| 1 | """Tests for the agent row schema and the workflow oracles."""
|
| 2 |
|
| 3 | from __future__ import annotations
|
| 4 |
|
| 5 | import sys
|
| 6 | import tempfile
|
| 7 | from pathlib import Path
|
| 8 |
|
| 9 | sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "scripts"))
|
| 10 |
|
| 11 | import unittest
|
| 12 | from pathlib import Path
|
| 13 |
|
| 14 | import agent_metrics_schema
|
| 15 | import agent_workflow
|
| 16 | import workflow_ab
|
| 17 |
|
| 18 |
|
| 19 | class SchemaTest(unittest.TestCase):
|
| 20 | def test_example_row_validates(self) -> None:
|
| 21 | row = agent_metrics_schema.example_row()
|
| 22 | self.assertEqual(agent_metrics_schema.validate_row(row), [])
|
| 23 |
|
| 24 | def test_missing_group_rejected(self) -> None:
|
| 25 | row = agent_metrics_schema.example_row()
|
| 26 | del row["turn_metrics"]
|
| 27 | errors = agent_metrics_schema.validate_row(row)
|
| 28 | self.assertTrue(any("turn_metrics" in error for error in errors))
|
| 29 |
|
| 30 | def test_unknown_property_rejected(self) -> None:
|
| 31 | row = agent_metrics_schema.example_row()
|
| 32 | row["token_metrics"]["surprise"] = 1
|
| 33 | errors = agent_metrics_schema.validate_row(row)
|
| 34 | self.assertTrue(any("surprise" in error for error in errors))
|
| 35 |
|
| 36 | def test_nulls_are_legal_for_unmeasured_signals(self) -> None:
|
| 37 | # ADR-0002: unmeasured must be expressible as null, not forced to zero.
|
| 38 | row = agent_metrics_schema.example_row()
|
| 39 | row["turn_metrics"]["assistant_turns_total"] = None
|
| 40 | row["bytes_metrics"]["bytes_hydrated"] = None
|
| 41 | row["wall_clock_metrics"]["tool_wait_ms"] = None
|
| 42 | self.assertEqual(agent_metrics_schema.validate_row(row), [])
|
| 43 |
|
| 44 |
|
| 45 | class OracleShapeTest(unittest.TestCase):
|
| 46 | def workflow(self, name: str) -> workflow_ab.Workflow:
|
| 47 | return workflow_ab.WORKFLOWS[name]
|
| 48 |
|
| 49 | def test_bugfix_requires_the_one_file(self) -> None:
|
| 50 | wf = self.workflow("bugfix_test_loop")
|
| 51 | self.assertTrue(agent_workflow.expected_change_ok(wf, ["app/pricing.py"]))
|
| 52 | self.assertTrue(
|
| 53 | agent_workflow.expected_change_ok(
|
| 54 | wf,
|
| 55 | [
|
| 56 | "app/pricing.py",
|
| 57 | "app/__pycache__/pricing.cpython-313.pyc",
|
| 58 | "tests/__pycache__/test_pricing.cpython-313.pyc",
|
| 59 | ],
|
| 60 | )
|
| 61 | )
|
| 62 | self.assertFalse(agent_workflow.expected_change_ok(wf, []))
|
| 63 | self.assertFalse(agent_workflow.expected_change_ok(wf, ["app/pricing.py", "README.md"]))
|
| 64 |
|
| 65 | def test_workspace_manifest_ignores_python_bytecode_caches(self) -> None:
|
| 66 | with tempfile.TemporaryDirectory() as tmp:
|
| 67 | root = Path(tmp)
|
| 68 | (root / "app" / "__pycache__").mkdir(parents=True)
|
| 69 | (root / "tests" / "__pycache__").mkdir(parents=True)
|
| 70 | (root / "app" / "pricing.py").write_text("def f(): pass\n")
|
| 71 | (root / "app" / "__pycache__" / "pricing.cpython-313.pyc").write_bytes(b"cache")
|
| 72 | (root / "tests" / "__pycache__" / "test_pricing.cpython-313.pyc").write_bytes(b"cache")
|
| 73 |
|
| 74 | manifest = agent_workflow.workspace_manifest(root)
|
| 75 |
|
| 76 | self.assertEqual(list(manifest), ["app/pricing.py"])
|
| 77 |
|
| 78 | def test_history_archaeology_is_read_only(self) -> None:
|
| 79 | wf = self.workflow("history_archaeology")
|
| 80 | self.assertTrue(agent_workflow.expected_change_ok(wf, []))
|
| 81 | self.assertFalse(agent_workflow.expected_change_ok(wf, ["app/pricing.py"]))
|
| 82 |
|
| 83 | def test_error_recovery_allows_only_readme(self) -> None:
|
| 84 | wf = self.workflow("vcs_error_recovery")
|
| 85 | self.assertTrue(agent_workflow.expected_change_ok(wf, ["README.md"]))
|
| 86 | self.assertFalse(agent_workflow.expected_change_ok(wf, ["app/pricing.py"]))
|
| 87 |
|
| 88 |
|
| 89 | class WorkflowCatalogTest(unittest.TestCase):
|
| 90 | def test_snapshot_accounting_in_steps(self) -> None:
|
| 91 | # Workflow steps must follow the snapshot convention from the
|
| 92 | # command-semantics contract: two git calls, one oak call.
|
| 93 | git = workflow_ab.Subject(name="git", kind="git", label="Git", bin=Path("/usr/bin/git"))
|
| 94 | oak = workflow_ab.Subject(name="oak", kind="oak", label="Oak", bin=Path("/usr/local/bin/oak"))
|
| 95 | git_ops = [s.operation for s in workflow_ab.vcs_workflow_steps(git, "agent-default")]
|
| 96 | oak_ops = [s.operation for s in workflow_ab.vcs_workflow_steps(oak, "agent-default")]
|
| 97 | self.assertIn("vcs.snapshot.add", git_ops)
|
| 98 | self.assertIn("vcs.snapshot.commit", git_ops)
|
| 99 | self.assertIn("vcs.snapshot", oak_ops)
|
| 100 | self.assertNotIn("vcs.snapshot.add", oak_ops)
|
| 101 |
|
| 102 |
|
| 103 | if __name__ == "__main__":
|
| 104 | unittest.main()
|