| 1 | """Stronger-test #10 — ``run`` / ``gate`` / ``report --format junit`` agree. |
| 2 | |
| 3 | B16 pinned the gate's PASS/FAIL accounting against ``sway run``'s |
| 4 | rendered table but didn't pin the ``report --format junit`` |
| 5 | re-rendering of saved JSON. Schema drift between the writer and the |
| 6 | JUnit emitter would produce inconsistent PASS/FAIL tallies across the |
| 7 | three surfaces — a CI dashboard reading the JUnit file and a developer |
| 8 | reading the terminal output would see different numbers. |
| 9 | |
| 10 | This test runs a controlled spec through all three CLI paths against |
| 11 | a stubbed dummy backend (the same pattern used by |
| 12 | ``test_sway_gate_exit_code.py``) and asserts the per-verdict tally |
| 13 | is identical on every surface. |
| 14 | """ |
| 15 | |
| 16 | from __future__ import annotations |
| 17 | |
| 18 | import json |
| 19 | import xml.etree.ElementTree as ET |
| 20 | from collections import Counter |
| 21 | from pathlib import Path |
| 22 | |
| 23 | import pytest |
| 24 | from typer.testing import CliRunner |
| 25 | |
| 26 | from dlm_sway.backends.dummy import DummyDifferentialBackend, DummyResponses |
| 27 | from dlm_sway.cli.app import app |
| 28 | |
| 29 | |
| 30 | @pytest.fixture |
| 31 | def stub_build_backend(monkeypatch: pytest.MonkeyPatch) -> None: |
| 32 | """Replace ``backends.build`` with a dummy-returning factory so the |
| 33 | CLI commands run without loading a real HF model. Same shape as |
| 34 | ``tests/integration/test_sway_gate_exit_code.py``'s fixture.""" |
| 35 | |
| 36 | def _factory(*_args: object, **_kwargs: object) -> DummyDifferentialBackend: |
| 37 | return DummyDifferentialBackend(base=DummyResponses(), ft=DummyResponses()) |
| 38 | |
| 39 | import dlm_sway.backends as backends_mod |
| 40 | |
| 41 | monkeypatch.setattr(backends_mod, "build", _factory) |
| 42 | |
| 43 | |
| 44 | def _write_spec(path: Path) -> None: |
| 45 | """Two delta_kl probes with different thresholds — one passes on |
| 46 | the dummy backend's synthesized divergence, one fails.""" |
| 47 | path.write_text( |
| 48 | """ |
| 49 | version: 1 |
| 50 | models: |
| 51 | base: |
| 52 | base: stub |
| 53 | kind: hf |
| 54 | adapter: /tmp/stub |
| 55 | ft: |
| 56 | base: stub |
| 57 | kind: hf |
| 58 | adapter: /tmp/stub |
| 59 | defaults: |
| 60 | seed: 0 |
| 61 | coverage_threshold: 0.0 |
| 62 | suite: |
| 63 | - name: dk_loose |
| 64 | kind: delta_kl |
| 65 | prompts: [p1, p2, p3, p4] |
| 66 | assert_mean_gte: 0.0 |
| 67 | - name: dk_strict |
| 68 | kind: delta_kl |
| 69 | prompts: [p1, p2, p3, p4] |
| 70 | assert_mean_gte: 100.0 |
| 71 | """.strip() |
| 72 | ) |
| 73 | |
| 74 | |
| 75 | def _verdicts_from_json(payload: dict[str, object]) -> Counter[str]: |
| 76 | probes = payload["probes"] |
| 77 | assert isinstance(probes, list) |
| 78 | return Counter(str(p["verdict"]) for p in probes) |
| 79 | |
| 80 | |
| 81 | def _verdicts_from_junit(xml_text: str) -> Counter[str]: |
| 82 | """<testcase> without child → PASS; with failure/error/skipped → mapped accordingly.""" |
| 83 | root = ET.fromstring(xml_text) |
| 84 | counts: Counter[str] = Counter() |
| 85 | for case in root.iter("testcase"): |
| 86 | if case.find("error") is not None: |
| 87 | counts["error"] += 1 |
| 88 | elif case.find("failure") is not None: |
| 89 | counts["fail"] += 1 |
| 90 | elif case.find("skipped") is not None: |
| 91 | counts["skip"] += 1 |
| 92 | else: |
| 93 | counts["pass"] += 1 |
| 94 | return counts |
| 95 | |
| 96 | |
| 97 | def _verdicts_from_stdout(stdout: str) -> Counter[str]: |
| 98 | """Lift per-row verdict tokens out of the markdown/terminal |
| 99 | table rendered to stdout. The verdict column carries a single |
| 100 | lowercased keyword in the fixed set.""" |
| 101 | return Counter( |
| 102 | tok for tok in stdout.split() if tok in {"pass", "fail", "skip", "error", "warn"} |
| 103 | ) |
| 104 | |
| 105 | |
| 106 | def test_gate_run_junit_agree_on_verdicts( |
| 107 | stub_build_backend: None, # noqa: ARG001 |
| 108 | tmp_path: Path, |
| 109 | ) -> None: |
| 110 | spec = tmp_path / "sway.yaml" |
| 111 | _write_spec(spec) |
| 112 | runner = CliRunner() |
| 113 | |
| 114 | # ``sway run`` writes the JSON we'll re-render as JUnit. |
| 115 | json_out = tmp_path / "run.json" |
| 116 | run = runner.invoke(app, ["run", str(spec), "--json", str(json_out)]) |
| 117 | assert run.exit_code == 0, run.stdout |
| 118 | assert json_out.exists() |
| 119 | run_payload = json.loads(json_out.read_text(encoding="utf-8")) |
| 120 | run_verdicts = _verdicts_from_json(run_payload) |
| 121 | |
| 122 | # ``sway gate`` surfaces PASS/FAIL in stdout even when it exits 1. |
| 123 | gate = runner.invoke(app, ["gate", str(spec)]) |
| 124 | assert gate.exit_code == 1, gate.stdout # one failing probe |
| 125 | gate_verdicts = _verdicts_from_stdout(gate.stdout) |
| 126 | |
| 127 | # ``sway report --format junit`` re-renders the JSON file. |
| 128 | junit_out = tmp_path / "run.junit.xml" |
| 129 | rpt = runner.invoke( |
| 130 | app, ["report", str(json_out), "--format", "junit", "--out", str(junit_out)] |
| 131 | ) |
| 132 | assert rpt.exit_code == 0, rpt.stdout |
| 133 | junit_verdicts = _verdicts_from_junit(junit_out.read_text(encoding="utf-8")) |
| 134 | |
| 135 | assert run_verdicts == gate_verdicts == junit_verdicts, ( |
| 136 | f"surfaces disagree:\n" |
| 137 | f" run: {dict(run_verdicts)}\n" |
| 138 | f" gate: {dict(gate_verdicts)}\n" |
| 139 | f" junit: {dict(junit_verdicts)}" |
| 140 | ) |
| 141 | assert run_verdicts["pass"] == 1 |
| 142 | assert run_verdicts["fail"] == 1 |