test_cmd_bisect_hardening.py
python
sha256:660fcac1df3ab28f61862e961890bd2ca8b754fa0242079d93ca1e25037ec8a6
chore(tests): add docstring to tests/__init__.py so rc14 tr…
Human
26 days ago
| 1 | """Comprehensive hardening tests for ``muse bisect``. |
| 2 | |
| 3 | Covers: |
| 4 | - Unit: _toml_escape, _load_state symlink guard, size cap, _save_state injection |
| 5 | - Security: branch TOML injection, symlink state file, oversized state, ANSI |
| 6 | sanitization, error routing to stderr, null bytes in refs |
| 7 | - JSON schema: all subcommands (start, bad, good, skip, log, reset, run) |
| 8 | - Integration: --json round-trips, get_bisect_next public API, session lifecycle |
| 9 | - E2E: symbol-scoped bisect, run subcommand NDJSON, reset --json, log --json |
| 10 | - Stress: 200-commit chain, concurrent read-only queries |
| 11 | """ |
| 12 | from __future__ import annotations |
| 13 | |
| 14 | import datetime |
| 15 | import json |
| 16 | import pathlib |
| 17 | import re |
| 18 | import threading |
| 19 | import uuid |
| 20 | from typing import TypedDict |
| 21 | |
| 22 | import pytest |
| 23 | |
| 24 | from muse.core.snapshot import compute_commit_id, compute_snapshot_id |
| 25 | from muse.core.store import CommitRecord, SnapshotRecord, write_commit, write_snapshot |
| 26 | from muse.core._types import Manifest |
| 27 | from tests.cli_test_helper import CliRunner, InvokeResult |
| 28 | |
| 29 | # Helpers to check store field names at import time; mypy will catch mismatches. |
| 30 | _SNAP_FIELDS: set[str] = {"snapshot_id", "manifest", "created_at"} |
| 31 | _COMMIT_FIELDS: set[str] = {"commit_id", "repo_id", "branch", "snapshot_id", "message", "committed_at"} |
| 32 | |
| 33 | runner = CliRunner() |
| 34 | |
| 35 | _ANSI_RE = re.compile(r"\x1b\[[0-9;]*m") |
| 36 | |
| 37 | |
| 38 | # --------------------------------------------------------------------------- |
| 39 | # Fixtures |
| 40 | # --------------------------------------------------------------------------- |
| 41 | |
| 42 | |
| 43 | def _make_repo(tmp_path: pathlib.Path) -> tuple[pathlib.Path, str]: |
| 44 | """Create a minimal Muse repo layout without calling muse init. |
| 45 | |
| 46 | Returns (repo_root, repo_id). |
| 47 | """ |
| 48 | repo_id = str(uuid.uuid4()) |
| 49 | muse = tmp_path / ".muse" |
| 50 | muse.mkdir() |
| 51 | (muse / "repo.json").write_text( |
| 52 | json.dumps({ |
| 53 | "repo_id": repo_id, |
| 54 | "domain": "code", |
| 55 | "default_branch": "main", |
| 56 | "created_at": "2026-01-01T00:00:00+00:00", |
| 57 | }) |
| 58 | ) |
| 59 | (muse / "HEAD").write_text("ref: refs/heads/main") |
| 60 | (muse / "refs" / "heads").mkdir(parents=True) |
| 61 | (muse / "snapshots").mkdir() |
| 62 | (muse / "commits").mkdir() |
| 63 | (muse / "objects").mkdir() |
| 64 | return tmp_path, repo_id |
| 65 | |
| 66 | |
| 67 | def _make_commit( |
| 68 | root: pathlib.Path, |
| 69 | repo_id: str, |
| 70 | *, |
| 71 | branch: str = "main", |
| 72 | message: str = "commit", |
| 73 | parent_id: str | None = None, |
| 74 | ) -> str: |
| 75 | """Write a synthetic commit and return its commit_id.""" |
| 76 | manifest: Manifest = {} |
| 77 | snap_id = compute_snapshot_id(manifest) |
| 78 | committed_at = datetime.datetime.now(datetime.timezone.utc) |
| 79 | commit_id = compute_commit_id( |
| 80 | parent_ids=[parent_id] if parent_id else [], |
| 81 | snapshot_id=snap_id, |
| 82 | message=message, |
| 83 | committed_at_iso=committed_at.isoformat(), |
| 84 | ) |
| 85 | snap = SnapshotRecord( |
| 86 | snapshot_id=snap_id, |
| 87 | manifest={}, |
| 88 | created_at=committed_at, |
| 89 | ) |
| 90 | write_snapshot(root, snap) |
| 91 | commit = CommitRecord( |
| 92 | commit_id=commit_id, |
| 93 | repo_id=repo_id, |
| 94 | parent_commit_id=parent_id, |
| 95 | parent2_commit_id=None, |
| 96 | snapshot_id=snap_id, |
| 97 | branch=branch, |
| 98 | message=message, |
| 99 | committed_at=committed_at, |
| 100 | ) |
| 101 | write_commit(root, commit) |
| 102 | ref_path = root / ".muse" / "refs" / "heads" / branch |
| 103 | ref_path.write_text(commit_id) |
| 104 | (root / ".muse" / "HEAD").write_text(f"ref: refs/heads/{branch}") |
| 105 | return commit_id |
| 106 | |
| 107 | |
| 108 | def _build_chain(root: pathlib.Path, repo_id: str, n: int) -> list[str]: |
| 109 | """Create n commits (linear chain) and return their IDs oldest-first.""" |
| 110 | ids: list[str] = [] |
| 111 | parent: str | None = None |
| 112 | for i in range(n): |
| 113 | cid = _make_commit(root, repo_id, message=f"commit {i}", parent_id=parent) |
| 114 | ids.append(cid) |
| 115 | parent = cid |
| 116 | return ids |
| 117 | |
| 118 | |
| 119 | def _invoke(root: pathlib.Path, args: list[str]) -> InvokeResult: |
| 120 | return runner.invoke(None, args, env={"MUSE_REPO_ROOT": str(root)}) |
| 121 | |
| 122 | |
| 123 | def _json_blob(output: str) -> str: |
| 124 | """Extract the first JSON object/array from mixed output.""" |
| 125 | for line in output.splitlines(): |
| 126 | stripped = line.strip() |
| 127 | if stripped.startswith("{") or stripped.startswith("["): |
| 128 | return stripped |
| 129 | return output.strip() |
| 130 | |
| 131 | |
| 132 | # --------------------------------------------------------------------------- |
| 133 | # Typed schema helpers |
| 134 | # --------------------------------------------------------------------------- |
| 135 | |
| 136 | |
| 137 | class _StepJson(TypedDict): |
| 138 | done: bool |
| 139 | first_bad: str | None |
| 140 | next_to_test: str | None |
| 141 | remaining_count: int |
| 142 | steps_remaining: int |
| 143 | verdict: str |
| 144 | symbol_changes: list[str] |
| 145 | |
| 146 | |
| 147 | class _LogJson(TypedDict): |
| 148 | active: bool |
| 149 | entries: list[str] |
| 150 | |
| 151 | |
| 152 | class _ResetJson(TypedDict): |
| 153 | reset: bool |
| 154 | |
| 155 | |
| 156 | class _RunStepJson(TypedDict): |
| 157 | step: int |
| 158 | testing: str |
| 159 | verdict: str |
| 160 | remaining_count: int |
| 161 | done: bool |
| 162 | |
| 163 | |
| 164 | class _RunDoneJson(TypedDict): |
| 165 | done: bool |
| 166 | first_bad: str | None |
| 167 | steps_taken: int |
| 168 | |
| 169 | |
| 170 | def _repo(tmp_path: pathlib.Path) -> tuple[pathlib.Path, str]: |
| 171 | """Alias for _make_repo for readability inside test methods.""" |
| 172 | return _make_repo(tmp_path) |
| 173 | |
| 174 | |
| 175 | def _parse_step(output: str) -> _StepJson: |
| 176 | raw = json.loads(_json_blob(output)) |
| 177 | assert isinstance(raw, dict) |
| 178 | done_val = raw["done"] |
| 179 | first_bad_val = raw["first_bad"] |
| 180 | next_to_test_val = raw["next_to_test"] |
| 181 | remaining_count_val = raw["remaining_count"] |
| 182 | steps_remaining_val = raw["steps_remaining"] |
| 183 | verdict_val = raw["verdict"] |
| 184 | symbol_changes_val = raw["symbol_changes"] |
| 185 | assert isinstance(done_val, bool) |
| 186 | assert first_bad_val is None or isinstance(first_bad_val, str) |
| 187 | assert next_to_test_val is None or isinstance(next_to_test_val, str) |
| 188 | assert isinstance(remaining_count_val, int) |
| 189 | assert isinstance(steps_remaining_val, int) |
| 190 | assert isinstance(verdict_val, str) |
| 191 | assert isinstance(symbol_changes_val, list) |
| 192 | return _StepJson( |
| 193 | done=done_val, |
| 194 | first_bad=first_bad_val, |
| 195 | next_to_test=next_to_test_val, |
| 196 | remaining_count=remaining_count_val, |
| 197 | steps_remaining=steps_remaining_val, |
| 198 | verdict=verdict_val, |
| 199 | symbol_changes=symbol_changes_val, |
| 200 | ) |
| 201 | |
| 202 | |
| 203 | def _parse_log(output: str) -> _LogJson: |
| 204 | raw = json.loads(_json_blob(output)) |
| 205 | assert isinstance(raw, dict) |
| 206 | active_val = raw["active"] |
| 207 | entries_val = raw["entries"] |
| 208 | assert isinstance(active_val, bool) |
| 209 | assert isinstance(entries_val, list) |
| 210 | return _LogJson(active=active_val, entries=entries_val) |
| 211 | |
| 212 | |
| 213 | def _parse_reset(output: str) -> _ResetJson: |
| 214 | raw = json.loads(_json_blob(output)) |
| 215 | assert isinstance(raw, dict) |
| 216 | reset_val = raw["reset"] |
| 217 | assert isinstance(reset_val, bool) |
| 218 | return _ResetJson(reset=reset_val) |
| 219 | |
| 220 | |
| 221 | # --------------------------------------------------------------------------- |
| 222 | # Unit — _toml_escape |
| 223 | # --------------------------------------------------------------------------- |
| 224 | |
| 225 | |
| 226 | class TestTomlEscape: |
| 227 | def test_plain_string_unchanged(self) -> None: |
| 228 | from muse.core.bisect import _toml_escape |
| 229 | |
| 230 | assert _toml_escape("feat/my-thing") == "feat/my-thing" |
| 231 | |
| 232 | def test_double_quote_escaped(self) -> None: |
| 233 | from muse.core.bisect import _toml_escape |
| 234 | |
| 235 | result = _toml_escape('branch"with"quotes') |
| 236 | # After escaping, no bare double-quotes remain (only \"). |
| 237 | assert '\\"' in result |
| 238 | |
| 239 | def test_backslash_escaped(self) -> None: |
| 240 | from muse.core.bisect import _toml_escape |
| 241 | |
| 242 | result = _toml_escape("branch\\with\\backslash") |
| 243 | assert result == "branch\\\\with\\\\backslash" |
| 244 | |
| 245 | def test_both_escaped(self) -> None: |
| 246 | from muse.core.bisect import _toml_escape |
| 247 | |
| 248 | result = _toml_escape('evil"; bad_id = "hacked') |
| 249 | assert '\\"' in result |
| 250 | assert "bad_id" in result # literal text preserved, just escaped |
| 251 | |
| 252 | |
| 253 | # --------------------------------------------------------------------------- |
| 254 | # Unit — _load_state security |
| 255 | # --------------------------------------------------------------------------- |
| 256 | |
| 257 | |
| 258 | class TestLoadStateSecurity: |
| 259 | def test_symlink_state_file_rejected(self, tmp_path: pathlib.Path) -> None: |
| 260 | """A symlink at the bisect state path must be silently ignored.""" |
| 261 | from muse.core.bisect import _load_state, _state_path |
| 262 | |
| 263 | root, _ = _make_repo(tmp_path) |
| 264 | target = tmp_path / "real_state.toml" |
| 265 | target.write_text('bad_id = "abc"\ngood_ids = []\nskipped_ids = []\nremaining = []\nlog = []\n') |
| 266 | state_path = _state_path(root) |
| 267 | state_path.symlink_to(target) |
| 268 | result = _load_state(root) |
| 269 | assert result is None |
| 270 | |
| 271 | def test_oversized_state_file_rejected(self, tmp_path: pathlib.Path) -> None: |
| 272 | """State files exceeding _MAX_STATE_BYTES must be rejected.""" |
| 273 | from muse.core.bisect import _MAX_STATE_BYTES, _load_state, _state_path |
| 274 | |
| 275 | root, _ = _make_repo(tmp_path) |
| 276 | state_path = _state_path(root) |
| 277 | huge = "x" * (_MAX_STATE_BYTES + 1) |
| 278 | state_path.write_text(huge) |
| 279 | result = _load_state(root) |
| 280 | assert result is None |
| 281 | |
| 282 | def test_corrupt_state_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 283 | from muse.core.bisect import _load_state, _state_path |
| 284 | |
| 285 | root, _ = _make_repo(tmp_path) |
| 286 | state_path = _state_path(root) |
| 287 | state_path.write_text("not valid toml ]] [[[ !!!") |
| 288 | result = _load_state(root) |
| 289 | assert result is None |
| 290 | |
| 291 | def test_missing_state_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 292 | from muse.core.bisect import _load_state |
| 293 | |
| 294 | root, _ = _make_repo(tmp_path) |
| 295 | result = _load_state(root) |
| 296 | assert result is None |
| 297 | |
| 298 | |
| 299 | # --------------------------------------------------------------------------- |
| 300 | # Unit — _save_state TOML injection |
| 301 | # --------------------------------------------------------------------------- |
| 302 | |
| 303 | |
| 304 | class TestSaveStateTomlInjection: |
| 305 | def test_branch_with_quote_survives_roundtrip(self, tmp_path: pathlib.Path) -> None: |
| 306 | """A branch name containing a double-quote must not corrupt the state file.""" |
| 307 | from muse.core.bisect import BisectStateDict, _load_state, _save_state |
| 308 | |
| 309 | root, _ = _make_repo(tmp_path) |
| 310 | state: BisectStateDict = { |
| 311 | "bad_id": "a" * 64, |
| 312 | "good_ids": ["b" * 64], |
| 313 | "skipped_ids": [], |
| 314 | "remaining": [], |
| 315 | "log": [], |
| 316 | "branch": 'evil"; bad_id = "injected', |
| 317 | } |
| 318 | _save_state(root, state) |
| 319 | loaded = _load_state(root) |
| 320 | assert loaded is not None |
| 321 | assert loaded.get("bad_id") == "a" * 64 |
| 322 | assert loaded.get("branch") == 'evil"; bad_id = "injected' |
| 323 | |
| 324 | def test_branch_with_backslash_survives_roundtrip(self, tmp_path: pathlib.Path) -> None: |
| 325 | from muse.core.bisect import BisectStateDict, _load_state, _save_state |
| 326 | |
| 327 | root, _ = _make_repo(tmp_path) |
| 328 | state: BisectStateDict = { |
| 329 | "bad_id": "c" * 64, |
| 330 | "good_ids": ["d" * 64], |
| 331 | "skipped_ids": [], |
| 332 | "remaining": [], |
| 333 | "log": [], |
| 334 | "branch": "feat\\\\weird", |
| 335 | } |
| 336 | _save_state(root, state) |
| 337 | loaded = _load_state(root) |
| 338 | assert loaded is not None |
| 339 | assert loaded.get("branch") == "feat\\\\weird" |
| 340 | |
| 341 | def test_symbol_filter_injection_survives_roundtrip(self, tmp_path: pathlib.Path) -> None: |
| 342 | from muse.core.bisect import BisectStateDict, _load_state, _save_state |
| 343 | |
| 344 | root, _ = _make_repo(tmp_path) |
| 345 | state: BisectStateDict = { |
| 346 | "bad_id": "e" * 64, |
| 347 | "good_ids": ["f" * 64], |
| 348 | "skipped_ids": [], |
| 349 | "remaining": [], |
| 350 | "log": [], |
| 351 | "symbol_filter": 'billing.py::Invoice"; bad_id = "EVIL', |
| 352 | } |
| 353 | _save_state(root, state) |
| 354 | loaded = _load_state(root) |
| 355 | assert loaded is not None |
| 356 | assert loaded.get("bad_id") == "e" * 64 |
| 357 | assert loaded.get("symbol_filter") == 'billing.py::Invoice"; bad_id = "EVIL' |
| 358 | |
| 359 | |
| 360 | # --------------------------------------------------------------------------- |
| 361 | # Unit — get_bisect_next public API |
| 362 | # --------------------------------------------------------------------------- |
| 363 | |
| 364 | |
| 365 | class TestGetBisectNext: |
| 366 | def test_no_session_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 367 | from muse.core.bisect import get_bisect_next |
| 368 | |
| 369 | root, _ = _make_repo(tmp_path) |
| 370 | nxt, sf = get_bisect_next(root) |
| 371 | assert nxt is None |
| 372 | assert sf == "" |
| 373 | |
| 374 | def test_returns_next_after_start(self, tmp_path: pathlib.Path) -> None: |
| 375 | from muse.core.bisect import get_bisect_next, start_bisect |
| 376 | |
| 377 | root, repo_id = _make_repo(tmp_path) |
| 378 | ids = _build_chain(root, repo_id, 5) |
| 379 | start_bisect(root, ids[-1], [ids[0]]) |
| 380 | nxt, sf = get_bisect_next(root) |
| 381 | assert nxt is not None |
| 382 | assert nxt in ids |
| 383 | assert sf == "" |
| 384 | |
| 385 | def test_returns_symbol_filter(self, tmp_path: pathlib.Path) -> None: |
| 386 | from muse.core.bisect import get_bisect_next, start_bisect |
| 387 | |
| 388 | root, repo_id = _make_repo(tmp_path) |
| 389 | ids = _build_chain(root, repo_id, 4) |
| 390 | # No commits touch this symbol, so remaining will be empty. |
| 391 | start_bisect(root, ids[-1], [ids[0]], symbol_filter="no_file.py::NoSymbol") |
| 392 | nxt, sf = get_bisect_next(root) |
| 393 | # Symbol filter is preserved regardless of whether next exists. |
| 394 | assert sf == "no_file.py::NoSymbol" |
| 395 | |
| 396 | |
| 397 | # --------------------------------------------------------------------------- |
| 398 | # Security — CLI error routing |
| 399 | # --------------------------------------------------------------------------- |
| 400 | |
| 401 | |
| 402 | class TestErrorRouting: |
| 403 | def test_bad_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 404 | root, _ = _make_repo(tmp_path) |
| 405 | result = _invoke(root, ["bisect", "bad"]) |
| 406 | assert result.exit_code != 0 |
| 407 | assert "No bisect session" in (result.stderr or result.output) |
| 408 | |
| 409 | def test_good_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 410 | root, _ = _make_repo(tmp_path) |
| 411 | result = _invoke(root, ["bisect", "good"]) |
| 412 | assert result.exit_code != 0 |
| 413 | assert "No bisect session" in (result.stderr or result.output) |
| 414 | |
| 415 | def test_skip_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 416 | root, _ = _make_repo(tmp_path) |
| 417 | result = _invoke(root, ["bisect", "skip"]) |
| 418 | assert result.exit_code != 0 |
| 419 | assert "No bisect session" in (result.stderr or result.output) |
| 420 | |
| 421 | def test_run_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 422 | root, _ = _make_repo(tmp_path) |
| 423 | result = _invoke(root, ["bisect", "run", "true"]) |
| 424 | assert result.exit_code != 0 |
| 425 | assert "No bisect session" in (result.stderr or result.output) |
| 426 | |
| 427 | def test_symbol_without_double_colon_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 428 | root, repo_id = _make_repo(tmp_path) |
| 429 | ids = _build_chain(root, repo_id, 2) |
| 430 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", "no_colon_here"]) |
| 431 | assert result.exit_code != 0 |
| 432 | assert "❌" in (result.stderr or result.output) |
| 433 | |
| 434 | def test_symbol_too_long_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 435 | root, repo_id = _make_repo(tmp_path) |
| 436 | ids = _build_chain(root, repo_id, 2) |
| 437 | long_sym = "f.py::" + "x" * 600 |
| 438 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", long_sym]) |
| 439 | assert result.exit_code != 0 |
| 440 | assert "too long" in (result.stderr or result.output) |
| 441 | |
| 442 | def test_double_start_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 443 | root, repo_id = _make_repo(tmp_path) |
| 444 | ids = _build_chain(root, repo_id, 3) |
| 445 | r1 = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 446 | assert r1.exit_code == 0 |
| 447 | r2 = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 448 | assert r2.exit_code != 0 |
| 449 | assert "already active" in (r2.stderr or r2.output) |
| 450 | |
| 451 | |
| 452 | # --------------------------------------------------------------------------- |
| 453 | # Security — ANSI sanitization in outputs |
| 454 | # --------------------------------------------------------------------------- |
| 455 | |
| 456 | |
| 457 | class TestAnsiSanitization: |
| 458 | def test_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 459 | root, repo_id = _make_repo(tmp_path) |
| 460 | ids = _build_chain(root, repo_id, 2) |
| 461 | ansi_ref = "\x1b[31mHEAD\x1b[0m" |
| 462 | result = _invoke(root, ["bisect", "start", "--bad", ansi_ref, "--good", ids[0]]) |
| 463 | assert _ANSI_RE.search(result.output) is None |
| 464 | |
| 465 | def test_ansi_in_symbol_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 466 | root, repo_id = _make_repo(tmp_path) |
| 467 | ids = _build_chain(root, repo_id, 2) |
| 468 | sym = "\x1b[31mfoo.py::Bar\x1b[0m" |
| 469 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", sym]) |
| 470 | assert _ANSI_RE.search(result.output) is None |
| 471 | |
| 472 | |
| 473 | # --------------------------------------------------------------------------- |
| 474 | # JSON schema — start |
| 475 | # --------------------------------------------------------------------------- |
| 476 | |
| 477 | |
| 478 | class TestJsonSchemaStart: |
| 479 | def test_start_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 480 | root, repo_id = _make_repo(tmp_path) |
| 481 | ids = _build_chain(root, repo_id, 5) |
| 482 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 483 | assert result.exit_code == 0 |
| 484 | parsed = _parse_step(result.output) |
| 485 | assert parsed["verdict"] == "started" |
| 486 | assert isinstance(parsed["done"], bool) |
| 487 | assert isinstance(parsed["remaining_count"], int) |
| 488 | assert parsed["remaining_count"] >= 0 |
| 489 | |
| 490 | def test_start_json_done_when_no_remaining(self, tmp_path: pathlib.Path) -> None: |
| 491 | """When bad and good are adjacent, start should report done=True immediately.""" |
| 492 | root, repo_id = _make_repo(tmp_path) |
| 493 | ids = _build_chain(root, repo_id, 2) |
| 494 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 495 | assert result.exit_code == 0 |
| 496 | parsed = _parse_step(result.output) |
| 497 | assert parsed["done"] is True |
| 498 | assert parsed["first_bad"] == ids[-1] |
| 499 | |
| 500 | def test_start_json_symbol_changes_list(self, tmp_path: pathlib.Path) -> None: |
| 501 | root, repo_id = _make_repo(tmp_path) |
| 502 | ids = _build_chain(root, repo_id, 4) |
| 503 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 504 | parsed = _parse_step(result.output) |
| 505 | assert isinstance(parsed["symbol_changes"], list) |
| 506 | |
| 507 | |
| 508 | # --------------------------------------------------------------------------- |
| 509 | # JSON schema — bad / good / skip |
| 510 | # --------------------------------------------------------------------------- |
| 511 | |
| 512 | |
| 513 | class TestJsonSchemaBadGoodSkip: |
| 514 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 515 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 516 | assert r.exit_code == 0 |
| 517 | |
| 518 | def test_bad_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 519 | root, repo_id = _make_repo(tmp_path) |
| 520 | ids = _build_chain(root, repo_id, 5) |
| 521 | self._start(root, ids) |
| 522 | midpoint = ids[len(ids) // 2] |
| 523 | result = _invoke(root, ["bisect", "bad", midpoint, "--json"]) |
| 524 | assert result.exit_code == 0 |
| 525 | parsed = _parse_step(result.output) |
| 526 | assert parsed["verdict"] == "bad" |
| 527 | |
| 528 | def test_good_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 529 | root, repo_id = _make_repo(tmp_path) |
| 530 | ids = _build_chain(root, repo_id, 5) |
| 531 | self._start(root, ids) |
| 532 | midpoint = ids[len(ids) // 2] |
| 533 | result = _invoke(root, ["bisect", "good", midpoint, "--json"]) |
| 534 | assert result.exit_code == 0 |
| 535 | parsed = _parse_step(result.output) |
| 536 | assert parsed["verdict"] == "good" |
| 537 | |
| 538 | def test_skip_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 539 | root, repo_id = _make_repo(tmp_path) |
| 540 | ids = _build_chain(root, repo_id, 5) |
| 541 | self._start(root, ids) |
| 542 | midpoint = ids[len(ids) // 2] |
| 543 | result = _invoke(root, ["bisect", "skip", midpoint, "--json"]) |
| 544 | assert result.exit_code == 0 |
| 545 | parsed = _parse_step(result.output) |
| 546 | assert parsed["verdict"] == "skip" |
| 547 | |
| 548 | |
| 549 | # --------------------------------------------------------------------------- |
| 550 | # JSON schema — log |
| 551 | # --------------------------------------------------------------------------- |
| 552 | |
| 553 | |
| 554 | class TestJsonSchemaLog: |
| 555 | def test_log_json_no_session(self, tmp_path: pathlib.Path) -> None: |
| 556 | root, _ = _make_repo(tmp_path) |
| 557 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 558 | assert result.exit_code == 0 |
| 559 | parsed = _parse_log(result.output) |
| 560 | assert parsed["active"] is False |
| 561 | assert parsed["entries"] == [] |
| 562 | |
| 563 | def test_log_json_after_start(self, tmp_path: pathlib.Path) -> None: |
| 564 | root, repo_id = _make_repo(tmp_path) |
| 565 | ids = _build_chain(root, repo_id, 4) |
| 566 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 567 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 568 | assert result.exit_code == 0 |
| 569 | parsed = _parse_log(result.output) |
| 570 | assert parsed["active"] is True |
| 571 | assert len(parsed["entries"]) >= 2 |
| 572 | |
| 573 | def test_log_json_entries_are_strings(self, tmp_path: pathlib.Path) -> None: |
| 574 | root, repo_id = _make_repo(tmp_path) |
| 575 | ids = _build_chain(root, repo_id, 3) |
| 576 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 577 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 578 | parsed = _parse_log(result.output) |
| 579 | for entry in parsed["entries"]: |
| 580 | assert isinstance(entry, str) |
| 581 | |
| 582 | |
| 583 | # --------------------------------------------------------------------------- |
| 584 | # JSON schema — reset |
| 585 | # --------------------------------------------------------------------------- |
| 586 | |
| 587 | |
| 588 | class TestJsonSchemaReset: |
| 589 | def test_reset_json_no_session(self, tmp_path: pathlib.Path) -> None: |
| 590 | root, _ = _make_repo(tmp_path) |
| 591 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 592 | assert result.exit_code == 0 |
| 593 | parsed = _parse_reset(result.output) |
| 594 | assert parsed["reset"] is True |
| 595 | |
| 596 | def test_reset_json_with_session(self, tmp_path: pathlib.Path) -> None: |
| 597 | root, repo_id = _make_repo(tmp_path) |
| 598 | ids = _build_chain(root, repo_id, 3) |
| 599 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 600 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 601 | assert result.exit_code == 0 |
| 602 | parsed = _parse_reset(result.output) |
| 603 | assert parsed["reset"] is True |
| 604 | |
| 605 | def test_reset_clears_active_flag(self, tmp_path: pathlib.Path) -> None: |
| 606 | root, repo_id = _make_repo(tmp_path) |
| 607 | ids = _build_chain(root, repo_id, 3) |
| 608 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 609 | _invoke(root, ["bisect", "reset", "--json"]) |
| 610 | log_result = _invoke(root, ["bisect", "log", "--json"]) |
| 611 | parsed = _parse_log(log_result.output) |
| 612 | assert parsed["active"] is False |
| 613 | |
| 614 | |
| 615 | # --------------------------------------------------------------------------- |
| 616 | # JSON schema — run (NDJSON) |
| 617 | # --------------------------------------------------------------------------- |
| 618 | |
| 619 | |
| 620 | class TestJsonSchemaRun: |
| 621 | def test_run_json_ndjson_format(self, tmp_path: pathlib.Path) -> None: |
| 622 | """``bisect run --json`` should emit valid NDJSON.""" |
| 623 | root, repo_id = _make_repo(tmp_path) |
| 624 | ids = _build_chain(root, repo_id, 6) |
| 625 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 626 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 627 | assert result.exit_code == 0 |
| 628 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 629 | assert len(lines) >= 1 |
| 630 | for raw_line in lines[:-1]: |
| 631 | step_raw = json.loads(raw_line) |
| 632 | assert "step" in step_raw |
| 633 | assert "verdict" in step_raw |
| 634 | assert "testing" in step_raw |
| 635 | assert "remaining_count" in step_raw |
| 636 | assert "done" in step_raw |
| 637 | done_raw = json.loads(lines[-1]) |
| 638 | done_val = done_raw["done"] |
| 639 | assert isinstance(done_val, bool) |
| 640 | steps_taken_val = done_raw["steps_taken"] |
| 641 | assert isinstance(steps_taken_val, int) |
| 642 | |
| 643 | def test_run_json_done_has_first_bad(self, tmp_path: pathlib.Path) -> None: |
| 644 | """With always-good command, first_bad on the done line should be set.""" |
| 645 | root, repo_id = _make_repo(tmp_path) |
| 646 | ids = _build_chain(root, repo_id, 4) |
| 647 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 648 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 649 | assert result.exit_code == 0 |
| 650 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 651 | done_raw = json.loads(lines[-1]) |
| 652 | done_val = done_raw["done"] |
| 653 | first_bad_val = done_raw["first_bad"] |
| 654 | if done_val: |
| 655 | assert first_bad_val is not None |
| 656 | |
| 657 | def test_run_json_steps_taken_increments(self, tmp_path: pathlib.Path) -> None: |
| 658 | root, repo_id = _make_repo(tmp_path) |
| 659 | ids = _build_chain(root, repo_id, 8) |
| 660 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 661 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 662 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 663 | done_raw = json.loads(lines[-1]) |
| 664 | steps_taken = done_raw["steps_taken"] |
| 665 | assert steps_taken >= 1 |
| 666 | |
| 667 | |
| 668 | # --------------------------------------------------------------------------- |
| 669 | # Integration — session lifecycle with --json |
| 670 | # --------------------------------------------------------------------------- |
| 671 | |
| 672 | |
| 673 | class TestIntegrationJson: |
| 674 | def test_start_bad_good_converge(self, tmp_path: pathlib.Path) -> None: |
| 675 | """A manual bisect session with --json converges to a first_bad.""" |
| 676 | root, repo_id = _make_repo(tmp_path) |
| 677 | ids = _build_chain(root, repo_id, 7) |
| 678 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 679 | assert r_start.exit_code == 0 |
| 680 | step = _parse_step(r_start.output) |
| 681 | if step["done"]: |
| 682 | assert step["first_bad"] is not None |
| 683 | return |
| 684 | for _ in range(20): |
| 685 | nxt = step["next_to_test"] |
| 686 | assert nxt is not None |
| 687 | r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 688 | assert r.exit_code == 0 |
| 689 | step = _parse_step(r.output) |
| 690 | if step["done"]: |
| 691 | assert step["first_bad"] is not None |
| 692 | return |
| 693 | pytest.fail("Bisect did not converge within 20 steps") |
| 694 | |
| 695 | def test_good_narrows_range(self, tmp_path: pathlib.Path) -> None: |
| 696 | root, repo_id = _make_repo(tmp_path) |
| 697 | ids = _build_chain(root, repo_id, 8) |
| 698 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 699 | midpoint = ids[len(ids) // 2] |
| 700 | r_good = _invoke(root, ["bisect", "good", midpoint, "--json"]) |
| 701 | step = _parse_step(r_good.output) |
| 702 | if not step["done"]: |
| 703 | assert step["remaining_count"] < len(ids) - 2 |
| 704 | |
| 705 | def test_log_grows_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 706 | root, repo_id = _make_repo(tmp_path) |
| 707 | ids = _build_chain(root, repo_id, 5) |
| 708 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 709 | midpoint = ids[len(ids) // 2] |
| 710 | _invoke(root, ["bisect", "bad", midpoint]) |
| 711 | r_log = _invoke(root, ["bisect", "log", "--json"]) |
| 712 | parsed = _parse_log(r_log.output) |
| 713 | # start logs 2 entries (bad+good); bad adds 1 more → at least 3. |
| 714 | assert len(parsed["entries"]) >= 3 |
| 715 | |
| 716 | def test_skip_excluded_from_remaining(self, tmp_path: pathlib.Path) -> None: |
| 717 | root, repo_id = _make_repo(tmp_path) |
| 718 | ids = _build_chain(root, repo_id, 6) |
| 719 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 720 | step_start = _parse_step(r_start.output) |
| 721 | if step_start["done"]: |
| 722 | return |
| 723 | nxt = step_start["next_to_test"] |
| 724 | assert nxt is not None |
| 725 | r_skip = _invoke(root, ["bisect", "skip", nxt, "--json"]) |
| 726 | step_skip = _parse_step(r_skip.output) |
| 727 | if not step_skip["done"]: |
| 728 | assert step_skip["next_to_test"] != nxt |
| 729 | |
| 730 | |
| 731 | # --------------------------------------------------------------------------- |
| 732 | # E2E — text (non-JSON) output still works |
| 733 | # --------------------------------------------------------------------------- |
| 734 | |
| 735 | |
| 736 | class TestE2EText: |
| 737 | def test_start_text_output_no_json(self, tmp_path: pathlib.Path) -> None: |
| 738 | root, repo_id = _make_repo(tmp_path) |
| 739 | ids = _build_chain(root, repo_id, 4) |
| 740 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 741 | assert result.exit_code == 0 |
| 742 | assert "Bisect session started" in result.output or "First bad commit" in result.output |
| 743 | |
| 744 | def test_bad_text_output(self, tmp_path: pathlib.Path) -> None: |
| 745 | root, repo_id = _make_repo(tmp_path) |
| 746 | ids = _build_chain(root, repo_id, 4) |
| 747 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 748 | midpoint = ids[len(ids) // 2] |
| 749 | result = _invoke(root, ["bisect", "bad", midpoint]) |
| 750 | assert result.exit_code == 0 |
| 751 | assert "bad" in result.output.lower() |
| 752 | |
| 753 | def test_log_text_shows_entries(self, tmp_path: pathlib.Path) -> None: |
| 754 | root, repo_id = _make_repo(tmp_path) |
| 755 | ids = _build_chain(root, repo_id, 3) |
| 756 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 757 | result = _invoke(root, ["bisect", "log"]) |
| 758 | assert result.exit_code == 0 |
| 759 | assert "Bisect log" in result.output |
| 760 | |
| 761 | def test_reset_text_output(self, tmp_path: pathlib.Path) -> None: |
| 762 | root, repo_id = _make_repo(tmp_path) |
| 763 | ids = _build_chain(root, repo_id, 2) |
| 764 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 765 | result = _invoke(root, ["bisect", "reset"]) |
| 766 | assert result.exit_code == 0 |
| 767 | assert "reset" in result.output.lower() |
| 768 | |
| 769 | def test_run_text_output_converges(self, tmp_path: pathlib.Path) -> None: |
| 770 | root, repo_id = _make_repo(tmp_path) |
| 771 | ids = _build_chain(root, repo_id, 5) |
| 772 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 773 | result = _invoke(root, ["bisect", "run", "true"]) |
| 774 | assert result.exit_code == 0 |
| 775 | assert "First bad commit" in result.output or "Bisect complete" in result.output |
| 776 | |
| 777 | def test_no_good_flag_fails_clearly(self, tmp_path: pathlib.Path) -> None: |
| 778 | root, repo_id = _make_repo(tmp_path) |
| 779 | ids = _build_chain(root, repo_id, 2) |
| 780 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 781 | assert result.exit_code != 0 |
| 782 | |
| 783 | def test_log_empty_when_no_session(self, tmp_path: pathlib.Path) -> None: |
| 784 | root, _ = _make_repo(tmp_path) |
| 785 | result = _invoke(root, ["bisect", "log"]) |
| 786 | assert result.exit_code == 0 |
| 787 | assert "No bisect log" in result.output |
| 788 | |
| 789 | |
| 790 | # --------------------------------------------------------------------------- |
| 791 | # E2E — symbol-scoped bisect |
| 792 | # --------------------------------------------------------------------------- |
| 793 | |
| 794 | |
| 795 | class TestSymbolScopedBisect: |
| 796 | def test_symbol_filter_no_matching_commits_warns(self, tmp_path: pathlib.Path) -> None: |
| 797 | root, repo_id = _make_repo(tmp_path) |
| 798 | ids = _build_chain(root, repo_id, 4) |
| 799 | result = _invoke( |
| 800 | root, |
| 801 | [ |
| 802 | "bisect", "start", |
| 803 | "--bad", ids[-1], |
| 804 | "--good", ids[0], |
| 805 | "--symbol", "ghost.py::GhostFunc", |
| 806 | ], |
| 807 | ) |
| 808 | assert result.exit_code == 0 |
| 809 | combined = result.output + (result.stderr or "") |
| 810 | assert "No commits" in combined or "First bad" in combined |
| 811 | |
| 812 | def test_symbol_filter_json_schema_preserved(self, tmp_path: pathlib.Path) -> None: |
| 813 | root, repo_id = _make_repo(tmp_path) |
| 814 | ids = _build_chain(root, repo_id, 5) |
| 815 | result = _invoke( |
| 816 | root, |
| 817 | [ |
| 818 | "bisect", "start", |
| 819 | "--bad", ids[-1], |
| 820 | "--good", ids[0], |
| 821 | "--symbol", "ghost.py::GhostFunc", |
| 822 | "--json", |
| 823 | ], |
| 824 | ) |
| 825 | assert result.exit_code == 0 |
| 826 | parsed = _parse_step(result.output) |
| 827 | assert isinstance(parsed["symbol_changes"], list) |
| 828 | |
| 829 | def test_symbol_filter_state_persisted(self, tmp_path: pathlib.Path) -> None: |
| 830 | """After start with --symbol, the symbol_filter must survive state reload.""" |
| 831 | from muse.core.bisect import _load_state |
| 832 | |
| 833 | root, repo_id = _make_repo(tmp_path) |
| 834 | ids = _build_chain(root, repo_id, 4) |
| 835 | _invoke( |
| 836 | root, |
| 837 | [ |
| 838 | "bisect", "start", |
| 839 | "--bad", ids[-1], |
| 840 | "--good", ids[0], |
| 841 | "--symbol", "billing.py::Invoice", |
| 842 | ], |
| 843 | ) |
| 844 | state = _load_state(root) |
| 845 | assert state is not None |
| 846 | assert state.get("symbol_filter") == "billing.py::Invoice" |
| 847 | |
| 848 | |
| 849 | # --------------------------------------------------------------------------- |
| 850 | # Stress — large commit chains |
| 851 | # --------------------------------------------------------------------------- |
| 852 | |
| 853 | |
| 854 | class TestStress: |
| 855 | def test_200_commit_chain_converges(self, tmp_path: pathlib.Path) -> None: |
| 856 | """Bisect over 200 commits must converge in ≤9 steps (log₂(200) ≈ 7.6).""" |
| 857 | root, repo_id = _make_repo(tmp_path) |
| 858 | ids = _build_chain(root, repo_id, 200) |
| 859 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 860 | |
| 861 | steps = 0 |
| 862 | for _ in range(10): |
| 863 | r = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 864 | assert r.exit_code == 0 |
| 865 | lines = [ln.strip() for ln in r.output.strip().splitlines() if ln.strip()] |
| 866 | if lines: |
| 867 | done_raw = json.loads(lines[-1]) |
| 868 | if done_raw.get("done"): |
| 869 | steps = done_raw.get("steps_taken", 0) |
| 870 | break |
| 871 | else: |
| 872 | pytest.fail("Bisect did not terminate within 10 run invocations") |
| 873 | assert steps <= 9, f"Expected ≤9 steps for 200 commits, got {steps}" |
| 874 | |
| 875 | def test_concurrent_log_reads_are_safe(self, tmp_path: pathlib.Path) -> None: |
| 876 | """Concurrent reads of bisect log must not crash.""" |
| 877 | root, repo_id = _make_repo(tmp_path) |
| 878 | ids = _build_chain(root, repo_id, 10) |
| 879 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 880 | |
| 881 | errors: list[str] = [] |
| 882 | |
| 883 | def _read_log() -> None: |
| 884 | from muse.core.bisect import get_bisect_log |
| 885 | try: |
| 886 | entries = get_bisect_log(root) |
| 887 | assert isinstance(entries, list) |
| 888 | except Exception as exc: |
| 889 | errors.append(str(exc)) |
| 890 | |
| 891 | threads = [threading.Thread(target=_read_log) for _ in range(20)] |
| 892 | for t in threads: |
| 893 | t.start() |
| 894 | for t in threads: |
| 895 | t.join() |
| 896 | |
| 897 | assert not errors, f"Concurrent read failures: {errors}" |
| 898 | |
| 899 | def test_50_step_manual_bisect_json(self, tmp_path: pathlib.Path) -> None: |
| 900 | """50 mark_bad calls on a 100-commit chain must all emit valid JSON.""" |
| 901 | root, repo_id = _make_repo(tmp_path) |
| 902 | ids = _build_chain(root, repo_id, 100) |
| 903 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 904 | assert r_start.exit_code == 0 |
| 905 | step = _parse_step(r_start.output) |
| 906 | |
| 907 | for _ in range(50): |
| 908 | if step["done"]: |
| 909 | assert step["first_bad"] is not None |
| 910 | return |
| 911 | nxt = step["next_to_test"] |
| 912 | assert nxt is not None |
| 913 | r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 914 | assert r.exit_code == 0 |
| 915 | step = _parse_step(r.output) |
| 916 | |
| 917 | assert step["done"] is True |
| 918 | |
| 919 | |
| 920 | # --------------------------------------------------------------------------- |
| 921 | # bisect start — Extended, Security, Stress |
| 922 | # --------------------------------------------------------------------------- |
| 923 | |
| 924 | |
| 925 | class TestBisectStartExtended: |
| 926 | """Extended unit / integration / e2e tests for muse bisect start.""" |
| 927 | |
| 928 | def test_start_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 929 | root, repo_id = _make_repo(tmp_path) |
| 930 | ids = _build_chain(root, repo_id, 5) |
| 931 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 932 | assert result.exit_code == 0 |
| 933 | |
| 934 | def test_start_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 935 | """-j is an accepted alias for --json.""" |
| 936 | root, repo_id = _make_repo(tmp_path) |
| 937 | ids = _build_chain(root, repo_id, 5) |
| 938 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "-j"]) |
| 939 | assert result.exit_code == 0 |
| 940 | parsed = _parse_step(result.output) |
| 941 | assert parsed["verdict"] == "started" |
| 942 | |
| 943 | def test_start_json_verdict_is_started(self, tmp_path: pathlib.Path) -> None: |
| 944 | root, repo_id = _make_repo(tmp_path) |
| 945 | ids = _build_chain(root, repo_id, 5) |
| 946 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 947 | assert result.exit_code == 0 |
| 948 | assert _parse_step(result.output)["verdict"] == "started" |
| 949 | |
| 950 | def test_start_json_done_false_with_remaining(self, tmp_path: pathlib.Path) -> None: |
| 951 | root, repo_id = _make_repo(tmp_path) |
| 952 | ids = _build_chain(root, repo_id, 5) |
| 953 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 954 | assert result.exit_code == 0 |
| 955 | parsed = _parse_step(result.output) |
| 956 | assert parsed["done"] is False |
| 957 | assert parsed["next_to_test"] is not None |
| 958 | |
| 959 | def test_start_json_done_true_when_adjacent(self, tmp_path: pathlib.Path) -> None: |
| 960 | root, repo_id = _make_repo(tmp_path) |
| 961 | ids = _build_chain(root, repo_id, 2) |
| 962 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 963 | assert result.exit_code == 0 |
| 964 | parsed = _parse_step(result.output) |
| 965 | assert parsed["done"] is True |
| 966 | assert parsed["first_bad"] == ids[-1] |
| 967 | |
| 968 | def test_start_json_remaining_count_positive(self, tmp_path: pathlib.Path) -> None: |
| 969 | root, repo_id = _make_repo(tmp_path) |
| 970 | ids = _build_chain(root, repo_id, 8) |
| 971 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 972 | assert result.exit_code == 0 |
| 973 | assert _parse_step(result.output)["remaining_count"] > 0 |
| 974 | |
| 975 | def test_start_json_steps_remaining_positive(self, tmp_path: pathlib.Path) -> None: |
| 976 | root, repo_id = _make_repo(tmp_path) |
| 977 | ids = _build_chain(root, repo_id, 8) |
| 978 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 979 | assert result.exit_code == 0 |
| 980 | assert _parse_step(result.output)["steps_remaining"] > 0 |
| 981 | |
| 982 | def test_start_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 983 | root, repo_id = _make_repo(tmp_path) |
| 984 | ids = _build_chain(root, repo_id, 5) |
| 985 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 986 | assert result.exit_code == 0 |
| 987 | d = json.loads(_json_blob(result.output)) |
| 988 | assert set(d.keys()) == {"done", "first_bad", "next_to_test", "remaining_count", |
| 989 | "steps_remaining", "verdict", "symbol_changes"} |
| 990 | |
| 991 | def test_start_multiple_good_refs(self, tmp_path: pathlib.Path) -> None: |
| 992 | root, repo_id = _make_repo(tmp_path) |
| 993 | ids = _build_chain(root, repo_id, 6) |
| 994 | result = _invoke( |
| 995 | root, |
| 996 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--good", ids[1], "--json"], |
| 997 | ) |
| 998 | assert result.exit_code == 0 |
| 999 | assert _parse_step(result.output)["verdict"] == "started" |
| 1000 | |
| 1001 | def test_start_no_good_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1002 | root, repo_id = _make_repo(tmp_path) |
| 1003 | ids = _build_chain(root, repo_id, 3) |
| 1004 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 1005 | assert result.exit_code == 1 |
| 1006 | |
| 1007 | def test_start_no_good_error_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 1008 | root, repo_id = _make_repo(tmp_path) |
| 1009 | ids = _build_chain(root, repo_id, 3) |
| 1010 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 1011 | assert result.exit_code != 0 |
| 1012 | combined = result.output + (result.stderr or "") |
| 1013 | assert "good" in combined.lower() |
| 1014 | |
| 1015 | def test_start_double_start_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1016 | root, repo_id = _make_repo(tmp_path) |
| 1017 | ids = _build_chain(root, repo_id, 5) |
| 1018 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1019 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1020 | assert result.exit_code == 1 |
| 1021 | |
| 1022 | def test_start_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1023 | empty = tmp_path / "not_a_repo" |
| 1024 | empty.mkdir() |
| 1025 | result = _invoke(empty, ["bisect", "start", "--bad", "abc", "--good", "def"]) |
| 1026 | assert result.exit_code == 2 |
| 1027 | |
| 1028 | def test_start_bad_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1029 | root, repo_id = _make_repo(tmp_path) |
| 1030 | ids = _build_chain(root, repo_id, 4) |
| 1031 | # HEAD points to ids[-1]; omit --bad |
| 1032 | result = _invoke(root, ["bisect", "start", "--good", ids[0], "--json"]) |
| 1033 | assert result.exit_code == 0 |
| 1034 | |
| 1035 | def test_start_text_mentions_session_started(self, tmp_path: pathlib.Path) -> None: |
| 1036 | root, repo_id = _make_repo(tmp_path) |
| 1037 | ids = _build_chain(root, repo_id, 5) |
| 1038 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1039 | assert result.exit_code == 0 |
| 1040 | assert "Bisect session started" in result.output or "First bad commit" in result.output |
| 1041 | |
| 1042 | def test_start_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1043 | root, repo_id = _make_repo(tmp_path) |
| 1044 | ids = _build_chain(root, repo_id, 5) |
| 1045 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1046 | assert result.exit_code == 0 |
| 1047 | assert not result.output.strip().startswith("{") |
| 1048 | |
| 1049 | def test_start_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1050 | root, _ = _make_repo(tmp_path) |
| 1051 | result = _invoke(root, ["bisect", "start", "--help"]) |
| 1052 | assert "Agent quickstart" in result.output or "binary" in result.output.lower() |
| 1053 | |
| 1054 | def test_start_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1055 | root, repo_id = _make_repo(tmp_path) |
| 1056 | ids = _build_chain(root, repo_id, 3) |
| 1057 | result = _invoke(root, ["bisect", "start", "--bad", "nonexistent_ref_abc123", "--good", ids[0]]) |
| 1058 | assert result.exit_code == 1 |
| 1059 | |
| 1060 | |
| 1061 | class TestBisectStartSecurity: |
| 1062 | """Security hardening tests for muse bisect start.""" |
| 1063 | |
| 1064 | def test_start_symbol_changes_no_ansi_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1065 | """symbol_changes entries are sanitized in JSON output.""" |
| 1066 | from unittest.mock import patch |
| 1067 | from muse.core.bisect import BisectResult |
| 1068 | root, repo_id = _make_repo(tmp_path) |
| 1069 | ids = _build_chain(root, repo_id, 5) |
| 1070 | injected = BisectResult( |
| 1071 | done=False, |
| 1072 | first_bad=None, |
| 1073 | next_to_test=ids[2], |
| 1074 | remaining_count=3, |
| 1075 | steps_remaining=2, |
| 1076 | verdict="started", |
| 1077 | symbol_changes=["add Invoice.compute\x1b[31mred\x1b[0m"], |
| 1078 | ) |
| 1079 | with patch("muse.cli.commands.bisect.start_bisect", return_value=injected): |
| 1080 | result = _invoke( |
| 1081 | root, |
| 1082 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], |
| 1083 | "--symbol", "billing.py::Invoice", "--json"], |
| 1084 | ) |
| 1085 | assert result.exit_code == 0 |
| 1086 | assert "\x1b" not in result.output |
| 1087 | |
| 1088 | def test_start_symbol_changes_no_ansi_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1089 | """symbol_changes entries are sanitized in text output.""" |
| 1090 | from unittest.mock import patch |
| 1091 | from muse.core.bisect import BisectResult |
| 1092 | root, repo_id = _make_repo(tmp_path) |
| 1093 | ids = _build_chain(root, repo_id, 5) |
| 1094 | injected = BisectResult( |
| 1095 | done=False, |
| 1096 | first_bad=None, |
| 1097 | next_to_test=ids[2], |
| 1098 | remaining_count=3, |
| 1099 | steps_remaining=2, |
| 1100 | verdict="started", |
| 1101 | symbol_changes=["add Invoice.compute\x1b[31mred\x1b[0m"], |
| 1102 | ) |
| 1103 | with patch("muse.cli.commands.bisect.start_bisect", return_value=injected): |
| 1104 | result = _invoke( |
| 1105 | root, |
| 1106 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], |
| 1107 | "--symbol", "billing.py::Invoice"], |
| 1108 | ) |
| 1109 | assert result.exit_code == 0 |
| 1110 | assert "\x1b" not in result.output |
| 1111 | |
| 1112 | def test_start_symbol_missing_separator_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1113 | """--symbol without '::' separator is rejected.""" |
| 1114 | root, repo_id = _make_repo(tmp_path) |
| 1115 | ids = _build_chain(root, repo_id, 3) |
| 1116 | result = _invoke( |
| 1117 | root, |
| 1118 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", "NoSeparator"], |
| 1119 | ) |
| 1120 | assert result.exit_code == 1 |
| 1121 | |
| 1122 | def test_start_symbol_too_long_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1123 | """--symbol exceeding max length is rejected.""" |
| 1124 | root, repo_id = _make_repo(tmp_path) |
| 1125 | ids = _build_chain(root, repo_id, 3) |
| 1126 | long_sym = "a" * 510 + "::b" |
| 1127 | result = _invoke( |
| 1128 | root, |
| 1129 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", long_sym], |
| 1130 | ) |
| 1131 | assert result.exit_code == 1 |
| 1132 | |
| 1133 | def test_start_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1134 | """JSON output is well-formed.""" |
| 1135 | root, repo_id = _make_repo(tmp_path) |
| 1136 | ids = _build_chain(root, repo_id, 5) |
| 1137 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1138 | assert result.exit_code == 0 |
| 1139 | d = json.loads(_json_blob(result.output)) |
| 1140 | assert isinstance(d, dict) |
| 1141 | |
| 1142 | def test_start_json_bool_fields_are_bool(self, tmp_path: pathlib.Path) -> None: |
| 1143 | """done field is always a bool, never int or string.""" |
| 1144 | root, repo_id = _make_repo(tmp_path) |
| 1145 | ids = _build_chain(root, repo_id, 5) |
| 1146 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1147 | assert result.exit_code == 0 |
| 1148 | d = json.loads(_json_blob(result.output)) |
| 1149 | assert isinstance(d["done"], bool) |
| 1150 | |
| 1151 | |
| 1152 | class TestBisectStartStress: |
| 1153 | """Performance and scale tests for muse bisect start.""" |
| 1154 | |
| 1155 | def test_start_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1156 | """Start over a 100-commit chain exits 0 and returns a midpoint.""" |
| 1157 | root, repo_id = _make_repo(tmp_path) |
| 1158 | ids = _build_chain(root, repo_id, 100) |
| 1159 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1160 | assert result.exit_code == 0 |
| 1161 | parsed = _parse_step(result.output) |
| 1162 | assert parsed["done"] is False |
| 1163 | assert parsed["remaining_count"] > 0 |
| 1164 | assert parsed["next_to_test"] is not None |
| 1165 | |
| 1166 | def test_start_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1167 | """Start over 100 commits completes within 5 seconds.""" |
| 1168 | import time |
| 1169 | root, repo_id = _make_repo(tmp_path) |
| 1170 | ids = _build_chain(root, repo_id, 100) |
| 1171 | t0 = time.monotonic() |
| 1172 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1173 | elapsed = time.monotonic() - t0 |
| 1174 | assert result.exit_code == 0 |
| 1175 | assert elapsed < 5.0, f"start over 100 commits took {elapsed:.2f}s" |
| 1176 | |
| 1177 | def test_start_midpoint_is_within_range(self, tmp_path: pathlib.Path) -> None: |
| 1178 | """The suggested midpoint falls strictly between good and bad.""" |
| 1179 | root, repo_id = _make_repo(tmp_path) |
| 1180 | ids = _build_chain(root, repo_id, 20) |
| 1181 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1182 | assert result.exit_code == 0 |
| 1183 | parsed = _parse_step(result.output) |
| 1184 | assert parsed["next_to_test"] not in (ids[0], ids[-1]) |
| 1185 | |
| 1186 | |
| 1187 | # --------------------------------------------------------------------------- |
| 1188 | # bisect bad — Extended, Security, Stress |
| 1189 | # --------------------------------------------------------------------------- |
| 1190 | |
| 1191 | |
| 1192 | class TestBisectBadExtended: |
| 1193 | """Extended unit / integration / e2e tests for muse bisect bad.""" |
| 1194 | |
| 1195 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1196 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1197 | assert r.exit_code == 0 |
| 1198 | |
| 1199 | def test_bad_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1200 | root, repo_id = _make_repo(tmp_path) |
| 1201 | ids = _build_chain(root, repo_id, 6) |
| 1202 | self._start(root, ids) |
| 1203 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2]]) |
| 1204 | assert result.exit_code == 0 |
| 1205 | |
| 1206 | def test_bad_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1207 | root, repo_id = _make_repo(tmp_path) |
| 1208 | ids = _build_chain(root, repo_id, 6) |
| 1209 | self._start(root, ids) |
| 1210 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "-j"]) |
| 1211 | assert result.exit_code == 0 |
| 1212 | parsed = _parse_step(result.output) |
| 1213 | assert parsed["verdict"] == "bad" |
| 1214 | |
| 1215 | def test_bad_json_verdict_is_bad(self, tmp_path: pathlib.Path) -> None: |
| 1216 | root, repo_id = _make_repo(tmp_path) |
| 1217 | ids = _build_chain(root, repo_id, 6) |
| 1218 | self._start(root, ids) |
| 1219 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1220 | assert result.exit_code == 0 |
| 1221 | assert _parse_step(result.output)["verdict"] == "bad" |
| 1222 | |
| 1223 | def test_bad_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1224 | root, repo_id = _make_repo(tmp_path) |
| 1225 | ids = _build_chain(root, repo_id, 6) |
| 1226 | self._start(root, ids) |
| 1227 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1228 | assert result.exit_code == 0 |
| 1229 | d = json.loads(_json_blob(result.output)) |
| 1230 | assert set(d.keys()) == {"done", "first_bad", "next_to_test", "remaining_count", |
| 1231 | "steps_remaining", "verdict", "symbol_changes"} |
| 1232 | |
| 1233 | def test_bad_reduces_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1234 | root, repo_id = _make_repo(tmp_path) |
| 1235 | ids = _build_chain(root, repo_id, 10) |
| 1236 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1237 | before = _parse_step(r.output)["remaining_count"] |
| 1238 | mid = _parse_step(r.output)["next_to_test"] |
| 1239 | result = _invoke(root, ["bisect", "bad", mid, "--json"]) |
| 1240 | assert result.exit_code == 0 |
| 1241 | after = _parse_step(result.output)["remaining_count"] |
| 1242 | assert after < before |
| 1243 | |
| 1244 | def test_bad_done_true_when_isolated(self, tmp_path: pathlib.Path) -> None: |
| 1245 | root, repo_id = _make_repo(tmp_path) |
| 1246 | ids = _build_chain(root, repo_id, 3) |
| 1247 | # With 3 commits: good=ids[0], bad=ids[2] → ids[1] is the only remaining |
| 1248 | self._start(root, ids) |
| 1249 | result = _invoke(root, ["bisect", "bad", ids[1], "--json"]) |
| 1250 | assert result.exit_code == 0 |
| 1251 | parsed = _parse_step(result.output) |
| 1252 | assert parsed["done"] is True |
| 1253 | assert parsed["first_bad"] is not None |
| 1254 | |
| 1255 | def test_bad_first_bad_set_when_done(self, tmp_path: pathlib.Path) -> None: |
| 1256 | root, repo_id = _make_repo(tmp_path) |
| 1257 | ids = _build_chain(root, repo_id, 3) |
| 1258 | self._start(root, ids) |
| 1259 | result = _invoke(root, ["bisect", "bad", ids[1], "--json"]) |
| 1260 | assert result.exit_code == 0 |
| 1261 | parsed = _parse_step(result.output) |
| 1262 | assert parsed["done"] is True |
| 1263 | assert isinstance(parsed["first_bad"], str) |
| 1264 | |
| 1265 | def test_bad_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1266 | root, repo_id = _make_repo(tmp_path) |
| 1267 | ids = _build_chain(root, repo_id, 5) |
| 1268 | self._start(root, ids) |
| 1269 | # HEAD points to ids[-1] (the known-bad); marking it bad again is valid |
| 1270 | result = _invoke(root, ["bisect", "bad", "--json"]) |
| 1271 | assert result.exit_code == 0 |
| 1272 | |
| 1273 | def test_bad_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1274 | root, _ = _make_repo(tmp_path) |
| 1275 | result = _invoke(root, ["bisect", "bad"]) |
| 1276 | assert result.exit_code == 1 |
| 1277 | |
| 1278 | def test_bad_no_session_error_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 1279 | root, _ = _make_repo(tmp_path) |
| 1280 | result = _invoke(root, ["bisect", "bad"]) |
| 1281 | assert result.exit_code != 0 |
| 1282 | combined = result.output + (result.stderr or "") |
| 1283 | assert "No bisect session" in combined |
| 1284 | |
| 1285 | def test_bad_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1286 | empty = tmp_path / "not_a_repo" |
| 1287 | empty.mkdir() |
| 1288 | result = _invoke(empty, ["bisect", "bad"]) |
| 1289 | assert result.exit_code == 2 |
| 1290 | |
| 1291 | def test_bad_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1292 | root, repo_id = _make_repo(tmp_path) |
| 1293 | ids = _build_chain(root, repo_id, 4) |
| 1294 | self._start(root, ids) |
| 1295 | result = _invoke(root, ["bisect", "bad", "deadbeef_nonexistent"]) |
| 1296 | assert result.exit_code == 1 |
| 1297 | |
| 1298 | def test_bad_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1299 | root, repo_id = _make_repo(tmp_path) |
| 1300 | ids = _build_chain(root, repo_id, 5) |
| 1301 | self._start(root, ids) |
| 1302 | mid = ids[len(ids) // 2] |
| 1303 | result = _invoke(root, ["bisect", "bad", mid]) |
| 1304 | assert result.exit_code == 0 |
| 1305 | assert mid[:12] in result.output |
| 1306 | |
| 1307 | def test_bad_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1308 | root, repo_id = _make_repo(tmp_path) |
| 1309 | ids = _build_chain(root, repo_id, 5) |
| 1310 | self._start(root, ids) |
| 1311 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2]]) |
| 1312 | assert result.exit_code == 0 |
| 1313 | assert not result.output.strip().startswith("{") |
| 1314 | |
| 1315 | def test_bad_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1316 | root, _ = _make_repo(tmp_path) |
| 1317 | result = _invoke(root, ["bisect", "bad", "--help"]) |
| 1318 | assert "Agent quickstart" in result.output or "regression" in result.output.lower() |
| 1319 | |
| 1320 | def test_bad_advances_bisect_log(self, tmp_path: pathlib.Path) -> None: |
| 1321 | """After marking bad, the bisect log records the verdict.""" |
| 1322 | from muse.core.bisect import _load_state |
| 1323 | root, repo_id = _make_repo(tmp_path) |
| 1324 | ids = _build_chain(root, repo_id, 6) |
| 1325 | self._start(root, ids) |
| 1326 | mid = ids[len(ids) // 2] |
| 1327 | _invoke(root, ["bisect", "bad", mid]) |
| 1328 | state = _load_state(root) |
| 1329 | assert state is not None |
| 1330 | assert any("bad" in entry for entry in state.get("log", [])) |
| 1331 | |
| 1332 | def test_bad_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1333 | root, repo_id = _make_repo(tmp_path) |
| 1334 | ids = _build_chain(root, repo_id, 5) |
| 1335 | self._start(root, ids) |
| 1336 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1337 | assert result.exit_code == 0 |
| 1338 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1339 | |
| 1340 | def test_bad_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1341 | root, repo_id = _make_repo(tmp_path) |
| 1342 | ids = _build_chain(root, repo_id, 5) |
| 1343 | self._start(root, ids) |
| 1344 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1345 | assert result.exit_code == 0 |
| 1346 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1347 | |
| 1348 | |
| 1349 | class TestBisectBadSecurity: |
| 1350 | """Security hardening tests for muse bisect bad.""" |
| 1351 | |
| 1352 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1353 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1354 | assert r.exit_code == 0 |
| 1355 | |
| 1356 | def test_bad_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1357 | root, repo_id = _make_repo(tmp_path) |
| 1358 | ids = _build_chain(root, repo_id, 5) |
| 1359 | self._start(root, ids) |
| 1360 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1361 | assert result.exit_code == 0 |
| 1362 | d = json.loads(_json_blob(result.output)) |
| 1363 | assert isinstance(d, dict) |
| 1364 | |
| 1365 | def test_bad_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1366 | root, repo_id = _make_repo(tmp_path) |
| 1367 | ids = _build_chain(root, repo_id, 5) |
| 1368 | self._start(root, ids) |
| 1369 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1370 | assert result.exit_code == 0 |
| 1371 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1372 | |
| 1373 | def test_bad_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1374 | """ANSI in symbol_changes entries stripped from JSON output.""" |
| 1375 | from unittest.mock import patch |
| 1376 | from muse.core.bisect import BisectResult |
| 1377 | root, repo_id = _make_repo(tmp_path) |
| 1378 | ids = _build_chain(root, repo_id, 5) |
| 1379 | self._start(root, ids) |
| 1380 | injected = BisectResult( |
| 1381 | done=False, |
| 1382 | first_bad=None, |
| 1383 | next_to_test=ids[2], |
| 1384 | remaining_count=2, |
| 1385 | steps_remaining=1, |
| 1386 | verdict="bad", |
| 1387 | symbol_changes=["modify func\x1b[31mred\x1b[0m"], |
| 1388 | ) |
| 1389 | with patch("muse.cli.commands.bisect.mark_bad", return_value=injected): |
| 1390 | result = _invoke(root, ["bisect", "bad", ids[2], "--json"]) |
| 1391 | assert "\x1b" not in result.output |
| 1392 | |
| 1393 | def test_bad_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1394 | """ANSI in symbol_changes entries stripped from text output.""" |
| 1395 | from unittest.mock import patch |
| 1396 | from muse.core.bisect import BisectResult |
| 1397 | root, repo_id = _make_repo(tmp_path) |
| 1398 | ids = _build_chain(root, repo_id, 5) |
| 1399 | self._start(root, ids) |
| 1400 | injected = BisectResult( |
| 1401 | done=False, |
| 1402 | first_bad=None, |
| 1403 | next_to_test=ids[2], |
| 1404 | remaining_count=2, |
| 1405 | steps_remaining=1, |
| 1406 | verdict="bad", |
| 1407 | symbol_changes=["modify func\x1b[31mred\x1b[0m"], |
| 1408 | ) |
| 1409 | with patch("muse.cli.commands.bisect.mark_bad", return_value=injected): |
| 1410 | result = _invoke(root, ["bisect", "bad", ids[2]]) |
| 1411 | assert "\x1b" not in result.output |
| 1412 | |
| 1413 | def test_bad_error_output_to_stderr_not_stdout(self, tmp_path: pathlib.Path) -> None: |
| 1414 | """Error messages go to stderr; stdout is clean on failure.""" |
| 1415 | root, _ = _make_repo(tmp_path) |
| 1416 | result = _invoke(root, ["bisect", "bad"]) |
| 1417 | assert result.exit_code != 0 |
| 1418 | # CliRunner mixes stderr into output; verify no JSON object was emitted |
| 1419 | assert not result.output.strip().startswith("{") |
| 1420 | |
| 1421 | def test_bad_ansi_in_ref_does_not_leak_to_output(self, tmp_path: pathlib.Path) -> None: |
| 1422 | """Passing an ANSI-injected ref does not leak escape codes to stdout.""" |
| 1423 | root, repo_id = _make_repo(tmp_path) |
| 1424 | ids = _build_chain(root, repo_id, 4) |
| 1425 | self._start(root, ids) |
| 1426 | result = _invoke(root, ["bisect", "bad", "\x1b[31mHEAD\x1b[0m"]) |
| 1427 | # Will fail (ref not found) but must not echo ANSI to stdout |
| 1428 | assert "\x1b" not in result.output |
| 1429 | |
| 1430 | |
| 1431 | class TestBisectBadStress: |
| 1432 | """Performance and scale tests for muse bisect bad.""" |
| 1433 | |
| 1434 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1435 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1436 | assert r.exit_code == 0 |
| 1437 | |
| 1438 | def test_bad_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1439 | """Marking bad on a 100-commit session exits 0 and advances the search.""" |
| 1440 | root, repo_id = _make_repo(tmp_path) |
| 1441 | ids = _build_chain(root, repo_id, 100) |
| 1442 | self._start(root, ids) |
| 1443 | result = _invoke(root, ["bisect", "bad", ids[50], "--json"]) |
| 1444 | assert result.exit_code == 0 |
| 1445 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1446 | |
| 1447 | def test_bad_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1448 | """Marking bad on a 100-commit session completes within 5 seconds.""" |
| 1449 | import time |
| 1450 | root, repo_id = _make_repo(tmp_path) |
| 1451 | ids = _build_chain(root, repo_id, 100) |
| 1452 | self._start(root, ids) |
| 1453 | t0 = time.monotonic() |
| 1454 | result = _invoke(root, ["bisect", "bad", ids[50], "--json"]) |
| 1455 | elapsed = time.monotonic() - t0 |
| 1456 | assert result.exit_code == 0 |
| 1457 | assert elapsed < 5.0, f"bisect bad on 100 commits took {elapsed:.2f}s" |
| 1458 | |
| 1459 | def test_bad_converges_full_session(self, tmp_path: pathlib.Path) -> None: |
| 1460 | """Marking next_to_test as bad on every step converges within log2(20) steps.""" |
| 1461 | root, repo_id = _make_repo(tmp_path) |
| 1462 | ids = _build_chain(root, repo_id, 20) |
| 1463 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1464 | assert r.exit_code == 0 |
| 1465 | parsed = _parse_step(r.output) |
| 1466 | done = parsed["done"] |
| 1467 | for _ in range(10): |
| 1468 | if done: |
| 1469 | break |
| 1470 | nxt = parsed["next_to_test"] |
| 1471 | assert nxt is not None |
| 1472 | next_r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 1473 | assert next_r.exit_code == 0 |
| 1474 | parsed = _parse_step(next_r.output) |
| 1475 | done = parsed["done"] |
| 1476 | assert done, "bisect did not converge within 10 bad steps on 20-commit chain" |
| 1477 | |
| 1478 | |
| 1479 | # --------------------------------------------------------------------------- |
| 1480 | # bisect good — Extended, Security, Stress |
| 1481 | # --------------------------------------------------------------------------- |
| 1482 | |
| 1483 | |
| 1484 | class TestBisectGoodExtended: |
| 1485 | """Extended unit / integration / e2e tests for muse bisect good.""" |
| 1486 | |
| 1487 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1488 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1489 | assert r.exit_code == 0 |
| 1490 | |
| 1491 | def test_good_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1492 | root, repo_id = _make_repo(tmp_path) |
| 1493 | ids = _build_chain(root, repo_id, 6) |
| 1494 | self._start(root, ids) |
| 1495 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1496 | assert result.exit_code == 0 |
| 1497 | |
| 1498 | def test_good_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1499 | root, repo_id = _make_repo(tmp_path) |
| 1500 | ids = _build_chain(root, repo_id, 6) |
| 1501 | self._start(root, ids) |
| 1502 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "-j"]) |
| 1503 | assert result.exit_code == 0 |
| 1504 | assert _parse_step(result.output)["verdict"] == "good" |
| 1505 | |
| 1506 | def test_good_json_verdict_is_good(self, tmp_path: pathlib.Path) -> None: |
| 1507 | root, repo_id = _make_repo(tmp_path) |
| 1508 | ids = _build_chain(root, repo_id, 6) |
| 1509 | self._start(root, ids) |
| 1510 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1511 | assert result.exit_code == 0 |
| 1512 | assert _parse_step(result.output)["verdict"] == "good" |
| 1513 | |
| 1514 | def test_good_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1515 | root, repo_id = _make_repo(tmp_path) |
| 1516 | ids = _build_chain(root, repo_id, 6) |
| 1517 | self._start(root, ids) |
| 1518 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1519 | assert result.exit_code == 0 |
| 1520 | d = json.loads(_json_blob(result.output)) |
| 1521 | assert set(d.keys()) == {"done", "first_bad", "next_to_test", "remaining_count", |
| 1522 | "steps_remaining", "verdict", "symbol_changes"} |
| 1523 | |
| 1524 | def test_good_reduces_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1525 | root, repo_id = _make_repo(tmp_path) |
| 1526 | ids = _build_chain(root, repo_id, 10) |
| 1527 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1528 | before = _parse_step(r.output)["remaining_count"] |
| 1529 | mid = _parse_step(r.output)["next_to_test"] |
| 1530 | result = _invoke(root, ["bisect", "good", mid, "--json"]) |
| 1531 | assert result.exit_code == 0 |
| 1532 | assert _parse_step(result.output)["remaining_count"] < before |
| 1533 | |
| 1534 | def test_good_done_true_when_isolated(self, tmp_path: pathlib.Path) -> None: |
| 1535 | """Marking the only remaining commit good isolates first bad immediately.""" |
| 1536 | root, repo_id = _make_repo(tmp_path) |
| 1537 | ids = _build_chain(root, repo_id, 3) |
| 1538 | # good=ids[0], bad=ids[2]: ids[1] is the midpoint; marking it good resolves |
| 1539 | self._start(root, ids) |
| 1540 | result = _invoke(root, ["bisect", "good", ids[1], "--json"]) |
| 1541 | assert result.exit_code == 0 |
| 1542 | parsed = _parse_step(result.output) |
| 1543 | assert parsed["done"] is True |
| 1544 | assert parsed["first_bad"] == ids[2] |
| 1545 | |
| 1546 | def test_good_first_bad_set_when_done(self, tmp_path: pathlib.Path) -> None: |
| 1547 | root, repo_id = _make_repo(tmp_path) |
| 1548 | ids = _build_chain(root, repo_id, 3) |
| 1549 | self._start(root, ids) |
| 1550 | result = _invoke(root, ["bisect", "good", ids[1], "--json"]) |
| 1551 | assert result.exit_code == 0 |
| 1552 | parsed = _parse_step(result.output) |
| 1553 | assert parsed["done"] is True |
| 1554 | assert isinstance(parsed["first_bad"], str) |
| 1555 | |
| 1556 | def test_good_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1557 | root, repo_id = _make_repo(tmp_path) |
| 1558 | ids = _build_chain(root, repo_id, 5) |
| 1559 | self._start(root, ids) |
| 1560 | # HEAD is ids[-1] (known bad); marking it good is legal but pushes bad boundary |
| 1561 | result = _invoke(root, ["bisect", "good", "--json"]) |
| 1562 | assert result.exit_code == 0 |
| 1563 | |
| 1564 | def test_good_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1565 | root, _ = _make_repo(tmp_path) |
| 1566 | result = _invoke(root, ["bisect", "good"]) |
| 1567 | assert result.exit_code == 1 |
| 1568 | |
| 1569 | def test_good_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 1570 | root, _ = _make_repo(tmp_path) |
| 1571 | result = _invoke(root, ["bisect", "good"]) |
| 1572 | combined = result.output + (result.stderr or "") |
| 1573 | assert "No bisect session" in combined |
| 1574 | |
| 1575 | def test_good_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1576 | empty = tmp_path / "not_a_repo" |
| 1577 | empty.mkdir() |
| 1578 | result = _invoke(empty, ["bisect", "good"]) |
| 1579 | assert result.exit_code == 2 |
| 1580 | |
| 1581 | def test_good_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1582 | root, repo_id = _make_repo(tmp_path) |
| 1583 | ids = _build_chain(root, repo_id, 4) |
| 1584 | self._start(root, ids) |
| 1585 | result = _invoke(root, ["bisect", "good", "deadbeef_nonexistent"]) |
| 1586 | assert result.exit_code == 1 |
| 1587 | |
| 1588 | def test_good_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1589 | root, repo_id = _make_repo(tmp_path) |
| 1590 | ids = _build_chain(root, repo_id, 5) |
| 1591 | self._start(root, ids) |
| 1592 | mid = ids[len(ids) // 2] |
| 1593 | result = _invoke(root, ["bisect", "good", mid]) |
| 1594 | assert result.exit_code == 0 |
| 1595 | assert mid[:12] in result.output |
| 1596 | |
| 1597 | def test_good_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1598 | root, repo_id = _make_repo(tmp_path) |
| 1599 | ids = _build_chain(root, repo_id, 5) |
| 1600 | self._start(root, ids) |
| 1601 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1602 | assert result.exit_code == 0 |
| 1603 | assert not result.output.strip().startswith("{") |
| 1604 | |
| 1605 | def test_good_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1606 | root, _ = _make_repo(tmp_path) |
| 1607 | result = _invoke(root, ["bisect", "good", "--help"]) |
| 1608 | assert "Agent quickstart" in result.output or "regression" in result.output.lower() |
| 1609 | |
| 1610 | def test_good_advances_bisect_log(self, tmp_path: pathlib.Path) -> None: |
| 1611 | from muse.core.bisect import _load_state |
| 1612 | root, repo_id = _make_repo(tmp_path) |
| 1613 | ids = _build_chain(root, repo_id, 6) |
| 1614 | self._start(root, ids) |
| 1615 | _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1616 | state = _load_state(root) |
| 1617 | assert state is not None |
| 1618 | assert any("good" in entry for entry in state.get("log", [])) |
| 1619 | |
| 1620 | def test_good_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1621 | root, repo_id = _make_repo(tmp_path) |
| 1622 | ids = _build_chain(root, repo_id, 5) |
| 1623 | self._start(root, ids) |
| 1624 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1625 | assert result.exit_code == 0 |
| 1626 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1627 | |
| 1628 | def test_good_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1629 | root, repo_id = _make_repo(tmp_path) |
| 1630 | ids = _build_chain(root, repo_id, 5) |
| 1631 | self._start(root, ids) |
| 1632 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1633 | assert result.exit_code == 0 |
| 1634 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1635 | |
| 1636 | |
| 1637 | class TestBisectGoodSecurity: |
| 1638 | """Security hardening tests for muse bisect good.""" |
| 1639 | |
| 1640 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1641 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1642 | assert r.exit_code == 0 |
| 1643 | |
| 1644 | def test_good_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1645 | root, repo_id = _make_repo(tmp_path) |
| 1646 | ids = _build_chain(root, repo_id, 5) |
| 1647 | self._start(root, ids) |
| 1648 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1649 | assert result.exit_code == 0 |
| 1650 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 1651 | |
| 1652 | def test_good_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1653 | root, repo_id = _make_repo(tmp_path) |
| 1654 | ids = _build_chain(root, repo_id, 5) |
| 1655 | self._start(root, ids) |
| 1656 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1657 | assert result.exit_code == 0 |
| 1658 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1659 | |
| 1660 | def test_good_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1661 | from unittest.mock import patch |
| 1662 | from muse.core.bisect import BisectResult |
| 1663 | root, repo_id = _make_repo(tmp_path) |
| 1664 | ids = _build_chain(root, repo_id, 5) |
| 1665 | self._start(root, ids) |
| 1666 | injected = BisectResult( |
| 1667 | done=False, |
| 1668 | first_bad=None, |
| 1669 | next_to_test=ids[2], |
| 1670 | remaining_count=2, |
| 1671 | steps_remaining=1, |
| 1672 | verdict="good", |
| 1673 | symbol_changes=["add func\x1b[32mgreen\x1b[0m"], |
| 1674 | ) |
| 1675 | with patch("muse.cli.commands.bisect.mark_good", return_value=injected): |
| 1676 | result = _invoke(root, ["bisect", "good", ids[2], "--json"]) |
| 1677 | assert "\x1b" not in result.output |
| 1678 | |
| 1679 | def test_good_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1680 | from unittest.mock import patch |
| 1681 | from muse.core.bisect import BisectResult |
| 1682 | root, repo_id = _make_repo(tmp_path) |
| 1683 | ids = _build_chain(root, repo_id, 5) |
| 1684 | self._start(root, ids) |
| 1685 | injected = BisectResult( |
| 1686 | done=False, |
| 1687 | first_bad=None, |
| 1688 | next_to_test=ids[2], |
| 1689 | remaining_count=2, |
| 1690 | steps_remaining=1, |
| 1691 | verdict="good", |
| 1692 | symbol_changes=["add func\x1b[32mgreen\x1b[0m"], |
| 1693 | ) |
| 1694 | with patch("muse.cli.commands.bisect.mark_good", return_value=injected): |
| 1695 | result = _invoke(root, ["bisect", "good", ids[2]]) |
| 1696 | assert "\x1b" not in result.output |
| 1697 | |
| 1698 | def test_good_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 1699 | root, _ = _make_repo(tmp_path) |
| 1700 | result = _invoke(root, ["bisect", "good"]) |
| 1701 | assert result.exit_code != 0 |
| 1702 | assert not result.output.strip().startswith("{") |
| 1703 | |
| 1704 | def test_good_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 1705 | root, repo_id = _make_repo(tmp_path) |
| 1706 | ids = _build_chain(root, repo_id, 4) |
| 1707 | self._start(root, ids) |
| 1708 | result = _invoke(root, ["bisect", "good", "\x1b[32mHEAD\x1b[0m"]) |
| 1709 | assert "\x1b" not in result.output |
| 1710 | |
| 1711 | |
| 1712 | class TestBisectGoodStress: |
| 1713 | """Performance and scale tests for muse bisect good.""" |
| 1714 | |
| 1715 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1716 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1717 | assert r.exit_code == 0 |
| 1718 | |
| 1719 | def test_good_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1720 | root, repo_id = _make_repo(tmp_path) |
| 1721 | ids = _build_chain(root, repo_id, 100) |
| 1722 | self._start(root, ids) |
| 1723 | result = _invoke(root, ["bisect", "good", ids[10], "--json"]) |
| 1724 | assert result.exit_code == 0 |
| 1725 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1726 | |
| 1727 | def test_good_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1728 | import time |
| 1729 | root, repo_id = _make_repo(tmp_path) |
| 1730 | ids = _build_chain(root, repo_id, 100) |
| 1731 | self._start(root, ids) |
| 1732 | t0 = time.monotonic() |
| 1733 | result = _invoke(root, ["bisect", "good", ids[10], "--json"]) |
| 1734 | elapsed = time.monotonic() - t0 |
| 1735 | assert result.exit_code == 0 |
| 1736 | assert elapsed < 5.0, f"bisect good on 100 commits took {elapsed:.2f}s" |
| 1737 | |
| 1738 | def test_good_converges_full_session(self, tmp_path: pathlib.Path) -> None: |
| 1739 | """Marking next_to_test as good on each step converges within log2(20) steps.""" |
| 1740 | root, repo_id = _make_repo(tmp_path) |
| 1741 | ids = _build_chain(root, repo_id, 20) |
| 1742 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1743 | assert r.exit_code == 0 |
| 1744 | parsed = _parse_step(r.output) |
| 1745 | done = parsed["done"] |
| 1746 | for _ in range(10): |
| 1747 | if done: |
| 1748 | break |
| 1749 | nxt = parsed["next_to_test"] |
| 1750 | assert nxt is not None |
| 1751 | next_r = _invoke(root, ["bisect", "good", nxt, "--json"]) |
| 1752 | assert next_r.exit_code == 0 |
| 1753 | parsed = _parse_step(next_r.output) |
| 1754 | done = parsed["done"] |
| 1755 | assert done, "bisect did not converge within 10 good steps on 20-commit chain" |
| 1756 | |
| 1757 | |
| 1758 | # --------------------------------------------------------------------------- |
| 1759 | # bisect skip — Extended, Security, Stress |
| 1760 | # --------------------------------------------------------------------------- |
| 1761 | |
| 1762 | |
| 1763 | class TestBisectSkipExtended: |
| 1764 | """Extended unit / integration / e2e tests for muse bisect skip.""" |
| 1765 | |
| 1766 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1767 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1768 | assert r.exit_code == 0 |
| 1769 | |
| 1770 | def test_skip_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1771 | root, repo_id = _make_repo(tmp_path) |
| 1772 | ids = _build_chain(root, repo_id, 6) |
| 1773 | self._start(root, ids) |
| 1774 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1775 | assert result.exit_code == 0 |
| 1776 | |
| 1777 | def test_skip_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1778 | root, repo_id = _make_repo(tmp_path) |
| 1779 | ids = _build_chain(root, repo_id, 6) |
| 1780 | self._start(root, ids) |
| 1781 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "-j"]) |
| 1782 | assert result.exit_code == 0 |
| 1783 | assert _parse_step(result.output)["verdict"] == "skip" |
| 1784 | |
| 1785 | def test_skip_json_verdict_is_skip(self, tmp_path: pathlib.Path) -> None: |
| 1786 | root, repo_id = _make_repo(tmp_path) |
| 1787 | ids = _build_chain(root, repo_id, 6) |
| 1788 | self._start(root, ids) |
| 1789 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1790 | assert result.exit_code == 0 |
| 1791 | assert _parse_step(result.output)["verdict"] == "skip" |
| 1792 | |
| 1793 | def test_skip_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1794 | root, repo_id = _make_repo(tmp_path) |
| 1795 | ids = _build_chain(root, repo_id, 6) |
| 1796 | self._start(root, ids) |
| 1797 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1798 | assert result.exit_code == 0 |
| 1799 | d = json.loads(_json_blob(result.output)) |
| 1800 | assert set(d.keys()) == {"done", "first_bad", "next_to_test", "remaining_count", |
| 1801 | "steps_remaining", "verdict", "symbol_changes"} |
| 1802 | |
| 1803 | def test_skip_removes_commit_from_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1804 | root, repo_id = _make_repo(tmp_path) |
| 1805 | ids = _build_chain(root, repo_id, 10) |
| 1806 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1807 | before = _parse_step(r.output)["remaining_count"] |
| 1808 | mid = _parse_step(r.output)["next_to_test"] |
| 1809 | result = _invoke(root, ["bisect", "skip", mid, "--json"]) |
| 1810 | assert result.exit_code == 0 |
| 1811 | assert _parse_step(result.output)["remaining_count"] < before |
| 1812 | |
| 1813 | def test_skip_persisted_in_state(self, tmp_path: pathlib.Path) -> None: |
| 1814 | from muse.core.bisect import _load_state |
| 1815 | root, repo_id = _make_repo(tmp_path) |
| 1816 | ids = _build_chain(root, repo_id, 6) |
| 1817 | self._start(root, ids) |
| 1818 | mid = ids[len(ids) // 2] |
| 1819 | _invoke(root, ["bisect", "skip", mid]) |
| 1820 | state = _load_state(root) |
| 1821 | assert state is not None |
| 1822 | assert mid in state.get("skipped_ids", []) |
| 1823 | |
| 1824 | def test_skip_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1825 | root, repo_id = _make_repo(tmp_path) |
| 1826 | ids = _build_chain(root, repo_id, 5) |
| 1827 | self._start(root, ids) |
| 1828 | result = _invoke(root, ["bisect", "skip", "--json"]) |
| 1829 | assert result.exit_code == 0 |
| 1830 | |
| 1831 | def test_skip_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1832 | root, _ = _make_repo(tmp_path) |
| 1833 | result = _invoke(root, ["bisect", "skip"]) |
| 1834 | assert result.exit_code == 1 |
| 1835 | |
| 1836 | def test_skip_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 1837 | root, _ = _make_repo(tmp_path) |
| 1838 | result = _invoke(root, ["bisect", "skip"]) |
| 1839 | combined = result.output + (result.stderr or "") |
| 1840 | assert "No bisect session" in combined |
| 1841 | |
| 1842 | def test_skip_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1843 | empty = tmp_path / "not_a_repo" |
| 1844 | empty.mkdir() |
| 1845 | result = _invoke(empty, ["bisect", "skip"]) |
| 1846 | assert result.exit_code == 2 |
| 1847 | |
| 1848 | def test_skip_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1849 | root, repo_id = _make_repo(tmp_path) |
| 1850 | ids = _build_chain(root, repo_id, 4) |
| 1851 | self._start(root, ids) |
| 1852 | result = _invoke(root, ["bisect", "skip", "deadbeef_nonexistent"]) |
| 1853 | assert result.exit_code == 1 |
| 1854 | |
| 1855 | def test_skip_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1856 | root, repo_id = _make_repo(tmp_path) |
| 1857 | ids = _build_chain(root, repo_id, 5) |
| 1858 | self._start(root, ids) |
| 1859 | mid = ids[len(ids) // 2] |
| 1860 | result = _invoke(root, ["bisect", "skip", mid]) |
| 1861 | assert result.exit_code == 0 |
| 1862 | assert mid[:12] in result.output |
| 1863 | |
| 1864 | def test_skip_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1865 | root, repo_id = _make_repo(tmp_path) |
| 1866 | ids = _build_chain(root, repo_id, 5) |
| 1867 | self._start(root, ids) |
| 1868 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1869 | assert result.exit_code == 0 |
| 1870 | assert not result.output.strip().startswith("{") |
| 1871 | |
| 1872 | def test_skip_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1873 | root, _ = _make_repo(tmp_path) |
| 1874 | result = _invoke(root, ["bisect", "skip", "--help"]) |
| 1875 | assert "Agent quickstart" in result.output or "125" in result.output |
| 1876 | |
| 1877 | def test_skip_advances_log(self, tmp_path: pathlib.Path) -> None: |
| 1878 | from muse.core.bisect import _load_state |
| 1879 | root, repo_id = _make_repo(tmp_path) |
| 1880 | ids = _build_chain(root, repo_id, 6) |
| 1881 | self._start(root, ids) |
| 1882 | _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1883 | state = _load_state(root) |
| 1884 | assert state is not None |
| 1885 | assert any("skip" in entry for entry in state.get("log", [])) |
| 1886 | |
| 1887 | def test_skip_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1888 | root, repo_id = _make_repo(tmp_path) |
| 1889 | ids = _build_chain(root, repo_id, 5) |
| 1890 | self._start(root, ids) |
| 1891 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1892 | assert result.exit_code == 0 |
| 1893 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1894 | |
| 1895 | def test_skip_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1896 | root, repo_id = _make_repo(tmp_path) |
| 1897 | ids = _build_chain(root, repo_id, 5) |
| 1898 | self._start(root, ids) |
| 1899 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1900 | assert result.exit_code == 0 |
| 1901 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1902 | |
| 1903 | def test_skip_multiple_commits(self, tmp_path: pathlib.Path) -> None: |
| 1904 | """Skipping several commits all land in skipped_ids.""" |
| 1905 | from muse.core.bisect import _load_state |
| 1906 | root, repo_id = _make_repo(tmp_path) |
| 1907 | ids = _build_chain(root, repo_id, 8) |
| 1908 | self._start(root, ids) |
| 1909 | for idx in (2, 3, 4): |
| 1910 | r = _invoke(root, ["bisect", "skip", ids[idx]]) |
| 1911 | assert r.exit_code == 0 |
| 1912 | state = _load_state(root) |
| 1913 | assert state is not None |
| 1914 | skipped = state.get("skipped_ids", []) |
| 1915 | assert all(ids[i] in skipped for i in (2, 3, 4)) |
| 1916 | |
| 1917 | |
| 1918 | class TestBisectSkipSecurity: |
| 1919 | """Security hardening tests for muse bisect skip.""" |
| 1920 | |
| 1921 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1922 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1923 | assert r.exit_code == 0 |
| 1924 | |
| 1925 | def test_skip_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1926 | root, repo_id = _make_repo(tmp_path) |
| 1927 | ids = _build_chain(root, repo_id, 5) |
| 1928 | self._start(root, ids) |
| 1929 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1930 | assert result.exit_code == 0 |
| 1931 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 1932 | |
| 1933 | def test_skip_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1934 | root, repo_id = _make_repo(tmp_path) |
| 1935 | ids = _build_chain(root, repo_id, 5) |
| 1936 | self._start(root, ids) |
| 1937 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1938 | assert result.exit_code == 0 |
| 1939 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1940 | |
| 1941 | def test_skip_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1942 | from unittest.mock import patch |
| 1943 | from muse.core.bisect import BisectResult |
| 1944 | root, repo_id = _make_repo(tmp_path) |
| 1945 | ids = _build_chain(root, repo_id, 5) |
| 1946 | self._start(root, ids) |
| 1947 | injected = BisectResult( |
| 1948 | done=False, |
| 1949 | first_bad=None, |
| 1950 | next_to_test=ids[2], |
| 1951 | remaining_count=2, |
| 1952 | steps_remaining=1, |
| 1953 | verdict="skip", |
| 1954 | symbol_changes=["modify func\x1b[33myellow\x1b[0m"], |
| 1955 | ) |
| 1956 | with patch("muse.cli.commands.bisect.skip_commit", return_value=injected): |
| 1957 | result = _invoke(root, ["bisect", "skip", ids[2], "--json"]) |
| 1958 | assert "\x1b" not in result.output |
| 1959 | |
| 1960 | def test_skip_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1961 | from unittest.mock import patch |
| 1962 | from muse.core.bisect import BisectResult |
| 1963 | root, repo_id = _make_repo(tmp_path) |
| 1964 | ids = _build_chain(root, repo_id, 5) |
| 1965 | self._start(root, ids) |
| 1966 | injected = BisectResult( |
| 1967 | done=False, |
| 1968 | first_bad=None, |
| 1969 | next_to_test=ids[2], |
| 1970 | remaining_count=2, |
| 1971 | steps_remaining=1, |
| 1972 | verdict="skip", |
| 1973 | symbol_changes=["modify func\x1b[33myellow\x1b[0m"], |
| 1974 | ) |
| 1975 | with patch("muse.cli.commands.bisect.skip_commit", return_value=injected): |
| 1976 | result = _invoke(root, ["bisect", "skip", ids[2]]) |
| 1977 | assert "\x1b" not in result.output |
| 1978 | |
| 1979 | def test_skip_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 1980 | root, _ = _make_repo(tmp_path) |
| 1981 | result = _invoke(root, ["bisect", "skip"]) |
| 1982 | assert result.exit_code != 0 |
| 1983 | assert not result.output.strip().startswith("{") |
| 1984 | |
| 1985 | def test_skip_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 1986 | root, repo_id = _make_repo(tmp_path) |
| 1987 | ids = _build_chain(root, repo_id, 4) |
| 1988 | self._start(root, ids) |
| 1989 | result = _invoke(root, ["bisect", "skip", "\x1b[33mHEAD\x1b[0m"]) |
| 1990 | assert "\x1b" not in result.output |
| 1991 | |
| 1992 | |
| 1993 | class TestBisectSkipStress: |
| 1994 | """Performance and scale tests for muse bisect skip.""" |
| 1995 | |
| 1996 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1997 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1998 | assert r.exit_code == 0 |
| 1999 | |
| 2000 | def test_skip_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 2001 | root, repo_id = _make_repo(tmp_path) |
| 2002 | ids = _build_chain(root, repo_id, 100) |
| 2003 | self._start(root, ids) |
| 2004 | result = _invoke(root, ["bisect", "skip", ids[50], "--json"]) |
| 2005 | assert result.exit_code == 0 |
| 2006 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 2007 | |
| 2008 | def test_skip_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 2009 | import time |
| 2010 | root, repo_id = _make_repo(tmp_path) |
| 2011 | ids = _build_chain(root, repo_id, 100) |
| 2012 | self._start(root, ids) |
| 2013 | t0 = time.monotonic() |
| 2014 | result = _invoke(root, ["bisect", "skip", ids[50], "--json"]) |
| 2015 | elapsed = time.monotonic() - t0 |
| 2016 | assert result.exit_code == 0 |
| 2017 | assert elapsed < 5.0, f"bisect skip on 100 commits took {elapsed:.2f}s" |
| 2018 | |
| 2019 | def test_skip_reduces_remaining_monotonically(self, tmp_path: pathlib.Path) -> None: |
| 2020 | """Each consecutive skip reduces remaining_count (non-increasing sequence).""" |
| 2021 | root, repo_id = _make_repo(tmp_path) |
| 2022 | ids = _build_chain(root, repo_id, 20) |
| 2023 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 2024 | assert r.exit_code == 0 |
| 2025 | counts = [_parse_step(r.output)["remaining_count"]] |
| 2026 | cur = r |
| 2027 | for _ in range(5): |
| 2028 | parsed = _parse_step(cur.output) |
| 2029 | if parsed["done"] or parsed["next_to_test"] is None: |
| 2030 | break |
| 2031 | nxt = parsed["next_to_test"] |
| 2032 | cur = _invoke(root, ["bisect", "skip", nxt, "--json"]) |
| 2033 | assert cur.exit_code == 0 |
| 2034 | counts.append(_parse_step(cur.output)["remaining_count"]) |
| 2035 | assert all(counts[i] >= counts[i + 1] for i in range(len(counts) - 1)) |
| 2036 | |
| 2037 | |
| 2038 | # --------------------------------------------------------------------------- |
| 2039 | # bisect run — Extended, Security, Stress |
| 2040 | # --------------------------------------------------------------------------- |
| 2041 | |
| 2042 | |
| 2043 | class TestBisectRunExtended: |
| 2044 | """Extended unit / integration / e2e tests for muse bisect run.""" |
| 2045 | |
| 2046 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2047 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2048 | assert r.exit_code == 0 |
| 2049 | |
| 2050 | def test_run_exits_0_with_true(self, tmp_path: pathlib.Path) -> None: |
| 2051 | root, repo_id = _make_repo(tmp_path) |
| 2052 | ids = _build_chain(root, repo_id, 5) |
| 2053 | self._start(root, ids) |
| 2054 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2055 | assert result.exit_code == 0 |
| 2056 | |
| 2057 | def test_run_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2058 | root, repo_id = _make_repo(tmp_path) |
| 2059 | ids = _build_chain(root, repo_id, 5) |
| 2060 | self._start(root, ids) |
| 2061 | result = _invoke(root, ["bisect", "run", "true", "-j"]) |
| 2062 | assert result.exit_code == 0 |
| 2063 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2064 | assert len(lines) >= 1 |
| 2065 | done_raw = json.loads(lines[-1]) |
| 2066 | assert done_raw["done"] is True |
| 2067 | |
| 2068 | def test_run_json_ndjson_step_keys(self, tmp_path: pathlib.Path) -> None: |
| 2069 | root, repo_id = _make_repo(tmp_path) |
| 2070 | ids = _build_chain(root, repo_id, 6) |
| 2071 | self._start(root, ids) |
| 2072 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2073 | assert result.exit_code == 0 |
| 2074 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2075 | if len(lines) > 1: |
| 2076 | step = json.loads(lines[0]) |
| 2077 | assert set(step.keys()) == {"step", "testing", "verdict", "remaining_count", "done"} |
| 2078 | |
| 2079 | def test_run_json_done_line_keys(self, tmp_path: pathlib.Path) -> None: |
| 2080 | root, repo_id = _make_repo(tmp_path) |
| 2081 | ids = _build_chain(root, repo_id, 5) |
| 2082 | self._start(root, ids) |
| 2083 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2084 | assert result.exit_code == 0 |
| 2085 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2086 | done = json.loads(lines[-1]) |
| 2087 | assert set(done.keys()) == {"done", "first_bad", "steps_taken"} |
| 2088 | |
| 2089 | def test_run_json_done_true_on_last_line(self, tmp_path: pathlib.Path) -> None: |
| 2090 | root, repo_id = _make_repo(tmp_path) |
| 2091 | ids = _build_chain(root, repo_id, 5) |
| 2092 | self._start(root, ids) |
| 2093 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2094 | assert result.exit_code == 0 |
| 2095 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2096 | assert json.loads(lines[-1])["done"] is True |
| 2097 | |
| 2098 | def test_run_json_steps_taken_positive(self, tmp_path: pathlib.Path) -> None: |
| 2099 | root, repo_id = _make_repo(tmp_path) |
| 2100 | ids = _build_chain(root, repo_id, 6) |
| 2101 | self._start(root, ids) |
| 2102 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2103 | assert result.exit_code == 0 |
| 2104 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2105 | assert json.loads(lines[-1])["steps_taken"] >= 1 |
| 2106 | |
| 2107 | def test_run_json_verdict_good_with_true(self, tmp_path: pathlib.Path) -> None: |
| 2108 | root, repo_id = _make_repo(tmp_path) |
| 2109 | ids = _build_chain(root, repo_id, 5) |
| 2110 | self._start(root, ids) |
| 2111 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2112 | assert result.exit_code == 0 |
| 2113 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2114 | step_lines = lines[:-1] |
| 2115 | assert all(json.loads(l)["verdict"] == "good" for l in step_lines) |
| 2116 | |
| 2117 | def test_run_json_verdict_bad_with_false(self, tmp_path: pathlib.Path) -> None: |
| 2118 | root, repo_id = _make_repo(tmp_path) |
| 2119 | ids = _build_chain(root, repo_id, 5) |
| 2120 | self._start(root, ids) |
| 2121 | result = _invoke(root, ["bisect", "run", "false", "--json"]) |
| 2122 | assert result.exit_code == 0 |
| 2123 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2124 | step_lines = lines[:-1] |
| 2125 | assert all(json.loads(l)["verdict"] == "bad" for l in step_lines) |
| 2126 | |
| 2127 | def test_run_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 2128 | root, _ = _make_repo(tmp_path) |
| 2129 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2130 | assert result.exit_code == 1 |
| 2131 | |
| 2132 | def test_run_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 2133 | root, _ = _make_repo(tmp_path) |
| 2134 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2135 | combined = result.output + (result.stderr or "") |
| 2136 | assert "No bisect session" in combined |
| 2137 | |
| 2138 | def test_run_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2139 | empty = tmp_path / "not_a_repo" |
| 2140 | empty.mkdir() |
| 2141 | result = _invoke(empty, ["bisect", "run", "true"]) |
| 2142 | assert result.exit_code == 2 |
| 2143 | |
| 2144 | def test_run_text_mentions_testing(self, tmp_path: pathlib.Path) -> None: |
| 2145 | root, repo_id = _make_repo(tmp_path) |
| 2146 | ids = _build_chain(root, repo_id, 5) |
| 2147 | self._start(root, ids) |
| 2148 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2149 | assert result.exit_code == 0 |
| 2150 | assert "Testing" in result.output or "→" in result.output |
| 2151 | |
| 2152 | def test_run_text_mentions_first_bad(self, tmp_path: pathlib.Path) -> None: |
| 2153 | root, repo_id = _make_repo(tmp_path) |
| 2154 | ids = _build_chain(root, repo_id, 5) |
| 2155 | self._start(root, ids) |
| 2156 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2157 | assert result.exit_code == 0 |
| 2158 | assert "First bad commit" in result.output or "Bisect complete" in result.output |
| 2159 | |
| 2160 | def test_run_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2161 | root, _ = _make_repo(tmp_path) |
| 2162 | result = _invoke(root, ["bisect", "run", "--help"]) |
| 2163 | assert "Agent quickstart" in result.output or "125" in result.output |
| 2164 | |
| 2165 | def test_run_json_step_numbers_increment(self, tmp_path: pathlib.Path) -> None: |
| 2166 | root, repo_id = _make_repo(tmp_path) |
| 2167 | ids = _build_chain(root, repo_id, 8) |
| 2168 | self._start(root, ids) |
| 2169 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2170 | assert result.exit_code == 0 |
| 2171 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2172 | step_nums = [json.loads(l)["step"] for l in lines[:-1]] |
| 2173 | assert step_nums == list(range(1, len(step_nums) + 1)) |
| 2174 | |
| 2175 | def test_run_json_remaining_nonincreasing(self, tmp_path: pathlib.Path) -> None: |
| 2176 | root, repo_id = _make_repo(tmp_path) |
| 2177 | ids = _build_chain(root, repo_id, 8) |
| 2178 | self._start(root, ids) |
| 2179 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2180 | assert result.exit_code == 0 |
| 2181 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2182 | counts = [json.loads(l)["remaining_count"] for l in lines[:-1]] |
| 2183 | assert all(counts[i] >= counts[i + 1] for i in range(len(counts) - 1)) |
| 2184 | |
| 2185 | def test_run_text_no_json_by_default(self, tmp_path: pathlib.Path) -> None: |
| 2186 | root, repo_id = _make_repo(tmp_path) |
| 2187 | ids = _build_chain(root, repo_id, 4) |
| 2188 | self._start(root, ids) |
| 2189 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2190 | assert result.exit_code == 0 |
| 2191 | # Text mode should not have a JSON object on a line by itself |
| 2192 | json_lines = [l for l in result.output.splitlines() if l.strip().startswith("{")] |
| 2193 | assert json_lines == [] |
| 2194 | |
| 2195 | def test_run_json_first_bad_set_on_done(self, tmp_path: pathlib.Path) -> None: |
| 2196 | root, repo_id = _make_repo(tmp_path) |
| 2197 | ids = _build_chain(root, repo_id, 5) |
| 2198 | self._start(root, ids) |
| 2199 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2200 | assert result.exit_code == 0 |
| 2201 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2202 | done = json.loads(lines[-1]) |
| 2203 | if done["done"]: |
| 2204 | assert done["first_bad"] is not None |
| 2205 | |
| 2206 | |
| 2207 | class TestBisectRunSecurity: |
| 2208 | """Security hardening tests for muse bisect run.""" |
| 2209 | |
| 2210 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2211 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2212 | assert r.exit_code == 0 |
| 2213 | |
| 2214 | def test_run_json_lines_are_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2215 | root, repo_id = _make_repo(tmp_path) |
| 2216 | ids = _build_chain(root, repo_id, 5) |
| 2217 | self._start(root, ids) |
| 2218 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2219 | assert result.exit_code == 0 |
| 2220 | for line in result.output.strip().splitlines(): |
| 2221 | if line.strip(): |
| 2222 | assert isinstance(json.loads(line.strip()), dict) |
| 2223 | |
| 2224 | def test_run_json_done_field_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2225 | root, repo_id = _make_repo(tmp_path) |
| 2226 | ids = _build_chain(root, repo_id, 5) |
| 2227 | self._start(root, ids) |
| 2228 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2229 | assert result.exit_code == 0 |
| 2230 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2231 | for line in lines: |
| 2232 | assert isinstance(json.loads(line)["done"], bool) |
| 2233 | |
| 2234 | def test_run_text_symbol_changes_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2235 | """ANSI codes in symbol_changes are stripped from text output during run.""" |
| 2236 | from unittest.mock import patch |
| 2237 | from muse.core.bisect import BisectResult |
| 2238 | root, repo_id = _make_repo(tmp_path) |
| 2239 | ids = _build_chain(root, repo_id, 5) |
| 2240 | self._start(root, ids) |
| 2241 | injected_result = BisectResult( |
| 2242 | done=True, |
| 2243 | first_bad=ids[2], |
| 2244 | next_to_test=None, |
| 2245 | remaining_count=0, |
| 2246 | steps_remaining=0, |
| 2247 | verdict="bad", |
| 2248 | symbol_changes=[], |
| 2249 | ) |
| 2250 | with patch("muse.cli.commands.bisect._symbol_ops_in_commit", |
| 2251 | return_value=["add func\x1b[31mred\x1b[0m"]), \ |
| 2252 | patch("muse.cli.commands.bisect.get_bisect_next", |
| 2253 | side_effect=[(ids[2], "billing.py::Invoice"), (None, "")]), \ |
| 2254 | patch("muse.cli.commands.bisect.run_bisect_command", |
| 2255 | return_value=injected_result): |
| 2256 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2257 | assert "\x1b" not in result.output |
| 2258 | |
| 2259 | def test_run_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 2260 | root, _ = _make_repo(tmp_path) |
| 2261 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2262 | assert result.exit_code != 0 |
| 2263 | assert not result.output.strip().startswith("{") |
| 2264 | |
| 2265 | def test_run_json_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2266 | root, repo_id = _make_repo(tmp_path) |
| 2267 | ids = _build_chain(root, repo_id, 5) |
| 2268 | self._start(root, ids) |
| 2269 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2270 | assert result.exit_code == 0 |
| 2271 | assert "\x1b" not in result.output |
| 2272 | |
| 2273 | def test_run_text_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2274 | root, repo_id = _make_repo(tmp_path) |
| 2275 | ids = _build_chain(root, repo_id, 5) |
| 2276 | self._start(root, ids) |
| 2277 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2278 | assert result.exit_code == 0 |
| 2279 | assert "\x1b" not in result.output |
| 2280 | |
| 2281 | |
| 2282 | class TestBisectRunStress: |
| 2283 | """Performance and scale tests for muse bisect run.""" |
| 2284 | |
| 2285 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2286 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2287 | assert r.exit_code == 0 |
| 2288 | |
| 2289 | def test_run_50_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 2290 | """run converges on a 50-commit chain with always-good command.""" |
| 2291 | root, repo_id = _make_repo(tmp_path) |
| 2292 | ids = _build_chain(root, repo_id, 50) |
| 2293 | self._start(root, ids) |
| 2294 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2295 | assert result.exit_code == 0 |
| 2296 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2297 | assert json.loads(lines[-1])["done"] is True |
| 2298 | |
| 2299 | def test_run_performance_20_commits(self, tmp_path: pathlib.Path) -> None: |
| 2300 | """run over 20 commits completes within 10 seconds.""" |
| 2301 | import time |
| 2302 | root, repo_id = _make_repo(tmp_path) |
| 2303 | ids = _build_chain(root, repo_id, 20) |
| 2304 | self._start(root, ids) |
| 2305 | t0 = time.monotonic() |
| 2306 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2307 | elapsed = time.monotonic() - t0 |
| 2308 | assert result.exit_code == 0 |
| 2309 | assert elapsed < 10.0, f"bisect run 20 commits took {elapsed:.2f}s" |
| 2310 | |
| 2311 | def test_run_steps_taken_within_log2(self, tmp_path: pathlib.Path) -> None: |
| 2312 | """Steps taken should be at most log2(n)+1 for an always-good command.""" |
| 2313 | import math |
| 2314 | root, repo_id = _make_repo(tmp_path) |
| 2315 | n = 32 |
| 2316 | ids = _build_chain(root, repo_id, n) |
| 2317 | self._start(root, ids) |
| 2318 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2319 | assert result.exit_code == 0 |
| 2320 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2321 | steps_taken = json.loads(lines[-1])["steps_taken"] |
| 2322 | assert steps_taken <= int(math.log2(n)) + 2 |
| 2323 | |
| 2324 | |
| 2325 | # --------------------------------------------------------------------------- |
| 2326 | # bisect log — Extended, Security, Stress |
| 2327 | # --------------------------------------------------------------------------- |
| 2328 | |
| 2329 | |
| 2330 | class TestBisectLogExtended: |
| 2331 | """Extended unit / integration / e2e tests for muse bisect log.""" |
| 2332 | |
| 2333 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2334 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2335 | assert r.exit_code == 0 |
| 2336 | |
| 2337 | def test_log_exits_0_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2338 | root, _ = _make_repo(tmp_path) |
| 2339 | result = _invoke(root, ["bisect", "log"]) |
| 2340 | assert result.exit_code == 0 |
| 2341 | |
| 2342 | def test_log_exits_0_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2343 | root, repo_id = _make_repo(tmp_path) |
| 2344 | ids = _build_chain(root, repo_id, 4) |
| 2345 | self._start(root, ids) |
| 2346 | result = _invoke(root, ["bisect", "log"]) |
| 2347 | assert result.exit_code == 0 |
| 2348 | |
| 2349 | def test_log_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2350 | root, _ = _make_repo(tmp_path) |
| 2351 | result = _invoke(root, ["bisect", "log", "-j"]) |
| 2352 | assert result.exit_code == 0 |
| 2353 | parsed = _parse_log(result.output) |
| 2354 | assert isinstance(parsed["active"], bool) |
| 2355 | |
| 2356 | def test_log_json_active_false_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2357 | root, _ = _make_repo(tmp_path) |
| 2358 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2359 | assert result.exit_code == 0 |
| 2360 | assert _parse_log(result.output)["active"] is False |
| 2361 | |
| 2362 | def test_log_json_active_true_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2363 | root, repo_id = _make_repo(tmp_path) |
| 2364 | ids = _build_chain(root, repo_id, 4) |
| 2365 | self._start(root, ids) |
| 2366 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2367 | assert result.exit_code == 0 |
| 2368 | assert _parse_log(result.output)["active"] is True |
| 2369 | |
| 2370 | def test_log_json_entries_empty_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2371 | root, _ = _make_repo(tmp_path) |
| 2372 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2373 | assert result.exit_code == 0 |
| 2374 | assert _parse_log(result.output)["entries"] == [] |
| 2375 | |
| 2376 | def test_log_json_entries_grow_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 2377 | root, repo_id = _make_repo(tmp_path) |
| 2378 | ids = _build_chain(root, repo_id, 6) |
| 2379 | self._start(root, ids) |
| 2380 | after_start = len(_parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"]) |
| 2381 | _invoke(root, ["bisect", "bad", ids[3]]) |
| 2382 | after_bad = len(_parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"]) |
| 2383 | assert after_bad > after_start |
| 2384 | |
| 2385 | def test_log_json_two_keys(self, tmp_path: pathlib.Path) -> None: |
| 2386 | root, _ = _make_repo(tmp_path) |
| 2387 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2388 | assert result.exit_code == 0 |
| 2389 | d = json.loads(_json_blob(result.output)) |
| 2390 | assert set(d.keys()) == {"active", "entries"} |
| 2391 | |
| 2392 | def test_log_json_start_records_bad_and_good(self, tmp_path: pathlib.Path) -> None: |
| 2393 | root, repo_id = _make_repo(tmp_path) |
| 2394 | ids = _build_chain(root, repo_id, 4) |
| 2395 | self._start(root, ids) |
| 2396 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2397 | verdicts = [e.split()[1] for e in entries if len(e.split()) >= 2] |
| 2398 | assert "bad" in verdicts |
| 2399 | assert "good" in verdicts |
| 2400 | |
| 2401 | def test_log_json_entries_contain_commit_ids(self, tmp_path: pathlib.Path) -> None: |
| 2402 | root, repo_id = _make_repo(tmp_path) |
| 2403 | ids = _build_chain(root, repo_id, 4) |
| 2404 | self._start(root, ids) |
| 2405 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2406 | for entry in entries: |
| 2407 | parts = entry.split() |
| 2408 | assert len(parts) >= 2 |
| 2409 | assert len(parts[0]) == 64 # full sha256 commit ID |
| 2410 | |
| 2411 | def test_log_json_entries_are_strings(self, tmp_path: pathlib.Path) -> None: |
| 2412 | root, repo_id = _make_repo(tmp_path) |
| 2413 | ids = _build_chain(root, repo_id, 3) |
| 2414 | self._start(root, ids) |
| 2415 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2416 | assert all(isinstance(e, str) for e in entries) |
| 2417 | |
| 2418 | def test_log_active_false_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2419 | root, repo_id = _make_repo(tmp_path) |
| 2420 | ids = _build_chain(root, repo_id, 3) |
| 2421 | self._start(root, ids) |
| 2422 | _invoke(root, ["bisect", "reset"]) |
| 2423 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2424 | assert result.exit_code == 0 |
| 2425 | assert _parse_log(result.output)["active"] is False |
| 2426 | |
| 2427 | def test_log_text_shows_bisect_log_header(self, tmp_path: pathlib.Path) -> None: |
| 2428 | root, repo_id = _make_repo(tmp_path) |
| 2429 | ids = _build_chain(root, repo_id, 4) |
| 2430 | self._start(root, ids) |
| 2431 | result = _invoke(root, ["bisect", "log"]) |
| 2432 | assert result.exit_code == 0 |
| 2433 | assert "Bisect log" in result.output |
| 2434 | |
| 2435 | def test_log_text_no_session_message(self, tmp_path: pathlib.Path) -> None: |
| 2436 | root, _ = _make_repo(tmp_path) |
| 2437 | result = _invoke(root, ["bisect", "log"]) |
| 2438 | assert result.exit_code == 0 |
| 2439 | assert "No bisect log" in result.output |
| 2440 | |
| 2441 | def test_log_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2442 | empty = tmp_path / "not_a_repo" |
| 2443 | empty.mkdir() |
| 2444 | result = _invoke(empty, ["bisect", "log"]) |
| 2445 | assert result.exit_code == 2 |
| 2446 | |
| 2447 | def test_log_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2448 | root, _ = _make_repo(tmp_path) |
| 2449 | result = _invoke(root, ["bisect", "log", "--help"]) |
| 2450 | assert "Agent quickstart" in result.output or "verdict" in result.output.lower() |
| 2451 | |
| 2452 | def test_log_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 2453 | root, repo_id = _make_repo(tmp_path) |
| 2454 | ids = _build_chain(root, repo_id, 4) |
| 2455 | self._start(root, ids) |
| 2456 | result = _invoke(root, ["bisect", "log"]) |
| 2457 | assert result.exit_code == 0 |
| 2458 | assert not any(l.strip().startswith("{") for l in result.output.splitlines()) |
| 2459 | |
| 2460 | |
| 2461 | class TestBisectLogSecurity: |
| 2462 | """Security hardening tests for muse bisect log.""" |
| 2463 | |
| 2464 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2465 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2466 | assert r.exit_code == 0 |
| 2467 | |
| 2468 | def test_log_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2469 | root, _ = _make_repo(tmp_path) |
| 2470 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2471 | assert result.exit_code == 0 |
| 2472 | d = json.loads(_json_blob(result.output)) |
| 2473 | assert isinstance(d, dict) |
| 2474 | |
| 2475 | def test_log_json_active_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2476 | root, _ = _make_repo(tmp_path) |
| 2477 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2478 | assert result.exit_code == 0 |
| 2479 | assert isinstance(json.loads(_json_blob(result.output))["active"], bool) |
| 2480 | |
| 2481 | def test_log_json_entries_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2482 | """ANSI codes injected into the log state are stripped from JSON output.""" |
| 2483 | from muse.core.bisect import _load_state, _save_state |
| 2484 | root, repo_id = _make_repo(tmp_path) |
| 2485 | ids = _build_chain(root, repo_id, 3) |
| 2486 | self._start(root, ids) |
| 2487 | # Tamper: inject ANSI into a log entry |
| 2488 | state = _load_state(root) |
| 2489 | assert state is not None |
| 2490 | state["log"].append(f"{ids[1]} bad\x1b[31m 2026-01-01T00:00:00\x1b[0m") |
| 2491 | _save_state(root, state) |
| 2492 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2493 | assert result.exit_code == 0 |
| 2494 | assert "\x1b" not in result.output |
| 2495 | |
| 2496 | def test_log_text_entries_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2497 | """ANSI codes in log entries are stripped from text output.""" |
| 2498 | from muse.core.bisect import _load_state, _save_state |
| 2499 | root, repo_id = _make_repo(tmp_path) |
| 2500 | ids = _build_chain(root, repo_id, 3) |
| 2501 | self._start(root, ids) |
| 2502 | state = _load_state(root) |
| 2503 | assert state is not None |
| 2504 | state["log"].append(f"{ids[1]} bad\x1b[31m 2026-01-01T00:00:00\x1b[0m") |
| 2505 | _save_state(root, state) |
| 2506 | result = _invoke(root, ["bisect", "log"]) |
| 2507 | assert result.exit_code == 0 |
| 2508 | assert "\x1b" not in result.output |
| 2509 | |
| 2510 | def test_log_json_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2511 | root, repo_id = _make_repo(tmp_path) |
| 2512 | ids = _build_chain(root, repo_id, 4) |
| 2513 | self._start(root, ids) |
| 2514 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2515 | assert result.exit_code == 0 |
| 2516 | assert "\x1b" not in result.output |
| 2517 | |
| 2518 | def test_log_text_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2519 | root, repo_id = _make_repo(tmp_path) |
| 2520 | ids = _build_chain(root, repo_id, 4) |
| 2521 | self._start(root, ids) |
| 2522 | result = _invoke(root, ["bisect", "log"]) |
| 2523 | assert result.exit_code == 0 |
| 2524 | assert "\x1b" not in result.output |
| 2525 | |
| 2526 | |
| 2527 | class TestBisectLogStress: |
| 2528 | """Performance and scale tests for muse bisect log.""" |
| 2529 | |
| 2530 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2531 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2532 | assert r.exit_code == 0 |
| 2533 | |
| 2534 | def test_log_100_commit_session(self, tmp_path: pathlib.Path) -> None: |
| 2535 | """Log on a 100-step session returns all entries.""" |
| 2536 | root, repo_id = _make_repo(tmp_path) |
| 2537 | ids = _build_chain(root, repo_id, 100) |
| 2538 | self._start(root, ids) |
| 2539 | # Apply 10 good verdicts to build up a log |
| 2540 | for i in range(1, 11): |
| 2541 | _invoke(root, ["bisect", "good", ids[i]]) |
| 2542 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2543 | assert result.exit_code == 0 |
| 2544 | entries = _parse_log(result.output)["entries"] |
| 2545 | # start adds 2 entries; 10 good verdicts add 10 more |
| 2546 | assert len(entries) >= 12 |
| 2547 | |
| 2548 | def test_log_performance_large_session(self, tmp_path: pathlib.Path) -> None: |
| 2549 | """Log on a large session completes within 5 seconds.""" |
| 2550 | import time |
| 2551 | root, repo_id = _make_repo(tmp_path) |
| 2552 | ids = _build_chain(root, repo_id, 50) |
| 2553 | self._start(root, ids) |
| 2554 | for i in range(1, 8): |
| 2555 | _invoke(root, ["bisect", "bad", ids[i]]) |
| 2556 | t0 = time.monotonic() |
| 2557 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2558 | elapsed = time.monotonic() - t0 |
| 2559 | assert result.exit_code == 0 |
| 2560 | assert elapsed < 5.0, f"bisect log took {elapsed:.2f}s" |
| 2561 | |
| 2562 | def test_log_concurrent_reads_consistent(self, tmp_path: pathlib.Path) -> None: |
| 2563 | """Concurrent log reads all return the same entry count.""" |
| 2564 | root, repo_id = _make_repo(tmp_path) |
| 2565 | ids = _build_chain(root, repo_id, 20) |
| 2566 | self._start(root, ids) |
| 2567 | _invoke(root, ["bisect", "bad", ids[10]]) |
| 2568 | counts: list[int] = [] |
| 2569 | errors: list[str] = [] |
| 2570 | lock = threading.Lock() |
| 2571 | |
| 2572 | def _run() -> None: |
| 2573 | r = _invoke(root, ["bisect", "log", "--json"]) |
| 2574 | with lock: |
| 2575 | if r.exit_code != 0: |
| 2576 | errors.append(r.output) |
| 2577 | return |
| 2578 | try: |
| 2579 | counts.append(len(_parse_log(r.output)["entries"])) |
| 2580 | except (json.JSONDecodeError, KeyError, ValueError) as exc: |
| 2581 | errors.append(f"parse error: {exc!r} output={r.output!r}") |
| 2582 | |
| 2583 | threads = [threading.Thread(target=_run) for _ in range(8)] |
| 2584 | for t in threads: |
| 2585 | t.start() |
| 2586 | for t in threads: |
| 2587 | t.join() |
| 2588 | assert not errors |
| 2589 | assert all(c == counts[0] for c in counts) |
| 2590 | |
| 2591 | |
| 2592 | # --------------------------------------------------------------------------- |
| 2593 | # bisect reset — Extended, Security, Stress |
| 2594 | # --------------------------------------------------------------------------- |
| 2595 | |
| 2596 | |
| 2597 | class TestBisectResetExtended: |
| 2598 | """Extended unit / integration / e2e tests for muse bisect reset.""" |
| 2599 | |
| 2600 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2601 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2602 | assert r.exit_code == 0 |
| 2603 | |
| 2604 | def test_reset_exits_0_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2605 | root, repo_id = _make_repo(tmp_path) |
| 2606 | ids = _build_chain(root, repo_id, 4) |
| 2607 | self._start(root, ids) |
| 2608 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2609 | |
| 2610 | def test_reset_exits_0_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2611 | root, _ = _make_repo(tmp_path) |
| 2612 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2613 | |
| 2614 | def test_reset_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2615 | root, _ = _make_repo(tmp_path) |
| 2616 | result = _invoke(root, ["bisect", "reset", "-j"]) |
| 2617 | assert result.exit_code == 0 |
| 2618 | assert _parse_reset(result.output)["reset"] is True |
| 2619 | |
| 2620 | def test_reset_json_reset_true(self, tmp_path: pathlib.Path) -> None: |
| 2621 | root, _ = _make_repo(tmp_path) |
| 2622 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2623 | assert result.exit_code == 0 |
| 2624 | assert _parse_reset(result.output)["reset"] is True |
| 2625 | |
| 2626 | def test_reset_json_single_key(self, tmp_path: pathlib.Path) -> None: |
| 2627 | root, _ = _make_repo(tmp_path) |
| 2628 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2629 | assert result.exit_code == 0 |
| 2630 | d = json.loads(_json_blob(result.output)) |
| 2631 | assert set(d.keys()) == {"reset"} |
| 2632 | |
| 2633 | def test_reset_clears_active_session(self, tmp_path: pathlib.Path) -> None: |
| 2634 | root, repo_id = _make_repo(tmp_path) |
| 2635 | ids = _build_chain(root, repo_id, 4) |
| 2636 | self._start(root, ids) |
| 2637 | _invoke(root, ["bisect", "reset"]) |
| 2638 | log_r = _invoke(root, ["bisect", "log", "--json"]) |
| 2639 | assert _parse_log(log_r.output)["active"] is False |
| 2640 | |
| 2641 | def test_reset_prevents_bad_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2642 | root, repo_id = _make_repo(tmp_path) |
| 2643 | ids = _build_chain(root, repo_id, 4) |
| 2644 | self._start(root, ids) |
| 2645 | _invoke(root, ["bisect", "reset"]) |
| 2646 | result = _invoke(root, ["bisect", "bad", ids[2]]) |
| 2647 | assert result.exit_code == 1 |
| 2648 | |
| 2649 | def test_reset_prevents_good_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2650 | root, repo_id = _make_repo(tmp_path) |
| 2651 | ids = _build_chain(root, repo_id, 4) |
| 2652 | self._start(root, ids) |
| 2653 | _invoke(root, ["bisect", "reset"]) |
| 2654 | assert _invoke(root, ["bisect", "good", ids[1]]).exit_code == 1 |
| 2655 | |
| 2656 | def test_reset_prevents_skip_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2657 | root, repo_id = _make_repo(tmp_path) |
| 2658 | ids = _build_chain(root, repo_id, 4) |
| 2659 | self._start(root, ids) |
| 2660 | _invoke(root, ["bisect", "reset"]) |
| 2661 | assert _invoke(root, ["bisect", "skip", ids[2]]).exit_code == 1 |
| 2662 | |
| 2663 | def test_reset_idempotent_double_reset(self, tmp_path: pathlib.Path) -> None: |
| 2664 | root, repo_id = _make_repo(tmp_path) |
| 2665 | ids = _build_chain(root, repo_id, 3) |
| 2666 | self._start(root, ids) |
| 2667 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2668 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2669 | |
| 2670 | def test_reset_allows_new_session_after(self, tmp_path: pathlib.Path) -> None: |
| 2671 | root, repo_id = _make_repo(tmp_path) |
| 2672 | ids = _build_chain(root, repo_id, 5) |
| 2673 | self._start(root, ids) |
| 2674 | _invoke(root, ["bisect", "reset"]) |
| 2675 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2676 | assert result.exit_code == 0 |
| 2677 | |
| 2678 | def test_reset_clears_log_entries(self, tmp_path: pathlib.Path) -> None: |
| 2679 | root, repo_id = _make_repo(tmp_path) |
| 2680 | ids = _build_chain(root, repo_id, 4) |
| 2681 | self._start(root, ids) |
| 2682 | _invoke(root, ["bisect", "bad", ids[2]]) |
| 2683 | _invoke(root, ["bisect", "reset"]) |
| 2684 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2685 | assert entries == [] |
| 2686 | |
| 2687 | def test_reset_text_output_mentions_reset(self, tmp_path: pathlib.Path) -> None: |
| 2688 | root, _ = _make_repo(tmp_path) |
| 2689 | result = _invoke(root, ["bisect", "reset"]) |
| 2690 | assert result.exit_code == 0 |
| 2691 | assert "reset" in result.output.lower() |
| 2692 | |
| 2693 | def test_reset_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 2694 | root, _ = _make_repo(tmp_path) |
| 2695 | result = _invoke(root, ["bisect", "reset"]) |
| 2696 | assert not result.output.strip().startswith("{") |
| 2697 | |
| 2698 | def test_reset_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2699 | empty = tmp_path / "not_a_repo" |
| 2700 | empty.mkdir() |
| 2701 | assert _invoke(empty, ["bisect", "reset"]).exit_code == 2 |
| 2702 | |
| 2703 | def test_reset_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2704 | root, _ = _make_repo(tmp_path) |
| 2705 | result = _invoke(root, ["bisect", "reset", "--help"]) |
| 2706 | assert "Agent quickstart" in result.output or "Idempotent" in result.output |
| 2707 | |
| 2708 | def test_reset_json_reset_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2709 | root, _ = _make_repo(tmp_path) |
| 2710 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2711 | assert result.exit_code == 0 |
| 2712 | assert isinstance(json.loads(_json_blob(result.output))["reset"], bool) |
| 2713 | |
| 2714 | def test_reset_mid_session_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 2715 | """Reset works correctly after several verdicts have been applied.""" |
| 2716 | root, repo_id = _make_repo(tmp_path) |
| 2717 | ids = _build_chain(root, repo_id, 10) |
| 2718 | self._start(root, ids) |
| 2719 | _invoke(root, ["bisect", "bad", ids[7]]) |
| 2720 | _invoke(root, ["bisect", "good", ids[3]]) |
| 2721 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2722 | assert result.exit_code == 0 |
| 2723 | assert _parse_reset(result.output)["reset"] is True |
| 2724 | assert _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["active"] is False |
| 2725 | |
| 2726 | |
| 2727 | class TestBisectResetSecurity: |
| 2728 | """Security hardening tests for muse bisect reset.""" |
| 2729 | |
| 2730 | def test_reset_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2731 | root, _ = _make_repo(tmp_path) |
| 2732 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2733 | assert result.exit_code == 0 |
| 2734 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 2735 | |
| 2736 | def test_reset_json_no_ansi(self, tmp_path: pathlib.Path) -> None: |
| 2737 | root, _ = _make_repo(tmp_path) |
| 2738 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2739 | assert result.exit_code == 0 |
| 2740 | assert "\x1b" not in result.output |
| 2741 | |
| 2742 | def test_reset_text_no_ansi(self, tmp_path: pathlib.Path) -> None: |
| 2743 | root, _ = _make_repo(tmp_path) |
| 2744 | result = _invoke(root, ["bisect", "reset"]) |
| 2745 | assert result.exit_code == 0 |
| 2746 | assert "\x1b" not in result.output |
| 2747 | |
| 2748 | def test_reset_state_file_removed(self, tmp_path: pathlib.Path) -> None: |
| 2749 | """After reset the state file no longer exists on disk.""" |
| 2750 | from muse.core.bisect import _state_path |
| 2751 | root, repo_id = _make_repo(tmp_path) |
| 2752 | ids = _build_chain(root, repo_id, 3) |
| 2753 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2754 | assert _state_path(root).exists() |
| 2755 | _invoke(root, ["bisect", "reset"]) |
| 2756 | assert not _state_path(root).exists() |
| 2757 | |
| 2758 | def test_reset_no_session_state_file_absent(self, tmp_path: pathlib.Path) -> None: |
| 2759 | """Reset with no state file is a safe no-op.""" |
| 2760 | from muse.core.bisect import _state_path |
| 2761 | root, _ = _make_repo(tmp_path) |
| 2762 | assert not _state_path(root).exists() |
| 2763 | result = _invoke(root, ["bisect", "reset"]) |
| 2764 | assert result.exit_code == 0 |
| 2765 | |
| 2766 | def test_reset_json_reset_value_true(self, tmp_path: pathlib.Path) -> None: |
| 2767 | """reset field is always true, never false or a truthy int.""" |
| 2768 | root, _ = _make_repo(tmp_path) |
| 2769 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2770 | assert result.exit_code == 0 |
| 2771 | assert json.loads(_json_blob(result.output))["reset"] is True |
| 2772 | |
| 2773 | |
| 2774 | class TestBisectResetStress: |
| 2775 | """Performance and scale tests for muse bisect reset.""" |
| 2776 | |
| 2777 | def test_reset_after_100_commit_session(self, tmp_path: pathlib.Path) -> None: |
| 2778 | """Reset clears state from a 100-commit session instantly.""" |
| 2779 | root, repo_id = _make_repo(tmp_path) |
| 2780 | ids = _build_chain(root, repo_id, 100) |
| 2781 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2782 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2783 | assert result.exit_code == 0 |
| 2784 | assert _parse_reset(result.output)["reset"] is True |
| 2785 | assert _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["active"] is False |
| 2786 | |
| 2787 | def test_reset_performance(self, tmp_path: pathlib.Path) -> None: |
| 2788 | """Reset completes within 2 seconds even after a large session.""" |
| 2789 | import time |
| 2790 | root, repo_id = _make_repo(tmp_path) |
| 2791 | ids = _build_chain(root, repo_id, 100) |
| 2792 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2793 | for i in range(1, 8): |
| 2794 | _invoke(root, ["bisect", "bad", ids[i]]) |
| 2795 | t0 = time.monotonic() |
| 2796 | result = _invoke(root, ["bisect", "reset"]) |
| 2797 | elapsed = time.monotonic() - t0 |
| 2798 | assert result.exit_code == 0 |
| 2799 | assert elapsed < 2.0, f"bisect reset took {elapsed:.2f}s" |
| 2800 | |
| 2801 | def test_reset_cycle_10_times(self, tmp_path: pathlib.Path) -> None: |
| 2802 | """Start → reset × 10 all succeed with no state leakage.""" |
| 2803 | root, repo_id = _make_repo(tmp_path) |
| 2804 | ids = _build_chain(root, repo_id, 6) |
| 2805 | for _ in range(10): |
| 2806 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2807 | assert r_start.exit_code == 0 |
| 2808 | r_reset = _invoke(root, ["bisect", "reset", "--json"]) |
| 2809 | assert r_reset.exit_code == 0 |
| 2810 | assert _parse_reset(r_reset.output)["reset"] is True |
File History
5 commits
sha256:660fcac1df3ab28f61862e961890bd2ca8b754fa0242079d93ca1e25037ec8a6
chore(tests): add docstring to tests/__init__.py so rc14 tr…
Human
26 days ago
sha256:d8316ffae901be06347e16ab55be11868eb519dd16ade3e8aa16a99e662f7e62
baseline: rc14 re-baseline after rc3 store corruption recovery
Human
patch
26 days ago
sha256:50d413a635f57761c3fce1f70251b957b6423821525bf330747ef4007339222e
Merge branch 'dev' into main
Human
29 days ago
sha256:fb67fed5a4d3e40de84bdd163de94ef1386570bef1dd1a020a732c8a038962ce
Merge branch 'dev' into main
Human
48 days ago
sha256:1c4b3e3a9a1f300774c3ee662b572a698d5fd405bf765a71e6011a2e9c3eaaaa
feat: Muse — version control for the agent era
Human
100 days ago