| 2234 |
|
| 2235 | branches: dict[str, str] = {}
|
| 2236 | payloads: dict[str, str] = {}
|
| 2237 | for phase in PHASES:
|
| 2238 | workspace = run_root / f"tl-{subject.name}-r{run_index}-{phase}"
|
| 2239 | branch = oakbench_remotes.disposable_branch(
|
| 2240 | str(metadata["bench_id"]), f"taskloop-{phase}", subject.name, run_index
|
| 2241 | )
|
| 2242 | branches[phase] = branch
|
| 2243 | payload = f"payload {phase} run={run_index} bench={metadata['bench_id']}"
|
| 2244 | payloads[phase] = payload
|
| 2245 | acquire = step_row(
|
| 2246 | subject, scenario, f"{phase}.acquire",
|
| 2247 | [vcs, "worktree", "add", str(workspace), "-b", branch], base, run_index, meta, admitted,
|
| 2248 | )
|
| 2249 | rows.append(acquire)
|
| 2250 | if int(acquire["returncode"]) != 0:
|
| 2251 | return rows
|
| 2252 | rows.append(edit_step(subject, scenario, f"{phase}.edit", workspace, payload, run_index, meta, admitted))
|
| 2253 | # Agent-default snapshot: git pays stage+commit as two calls.
|
| 2254 | rows.append(step_row(subject, scenario, f"{phase}.snapshot.add", [vcs, "add", "."], workspace, run_index, meta, admitted))
|
| 2255 | rows.append(
|
| 2256 | step_row(
|
| 2257 | subject, scenario, f"{phase}.snapshot.commit",
|
| 2258 | [vcs, "commit", "-q", "-m", f"bench task loop {metadata['bench_id']} r{run_index} {phase}"],
|
| 2259 | workspace, run_index, meta, admitted,
|
| 2260 | )
|
| 2261 | )
|
| 2262 | rows.append(
|
| 2263 | step_row(
|
| 2264 | subject, scenario, f"{phase}.publish",
|
| 2265 | [vcs, "push", "-q", "-u", "origin", branch], workspace, run_index, meta, admitted,
|
| 2266 | )
|
| 2267 | )
|
| 2268 | # No task.describe: git's description is the commit message already
|
| 2269 | # paid in {phase}.snapshot.commit (asymmetry documented in the lane doc).
|
| 2270 | rows.append(
|
| 2271 | step_row(
|
| 2272 | subject, scenario, f"{phase}.end",
|
| 2273 | [vcs, "worktree", "remove", str(workspace)], base, run_index, meta, admitted,
|
| 2274 | )
|
| 2275 | )
|
| 2276 |
|
| 2277 | # Untimed oracle: payloads visible on the remote branches.
|
| 2278 | oracle_ok: bool | None = True
|
| 2279 | for phase in PHASES:
|
| 2280 | fetch = run_untimed([vcs, "fetch", "-q", "origin", branches[phase]], base)
|
| 2281 | show = run_untimed([vcs, "show", f"FETCH_HEAD:task-notes.md"], base)
|
| 2282 | if fetch.returncode != 0 or payloads[phase] not in show.stdout:
|
| 2283 | oracle_ok = False
|
| 2284 | rows.append(
|
| 2285 | {
|
| 2286 | **phase_total_row(rows, subject, scenario, run_index, meta, "task_loop.total"),
|
| 2287 | "oracle_payload_on_remote_branch": oracle_ok,
|
| 2288 | "workspace_branches": branches,
|
| 2289 | "measurement_source": "fetch_plus_show_grep",
|
| 2290 | }
|
| 2291 | )
|
| 2292 | return rows
|
| 2293 |
|
| 2294 |
|
| 2295 | def loop_summary_fields(rows: list[dict[str, Any]]) -> dict[str, Any]:
|
| 2296 | """followup-vs-task warm deltas computed from this run's step rows."""
|
| 2297 | def op_ms(operation: str) -> float | None:
|
| 2298 | for row in rows:
|
| 2299 | if row.get("operation") == operation:
|
| 2300 | return float(row.get("elapsed_ms", 0.0))
|
| 2301 | return None
|
| 2302 |
|
| 2303 | def phase_ms(prefix: str) -> float:
|
| 2304 | return sum(
|
| 2305 | float(row.get("elapsed_ms", 0.0))
|
| 2306 | for row in rows
|
| 2307 | if str(row.get("operation", "")).startswith(prefix + ".")
|
| 2308 | )
|
| 2309 |
|
| 2310 | task_acquire = op_ms("task.acquire")
|
| 2311 | followup_acquire = op_ms("followup.acquire")
|
| 2312 | fields: dict[str, Any] = {
|
| 2313 | "task_phase_ms": round(phase_ms("task"), 3),
|
| 2314 | "followup_phase_ms": round(phase_ms("followup"), 3),
|
| 2315 | }
|
| 2316 | if task_acquire is not None and followup_acquire is not None:
|
| 2317 | fields["followup_warm_delta_ms"] = round(followup_acquire - task_acquire, 3)
|
| 2318 | fields["acquire_ms_task"] = round(task_acquire, 3)
|
| 2319 | fields["acquire_ms_followup"] = round(followup_acquire, 3)
|
| 2320 | else:
|
| 2321 | fields["followup_warm_delta_ms"] = None
|
| 2322 | return fields
|
| 2323 |
|
| 2324 |
|
| 2325 | def summary_text(rows: list[dict[str, Any]]) -> str:
|
| 2326 | totals = [row for row in rows if row.get("operation") == "task_loop.total"]
|
| 2327 | lines = [
|
| 2328 | "# Task Loop Summary",
|
| 2329 | "",
|
| 2330 | "Whole-task cost of the signature motion: acquire workspace -> edit -> snapshot ->",
|
| 2331 | "publish -> describe -> end, then a follow-up task in a fresh workspace.",
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2332 | "Command sequences are NOT step-equivalent across subjects",
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2333 | "(oak pays desc; git's description is its commit message). Latency across unlike",
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2334 | "transports is never comparable; scripted call equivalents and estimated token proxies are.",
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2335 | "",
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2336 | "| Subject | Run | Transport | Wall ms | Scripted call equivalents | Cost-wt tokens (env) | Warm acquire delta ms | Oracle |",
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2337 | "| --- | ---: | --- | ---: | ---: | ---: | ---: | --- |",
|
| 2338 | ]
|
| 2339 | for row in totals:
|
| 2340 | oracle = row.get("oracle_payload_on_remote_branch")
|
| 2341 | lines.append(
|
| 2342 | "| `{}` | {} | {} | {} | {} | {} | {} | {} |".format(
|
| 2343 | row["subject"],
|
| 2344 | row["run"],
|
| 2345 | row.get("remote_transport", "unmeasured"),
|
| 2346 | fmt_ms(row.get("elapsed_ms")),
|
| 2347 | row.get("tool_call_count", "unmeasured"),
|
| 2348 | row.get("estimated_cost_weighted_tokens_with_envelope", "unmeasured"),
|
| 2349 | row.get("followup_warm_delta_ms", "unmeasured"),
|
| 2350 | "pass" if oracle else ("FAIL" if oracle is False else "unmeasured"),
|
| 2351 | )
|
| 2352 | )
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2353 | skips = [
|
| 2354 | row for row in rows
|
| 2355 | if row.get("skipped") and row.get("scenario") == TASK_LOOP_SCENARIO
|
| 2356 | ]
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2357 | for row in skips:
|
| 2358 | lines.append(f"| `{row['subject']}` | {row['run']} | skipped | | | | | {row['skip_reason'][:70]} |")
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2359 |
|
| 2360 | agent_totals = [
|
| 2361 | row
|
| 2362 | for row in rows
|
| 2363 | if row.get("operation") in ("agent_task.total", "agent_task_oak_local.total")
|
| 2364 | ]
|
| 2365 | if agent_totals:
|
| 2366 | lines.extend(
|
| 2367 | [
|
| 2368 | "",
|
| 2369 | "## Cold clone and warm isolated task",
|
| 2370 | "",
|
| 2371 | "Local-file Git rows compare a fresh full clone with a worktree sharing",
|
| 2372 | "the first clone's object store. Oak local rows use two independent fresh",
|
| 2373 | "clones against an owned loopback server; they do not claim a shared cache.",
|
| 2374 | "OS cache and network bytes remain unknown. Hosted CI/review are skips.",
|
| 2375 | "",
|
| 2376 | "| Subject | Run | Wall ms | Scripted call equivalents | Output bytes | Coverage | Oracle |",
|
| 2377 | "| --- | ---: | ---: | ---: | ---: | --- | --- |",
|
| 2378 | ]
|
| 2379 | )
|
| 2380 | for row in agent_totals:
|
| 2381 | lines.append(
|
| 2382 | "| `{}` | {} | {} | {} | {} | {} | {} |".format(
|
| 2383 | row["subject"],
|
| 2384 | row["run"],
|
| 2385 | fmt_ms(row.get("elapsed_ms")),
|
| 2386 | row.get("tool_call_count", "unmeasured"),
|
| 2387 | row.get("raw_output_bytes", "unmeasured"),
|
| 2388 | "complete" if row.get("coverage_complete") else "partial",
|
| 2389 | "pass" if row.get("oracle_payload_on_remote_branch") else "FAIL",
|
| 2390 | )
|
| 2391 | )
|
| 2392 |
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 2393 | materialization_totals = [
|
| 2394 | row
|
| 2395 | for row in rows
|
| 2396 | if row.get("operation") == "legacy_selected_clone_materialization.total"
|
| 2397 | ]
|
| 2398 | if materialization_totals:
|
| 2399 | lines.extend(
|
| 2400 | [
|
| 2401 | "",
|
| 2402 | "## Legacy selected-branch materialization",
|
| 2403 | "",
|
| 2404 | "One fixed loopback Serve and one immutable fixture are shared across subjects.",
|
| 2405 | "Fresh destinations do not imply a cold OS page cache. Timing claims require",
|
| 2406 | "separate interleaved A/A noise and A/B campaigns.",
|
| 2407 | "",
|
| 2408 | "| Subject | Run | Clone ms | First status ms | First no-op commit ms | Passes | Harness controls met | Oracle |",
|
| 2409 | "| --- | ---: | ---: | ---: | ---: | ---: | --- | --- |",
|
| 2410 | ]
|
| 2411 | )
|
| 2412 | for row in materialization_totals:
|
| 2413 | lines.append(
|
| 2414 | "| `{}` | {} | {} | {} | {} | {} | {} | {} |".format(
|
| 2415 | row["subject"],
|
| 2416 | row["run"],
|
| 2417 | fmt_ms(row.get("clone_elapsed_ms")),
|
| 2418 | fmt_ms(row.get("first_status_elapsed_ms")),
|
| 2419 | fmt_ms(row.get("first_noop_commit_elapsed_ms")),
|
| 2420 | row.get("materialization_passes_observed", "unmeasured"),
|
| 2421 | "yes" if row.get("materialization_timing_controls_met") else "no",
|
| 2422 | "pass" if row.get("oracle_passed") else "FAIL",
|
| 2423 | )
|
| 2424 | )
|
| 2425 |
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2426 | interrupted_totals = [
|
| 2427 | row for row in rows if row.get("operation") == "interrupted_publication.total"
|
| 2428 | ]
|
| 2429 | if interrupted_totals:
|
| 2430 | lines.extend(
|
| 2431 | [
|
| 2432 | "",
|
| 2433 | "## Interrupted publication",
|
| 2434 | "",
|
| 2435 | "A loopback fixture validates the request framing and drops the response.",
|
| 2436 | "Fixture delivery knowledge is reported separately from the CLI's own outcome",
|
| 2437 | "claim. Request receipt is not durability proof; the harness never retries.",
|
| 2438 | "",
|
| 2439 | "| Subject | Run | Wall ms | Scripted call equivalents | Output bytes | CLI outcome | Fixture delivery | POSTs | Request bytes | Head/work preserved |",
|
| 2440 | "| --- | ---: | ---: | ---: | ---: | --- | --- | ---: | ---: | --- |",
|
| 2441 | ]
|
| 2442 | )
|
| 2443 | for row in interrupted_totals:
|
| 2444 | lines.append(
|
| 2445 | "| `{}` | {} | {} | {} | {} | `{}` | `{}` | {} | {} | {}/{} |".format(
|
| 2446 | row["subject"],
|
| 2447 | row["run"],
|
| 2448 | fmt_ms(row.get("elapsed_ms")),
|
| 2449 | row.get("tool_call_count", "unmeasured"),
|
| 2450 | row.get("raw_output_bytes", "unmeasured"),
|
| 2451 | row.get("publication_outcome", "unmeasured"),
|
| 2452 | row.get("fixture_delivery_state", "unmeasured"),
|
| 2453 | row.get("mutation_requests_observed", "unmeasured"),
|
| 2454 | row.get("network_request_bytes_observed", "unmeasured"),
|
| 2455 | "yes" if row.get("local_head_preserved") else "no",
|
| 2456 | "yes" if row.get("local_work_preserved") else "no",
|
| 2457 | )
|
| 2458 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2459 | return "\n".join(lines) + "\n"
|
| 2460 |
|
| 2461 |
|
| 2462 | def fmt_ms(value: Any) -> str:
|
| 2463 | try:
|
| 2464 | return f"{float(value):.0f}"
|
| 2465 | except (TypeError, ValueError):
|
| 2466 | return "unmeasured"
|
| 2467 |
|
| 2468 |
|
| 2469 | def main() -> int:
|
| 2470 | args = parse_args()
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2471 | scenarios = [item.strip() for item in args.scenarios.split(",") if item.strip()]
|
| 2472 | known_scenarios = {
|
| 2473 | TASK_LOOP_SCENARIO,
|
| 2474 | AGENT_TASK_SCENARIO,
|
| 2475 | AGENT_TASK_OAK_LOCAL_SCENARIO,
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 2476 | LEGACY_SELECTED_MATERIALIZATION_SCENARIO,
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2477 | INTERRUPTED_PUBLICATION_SCENARIO,
|
| 2478 | }
|
| 2479 | unknown_scenarios = sorted(set(scenarios) - known_scenarios)
|
| 2480 | if unknown_scenarios:
|
| 2481 | raise SystemExit(
|
| 2482 | "Unknown scenarios: " + ", ".join(unknown_scenarios)
|
| 2483 | + "; valid: " + ", ".join(sorted(known_scenarios))
|
| 2484 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2485 | subjects = load_subjects(args)
|
| 2486 | timestamp = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
|
| 2487 | args.results.mkdir(parents=True, exist_ok=True)
|
| 2488 | args.workdir.mkdir(parents=True, exist_ok=True)
|
| 2489 | run_root = args.workdir / "runs" / timestamp
|
| 2490 | run_root.mkdir(parents=True, exist_ok=True)
|
| 2491 |
|
| 2492 | metadata = {
|
| 2493 | "bench_id": timestamp,
|
| 2494 | "profile": "task-loop",
|
| 2495 | "benchmark_track": "agent-default",
|
| 2496 | "timestamp_utc": timestamp,
|
| 2497 | "host": platform.node(),
|
| 2498 | "platform": platform.platform(),
|
| 2499 | "machine": platform.machine(),
|
| 2500 | "python": platform.python_version(),
|
| 2501 | "env_isolation_version": ENV_ISOLATION_VERSION,
|
| 2502 | "subject_versions": subject_versions(subjects),
|
| 2503 | "subject_details": subject_details(subjects),
|
| 2504 | "source": source_metadata(args.oak_repo),
|
| 2505 | "git_remote_mode": args.git_remote,
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2506 | "trial_order_randomized": args.randomize_subject_order,
|
| 2507 | "trial_order_seed": args.random_seed if args.randomize_subject_order else None,
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2508 | }
|
| 2509 |
|
| 2510 | rows: list[dict[str, Any]] = []
|
| 2511 | with measurement_lock("task_loop") as lock_info:
|
| 2512 | metadata["measurement_lock_wait_ms"] = lock_info.wait_ms
|
| 2513 | metadata["measurement_lock"] = "held" if lock_info.enabled else "disabled"
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2514 | order_rng = random.Random(args.random_seed)
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 2515 | if LEGACY_SELECTED_MATERIALIZATION_SCENARIO in scenarios:
|
| 2516 | rows.extend(
|
| 2517 | run_legacy_selected_materialization_campaign(
|
| 2518 | subjects,
|
| 2519 | args.oak_serve_bin,
|
| 2520 | args.oak_serve_source,
|
| 2521 | args.oak_serve_source_head,
|
| 2522 | args.materialization_baseline_subject,
|
| 2523 | run_root,
|
| 2524 | args.runs,
|
| 2525 | metadata,
|
| 2526 | args.admitted_output_chars,
|
| 2527 | args.randomize_subject_order,
|
| 2528 | args.random_seed,
|
| 2529 | args.materialization_timing_controls_met,
|
| 2530 | )
|
| 2531 | )
|
| 2532 | ordinary_scenarios = [
|
| 2533 | scenario for scenario in scenarios if scenario != LEGACY_SELECTED_MATERIALIZATION_SCENARIO
|
| 2534 | ]
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2535 | for run_index in range(args.runs):
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2536 | run_subjects = list(subjects)
|
| 2537 | if args.randomize_subject_order:
|
| 2538 | order_rng.shuffle(run_subjects)
|
| 2539 | run_metadata = {
|
| 2540 | **metadata,
|
| 2541 | "run_subject_order": [subject.name for subject in run_subjects],
|
| 2542 | }
|
| 2543 | for subject in run_subjects:
|
a9e8fa39d0a0
Add a controlled actual-Serve benchmark lane for
13 days ago
| 2544 | for scenario in ordinary_scenarios:
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2545 | print(f"[run] {scenario} run={run_index} subject={subject.name}", flush=True)
|
| 2546 | if scenario == INTERRUPTED_PUBLICATION_SCENARIO:
|
| 2547 | rows.extend(
|
| 2548 | run_interrupted_publication(
|
| 2549 | subject, run_root, run_index, run_metadata, args.admitted_output_chars
|
| 2550 | )
|
| 2551 | )
|
| 2552 | continue
|
| 2553 | if scenario == AGENT_TASK_SCENARIO:
|
| 2554 | if subject.kind != "git":
|
| 2555 | rows.append(
|
| 2556 | unmeasured_step_row(
|
| 2557 | subject,
|
| 2558 | scenario,
|
| 2559 | "agent_task.skipped",
|
| 2560 | run_index,
|
| 2561 | run_metadata,
|
| 2562 | "Oak end-to-end publication and exact CI require an explicitly configured disposable staging lane; production mutation is forbidden",
|
| 2563 | )
|
| 2564 | )
|
| 2565 | continue
|
| 2566 | remote_create_start = time.perf_counter()
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2567 | remote = oakbench_remotes.make_git_bare_remote(
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2568 | subject.bin, run_root, f"agent-task-remote-r{run_index}.git"
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2569 | )
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2570 | remote_create_ms = (time.perf_counter() - remote_create_start) * 1000
|
| 2571 | rows.extend(
|
| 2572 | run_git_agent_task_baseline(
|
| 2573 | subject,
|
| 2574 | remote,
|
| 2575 | run_root,
|
| 2576 | run_index,
|
| 2577 | run_metadata,
|
| 2578 | args.admitted_output_chars,
|
| 2579 | remote_create_ms,
|
| 2580 | )
|
| 2581 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2582 | continue
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2583 | if scenario == AGENT_TASK_OAK_LOCAL_SCENARIO:
|
| 2584 | if subject.kind != "oak":
|
| 2585 | rows.append(
|
| 2586 | unmeasured_step_row(
|
| 2587 | subject,
|
| 2588 | scenario,
|
| 2589 | "agent_task_oak_local.skipped",
|
| 2590 | run_index,
|
| 2591 | run_metadata,
|
| 2592 | "owned local oak serve scenario applies only to Oak subjects",
|
| 2593 | )
|
| 2594 | )
|
| 2595 | continue
|
| 2596 | rows.extend(
|
| 2597 | run_oak_local_agent_task_baseline(
|
| 2598 | subject,
|
| 2599 | run_root,
|
| 2600 | run_index,
|
| 2601 | run_metadata,
|
| 2602 | args.admitted_output_chars,
|
| 2603 | )
|
| 2604 | )
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2605 | continue
|
13de5e5480b3
Measure bounded end-to-end agent task costs with
13 days ago
| 2606 | if subject.kind == "git":
|
| 2607 | if args.git_remote == "github":
|
| 2608 | remote = oakbench_remotes.resolve_git_github_remote()
|
| 2609 | else:
|
| 2610 | remote = oakbench_remotes.make_git_bare_remote(
|
| 2611 | subject.bin, run_root, f"task-loop-remote-r{run_index}.git"
|
| 2612 | )
|
| 2613 | if not remote.resolved:
|
| 2614 | rows.append(skip_row(subject, TASK_LOOP_SCENARIO, run_index, run_metadata, remote.skip_reason or ""))
|
| 2615 | continue
|
| 2616 | subject_rows = run_git_loop(subject, remote, run_root, run_index, run_metadata, args.admitted_output_chars)
|
| 2617 | else:
|
| 2618 | remote = oakbench_remotes.resolve_oak_remote("sync")
|
| 2619 | if not remote.resolved:
|
| 2620 | rows.append(skip_row(subject, TASK_LOOP_SCENARIO, run_index, run_metadata, remote.skip_reason or ""))
|
| 2621 | continue
|
| 2622 | subject_rows = run_oak_loop(subject, remote, run_root, run_index, run_metadata, args.admitted_output_chars)
|
| 2623 | total = next((row for row in subject_rows if row.get("operation") == "task_loop.total"), None)
|
| 2624 | if total is not None:
|
| 2625 | total.update(loop_summary_fields(subject_rows))
|
| 2626 | rows.extend(subject_rows)
|
88e86b3faccd
Rebuild benchmarks as a clean single-root reposi
2 months ago
| 2627 |
|
| 2628 | store = ResultsStore(args.results, lane="task-loop")
|
| 2629 | raw_path, summary_path = store.write(timestamp, rows, summary_text(rows))
|
| 2630 | if not args.keep_workdirs:
|
| 2631 | shutil.rmtree(run_root, ignore_errors=True)
|
| 2632 | print(f"[result] {raw_path}")
|
| 2633 | print(f"[summary] {summary_path}")
|
| 2634 | return 1 if has_lane_failures(rows) else 0
|
| 2635 |
|
| 2636 |
|
| 2637 | if __name__ == "__main__":
|
| 2638 | raise SystemExit(main())
|