test_cmd_bisect.py
python
sha256:d8316ffae901be06347e16ab55be11868eb519dd16ade3e8aa16a99e662f7e62
baseline: rc14 re-baseline after rc3 store corruption recovery
Human
patch
26 days ago
| 1 | """Comprehensive tests for ``muse bisect`` — binary search for bad commits. |
| 2 | |
| 3 | Coverage: |
| 4 | - Unit: bisect core functions (start, mark, skip, reset) |
| 5 | - Integration: CLI subcommands (start, bad, good, skip, log, reset) |
| 6 | - E2E: full bisect workflow resolving to a first-bad commit |
| 7 | - Security: invalid refs, session guard (no double-start), ref sanitization |
| 8 | - Stress: deep commit history bisect |
| 9 | """ |
| 10 | |
| 11 | from __future__ import annotations |
| 12 | |
| 13 | import datetime |
| 14 | import json |
| 15 | import pathlib |
| 16 | import uuid |
| 17 | |
| 18 | import pytest |
| 19 | from tests.cli_test_helper import CliRunner |
| 20 | |
| 21 | cli = None # argparse migration — CliRunner ignores this arg |
| 22 | from muse.core.store import CommitRecord, SnapshotRecord, write_commit, write_snapshot |
| 23 | from muse.core.snapshot import compute_commit_id, compute_snapshot_id |
| 24 | from muse.core._types import Manifest |
| 25 | |
| 26 | runner = CliRunner() |
| 27 | |
| 28 | |
| 29 | # --------------------------------------------------------------------------- |
| 30 | # Helpers |
| 31 | # --------------------------------------------------------------------------- |
| 32 | |
| 33 | |
| 34 | def _init_repo(tmp_path: pathlib.Path) -> tuple[pathlib.Path, str]: |
| 35 | repo_id = str(uuid.uuid4()) |
| 36 | muse = tmp_path / ".muse" |
| 37 | muse.mkdir() |
| 38 | (muse / "repo.json").write_text( |
| 39 | json.dumps({"repo_id": repo_id, "domain": "midi", |
| 40 | "default_branch": "main", |
| 41 | "created_at": "2026-01-01T00:00:00+00:00"}) |
| 42 | ) |
| 43 | (muse / "HEAD").write_text("ref: refs/heads/main") |
| 44 | (muse / "refs" / "heads").mkdir(parents=True) |
| 45 | (muse / "snapshots").mkdir() |
| 46 | (muse / "commits").mkdir() |
| 47 | (muse / "objects").mkdir() |
| 48 | return tmp_path, repo_id |
| 49 | |
| 50 | |
| 51 | def _env(root: pathlib.Path) -> Manifest: |
| 52 | return {"MUSE_REPO_ROOT": str(root)} |
| 53 | |
| 54 | |
| 55 | def _make_commit( |
| 56 | root: pathlib.Path, |
| 57 | repo_id: str, |
| 58 | *, |
| 59 | branch: str = "main", |
| 60 | message: str = "commit", |
| 61 | parent_id: str | None = None, |
| 62 | ) -> str: |
| 63 | manifest: Manifest = {} |
| 64 | snap_id = compute_snapshot_id(manifest) |
| 65 | committed_at = datetime.datetime.now(datetime.timezone.utc) |
| 66 | commit_id = compute_commit_id( |
| 67 | parent_ids=[parent_id] if parent_id else [], |
| 68 | snapshot_id=snap_id, |
| 69 | message=message, |
| 70 | committed_at_iso=committed_at.isoformat(), |
| 71 | ) |
| 72 | write_snapshot(root, SnapshotRecord(snapshot_id=snap_id, manifest=manifest)) |
| 73 | write_commit(root, CommitRecord( |
| 74 | commit_id=commit_id, |
| 75 | repo_id=repo_id, |
| 76 | branch=branch, |
| 77 | snapshot_id=snap_id, |
| 78 | message=message, |
| 79 | committed_at=committed_at, |
| 80 | parent_commit_id=parent_id, |
| 81 | )) |
| 82 | ref_file = root / ".muse" / "refs" / "heads" / branch |
| 83 | ref_file.parent.mkdir(parents=True, exist_ok=True) |
| 84 | ref_file.write_text(commit_id) |
| 85 | return commit_id |
| 86 | |
| 87 | |
| 88 | def _make_chain(root: pathlib.Path, repo_id: str, n: int) -> list[str]: |
| 89 | """Create a linear chain of n commits; return commit IDs oldest-first.""" |
| 90 | ids: list[str] = [] |
| 91 | parent: str | None = None |
| 92 | for i in range(n): |
| 93 | cid = _make_commit(root, repo_id, message=f"commit-{i}", parent_id=parent) |
| 94 | ids.append(cid) |
| 95 | parent = cid |
| 96 | return ids |
| 97 | |
| 98 | |
| 99 | # --------------------------------------------------------------------------- |
| 100 | # Unit tests — core bisect logic |
| 101 | # --------------------------------------------------------------------------- |
| 102 | |
| 103 | |
| 104 | class TestBisectCore: |
| 105 | def test_start_bisect_returns_result(self, tmp_path: pathlib.Path) -> None: |
| 106 | root, repo_id = _init_repo(tmp_path) |
| 107 | ids = _make_chain(root, repo_id, 4) |
| 108 | from muse.core.bisect import start_bisect |
| 109 | result = start_bisect(root, ids[-1], [ids[0]], branch="main") |
| 110 | assert result.next_to_test is not None or result.done |
| 111 | |
| 112 | def test_mark_bad_advances_search(self, tmp_path: pathlib.Path) -> None: |
| 113 | root, repo_id = _init_repo(tmp_path) |
| 114 | ids = _make_chain(root, repo_id, 8) |
| 115 | from muse.core.bisect import mark_bad, start_bisect |
| 116 | start_bisect(root, ids[-1], [ids[0]], branch="main") |
| 117 | result = mark_bad(root, ids[-1]) |
| 118 | assert not result.done or result.first_bad is not None |
| 119 | |
| 120 | def test_mark_good_advances_search(self, tmp_path: pathlib.Path) -> None: |
| 121 | root, repo_id = _init_repo(tmp_path) |
| 122 | ids = _make_chain(root, repo_id, 8) |
| 123 | from muse.core.bisect import mark_good, start_bisect |
| 124 | start_bisect(root, ids[-1], [ids[0]], branch="main") |
| 125 | result = mark_good(root, ids[0]) |
| 126 | assert result is not None |
| 127 | |
| 128 | def test_reset_clears_state(self, tmp_path: pathlib.Path) -> None: |
| 129 | root, repo_id = _init_repo(tmp_path) |
| 130 | ids = _make_chain(root, repo_id, 4) |
| 131 | from muse.core.bisect import is_bisect_active, reset_bisect, start_bisect |
| 132 | start_bisect(root, ids[-1], [ids[0]], branch="main") |
| 133 | assert is_bisect_active(root) |
| 134 | reset_bisect(root) |
| 135 | assert not is_bisect_active(root) |
| 136 | |
| 137 | def test_bisect_log_records_events(self, tmp_path: pathlib.Path) -> None: |
| 138 | root, repo_id = _init_repo(tmp_path) |
| 139 | ids = _make_chain(root, repo_id, 4) |
| 140 | from muse.core.bisect import get_bisect_log, start_bisect |
| 141 | start_bisect(root, ids[-1], [ids[0]], branch="main") |
| 142 | log = get_bisect_log(root) |
| 143 | assert len(log) > 0 |
| 144 | |
| 145 | |
| 146 | # --------------------------------------------------------------------------- |
| 147 | # Integration tests — CLI subcommands |
| 148 | # --------------------------------------------------------------------------- |
| 149 | |
| 150 | |
| 151 | class TestBisectCLI: |
| 152 | def test_start_requires_good_ref(self, tmp_path: pathlib.Path) -> None: |
| 153 | root, repo_id = _init_repo(tmp_path) |
| 154 | ids = _make_chain(root, repo_id, 2) |
| 155 | result = runner.invoke( |
| 156 | cli, ["bisect", "start", "--bad", ids[-1]], |
| 157 | env=_env(root) |
| 158 | ) |
| 159 | assert result.exit_code != 0 |
| 160 | assert "good" in result.output.lower() |
| 161 | |
| 162 | def test_start_with_bad_and_good(self, tmp_path: pathlib.Path) -> None: |
| 163 | root, repo_id = _init_repo(tmp_path) |
| 164 | ids = _make_chain(root, repo_id, 4) |
| 165 | result = runner.invoke( |
| 166 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 167 | env=_env(root), catch_exceptions=False |
| 168 | ) |
| 169 | assert result.exit_code == 0 |
| 170 | assert "started" in result.output.lower() or "next" in result.output.lower() |
| 171 | |
| 172 | def test_bad_without_session_fails(self, tmp_path: pathlib.Path) -> None: |
| 173 | root, repo_id = _init_repo(tmp_path) |
| 174 | ids = _make_chain(root, repo_id, 2) |
| 175 | result = runner.invoke(cli, ["bisect", "bad", ids[-1]], env=_env(root)) |
| 176 | assert result.exit_code != 0 |
| 177 | |
| 178 | def test_good_without_session_fails(self, tmp_path: pathlib.Path) -> None: |
| 179 | root, repo_id = _init_repo(tmp_path) |
| 180 | ids = _make_chain(root, repo_id, 2) |
| 181 | result = runner.invoke(cli, ["bisect", "good", ids[0]], env=_env(root)) |
| 182 | assert result.exit_code != 0 |
| 183 | |
| 184 | def test_skip_without_session_fails(self, tmp_path: pathlib.Path) -> None: |
| 185 | root, repo_id = _init_repo(tmp_path) |
| 186 | ids = _make_chain(root, repo_id, 2) |
| 187 | result = runner.invoke(cli, ["bisect", "skip", ids[0]], env=_env(root)) |
| 188 | assert result.exit_code != 0 |
| 189 | |
| 190 | def test_reset_clears_session(self, tmp_path: pathlib.Path) -> None: |
| 191 | root, repo_id = _init_repo(tmp_path) |
| 192 | ids = _make_chain(root, repo_id, 4) |
| 193 | runner.invoke( |
| 194 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 195 | env=_env(root) |
| 196 | ) |
| 197 | result = runner.invoke(cli, ["bisect", "reset"], env=_env(root), catch_exceptions=False) |
| 198 | assert result.exit_code == 0 |
| 199 | # After reset, bad should fail |
| 200 | result2 = runner.invoke(cli, ["bisect", "bad", ids[-1]], env=_env(root)) |
| 201 | assert result2.exit_code != 0 |
| 202 | |
| 203 | def test_log_shows_entries(self, tmp_path: pathlib.Path) -> None: |
| 204 | root, repo_id = _init_repo(tmp_path) |
| 205 | ids = _make_chain(root, repo_id, 4) |
| 206 | runner.invoke( |
| 207 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 208 | env=_env(root) |
| 209 | ) |
| 210 | result = runner.invoke(cli, ["bisect", "log"], env=_env(root), catch_exceptions=False) |
| 211 | assert result.exit_code == 0 |
| 212 | |
| 213 | def test_double_start_fails(self, tmp_path: pathlib.Path) -> None: |
| 214 | root, repo_id = _init_repo(tmp_path) |
| 215 | ids = _make_chain(root, repo_id, 4) |
| 216 | runner.invoke( |
| 217 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 218 | env=_env(root) |
| 219 | ) |
| 220 | result = runner.invoke( |
| 221 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 222 | env=_env(root) |
| 223 | ) |
| 224 | assert result.exit_code != 0 |
| 225 | assert "already" in result.output.lower() |
| 226 | |
| 227 | def test_bad_invalid_ref_fails(self, tmp_path: pathlib.Path) -> None: |
| 228 | root, repo_id = _init_repo(tmp_path) |
| 229 | ids = _make_chain(root, repo_id, 4) |
| 230 | runner.invoke( |
| 231 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 232 | env=_env(root) |
| 233 | ) |
| 234 | result = runner.invoke(cli, ["bisect", "bad", "deadbeef"], env=_env(root)) |
| 235 | assert result.exit_code != 0 |
| 236 | |
| 237 | def test_reset_without_session_succeeds(self, tmp_path: pathlib.Path) -> None: |
| 238 | """reset when no session is active should not crash.""" |
| 239 | root, _ = _init_repo(tmp_path) |
| 240 | result = runner.invoke(cli, ["bisect", "reset"], env=_env(root), catch_exceptions=False) |
| 241 | assert result.exit_code == 0 |
| 242 | |
| 243 | def test_log_empty_without_session(self, tmp_path: pathlib.Path) -> None: |
| 244 | root, _ = _init_repo(tmp_path) |
| 245 | result = runner.invoke(cli, ["bisect", "log"], env=_env(root), catch_exceptions=False) |
| 246 | assert result.exit_code == 0 |
| 247 | assert "no bisect" in result.output.lower() or result.output.strip() == "" or "no" in result.output.lower() |
| 248 | |
| 249 | |
| 250 | # --------------------------------------------------------------------------- |
| 251 | # E2E tests |
| 252 | # --------------------------------------------------------------------------- |
| 253 | |
| 254 | |
| 255 | class TestBisectE2E: |
| 256 | def test_full_bisect_workflow_2_commits(self, tmp_path: pathlib.Path) -> None: |
| 257 | """Start → mark good → mark bad → find first bad commit.""" |
| 258 | root, repo_id = _init_repo(tmp_path) |
| 259 | ids = _make_chain(root, repo_id, 2) |
| 260 | good_id, bad_id = ids[0], ids[1] |
| 261 | |
| 262 | runner.invoke( |
| 263 | cli, ["bisect", "start", "--bad", bad_id, "--good", good_id], |
| 264 | env=_env(root) |
| 265 | ) |
| 266 | # With only 2 commits, bisect should already identify bad_id |
| 267 | from muse.core.bisect import get_bisect_log |
| 268 | log = get_bisect_log(root) |
| 269 | assert len(log) >= 1 |
| 270 | |
| 271 | def test_full_bisect_workflow_many_commits(self, tmp_path: pathlib.Path) -> None: |
| 272 | """With a chain of 8 commits, bisect converges without error.""" |
| 273 | root, repo_id = _init_repo(tmp_path) |
| 274 | ids = _make_chain(root, repo_id, 8) |
| 275 | |
| 276 | runner.invoke( |
| 277 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 278 | env=_env(root), catch_exceptions=False |
| 279 | ) |
| 280 | |
| 281 | from muse.core.bisect import _load_state, is_bisect_active, mark_bad, mark_good |
| 282 | # Simulate binary search: assume the bug was introduced at ids[4] |
| 283 | max_steps = 20 |
| 284 | steps = 0 |
| 285 | done = False |
| 286 | while is_bisect_active(root) and steps < max_steps and not done: |
| 287 | state = _load_state(root) |
| 288 | if state is None: |
| 289 | break |
| 290 | remaining = state.get("remaining", []) |
| 291 | if not remaining: |
| 292 | break |
| 293 | mid = remaining[len(remaining) // 2] |
| 294 | # ids[4] and later are "bad" |
| 295 | if mid in ids[4:]: |
| 296 | result = mark_bad(root, mid) |
| 297 | else: |
| 298 | result = mark_good(root, mid) |
| 299 | done = result.done |
| 300 | steps += 1 |
| 301 | |
| 302 | # Bisect should have converged or be close |
| 303 | assert done or steps < max_steps |
| 304 | |
| 305 | |
| 306 | # --------------------------------------------------------------------------- |
| 307 | # Security tests |
| 308 | # --------------------------------------------------------------------------- |
| 309 | |
| 310 | |
| 311 | class TestBisectSecurity: |
| 312 | def test_ref_with_control_chars_is_rejected(self, tmp_path: pathlib.Path) -> None: |
| 313 | root, repo_id = _init_repo(tmp_path) |
| 314 | ids = _make_chain(root, repo_id, 2) |
| 315 | runner.invoke( |
| 316 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 317 | env=_env(root) |
| 318 | ) |
| 319 | # Inject control chars in a bad ref |
| 320 | result = runner.invoke(cli, ["bisect", "bad", "\x1b[31minjection\x1b[0m"], env=_env(root)) |
| 321 | assert result.exit_code != 0 |
| 322 | |
| 323 | def test_output_contains_no_ansi_on_invalid_ref(self, tmp_path: pathlib.Path) -> None: |
| 324 | root, repo_id = _init_repo(tmp_path) |
| 325 | ids = _make_chain(root, repo_id, 2) |
| 326 | runner.invoke( |
| 327 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 328 | env=_env(root) |
| 329 | ) |
| 330 | result = runner.invoke(cli, ["bisect", "bad", "nonexistent-ref\x1b[31m"], env=_env(root)) |
| 331 | assert "\x1b[31m" not in result.output |
| 332 | |
| 333 | |
| 334 | # --------------------------------------------------------------------------- |
| 335 | # Stress tests |
| 336 | # --------------------------------------------------------------------------- |
| 337 | |
| 338 | |
| 339 | class TestBisectStress: |
| 340 | def test_bisect_50_commit_chain(self, tmp_path: pathlib.Path) -> None: |
| 341 | """A 50-commit chain converges within log2(50) + 2 ≈ 8 steps.""" |
| 342 | root, repo_id = _init_repo(tmp_path) |
| 343 | ids = _make_chain(root, repo_id, 50) |
| 344 | bad_start = 25 # regression introduced at index 25 |
| 345 | |
| 346 | result = runner.invoke( |
| 347 | cli, ["bisect", "start", "--bad", ids[-1], "--good", ids[0]], |
| 348 | env=_env(root) |
| 349 | ) |
| 350 | assert result.exit_code == 0 |
| 351 | |
| 352 | from muse.core.bisect import _load_state, is_bisect_active, mark_bad, mark_good |
| 353 | max_steps = 10 # ceil(log2(48)) = 6; allow generous headroom |
| 354 | steps = 0 |
| 355 | done = False |
| 356 | while is_bisect_active(root) and steps < max_steps and not done: |
| 357 | state = _load_state(root) |
| 358 | if state is None: |
| 359 | break |
| 360 | remaining = state.get("remaining", []) |
| 361 | if not remaining: |
| 362 | break |
| 363 | mid = remaining[len(remaining) // 2] |
| 364 | idx = ids.index(mid) if mid in ids else -1 |
| 365 | if idx >= bad_start: |
| 366 | result = mark_bad(root, mid) |
| 367 | else: |
| 368 | result = mark_good(root, mid) |
| 369 | done = result.done |
| 370 | steps += 1 |
| 371 | |
| 372 | assert done or steps < max_steps, f"Bisect failed to converge in {steps} steps" |
File History
4 commits
sha256:d8316ffae901be06347e16ab55be11868eb519dd16ade3e8aa16a99e662f7e62
baseline: rc14 re-baseline after rc3 store corruption recovery
Human
patch
26 days ago
sha256:50d413a635f57761c3fce1f70251b957b6423821525bf330747ef4007339222e
Merge branch 'dev' into main
Human
29 days ago
sha256:fb67fed5a4d3e40de84bdd163de94ef1386570bef1dd1a020a732c8a038962ce
Merge branch 'dev' into main
Human
48 days ago
sha256:1c4b3e3a9a1f300774c3ee662b572a698d5fd405bf765a71e6011a2e9c3eaaaa
feat: Muse — version control for the agent era
Human
100 days ago