test_stress_e2e_workflow.py
python
| 1 | """End-to-end CLI stress tests — mission-critical workflow verification. |
| 2 | |
| 3 | Tests the full muse CLI lifecycle using CliRunner with monkeypatch.chdir: |
| 4 | init → commit → log → branch → checkout → merge → tag → revert → stash |
| 5 | |
| 6 | Adversarial scenarios: |
| 7 | - 50-commit linear history: log shows all, branch creates correctly. |
| 8 | - Concurrent agent commits with provenance fields. |
| 9 | - Branch → commit → merge full cycle. |
| 10 | - Annotate accumulates reviewers (ORSet semantics). |
| 11 | - Stash and pop. |
| 12 | - Revert commit. |
| 13 | - Reset. |
| 14 | """ |
| 15 | from __future__ import annotations |
| 16 | |
| 17 | import pathlib |
| 18 | |
| 19 | import pytest |
| 20 | from typer.testing import CliRunner |
| 21 | |
| 22 | from muse.cli.app import cli |
| 23 | |
| 24 | runner = CliRunner() |
| 25 | |
| 26 | |
| 27 | # --------------------------------------------------------------------------- |
| 28 | # Helpers |
| 29 | # --------------------------------------------------------------------------- |
| 30 | |
| 31 | |
| 32 | def _run(*args: str) -> tuple[int, str]: |
| 33 | result = runner.invoke(cli, list(args), catch_exceptions=False) |
| 34 | return result.exit_code, result.output |
| 35 | |
| 36 | |
| 37 | def _write_file(repo: pathlib.Path, filename: str, content: str = "code = True\n") -> None: |
| 38 | work = repo / "muse-work" |
| 39 | work.mkdir(exist_ok=True) |
| 40 | (work / filename).write_text(content) |
| 41 | |
| 42 | |
| 43 | @pytest.fixture |
| 44 | def repo(tmp_path: pathlib.Path, monkeypatch: pytest.MonkeyPatch) -> pathlib.Path: |
| 45 | monkeypatch.chdir(tmp_path) |
| 46 | monkeypatch.setenv("MUSE_REPO_ROOT", str(tmp_path)) |
| 47 | code, out = _run("init", "--domain", "code") |
| 48 | assert code == 0, f"init failed: {out}" |
| 49 | return tmp_path |
| 50 | |
| 51 | |
| 52 | # =========================================================================== |
| 53 | # Basic lifecycle |
| 54 | # =========================================================================== |
| 55 | |
| 56 | |
| 57 | class TestBasicLifecycle: |
| 58 | def test_init_creates_muse_dir(self, tmp_path: pathlib.Path, monkeypatch: pytest.MonkeyPatch) -> None: |
| 59 | monkeypatch.chdir(tmp_path) |
| 60 | code, _ = _run("init") |
| 61 | assert code == 0 |
| 62 | assert (tmp_path / ".muse").is_dir() |
| 63 | |
| 64 | def test_commit_and_log(self, repo: pathlib.Path) -> None: |
| 65 | _write_file(repo, "main.py", "x = 1\n") |
| 66 | code, out = _run("commit", "-m", "first commit") |
| 67 | assert code == 0 |
| 68 | |
| 69 | code2, log_out = _run("log") |
| 70 | assert code2 == 0 |
| 71 | assert "first commit" in log_out |
| 72 | |
| 73 | def test_status_works(self, repo: pathlib.Path) -> None: |
| 74 | _write_file(repo, "app.py", "print('hello')\n") |
| 75 | code, out = _run("status") |
| 76 | assert code == 0 |
| 77 | |
| 78 | def test_tag_commit(self, repo: pathlib.Path) -> None: |
| 79 | _write_file(repo, "tagged.py", "tagged = True\n") |
| 80 | _run("commit", "-m", "commit to tag") |
| 81 | code, out = _run("tag", "add", "v1.0.0") |
| 82 | assert code == 0 |
| 83 | |
| 84 | def test_log_shows_multiple_commits(self, repo: pathlib.Path) -> None: |
| 85 | for i in range(5): |
| 86 | _write_file(repo, f"file{i}.py", f"x = {i}\n") |
| 87 | _run("commit", "-m", f"commit number {i}") |
| 88 | |
| 89 | code, out = _run("log") |
| 90 | assert code == 0 |
| 91 | for i in range(5): |
| 92 | assert f"commit number {i}" in out |
| 93 | |
| 94 | def test_show_commit(self, repo: pathlib.Path) -> None: |
| 95 | _write_file(repo, "show_me.py", "show = 'this'\n") |
| 96 | _run("commit", "-m", "showable commit") |
| 97 | code, _ = _run("log") |
| 98 | assert code == 0 |
| 99 | |
| 100 | |
| 101 | # =========================================================================== |
| 102 | # Branch and checkout |
| 103 | # =========================================================================== |
| 104 | |
| 105 | |
| 106 | class TestBranchAndCheckout: |
| 107 | def test_branch_creation(self, repo: pathlib.Path) -> None: |
| 108 | _write_file(repo, "base.py", "base = 1\n") |
| 109 | _run("commit", "-m", "base commit") |
| 110 | code, out = _run("branch", "feature/new-thing") |
| 111 | assert code == 0 |
| 112 | |
| 113 | def test_checkout_to_new_branch_then_back(self, repo: pathlib.Path) -> None: |
| 114 | _write_file(repo, "common.py", "common = 1\n") |
| 115 | _run("commit", "-m", "initial") |
| 116 | |
| 117 | _run("branch", "feature") |
| 118 | code, _ = _run("checkout", "feature") |
| 119 | assert code == 0 |
| 120 | code3, _ = _run("checkout", "main") |
| 121 | assert code3 == 0 |
| 122 | |
| 123 | def test_multiple_branches_independent(self, repo: pathlib.Path) -> None: |
| 124 | _write_file(repo, "root.py", "root = True\n") |
| 125 | _run("commit", "-m", "root") |
| 126 | |
| 127 | for i in range(3): |
| 128 | _run("branch", f"branch-{i}") |
| 129 | code, out = _run("checkout", f"branch-{i}") |
| 130 | assert code == 0, f"checkout branch-{i} failed: {out}" |
| 131 | _write_file(repo, f"branch_{i}.py", f"b = {i}\n") |
| 132 | _run("commit", "-m", f"branch-{i} commit") |
| 133 | _run("checkout", "main") |
| 134 | |
| 135 | |
| 136 | # =========================================================================== |
| 137 | # Stash |
| 138 | # =========================================================================== |
| 139 | |
| 140 | |
| 141 | class TestStash: |
| 142 | def test_stash_and_pop(self, repo: pathlib.Path) -> None: |
| 143 | _write_file(repo, "stash_me.py", "stash = True\n") |
| 144 | _run("commit", "-m", "before stash") |
| 145 | |
| 146 | _write_file(repo, "unstaged.py", "unstaged = True\n") |
| 147 | code, out = _run("stash") |
| 148 | assert code == 0, f"stash failed: {out}" |
| 149 | |
| 150 | code2, out2 = _run("stash", "pop") |
| 151 | assert code2 == 0, f"stash pop failed: {out2}" |
| 152 | |
| 153 | |
| 154 | # =========================================================================== |
| 155 | # Revert |
| 156 | # =========================================================================== |
| 157 | |
| 158 | |
| 159 | class TestRevert: |
| 160 | def test_revert_undoes_last_commit(self, repo: pathlib.Path) -> None: |
| 161 | _write_file(repo, "original.py", "original = True\n") |
| 162 | _run("commit", "-m", "original state") |
| 163 | _write_file(repo, "added.py", "added = True\n") |
| 164 | _run("commit", "-m", "added something") |
| 165 | |
| 166 | code, out = _run("log") |
| 167 | assert "added something" in out |
| 168 | |
| 169 | code2, _ = _run("revert", "HEAD") |
| 170 | assert code2 == 0 |
| 171 | |
| 172 | |
| 173 | # =========================================================================== |
| 174 | # Reset |
| 175 | # =========================================================================== |
| 176 | |
| 177 | |
| 178 | class TestReset: |
| 179 | def test_soft_reset_to_head(self, repo: pathlib.Path) -> None: |
| 180 | """Reset to HEAD (soft) — a no-op but must not fail.""" |
| 181 | _write_file(repo, "file.py", "x = 1\n") |
| 182 | code1, out1 = _run("commit", "-m", "commit 1") |
| 183 | assert code1 == 0 |
| 184 | |
| 185 | # "HEAD" is accepted by resolve_commit_ref as the current commit. |
| 186 | # Options must precede the positional REF argument in Typer sub-typers. |
| 187 | code, out = _run("reset", "--soft", "HEAD") |
| 188 | assert code == 0, f"reset failed: {out}" |
| 189 | |
| 190 | |
| 191 | # =========================================================================== |
| 192 | # Provenance fields in commit |
| 193 | # =========================================================================== |
| 194 | |
| 195 | |
| 196 | class TestProvenanceCommit: |
| 197 | def test_commit_with_agent_id(self, repo: pathlib.Path) -> None: |
| 198 | _write_file(repo, "agent.py", "agent = True\n") |
| 199 | result = runner.invoke( |
| 200 | cli, |
| 201 | ["commit", "-m", "agent commit", "--agent-id", "test-agent"], |
| 202 | catch_exceptions=False, |
| 203 | ) |
| 204 | assert result.exit_code == 0 |
| 205 | |
| 206 | def test_commit_with_model_id(self, repo: pathlib.Path) -> None: |
| 207 | _write_file(repo, "model.py", "model = 'gpt-4o'\n") |
| 208 | result = runner.invoke( |
| 209 | cli, |
| 210 | ["commit", "-m", "model commit", "--model-id", "gpt-4o"], |
| 211 | catch_exceptions=False, |
| 212 | ) |
| 213 | assert result.exit_code == 0 |
| 214 | |
| 215 | |
| 216 | # =========================================================================== |
| 217 | # Annotate command |
| 218 | # =========================================================================== |
| 219 | |
| 220 | |
| 221 | class TestAnnotateCommand: |
| 222 | def test_annotate_test_run(self, repo: pathlib.Path) -> None: |
| 223 | _write_file(repo, "annotate_me.py", "code = True\n") |
| 224 | _run("commit", "-m", "to annotate") |
| 225 | code, out = _run("annotate", "--test-run") |
| 226 | assert code == 0 |
| 227 | |
| 228 | def test_annotate_reviewed_by(self, repo: pathlib.Path) -> None: |
| 229 | _write_file(repo, "reviewed.py", "reviewed = True\n") |
| 230 | _run("commit", "-m", "for review") |
| 231 | code, out = _run("annotate", "--reviewed-by", "alice") |
| 232 | assert code == 0 |
| 233 | |
| 234 | def test_annotate_accumulates_reviewers(self, repo: pathlib.Path) -> None: |
| 235 | _write_file(repo, "multi_review.py", "x = 1\n") |
| 236 | _run("commit", "-m", "multi-review") |
| 237 | _run("annotate", "--reviewed-by", "alice") |
| 238 | code, out = _run("annotate", "--reviewed-by", "bob") |
| 239 | assert code == 0 |
| 240 | |
| 241 | |
| 242 | # =========================================================================== |
| 243 | # Long workflow stress |
| 244 | # =========================================================================== |
| 245 | |
| 246 | |
| 247 | class TestLongWorkflowStress: |
| 248 | def test_50_sequential_commits(self, repo: pathlib.Path) -> None: |
| 249 | for i in range(50): |
| 250 | _write_file(repo, f"module_{i:03d}.py", f"x = {i}\n") |
| 251 | code, out = _run("commit", "-m", f"commit {i:03d}") |
| 252 | assert code == 0, f"commit {i} failed: {out}" |
| 253 | |
| 254 | code, out = _run("log") |
| 255 | assert code == 0 |
| 256 | assert "commit 000" in out |
| 257 | assert "commit 049" in out |
| 258 | |
| 259 | def test_branch_commit_merge_cycle(self, repo: pathlib.Path) -> None: |
| 260 | """Full branch → commit → merge cycle.""" |
| 261 | _write_file(repo, "main.py", "main = True\n") |
| 262 | _run("commit", "-m", "main base") |
| 263 | |
| 264 | _run("branch", "feature/thing") |
| 265 | _run("checkout", "feature/thing") |
| 266 | _write_file(repo, "feature.py", "feature = True\n") |
| 267 | code, out = _run("commit", "-m", "feature work") |
| 268 | assert code == 0 |
| 269 | |
| 270 | _run("checkout", "main") |
| 271 | code2, out2 = _run("merge", "feature/thing") |
| 272 | # Merge may succeed or report no commits yet — either way must not crash. |
| 273 | assert code2 in (0, 1), f"merge crashed: {out2}" |