Mechanical refactor transform: mirror autonomy test layout

transform_autonomy_test_layout.py

#!/usr/bin/env python3
"""Reproduce the F-X1-1 autonomy test layout cleanup.

Run from the repo root: python3 /tmp/transform_autonomy_test_layout.py
"""

import sys
from pathlib import Path

sys.path.append(".claude/skills/mechanical-refactor-verify")
from mechanical_refactor_verify_utils import git_add_and_commit, verify_mechanical_refactor

BASE_COMMIT = "9bfa2e4e4e38465cd7d5aae77c4a3f0a9889e459"
TARGET_COMMIT = "6593af78182bf0e8df4973b6749401bdda4d0ce1"


def _replace(path: Path, old: str, new: str) -> None:
    content = path.read_text()
    updated = content.replace(old, new)
    if updated == content:
        raise RuntimeError(f"replacement not found in {path}: {old!r}")
    path.write_text(updated)


def _move(dir_root: Path, source_name: str, target_name: str) -> None:
    source = dir_root / source_name
    target = dir_root / target_name
    target.parent.mkdir(parents=True, exist_ok=True)
    source.rename(target)


def _logging_tests(dir_root: Path) -> None:
    target_name = "autonomy/tests/loggings/test_rollout_metrics.py"
    _move(dir_root, "autonomy/tests/test_rollout_logging.py", target_name)
    _replace(
        dir_root / target_name,
        "autonomy.tests.test_rollout_logging.reference_rollout_metric_reward",
        "autonomy.tests.loggings.test_rollout_metrics.reference_rollout_metric_reward",
    )
    _replace(
        dir_root / "autonomy/README_CLAUDE.md",
        "[test_rollout_logging.py](tests/test_rollout_logging.py)",
        "[test_rollout_metrics.py](tests/loggings/test_rollout_metrics.py)",
    )
    git_add_and_commit("[tests, logging] refactor: mirror rollout logging tests", cwd=str(dir_root))


def _async_rollout_tests(dir_root: Path) -> None:
    for test_name in (
        "test_distributed_rollout.py",
        "test_fully_async_rollout.py",
        "test_generate_interface.py",
        "test_strands_distributed_rollout.py",
    ):
        _move(
            dir_root,
            f"autonomy/tests/{test_name}",
            f"autonomy/tests/rollouts/{test_name}",
        )

    readme = dir_root / "autonomy/README_CLAUDE.md"
    _replace(
        readme,
        "│   ├── test_generate_interface.py",
        "│   ├── rollouts/test_generate_interface.py",
    )
    _replace(
        readme,
        "│   ├── test_distributed_rollout.py",
        "│   ├── rollouts/test_distributed_rollout.py",
    )
    for test_name in (
        "test_distributed_rollout.py",
        "test_fully_async_rollout.py",
        "test_generate_interface.py",
    ):
        _replace(
            readme,
            f"tests/{test_name}",
            f"tests/rollouts/{test_name}",
        )
    git_add_and_commit("[tests, rollout] refactor: mirror async rollout tests", cwd=str(dir_root))


def _multi_turn_tests(dir_root: Path) -> None:
    moves = {
        "autonomy/tests/test_rollout.py": "autonomy/tests/rollouts/test_multi_turn_vlm.py",
        "autonomy/tests/test_nova_act_prompt_golden.py": (
            "autonomy/tests/rollouts/test_nova_act_prompt_golden.py"
        ),
        "autonomy/tests/test_total_generation_length.py": (
            "autonomy/tests/rollouts/test_total_generation_length.py"
        ),
        "autonomy/tests/fake_env.py": "autonomy/tests/fixtures/fake_env.py",
    }
    for source_name, target_name in moves.items():
        _move(dir_root, source_name, target_name)

    for relative_path in (
        "autonomy/tests/rollouts/test_multi_turn_vlm.py",
        "autonomy/tests/rollouts/test_nova_act_prompt_golden.py",
    ):
        _replace(
            dir_root / relative_path,
            "autonomy.tests.fake_env",
            "autonomy.tests.fixtures.fake_env",
        )

    for relative_path in ("autonomy/runbook.md", "autonomy/README_CLAUDE.md"):
        _replace(
            dir_root / relative_path,
            "autonomy/tests/test_rollout.py",
            "autonomy/tests/rollouts/test_multi_turn_vlm.py",
        )

    readme = dir_root / "autonomy/README_CLAUDE.md"
    replacements = {
        "│   ├── test_rollout.py": "│   ├── rollouts/test_multi_turn_vlm.py",
        "│   ├── fake_env.py": "│   ├── fixtures/fake_env.py",
        "[test_rollout.py](tests/test_rollout.py)": (
            "[test_multi_turn_vlm.py](tests/rollouts/test_multi_turn_vlm.py)"
        ),
        "[fake_env.py](tests/fake_env.py)": (
            "[fake_env.py](tests/fixtures/fake_env.py)"
        ),
    }
    for old, new in replacements.items():
        _replace(readme, old, new)
    _replace(
        dir_root / "autonomy/tests/run_tests.sh",
        "autonomy/tests/test_rollout.py",
        "autonomy/tests/rollouts/test_multi_turn_vlm.py",
    )
    git_add_and_commit("[tests, rollout] refactor: mirror multi-turn rollout tests", cwd=str(dir_root))


def _agent_trainer_tests(dir_root: Path) -> None:
    _move(
        dir_root,
        "autonomy/tests/test_strands_rollout.py",
        "autonomy/tests/rollouts/test_agent_rollout.py",
    )
    trainers_dir = dir_root / "autonomy/tests/trainers"
    (trainers_dir / "__init__.py").write_text("\n")
    for test_name in (
        "test_distill_only.py",
        "test_reloadable_process_group.py",
        "test_train_dispose_try_finally.py",
    ):
        _move(
            dir_root,
            f"autonomy/tests/{test_name}",
            f"autonomy/tests/trainers/{test_name}",
        )

    _replace(trainers_dir / "test_distill_only.py", "parents[2]", "parents[3]")
    _replace(
        trainers_dir / "test_train_dispose_try_finally.py",
        "autonomy.tests.test_distill_only",
        "autonomy.tests.trainers.test_distill_only",
    )
    _replace(
        trainers_dir / "test_train_dispose_try_finally.py",
        "parents[2]",
        "parents[3]",
    )

    readme = dir_root / "autonomy/README_CLAUDE.md"
    _replace(
        readme,
        "│   ├── test_strands_rollout.py  # Strands rollout integration tests with typed interface",
        "│   ├── rollouts/test_agent_rollout.py  # Agent rollout integration tests with typed interface",
    )
    _replace(
        readme,
        "[test_strands_rollout.py](tests/test_strands_rollout.py) - Strands rollout integration",
        "[test_agent_rollout.py](tests/rollouts/test_agent_rollout.py) - Agent rollout integration",
    )
    git_add_and_commit(
        "[tests, rollout, trainer] refactor: mirror agent and trainer tests",
        cwd=str(dir_root),
    )


def _environment_tests(dir_root: Path) -> None:
    moves = {
        "autonomy/tests/test_podman_provider.py": "autonomy/tests/env_providers/test_podman_provider.py",
        "autonomy/tests/math_verifier_report.txt": "autonomy/tests/envs/math_verifier_report.txt",
        "autonomy/tests/math_verifier_test_data.json": "autonomy/tests/envs/math_verifier_test_data.json",
        "autonomy/tests/test_pydantic_typing.py": "autonomy/tests/envs/test_coding_metadata_validation.py",
        "autonomy/tests/test_harbor_env.py": "autonomy/tests/envs/test_harbor_env.py",
        "autonomy/tests/test_math_verifier.py": "autonomy/tests/envs/test_math_env.py",
        "autonomy/tests/test_nova_act_env.py": "autonomy/tests/envs/test_nova_act_env_pool.py",
        "autonomy/tests/test_oss_nova_act_env.py": "autonomy/tests/envs/test_oss_nova_act_env.py",
        "autonomy/tests/test_parsing_assertions.py": "autonomy/tests/envs/test_oss_nova_act_parsing.py",
        "autonomy/tests/test_verifier_env_dispatch.py": "autonomy/tests/envs/test_verifier_env_dispatch.py",
        "autonomy/tests/test_verifier_utils.py": "autonomy/tests/envs/test_verifier_utils.py",
        "autonomy/tests/local_provider.py": "autonomy/tests/fixtures/local_provider.py",
    }
    for source_name, target_name in moves.items():
        _move(dir_root, source_name, target_name)

    local_provider_import = "from autonomy.tests.local_provider import LocalProvider"
    fixture_provider_import = "from autonomy.tests.fixtures.local_provider import LocalProvider"
    for relative_path in (
        "autonomy/README_CLAUDE.md",
        "autonomy/tests/env_providers/test_service_endpoints.py",
        "autonomy/tests/envs/test_coding_metadata_validation.py",
        "autonomy/tests/envs/test_verifier_utils.py",
        "autonomy/tests/fixtures/local_provider.py",
        "autonomy/tests/utils/test_code_verifier.py",
        "autonomy/tests/utils/test_code_verifier_sample.py",
    ):
        _replace(dir_root / relative_path, local_provider_import, fixture_provider_import)

    for relative_path in (
        "autonomy/tests/envs/test_coding_metadata_validation.py",
        "autonomy/tests/envs/test_verifier_utils.py",
    ):
        _replace(
            dir_root / relative_path,
            (
                "from autonomy.tests.fixtures.verify import verify_via_env\n"
                "from autonomy.tests.fixtures.local_provider import LocalProvider"
            ),
            (
                "from autonomy.tests.fixtures.local_provider import LocalProvider\n"
                "from autonomy.tests.fixtures.verify import verify_via_env"
            ),
        )

    for relative_path in (
        "autonomy/tests/envs/test_coding_metadata_validation.py",
        "autonomy/tests/envs/test_math_env.py",
        "autonomy/tests/envs/test_verifier_utils.py",
    ):
        _replace(
            dir_root / relative_path,
            "slime_root = Path(__file__).parent.parent.parent.parent",
            "slime_root = Path(__file__).parent.parent.parent.parent.parent",
        )

    _replace(
        dir_root / "autonomy/tests/envs/test_harbor_env.py",
        "autonomy/tests/test_harbor_env.py",
        "autonomy/tests/envs/test_harbor_env.py",
    )
    _replace(
        dir_root / "autonomy/tests/fixtures/local_provider.py",
        "intentionally placed in autonomy/tests/ to prevent",
        "intentionally placed in autonomy/tests/fixtures/ to prevent",
    )
    _replace(
        dir_root / "autonomy/tests/run_tests.sh",
        "autonomy/tests/test_verifier_utils.py",
        "autonomy/tests/envs/test_verifier_utils.py",
    )
    _replace(
        dir_root / "autonomy/tests/envs/test_math_env.py",
        'report_path = Path("./math_verifier_report.txt")',
        'report_path = Path(__file__).parent / "math_verifier_report.txt"',
    )

    readme = dir_root / "autonomy/README_CLAUDE.md"
    replacements = {
        "│   ├── local_provider.py          #": "│   ├── fixtures/local_provider.py #",
        "│   ├── test_math_verifier.py": "│   ├── envs/test_math_env.py",
        "│   ├── test_verifier_utils.py        #": "│   ├── envs/test_verifier_utils.py   #",
        "[test_math_verifier.py](tests/test_math_verifier.py)": (
            "[test_math_env.py](tests/envs/test_math_env.py)"
        ),
        "`AutoGrader_API_Key` for live tests": "`OPENROUTER_API_KEY` for live tests",
        "[test_code_verifier.py](tests/test_code_verifier.py)": (
            "[test_code_verifier.py](tests/utils/test_code_verifier.py)"
        ),
        "[code_verifier_test_data.json](tests/code_verifier_test_data.json)": (
            "[code_verifier_test_data.json](tests/utils/code_verifier_test_data.json)"
        ),
        "[test_code_verifier_sample.py](tests/test_code_verifier_sample.py)": (
            "[test_code_verifier_sample.py](tests/utils/test_code_verifier_sample.py)"
        ),
        "[test_verifier_utils.py](tests/test_verifier_utils.py)": (
            "[test_verifier_utils.py](tests/envs/test_verifier_utils.py)"
        ),
        "[tests/local_provider.py](tests/local_provider.py)": (
            "[tests/fixtures/local_provider.py](tests/fixtures/local_provider.py)"
        ),
        "placed in `autonomy/tests/` to prevent": (
            "placed in `autonomy/tests/fixtures/` to prevent"
        ),
    }
    for old, new in replacements.items():
        _replace(readme, old, new)
    git_add_and_commit("[tests, env] refactor: mirror environment tests", cwd=str(dir_root))


def _core_tests(dir_root: Path) -> None:
    moves = {
        "autonomy/tests/test_circular_imports.py": "autonomy/tests/utils/test_circular_imports.py",
        "autonomy/tests/test_convert_to_sft.py": "autonomy/tests/data/rl/distill/test_convert_to_sft.py",
        "autonomy/tests/test_data_config.py": "autonomy/tests/dataloaders/test_data_config.py",
        "autonomy/tests/test_data_mixture_source.py": "autonomy/tests/dataloaders/test_data_mixture_source.py",
        "autonomy/tests/test_dedupe_embeddings.py": "autonomy/tests/data/rl/dedupe/test_dedupe_embeddings.py",
        "autonomy/tests/test_filter_distilled.py": "autonomy/tests/data/rl/distill/test_filter_distilled.py",
        "autonomy/tests/test_image_utils.py": "autonomy/tests/utils/test_image_utils.py",
        "autonomy/tests/test_import_linter_sync.py": "autonomy/tests/scripts/test_import_linter_sync.py",
        "autonomy/tests/test_model_registration.py": "autonomy/tests/models/test_model_registration.py",
        "autonomy/tests/test_oss_adapter.py": "autonomy/tests/models/oss/test_adapter.py",
        "autonomy/tests/test_reward.py": "autonomy/tests/utils/test_reward_postprocessing.py",
        "autonomy/tests/test_s3_utils.py": "autonomy/tests/utils/test_s3_utils.py",
        "autonomy/tests/test_standalone_imports.py": "autonomy/tests/utils/coding_utils/test_standalone_imports.py",
        "autonomy/tests/test_types.py": "autonomy/tests/utils/test_slime_args_model.py",
        "autonomy/tests/test_wandb_pytest_guard.py": "autonomy/tests/loggings/test_wandb_pytest_guard.py",
    }
    for source_name, target_name in moves.items():
        _move(dir_root, source_name, target_name)

    _replace(
        dir_root / "autonomy/tests/data/rl/dedupe/test_dedupe_embeddings.py",
        "slime_root = Path(__file__).parent.parent.parent.parent",
        "slime_root = Path(__file__).parent.parent.parent.parent.parent.parent.parent",
    )
    for name in ("PYPROJECT", "PACKAGES_ROOT"):
        _replace(
            dir_root / "autonomy/tests/scripts/test_import_linter_sync.py",
            f'{name} = Path(__file__).resolve().parent.parent /',
            f'{name} = Path(__file__).resolve().parent.parent.parent /',
        )
    _replace(
        dir_root / "autonomy/tests/utils/coding_utils/test_standalone_imports.py",
        'Path(__file__).parent.parent / "utils" / "coding_utils" / "standalone"',
        'Path(__file__).resolve().parents[3] / "utils" / "coding_utils" / "standalone"',
    )
    _replace(
        dir_root / "autonomy/pyproject.toml",
        "autonomy/tests/test_import_linter_sync.py",
        "autonomy/tests/scripts/test_import_linter_sync.py",
    )
    for relative_path in ("autonomy/README_CLAUDE.md", "autonomy/tests/run_tests.sh"):
        _replace(
            dir_root / relative_path,
            "autonomy/tests/test_reward.py",
            "autonomy/tests/utils/test_reward_postprocessing.py",
        )

    readme = dir_root / "autonomy/README_CLAUDE.md"
    replacements = {
        "│   ├── test_dedupe_embeddings.py     #": (
            "│   ├── data/rl/dedupe/test_dedupe_embeddings.py #"
        ),
        "│   ├── test_types.py            #": "│   ├── utils/test_slime_args_model.py #",
        "│   ├── test_reward.py": "│   ├── utils/test_reward_postprocessing.py",
        "[test_reward.py](tests/test_reward.py)": (
            "[test_reward_postprocessing.py](tests/utils/test_reward_postprocessing.py)"
        ),
        "[tests/test_filter_distilled.py](tests/test_filter_distilled.py)": (
            "[test_filter_distilled.py](tests/data/rl/distill/test_filter_distilled.py)"
        ),
        "[tests/test_verify_distilled.py](tests/test_verify_distilled.py)": (
            "[test_verify_distilled.py](tests/data/rl/distill/test_verify_distilled.py)"
        ),
        "[tests/test_convert_to_sft.py](tests/test_convert_to_sft.py)": (
            "[test_convert_to_sft.py](tests/data/rl/distill/test_convert_to_sft.py)"
        ),
        "[test_dedupe_embeddings.py](tests/test_dedupe_embeddings.py)": (
            "[test_dedupe_embeddings.py](tests/data/rl/dedupe/test_dedupe_embeddings.py)"
        ),
        "[test_filter_distilled.py](tests/test_filter_distilled.py)": (
            "[test_filter_distilled.py](tests/data/rl/distill/test_filter_distilled.py)"
        ),
        "[test_verify_distilled.py](tests/test_verify_distilled.py)": (
            "[test_verify_distilled.py](tests/data/rl/distill/test_verify_distilled.py)"
        ),
        "[test_types.py](tests/test_types.py)": (
            "[test_slime_args_model.py](tests/utils/test_slime_args_model.py)"
        ),
    }
    for old, new in replacements.items():
        _replace(readme, old, new)
    for stale_line in (
        "- [test_oss_trajectory.py](tests/test_oss_trajectory.py) - OSSTrajectory message parsing tests:\n",
        "  - Nova Act SDK observation parsing (type 1=text, type 2=image)\n",
        "  - Context-aware extraction (first observation vs subsequent observations)\n",
        "  - Robust marker detection with quoted text handling\n",
    ):
        _replace(readme, stale_line, "")
    git_add_and_commit("[tests] refactor: mirror core test layout", cwd=str(dir_root))


def _launcher_ci_tests(dir_root: Path) -> None:
    _move(
        dir_root,
        "autonomy/tests/test_launcher.py",
        "autonomy/tests/launcher/test_launcher.py",
    )
    _move(
        dir_root,
        "autonomy/tests/test_autonomy_image_tests_workflow.py",
        "autonomy/tests/scripts/ci/test_autonomy_image_tests_workflow.py",
    )
    _replace(
        dir_root / "autonomy/tests/scripts/ci/test_autonomy_image_tests_workflow.py",
        "REPO_ROOT = Path(__file__).resolve().parents[2]",
        "REPO_ROOT = Path(__file__).resolve().parents[4]",
    )
    _replace(
        dir_root / "autonomy/README_CLAUDE.md",
        "│   ├── test_launcher.py   #",
        "│   ├── launcher/test_launcher.py #",
    )
    _replace(
        dir_root / "autonomy/README_CLAUDE.md",
        "[test_launcher.py](tests/test_launcher.py)",
        "[test_launcher.py](tests/launcher/test_launcher.py)",
    )
    _replace(
        dir_root / ".claude/skills/gpu-ci/SKILL.md",
        "autonomy/tests/test_autonomy_image_tests_workflow.py",
        "autonomy/tests/scripts/ci/test_autonomy_image_tests_workflow.py",
    )
    git_add_and_commit("[tests, ci] refactor: mirror launcher and workflow tests", cwd=str(dir_root))


def _package_support_tests(dir_root: Path) -> None:
    _move(
        dir_root,
        "autonomy/tests/test_container_utils.py",
        "autonomy/tests/utils/test_container_utils.py",
    )
    _move(
        dir_root,
        "autonomy/tests/test_package_dist_names.py",
        "autonomy/tests/packages/test_dist_names.py",
    )
    (dir_root / "autonomy/tests/packages/__init__.py").write_text("\n")

    _replace(
        dir_root / "autonomy/tests/packages/test_dist_names.py",
        'PACKAGES_ROOT = Path(__file__).resolve().parent.parent / "packages"',
        'PACKAGES_ROOT = Path(__file__).resolve().parent.parent.parent / "packages"',
    )
    for relative_path in (
        ".claude/skills/add-subpackage/SKILL.md",
        "autonomy/packages/public/README.md",
        "autonomy/packages/public/circuit_client/README.md",
    ):
        _replace(
            dir_root / relative_path,
            "autonomy/tests/test_package_dist_names.py",
            "autonomy/tests/packages/test_dist_names.py",
        )
    _replace(
        dir_root / "autonomy/packages/internal/gym_proxy/tests/test_proxy_integration.py",
        "autonomy/tests/\ntest_container_utils.py",
        "autonomy/tests/utils/\ntest_container_utils.py",
    )
    _replace(
        dir_root / ".github/CODEOWNERS",
        "/autonomy/packages/ @GymPod/SlimeCodeowners @GymPod/NextAct\n",
        (
            "/autonomy/packages/ @GymPod/SlimeCodeowners @GymPod/NextAct\n"
            "/autonomy/tests/packages/ @GymPod/SlimeCodeowners @GymPod/NextAct\n"
        ),
    )
    _replace(
        dir_root / "autonomy/packages/internal/circuit_internal/ARCHITECTURE.md",
        (
            "`autonomy/tests/test_package_isolation.py` enforces the autonomy boundary\n"
            "mechanically. The `slime.*` / `strands.*` boundary is enforced by code\n"
            "review and grep. If you find yourself wanting to import either inside this\n"
            "package, the work belongs in the wrapper instead — see\n"
        ),
        (
            "The import-linter contracts in `autonomy/pyproject.toml` enforce the autonomy\n"
            "boundary mechanically, with\n"
            "`autonomy/tests/scripts/test_import_linter_sync.py` keeping the package list in\n"
            "sync. The `slime.*` / `strands.*` boundary is enforced by code review and grep.\n"
            "If you find yourself wanting to import either inside this package, the work\n"
            "belongs in the wrapper instead — see\n"
        ),
    )
    git_add_and_commit("[tests, packages] refactor: mirror package support tests", cwd=str(dir_root))


def _sft_tool_use_test(dir_root: Path) -> None:
    _move(
        dir_root,
        "autonomy/tests/test_sft_tooluse.py",
        "autonomy/tests/data/sft/tooluse/test_build_sft_tooluse.py",
    )
    readme = dir_root / "autonomy/README_CLAUDE.md"
    _replace(
        readme,
        "tests/test_sft_tooluse.py",
        "tests/data/sft/tooluse/test_build_sft_tooluse.py",
    )
    _replace(readme, "[test_sft_tooluse.py]", "[test_build_sft_tooluse.py]")
    git_add_and_commit("[tests, data] refactor: mirror SFT tool-use test", cwd=str(dir_root))


def _review_followup(dir_root: Path) -> None:
    _replace(
        dir_root / ".claude/skills/add-subpackage/SKILL.md",
        "test_package_dist_names.py enforces",
        "test_dist_names.py enforces",
    )
    git_add_and_commit(
        "[tests, packages] docs: update mirrored test reference",
        cwd=str(dir_root),
    )


def _package_marker_followup(dir_root: Path) -> None:
    for relative_path in (
        "autonomy/tests/data/sft/__init__.py",
        "autonomy/tests/data/sft/tooluse/__init__.py",
        "autonomy/tests/models/oss/__init__.py",
    ):
        (dir_root / relative_path).write_text("\n")
    git_add_and_commit(
        "[tests] chore: add mirrored test package markers",
        cwd=str(dir_root),
    )


def transform(dir_root: Path) -> None:
    """Apply the nine atomic F-X1-1 slices and review follow-ups."""
    _logging_tests(dir_root)
    _async_rollout_tests(dir_root)
    _multi_turn_tests(dir_root)
    _agent_trainer_tests(dir_root)
    _environment_tests(dir_root)
    _core_tests(dir_root)
    _launcher_ci_tests(dir_root)
    _package_support_tests(dir_root)
    _sft_tool_use_test(dir_root)
    _review_followup(dir_root)
    _package_marker_followup(dir_root)


if __name__ == "__main__":
    verify_mechanical_refactor(
        base_commit=BASE_COMMIT,
        target_commit=TARGET_COMMIT,
        transform=transform,
    )