test_cmd_bisect_hardening.py
python
sha256:51ce277f663e01a43eaffbe77509b1de7ac2d4251b55d23306304bcdeb92c90d
feat(pack): delta-encode snapshots in MPackBundle wire format
Sonnet 4.6
minor
⚠ breaking
122 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 | from typing import TypedDict |
| 20 | |
| 21 | import pytest |
| 22 | |
| 23 | from muse.core.snapshot import compute_commit_id, compute_snapshot_id |
| 24 | from muse.core.store import CommitRecord, SnapshotRecord, write_commit, write_snapshot |
| 25 | from muse.core.types import Manifest, fake_id, short_id |
| 26 | from tests.cli_test_helper import CliRunner, InvokeResult |
| 27 | |
| 28 | # Helpers to check store field names at import time; mypy will catch mismatches. |
| 29 | _SNAP_FIELDS: set[str] = {"snapshot_id", "manifest", "created_at"} |
| 30 | _COMMIT_FIELDS: set[str] = {"commit_id", "repo_id", "branch", "snapshot_id", "message", "committed_at"} |
| 31 | |
| 32 | runner = CliRunner() |
| 33 | |
| 34 | _ANSI_RE = re.compile(r"\x1b\[[0-9;]*m") |
| 35 | |
| 36 | |
| 37 | # --------------------------------------------------------------------------- |
| 38 | # Fixtures |
| 39 | # --------------------------------------------------------------------------- |
| 40 | |
| 41 | |
| 42 | def _make_repo(tmp_path: pathlib.Path) -> tuple[pathlib.Path, str]: |
| 43 | """Create a minimal Muse repo layout without calling muse init. |
| 44 | |
| 45 | Returns (repo_root, repo_id). |
| 46 | """ |
| 47 | repo_id = fake_id("repo") |
| 48 | muse = muse_dir(tmp_path) |
| 49 | muse.mkdir() |
| 50 | (muse / "repo.json").write_text( |
| 51 | json.dumps({ |
| 52 | "repo_id": repo_id, |
| 53 | "domain": "code", |
| 54 | "default_branch": "main", |
| 55 | "created_at": "2026-01-01T00:00:00+00:00", |
| 56 | }) |
| 57 | ) |
| 58 | (muse / "HEAD").write_text("ref: refs/heads/main") |
| 59 | (muse / "refs" / "heads").mkdir(parents=True) |
| 60 | (muse / "snapshots").mkdir() |
| 61 | (muse / "commits").mkdir() |
| 62 | (muse / "objects").mkdir() |
| 63 | return tmp_path, repo_id |
| 64 | |
| 65 | |
| 66 | def _make_commit( |
| 67 | root: pathlib.Path, |
| 68 | repo_id: str, |
| 69 | *, |
| 70 | branch: str = "main", |
| 71 | message: str = "commit", |
| 72 | parent_id: str | None = None, |
| 73 | ) -> str: |
| 74 | """Write a synthetic commit and return its commit_id.""" |
| 75 | manifest: Manifest = {} |
| 76 | snap_id = compute_snapshot_id(manifest) |
| 77 | committed_at = datetime.datetime.now(datetime.timezone.utc) |
| 78 | commit_id = compute_commit_id( parent_ids=[parent_id] if parent_id else [], |
| 79 | snapshot_id=snap_id, |
| 80 | message=message, |
| 81 | committed_at_iso=committed_at.isoformat(), |
| 82 | ) |
| 83 | snap = SnapshotRecord( |
| 84 | snapshot_id=snap_id, |
| 85 | manifest={}, |
| 86 | created_at=committed_at, |
| 87 | ) |
| 88 | write_snapshot(root, snap) |
| 89 | commit = CommitRecord( |
| 90 | commit_id=commit_id, |
| 91 | repo_id="test-repo", |
| 92 | parent_commit_id=parent_id, |
| 93 | parent2_commit_id=None, |
| 94 | snapshot_id=snap_id, |
| 95 | branch=branch, |
| 96 | message=message, |
| 97 | committed_at=committed_at, |
| 98 | ) |
| 99 | write_commit(root, commit) |
| 100 | branch_ref = ref_path(root, branch) |
| 101 | branch_ref.write_text(commit_id) |
| 102 | (head_path(root)).write_text(f"ref: refs/heads/{branch}") |
| 103 | return commit_id |
| 104 | |
| 105 | |
| 106 | def _build_chain(root: pathlib.Path, repo_id: str, n: int) -> list[str]: |
| 107 | """Create n commits (linear chain) and return their IDs oldest-first.""" |
| 108 | ids: list[str] = [] |
| 109 | parent: str | None = None |
| 110 | for i in range(n): |
| 111 | cid = _make_commit(root, repo_id, message=f"commit {i}", parent_id=parent) |
| 112 | ids.append(cid) |
| 113 | parent = cid |
| 114 | return ids |
| 115 | |
| 116 | |
| 117 | def _invoke(root: pathlib.Path, args: list[str]) -> InvokeResult: |
| 118 | return runner.invoke(None, args, env={"MUSE_REPO_ROOT": str(root)}) |
| 119 | |
| 120 | |
| 121 | def _json_blob(output: str) -> str: |
| 122 | """Extract the first complete JSON object/array from mixed output. |
| 123 | |
| 124 | Handles both compact (single-line) and pretty-printed (multi-line) JSON. |
| 125 | Falls back to line-by-line extraction for NDJSON streams. |
| 126 | """ |
| 127 | stripped = output.strip() |
| 128 | # Fast path: try the whole output (works for pretty-printed single objects) |
| 129 | try: |
| 130 | json.loads(stripped) |
| 131 | return stripped |
| 132 | except json.JSONDecodeError: |
| 133 | pass |
| 134 | # Fallback: find the first JSON line (NDJSON or compact output mixed with text) |
| 135 | for line in output.splitlines(): |
| 136 | s = line.strip() |
| 137 | if s.startswith("{") or s.startswith("["): |
| 138 | return s |
| 139 | return stripped |
| 140 | |
| 141 | |
| 142 | # --------------------------------------------------------------------------- |
| 143 | # Typed schema helpers |
| 144 | # --------------------------------------------------------------------------- |
| 145 | |
| 146 | |
| 147 | class _StepJson(TypedDict): |
| 148 | done: bool |
| 149 | first_bad: str | None |
| 150 | next_to_test: str | None |
| 151 | remaining_count: int |
| 152 | steps_remaining: int |
| 153 | verdict: str |
| 154 | symbol_changes: list[str] |
| 155 | |
| 156 | |
| 157 | class _LogEntryJson(TypedDict): |
| 158 | commit_id: str |
| 159 | verdict: str |
| 160 | timestamp: str |
| 161 | |
| 162 | |
| 163 | class _LogJson(TypedDict): |
| 164 | active: bool |
| 165 | entries: list[_LogEntryJson] |
| 166 | |
| 167 | |
| 168 | class _ResetJson(TypedDict): |
| 169 | reset: bool |
| 170 | |
| 171 | |
| 172 | class _RunStepJson(TypedDict): |
| 173 | step: int |
| 174 | testing: str |
| 175 | verdict: str |
| 176 | remaining_count: int |
| 177 | done: bool |
| 178 | symbol_changes: list[str] |
| 179 | |
| 180 | |
| 181 | class _RunDoneJson(TypedDict): |
| 182 | done: bool |
| 183 | first_bad: str | None |
| 184 | steps_taken: int |
| 185 | |
| 186 | |
| 187 | def _repo(tmp_path: pathlib.Path) -> tuple[pathlib.Path, str]: |
| 188 | """Alias for _make_repo for readability inside test methods.""" |
| 189 | return _make_repo(tmp_path) |
| 190 | |
| 191 | |
| 192 | def _parse_step(output: str) -> _StepJson: |
| 193 | raw = json.loads(_json_blob(output)) |
| 194 | assert isinstance(raw, dict) |
| 195 | done_val = raw["done"] |
| 196 | first_bad_val = raw["first_bad"] |
| 197 | next_to_test_val = raw["next_to_test"] |
| 198 | remaining_count_val = raw["remaining_count"] |
| 199 | steps_remaining_val = raw["steps_remaining"] |
| 200 | verdict_val = raw["verdict"] |
| 201 | symbol_changes_val = raw["symbol_changes"] |
| 202 | assert isinstance(done_val, bool) |
| 203 | assert first_bad_val is None or isinstance(first_bad_val, str) |
| 204 | assert next_to_test_val is None or isinstance(next_to_test_val, str) |
| 205 | assert isinstance(remaining_count_val, int) |
| 206 | assert isinstance(steps_remaining_val, int) |
| 207 | assert isinstance(verdict_val, str) |
| 208 | assert isinstance(symbol_changes_val, list) |
| 209 | return _StepJson( |
| 210 | done=done_val, |
| 211 | first_bad=first_bad_val, |
| 212 | next_to_test=next_to_test_val, |
| 213 | remaining_count=remaining_count_val, |
| 214 | steps_remaining=steps_remaining_val, |
| 215 | verdict=verdict_val, |
| 216 | symbol_changes=symbol_changes_val, |
| 217 | ) |
| 218 | |
| 219 | |
| 220 | def _parse_log(output: str) -> _LogJson: |
| 221 | raw = json.loads(_json_blob(output)) |
| 222 | assert isinstance(raw, dict) |
| 223 | active_val = raw["active"] |
| 224 | entries_val = raw["entries"] |
| 225 | assert isinstance(active_val, bool) |
| 226 | assert isinstance(entries_val, list) |
| 227 | return _LogJson(active=active_val, entries=entries_val) |
| 228 | |
| 229 | |
| 230 | def _parse_reset(output: str) -> _ResetJson: |
| 231 | raw = json.loads(_json_blob(output)) |
| 232 | assert isinstance(raw, dict) |
| 233 | reset_val = raw["reset"] |
| 234 | assert isinstance(reset_val, bool) |
| 235 | return _ResetJson(reset=reset_val) |
| 236 | |
| 237 | |
| 238 | # --------------------------------------------------------------------------- |
| 239 | # Unit — _toml_escape |
| 240 | # --------------------------------------------------------------------------- |
| 241 | |
| 242 | |
| 243 | class TestTomlEscape: |
| 244 | def test_plain_string_unchanged(self) -> None: |
| 245 | from muse.core.bisect import _toml_escape |
| 246 | |
| 247 | assert _toml_escape("feat/my-thing") == "feat/my-thing" |
| 248 | |
| 249 | def test_double_quote_escaped(self) -> None: |
| 250 | from muse.core.bisect import _toml_escape |
| 251 | |
| 252 | result = _toml_escape('branch"with"quotes') |
| 253 | # After escaping, no bare double-quotes remain (only \"). |
| 254 | assert '\\"' in result |
| 255 | |
| 256 | def test_backslash_escaped(self) -> None: |
| 257 | from muse.core.bisect import _toml_escape |
| 258 | |
| 259 | result = _toml_escape("branch\\with\\backslash") |
| 260 | assert result == "branch\\\\with\\\\backslash" |
| 261 | |
| 262 | def test_both_escaped(self) -> None: |
| 263 | from muse.core.bisect import _toml_escape |
| 264 | |
| 265 | result = _toml_escape('malicious"; bad_id = "hacked') |
| 266 | assert '\\"' in result |
| 267 | assert "bad_id" in result # literal text preserved, just escaped |
| 268 | |
| 269 | |
| 270 | # --------------------------------------------------------------------------- |
| 271 | # Unit — _load_state security |
| 272 | # --------------------------------------------------------------------------- |
| 273 | |
| 274 | |
| 275 | class TestLoadStateSecurity: |
| 276 | def test_symlink_state_file_rejected(self, tmp_path: pathlib.Path) -> None: |
| 277 | """A symlink at the bisect state path must be silently ignored.""" |
| 278 | from muse.core.bisect import _load_state, _state_path |
| 279 | |
| 280 | root, _ = _make_repo(tmp_path) |
| 281 | target = tmp_path / "real_state.toml" |
| 282 | target.write_text('bad_id = "abc"\ngood_ids = []\nskipped_ids = []\nremaining = []\nlog = []\n') |
| 283 | state_path = _state_path(root) |
| 284 | state_path.symlink_to(target) |
| 285 | result = _load_state(root) |
| 286 | assert result is None |
| 287 | |
| 288 | def test_oversized_state_file_rejected(self, tmp_path: pathlib.Path) -> None: |
| 289 | """State files exceeding _MAX_STATE_BYTES must be rejected.""" |
| 290 | from muse.core.bisect import _MAX_STATE_BYTES, _load_state, _state_path |
| 291 | |
| 292 | root, _ = _make_repo(tmp_path) |
| 293 | state_path = _state_path(root) |
| 294 | huge = "x" * (_MAX_STATE_BYTES + 1) |
| 295 | state_path.write_text(huge) |
| 296 | result = _load_state(root) |
| 297 | assert result is None |
| 298 | |
| 299 | def test_corrupt_state_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 300 | from muse.core.bisect import _load_state, _state_path |
| 301 | |
| 302 | root, _ = _make_repo(tmp_path) |
| 303 | state_path = _state_path(root) |
| 304 | state_path.write_text("not valid toml ]] [[[ !!!") |
| 305 | result = _load_state(root) |
| 306 | assert result is None |
| 307 | |
| 308 | def test_missing_state_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 309 | from muse.core.bisect import _load_state |
| 310 | |
| 311 | root, _ = _make_repo(tmp_path) |
| 312 | result = _load_state(root) |
| 313 | assert result is None |
| 314 | |
| 315 | |
| 316 | # --------------------------------------------------------------------------- |
| 317 | # Unit — _save_state TOML injection |
| 318 | # --------------------------------------------------------------------------- |
| 319 | |
| 320 | |
| 321 | class TestSaveStateTomlInjection: |
| 322 | def test_branch_with_quote_survives_roundtrip(self, tmp_path: pathlib.Path) -> None: |
| 323 | """A branch name containing a double-quote must not corrupt the state file.""" |
| 324 | from muse.core.bisect import BisectStateDict, _load_state, _save_state |
| 325 | |
| 326 | root, _ = _make_repo(tmp_path) |
| 327 | state: BisectStateDict = { |
| 328 | "bad_id": "a" * 64, |
| 329 | "good_ids": ["b" * 64], |
| 330 | "skipped_ids": [], |
| 331 | "remaining": [], |
| 332 | "log": [], |
| 333 | "branch": 'malicious"; bad_id = "injected', |
| 334 | } |
| 335 | _save_state(root, state) |
| 336 | loaded = _load_state(root) |
| 337 | assert loaded is not None |
| 338 | assert loaded.get("bad_id") == "a" * 64 |
| 339 | assert loaded.get("branch") == 'malicious"; bad_id = "injected' |
| 340 | |
| 341 | def test_branch_with_backslash_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": "c" * 64, |
| 347 | "good_ids": ["d" * 64], |
| 348 | "skipped_ids": [], |
| 349 | "remaining": [], |
| 350 | "log": [], |
| 351 | "branch": "feat\\\\weird", |
| 352 | } |
| 353 | _save_state(root, state) |
| 354 | loaded = _load_state(root) |
| 355 | assert loaded is not None |
| 356 | assert loaded.get("branch") == "feat\\\\weird" |
| 357 | |
| 358 | def test_symbol_filter_injection_survives_roundtrip(self, tmp_path: pathlib.Path) -> None: |
| 359 | from muse.core.bisect import BisectStateDict, _load_state, _save_state |
| 360 | |
| 361 | root, _ = _make_repo(tmp_path) |
| 362 | state: BisectStateDict = { |
| 363 | "bad_id": "e" * 64, |
| 364 | "good_ids": ["f" * 64], |
| 365 | "skipped_ids": [], |
| 366 | "remaining": [], |
| 367 | "log": [], |
| 368 | "symbol_filter": 'billing.py::Invoice"; bad_id = "EVIL', |
| 369 | } |
| 370 | _save_state(root, state) |
| 371 | loaded = _load_state(root) |
| 372 | assert loaded is not None |
| 373 | assert loaded.get("bad_id") == "e" * 64 |
| 374 | assert loaded.get("symbol_filter") == 'billing.py::Invoice"; bad_id = "EVIL' |
| 375 | |
| 376 | |
| 377 | # --------------------------------------------------------------------------- |
| 378 | # Unit — get_bisect_next public API |
| 379 | # --------------------------------------------------------------------------- |
| 380 | |
| 381 | |
| 382 | class TestGetBisectNext: |
| 383 | def test_no_session_returns_none(self, tmp_path: pathlib.Path) -> None: |
| 384 | from muse.core.bisect import get_bisect_next |
| 385 | |
| 386 | root, _ = _make_repo(tmp_path) |
| 387 | nxt, sf = get_bisect_next(root) |
| 388 | assert nxt is None |
| 389 | assert sf == "" |
| 390 | |
| 391 | def test_returns_next_after_start(self, tmp_path: pathlib.Path) -> None: |
| 392 | from muse.core.bisect import get_bisect_next, start_bisect |
| 393 | |
| 394 | root, repo_id = _make_repo(tmp_path) |
| 395 | ids = _build_chain(root, repo_id, 5) |
| 396 | start_bisect(root, ids[-1], [ids[0]]) |
| 397 | nxt, sf = get_bisect_next(root) |
| 398 | assert nxt is not None |
| 399 | assert nxt in ids |
| 400 | assert sf == "" |
| 401 | |
| 402 | def test_returns_symbol_filter(self, tmp_path: pathlib.Path) -> None: |
| 403 | from muse.core.bisect import get_bisect_next, start_bisect |
| 404 | |
| 405 | root, repo_id = _make_repo(tmp_path) |
| 406 | ids = _build_chain(root, repo_id, 4) |
| 407 | # No commits touch this symbol, so remaining will be empty. |
| 408 | start_bisect(root, ids[-1], [ids[0]], symbol_filter="no_file.py::NoSymbol") |
| 409 | nxt, sf = get_bisect_next(root) |
| 410 | # Symbol filter is preserved regardless of whether next exists. |
| 411 | assert sf == "no_file.py::NoSymbol" |
| 412 | |
| 413 | |
| 414 | # --------------------------------------------------------------------------- |
| 415 | # Security — CLI error routing |
| 416 | # --------------------------------------------------------------------------- |
| 417 | |
| 418 | |
| 419 | class TestErrorRouting: |
| 420 | def test_bad_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 421 | root, _ = _make_repo(tmp_path) |
| 422 | result = _invoke(root, ["bisect", "bad"]) |
| 423 | assert result.exit_code != 0 |
| 424 | assert "No bisect session" in (result.stderr or result.output) |
| 425 | |
| 426 | def test_good_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 427 | root, _ = _make_repo(tmp_path) |
| 428 | result = _invoke(root, ["bisect", "good"]) |
| 429 | assert result.exit_code != 0 |
| 430 | assert "No bisect session" in (result.stderr or result.output) |
| 431 | |
| 432 | def test_skip_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 433 | root, _ = _make_repo(tmp_path) |
| 434 | result = _invoke(root, ["bisect", "skip"]) |
| 435 | assert result.exit_code != 0 |
| 436 | assert "No bisect session" in (result.stderr or result.output) |
| 437 | |
| 438 | def test_run_without_session_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 439 | root, _ = _make_repo(tmp_path) |
| 440 | result = _invoke(root, ["bisect", "run", "true"]) |
| 441 | assert result.exit_code != 0 |
| 442 | assert "No bisect session" in (result.stderr or result.output) |
| 443 | |
| 444 | def test_symbol_without_double_colon_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 445 | root, repo_id = _make_repo(tmp_path) |
| 446 | ids = _build_chain(root, repo_id, 2) |
| 447 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", "no_colon_here"]) |
| 448 | assert result.exit_code != 0 |
| 449 | assert "❌" in (result.stderr or result.output) |
| 450 | |
| 451 | def test_symbol_too_long_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 452 | root, repo_id = _make_repo(tmp_path) |
| 453 | ids = _build_chain(root, repo_id, 2) |
| 454 | long_sym = f"f.py::{'x' * 600}" |
| 455 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", long_sym]) |
| 456 | assert result.exit_code != 0 |
| 457 | assert "too long" in (result.stderr or result.output) |
| 458 | |
| 459 | def test_double_start_goes_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 460 | root, repo_id = _make_repo(tmp_path) |
| 461 | ids = _build_chain(root, repo_id, 3) |
| 462 | r1 = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 463 | assert r1.exit_code == 0 |
| 464 | r2 = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 465 | assert r2.exit_code != 0 |
| 466 | assert "already active" in (r2.stderr or r2.output) |
| 467 | |
| 468 | |
| 469 | # --------------------------------------------------------------------------- |
| 470 | # Security — ANSI sanitization in outputs |
| 471 | # --------------------------------------------------------------------------- |
| 472 | |
| 473 | |
| 474 | class TestAnsiSanitization: |
| 475 | def test_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 476 | root, repo_id = _make_repo(tmp_path) |
| 477 | ids = _build_chain(root, repo_id, 2) |
| 478 | ansi_ref = "\x1b[31mHEAD\x1b[0m" |
| 479 | result = _invoke(root, ["bisect", "start", "--bad", ansi_ref, "--good", ids[0]]) |
| 480 | assert _ANSI_RE.search(result.output) is None |
| 481 | |
| 482 | def test_ansi_in_symbol_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 483 | root, repo_id = _make_repo(tmp_path) |
| 484 | ids = _build_chain(root, repo_id, 2) |
| 485 | sym = "\x1b[31mfoo.py::Bar\x1b[0m" |
| 486 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", sym]) |
| 487 | assert _ANSI_RE.search(result.output) is None |
| 488 | |
| 489 | |
| 490 | # --------------------------------------------------------------------------- |
| 491 | # JSON schema — start |
| 492 | # --------------------------------------------------------------------------- |
| 493 | |
| 494 | |
| 495 | class TestJsonSchemaStart: |
| 496 | def test_start_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 497 | root, repo_id = _make_repo(tmp_path) |
| 498 | ids = _build_chain(root, repo_id, 5) |
| 499 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 500 | assert result.exit_code == 0 |
| 501 | parsed = _parse_step(result.output) |
| 502 | assert parsed["verdict"] == "started" |
| 503 | assert isinstance(parsed["done"], bool) |
| 504 | assert isinstance(parsed["remaining_count"], int) |
| 505 | assert parsed["remaining_count"] >= 0 |
| 506 | |
| 507 | def test_start_json_done_when_no_remaining(self, tmp_path: pathlib.Path) -> None: |
| 508 | """When bad and good are adjacent, start should report done=True immediately.""" |
| 509 | root, repo_id = _make_repo(tmp_path) |
| 510 | ids = _build_chain(root, repo_id, 2) |
| 511 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 512 | assert result.exit_code == 0 |
| 513 | parsed = _parse_step(result.output) |
| 514 | assert parsed["done"] is True |
| 515 | assert parsed["first_bad"] == ids[-1] |
| 516 | |
| 517 | def test_start_json_symbol_changes_list(self, tmp_path: pathlib.Path) -> None: |
| 518 | root, repo_id = _make_repo(tmp_path) |
| 519 | ids = _build_chain(root, repo_id, 4) |
| 520 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 521 | parsed = _parse_step(result.output) |
| 522 | assert isinstance(parsed["symbol_changes"], list) |
| 523 | |
| 524 | |
| 525 | # --------------------------------------------------------------------------- |
| 526 | # JSON schema — bad / good / skip |
| 527 | # --------------------------------------------------------------------------- |
| 528 | |
| 529 | |
| 530 | class TestJsonSchemaBadGoodSkip: |
| 531 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 532 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 533 | assert r.exit_code == 0 |
| 534 | |
| 535 | def test_bad_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 536 | root, repo_id = _make_repo(tmp_path) |
| 537 | ids = _build_chain(root, repo_id, 5) |
| 538 | self._start(root, ids) |
| 539 | midpoint = ids[len(ids) // 2] |
| 540 | result = _invoke(root, ["bisect", "bad", midpoint, "--json"]) |
| 541 | assert result.exit_code == 0 |
| 542 | parsed = _parse_step(result.output) |
| 543 | assert parsed["verdict"] == "bad" |
| 544 | |
| 545 | def test_good_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 546 | root, repo_id = _make_repo(tmp_path) |
| 547 | ids = _build_chain(root, repo_id, 5) |
| 548 | self._start(root, ids) |
| 549 | midpoint = ids[len(ids) // 2] |
| 550 | result = _invoke(root, ["bisect", "good", midpoint, "--json"]) |
| 551 | assert result.exit_code == 0 |
| 552 | parsed = _parse_step(result.output) |
| 553 | assert parsed["verdict"] == "good" |
| 554 | |
| 555 | def test_skip_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 556 | root, repo_id = _make_repo(tmp_path) |
| 557 | ids = _build_chain(root, repo_id, 5) |
| 558 | self._start(root, ids) |
| 559 | midpoint = ids[len(ids) // 2] |
| 560 | result = _invoke(root, ["bisect", "skip", midpoint, "--json"]) |
| 561 | assert result.exit_code == 0 |
| 562 | parsed = _parse_step(result.output) |
| 563 | assert parsed["verdict"] == "skip" |
| 564 | |
| 565 | |
| 566 | # --------------------------------------------------------------------------- |
| 567 | # JSON schema — log |
| 568 | # --------------------------------------------------------------------------- |
| 569 | |
| 570 | |
| 571 | class TestJsonSchemaLog: |
| 572 | def test_log_json_no_session(self, tmp_path: pathlib.Path) -> None: |
| 573 | root, _ = _make_repo(tmp_path) |
| 574 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 575 | assert result.exit_code == 0 |
| 576 | parsed = _parse_log(result.output) |
| 577 | assert parsed["active"] is False |
| 578 | assert parsed["entries"] == [] |
| 579 | |
| 580 | def test_log_json_after_start(self, tmp_path: pathlib.Path) -> None: |
| 581 | root, repo_id = _make_repo(tmp_path) |
| 582 | ids = _build_chain(root, repo_id, 4) |
| 583 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 584 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 585 | assert result.exit_code == 0 |
| 586 | parsed = _parse_log(result.output) |
| 587 | assert parsed["active"] is True |
| 588 | assert len(parsed["entries"]) >= 2 |
| 589 | |
| 590 | def test_log_json_entries_are_dicts(self, tmp_path: pathlib.Path) -> None: |
| 591 | root, repo_id = _make_repo(tmp_path) |
| 592 | ids = _build_chain(root, repo_id, 3) |
| 593 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 594 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 595 | parsed = _parse_log(result.output) |
| 596 | for entry in parsed["entries"]: |
| 597 | assert isinstance(entry, dict) |
| 598 | assert set(entry.keys()) == {"commit_id", "verdict", "timestamp"} |
| 599 | |
| 600 | |
| 601 | # --------------------------------------------------------------------------- |
| 602 | # JSON schema — reset |
| 603 | # --------------------------------------------------------------------------- |
| 604 | |
| 605 | |
| 606 | class TestJsonSchemaReset: |
| 607 | def test_reset_json_no_session(self, tmp_path: pathlib.Path) -> None: |
| 608 | root, _ = _make_repo(tmp_path) |
| 609 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 610 | assert result.exit_code == 0 |
| 611 | parsed = _parse_reset(result.output) |
| 612 | assert parsed["reset"] is True |
| 613 | |
| 614 | def test_reset_json_with_session(self, tmp_path: pathlib.Path) -> None: |
| 615 | root, repo_id = _make_repo(tmp_path) |
| 616 | ids = _build_chain(root, repo_id, 3) |
| 617 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 618 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 619 | assert result.exit_code == 0 |
| 620 | parsed = _parse_reset(result.output) |
| 621 | assert parsed["reset"] is True |
| 622 | |
| 623 | def test_reset_clears_active_flag(self, tmp_path: pathlib.Path) -> None: |
| 624 | root, repo_id = _make_repo(tmp_path) |
| 625 | ids = _build_chain(root, repo_id, 3) |
| 626 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 627 | _invoke(root, ["bisect", "reset", "--json"]) |
| 628 | log_result = _invoke(root, ["bisect", "log", "--json"]) |
| 629 | parsed = _parse_log(log_result.output) |
| 630 | assert parsed["active"] is False |
| 631 | |
| 632 | |
| 633 | # --------------------------------------------------------------------------- |
| 634 | # JSON schema — run (NDJSON) |
| 635 | # --------------------------------------------------------------------------- |
| 636 | |
| 637 | |
| 638 | class TestJsonSchemaRun: |
| 639 | def test_run_json_ndjson_format(self, tmp_path: pathlib.Path) -> None: |
| 640 | """``bisect run --json`` should emit valid NDJSON.""" |
| 641 | root, repo_id = _make_repo(tmp_path) |
| 642 | ids = _build_chain(root, repo_id, 6) |
| 643 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 644 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 645 | assert result.exit_code == 0 |
| 646 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 647 | assert len(lines) >= 1 |
| 648 | for raw_line in lines[:-1]: |
| 649 | step_raw = json.loads(raw_line) |
| 650 | assert "step" in step_raw |
| 651 | assert "verdict" in step_raw |
| 652 | assert "testing" in step_raw |
| 653 | assert "remaining_count" in step_raw |
| 654 | assert "done" in step_raw |
| 655 | done_raw = json.loads(lines[-1]) |
| 656 | done_val = done_raw["done"] |
| 657 | assert isinstance(done_val, bool) |
| 658 | steps_taken_val = done_raw["steps_taken"] |
| 659 | assert isinstance(steps_taken_val, int) |
| 660 | |
| 661 | def test_run_json_done_has_first_bad(self, tmp_path: pathlib.Path) -> None: |
| 662 | """With always-good command, first_bad on the done line should be set.""" |
| 663 | root, repo_id = _make_repo(tmp_path) |
| 664 | ids = _build_chain(root, repo_id, 4) |
| 665 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 666 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 667 | assert result.exit_code == 0 |
| 668 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 669 | done_raw = json.loads(lines[-1]) |
| 670 | done_val = done_raw["done"] |
| 671 | first_bad_val = done_raw["first_bad"] |
| 672 | if done_val: |
| 673 | assert first_bad_val is not None |
| 674 | |
| 675 | def test_run_json_steps_taken_increments(self, tmp_path: pathlib.Path) -> None: |
| 676 | root, repo_id = _make_repo(tmp_path) |
| 677 | ids = _build_chain(root, repo_id, 8) |
| 678 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 679 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 680 | lines = [ln.strip() for ln in result.output.strip().splitlines() if ln.strip()] |
| 681 | done_raw = json.loads(lines[-1]) |
| 682 | steps_taken = done_raw["steps_taken"] |
| 683 | assert steps_taken >= 1 |
| 684 | |
| 685 | |
| 686 | # --------------------------------------------------------------------------- |
| 687 | # Integration — session lifecycle with --json |
| 688 | # --------------------------------------------------------------------------- |
| 689 | |
| 690 | |
| 691 | class TestIntegrationJson: |
| 692 | def test_start_bad_good_converge(self, tmp_path: pathlib.Path) -> None: |
| 693 | """A manual bisect session with --json converges to a first_bad.""" |
| 694 | root, repo_id = _make_repo(tmp_path) |
| 695 | ids = _build_chain(root, repo_id, 7) |
| 696 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 697 | assert r_start.exit_code == 0 |
| 698 | step = _parse_step(r_start.output) |
| 699 | if step["done"]: |
| 700 | assert step["first_bad"] is not None |
| 701 | return |
| 702 | for _ in range(20): |
| 703 | nxt = step["next_to_test"] |
| 704 | assert nxt is not None |
| 705 | r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 706 | assert r.exit_code == 0 |
| 707 | step = _parse_step(r.output) |
| 708 | if step["done"]: |
| 709 | assert step["first_bad"] is not None |
| 710 | return |
| 711 | pytest.fail("Bisect did not converge within 20 steps") |
| 712 | |
| 713 | def test_good_narrows_range(self, tmp_path: pathlib.Path) -> None: |
| 714 | root, repo_id = _make_repo(tmp_path) |
| 715 | ids = _build_chain(root, repo_id, 8) |
| 716 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 717 | midpoint = ids[len(ids) // 2] |
| 718 | r_good = _invoke(root, ["bisect", "good", midpoint, "--json"]) |
| 719 | step = _parse_step(r_good.output) |
| 720 | if not step["done"]: |
| 721 | assert step["remaining_count"] < len(ids) - 2 |
| 722 | |
| 723 | def test_log_grows_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 724 | root, repo_id = _make_repo(tmp_path) |
| 725 | ids = _build_chain(root, repo_id, 5) |
| 726 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 727 | midpoint = ids[len(ids) // 2] |
| 728 | _invoke(root, ["bisect", "bad", midpoint]) |
| 729 | r_log = _invoke(root, ["bisect", "log", "--json"]) |
| 730 | parsed = _parse_log(r_log.output) |
| 731 | # start logs 2 entries (bad+good); bad adds 1 more → at least 3. |
| 732 | assert len(parsed["entries"]) >= 3 |
| 733 | |
| 734 | def test_skip_excluded_from_remaining(self, tmp_path: pathlib.Path) -> None: |
| 735 | root, repo_id = _make_repo(tmp_path) |
| 736 | ids = _build_chain(root, repo_id, 6) |
| 737 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 738 | step_start = _parse_step(r_start.output) |
| 739 | if step_start["done"]: |
| 740 | return |
| 741 | nxt = step_start["next_to_test"] |
| 742 | assert nxt is not None |
| 743 | r_skip = _invoke(root, ["bisect", "skip", nxt, "--json"]) |
| 744 | step_skip = _parse_step(r_skip.output) |
| 745 | if not step_skip["done"]: |
| 746 | assert step_skip["next_to_test"] != nxt |
| 747 | |
| 748 | |
| 749 | # --------------------------------------------------------------------------- |
| 750 | # E2E — text (non-JSON) output still works |
| 751 | # --------------------------------------------------------------------------- |
| 752 | |
| 753 | |
| 754 | class TestE2EText: |
| 755 | def test_start_text_output_no_json(self, tmp_path: pathlib.Path) -> None: |
| 756 | root, repo_id = _make_repo(tmp_path) |
| 757 | ids = _build_chain(root, repo_id, 4) |
| 758 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 759 | assert result.exit_code == 0 |
| 760 | assert "Bisect session started" in result.output or "First bad commit" in result.output |
| 761 | |
| 762 | def test_bad_text_output(self, tmp_path: pathlib.Path) -> None: |
| 763 | root, repo_id = _make_repo(tmp_path) |
| 764 | ids = _build_chain(root, repo_id, 4) |
| 765 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 766 | midpoint = ids[len(ids) // 2] |
| 767 | result = _invoke(root, ["bisect", "bad", midpoint]) |
| 768 | assert result.exit_code == 0 |
| 769 | assert "bad" in result.output.lower() |
| 770 | |
| 771 | def test_log_text_shows_entries(self, tmp_path: pathlib.Path) -> None: |
| 772 | root, repo_id = _make_repo(tmp_path) |
| 773 | ids = _build_chain(root, repo_id, 3) |
| 774 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 775 | result = _invoke(root, ["bisect", "log"]) |
| 776 | assert result.exit_code == 0 |
| 777 | assert "Bisect log" in result.output |
| 778 | |
| 779 | def test_reset_text_output(self, tmp_path: pathlib.Path) -> None: |
| 780 | root, repo_id = _make_repo(tmp_path) |
| 781 | ids = _build_chain(root, repo_id, 2) |
| 782 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 783 | result = _invoke(root, ["bisect", "reset"]) |
| 784 | assert result.exit_code == 0 |
| 785 | assert "reset" in result.output.lower() |
| 786 | |
| 787 | def test_run_text_output_converges(self, tmp_path: pathlib.Path) -> None: |
| 788 | root, repo_id = _make_repo(tmp_path) |
| 789 | ids = _build_chain(root, repo_id, 5) |
| 790 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 791 | result = _invoke(root, ["bisect", "run", "true"]) |
| 792 | assert result.exit_code == 0 |
| 793 | assert "First bad commit" in result.output or "Bisect complete" in result.output |
| 794 | |
| 795 | def test_no_good_flag_fails_clearly(self, tmp_path: pathlib.Path) -> None: |
| 796 | root, repo_id = _make_repo(tmp_path) |
| 797 | ids = _build_chain(root, repo_id, 2) |
| 798 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 799 | assert result.exit_code != 0 |
| 800 | |
| 801 | def test_log_empty_when_no_session(self, tmp_path: pathlib.Path) -> None: |
| 802 | root, _ = _make_repo(tmp_path) |
| 803 | result = _invoke(root, ["bisect", "log"]) |
| 804 | assert result.exit_code == 0 |
| 805 | assert "No bisect log" in result.output |
| 806 | |
| 807 | |
| 808 | # --------------------------------------------------------------------------- |
| 809 | # E2E — symbol-scoped bisect |
| 810 | # --------------------------------------------------------------------------- |
| 811 | |
| 812 | |
| 813 | class TestSymbolScopedBisect: |
| 814 | def test_symbol_filter_no_matching_commits_warns(self, tmp_path: pathlib.Path) -> None: |
| 815 | root, repo_id = _make_repo(tmp_path) |
| 816 | ids = _build_chain(root, repo_id, 4) |
| 817 | result = _invoke( |
| 818 | root, |
| 819 | [ |
| 820 | "bisect", "start", |
| 821 | "--bad", ids[-1], |
| 822 | "--good", ids[0], |
| 823 | "--symbol", "ghost.py::GhostFunc", |
| 824 | ], |
| 825 | ) |
| 826 | assert result.exit_code == 0 |
| 827 | combined = result.output + (result.stderr or "") |
| 828 | assert "No commits" in combined or "First bad" in combined |
| 829 | |
| 830 | def test_symbol_filter_json_schema_preserved(self, tmp_path: pathlib.Path) -> None: |
| 831 | root, repo_id = _make_repo(tmp_path) |
| 832 | ids = _build_chain(root, repo_id, 5) |
| 833 | result = _invoke( |
| 834 | root, |
| 835 | [ |
| 836 | "bisect", "start", |
| 837 | "--bad", ids[-1], |
| 838 | "--good", ids[0], |
| 839 | "--symbol", "ghost.py::GhostFunc", |
| 840 | "--json", |
| 841 | ], |
| 842 | ) |
| 843 | assert result.exit_code == 0 |
| 844 | parsed = _parse_step(result.output) |
| 845 | assert isinstance(parsed["symbol_changes"], list) |
| 846 | |
| 847 | def test_symbol_filter_state_persisted(self, tmp_path: pathlib.Path) -> None: |
| 848 | """After start with --symbol, the symbol_filter must survive state reload.""" |
| 849 | from muse.core.bisect import _load_state |
| 850 | |
| 851 | root, repo_id = _make_repo(tmp_path) |
| 852 | ids = _build_chain(root, repo_id, 4) |
| 853 | _invoke( |
| 854 | root, |
| 855 | [ |
| 856 | "bisect", "start", |
| 857 | "--bad", ids[-1], |
| 858 | "--good", ids[0], |
| 859 | "--symbol", "billing.py::Invoice", |
| 860 | ], |
| 861 | ) |
| 862 | state = _load_state(root) |
| 863 | assert state is not None |
| 864 | assert state.get("symbol_filter") == "billing.py::Invoice" |
| 865 | |
| 866 | |
| 867 | # --------------------------------------------------------------------------- |
| 868 | # Stress — large commit chains |
| 869 | # --------------------------------------------------------------------------- |
| 870 | |
| 871 | |
| 872 | class TestStress: |
| 873 | def test_200_commit_chain_converges(self, tmp_path: pathlib.Path) -> None: |
| 874 | """Bisect over 200 commits must converge in ≤9 steps (log₂(200) ≈ 7.6).""" |
| 875 | root, repo_id = _make_repo(tmp_path) |
| 876 | ids = _build_chain(root, repo_id, 200) |
| 877 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 878 | |
| 879 | steps = 0 |
| 880 | for _ in range(10): |
| 881 | r = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 882 | assert r.exit_code == 0 |
| 883 | lines = [ln.strip() for ln in r.output.strip().splitlines() if ln.strip()] |
| 884 | if lines: |
| 885 | done_raw = json.loads(lines[-1]) |
| 886 | if done_raw.get("done"): |
| 887 | steps = done_raw.get("steps_taken", 0) |
| 888 | break |
| 889 | else: |
| 890 | pytest.fail("Bisect did not terminate within 10 run invocations") |
| 891 | assert steps <= 9, f"Expected ≤9 steps for 200 commits, got {steps}" |
| 892 | |
| 893 | def test_concurrent_log_reads_are_safe(self, tmp_path: pathlib.Path) -> None: |
| 894 | """Concurrent reads of bisect log must not crash.""" |
| 895 | root, repo_id = _make_repo(tmp_path) |
| 896 | ids = _build_chain(root, repo_id, 10) |
| 897 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 898 | |
| 899 | errors: list[str] = [] |
| 900 | |
| 901 | def _read_log() -> None: |
| 902 | from muse.core.bisect import get_bisect_log |
| 903 | try: |
| 904 | entries = get_bisect_log(root) |
| 905 | assert isinstance(entries, list) |
| 906 | except Exception as exc: |
| 907 | errors.append(str(exc)) |
| 908 | |
| 909 | threads = [threading.Thread(target=_read_log) for _ in range(20)] |
| 910 | for t in threads: |
| 911 | t.start() |
| 912 | for t in threads: |
| 913 | t.join() |
| 914 | |
| 915 | assert not errors, f"Concurrent read failures: {errors}" |
| 916 | |
| 917 | def test_50_step_manual_bisect_json(self, tmp_path: pathlib.Path) -> None: |
| 918 | """50 mark_bad calls on a 100-commit chain must all emit valid JSON.""" |
| 919 | root, repo_id = _make_repo(tmp_path) |
| 920 | ids = _build_chain(root, repo_id, 100) |
| 921 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 922 | assert r_start.exit_code == 0 |
| 923 | step = _parse_step(r_start.output) |
| 924 | |
| 925 | for _ in range(50): |
| 926 | if step["done"]: |
| 927 | assert step["first_bad"] is not None |
| 928 | return |
| 929 | nxt = step["next_to_test"] |
| 930 | assert nxt is not None |
| 931 | r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 932 | assert r.exit_code == 0 |
| 933 | step = _parse_step(r.output) |
| 934 | |
| 935 | assert step["done"] is True |
| 936 | |
| 937 | |
| 938 | # --------------------------------------------------------------------------- |
| 939 | # bisect start — Extended, Security, Stress |
| 940 | # --------------------------------------------------------------------------- |
| 941 | |
| 942 | |
| 943 | class TestBisectStartExtended: |
| 944 | """Extended unit / integration / e2e tests for muse bisect start.""" |
| 945 | |
| 946 | def test_start_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 947 | root, repo_id = _make_repo(tmp_path) |
| 948 | ids = _build_chain(root, repo_id, 5) |
| 949 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 950 | assert result.exit_code == 0 |
| 951 | |
| 952 | def test_start_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 953 | """-j is an accepted alias for --json.""" |
| 954 | root, repo_id = _make_repo(tmp_path) |
| 955 | ids = _build_chain(root, repo_id, 5) |
| 956 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "-j"]) |
| 957 | assert result.exit_code == 0 |
| 958 | parsed = _parse_step(result.output) |
| 959 | assert parsed["verdict"] == "started" |
| 960 | |
| 961 | def test_start_json_verdict_is_started(self, tmp_path: pathlib.Path) -> None: |
| 962 | root, repo_id = _make_repo(tmp_path) |
| 963 | ids = _build_chain(root, repo_id, 5) |
| 964 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 965 | assert result.exit_code == 0 |
| 966 | assert _parse_step(result.output)["verdict"] == "started" |
| 967 | |
| 968 | def test_start_json_done_false_with_remaining(self, tmp_path: pathlib.Path) -> None: |
| 969 | root, repo_id = _make_repo(tmp_path) |
| 970 | ids = _build_chain(root, repo_id, 5) |
| 971 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 972 | assert result.exit_code == 0 |
| 973 | parsed = _parse_step(result.output) |
| 974 | assert parsed["done"] is False |
| 975 | assert parsed["next_to_test"] is not None |
| 976 | |
| 977 | def test_start_json_done_true_when_adjacent(self, tmp_path: pathlib.Path) -> None: |
| 978 | root, repo_id = _make_repo(tmp_path) |
| 979 | ids = _build_chain(root, repo_id, 2) |
| 980 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 981 | assert result.exit_code == 0 |
| 982 | parsed = _parse_step(result.output) |
| 983 | assert parsed["done"] is True |
| 984 | assert parsed["first_bad"] == ids[-1] |
| 985 | |
| 986 | def test_start_json_remaining_count_positive(self, tmp_path: pathlib.Path) -> None: |
| 987 | root, repo_id = _make_repo(tmp_path) |
| 988 | ids = _build_chain(root, repo_id, 8) |
| 989 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 990 | assert result.exit_code == 0 |
| 991 | assert _parse_step(result.output)["remaining_count"] > 0 |
| 992 | |
| 993 | def test_start_json_steps_remaining_positive(self, tmp_path: pathlib.Path) -> None: |
| 994 | root, repo_id = _make_repo(tmp_path) |
| 995 | ids = _build_chain(root, repo_id, 8) |
| 996 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 997 | assert result.exit_code == 0 |
| 998 | assert _parse_step(result.output)["steps_remaining"] > 0 |
| 999 | |
| 1000 | def test_start_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1001 | root, repo_id = _make_repo(tmp_path) |
| 1002 | ids = _build_chain(root, repo_id, 5) |
| 1003 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1004 | assert result.exit_code == 0 |
| 1005 | d = json.loads(_json_blob(result.output)) |
| 1006 | assert {"done", "first_bad", "next_to_test", "remaining_count", |
| 1007 | "steps_remaining", "verdict", "symbol_changes"} <= set(d.keys()) |
| 1008 | |
| 1009 | def test_start_multiple_good_refs(self, tmp_path: pathlib.Path) -> None: |
| 1010 | root, repo_id = _make_repo(tmp_path) |
| 1011 | ids = _build_chain(root, repo_id, 6) |
| 1012 | result = _invoke( |
| 1013 | root, |
| 1014 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--good", ids[1], "--json"], |
| 1015 | ) |
| 1016 | assert result.exit_code == 0 |
| 1017 | assert _parse_step(result.output)["verdict"] == "started" |
| 1018 | |
| 1019 | def test_start_no_good_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1020 | root, repo_id = _make_repo(tmp_path) |
| 1021 | ids = _build_chain(root, repo_id, 3) |
| 1022 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 1023 | assert result.exit_code == 1 |
| 1024 | |
| 1025 | def test_start_no_good_error_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 1026 | root, repo_id = _make_repo(tmp_path) |
| 1027 | ids = _build_chain(root, repo_id, 3) |
| 1028 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1]]) |
| 1029 | assert result.exit_code != 0 |
| 1030 | combined = result.output + (result.stderr or "") |
| 1031 | assert "good" in combined.lower() |
| 1032 | |
| 1033 | def test_start_double_start_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1034 | root, repo_id = _make_repo(tmp_path) |
| 1035 | ids = _build_chain(root, repo_id, 5) |
| 1036 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1037 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1038 | assert result.exit_code == 1 |
| 1039 | |
| 1040 | def test_start_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1041 | empty = tmp_path / "not_a_repo" |
| 1042 | empty.mkdir() |
| 1043 | result = _invoke(empty, ["bisect", "start", "--bad", "abc", "--good", "def"]) |
| 1044 | assert result.exit_code == 2 |
| 1045 | |
| 1046 | def test_start_bad_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1047 | root, repo_id = _make_repo(tmp_path) |
| 1048 | ids = _build_chain(root, repo_id, 4) |
| 1049 | # HEAD points to ids[-1]; omit --bad |
| 1050 | result = _invoke(root, ["bisect", "start", "--good", ids[0], "--json"]) |
| 1051 | assert result.exit_code == 0 |
| 1052 | |
| 1053 | def test_start_text_mentions_session_started(self, tmp_path: pathlib.Path) -> None: |
| 1054 | root, repo_id = _make_repo(tmp_path) |
| 1055 | ids = _build_chain(root, repo_id, 5) |
| 1056 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1057 | assert result.exit_code == 0 |
| 1058 | assert "Bisect session started" in result.output or "First bad commit" in result.output |
| 1059 | |
| 1060 | def test_start_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1061 | root, repo_id = _make_repo(tmp_path) |
| 1062 | ids = _build_chain(root, repo_id, 5) |
| 1063 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1064 | assert result.exit_code == 0 |
| 1065 | assert not result.output.strip().startswith("{") |
| 1066 | |
| 1067 | def test_start_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1068 | root, _ = _make_repo(tmp_path) |
| 1069 | result = _invoke(root, ["bisect", "start", "--help"]) |
| 1070 | assert "Agent quickstart" in result.output or "binary" in result.output.lower() |
| 1071 | |
| 1072 | def test_start_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1073 | root, repo_id = _make_repo(tmp_path) |
| 1074 | ids = _build_chain(root, repo_id, 3) |
| 1075 | result = _invoke(root, ["bisect", "start", "--bad", "nonexistent_ref_abc123", "--good", ids[0]]) |
| 1076 | assert result.exit_code == 1 |
| 1077 | |
| 1078 | |
| 1079 | class TestBisectStartSecurity: |
| 1080 | """Security hardening tests for muse bisect start.""" |
| 1081 | |
| 1082 | def test_start_symbol_changes_no_ansi_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1083 | """symbol_changes entries are sanitized in JSON output.""" |
| 1084 | from unittest.mock import patch |
| 1085 | from muse.core.bisect import BisectResult |
| 1086 | root, repo_id = _make_repo(tmp_path) |
| 1087 | ids = _build_chain(root, repo_id, 5) |
| 1088 | injected = BisectResult( |
| 1089 | done=False, |
| 1090 | first_bad=None, |
| 1091 | next_to_test=ids[2], |
| 1092 | remaining_count=3, |
| 1093 | steps_remaining=2, |
| 1094 | verdict="started", |
| 1095 | symbol_changes=["add Invoice.compute\x1b[31mred\x1b[0m"], |
| 1096 | ) |
| 1097 | with patch("muse.cli.commands.bisect.start_bisect", return_value=injected): |
| 1098 | result = _invoke( |
| 1099 | root, |
| 1100 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], |
| 1101 | "--symbol", "billing.py::Invoice", "--json"], |
| 1102 | ) |
| 1103 | assert result.exit_code == 0 |
| 1104 | assert "\x1b" not in result.output |
| 1105 | |
| 1106 | def test_start_symbol_changes_no_ansi_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1107 | """symbol_changes entries are sanitized in text output.""" |
| 1108 | from unittest.mock import patch |
| 1109 | from muse.core.bisect import BisectResult |
| 1110 | root, repo_id = _make_repo(tmp_path) |
| 1111 | ids = _build_chain(root, repo_id, 5) |
| 1112 | injected = BisectResult( |
| 1113 | done=False, |
| 1114 | first_bad=None, |
| 1115 | next_to_test=ids[2], |
| 1116 | remaining_count=3, |
| 1117 | steps_remaining=2, |
| 1118 | verdict="started", |
| 1119 | symbol_changes=["add Invoice.compute\x1b[31mred\x1b[0m"], |
| 1120 | ) |
| 1121 | with patch("muse.cli.commands.bisect.start_bisect", return_value=injected): |
| 1122 | result = _invoke( |
| 1123 | root, |
| 1124 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], |
| 1125 | "--symbol", "billing.py::Invoice"], |
| 1126 | ) |
| 1127 | assert result.exit_code == 0 |
| 1128 | assert "\x1b" not in result.output |
| 1129 | |
| 1130 | def test_start_symbol_missing_separator_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1131 | """--symbol without '::' separator is rejected.""" |
| 1132 | root, repo_id = _make_repo(tmp_path) |
| 1133 | ids = _build_chain(root, repo_id, 3) |
| 1134 | result = _invoke( |
| 1135 | root, |
| 1136 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", "NoSeparator"], |
| 1137 | ) |
| 1138 | assert result.exit_code == 1 |
| 1139 | |
| 1140 | def test_start_symbol_too_long_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1141 | """--symbol exceeding max length is rejected.""" |
| 1142 | root, repo_id = _make_repo(tmp_path) |
| 1143 | ids = _build_chain(root, repo_id, 3) |
| 1144 | long_sym = "a" * 510 + "::b" |
| 1145 | result = _invoke( |
| 1146 | root, |
| 1147 | ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--symbol", long_sym], |
| 1148 | ) |
| 1149 | assert result.exit_code == 1 |
| 1150 | |
| 1151 | def test_start_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1152 | """JSON output is well-formed.""" |
| 1153 | root, repo_id = _make_repo(tmp_path) |
| 1154 | ids = _build_chain(root, repo_id, 5) |
| 1155 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1156 | assert result.exit_code == 0 |
| 1157 | d = json.loads(_json_blob(result.output)) |
| 1158 | assert isinstance(d, dict) |
| 1159 | |
| 1160 | def test_start_json_bool_fields_are_bool(self, tmp_path: pathlib.Path) -> None: |
| 1161 | """done field is always a bool, never int or string.""" |
| 1162 | root, repo_id = _make_repo(tmp_path) |
| 1163 | ids = _build_chain(root, repo_id, 5) |
| 1164 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1165 | assert result.exit_code == 0 |
| 1166 | d = json.loads(_json_blob(result.output)) |
| 1167 | assert isinstance(d["done"], bool) |
| 1168 | |
| 1169 | |
| 1170 | class TestBisectStartStress: |
| 1171 | """Performance and scale tests for muse bisect start.""" |
| 1172 | |
| 1173 | def test_start_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1174 | """Start over a 100-commit chain exits 0 and returns a midpoint.""" |
| 1175 | root, repo_id = _make_repo(tmp_path) |
| 1176 | ids = _build_chain(root, repo_id, 100) |
| 1177 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1178 | assert result.exit_code == 0 |
| 1179 | parsed = _parse_step(result.output) |
| 1180 | assert parsed["done"] is False |
| 1181 | assert parsed["remaining_count"] > 0 |
| 1182 | assert parsed["next_to_test"] is not None |
| 1183 | |
| 1184 | def test_start_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1185 | """Start over 100 commits completes within 5 seconds.""" |
| 1186 | import time |
| 1187 | root, repo_id = _make_repo(tmp_path) |
| 1188 | ids = _build_chain(root, repo_id, 100) |
| 1189 | t0 = time.monotonic() |
| 1190 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1191 | elapsed = time.monotonic() - t0 |
| 1192 | assert result.exit_code == 0 |
| 1193 | assert elapsed < 5.0, f"start over 100 commits took {elapsed:.2f}s" |
| 1194 | |
| 1195 | def test_start_midpoint_is_within_range(self, tmp_path: pathlib.Path) -> None: |
| 1196 | """The suggested midpoint falls strictly between good and bad.""" |
| 1197 | root, repo_id = _make_repo(tmp_path) |
| 1198 | ids = _build_chain(root, repo_id, 20) |
| 1199 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1200 | assert result.exit_code == 0 |
| 1201 | parsed = _parse_step(result.output) |
| 1202 | assert parsed["next_to_test"] not in (ids[0], ids[-1]) |
| 1203 | |
| 1204 | |
| 1205 | # --------------------------------------------------------------------------- |
| 1206 | # bisect bad — Extended, Security, Stress |
| 1207 | # --------------------------------------------------------------------------- |
| 1208 | |
| 1209 | |
| 1210 | class TestBisectBadExtended: |
| 1211 | """Extended unit / integration / e2e tests for muse bisect bad.""" |
| 1212 | |
| 1213 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1214 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1215 | assert r.exit_code == 0 |
| 1216 | |
| 1217 | def test_bad_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1218 | root, repo_id = _make_repo(tmp_path) |
| 1219 | ids = _build_chain(root, repo_id, 6) |
| 1220 | self._start(root, ids) |
| 1221 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2]]) |
| 1222 | assert result.exit_code == 0 |
| 1223 | |
| 1224 | def test_bad_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1225 | root, repo_id = _make_repo(tmp_path) |
| 1226 | ids = _build_chain(root, repo_id, 6) |
| 1227 | self._start(root, ids) |
| 1228 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "-j"]) |
| 1229 | assert result.exit_code == 0 |
| 1230 | parsed = _parse_step(result.output) |
| 1231 | assert parsed["verdict"] == "bad" |
| 1232 | |
| 1233 | def test_bad_json_verdict_is_bad(self, tmp_path: pathlib.Path) -> None: |
| 1234 | root, repo_id = _make_repo(tmp_path) |
| 1235 | ids = _build_chain(root, repo_id, 6) |
| 1236 | self._start(root, ids) |
| 1237 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1238 | assert result.exit_code == 0 |
| 1239 | assert _parse_step(result.output)["verdict"] == "bad" |
| 1240 | |
| 1241 | def test_bad_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1242 | root, repo_id = _make_repo(tmp_path) |
| 1243 | ids = _build_chain(root, repo_id, 6) |
| 1244 | self._start(root, ids) |
| 1245 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1246 | assert result.exit_code == 0 |
| 1247 | d = json.loads(_json_blob(result.output)) |
| 1248 | assert {"done", "first_bad", "next_to_test", "remaining_count", |
| 1249 | "steps_remaining", "verdict", "symbol_changes"} <= set(d.keys()) |
| 1250 | |
| 1251 | def test_bad_reduces_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1252 | root, repo_id = _make_repo(tmp_path) |
| 1253 | ids = _build_chain(root, repo_id, 10) |
| 1254 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1255 | before = _parse_step(r.output)["remaining_count"] |
| 1256 | mid = _parse_step(r.output)["next_to_test"] |
| 1257 | result = _invoke(root, ["bisect", "bad", mid, "--json"]) |
| 1258 | assert result.exit_code == 0 |
| 1259 | after = _parse_step(result.output)["remaining_count"] |
| 1260 | assert after < before |
| 1261 | |
| 1262 | def test_bad_done_true_when_isolated(self, tmp_path: pathlib.Path) -> None: |
| 1263 | root, repo_id = _make_repo(tmp_path) |
| 1264 | ids = _build_chain(root, repo_id, 3) |
| 1265 | # With 3 commits: good=ids[0], bad=ids[2] → ids[1] is the only remaining |
| 1266 | self._start(root, ids) |
| 1267 | result = _invoke(root, ["bisect", "bad", ids[1], "--json"]) |
| 1268 | assert result.exit_code == 0 |
| 1269 | parsed = _parse_step(result.output) |
| 1270 | assert parsed["done"] is True |
| 1271 | assert parsed["first_bad"] is not None |
| 1272 | |
| 1273 | def test_bad_first_bad_set_when_done(self, tmp_path: pathlib.Path) -> None: |
| 1274 | root, repo_id = _make_repo(tmp_path) |
| 1275 | ids = _build_chain(root, repo_id, 3) |
| 1276 | self._start(root, ids) |
| 1277 | result = _invoke(root, ["bisect", "bad", ids[1], "--json"]) |
| 1278 | assert result.exit_code == 0 |
| 1279 | parsed = _parse_step(result.output) |
| 1280 | assert parsed["done"] is True |
| 1281 | assert isinstance(parsed["first_bad"], str) |
| 1282 | |
| 1283 | def test_bad_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1284 | root, repo_id = _make_repo(tmp_path) |
| 1285 | ids = _build_chain(root, repo_id, 5) |
| 1286 | self._start(root, ids) |
| 1287 | # HEAD points to ids[-1] (the known-bad); marking it bad again is valid |
| 1288 | result = _invoke(root, ["bisect", "bad", "--json"]) |
| 1289 | assert result.exit_code == 0 |
| 1290 | |
| 1291 | def test_bad_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1292 | root, _ = _make_repo(tmp_path) |
| 1293 | result = _invoke(root, ["bisect", "bad"]) |
| 1294 | assert result.exit_code == 1 |
| 1295 | |
| 1296 | def test_bad_no_session_error_to_stderr(self, tmp_path: pathlib.Path) -> None: |
| 1297 | root, _ = _make_repo(tmp_path) |
| 1298 | result = _invoke(root, ["bisect", "bad"]) |
| 1299 | assert result.exit_code != 0 |
| 1300 | combined = result.output + (result.stderr or "") |
| 1301 | assert "No bisect session" in combined |
| 1302 | |
| 1303 | def test_bad_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1304 | empty = tmp_path / "not_a_repo" |
| 1305 | empty.mkdir() |
| 1306 | result = _invoke(empty, ["bisect", "bad"]) |
| 1307 | assert result.exit_code == 2 |
| 1308 | |
| 1309 | def test_bad_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1310 | root, repo_id = _make_repo(tmp_path) |
| 1311 | ids = _build_chain(root, repo_id, 4) |
| 1312 | self._start(root, ids) |
| 1313 | result = _invoke(root, ["bisect", "bad", "deadbeef_nonexistent"]) |
| 1314 | assert result.exit_code == 1 |
| 1315 | |
| 1316 | def test_bad_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1317 | root, repo_id = _make_repo(tmp_path) |
| 1318 | ids = _build_chain(root, repo_id, 5) |
| 1319 | self._start(root, ids) |
| 1320 | mid = ids[len(ids) // 2] |
| 1321 | result = _invoke(root, ["bisect", "bad", mid]) |
| 1322 | assert result.exit_code == 0 |
| 1323 | assert short_id(mid) in result.output |
| 1324 | |
| 1325 | def test_bad_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1326 | root, repo_id = _make_repo(tmp_path) |
| 1327 | ids = _build_chain(root, repo_id, 5) |
| 1328 | self._start(root, ids) |
| 1329 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2]]) |
| 1330 | assert result.exit_code == 0 |
| 1331 | assert not result.output.strip().startswith("{") |
| 1332 | |
| 1333 | def test_bad_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1334 | root, _ = _make_repo(tmp_path) |
| 1335 | result = _invoke(root, ["bisect", "bad", "--help"]) |
| 1336 | assert "Agent quickstart" in result.output or "regression" in result.output.lower() |
| 1337 | |
| 1338 | def test_bad_advances_bisect_log(self, tmp_path: pathlib.Path) -> None: |
| 1339 | """After marking bad, the bisect log records the verdict.""" |
| 1340 | from muse.core.bisect import _load_state |
| 1341 | root, repo_id = _make_repo(tmp_path) |
| 1342 | ids = _build_chain(root, repo_id, 6) |
| 1343 | self._start(root, ids) |
| 1344 | mid = ids[len(ids) // 2] |
| 1345 | _invoke(root, ["bisect", "bad", mid]) |
| 1346 | state = _load_state(root) |
| 1347 | assert state is not None |
| 1348 | assert any("bad" in entry for entry in state.get("log", [])) |
| 1349 | |
| 1350 | def test_bad_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1351 | root, repo_id = _make_repo(tmp_path) |
| 1352 | ids = _build_chain(root, repo_id, 5) |
| 1353 | self._start(root, ids) |
| 1354 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1355 | assert result.exit_code == 0 |
| 1356 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1357 | |
| 1358 | def test_bad_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1359 | root, repo_id = _make_repo(tmp_path) |
| 1360 | ids = _build_chain(root, repo_id, 5) |
| 1361 | self._start(root, ids) |
| 1362 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1363 | assert result.exit_code == 0 |
| 1364 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1365 | |
| 1366 | |
| 1367 | class TestBisectBadSecurity: |
| 1368 | """Security hardening tests for muse bisect bad.""" |
| 1369 | |
| 1370 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1371 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1372 | assert r.exit_code == 0 |
| 1373 | |
| 1374 | def test_bad_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1375 | root, repo_id = _make_repo(tmp_path) |
| 1376 | ids = _build_chain(root, repo_id, 5) |
| 1377 | self._start(root, ids) |
| 1378 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1379 | assert result.exit_code == 0 |
| 1380 | d = json.loads(_json_blob(result.output)) |
| 1381 | assert isinstance(d, dict) |
| 1382 | |
| 1383 | def test_bad_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1384 | root, repo_id = _make_repo(tmp_path) |
| 1385 | ids = _build_chain(root, repo_id, 5) |
| 1386 | self._start(root, ids) |
| 1387 | result = _invoke(root, ["bisect", "bad", ids[len(ids) // 2], "--json"]) |
| 1388 | assert result.exit_code == 0 |
| 1389 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1390 | |
| 1391 | def test_bad_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1392 | """ANSI in symbol_changes entries stripped from JSON output.""" |
| 1393 | from unittest.mock import patch |
| 1394 | from muse.core.bisect import BisectResult |
| 1395 | root, repo_id = _make_repo(tmp_path) |
| 1396 | ids = _build_chain(root, repo_id, 5) |
| 1397 | self._start(root, ids) |
| 1398 | injected = BisectResult( |
| 1399 | done=False, |
| 1400 | first_bad=None, |
| 1401 | next_to_test=ids[2], |
| 1402 | remaining_count=2, |
| 1403 | steps_remaining=1, |
| 1404 | verdict="bad", |
| 1405 | symbol_changes=["modify func\x1b[31mred\x1b[0m"], |
| 1406 | ) |
| 1407 | with patch("muse.cli.commands.bisect.mark_bad", return_value=injected): |
| 1408 | result = _invoke(root, ["bisect", "bad", ids[2], "--json"]) |
| 1409 | assert "\x1b" not in result.output |
| 1410 | |
| 1411 | def test_bad_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1412 | """ANSI in symbol_changes entries stripped from text output.""" |
| 1413 | from unittest.mock import patch |
| 1414 | from muse.core.bisect import BisectResult |
| 1415 | root, repo_id = _make_repo(tmp_path) |
| 1416 | ids = _build_chain(root, repo_id, 5) |
| 1417 | self._start(root, ids) |
| 1418 | injected = BisectResult( |
| 1419 | done=False, |
| 1420 | first_bad=None, |
| 1421 | next_to_test=ids[2], |
| 1422 | remaining_count=2, |
| 1423 | steps_remaining=1, |
| 1424 | verdict="bad", |
| 1425 | symbol_changes=["modify func\x1b[31mred\x1b[0m"], |
| 1426 | ) |
| 1427 | with patch("muse.cli.commands.bisect.mark_bad", return_value=injected): |
| 1428 | result = _invoke(root, ["bisect", "bad", ids[2]]) |
| 1429 | assert "\x1b" not in result.output |
| 1430 | |
| 1431 | def test_bad_error_output_to_stderr_not_stdout(self, tmp_path: pathlib.Path) -> None: |
| 1432 | """Error messages go to stderr; stdout is clean on failure.""" |
| 1433 | root, _ = _make_repo(tmp_path) |
| 1434 | result = _invoke(root, ["bisect", "bad"]) |
| 1435 | assert result.exit_code != 0 |
| 1436 | # CliRunner mixes stderr into output; verify no JSON object was emitted |
| 1437 | assert not result.output.strip().startswith("{") |
| 1438 | |
| 1439 | def test_bad_ansi_in_ref_does_not_leak_to_output(self, tmp_path: pathlib.Path) -> None: |
| 1440 | """Passing an ANSI-injected ref does not leak escape codes to stdout.""" |
| 1441 | root, repo_id = _make_repo(tmp_path) |
| 1442 | ids = _build_chain(root, repo_id, 4) |
| 1443 | self._start(root, ids) |
| 1444 | result = _invoke(root, ["bisect", "bad", "\x1b[31mHEAD\x1b[0m"]) |
| 1445 | # Will fail (ref not found) but must not echo ANSI to stdout |
| 1446 | assert "\x1b" not in result.output |
| 1447 | |
| 1448 | |
| 1449 | class TestBisectBadStress: |
| 1450 | """Performance and scale tests for muse bisect bad.""" |
| 1451 | |
| 1452 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1453 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1454 | assert r.exit_code == 0 |
| 1455 | |
| 1456 | def test_bad_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1457 | """Marking bad on a 100-commit session exits 0 and advances the search.""" |
| 1458 | root, repo_id = _make_repo(tmp_path) |
| 1459 | ids = _build_chain(root, repo_id, 100) |
| 1460 | self._start(root, ids) |
| 1461 | result = _invoke(root, ["bisect", "bad", ids[50], "--json"]) |
| 1462 | assert result.exit_code == 0 |
| 1463 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1464 | |
| 1465 | def test_bad_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1466 | """Marking bad on a 100-commit session completes within 5 seconds.""" |
| 1467 | import time |
| 1468 | root, repo_id = _make_repo(tmp_path) |
| 1469 | ids = _build_chain(root, repo_id, 100) |
| 1470 | self._start(root, ids) |
| 1471 | t0 = time.monotonic() |
| 1472 | result = _invoke(root, ["bisect", "bad", ids[50], "--json"]) |
| 1473 | elapsed = time.monotonic() - t0 |
| 1474 | assert result.exit_code == 0 |
| 1475 | assert elapsed < 5.0, f"bisect bad on 100 commits took {elapsed:.2f}s" |
| 1476 | |
| 1477 | def test_bad_converges_full_session(self, tmp_path: pathlib.Path) -> None: |
| 1478 | """Marking next_to_test as bad on every step converges within log2(20) steps.""" |
| 1479 | root, repo_id = _make_repo(tmp_path) |
| 1480 | ids = _build_chain(root, repo_id, 20) |
| 1481 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1482 | assert r.exit_code == 0 |
| 1483 | parsed = _parse_step(r.output) |
| 1484 | done = parsed["done"] |
| 1485 | for _ in range(10): |
| 1486 | if done: |
| 1487 | break |
| 1488 | nxt = parsed["next_to_test"] |
| 1489 | assert nxt is not None |
| 1490 | next_r = _invoke(root, ["bisect", "bad", nxt, "--json"]) |
| 1491 | assert next_r.exit_code == 0 |
| 1492 | parsed = _parse_step(next_r.output) |
| 1493 | done = parsed["done"] |
| 1494 | assert done, "bisect did not converge within 10 bad steps on 20-commit chain" |
| 1495 | |
| 1496 | |
| 1497 | # --------------------------------------------------------------------------- |
| 1498 | # bisect good — Extended, Security, Stress |
| 1499 | # --------------------------------------------------------------------------- |
| 1500 | |
| 1501 | |
| 1502 | class TestBisectGoodExtended: |
| 1503 | """Extended unit / integration / e2e tests for muse bisect good.""" |
| 1504 | |
| 1505 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1506 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1507 | assert r.exit_code == 0 |
| 1508 | |
| 1509 | def test_good_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1510 | root, repo_id = _make_repo(tmp_path) |
| 1511 | ids = _build_chain(root, repo_id, 6) |
| 1512 | self._start(root, ids) |
| 1513 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1514 | assert result.exit_code == 0 |
| 1515 | |
| 1516 | def test_good_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1517 | root, repo_id = _make_repo(tmp_path) |
| 1518 | ids = _build_chain(root, repo_id, 6) |
| 1519 | self._start(root, ids) |
| 1520 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "-j"]) |
| 1521 | assert result.exit_code == 0 |
| 1522 | assert _parse_step(result.output)["verdict"] == "good" |
| 1523 | |
| 1524 | def test_good_json_verdict_is_good(self, tmp_path: pathlib.Path) -> None: |
| 1525 | root, repo_id = _make_repo(tmp_path) |
| 1526 | ids = _build_chain(root, repo_id, 6) |
| 1527 | self._start(root, ids) |
| 1528 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1529 | assert result.exit_code == 0 |
| 1530 | assert _parse_step(result.output)["verdict"] == "good" |
| 1531 | |
| 1532 | def test_good_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1533 | root, repo_id = _make_repo(tmp_path) |
| 1534 | ids = _build_chain(root, repo_id, 6) |
| 1535 | self._start(root, ids) |
| 1536 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1537 | assert result.exit_code == 0 |
| 1538 | d = json.loads(_json_blob(result.output)) |
| 1539 | assert {"done", "first_bad", "next_to_test", "remaining_count", |
| 1540 | "steps_remaining", "verdict", "symbol_changes"} <= set(d.keys()) |
| 1541 | |
| 1542 | def test_good_reduces_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1543 | root, repo_id = _make_repo(tmp_path) |
| 1544 | ids = _build_chain(root, repo_id, 10) |
| 1545 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1546 | before = _parse_step(r.output)["remaining_count"] |
| 1547 | mid = _parse_step(r.output)["next_to_test"] |
| 1548 | result = _invoke(root, ["bisect", "good", mid, "--json"]) |
| 1549 | assert result.exit_code == 0 |
| 1550 | assert _parse_step(result.output)["remaining_count"] < before |
| 1551 | |
| 1552 | def test_good_done_true_when_isolated(self, tmp_path: pathlib.Path) -> None: |
| 1553 | """Marking the only remaining commit good isolates first bad immediately.""" |
| 1554 | root, repo_id = _make_repo(tmp_path) |
| 1555 | ids = _build_chain(root, repo_id, 3) |
| 1556 | # good=ids[0], bad=ids[2]: ids[1] is the midpoint; marking it good resolves |
| 1557 | self._start(root, ids) |
| 1558 | result = _invoke(root, ["bisect", "good", ids[1], "--json"]) |
| 1559 | assert result.exit_code == 0 |
| 1560 | parsed = _parse_step(result.output) |
| 1561 | assert parsed["done"] is True |
| 1562 | assert parsed["first_bad"] == ids[2] |
| 1563 | |
| 1564 | def test_good_first_bad_set_when_done(self, tmp_path: pathlib.Path) -> None: |
| 1565 | root, repo_id = _make_repo(tmp_path) |
| 1566 | ids = _build_chain(root, repo_id, 3) |
| 1567 | self._start(root, ids) |
| 1568 | result = _invoke(root, ["bisect", "good", ids[1], "--json"]) |
| 1569 | assert result.exit_code == 0 |
| 1570 | parsed = _parse_step(result.output) |
| 1571 | assert parsed["done"] is True |
| 1572 | assert isinstance(parsed["first_bad"], str) |
| 1573 | |
| 1574 | def test_good_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1575 | root, repo_id = _make_repo(tmp_path) |
| 1576 | ids = _build_chain(root, repo_id, 5) |
| 1577 | self._start(root, ids) |
| 1578 | # HEAD is ids[-1] (known bad); marking it good is legal but pushes bad boundary |
| 1579 | result = _invoke(root, ["bisect", "good", "--json"]) |
| 1580 | assert result.exit_code == 0 |
| 1581 | |
| 1582 | def test_good_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1583 | root, _ = _make_repo(tmp_path) |
| 1584 | result = _invoke(root, ["bisect", "good"]) |
| 1585 | assert result.exit_code == 1 |
| 1586 | |
| 1587 | def test_good_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 1588 | root, _ = _make_repo(tmp_path) |
| 1589 | result = _invoke(root, ["bisect", "good"]) |
| 1590 | combined = result.output + (result.stderr or "") |
| 1591 | assert "No bisect session" in combined |
| 1592 | |
| 1593 | def test_good_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1594 | empty = tmp_path / "not_a_repo" |
| 1595 | empty.mkdir() |
| 1596 | result = _invoke(empty, ["bisect", "good"]) |
| 1597 | assert result.exit_code == 2 |
| 1598 | |
| 1599 | def test_good_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1600 | root, repo_id = _make_repo(tmp_path) |
| 1601 | ids = _build_chain(root, repo_id, 4) |
| 1602 | self._start(root, ids) |
| 1603 | result = _invoke(root, ["bisect", "good", "deadbeef_nonexistent"]) |
| 1604 | assert result.exit_code == 1 |
| 1605 | |
| 1606 | def test_good_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1607 | root, repo_id = _make_repo(tmp_path) |
| 1608 | ids = _build_chain(root, repo_id, 5) |
| 1609 | self._start(root, ids) |
| 1610 | mid = ids[len(ids) // 2] |
| 1611 | result = _invoke(root, ["bisect", "good", mid]) |
| 1612 | assert result.exit_code == 0 |
| 1613 | assert short_id(mid) in result.output |
| 1614 | |
| 1615 | def test_good_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1616 | root, repo_id = _make_repo(tmp_path) |
| 1617 | ids = _build_chain(root, repo_id, 5) |
| 1618 | self._start(root, ids) |
| 1619 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1620 | assert result.exit_code == 0 |
| 1621 | assert not result.output.strip().startswith("{") |
| 1622 | |
| 1623 | def test_good_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1624 | root, _ = _make_repo(tmp_path) |
| 1625 | result = _invoke(root, ["bisect", "good", "--help"]) |
| 1626 | assert "Agent quickstart" in result.output or "regression" in result.output.lower() |
| 1627 | |
| 1628 | def test_good_advances_bisect_log(self, tmp_path: pathlib.Path) -> None: |
| 1629 | from muse.core.bisect import _load_state |
| 1630 | root, repo_id = _make_repo(tmp_path) |
| 1631 | ids = _build_chain(root, repo_id, 6) |
| 1632 | self._start(root, ids) |
| 1633 | _invoke(root, ["bisect", "good", ids[len(ids) // 2]]) |
| 1634 | state = _load_state(root) |
| 1635 | assert state is not None |
| 1636 | assert any("good" in entry for entry in state.get("log", [])) |
| 1637 | |
| 1638 | def test_good_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1639 | root, repo_id = _make_repo(tmp_path) |
| 1640 | ids = _build_chain(root, repo_id, 5) |
| 1641 | self._start(root, ids) |
| 1642 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1643 | assert result.exit_code == 0 |
| 1644 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1645 | |
| 1646 | def test_good_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1647 | root, repo_id = _make_repo(tmp_path) |
| 1648 | ids = _build_chain(root, repo_id, 5) |
| 1649 | self._start(root, ids) |
| 1650 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1651 | assert result.exit_code == 0 |
| 1652 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1653 | |
| 1654 | |
| 1655 | class TestBisectGoodSecurity: |
| 1656 | """Security hardening tests for muse bisect good.""" |
| 1657 | |
| 1658 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1659 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1660 | assert r.exit_code == 0 |
| 1661 | |
| 1662 | def test_good_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1663 | root, repo_id = _make_repo(tmp_path) |
| 1664 | ids = _build_chain(root, repo_id, 5) |
| 1665 | self._start(root, ids) |
| 1666 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1667 | assert result.exit_code == 0 |
| 1668 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 1669 | |
| 1670 | def test_good_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1671 | root, repo_id = _make_repo(tmp_path) |
| 1672 | ids = _build_chain(root, repo_id, 5) |
| 1673 | self._start(root, ids) |
| 1674 | result = _invoke(root, ["bisect", "good", ids[len(ids) // 2], "--json"]) |
| 1675 | assert result.exit_code == 0 |
| 1676 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1677 | |
| 1678 | def test_good_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1679 | from unittest.mock import patch |
| 1680 | from muse.core.bisect import BisectResult |
| 1681 | root, repo_id = _make_repo(tmp_path) |
| 1682 | ids = _build_chain(root, repo_id, 5) |
| 1683 | self._start(root, ids) |
| 1684 | injected = BisectResult( |
| 1685 | done=False, |
| 1686 | first_bad=None, |
| 1687 | next_to_test=ids[2], |
| 1688 | remaining_count=2, |
| 1689 | steps_remaining=1, |
| 1690 | verdict="good", |
| 1691 | symbol_changes=["add func\x1b[32mgreen\x1b[0m"], |
| 1692 | ) |
| 1693 | with patch("muse.cli.commands.bisect.mark_good", return_value=injected): |
| 1694 | result = _invoke(root, ["bisect", "good", ids[2], "--json"]) |
| 1695 | assert "\x1b" not in result.output |
| 1696 | |
| 1697 | def test_good_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1698 | from unittest.mock import patch |
| 1699 | from muse.core.bisect import BisectResult |
| 1700 | root, repo_id = _make_repo(tmp_path) |
| 1701 | ids = _build_chain(root, repo_id, 5) |
| 1702 | self._start(root, ids) |
| 1703 | injected = BisectResult( |
| 1704 | done=False, |
| 1705 | first_bad=None, |
| 1706 | next_to_test=ids[2], |
| 1707 | remaining_count=2, |
| 1708 | steps_remaining=1, |
| 1709 | verdict="good", |
| 1710 | symbol_changes=["add func\x1b[32mgreen\x1b[0m"], |
| 1711 | ) |
| 1712 | with patch("muse.cli.commands.bisect.mark_good", return_value=injected): |
| 1713 | result = _invoke(root, ["bisect", "good", ids[2]]) |
| 1714 | assert "\x1b" not in result.output |
| 1715 | |
| 1716 | def test_good_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 1717 | root, _ = _make_repo(tmp_path) |
| 1718 | result = _invoke(root, ["bisect", "good"]) |
| 1719 | assert result.exit_code != 0 |
| 1720 | assert not result.output.strip().startswith("{") |
| 1721 | |
| 1722 | def test_good_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 1723 | root, repo_id = _make_repo(tmp_path) |
| 1724 | ids = _build_chain(root, repo_id, 4) |
| 1725 | self._start(root, ids) |
| 1726 | result = _invoke(root, ["bisect", "good", "\x1b[32mHEAD\x1b[0m"]) |
| 1727 | assert "\x1b" not in result.output |
| 1728 | |
| 1729 | |
| 1730 | class TestBisectGoodStress: |
| 1731 | """Performance and scale tests for muse bisect good.""" |
| 1732 | |
| 1733 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1734 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1735 | assert r.exit_code == 0 |
| 1736 | |
| 1737 | def test_good_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 1738 | root, repo_id = _make_repo(tmp_path) |
| 1739 | ids = _build_chain(root, repo_id, 100) |
| 1740 | self._start(root, ids) |
| 1741 | result = _invoke(root, ["bisect", "good", ids[10], "--json"]) |
| 1742 | assert result.exit_code == 0 |
| 1743 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1744 | |
| 1745 | def test_good_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 1746 | import time |
| 1747 | root, repo_id = _make_repo(tmp_path) |
| 1748 | ids = _build_chain(root, repo_id, 100) |
| 1749 | self._start(root, ids) |
| 1750 | t0 = time.monotonic() |
| 1751 | result = _invoke(root, ["bisect", "good", ids[10], "--json"]) |
| 1752 | elapsed = time.monotonic() - t0 |
| 1753 | assert result.exit_code == 0 |
| 1754 | assert elapsed < 5.0, f"bisect good on 100 commits took {elapsed:.2f}s" |
| 1755 | |
| 1756 | def test_good_converges_full_session(self, tmp_path: pathlib.Path) -> None: |
| 1757 | """Marking next_to_test as good on each step converges within log2(20) steps.""" |
| 1758 | root, repo_id = _make_repo(tmp_path) |
| 1759 | ids = _build_chain(root, repo_id, 20) |
| 1760 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1761 | assert r.exit_code == 0 |
| 1762 | parsed = _parse_step(r.output) |
| 1763 | done = parsed["done"] |
| 1764 | for _ in range(10): |
| 1765 | if done: |
| 1766 | break |
| 1767 | nxt = parsed["next_to_test"] |
| 1768 | assert nxt is not None |
| 1769 | next_r = _invoke(root, ["bisect", "good", nxt, "--json"]) |
| 1770 | assert next_r.exit_code == 0 |
| 1771 | parsed = _parse_step(next_r.output) |
| 1772 | done = parsed["done"] |
| 1773 | assert done, "bisect did not converge within 10 good steps on 20-commit chain" |
| 1774 | |
| 1775 | |
| 1776 | # --------------------------------------------------------------------------- |
| 1777 | # bisect skip — Extended, Security, Stress |
| 1778 | # --------------------------------------------------------------------------- |
| 1779 | |
| 1780 | |
| 1781 | class TestBisectSkipExtended: |
| 1782 | """Extended unit / integration / e2e tests for muse bisect skip.""" |
| 1783 | |
| 1784 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1785 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1786 | assert r.exit_code == 0 |
| 1787 | |
| 1788 | def test_skip_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 1789 | root, repo_id = _make_repo(tmp_path) |
| 1790 | ids = _build_chain(root, repo_id, 6) |
| 1791 | self._start(root, ids) |
| 1792 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1793 | assert result.exit_code == 0 |
| 1794 | |
| 1795 | def test_skip_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 1796 | root, repo_id = _make_repo(tmp_path) |
| 1797 | ids = _build_chain(root, repo_id, 6) |
| 1798 | self._start(root, ids) |
| 1799 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "-j"]) |
| 1800 | assert result.exit_code == 0 |
| 1801 | assert _parse_step(result.output)["verdict"] == "skip" |
| 1802 | |
| 1803 | def test_skip_json_verdict_is_skip(self, tmp_path: pathlib.Path) -> None: |
| 1804 | root, repo_id = _make_repo(tmp_path) |
| 1805 | ids = _build_chain(root, repo_id, 6) |
| 1806 | self._start(root, ids) |
| 1807 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1808 | assert result.exit_code == 0 |
| 1809 | assert _parse_step(result.output)["verdict"] == "skip" |
| 1810 | |
| 1811 | def test_skip_json_all_seven_keys(self, tmp_path: pathlib.Path) -> None: |
| 1812 | root, repo_id = _make_repo(tmp_path) |
| 1813 | ids = _build_chain(root, repo_id, 6) |
| 1814 | self._start(root, ids) |
| 1815 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1816 | assert result.exit_code == 0 |
| 1817 | d = json.loads(_json_blob(result.output)) |
| 1818 | assert {"done", "first_bad", "next_to_test", "remaining_count", |
| 1819 | "steps_remaining", "verdict", "symbol_changes"} <= set(d.keys()) |
| 1820 | |
| 1821 | def test_skip_removes_commit_from_remaining(self, tmp_path: pathlib.Path) -> None: |
| 1822 | root, repo_id = _make_repo(tmp_path) |
| 1823 | ids = _build_chain(root, repo_id, 10) |
| 1824 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 1825 | before = _parse_step(r.output)["remaining_count"] |
| 1826 | mid = _parse_step(r.output)["next_to_test"] |
| 1827 | result = _invoke(root, ["bisect", "skip", mid, "--json"]) |
| 1828 | assert result.exit_code == 0 |
| 1829 | assert _parse_step(result.output)["remaining_count"] < before |
| 1830 | |
| 1831 | def test_skip_persisted_in_state(self, tmp_path: pathlib.Path) -> None: |
| 1832 | from muse.core.bisect import _load_state |
| 1833 | root, repo_id = _make_repo(tmp_path) |
| 1834 | ids = _build_chain(root, repo_id, 6) |
| 1835 | self._start(root, ids) |
| 1836 | mid = ids[len(ids) // 2] |
| 1837 | _invoke(root, ["bisect", "skip", mid]) |
| 1838 | state = _load_state(root) |
| 1839 | assert state is not None |
| 1840 | assert mid in state.get("skipped_ids", []) |
| 1841 | |
| 1842 | def test_skip_defaults_to_head(self, tmp_path: pathlib.Path) -> None: |
| 1843 | root, repo_id = _make_repo(tmp_path) |
| 1844 | ids = _build_chain(root, repo_id, 5) |
| 1845 | self._start(root, ids) |
| 1846 | result = _invoke(root, ["bisect", "skip", "--json"]) |
| 1847 | assert result.exit_code == 0 |
| 1848 | |
| 1849 | def test_skip_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1850 | root, _ = _make_repo(tmp_path) |
| 1851 | result = _invoke(root, ["bisect", "skip"]) |
| 1852 | assert result.exit_code == 1 |
| 1853 | |
| 1854 | def test_skip_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 1855 | root, _ = _make_repo(tmp_path) |
| 1856 | result = _invoke(root, ["bisect", "skip"]) |
| 1857 | combined = result.output + (result.stderr or "") |
| 1858 | assert "No bisect session" in combined |
| 1859 | |
| 1860 | def test_skip_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 1861 | empty = tmp_path / "not_a_repo" |
| 1862 | empty.mkdir() |
| 1863 | result = _invoke(empty, ["bisect", "skip"]) |
| 1864 | assert result.exit_code == 2 |
| 1865 | |
| 1866 | def test_skip_invalid_ref_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 1867 | root, repo_id = _make_repo(tmp_path) |
| 1868 | ids = _build_chain(root, repo_id, 4) |
| 1869 | self._start(root, ids) |
| 1870 | result = _invoke(root, ["bisect", "skip", "deadbeef_nonexistent"]) |
| 1871 | assert result.exit_code == 1 |
| 1872 | |
| 1873 | def test_skip_text_mentions_commit(self, tmp_path: pathlib.Path) -> None: |
| 1874 | root, repo_id = _make_repo(tmp_path) |
| 1875 | ids = _build_chain(root, repo_id, 5) |
| 1876 | self._start(root, ids) |
| 1877 | mid = ids[len(ids) // 2] |
| 1878 | result = _invoke(root, ["bisect", "skip", mid]) |
| 1879 | assert result.exit_code == 0 |
| 1880 | assert short_id(mid) in result.output |
| 1881 | |
| 1882 | def test_skip_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 1883 | root, repo_id = _make_repo(tmp_path) |
| 1884 | ids = _build_chain(root, repo_id, 5) |
| 1885 | self._start(root, ids) |
| 1886 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1887 | assert result.exit_code == 0 |
| 1888 | assert not result.output.strip().startswith("{") |
| 1889 | |
| 1890 | def test_skip_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 1891 | root, _ = _make_repo(tmp_path) |
| 1892 | result = _invoke(root, ["bisect", "skip", "--help"]) |
| 1893 | assert "Agent quickstart" in result.output or "125" in result.output |
| 1894 | |
| 1895 | def test_skip_advances_log(self, tmp_path: pathlib.Path) -> None: |
| 1896 | from muse.core.bisect import _load_state |
| 1897 | root, repo_id = _make_repo(tmp_path) |
| 1898 | ids = _build_chain(root, repo_id, 6) |
| 1899 | self._start(root, ids) |
| 1900 | _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 1901 | state = _load_state(root) |
| 1902 | assert state is not None |
| 1903 | assert any("skip" in entry for entry in state.get("log", [])) |
| 1904 | |
| 1905 | def test_skip_remaining_count_not_negative(self, tmp_path: pathlib.Path) -> None: |
| 1906 | root, repo_id = _make_repo(tmp_path) |
| 1907 | ids = _build_chain(root, repo_id, 5) |
| 1908 | self._start(root, ids) |
| 1909 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1910 | assert result.exit_code == 0 |
| 1911 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 1912 | |
| 1913 | def test_skip_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 1914 | root, repo_id = _make_repo(tmp_path) |
| 1915 | ids = _build_chain(root, repo_id, 5) |
| 1916 | self._start(root, ids) |
| 1917 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1918 | assert result.exit_code == 0 |
| 1919 | assert isinstance(_parse_step(result.output)["symbol_changes"], list) |
| 1920 | |
| 1921 | def test_skip_multiple_commits(self, tmp_path: pathlib.Path) -> None: |
| 1922 | """Skipping several commits all land in skipped_ids.""" |
| 1923 | from muse.core.bisect import _load_state |
| 1924 | root, repo_id = _make_repo(tmp_path) |
| 1925 | ids = _build_chain(root, repo_id, 8) |
| 1926 | self._start(root, ids) |
| 1927 | for idx in (2, 3, 4): |
| 1928 | r = _invoke(root, ["bisect", "skip", ids[idx]]) |
| 1929 | assert r.exit_code == 0 |
| 1930 | state = _load_state(root) |
| 1931 | assert state is not None |
| 1932 | skipped = state.get("skipped_ids", []) |
| 1933 | assert all(ids[i] in skipped for i in (2, 3, 4)) |
| 1934 | |
| 1935 | |
| 1936 | class TestBisectSkipSecurity: |
| 1937 | """Security hardening tests for muse bisect skip.""" |
| 1938 | |
| 1939 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 1940 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 1941 | assert r.exit_code == 0 |
| 1942 | |
| 1943 | def test_skip_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 1944 | root, repo_id = _make_repo(tmp_path) |
| 1945 | ids = _build_chain(root, repo_id, 5) |
| 1946 | self._start(root, ids) |
| 1947 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1948 | assert result.exit_code == 0 |
| 1949 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 1950 | |
| 1951 | def test_skip_json_done_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 1952 | root, repo_id = _make_repo(tmp_path) |
| 1953 | ids = _build_chain(root, repo_id, 5) |
| 1954 | self._start(root, ids) |
| 1955 | result = _invoke(root, ["bisect", "skip", ids[len(ids) // 2], "--json"]) |
| 1956 | assert result.exit_code == 0 |
| 1957 | assert isinstance(json.loads(_json_blob(result.output))["done"], bool) |
| 1958 | |
| 1959 | def test_skip_symbol_changes_sanitized_in_json(self, tmp_path: pathlib.Path) -> None: |
| 1960 | from unittest.mock import patch |
| 1961 | from muse.core.bisect import BisectResult |
| 1962 | root, repo_id = _make_repo(tmp_path) |
| 1963 | ids = _build_chain(root, repo_id, 5) |
| 1964 | self._start(root, ids) |
| 1965 | injected = BisectResult( |
| 1966 | done=False, |
| 1967 | first_bad=None, |
| 1968 | next_to_test=ids[2], |
| 1969 | remaining_count=2, |
| 1970 | steps_remaining=1, |
| 1971 | verdict="skip", |
| 1972 | symbol_changes=["modify func\x1b[33myellow\x1b[0m"], |
| 1973 | ) |
| 1974 | with patch("muse.cli.commands.bisect.skip_commit", return_value=injected): |
| 1975 | result = _invoke(root, ["bisect", "skip", ids[2], "--json"]) |
| 1976 | assert "\x1b" not in result.output |
| 1977 | |
| 1978 | def test_skip_symbol_changes_sanitized_in_text(self, tmp_path: pathlib.Path) -> None: |
| 1979 | from unittest.mock import patch |
| 1980 | from muse.core.bisect import BisectResult |
| 1981 | root, repo_id = _make_repo(tmp_path) |
| 1982 | ids = _build_chain(root, repo_id, 5) |
| 1983 | self._start(root, ids) |
| 1984 | injected = BisectResult( |
| 1985 | done=False, |
| 1986 | first_bad=None, |
| 1987 | next_to_test=ids[2], |
| 1988 | remaining_count=2, |
| 1989 | steps_remaining=1, |
| 1990 | verdict="skip", |
| 1991 | symbol_changes=["modify func\x1b[33myellow\x1b[0m"], |
| 1992 | ) |
| 1993 | with patch("muse.cli.commands.bisect.skip_commit", return_value=injected): |
| 1994 | result = _invoke(root, ["bisect", "skip", ids[2]]) |
| 1995 | assert "\x1b" not in result.output |
| 1996 | |
| 1997 | def test_skip_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 1998 | root, _ = _make_repo(tmp_path) |
| 1999 | result = _invoke(root, ["bisect", "skip"]) |
| 2000 | assert result.exit_code != 0 |
| 2001 | assert not result.output.strip().startswith("{") |
| 2002 | |
| 2003 | def test_skip_ansi_in_ref_does_not_leak(self, tmp_path: pathlib.Path) -> None: |
| 2004 | root, repo_id = _make_repo(tmp_path) |
| 2005 | ids = _build_chain(root, repo_id, 4) |
| 2006 | self._start(root, ids) |
| 2007 | result = _invoke(root, ["bisect", "skip", "\x1b[33mHEAD\x1b[0m"]) |
| 2008 | assert "\x1b" not in result.output |
| 2009 | |
| 2010 | |
| 2011 | class TestBisectSkipStress: |
| 2012 | """Performance and scale tests for muse bisect skip.""" |
| 2013 | |
| 2014 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2015 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2016 | assert r.exit_code == 0 |
| 2017 | |
| 2018 | def test_skip_on_100_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 2019 | root, repo_id = _make_repo(tmp_path) |
| 2020 | ids = _build_chain(root, repo_id, 100) |
| 2021 | self._start(root, ids) |
| 2022 | result = _invoke(root, ["bisect", "skip", ids[50], "--json"]) |
| 2023 | assert result.exit_code == 0 |
| 2024 | assert _parse_step(result.output)["remaining_count"] >= 0 |
| 2025 | |
| 2026 | def test_skip_performance_100_commits(self, tmp_path: pathlib.Path) -> None: |
| 2027 | import time |
| 2028 | root, repo_id = _make_repo(tmp_path) |
| 2029 | ids = _build_chain(root, repo_id, 100) |
| 2030 | self._start(root, ids) |
| 2031 | t0 = time.monotonic() |
| 2032 | result = _invoke(root, ["bisect", "skip", ids[50], "--json"]) |
| 2033 | elapsed = time.monotonic() - t0 |
| 2034 | assert result.exit_code == 0 |
| 2035 | assert elapsed < 5.0, f"bisect skip on 100 commits took {elapsed:.2f}s" |
| 2036 | |
| 2037 | def test_skip_reduces_remaining_monotonically(self, tmp_path: pathlib.Path) -> None: |
| 2038 | """Each consecutive skip reduces remaining_count (non-increasing sequence).""" |
| 2039 | root, repo_id = _make_repo(tmp_path) |
| 2040 | ids = _build_chain(root, repo_id, 20) |
| 2041 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 2042 | assert r.exit_code == 0 |
| 2043 | counts = [_parse_step(r.output)["remaining_count"]] |
| 2044 | cur = r |
| 2045 | for _ in range(5): |
| 2046 | parsed = _parse_step(cur.output) |
| 2047 | if parsed["done"] or parsed["next_to_test"] is None: |
| 2048 | break |
| 2049 | nxt = parsed["next_to_test"] |
| 2050 | cur = _invoke(root, ["bisect", "skip", nxt, "--json"]) |
| 2051 | assert cur.exit_code == 0 |
| 2052 | counts.append(_parse_step(cur.output)["remaining_count"]) |
| 2053 | assert all(counts[i] >= counts[i + 1] for i in range(len(counts) - 1)) |
| 2054 | |
| 2055 | |
| 2056 | # --------------------------------------------------------------------------- |
| 2057 | # bisect run — Extended, Security, Stress |
| 2058 | # --------------------------------------------------------------------------- |
| 2059 | |
| 2060 | |
| 2061 | class TestBisectRunExtended: |
| 2062 | """Extended unit / integration / e2e tests for muse bisect run.""" |
| 2063 | |
| 2064 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2065 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2066 | assert r.exit_code == 0 |
| 2067 | |
| 2068 | def test_run_exits_0_with_true(self, tmp_path: pathlib.Path) -> None: |
| 2069 | root, repo_id = _make_repo(tmp_path) |
| 2070 | ids = _build_chain(root, repo_id, 5) |
| 2071 | self._start(root, ids) |
| 2072 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2073 | assert result.exit_code == 0 |
| 2074 | |
| 2075 | def test_run_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2076 | root, repo_id = _make_repo(tmp_path) |
| 2077 | ids = _build_chain(root, repo_id, 5) |
| 2078 | self._start(root, ids) |
| 2079 | result = _invoke(root, ["bisect", "run", "true", "-j"]) |
| 2080 | assert result.exit_code == 0 |
| 2081 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2082 | assert len(lines) >= 1 |
| 2083 | done_raw = json.loads(lines[-1]) |
| 2084 | assert done_raw["done"] is True |
| 2085 | |
| 2086 | def test_run_json_ndjson_step_keys(self, tmp_path: pathlib.Path) -> None: |
| 2087 | root, repo_id = _make_repo(tmp_path) |
| 2088 | ids = _build_chain(root, repo_id, 6) |
| 2089 | self._start(root, ids) |
| 2090 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2091 | assert result.exit_code == 0 |
| 2092 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2093 | if len(lines) > 1: |
| 2094 | step = json.loads(lines[0]) |
| 2095 | assert {"step", "testing", "verdict", "remaining_count", "done", "symbol_changes"} <= set(step.keys()) |
| 2096 | |
| 2097 | def test_run_json_done_line_keys(self, tmp_path: pathlib.Path) -> None: |
| 2098 | root, repo_id = _make_repo(tmp_path) |
| 2099 | ids = _build_chain(root, repo_id, 5) |
| 2100 | self._start(root, ids) |
| 2101 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2102 | assert result.exit_code == 0 |
| 2103 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2104 | done = json.loads(lines[-1]) |
| 2105 | assert set(done.keys()) == {"done", "first_bad", "steps_taken"} |
| 2106 | |
| 2107 | def test_run_json_done_true_on_last_line(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 | assert json.loads(lines[-1])["done"] is True |
| 2115 | |
| 2116 | def test_run_json_steps_taken_positive(self, tmp_path: pathlib.Path) -> None: |
| 2117 | root, repo_id = _make_repo(tmp_path) |
| 2118 | ids = _build_chain(root, repo_id, 6) |
| 2119 | self._start(root, ids) |
| 2120 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2121 | assert result.exit_code == 0 |
| 2122 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2123 | assert json.loads(lines[-1])["steps_taken"] >= 1 |
| 2124 | |
| 2125 | def test_run_json_verdict_good_with_true(self, tmp_path: pathlib.Path) -> None: |
| 2126 | root, repo_id = _make_repo(tmp_path) |
| 2127 | ids = _build_chain(root, repo_id, 5) |
| 2128 | self._start(root, ids) |
| 2129 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2130 | assert result.exit_code == 0 |
| 2131 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2132 | step_lines = lines[:-1] |
| 2133 | assert all(json.loads(l)["verdict"] == "good" for l in step_lines) |
| 2134 | |
| 2135 | def test_run_json_verdict_bad_with_false(self, tmp_path: pathlib.Path) -> None: |
| 2136 | root, repo_id = _make_repo(tmp_path) |
| 2137 | ids = _build_chain(root, repo_id, 5) |
| 2138 | self._start(root, ids) |
| 2139 | result = _invoke(root, ["bisect", "run", "false", "--json"]) |
| 2140 | assert result.exit_code == 0 |
| 2141 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2142 | step_lines = lines[:-1] |
| 2143 | assert all(json.loads(l)["verdict"] == "bad" for l in step_lines) |
| 2144 | |
| 2145 | def test_run_no_session_exits_1(self, tmp_path: pathlib.Path) -> None: |
| 2146 | root, _ = _make_repo(tmp_path) |
| 2147 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2148 | assert result.exit_code == 1 |
| 2149 | |
| 2150 | def test_run_no_session_error_message(self, tmp_path: pathlib.Path) -> None: |
| 2151 | root, _ = _make_repo(tmp_path) |
| 2152 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2153 | combined = result.output + (result.stderr or "") |
| 2154 | assert "No bisect session" in combined |
| 2155 | |
| 2156 | def test_run_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2157 | empty = tmp_path / "not_a_repo" |
| 2158 | empty.mkdir() |
| 2159 | result = _invoke(empty, ["bisect", "run", "true"]) |
| 2160 | assert result.exit_code == 2 |
| 2161 | |
| 2162 | def test_run_text_mentions_testing(self, tmp_path: pathlib.Path) -> None: |
| 2163 | root, repo_id = _make_repo(tmp_path) |
| 2164 | ids = _build_chain(root, repo_id, 5) |
| 2165 | self._start(root, ids) |
| 2166 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2167 | assert result.exit_code == 0 |
| 2168 | assert "Testing" in result.output or "→" in result.output |
| 2169 | |
| 2170 | def test_run_text_mentions_first_bad(self, tmp_path: pathlib.Path) -> None: |
| 2171 | root, repo_id = _make_repo(tmp_path) |
| 2172 | ids = _build_chain(root, repo_id, 5) |
| 2173 | self._start(root, ids) |
| 2174 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2175 | assert result.exit_code == 0 |
| 2176 | assert "First bad commit" in result.output or "Bisect complete" in result.output |
| 2177 | |
| 2178 | def test_run_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2179 | root, _ = _make_repo(tmp_path) |
| 2180 | result = _invoke(root, ["bisect", "run", "--help"]) |
| 2181 | assert "Agent quickstart" in result.output or "125" in result.output |
| 2182 | |
| 2183 | def test_run_json_step_numbers_increment(self, tmp_path: pathlib.Path) -> None: |
| 2184 | root, repo_id = _make_repo(tmp_path) |
| 2185 | ids = _build_chain(root, repo_id, 8) |
| 2186 | self._start(root, ids) |
| 2187 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2188 | assert result.exit_code == 0 |
| 2189 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2190 | step_nums = [json.loads(l)["step"] for l in lines[:-1]] |
| 2191 | assert step_nums == list(range(1, len(step_nums) + 1)) |
| 2192 | |
| 2193 | def test_run_json_remaining_nonincreasing(self, tmp_path: pathlib.Path) -> None: |
| 2194 | root, repo_id = _make_repo(tmp_path) |
| 2195 | ids = _build_chain(root, repo_id, 8) |
| 2196 | self._start(root, ids) |
| 2197 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2198 | assert result.exit_code == 0 |
| 2199 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2200 | counts = [json.loads(l)["remaining_count"] for l in lines[:-1]] |
| 2201 | assert all(counts[i] >= counts[i + 1] for i in range(len(counts) - 1)) |
| 2202 | |
| 2203 | def test_run_text_no_json_by_default(self, tmp_path: pathlib.Path) -> None: |
| 2204 | root, repo_id = _make_repo(tmp_path) |
| 2205 | ids = _build_chain(root, repo_id, 4) |
| 2206 | self._start(root, ids) |
| 2207 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2208 | assert result.exit_code == 0 |
| 2209 | # Text mode should not have a JSON object on a line by itself |
| 2210 | json_lines = [l for l in result.output.splitlines() if l.strip().startswith("{")] |
| 2211 | assert json_lines == [] |
| 2212 | |
| 2213 | def test_run_json_first_bad_set_on_done(self, tmp_path: pathlib.Path) -> None: |
| 2214 | root, repo_id = _make_repo(tmp_path) |
| 2215 | ids = _build_chain(root, repo_id, 5) |
| 2216 | self._start(root, ids) |
| 2217 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2218 | assert result.exit_code == 0 |
| 2219 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2220 | done = json.loads(lines[-1]) |
| 2221 | if done["done"]: |
| 2222 | assert done["first_bad"] is not None |
| 2223 | |
| 2224 | |
| 2225 | class TestBisectRunSecurity: |
| 2226 | """Security hardening tests for muse bisect run.""" |
| 2227 | |
| 2228 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2229 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2230 | assert r.exit_code == 0 |
| 2231 | |
| 2232 | def test_run_json_lines_are_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2233 | root, repo_id = _make_repo(tmp_path) |
| 2234 | ids = _build_chain(root, repo_id, 5) |
| 2235 | self._start(root, ids) |
| 2236 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2237 | assert result.exit_code == 0 |
| 2238 | for line in result.output.strip().splitlines(): |
| 2239 | if line.strip(): |
| 2240 | assert isinstance(json.loads(line.strip()), dict) |
| 2241 | |
| 2242 | def test_run_json_done_field_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2243 | root, repo_id = _make_repo(tmp_path) |
| 2244 | ids = _build_chain(root, repo_id, 5) |
| 2245 | self._start(root, ids) |
| 2246 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2247 | assert result.exit_code == 0 |
| 2248 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2249 | for line in lines: |
| 2250 | assert isinstance(json.loads(line)["done"], bool) |
| 2251 | |
| 2252 | def test_run_text_symbol_changes_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2253 | """ANSI codes in symbol_changes are stripped from text output during run.""" |
| 2254 | from unittest.mock import patch |
| 2255 | from muse.core.bisect import BisectResult |
| 2256 | root, repo_id = _make_repo(tmp_path) |
| 2257 | ids = _build_chain(root, repo_id, 5) |
| 2258 | self._start(root, ids) |
| 2259 | injected_result = BisectResult( |
| 2260 | done=True, |
| 2261 | first_bad=ids[2], |
| 2262 | next_to_test=None, |
| 2263 | remaining_count=0, |
| 2264 | steps_remaining=0, |
| 2265 | verdict="bad", |
| 2266 | symbol_changes=[], |
| 2267 | ) |
| 2268 | with patch("muse.cli.commands.bisect._symbol_ops_in_commit", |
| 2269 | return_value=["add func\x1b[31mred\x1b[0m"]), \ |
| 2270 | patch("muse.cli.commands.bisect.get_bisect_next", |
| 2271 | side_effect=[(ids[2], "billing.py::Invoice"), (None, "")]), \ |
| 2272 | patch("muse.cli.commands.bisect.run_bisect_command", |
| 2273 | return_value=injected_result): |
| 2274 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2275 | assert "\x1b" not in result.output |
| 2276 | |
| 2277 | def test_run_error_no_json_on_failure(self, tmp_path: pathlib.Path) -> None: |
| 2278 | root, _ = _make_repo(tmp_path) |
| 2279 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2280 | assert result.exit_code != 0 |
| 2281 | assert not result.output.strip().startswith("{") |
| 2282 | |
| 2283 | def test_run_json_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2284 | root, repo_id = _make_repo(tmp_path) |
| 2285 | ids = _build_chain(root, repo_id, 5) |
| 2286 | self._start(root, ids) |
| 2287 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2288 | assert result.exit_code == 0 |
| 2289 | assert "\x1b" not in result.output |
| 2290 | |
| 2291 | def test_run_text_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2292 | root, repo_id = _make_repo(tmp_path) |
| 2293 | ids = _build_chain(root, repo_id, 5) |
| 2294 | self._start(root, ids) |
| 2295 | result = _invoke(root, ["bisect", "run", "true"]) |
| 2296 | assert result.exit_code == 0 |
| 2297 | assert "\x1b" not in result.output |
| 2298 | |
| 2299 | |
| 2300 | class TestBisectRunStress: |
| 2301 | """Performance and scale tests for muse bisect run.""" |
| 2302 | |
| 2303 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2304 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2305 | assert r.exit_code == 0 |
| 2306 | |
| 2307 | def test_run_50_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 2308 | """run converges on a 50-commit chain with always-good command.""" |
| 2309 | root, repo_id = _make_repo(tmp_path) |
| 2310 | ids = _build_chain(root, repo_id, 50) |
| 2311 | self._start(root, ids) |
| 2312 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2313 | assert result.exit_code == 0 |
| 2314 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2315 | assert json.loads(lines[-1])["done"] is True |
| 2316 | |
| 2317 | def test_run_performance_20_commits(self, tmp_path: pathlib.Path) -> None: |
| 2318 | """run over 20 commits completes within 10 seconds.""" |
| 2319 | import time |
| 2320 | root, repo_id = _make_repo(tmp_path) |
| 2321 | ids = _build_chain(root, repo_id, 20) |
| 2322 | self._start(root, ids) |
| 2323 | t0 = time.monotonic() |
| 2324 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2325 | elapsed = time.monotonic() - t0 |
| 2326 | assert result.exit_code == 0 |
| 2327 | assert elapsed < 10.0, f"bisect run 20 commits took {elapsed:.2f}s" |
| 2328 | |
| 2329 | def test_run_steps_taken_within_log2(self, tmp_path: pathlib.Path) -> None: |
| 2330 | """Steps taken should be at most log2(n)+1 for an always-good command.""" |
| 2331 | import math |
| 2332 | root, repo_id = _make_repo(tmp_path) |
| 2333 | n = 32 |
| 2334 | ids = _build_chain(root, repo_id, n) |
| 2335 | self._start(root, ids) |
| 2336 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 2337 | assert result.exit_code == 0 |
| 2338 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 2339 | steps_taken = json.loads(lines[-1])["steps_taken"] |
| 2340 | assert steps_taken <= int(math.log2(n)) + 2 |
| 2341 | |
| 2342 | |
| 2343 | # --------------------------------------------------------------------------- |
| 2344 | # bisect log — Extended, Security, Stress |
| 2345 | # --------------------------------------------------------------------------- |
| 2346 | |
| 2347 | |
| 2348 | class TestBisectLogExtended: |
| 2349 | """Extended unit / integration / e2e tests for muse bisect log.""" |
| 2350 | |
| 2351 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2352 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2353 | assert r.exit_code == 0 |
| 2354 | |
| 2355 | def test_log_exits_0_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2356 | root, _ = _make_repo(tmp_path) |
| 2357 | result = _invoke(root, ["bisect", "log"]) |
| 2358 | assert result.exit_code == 0 |
| 2359 | |
| 2360 | def test_log_exits_0_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2361 | root, repo_id = _make_repo(tmp_path) |
| 2362 | ids = _build_chain(root, repo_id, 4) |
| 2363 | self._start(root, ids) |
| 2364 | result = _invoke(root, ["bisect", "log"]) |
| 2365 | assert result.exit_code == 0 |
| 2366 | |
| 2367 | def test_log_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2368 | root, _ = _make_repo(tmp_path) |
| 2369 | result = _invoke(root, ["bisect", "log", "-j"]) |
| 2370 | assert result.exit_code == 0 |
| 2371 | parsed = _parse_log(result.output) |
| 2372 | assert isinstance(parsed["active"], bool) |
| 2373 | |
| 2374 | def test_log_json_active_false_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2375 | root, _ = _make_repo(tmp_path) |
| 2376 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2377 | assert result.exit_code == 0 |
| 2378 | assert _parse_log(result.output)["active"] is False |
| 2379 | |
| 2380 | def test_log_json_active_true_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2381 | root, repo_id = _make_repo(tmp_path) |
| 2382 | ids = _build_chain(root, repo_id, 4) |
| 2383 | self._start(root, ids) |
| 2384 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2385 | assert result.exit_code == 0 |
| 2386 | assert _parse_log(result.output)["active"] is True |
| 2387 | |
| 2388 | def test_log_json_entries_empty_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2389 | root, _ = _make_repo(tmp_path) |
| 2390 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2391 | assert result.exit_code == 0 |
| 2392 | assert _parse_log(result.output)["entries"] == [] |
| 2393 | |
| 2394 | def test_log_json_entries_grow_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 2395 | root, repo_id = _make_repo(tmp_path) |
| 2396 | ids = _build_chain(root, repo_id, 6) |
| 2397 | self._start(root, ids) |
| 2398 | after_start = len(_parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"]) |
| 2399 | _invoke(root, ["bisect", "bad", ids[3]]) |
| 2400 | after_bad = len(_parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"]) |
| 2401 | assert after_bad > after_start |
| 2402 | |
| 2403 | def test_log_json_two_keys(self, tmp_path: pathlib.Path) -> None: |
| 2404 | root, _ = _make_repo(tmp_path) |
| 2405 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2406 | assert result.exit_code == 0 |
| 2407 | d = json.loads(_json_blob(result.output)) |
| 2408 | assert {"active", "entries"} <= set(d.keys()) |
| 2409 | |
| 2410 | def test_log_json_start_records_bad_and_good(self, tmp_path: pathlib.Path) -> None: |
| 2411 | root, repo_id = _make_repo(tmp_path) |
| 2412 | ids = _build_chain(root, repo_id, 4) |
| 2413 | self._start(root, ids) |
| 2414 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2415 | verdicts = [e["verdict"] for e in entries] |
| 2416 | assert "bad" in verdicts |
| 2417 | assert "good" in verdicts |
| 2418 | |
| 2419 | def test_log_json_entries_contain_commit_ids(self, tmp_path: pathlib.Path) -> None: |
| 2420 | root, repo_id = _make_repo(tmp_path) |
| 2421 | ids = _build_chain(root, repo_id, 4) |
| 2422 | self._start(root, ids) |
| 2423 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2424 | for entry in entries: |
| 2425 | # commit_id is stored with the sha256: prefix (71 chars total) |
| 2426 | assert entry["commit_id"].startswith("sha256:") |
| 2427 | |
| 2428 | def test_log_json_entries_are_dicts(self, tmp_path: pathlib.Path) -> None: |
| 2429 | root, repo_id = _make_repo(tmp_path) |
| 2430 | ids = _build_chain(root, repo_id, 3) |
| 2431 | self._start(root, ids) |
| 2432 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2433 | for e in entries: |
| 2434 | assert isinstance(e, dict) |
| 2435 | assert "commit_id" in e |
| 2436 | assert "verdict" in e |
| 2437 | assert "timestamp" in e |
| 2438 | |
| 2439 | def test_log_active_false_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2440 | root, repo_id = _make_repo(tmp_path) |
| 2441 | ids = _build_chain(root, repo_id, 3) |
| 2442 | self._start(root, ids) |
| 2443 | _invoke(root, ["bisect", "reset"]) |
| 2444 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2445 | assert result.exit_code == 0 |
| 2446 | assert _parse_log(result.output)["active"] is False |
| 2447 | |
| 2448 | def test_log_text_shows_bisect_log_header(self, tmp_path: pathlib.Path) -> None: |
| 2449 | root, repo_id = _make_repo(tmp_path) |
| 2450 | ids = _build_chain(root, repo_id, 4) |
| 2451 | self._start(root, ids) |
| 2452 | result = _invoke(root, ["bisect", "log"]) |
| 2453 | assert result.exit_code == 0 |
| 2454 | assert "Bisect log" in result.output |
| 2455 | |
| 2456 | def test_log_text_no_session_message(self, tmp_path: pathlib.Path) -> None: |
| 2457 | root, _ = _make_repo(tmp_path) |
| 2458 | result = _invoke(root, ["bisect", "log"]) |
| 2459 | assert result.exit_code == 0 |
| 2460 | assert "No bisect log" in result.output |
| 2461 | |
| 2462 | def test_log_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2463 | empty = tmp_path / "not_a_repo" |
| 2464 | empty.mkdir() |
| 2465 | result = _invoke(empty, ["bisect", "log"]) |
| 2466 | assert result.exit_code == 2 |
| 2467 | |
| 2468 | def test_log_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2469 | root, _ = _make_repo(tmp_path) |
| 2470 | result = _invoke(root, ["bisect", "log", "--help"]) |
| 2471 | assert "Agent quickstart" in result.output or "verdict" in result.output.lower() |
| 2472 | |
| 2473 | def test_log_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 2474 | root, repo_id = _make_repo(tmp_path) |
| 2475 | ids = _build_chain(root, repo_id, 4) |
| 2476 | self._start(root, ids) |
| 2477 | result = _invoke(root, ["bisect", "log"]) |
| 2478 | assert result.exit_code == 0 |
| 2479 | assert not any(l.strip().startswith("{") for l in result.output.splitlines()) |
| 2480 | |
| 2481 | |
| 2482 | class TestBisectLogSecurity: |
| 2483 | """Security hardening tests for muse bisect log.""" |
| 2484 | |
| 2485 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2486 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2487 | assert r.exit_code == 0 |
| 2488 | |
| 2489 | def test_log_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2490 | root, _ = _make_repo(tmp_path) |
| 2491 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2492 | assert result.exit_code == 0 |
| 2493 | d = json.loads(_json_blob(result.output)) |
| 2494 | assert isinstance(d, dict) |
| 2495 | |
| 2496 | def test_log_json_active_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2497 | root, _ = _make_repo(tmp_path) |
| 2498 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2499 | assert result.exit_code == 0 |
| 2500 | assert isinstance(json.loads(_json_blob(result.output))["active"], bool) |
| 2501 | |
| 2502 | def test_log_json_entries_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2503 | """ANSI codes injected into the log state are stripped from JSON output.""" |
| 2504 | from muse.core.bisect import _load_state, _save_state |
| 2505 | root, repo_id = _make_repo(tmp_path) |
| 2506 | ids = _build_chain(root, repo_id, 3) |
| 2507 | self._start(root, ids) |
| 2508 | # Tamper: inject ANSI into a log entry |
| 2509 | state = _load_state(root) |
| 2510 | assert state is not None |
| 2511 | state["log"].append(f"{ids[1]} bad\x1b[31m 2026-01-01T00:00:00\x1b[0m") |
| 2512 | _save_state(root, state) |
| 2513 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2514 | assert result.exit_code == 0 |
| 2515 | assert "\x1b" not in result.output |
| 2516 | |
| 2517 | def test_log_text_entries_sanitized(self, tmp_path: pathlib.Path) -> None: |
| 2518 | """ANSI codes in log entries are stripped from text output.""" |
| 2519 | from muse.core.bisect import _load_state, _save_state |
| 2520 | root, repo_id = _make_repo(tmp_path) |
| 2521 | ids = _build_chain(root, repo_id, 3) |
| 2522 | self._start(root, ids) |
| 2523 | state = _load_state(root) |
| 2524 | assert state is not None |
| 2525 | state["log"].append(f"{ids[1]} bad\x1b[31m 2026-01-01T00:00:00\x1b[0m") |
| 2526 | _save_state(root, state) |
| 2527 | result = _invoke(root, ["bisect", "log"]) |
| 2528 | assert result.exit_code == 0 |
| 2529 | assert "\x1b" not in result.output |
| 2530 | |
| 2531 | def test_log_json_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2532 | root, repo_id = _make_repo(tmp_path) |
| 2533 | ids = _build_chain(root, repo_id, 4) |
| 2534 | self._start(root, ids) |
| 2535 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2536 | assert result.exit_code == 0 |
| 2537 | assert "\x1b" not in result.output |
| 2538 | |
| 2539 | def test_log_text_no_ansi_in_output(self, tmp_path: pathlib.Path) -> None: |
| 2540 | root, repo_id = _make_repo(tmp_path) |
| 2541 | ids = _build_chain(root, repo_id, 4) |
| 2542 | self._start(root, ids) |
| 2543 | result = _invoke(root, ["bisect", "log"]) |
| 2544 | assert result.exit_code == 0 |
| 2545 | assert "\x1b" not in result.output |
| 2546 | |
| 2547 | |
| 2548 | class TestBisectLogStress: |
| 2549 | """Performance and scale tests for muse bisect log.""" |
| 2550 | |
| 2551 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2552 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2553 | assert r.exit_code == 0 |
| 2554 | |
| 2555 | def test_log_100_commit_session(self, tmp_path: pathlib.Path) -> None: |
| 2556 | """Log on a 100-step session returns all entries.""" |
| 2557 | root, repo_id = _make_repo(tmp_path) |
| 2558 | ids = _build_chain(root, repo_id, 100) |
| 2559 | self._start(root, ids) |
| 2560 | # Apply 10 good verdicts to build up a log |
| 2561 | for i in range(1, 11): |
| 2562 | _invoke(root, ["bisect", "good", ids[i]]) |
| 2563 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2564 | assert result.exit_code == 0 |
| 2565 | entries = _parse_log(result.output)["entries"] |
| 2566 | # start adds 2 entries; 10 good verdicts add 10 more |
| 2567 | assert len(entries) >= 12 |
| 2568 | |
| 2569 | def test_log_performance_large_session(self, tmp_path: pathlib.Path) -> None: |
| 2570 | """Log on a large session completes within 5 seconds.""" |
| 2571 | import time |
| 2572 | root, repo_id = _make_repo(tmp_path) |
| 2573 | ids = _build_chain(root, repo_id, 50) |
| 2574 | self._start(root, ids) |
| 2575 | for i in range(1, 8): |
| 2576 | _invoke(root, ["bisect", "bad", ids[i]]) |
| 2577 | t0 = time.monotonic() |
| 2578 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 2579 | elapsed = time.monotonic() - t0 |
| 2580 | assert result.exit_code == 0 |
| 2581 | assert elapsed < 5.0, f"bisect log took {elapsed:.2f}s" |
| 2582 | |
| 2583 | def test_log_concurrent_reads_consistent(self, tmp_path: pathlib.Path) -> None: |
| 2584 | """Concurrent log reads all return the same entry count.""" |
| 2585 | root, repo_id = _make_repo(tmp_path) |
| 2586 | ids = _build_chain(root, repo_id, 20) |
| 2587 | self._start(root, ids) |
| 2588 | _invoke(root, ["bisect", "bad", ids[10]]) |
| 2589 | counts: list[int] = [] |
| 2590 | errors: list[str] = [] |
| 2591 | lock = threading.Lock() |
| 2592 | |
| 2593 | def _run() -> None: |
| 2594 | r = _invoke(root, ["bisect", "log", "--json"]) |
| 2595 | with lock: |
| 2596 | if r.exit_code != 0: |
| 2597 | errors.append(r.output) |
| 2598 | return |
| 2599 | try: |
| 2600 | counts.append(len(_parse_log(r.output)["entries"])) |
| 2601 | except (json.JSONDecodeError, KeyError, ValueError) as exc: |
| 2602 | errors.append(f"parse error: {exc!r} output={r.output!r}") |
| 2603 | |
| 2604 | threads = [threading.Thread(target=_run) for _ in range(8)] |
| 2605 | for t in threads: |
| 2606 | t.start() |
| 2607 | for t in threads: |
| 2608 | t.join() |
| 2609 | assert not errors |
| 2610 | assert all(c == counts[0] for c in counts) |
| 2611 | |
| 2612 | |
| 2613 | # --------------------------------------------------------------------------- |
| 2614 | # bisect reset — Extended, Security, Stress |
| 2615 | # --------------------------------------------------------------------------- |
| 2616 | |
| 2617 | |
| 2618 | class TestBisectResetExtended: |
| 2619 | """Extended unit / integration / e2e tests for muse bisect reset.""" |
| 2620 | |
| 2621 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2622 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2623 | assert r.exit_code == 0 |
| 2624 | |
| 2625 | def test_reset_exits_0_with_session(self, tmp_path: pathlib.Path) -> None: |
| 2626 | root, repo_id = _make_repo(tmp_path) |
| 2627 | ids = _build_chain(root, repo_id, 4) |
| 2628 | self._start(root, ids) |
| 2629 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2630 | |
| 2631 | def test_reset_exits_0_no_session(self, tmp_path: pathlib.Path) -> None: |
| 2632 | root, _ = _make_repo(tmp_path) |
| 2633 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2634 | |
| 2635 | def test_reset_j_alias_works(self, tmp_path: pathlib.Path) -> None: |
| 2636 | root, _ = _make_repo(tmp_path) |
| 2637 | result = _invoke(root, ["bisect", "reset", "-j"]) |
| 2638 | assert result.exit_code == 0 |
| 2639 | assert _parse_reset(result.output)["reset"] is True |
| 2640 | |
| 2641 | def test_reset_json_reset_true(self, tmp_path: pathlib.Path) -> None: |
| 2642 | root, _ = _make_repo(tmp_path) |
| 2643 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2644 | assert result.exit_code == 0 |
| 2645 | assert _parse_reset(result.output)["reset"] is True |
| 2646 | |
| 2647 | def test_reset_json_single_key(self, tmp_path: pathlib.Path) -> None: |
| 2648 | root, _ = _make_repo(tmp_path) |
| 2649 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2650 | assert result.exit_code == 0 |
| 2651 | d = json.loads(_json_blob(result.output)) |
| 2652 | assert {"reset"} <= set(d.keys()) |
| 2653 | |
| 2654 | def test_reset_clears_active_session(self, tmp_path: pathlib.Path) -> None: |
| 2655 | root, repo_id = _make_repo(tmp_path) |
| 2656 | ids = _build_chain(root, repo_id, 4) |
| 2657 | self._start(root, ids) |
| 2658 | _invoke(root, ["bisect", "reset"]) |
| 2659 | log_r = _invoke(root, ["bisect", "log", "--json"]) |
| 2660 | assert _parse_log(log_r.output)["active"] is False |
| 2661 | |
| 2662 | def test_reset_prevents_bad_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2663 | root, repo_id = _make_repo(tmp_path) |
| 2664 | ids = _build_chain(root, repo_id, 4) |
| 2665 | self._start(root, ids) |
| 2666 | _invoke(root, ["bisect", "reset"]) |
| 2667 | result = _invoke(root, ["bisect", "bad", ids[2]]) |
| 2668 | assert result.exit_code == 1 |
| 2669 | |
| 2670 | def test_reset_prevents_good_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2671 | root, repo_id = _make_repo(tmp_path) |
| 2672 | ids = _build_chain(root, repo_id, 4) |
| 2673 | self._start(root, ids) |
| 2674 | _invoke(root, ["bisect", "reset"]) |
| 2675 | assert _invoke(root, ["bisect", "good", ids[1]]).exit_code == 1 |
| 2676 | |
| 2677 | def test_reset_prevents_skip_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2678 | root, repo_id = _make_repo(tmp_path) |
| 2679 | ids = _build_chain(root, repo_id, 4) |
| 2680 | self._start(root, ids) |
| 2681 | _invoke(root, ["bisect", "reset"]) |
| 2682 | assert _invoke(root, ["bisect", "skip", ids[2]]).exit_code == 1 |
| 2683 | |
| 2684 | def test_reset_idempotent_double_reset(self, tmp_path: pathlib.Path) -> None: |
| 2685 | root, repo_id = _make_repo(tmp_path) |
| 2686 | ids = _build_chain(root, repo_id, 3) |
| 2687 | self._start(root, ids) |
| 2688 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2689 | assert _invoke(root, ["bisect", "reset"]).exit_code == 0 |
| 2690 | |
| 2691 | def test_reset_allows_new_session_after(self, tmp_path: pathlib.Path) -> None: |
| 2692 | root, repo_id = _make_repo(tmp_path) |
| 2693 | ids = _build_chain(root, repo_id, 5) |
| 2694 | self._start(root, ids) |
| 2695 | _invoke(root, ["bisect", "reset"]) |
| 2696 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2697 | assert result.exit_code == 0 |
| 2698 | |
| 2699 | def test_reset_clears_log_entries(self, tmp_path: pathlib.Path) -> None: |
| 2700 | root, repo_id = _make_repo(tmp_path) |
| 2701 | ids = _build_chain(root, repo_id, 4) |
| 2702 | self._start(root, ids) |
| 2703 | _invoke(root, ["bisect", "bad", ids[2]]) |
| 2704 | _invoke(root, ["bisect", "reset"]) |
| 2705 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2706 | assert entries == [] |
| 2707 | |
| 2708 | def test_reset_text_output_mentions_reset(self, tmp_path: pathlib.Path) -> None: |
| 2709 | root, _ = _make_repo(tmp_path) |
| 2710 | result = _invoke(root, ["bisect", "reset"]) |
| 2711 | assert result.exit_code == 0 |
| 2712 | assert "reset" in result.output.lower() |
| 2713 | |
| 2714 | def test_reset_text_no_json_object(self, tmp_path: pathlib.Path) -> None: |
| 2715 | root, _ = _make_repo(tmp_path) |
| 2716 | result = _invoke(root, ["bisect", "reset"]) |
| 2717 | assert not result.output.strip().startswith("{") |
| 2718 | |
| 2719 | def test_reset_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2720 | empty = tmp_path / "not_a_repo" |
| 2721 | empty.mkdir() |
| 2722 | assert _invoke(empty, ["bisect", "reset"]).exit_code == 2 |
| 2723 | |
| 2724 | def test_reset_help_description_present(self, tmp_path: pathlib.Path) -> None: |
| 2725 | root, _ = _make_repo(tmp_path) |
| 2726 | result = _invoke(root, ["bisect", "reset", "--help"]) |
| 2727 | assert "Agent quickstart" in result.output or "Idempotent" in result.output |
| 2728 | |
| 2729 | def test_reset_json_reset_is_bool(self, tmp_path: pathlib.Path) -> None: |
| 2730 | root, _ = _make_repo(tmp_path) |
| 2731 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2732 | assert result.exit_code == 0 |
| 2733 | assert isinstance(json.loads(_json_blob(result.output))["reset"], bool) |
| 2734 | |
| 2735 | def test_reset_mid_session_with_verdicts(self, tmp_path: pathlib.Path) -> None: |
| 2736 | """Reset works correctly after several verdicts have been applied.""" |
| 2737 | root, repo_id = _make_repo(tmp_path) |
| 2738 | ids = _build_chain(root, repo_id, 10) |
| 2739 | self._start(root, ids) |
| 2740 | _invoke(root, ["bisect", "bad", ids[7]]) |
| 2741 | _invoke(root, ["bisect", "good", ids[3]]) |
| 2742 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2743 | assert result.exit_code == 0 |
| 2744 | assert _parse_reset(result.output)["reset"] is True |
| 2745 | assert _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["active"] is False |
| 2746 | |
| 2747 | |
| 2748 | class TestBisectResetSecurity: |
| 2749 | """Security hardening tests for muse bisect reset.""" |
| 2750 | |
| 2751 | def test_reset_json_is_valid_json(self, tmp_path: pathlib.Path) -> None: |
| 2752 | root, _ = _make_repo(tmp_path) |
| 2753 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2754 | assert result.exit_code == 0 |
| 2755 | assert isinstance(json.loads(_json_blob(result.output)), dict) |
| 2756 | |
| 2757 | def test_reset_json_no_ansi(self, tmp_path: pathlib.Path) -> None: |
| 2758 | root, _ = _make_repo(tmp_path) |
| 2759 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2760 | assert result.exit_code == 0 |
| 2761 | assert "\x1b" not in result.output |
| 2762 | |
| 2763 | def test_reset_text_no_ansi(self, tmp_path: pathlib.Path) -> None: |
| 2764 | root, _ = _make_repo(tmp_path) |
| 2765 | result = _invoke(root, ["bisect", "reset"]) |
| 2766 | assert result.exit_code == 0 |
| 2767 | assert "\x1b" not in result.output |
| 2768 | |
| 2769 | def test_reset_state_file_removed(self, tmp_path: pathlib.Path) -> None: |
| 2770 | """After reset the state file no longer exists on disk.""" |
| 2771 | from muse.core.bisect import _state_path |
| 2772 | root, repo_id = _make_repo(tmp_path) |
| 2773 | ids = _build_chain(root, repo_id, 3) |
| 2774 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2775 | assert _state_path(root).exists() |
| 2776 | _invoke(root, ["bisect", "reset"]) |
| 2777 | assert not _state_path(root).exists() |
| 2778 | |
| 2779 | def test_reset_no_session_state_file_absent(self, tmp_path: pathlib.Path) -> None: |
| 2780 | """Reset with no state file is a safe no-op.""" |
| 2781 | from muse.core.bisect import _state_path |
| 2782 | root, _ = _make_repo(tmp_path) |
| 2783 | assert not _state_path(root).exists() |
| 2784 | result = _invoke(root, ["bisect", "reset"]) |
| 2785 | assert result.exit_code == 0 |
| 2786 | |
| 2787 | def test_reset_json_reset_value_true(self, tmp_path: pathlib.Path) -> None: |
| 2788 | """reset field is always true, never false or a truthy int.""" |
| 2789 | root, _ = _make_repo(tmp_path) |
| 2790 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2791 | assert result.exit_code == 0 |
| 2792 | assert json.loads(_json_blob(result.output))["reset"] is True |
| 2793 | |
| 2794 | |
| 2795 | class TestBisectResetStress: |
| 2796 | """Performance and scale tests for muse bisect reset.""" |
| 2797 | |
| 2798 | def test_reset_after_100_commit_session(self, tmp_path: pathlib.Path) -> None: |
| 2799 | """Reset clears state from a 100-commit session instantly.""" |
| 2800 | root, repo_id = _make_repo(tmp_path) |
| 2801 | ids = _build_chain(root, repo_id, 100) |
| 2802 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2803 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 2804 | assert result.exit_code == 0 |
| 2805 | assert _parse_reset(result.output)["reset"] is True |
| 2806 | assert _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["active"] is False |
| 2807 | |
| 2808 | def test_reset_performance(self, tmp_path: pathlib.Path) -> None: |
| 2809 | """Reset completes within 2 seconds even after a large session.""" |
| 2810 | import time |
| 2811 | root, repo_id = _make_repo(tmp_path) |
| 2812 | ids = _build_chain(root, repo_id, 100) |
| 2813 | _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2814 | for i in range(1, 8): |
| 2815 | _invoke(root, ["bisect", "bad", ids[i]]) |
| 2816 | t0 = time.monotonic() |
| 2817 | result = _invoke(root, ["bisect", "reset"]) |
| 2818 | elapsed = time.monotonic() - t0 |
| 2819 | assert result.exit_code == 0 |
| 2820 | assert elapsed < 2.0, f"bisect reset took {elapsed:.2f}s" |
| 2821 | |
| 2822 | def test_reset_cycle_10_times(self, tmp_path: pathlib.Path) -> None: |
| 2823 | """Start → reset × 10 all succeed with no state leakage.""" |
| 2824 | root, repo_id = _make_repo(tmp_path) |
| 2825 | ids = _build_chain(root, repo_id, 6) |
| 2826 | for _ in range(10): |
| 2827 | r_start = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2828 | assert r_start.exit_code == 0 |
| 2829 | r_reset = _invoke(root, ["bisect", "reset", "--json"]) |
| 2830 | assert r_reset.exit_code == 0 |
| 2831 | assert _parse_reset(r_reset.output)["reset"] is True |
| 2832 | |
| 2833 | |
| 2834 | # =========================================================================== |
| 2835 | # New feature tests — status, structured log, timeout, symbol_changes in run |
| 2836 | # =========================================================================== |
| 2837 | |
| 2838 | |
| 2839 | class TestBisectStatus: |
| 2840 | """Tests for the new ``muse bisect status`` subcommand.""" |
| 2841 | |
| 2842 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2843 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2844 | assert r.exit_code == 0 |
| 2845 | |
| 2846 | # ── Unit: no session ──────────────────────────────────────────────────── |
| 2847 | |
| 2848 | def test_status_no_session_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 2849 | root, _ = _make_repo(tmp_path) |
| 2850 | result = _invoke(root, ["bisect", "status"]) |
| 2851 | assert result.exit_code == 0 |
| 2852 | |
| 2853 | def test_status_no_session_json_active_false(self, tmp_path: pathlib.Path) -> None: |
| 2854 | root, _ = _make_repo(tmp_path) |
| 2855 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2856 | assert result.exit_code == 0 |
| 2857 | d = json.loads(result.output.strip()) |
| 2858 | assert d["active"] is False |
| 2859 | |
| 2860 | def test_status_no_session_json_only_active_key(self, tmp_path: pathlib.Path) -> None: |
| 2861 | root, _ = _make_repo(tmp_path) |
| 2862 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2863 | assert result.exit_code == 0 |
| 2864 | d = json.loads(result.output.strip()) |
| 2865 | assert {"active"} <= set(d.keys()) |
| 2866 | |
| 2867 | # ── Integration: active session ───────────────────────────────────────── |
| 2868 | |
| 2869 | def test_status_active_session_exits_0(self, tmp_path: pathlib.Path) -> None: |
| 2870 | root, repo_id = _make_repo(tmp_path) |
| 2871 | ids = _build_chain(root, repo_id, 6) |
| 2872 | self._start(root, ids) |
| 2873 | result = _invoke(root, ["bisect", "status"]) |
| 2874 | assert result.exit_code == 0 |
| 2875 | |
| 2876 | def test_status_active_json_schema(self, tmp_path: pathlib.Path) -> None: |
| 2877 | root, repo_id = _make_repo(tmp_path) |
| 2878 | ids = _build_chain(root, repo_id, 6) |
| 2879 | self._start(root, ids) |
| 2880 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2881 | assert result.exit_code == 0 |
| 2882 | d = json.loads(result.output.strip()) |
| 2883 | assert d["active"] is True |
| 2884 | assert "bad_id" in d |
| 2885 | assert "good_ids" in d |
| 2886 | assert "remaining_count" in d |
| 2887 | assert "steps_remaining" in d |
| 2888 | assert "skipped_count" in d |
| 2889 | assert "symbol_filter" in d |
| 2890 | |
| 2891 | def test_status_active_remaining_count_positive(self, tmp_path: pathlib.Path) -> None: |
| 2892 | root, repo_id = _make_repo(tmp_path) |
| 2893 | ids = _build_chain(root, repo_id, 8) |
| 2894 | self._start(root, ids) |
| 2895 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2896 | d = json.loads(result.output.strip()) |
| 2897 | assert d["remaining_count"] > 0 |
| 2898 | |
| 2899 | def test_status_bad_id_matches_session(self, tmp_path: pathlib.Path) -> None: |
| 2900 | root, repo_id = _make_repo(tmp_path) |
| 2901 | ids = _build_chain(root, repo_id, 5) |
| 2902 | self._start(root, ids) |
| 2903 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2904 | d = json.loads(result.output.strip()) |
| 2905 | assert d["bad_id"] == ids[-1] |
| 2906 | |
| 2907 | def test_status_skipped_count_increments(self, tmp_path: pathlib.Path) -> None: |
| 2908 | root, repo_id = _make_repo(tmp_path) |
| 2909 | ids = _build_chain(root, repo_id, 8) |
| 2910 | self._start(root, ids) |
| 2911 | before = json.loads( |
| 2912 | _invoke(root, ["bisect", "status", "--json"]).output.strip() |
| 2913 | )["skipped_count"] |
| 2914 | # Skip the midpoint |
| 2915 | next_id = json.loads( |
| 2916 | _invoke(root, ["bisect", "status", "--json"]).output.strip() |
| 2917 | ) |
| 2918 | _invoke(root, ["bisect", "skip", ids[len(ids) // 2]]) |
| 2919 | after = json.loads( |
| 2920 | _invoke(root, ["bisect", "status", "--json"]).output.strip() |
| 2921 | )["skipped_count"] |
| 2922 | assert after == before + 1 |
| 2923 | |
| 2924 | def test_status_active_false_after_reset(self, tmp_path: pathlib.Path) -> None: |
| 2925 | root, repo_id = _make_repo(tmp_path) |
| 2926 | ids = _build_chain(root, repo_id, 5) |
| 2927 | self._start(root, ids) |
| 2928 | _invoke(root, ["bisect", "reset"]) |
| 2929 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2930 | d = json.loads(result.output.strip()) |
| 2931 | assert d["active"] is False |
| 2932 | |
| 2933 | # ── Security ──────────────────────────────────────────────────────────── |
| 2934 | |
| 2935 | def test_status_outside_repo_exits_2(self, tmp_path: pathlib.Path) -> None: |
| 2936 | empty = tmp_path / "not_a_repo" |
| 2937 | empty.mkdir() |
| 2938 | result = _invoke(empty, ["bisect", "status"]) |
| 2939 | assert result.exit_code == 2 |
| 2940 | |
| 2941 | def test_status_json_is_compact(self, tmp_path: pathlib.Path) -> None: |
| 2942 | """JSON output is compact single-line.""" |
| 2943 | root, repo_id = _make_repo(tmp_path) |
| 2944 | ids = _build_chain(root, repo_id, 5) |
| 2945 | self._start(root, ids) |
| 2946 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2947 | assert result.exit_code == 0 |
| 2948 | json.loads(result.output) |
| 2949 | |
| 2950 | def test_status_json_no_ansi(self, tmp_path: pathlib.Path) -> None: |
| 2951 | root, repo_id = _make_repo(tmp_path) |
| 2952 | ids = _build_chain(root, repo_id, 5) |
| 2953 | self._start(root, ids) |
| 2954 | result = _invoke(root, ["bisect", "status", "--json"]) |
| 2955 | assert "\x1b" not in result.output |
| 2956 | |
| 2957 | def test_status_text_no_session_message(self, tmp_path: pathlib.Path) -> None: |
| 2958 | root, _ = _make_repo(tmp_path) |
| 2959 | result = _invoke(root, ["bisect", "status"]) |
| 2960 | assert "No bisect session" in result.output or "no bisect session" in result.output.lower() |
| 2961 | |
| 2962 | def test_status_text_active_shows_remaining(self, tmp_path: pathlib.Path) -> None: |
| 2963 | root, repo_id = _make_repo(tmp_path) |
| 2964 | ids = _build_chain(root, repo_id, 6) |
| 2965 | self._start(root, ids) |
| 2966 | result = _invoke(root, ["bisect", "status"]) |
| 2967 | assert "remaining" in result.output.lower() |
| 2968 | |
| 2969 | |
| 2970 | class TestBisectLogStructured: |
| 2971 | """Tests verifying the new structured log entry schema.""" |
| 2972 | |
| 2973 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 2974 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 2975 | assert r.exit_code == 0 |
| 2976 | |
| 2977 | def test_log_entry_has_three_keys(self, tmp_path: pathlib.Path) -> None: |
| 2978 | root, repo_id = _make_repo(tmp_path) |
| 2979 | ids = _build_chain(root, repo_id, 4) |
| 2980 | self._start(root, ids) |
| 2981 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2982 | assert len(entries) >= 2 |
| 2983 | for e in entries: |
| 2984 | assert set(e.keys()) == {"commit_id", "verdict", "timestamp"} |
| 2985 | |
| 2986 | def test_log_entry_verdict_values(self, tmp_path: pathlib.Path) -> None: |
| 2987 | root, repo_id = _make_repo(tmp_path) |
| 2988 | ids = _build_chain(root, repo_id, 4) |
| 2989 | self._start(root, ids) |
| 2990 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2991 | verdicts = {e["verdict"] for e in entries} |
| 2992 | assert verdicts <= {"bad", "good", "skip"} |
| 2993 | |
| 2994 | def test_log_entry_timestamp_is_iso8601(self, tmp_path: pathlib.Path) -> None: |
| 2995 | root, repo_id = _make_repo(tmp_path) |
| 2996 | ids = _build_chain(root, repo_id, 4) |
| 2997 | self._start(root, ids) |
| 2998 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 2999 | for e in entries: |
| 3000 | # ISO8601 timestamps contain 'T' separating date from time |
| 3001 | assert "T" in e["timestamp"] or e["timestamp"] == "" |
| 3002 | |
| 3003 | def test_log_skip_entry_appears_after_skip(self, tmp_path: pathlib.Path) -> None: |
| 3004 | root, repo_id = _make_repo(tmp_path) |
| 3005 | ids = _build_chain(root, repo_id, 6) |
| 3006 | self._start(root, ids) |
| 3007 | _invoke(root, ["bisect", "skip", ids[2]]) |
| 3008 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 3009 | verdicts = [e["verdict"] for e in entries] |
| 3010 | assert "skip" in verdicts |
| 3011 | |
| 3012 | def test_log_entry_commit_ids_in_session_ids(self, tmp_path: pathlib.Path) -> None: |
| 3013 | root, repo_id = _make_repo(tmp_path) |
| 3014 | ids = _build_chain(root, repo_id, 4) |
| 3015 | self._start(root, ids) |
| 3016 | entries = _parse_log(_invoke(root, ["bisect", "log", "--json"]).output)["entries"] |
| 3017 | entry_ids = {e["commit_id"] for e in entries} |
| 3018 | # bad and good commit IDs from start should appear in log |
| 3019 | assert ids[-1] in entry_ids # bad |
| 3020 | assert ids[0] in entry_ids # good |
| 3021 | |
| 3022 | |
| 3023 | class TestBisectRunTimeout: |
| 3024 | """Tests for ``--timeout`` on ``muse bisect run``.""" |
| 3025 | |
| 3026 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 3027 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 3028 | assert r.exit_code == 0 |
| 3029 | |
| 3030 | def test_timeout_flag_accepted(self, tmp_path: pathlib.Path) -> None: |
| 3031 | """--timeout is a valid flag that doesn't crash the parser.""" |
| 3032 | root, repo_id = _make_repo(tmp_path) |
| 3033 | ids = _build_chain(root, repo_id, 4) |
| 3034 | self._start(root, ids) |
| 3035 | result = _invoke(root, ["bisect", "run", "true", "--timeout", "30"]) |
| 3036 | assert result.exit_code == 0 |
| 3037 | |
| 3038 | def test_timeout_fast_command_succeeds(self, tmp_path: pathlib.Path) -> None: |
| 3039 | """A command that finishes well within the timeout is treated normally.""" |
| 3040 | root, repo_id = _make_repo(tmp_path) |
| 3041 | ids = _build_chain(root, repo_id, 5) |
| 3042 | self._start(root, ids) |
| 3043 | result = _invoke(root, ["bisect", "run", "true", "--timeout", "10"]) |
| 3044 | assert result.exit_code == 0 |
| 3045 | |
| 3046 | def test_timeout_triggers_skip(self, tmp_path: pathlib.Path) -> None: |
| 3047 | """A command that exceeds --timeout is treated as skip (exit 125).""" |
| 3048 | from muse.core.bisect import run_bisect_command |
| 3049 | import tempfile |
| 3050 | |
| 3051 | with tempfile.TemporaryDirectory() as td: |
| 3052 | root_path = pathlib.Path(td) |
| 3053 | # We test the core directly to avoid actually sleeping in a test. |
| 3054 | # Patch subprocess.run to raise TimeoutExpired. |
| 3055 | import unittest.mock as mock |
| 3056 | from muse.core.bisect import _SKIP_EXIT_CODE |
| 3057 | # Build a minimal state so _apply_verdict can run. |
| 3058 | import datetime |
| 3059 | from muse.core.snapshot import compute_commit_id, compute_snapshot_id |
| 3060 | from muse.core.store import CommitRecord, SnapshotRecord, write_commit, write_snapshot |
| 3061 | from muse.core.bisect import start_bisect |
| 3062 | |
| 3063 | repo_id = fake_id("repo") |
| 3064 | dot_muse = muse_dir(root_path) |
| 3065 | dot_muse.mkdir() |
| 3066 | (dot_muse / "repo.json").write_text(json.dumps({ |
| 3067 | "repo_id": repo_id, "domain": "code", |
| 3068 | "default_branch": "main", "created_at": "2026-01-01T00:00:00+00:00", |
| 3069 | })) |
| 3070 | (dot_muse / "HEAD").write_text("ref: refs/heads/main") |
| 3071 | (dot_muse / "refs" / "heads").mkdir(parents=True) |
| 3072 | (dot_muse / "snapshots").mkdir() |
| 3073 | (dot_muse / "commits").mkdir() |
| 3074 | (dot_muse / "objects").mkdir() |
| 3075 | |
| 3076 | ids: list[str] = [] |
| 3077 | parent = None |
| 3078 | for i in range(4): |
| 3079 | manifest = {} |
| 3080 | snap_id = compute_snapshot_id(manifest) |
| 3081 | committed_at = datetime.datetime.now(datetime.timezone.utc) |
| 3082 | commit_id = compute_commit_id( parent_ids=[parent] if parent else [], |
| 3083 | snapshot_id=snap_id, |
| 3084 | message=f"c{i}", |
| 3085 | committed_at_iso=committed_at.isoformat(), |
| 3086 | ) |
| 3087 | write_snapshot(root_path, SnapshotRecord(snapshot_id=snap_id, manifest={}, created_at=committed_at)) |
| 3088 | write_commit(root_path, CommitRecord( |
| 3089 | commit_id=commit_id, repo_id=repo_id, |
| 3090 | parent_commit_id=parent, parent2_commit_id=None, |
| 3091 | snapshot_id=snap_id, branch="main", message=f"c{i}", |
| 3092 | committed_at=committed_at, |
| 3093 | )) |
| 3094 | (dot_muse / "refs" / "heads" / "main").write_text(commit_id) |
| 3095 | ids.append(commit_id) |
| 3096 | parent = commit_id |
| 3097 | |
| 3098 | start_bisect(root_path, ids[-1], [ids[0]]) |
| 3099 | |
| 3100 | import subprocess |
| 3101 | with mock.patch("subprocess.run", side_effect=subprocess.TimeoutExpired("cmd", 1)): |
| 3102 | result = run_bisect_command(root_path, "sleep 99", ids[2], timeout=1) |
| 3103 | assert result.verdict == "skip" |
| 3104 | |
| 3105 | def test_timeout_short_alias(self, tmp_path: pathlib.Path) -> None: |
| 3106 | """-t is the short alias for --timeout.""" |
| 3107 | root, repo_id = _make_repo(tmp_path) |
| 3108 | ids = _build_chain(root, repo_id, 4) |
| 3109 | self._start(root, ids) |
| 3110 | result = _invoke(root, ["bisect", "run", "true", "-t", "10"]) |
| 3111 | assert result.exit_code == 0 |
| 3112 | |
| 3113 | |
| 3114 | class TestBisectRunStepSymbolChanges: |
| 3115 | """Tests verifying symbol_changes is present in NDJSON step lines.""" |
| 3116 | |
| 3117 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 3118 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 3119 | assert r.exit_code == 0 |
| 3120 | |
| 3121 | def test_step_json_has_symbol_changes_key(self, tmp_path: pathlib.Path) -> None: |
| 3122 | root, repo_id = _make_repo(tmp_path) |
| 3123 | ids = _build_chain(root, repo_id, 6) |
| 3124 | self._start(root, ids) |
| 3125 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 3126 | assert result.exit_code == 0 |
| 3127 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 3128 | step_lines = [l for l in lines if '"step"' in l] |
| 3129 | if step_lines: |
| 3130 | step = json.loads(step_lines[0]) |
| 3131 | assert "symbol_changes" in step |
| 3132 | |
| 3133 | def test_step_symbol_changes_is_list(self, tmp_path: pathlib.Path) -> None: |
| 3134 | root, repo_id = _make_repo(tmp_path) |
| 3135 | ids = _build_chain(root, repo_id, 6) |
| 3136 | self._start(root, ids) |
| 3137 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 3138 | assert result.exit_code == 0 |
| 3139 | lines = [l.strip() for l in result.output.strip().splitlines() if l.strip()] |
| 3140 | for line in lines: |
| 3141 | obj = json.loads(line) |
| 3142 | if "symbol_changes" in obj: |
| 3143 | assert isinstance(obj["symbol_changes"], list) |
| 3144 | |
| 3145 | def test_step_ndjson_stays_compact(self, tmp_path: pathlib.Path) -> None: |
| 3146 | """NDJSON step lines must be single-line (not pretty-printed).""" |
| 3147 | root, repo_id = _make_repo(tmp_path) |
| 3148 | ids = _build_chain(root, repo_id, 6) |
| 3149 | self._start(root, ids) |
| 3150 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 3151 | assert result.exit_code == 0 |
| 3152 | for line in result.output.strip().splitlines(): |
| 3153 | line = line.strip() |
| 3154 | if not line: |
| 3155 | continue |
| 3156 | # Every non-empty line must be valid JSON on its own |
| 3157 | obj = json.loads(line) |
| 3158 | assert isinstance(obj, dict) |
| 3159 | |
| 3160 | |
| 3161 | class TestBisectJsonCompact: |
| 3162 | """Tests verifying compact single-line JSON on single-object subcommands.""" |
| 3163 | |
| 3164 | def _start(self, root: pathlib.Path, ids: list[str]) -> None: |
| 3165 | r = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]]) |
| 3166 | assert r.exit_code == 0 |
| 3167 | |
| 3168 | def test_start_json_is_compact(self, tmp_path: pathlib.Path) -> None: |
| 3169 | root, repo_id = _make_repo(tmp_path) |
| 3170 | ids = _build_chain(root, repo_id, 4) |
| 3171 | result = _invoke(root, ["bisect", "start", "--bad", ids[-1], "--good", ids[0], "--json"]) |
| 3172 | assert result.exit_code == 0 |
| 3173 | json.loads(result.output) |
| 3174 | |
| 3175 | def test_bad_json_is_compact(self, tmp_path: pathlib.Path) -> None: |
| 3176 | root, repo_id = _make_repo(tmp_path) |
| 3177 | ids = _build_chain(root, repo_id, 5) |
| 3178 | self._start(root, ids) |
| 3179 | result = _invoke(root, ["bisect", "bad", ids[-1], "--json"]) |
| 3180 | assert result.exit_code == 0 |
| 3181 | json.loads(result.output) |
| 3182 | |
| 3183 | def test_log_json_is_compact(self, tmp_path: pathlib.Path) -> None: |
| 3184 | root, repo_id = _make_repo(tmp_path) |
| 3185 | ids = _build_chain(root, repo_id, 4) |
| 3186 | self._start(root, ids) |
| 3187 | result = _invoke(root, ["bisect", "log", "--json"]) |
| 3188 | assert result.exit_code == 0 |
| 3189 | json.loads(result.output) |
| 3190 | |
| 3191 | def test_reset_json_is_compact(self, tmp_path: pathlib.Path) -> None: |
| 3192 | root, repo_id = _make_repo(tmp_path) |
| 3193 | result = _invoke(root, ["bisect", "reset", "--json"]) |
| 3194 | assert result.exit_code == 0 |
| 3195 | json.loads(result.output) |
| 3196 | |
| 3197 | def test_run_json_ndjson_lines_are_compact(self, tmp_path: pathlib.Path) -> None: |
| 3198 | """run --json emits NDJSON: each line is a compact single-line JSON object.""" |
| 3199 | root, repo_id = _make_repo(tmp_path) |
| 3200 | ids = _build_chain(root, repo_id, 4) |
| 3201 | self._start(root, ids) |
| 3202 | result = _invoke(root, ["bisect", "run", "true", "--json"]) |
| 3203 | assert result.exit_code == 0 |
| 3204 | for line in result.output.strip().splitlines(): |
| 3205 | line = line.strip() |
| 3206 | if not line: |
| 3207 | continue |
| 3208 | # Single-line JSON: no embedded newlines, parseable as-is |
| 3209 | obj = json.loads(line) |
| 3210 | assert isinstance(obj, dict) |
| 3211 | |
| 3212 | |
| 3213 | # --------------------------------------------------------------------------- |
| 3214 | # Flag registration tests |
| 3215 | # --------------------------------------------------------------------------- |
| 3216 | |
| 3217 | import argparse as _argparse |
| 3218 | from muse.cli.commands.bisect import register as _register_bisect |
| 3219 | from muse.core.paths import head_path, muse_dir, ref_path |
| 3220 | |
| 3221 | |
| 3222 | def _parse_bisect(*args: str) -> _argparse.Namespace: |
| 3223 | """Build an argument parser via register() and parse args.""" |
| 3224 | root_p = _argparse.ArgumentParser() |
| 3225 | subs = root_p.add_subparsers(dest="cmd") |
| 3226 | _register_bisect(subs) |
| 3227 | return root_p.parse_args(["bisect", *args]) |
| 3228 | |
| 3229 | |
| 3230 | class TestRegisterFlags: |
| 3231 | # ── bad ───────────────────────────────────────────────────────────────── |
| 3232 | def test_bad_default_json_out_is_false(self) -> None: |
| 3233 | ns = _parse_bisect("bad") |
| 3234 | assert ns.json_out is False |
| 3235 | |
| 3236 | def test_bad_json_flag_sets_json_out(self) -> None: |
| 3237 | ns = _parse_bisect("bad", "--json") |
| 3238 | assert ns.json_out is True |
| 3239 | |
| 3240 | def test_bad_j_shorthand_sets_json_out(self) -> None: |
| 3241 | ns = _parse_bisect("bad", "-j") |
| 3242 | assert ns.json_out is True |
| 3243 | |
| 3244 | # ── good ──────────────────────────────────────────────────────────────── |
| 3245 | def test_good_default_json_out_is_false(self) -> None: |
| 3246 | ns = _parse_bisect("good") |
| 3247 | assert ns.json_out is False |
| 3248 | |
| 3249 | def test_good_json_flag_sets_json_out(self) -> None: |
| 3250 | ns = _parse_bisect("good", "--json") |
| 3251 | assert ns.json_out is True |
| 3252 | |
| 3253 | def test_good_j_shorthand_sets_json_out(self) -> None: |
| 3254 | ns = _parse_bisect("good", "-j") |
| 3255 | assert ns.json_out is True |
| 3256 | |
| 3257 | # ── log ───────────────────────────────────────────────────────────────── |
| 3258 | def test_log_default_json_out_is_false(self) -> None: |
| 3259 | ns = _parse_bisect("log") |
| 3260 | assert ns.json_out is False |
| 3261 | |
| 3262 | def test_log_j_shorthand_sets_json_out(self) -> None: |
| 3263 | ns = _parse_bisect("log", "-j") |
| 3264 | assert ns.json_out is True |
| 3265 | |
| 3266 | # ── reset ──────────────────────────────────────────────────────────────── |
| 3267 | def test_reset_default_json_out_is_false(self) -> None: |
| 3268 | ns = _parse_bisect("reset") |
| 3269 | assert ns.json_out is False |
| 3270 | |
| 3271 | def test_reset_j_shorthand_sets_json_out(self) -> None: |
| 3272 | ns = _parse_bisect("reset", "-j") |
| 3273 | assert ns.json_out is True |
| 3274 | |
| 3275 | # ── run ───────────────────────────────────────────────────────────────── |
| 3276 | def test_run_default_json_out_is_false(self) -> None: |
| 3277 | ns = _parse_bisect("run", "pytest -x") |
| 3278 | assert ns.json_out is False |
| 3279 | |
| 3280 | def test_run_j_shorthand_sets_json_out(self) -> None: |
| 3281 | ns = _parse_bisect("run", "pytest -x", "-j") |
| 3282 | assert ns.json_out is True |
| 3283 | |
| 3284 | # ── skip ───────────────────────────────────────────────────────────────── |
| 3285 | def test_skip_default_json_out_is_false(self) -> None: |
| 3286 | ns = _parse_bisect("skip") |
| 3287 | assert ns.json_out is False |
| 3288 | |
| 3289 | def test_skip_j_shorthand_sets_json_out(self) -> None: |
| 3290 | ns = _parse_bisect("skip", "-j") |
| 3291 | assert ns.json_out is True |
| 3292 | |
| 3293 | # ── start ──────────────────────────────────────────────────────────────── |
| 3294 | def test_start_default_json_out_is_false(self) -> None: |
| 3295 | ns = _parse_bisect("start", "--bad", "HEAD", "--good", "v1.0.0") |
| 3296 | assert ns.json_out is False |
| 3297 | |
| 3298 | def test_start_j_shorthand_sets_json_out(self) -> None: |
| 3299 | ns = _parse_bisect("start", "--bad", "HEAD", "--good", "v1.0.0", "-j") |
| 3300 | assert ns.json_out is True |
| 3301 | |
| 3302 | def test_start_bad_flag(self) -> None: |
| 3303 | ns = _parse_bisect("start", "--bad", "HEAD", "--good", "v1.0.0") |
| 3304 | assert ns.bad == "HEAD" |
| 3305 | |
| 3306 | def test_start_good_flag(self) -> None: |
| 3307 | ns = _parse_bisect("start", "--bad", "HEAD", "--good", "v1.0.0") |
| 3308 | assert ns.good == ["v1.0.0"] |
| 3309 | |
| 3310 | # ── status ─────────────────────────────────────────────────────────────── |
| 3311 | def test_status_default_json_out_is_false(self) -> None: |
| 3312 | ns = _parse_bisect("status") |
| 3313 | assert ns.json_out is False |
| 3314 | |
| 3315 | def test_status_j_shorthand_sets_json_out(self) -> None: |
| 3316 | ns = _parse_bisect("status", "-j") |
| 3317 | assert ns.json_out is True |
File History
1 commit
sha256:51ce277f663e01a43eaffbe77509b1de7ac2d4251b55d23306304bcdeb92c90d
feat(pack): delta-encode snapshots in MPackBundle wire format
Sonnet 4.6
minor
⚠
122 days ago