| 1 | """Tests for the task-loop lane's derived fields and lane registration."""
|
| 2 |
|
| 3 | from __future__ import annotations
|
| 4 |
|
| 5 | import sys
|
| 6 | import os
|
| 7 | import shutil
|
| 8 | import socket
|
| 9 | import tempfile
|
| 10 | import time
|
| 11 | import unittest
|
| 12 | import http.client
|
| 13 | from pathlib import Path
|
| 14 | from unittest.mock import patch
|
| 15 |
|
| 16 | sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "scripts"))
|
| 17 |
|
| 18 | import task_loop as lane
|
| 19 | from oakbench.execution import TimedCapture
|
| 20 | from oakbench.rows import LANE_REQUIRED_KEYS
|
| 21 | from oakbench.subjects import Subject
|
| 22 | from oakbench.remotes import RemoteResolution
|
| 23 |
|
| 24 |
|
| 25 | class LoopSummaryFieldsTest(unittest.TestCase):
|
| 26 | def test_warm_delta_from_acquire_pair(self) -> None:
|
| 27 | rows = [
|
| 28 | {"operation": "task.acquire", "elapsed_ms": 1000.0},
|
| 29 | {"operation": "task.edit", "elapsed_ms": 5.0},
|
| 30 | {"operation": "followup.acquire", "elapsed_ms": 600.0},
|
| 31 | {"operation": "followup.edit", "elapsed_ms": 5.0},
|
| 32 | ]
|
| 33 | fields = lane.loop_summary_fields(rows)
|
| 34 | self.assertEqual(fields["followup_warm_delta_ms"], -400.0)
|
| 35 | self.assertEqual(fields["task_phase_ms"], 1005.0)
|
| 36 | self.assertEqual(fields["followup_phase_ms"], 605.0)
|
| 37 |
|
| 38 | def test_missing_acquire_yields_null_not_zero(self) -> None:
|
| 39 | fields = lane.loop_summary_fields([{"operation": "task.edit", "elapsed_ms": 5.0}])
|
| 40 | self.assertIsNone(fields["followup_warm_delta_ms"])
|
| 41 |
|
| 42 |
|
| 43 | class LaneRegistrationTest(unittest.TestCase):
|
| 44 | def test_task_loop_lane_registered(self) -> None:
|
| 45 | self.assertIn("task-loop", LANE_REQUIRED_KEYS)
|
| 46 | self.assertIn("command", LANE_REQUIRED_KEYS["task-loop"])
|
| 47 |
|
| 48 |
|
| 49 | class VirtualBranchReTest(unittest.TestCase):
|
| 50 | def test_parses_mount_status(self) -> None:
|
| 51 | text = "Mount: oak/bench-sync-tmp@abc (virtual branch tl-x-r0-task--9f00aa11)"
|
| 52 | match = lane.VIRTUAL_BRANCH_RE.search(text)
|
| 53 | self.assertEqual(match.group(1), "tl-x-r0-task--9f00aa11")
|
| 54 |
|
| 55 |
|
| 56 | class FailurePolicyTest(unittest.TestCase):
|
| 57 | def test_step_row_never_treats_a_process_timeout_as_an_expected_error(self) -> None:
|
| 58 | subject = Subject("oak", "oak", "Oak", Path("/usr/bin/oak"))
|
| 59 | capture = TimedCapture(
|
| 60 | elapsed_ms=100.0,
|
| 61 | returncode=124,
|
| 62 | stdout_text="",
|
| 63 | stdout_bytes=0,
|
| 64 | stdout_truncated=False,
|
| 65 | stderr_text="",
|
| 66 | stderr_bytes=0,
|
| 67 | stderr_truncated=False,
|
| 68 | timed_out=True,
|
| 69 | )
|
| 70 | with tempfile.TemporaryDirectory() as tmp, patch.object(
|
| 71 | lane, "run_timed", return_value=capture
|
| 72 | ) as timed:
|
| 73 | row = lane.step_row(
|
| 74 | subject,
|
| 75 | "interrupted_publication_after_request_v1",
|
| 76 | "publication.attempt",
|
| 77 | ["oak", "push"],
|
| 78 | Path(tmp),
|
| 79 | 0,
|
| 80 | {"bench_id": "test"},
|
| 81 | 100,
|
| 82 | expected_returncodes=tuple(range(1, 128)),
|
| 83 | timeout_seconds=0.1,
|
| 84 | )
|
| 85 |
|
| 86 | self.assertEqual(timed.call_args.kwargs["timeout_seconds"], 0.1)
|
| 87 | self.assertTrue(row["timed_out"])
|
| 88 | self.assertEqual(row["returncode"], 124)
|
| 89 |
|
| 90 | def test_step_row_marks_unexpected_zero_exit_as_failure(self) -> None:
|
| 91 | subject = Subject("git", "git", "Git", Path("/usr/bin/git"))
|
| 92 | capture = TimedCapture(
|
| 93 | elapsed_ms=1.0,
|
| 94 | returncode=0,
|
| 95 | stdout_text="",
|
| 96 | stdout_bytes=0,
|
| 97 | stdout_truncated=False,
|
| 98 | stderr_text="",
|
| 99 | stderr_bytes=0,
|
| 100 | stderr_truncated=False,
|
| 101 | )
|
| 102 | with tempfile.TemporaryDirectory() as tmp, patch.object(lane, "run_timed", return_value=capture):
|
| 103 | row = lane.step_row(
|
| 104 | subject,
|
| 105 | "task_loop",
|
| 106 | "expect.fail",
|
| 107 | ["sh", "-c", "true"],
|
| 108 | Path(tmp),
|
| 109 | 0,
|
| 110 | {
|
| 111 | "bench_id": "test",
|
| 112 | "profile": "task-loop",
|
| 113 | "timestamp_utc": "test",
|
| 114 | "benchmark_track": "agent-default",
|
| 115 | },
|
| 116 | 100,
|
| 117 | expected_returncodes=(1,),
|
| 118 | )
|
| 119 |
|
| 120 | self.assertEqual(row["process_returncode"], 0)
|
| 121 | self.assertEqual(row["returncode"], 1)
|
| 122 |
|
| 123 | def test_oracle_false_marks_total_failed(self) -> None:
|
| 124 | total = {
|
| 125 | **lane.phase_total_row(
|
| 126 | [{"operation": "task.acquire", "elapsed_ms": 1.0, "returncode": 0, "command": []}],
|
| 127 | Subject("oak", "oak", "Oak", Path("/usr/bin/oak")),
|
| 128 | "task_loop",
|
| 129 | 0,
|
| 130 | {
|
| 131 | "bench_id": "test",
|
| 132 | "profile": "task-loop",
|
| 133 | "timestamp_utc": "test",
|
| 134 | "benchmark_track": "agent-default",
|
| 135 | },
|
| 136 | "task_loop.total",
|
| 137 | ),
|
| 138 | "oracle_payload_on_remote_branch": False,
|
| 139 | }
|
| 140 | lane.enforce_oracle_result(total, False)
|
| 141 | self.assertEqual(total["returncode"], 1)
|
| 142 | self.assertTrue(lane.has_lane_failures([total]))
|
| 143 |
|
| 144 | def test_lane_failures_ignore_skips_but_fail_on_real_errors(self) -> None:
|
| 145 | self.assertFalse(lane.has_lane_failures([{"returncode": lane.SKIP_RETURNCODE, "skipped": True}]))
|
| 146 | self.assertTrue(lane.has_lane_failures([{"returncode": 1}]))
|
| 147 |
|
| 148 |
|
| 149 | class LocalGitSeedTest(unittest.TestCase):
|
| 150 | def test_seeded_local_remote_advertises_main_as_head(self) -> None:
|
| 151 | subject = Subject("git", "git", "Git", Path("/usr/bin/git"))
|
| 152 | with tempfile.TemporaryDirectory() as tmp:
|
| 153 | root = Path(tmp)
|
| 154 | remote_path = root / "remote.git"
|
| 155 | init = lane.run_untimed(
|
| 156 | [str(subject.bin), "init", "--bare", "-q", str(remote_path)], root
|
| 157 | )
|
| 158 | self.assertEqual(init.returncode, 0, init.stderr)
|
| 159 | remote = RemoteResolution(
|
| 160 | repo=str(remote_path),
|
| 161 | transport="local_file",
|
| 162 | server="local-bare",
|
| 163 | env_var="generated",
|
| 164 | )
|
| 165 |
|
| 166 | error = lane.ensure_git_remote_seeded(subject, remote, root)
|
| 167 |
|
| 168 | self.assertIsNone(error)
|
| 169 | advertised = lane.run_untimed(
|
| 170 | [str(subject.bin), "symbolic-ref", "HEAD"], remote_path
|
| 171 | )
|
| 172 | self.assertEqual(advertised.returncode, 0, advertised.stderr)
|
| 173 | self.assertEqual(advertised.stdout.strip(), "refs/heads/main")
|
| 174 |
|
| 175 |
|
| 176 | class AgentTaskEvidenceTest(unittest.TestCase):
|
| 177 | @unittest.skipUnless(
|
| 178 | all(
|
| 179 | os.environ.get(key)
|
| 180 | for key in (
|
| 181 | "OAK_BENCH_TEST_MATERIALIZATION_BASELINE",
|
| 182 | "OAK_BENCH_TEST_MATERIALIZATION_CANDIDATE",
|
| 183 | "OAK_BENCH_TEST_MATERIALIZATION_SERVE",
|
| 184 | "OAK_BENCH_TEST_MATERIALIZATION_SERVE_HEAD",
|
| 185 | )
|
| 186 | ),
|
| 187 | "set all OAK_BENCH_TEST_MATERIALIZATION_* paths for the actual Serve tracer",
|
| 188 | )
|
| 189 | def test_legacy_selected_materialization_actual_serve_tracer(self) -> None:
|
| 190 | baseline = Subject(
|
| 191 | "oak_baseline",
|
| 192 | "oak",
|
| 193 | "Oak baseline",
|
| 194 | Path(os.environ["OAK_BENCH_TEST_MATERIALIZATION_BASELINE"]),
|
| 195 | )
|
| 196 | candidate = Subject(
|
| 197 | "oak_candidate",
|
| 198 | "oak",
|
| 199 | "Oak candidate",
|
| 200 | Path(os.environ["OAK_BENCH_TEST_MATERIALIZATION_CANDIDATE"]),
|
| 201 | )
|
| 202 | with tempfile.TemporaryDirectory() as tmp:
|
| 203 | rows = lane.run_legacy_selected_materialization_campaign(
|
| 204 | [baseline, candidate],
|
| 205 | Path(os.environ["OAK_BENCH_TEST_MATERIALIZATION_SERVE"]),
|
| 206 | (
|
| 207 | Path(os.environ["OAK_BENCH_TEST_MATERIALIZATION_SERVE_SOURCE"])
|
| 208 | if os.environ.get("OAK_BENCH_TEST_MATERIALIZATION_SERVE_SOURCE")
|
| 209 | else None
|
| 210 | ),
|
| 211 | os.environ["OAK_BENCH_TEST_MATERIALIZATION_SERVE_HEAD"],
|
| 212 | "oak_baseline",
|
| 213 | Path(tmp),
|
| 214 | 1,
|
| 215 | {
|
| 216 | "bench_id": "test",
|
| 217 | "profile": "task-loop",
|
| 218 | "timestamp_utc": "test",
|
| 219 | "benchmark_track": "oak-local-loopback-materialization",
|
| 220 | },
|
| 221 | 20_000,
|
| 222 | False,
|
| 223 | 468,
|
| 224 | False,
|
| 225 | )
|
| 226 |
|
| 227 | totals = [
|
| 228 | row
|
| 229 | for row in rows
|
| 230 | if row.get("operation") == "legacy_selected_clone_materialization.total"
|
| 231 | ]
|
| 232 | self.assertEqual([row["subject"] for row in totals], ["oak_baseline", "oak_candidate"])
|
| 233 | self.assertEqual([row["materialization_passes_observed"] for row in totals], [2, 1])
|
| 234 | self.assertTrue(all(row["oracle_passed"] for row in totals))
|
| 235 | self.assertTrue(all(row["oracle_network_work_matches_baseline"] for row in totals))
|
| 236 | self.assertEqual(
|
| 237 | [row["stat_cache_metadata_valid_before_status"] for row in totals],
|
| 238 | [False, True],
|
| 239 | )
|
| 240 | self.assertTrue(all(row["materialization_timing_claimable"] is False for row in totals))
|
| 241 |
|
| 242 | def test_legacy_selected_receipt_requires_exact_two_phase_route(self) -> None:
|
| 243 | receipt = {
|
| 244 | "result": {"snapshot_bound": False, "head": "a" * 64, "manifest": "b" * 64},
|
| 245 | "observed": {
|
| 246 | "pull_phases": [
|
| 247 | {"name": "initial", "materialization_passes": 0},
|
| 248 | {"name": "legacy_selected_branch", "materialization_passes": 1},
|
| 249 | ],
|
| 250 | "materialization": {"passes_observed": 2, "written_paths": 194},
|
| 251 | },
|
| 252 | }
|
| 253 |
|
| 254 | evidence = lane.legacy_selected_receipt_evidence(receipt)
|
| 255 |
|
| 256 | self.assertTrue(evidence["legacy_route_observed"])
|
| 257 | self.assertEqual(evidence["materialization_passes_observed"], 2)
|
| 258 | self.assertEqual(evidence["head"], "a" * 64)
|
| 259 | self.assertEqual(evidence["manifest"], "b" * 64)
|
| 260 |
|
| 261 | def test_materialization_fixture_environment_drops_inherited_oak_configuration(self) -> None:
|
| 262 | with tempfile.TemporaryDirectory() as tmp, patch.dict(
|
| 263 | os.environ,
|
| 264 | {
|
| 265 | "HOME": "/original/home",
|
| 266 | "OAK_API_KEY": "secret",
|
| 267 | "OAK_REMOTE": "https://production.invalid",
|
| 268 | "GITHUB_TOKEN": "github-secret",
|
| 269 | "AWS_SECRET_ACCESS_KEY": "aws-secret",
|
| 270 | },
|
| 271 | ):
|
| 272 | env = lane.isolated_oak_fixture_env(Path(tmp) / "config")
|
| 273 |
|
| 274 | self.assertEqual(env["HOME"], "/original/home")
|
| 275 | self.assertNotIn("OAK_API_KEY", env)
|
| 276 | self.assertNotIn("OAK_REMOTE", env)
|
| 277 | self.assertNotIn("GITHUB_TOKEN", env)
|
| 278 | self.assertNotIn("AWS_SECRET_ACCESS_KEY", env)
|
| 279 | self.assertEqual(env["OAK_AUTHOR"], "oak-materialization-bench")
|
| 280 | self.assertEqual(env["OAK_NO_UPDATE_CHECK"], "1")
|
| 281 |
|
| 282 | def test_materialization_timing_requires_explicit_full_control_contract(self) -> None:
|
| 283 | held = {"measurement_lock": "held"}
|
| 284 | self.assertTrue(lane.materialization_timing_is_claimable(held, 10, True, True))
|
| 285 | self.assertFalse(lane.materialization_timing_is_claimable(held, 9, True, True))
|
| 286 | self.assertFalse(lane.materialization_timing_is_claimable(held, 10, False, True))
|
| 287 | self.assertFalse(
|
| 288 | lane.materialization_timing_is_claimable(
|
| 289 | {"measurement_lock": "disabled"}, 10, True, True
|
| 290 | )
|
| 291 | )
|
| 292 | self.assertFalse(lane.materialization_timing_is_claimable(held, 10, True, False))
|
| 293 |
|
| 294 | def test_materialization_campaign_refuses_relative_or_nonexecutable_serve(self) -> None:
|
| 295 | subject = Subject("control", "oak", "Control", Path(sys.executable))
|
| 296 | metadata = {"measurement_lock": "held"}
|
| 297 | with tempfile.TemporaryDirectory() as tmp:
|
| 298 | root = Path(tmp)
|
| 299 | relative = lane.run_legacy_selected_materialization_campaign(
|
| 300 | [subject],
|
| 301 | Path("oak"),
|
| 302 | None,
|
| 303 | "a" * 64,
|
| 304 | "control",
|
| 305 | root,
|
| 306 | 1,
|
| 307 | metadata,
|
| 308 | 100,
|
| 309 | False,
|
| 310 | 1,
|
| 311 | False,
|
| 312 | )
|
| 313 | nonexecutable = root / "serve"
|
| 314 | nonexecutable.write_text("not executable")
|
| 315 | refused = lane.run_legacy_selected_materialization_campaign(
|
| 316 | [subject],
|
| 317 | nonexecutable,
|
| 318 | None,
|
| 319 | "a" * 64,
|
| 320 | "control",
|
| 321 | root,
|
| 322 | 1,
|
| 323 | metadata,
|
| 324 | 100,
|
| 325 | False,
|
| 326 | 1,
|
| 327 | False,
|
| 328 | )
|
| 329 |
|
| 330 | self.assertEqual(relative[0]["returncode"], lane.SKIP_RETURNCODE)
|
| 331 | self.assertEqual(refused[0]["returncode"], lane.SKIP_RETURNCODE)
|
| 332 |
|
| 333 | def test_materialization_campaign_requires_named_control(self) -> None:
|
| 334 | subject = Subject("candidate", "oak", "Candidate", Path(sys.executable))
|
| 335 | with tempfile.TemporaryDirectory() as tmp:
|
| 336 | rows = lane.run_legacy_selected_materialization_campaign(
|
| 337 | [subject],
|
| 338 | Path(sys.executable),
|
| 339 | None,
|
| 340 | "a" * 64,
|
| 341 | "missing",
|
| 342 | Path(tmp),
|
| 343 | 1,
|
| 344 | {"measurement_lock": "held"},
|
| 345 | 100,
|
| 346 | False,
|
| 347 | 1,
|
| 348 | False,
|
| 349 | )
|
| 350 |
|
| 351 | self.assertEqual(rows[0]["returncode"], lane.SKIP_RETURNCODE)
|
| 352 | self.assertIn("baseline-subject", rows[0]["skip_reason"])
|
| 353 |
|
| 354 | def test_normalized_worktree_snapshot_records_bytes_modes_and_ignores_vcs_state(self) -> None:
|
| 355 | with tempfile.TemporaryDirectory() as tmp:
|
| 356 | root = Path(tmp)
|
| 357 | (root / ".oak").mkdir()
|
| 358 | (root / ".oak/oak.db").write_bytes(b"private")
|
| 359 | (root / "nested").mkdir()
|
| 360 | file_path = root / "nested/tool.sh"
|
| 361 | file_path.write_bytes(b"#!/bin/sh\n")
|
| 362 | file_path.chmod(0o755)
|
| 363 |
|
| 364 | rows = lane.normalized_worktree_snapshot(root)
|
| 365 |
|
| 366 | self.assertEqual(len(rows), 1)
|
| 367 | self.assertEqual(rows[0]["path"], "nested/tool.sh")
|
| 368 | self.assertEqual(rows[0]["mode"], 0o755)
|
| 369 | self.assertEqual(rows[0]["kind"], "file")
|
| 370 | self.assertEqual(rows[0]["sha256"], "a8076d3d28d21e02012b20eaf7dbf75409a6277134439025f282e368e3305abf")
|
| 371 |
|
| 372 | @unittest.skipUnless(
|
| 373 | os.environ.get("OAK_BENCH_TEST_OAK_BIN"),
|
| 374 | "set OAK_BENCH_TEST_OAK_BIN to exercise the real local Oak server",
|
| 375 | )
|
| 376 | def test_local_oak_agent_task_uses_fresh_checkouts_and_remote_oracles(self) -> None:
|
| 377 | oak_bin = Path(os.environ["OAK_BENCH_TEST_OAK_BIN"])
|
| 378 | subject = Subject("oak_local", "oak", "Oak local", oak_bin)
|
| 379 | with tempfile.TemporaryDirectory() as tmp:
|
| 380 | rows = lane.run_oak_local_agent_task_baseline(
|
| 381 | subject,
|
| 382 | Path(tmp),
|
| 383 | 0,
|
| 384 | {
|
| 385 | "bench_id": "test",
|
| 386 | "profile": "task-loop",
|
| 387 | "timestamp_utc": "test",
|
| 388 | "benchmark_track": "agent-default",
|
| 389 | },
|
| 390 | 20_000,
|
| 391 | )
|
| 392 |
|
| 393 | total = next(row for row in rows if row["operation"] == "agent_task_oak_local.total")
|
| 394 | self.assertEqual(total["returncode"], 0)
|
| 395 | self.assertFalse(total["coverage_complete"])
|
| 396 | self.assertTrue(total["oracle_payload_on_remote_branch"])
|
| 397 | self.assertTrue(total["oracle_branch_head_matches"])
|
| 398 | self.assertTrue(total["server_process_stopped"])
|
| 399 | self.assertEqual(
|
| 400 | total["acquisition_models"],
|
| 401 | {"cold": "fresh_full_clone", "followup": "second_fresh_full_clone"},
|
| 402 | )
|
| 403 | self.assertEqual(total["repository_cache_state"]["followup"], "no_shared_object_store")
|
| 404 | self.assertIn("cold.ci.observe", total["unmeasured_operations"])
|
| 405 | self.assertIn("followup.review_handoff", total["unmeasured_operations"])
|
| 406 |
|
| 407 | def test_total_keeps_skips_unmeasured_and_aggregates_costs(self) -> None:
|
| 408 | subject = Subject("git", "git", "Git", Path("/usr/bin/git"))
|
| 409 | rows = [
|
| 410 | {
|
| 411 | "operation": "task.acquire",
|
| 412 | "elapsed_ms": 10.0,
|
| 413 | "returncode": 0,
|
| 414 | "command": ["git", "clone"],
|
| 415 | "tool_call_count": 1,
|
| 416 | "stdout_bytes": 7,
|
| 417 | "stderr_bytes": 0,
|
| 418 | "retry_count": 0,
|
| 419 | "human_action_count": 0,
|
| 420 | },
|
| 421 | {
|
| 422 | "operation": "task.ci.observe",
|
| 423 | "elapsed_ms": 0.0,
|
| 424 | "returncode": lane.SKIP_RETURNCODE,
|
| 425 | "command": [],
|
| 426 | "skipped": True,
|
| 427 | "skip_reason": "local transport has no hosted CI provider",
|
| 428 | "tool_call_count": 0,
|
| 429 | "stdout_bytes": None,
|
| 430 | "stderr_bytes": None,
|
| 431 | "retry_count": 0,
|
| 432 | "human_action_count": 0,
|
| 433 | },
|
| 434 | ]
|
| 435 |
|
| 436 | total = lane.phase_total_row(
|
| 437 | rows,
|
| 438 | subject,
|
| 439 | "agent_task_baseline_v1",
|
| 440 | 0,
|
| 441 | {
|
| 442 | "bench_id": "test",
|
| 443 | "profile": "task-loop",
|
| 444 | "timestamp_utc": "test",
|
| 445 | "benchmark_track": "agent-default",
|
| 446 | },
|
| 447 | "agent_task.total",
|
| 448 | )
|
| 449 |
|
| 450 | self.assertEqual(total["returncode"], 0)
|
| 451 | self.assertFalse(total["coverage_complete"])
|
| 452 | self.assertEqual(total["unmeasured_operations"], ["task.ci.observe"])
|
| 453 | self.assertEqual(total["raw_output_bytes"], 7)
|
| 454 | self.assertEqual(total["retry_count"], 0)
|
| 455 | self.assertEqual(total["human_action_count"], 0)
|
| 456 | self.assertEqual(total["phase_elapsed_ms"], {"task": 10.0})
|
| 457 |
|
| 458 | def test_interrupted_publication_summary_never_claims_retry(self) -> None:
|
| 459 | fields = lane.interrupted_publication_fields(
|
| 460 | request_count=1,
|
| 461 | request_bytes=4096,
|
| 462 | declared_request_bytes=4096,
|
| 463 | request_framing_valid=True,
|
| 464 | request_body_timed_out=False,
|
| 465 | local_head_before="a" * 64,
|
| 466 | local_head_after="a" * 64,
|
| 467 | status_after={"needs_push": True, "dirty": False},
|
| 468 | worktree_payload_before="payload-hash",
|
| 469 | worktree_payload_after="payload-hash",
|
| 470 | cli_observation={"reported_outcome": "unknown", "observation": "json_error_unknown"},
|
| 471 | )
|
| 472 |
|
| 473 | self.assertEqual(fields["publication_outcome"], "unknown")
|
| 474 | self.assertEqual(fields["mutation_requests_observed"], 1)
|
| 475 | self.assertEqual(fields["blind_replay_requests"], 0)
|
| 476 | self.assertTrue(fields["local_head_preserved"])
|
| 477 | self.assertTrue(fields["local_work_preserved"])
|
| 478 | self.assertTrue(fields["worktree_payload_preserved"])
|
| 479 | self.assertEqual(fields["network_request_bytes_observed"], 4096)
|
| 480 |
|
| 481 | def test_interruption_separates_cli_report_from_fixture_delivery_knowledge(self) -> None:
|
| 482 | generic_error = {
|
| 483 | "stdout": '{"schema_version":1,"error":{"code":"http","message":"connection closed"}}',
|
| 484 | "stderr": "",
|
| 485 | "process_returncode": 6,
|
| 486 | "timed_out": False,
|
| 487 | }
|
| 488 | cli = lane.cli_publication_observation(generic_error, json_requested=True)
|
| 489 | self.assertEqual(cli["reported_outcome"], "unreported")
|
| 490 | self.assertEqual(cli["observation"], "json_error_without_publication_outcome")
|
| 491 |
|
| 492 | fields = lane.interrupted_publication_fields(
|
| 493 | request_count=1,
|
| 494 | request_bytes=3,
|
| 495 | declared_request_bytes=100,
|
| 496 | request_framing_valid=True,
|
| 497 | request_body_timed_out=False,
|
| 498 | local_head_before="a" * 64,
|
| 499 | local_head_after="a" * 64,
|
| 500 | status_after={"needs_push": True, "dirty": False},
|
| 501 | worktree_payload_before="payload",
|
| 502 | worktree_payload_after="payload",
|
| 503 | cli_observation=cli,
|
| 504 | )
|
| 505 |
|
| 506 | self.assertEqual(fields["publication_outcome"], "unreported")
|
| 507 | self.assertEqual(fields["cli_observation"], "json_error_without_publication_outcome")
|
| 508 | self.assertEqual(fields["fixture_delivery_state"], "partial_request_observed")
|
| 509 | self.assertEqual(
|
| 510 | fields["fixture_outcome_knowledge"],
|
| 511 | "indeterminate_after_partial_request_disconnect",
|
| 512 | )
|
| 513 |
|
| 514 | def test_interruption_requires_exact_payload_and_unpublished_state(self) -> None:
|
| 515 | fields = lane.interrupted_publication_fields(
|
| 516 | request_count=1,
|
| 517 | request_bytes=100,
|
| 518 | declared_request_bytes=100,
|
| 519 | request_framing_valid=True,
|
| 520 | request_body_timed_out=False,
|
| 521 | local_head_before="a" * 64,
|
| 522 | local_head_after="a" * 64,
|
| 523 | status_after={"needs_push": True, "dirty": False},
|
| 524 | worktree_payload_before="before-hash",
|
| 525 | worktree_payload_after="after-hash",
|
| 526 | cli_observation={"reported_outcome": "unknown", "observation": "json_error_unknown"},
|
| 527 | )
|
| 528 |
|
| 529 | self.assertTrue(fields["unpublished_state_preserved"])
|
| 530 | self.assertFalse(fields["worktree_payload_preserved"])
|
| 531 | self.assertFalse(fields["local_work_preserved"])
|
| 532 |
|
| 533 | def test_interruption_server_reads_one_complete_push_then_drops_reply(self) -> None:
|
| 534 | server, thread = lane.start_interrupted_publication_server()
|
| 535 | host, port = server.server_address[:2]
|
| 536 | try:
|
| 537 | connection = http.client.HTTPConnection(host, port, timeout=2)
|
| 538 | connection.request("POST", "/api/repos", body=b"{}", headers={"Content-Type": "application/json"})
|
| 539 | response = connection.getresponse()
|
| 540 | self.assertEqual(response.status, 200)
|
| 541 | response.read()
|
| 542 | connection.close()
|
| 543 |
|
| 544 | connection = http.client.HTTPConnection(host, port, timeout=2)
|
| 545 | connection.request(
|
| 546 | "POST",
|
| 547 | "/api/oak/repo/push",
|
| 548 | body=b"fixture-payload",
|
| 549 | headers={"Content-Type": "application/json"},
|
| 550 | )
|
| 551 | with self.assertRaises(http.client.RemoteDisconnected):
|
| 552 | connection.getresponse()
|
| 553 | connection.close()
|
| 554 | finally:
|
| 555 | server.shutdown()
|
| 556 | server.server_close()
|
| 557 | thread.join(timeout=2)
|
| 558 |
|
| 559 | self.assertEqual(server.push_request_count, 1)
|
| 560 | self.assertEqual(server.push_request_bytes, len(b"fixture-payload"))
|
| 561 | self.assertTrue(server.request_complete_before_disconnect)
|
| 562 |
|
| 563 | def test_interruption_server_never_certifies_a_truncated_request(self) -> None:
|
| 564 | server, thread = lane.start_interrupted_publication_server()
|
| 565 | host, port = server.server_address[:2]
|
| 566 | try:
|
| 567 | with socket.create_connection((host, port), timeout=2) as connection:
|
| 568 | connection.sendall(
|
| 569 | b"POST /api/oak/repo/push HTTP/1.1\r\n"
|
| 570 | b"Host: local\r\n"
|
| 571 | b"Content-Length: 100\r\n"
|
| 572 | b"Content-Type: application/json\r\n\r\n"
|
| 573 | b"abc"
|
| 574 | )
|
| 575 | connection.shutdown(socket.SHUT_WR)
|
| 576 | connection.recv(1024)
|
| 577 | finally:
|
| 578 | server.shutdown()
|
| 579 | server.server_close()
|
| 580 | thread.join(timeout=2)
|
| 581 |
|
| 582 | self.assertEqual(server.push_request_count, 1)
|
| 583 | self.assertEqual(server.push_request_bytes, 3)
|
| 584 | self.assertEqual(server.push_declared_request_bytes, 100)
|
| 585 | self.assertTrue(server.push_request_framing_valid)
|
| 586 | self.assertFalse(server.request_complete_before_disconnect)
|
| 587 |
|
| 588 | def test_interruption_server_rejects_absent_and_invalid_framing(self) -> None:
|
| 589 | for framing in (
|
| 590 | b"",
|
| 591 | b"Content-Length: nope\r\n",
|
| 592 | b"Content-Length: +3\r\n",
|
| 593 | b"Content-Length: 3\r\nContent-Length: 3\r\n",
|
| 594 | f"Content-Length: {lane.INTERRUPTED_PUBLICATION_MAX_REQUEST_BYTES + 1}\r\n".encode(),
|
| 595 | ):
|
| 596 | with self.subTest(framing=framing):
|
| 597 | server, thread = lane.start_interrupted_publication_server()
|
| 598 | host, port = server.server_address[:2]
|
| 599 | try:
|
| 600 | with socket.create_connection((host, port), timeout=2) as connection:
|
| 601 | connection.sendall(
|
| 602 | b"POST /api/oak/repo/push HTTP/1.1\r\n"
|
| 603 | b"Host: local\r\n"
|
| 604 | + framing
|
| 605 | + b"Content-Type: application/json\r\n\r\nabc"
|
| 606 | )
|
| 607 | connection.shutdown(socket.SHUT_WR)
|
| 608 | connection.recv(1024)
|
| 609 | finally:
|
| 610 | server.shutdown()
|
| 611 | server.server_close()
|
| 612 | thread.join(timeout=2)
|
| 613 |
|
| 614 | self.assertEqual(server.push_request_count, 1)
|
| 615 | self.assertIsNone(server.push_declared_request_bytes)
|
| 616 | self.assertFalse(server.push_request_framing_valid)
|
| 617 | self.assertFalse(server.request_complete_before_disconnect)
|
| 618 |
|
| 619 | def test_interruption_server_accepts_http_ows_around_content_length(self) -> None:
|
| 620 | server, thread = lane.start_interrupted_publication_server()
|
| 621 | host, port = server.server_address[:2]
|
| 622 | try:
|
| 623 | with socket.create_connection((host, port), timeout=2) as connection:
|
| 624 | connection.sendall(
|
| 625 | b"POST /api/oak/repo/push HTTP/1.1\r\n"
|
| 626 | b"Host: local\r\n"
|
| 627 | b"Content-Length:\t3 \t\r\n\r\nabc"
|
| 628 | )
|
| 629 | connection.shutdown(socket.SHUT_WR)
|
| 630 | connection.recv(1024)
|
| 631 | finally:
|
| 632 | server.shutdown()
|
| 633 | server.server_close()
|
| 634 | thread.join(timeout=2)
|
| 635 |
|
| 636 | self.assertEqual(server.push_declared_request_bytes, 3)
|
| 637 | self.assertTrue(server.push_request_framing_valid)
|
| 638 | self.assertTrue(server.request_complete_before_disconnect)
|
| 639 |
|
| 640 | def test_interruption_server_bounds_a_partial_request_while_sender_stays_open(self) -> None:
|
| 641 | server, thread = lane.start_interrupted_publication_server()
|
| 642 | host, port = server.server_address[:2]
|
| 643 | started = time.monotonic()
|
| 644 | try:
|
| 645 | with socket.create_connection((host, port), timeout=4) as connection:
|
| 646 | connection.sendall(
|
| 647 | b"POST /api/oak/repo/push HTTP/1.1\r\n"
|
| 648 | b"Host: local\r\n"
|
| 649 | b"Content-Length: 100\r\n\r\nabc"
|
| 650 | )
|
| 651 | connection.recv(1024)
|
| 652 | finally:
|
| 653 | server.shutdown()
|
| 654 | server.server_close()
|
| 655 | thread.join(timeout=2)
|
| 656 |
|
| 657 | self.assertLess(time.monotonic() - started, 4)
|
| 658 | self.assertEqual(server.push_request_bytes, 3)
|
| 659 | self.assertTrue(server.push_request_framing_valid)
|
| 660 | self.assertTrue(server.push_request_body_timed_out)
|
| 661 | self.assertFalse(server.request_complete_before_disconnect)
|
| 662 |
|
| 663 | @unittest.skipIf(shutil.which("git") is None, "git not installed")
|
| 664 | def test_local_git_agent_task_records_cold_warm_and_unmeasured_ci(self) -> None:
|
| 665 | subject = Subject("git", "git", "Git", Path(shutil.which("git") or "git"))
|
| 666 | with tempfile.TemporaryDirectory() as tmp:
|
| 667 | root = Path(tmp)
|
| 668 | remote = lane.oakbench_remotes.make_git_bare_remote(
|
| 669 | subject.bin, root, "agent-task.git"
|
| 670 | )
|
| 671 | rows = lane.run_git_agent_task_baseline(
|
| 672 | subject,
|
| 673 | remote,
|
| 674 | root,
|
| 675 | 0,
|
| 676 | {
|
| 677 | "bench_id": "test",
|
| 678 | "profile": "task-loop",
|
| 679 | "timestamp_utc": "test",
|
| 680 | "benchmark_track": "agent-default",
|
| 681 | },
|
| 682 | 20_000,
|
| 683 | )
|
| 684 |
|
| 685 | operations = [row["operation"] for row in rows]
|
| 686 | self.assertIn("cold.acquire", operations)
|
| 687 | self.assertIn("warm.acquire", operations)
|
| 688 | self.assertIn("cold.ci.observe", operations)
|
| 689 | total = next(row for row in rows if row["operation"] == "agent_task.total")
|
| 690 | self.assertEqual(total["returncode"], 0)
|
| 691 | self.assertFalse(total["coverage_complete"])
|
| 692 | self.assertEqual(total["oracle_payload_on_remote_branch"], True)
|
| 693 | self.assertEqual(total["acquisition_models"], {
|
| 694 | "cold": "fresh_full_clone",
|
| 695 | "warm": "shared_object_worktree",
|
| 696 | })
|
| 697 | ci = next(row for row in rows if row["operation"] == "cold.ci.observe")
|
| 698 | self.assertEqual(ci["returncode"], lane.SKIP_RETURNCODE)
|
| 699 | self.assertIn("no hosted CI", ci["skip_reason"])
|
| 700 |
|
| 701 |
|
| 702 | if __name__ == "__main__":
|
| 703 | unittest.main()
|