GEWU/tests/domain/test_run_checkpoint.py

import copy
import hashlib
import json
import stat
from pathlib import Path

import agents.gewu.scripts.run_checkpoint as checkpoint_module
import pytest
from agents.gewu.scripts.run_checkpoint import (
    create_checkpoint,
    load_checkpoint,
    update_checkpoint,
)


def test_run_checkpoint_records_scope_coverage_receipt_and_pins(tmp_path: Path) -> None:
    workspace = tmp_path / "research-workspace"
    checkpoint = create_checkpoint(
        workspace,
        run_id="run-2026-01",
        objective="Compare audience response",
        projects=["dome-performance"],
        include_shared=True,
        lens="Audience behavior",
        budgets={"sources": 8},
    )
    path = workspace / "Runs" / "run-2026-01.json"
    assert stat.S_IMODE(path.stat().st_mode) == 0o600
    assert checkpoint["pack"]["snapshot_digest"]

    updated = update_checkpoint(
        workspace,
        "run-2026-01",
        {
            "stage": "report",
            "sources": [
                {
                    "source_revision_id": "source:r1",
                    "extraction_id": "a" * 64,
                    "status": "extracted",
                }
            ],
            "coverage": {
                "selected": ["section-1", "section-2"],
                "included": ["section-1"],
                "excluded": [{"item": "section-2", "reason": "outside project lens"}],
                "skipped": [],
            },
            "pending_operations": [
                {
                    "operation_id": "proposal-op-1",
                    "tool_name": "proposal_submit",
                    "input_sha256": "b" * 64,
                    "status": "complete",
                }
            ],
            "proposals": [
                {
                    "proposal_id": "proposal-1",
                    "revision": 1,
                    "status": "accepted",
                    "receipt": {
                        "proposal_id": "proposal-1",
                        "revision": 1,
                        "state": "accepted",
                        "commit_id": "c" * 40,
                    },
                }
            ],
            "outputs": [{"sha256": "d" * 64}],
        },
    )
    resumed = load_checkpoint(workspace, "run-2026-01")
    assert resumed == updated
    assert resumed["coverage"]["excluded"][0]["reason"] == "outside project lens"
    assert resumed["proposals"][0]["receipt"]["commit_id"] == "c" * 40
    with pytest.raises(ValueError, match="does not match"):
        update_checkpoint(
            workspace,
            "run-2026-01",
            {
                "proposals": [
                    {
                        "proposal_id": "proposal-1",
                        "revision": 1,
                        "status": "accepted",
                        "receipt": {
                            "proposal_id": "proposal-other",
                            "revision": 1,
                            "state": "accepted",
                            "commit_id": "c" * 40,
                        },
                    }
                ]
            },
        )


def test_run_checkpoint_rejects_secrets_bad_progress_and_snapshot_drift(tmp_path: Path) -> None:
    workspace = tmp_path / "research-workspace"
    create_checkpoint(
        workspace,
        run_id="run-drift-01",
        objective="Test resume pinning",
        projects=["dome-performance"],
        include_shared=False,
        lens="A bounded lens",
    )
    with pytest.raises(ValueError, match="credentials or source bodies"):
        update_checkpoint(workspace, "run-drift-01", {"last_failure": {"api_key": "secret"}})
    with pytest.raises(ValueError, match="SHA-256"):
        update_checkpoint(
            workspace,
            "run-drift-01",
            {
                "outputs": [{"sha256": "not-a-digest"}],
                "stage": "report",
            },
        )

    path = workspace / "Runs" / "run-drift-01.json"
    checkpoint = json.loads(path.read_text(encoding="utf-8"))
    checkpoint["pack"]["snapshot_digest"] = "0" * 64
    path.write_text(json.dumps(checkpoint), encoding="utf-8")
    with pytest.raises(ValueError, match="pack changed"):
        load_checkpoint(workspace, "run-drift-01")


def test_existing_run_cannot_resume_after_verified_pack_upgrade(
    tmp_path: Path,
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    workspace = tmp_path / "research-workspace"
    create_checkpoint(
        workspace,
        run_id="run-pack-update-01",
        objective="Resume with pinned modules",
        projects=["dome-performance"],
        include_shared=True,
        lens="Keep the original extraction and evidence rules",
    )
    checkpoint_path = workspace / "Runs" / "run-pack-update-01.json"
    original_checkpoint = checkpoint_path.read_bytes()

    updated_manifest = copy.deepcopy(checkpoint_module.load_manifest())
    updated_manifest["pack_version"] = "0.7.0"
    updated_manifest["modules"][0]["implementation_version"] += ".updated"
    modules_json = json.dumps(
        updated_manifest["modules"], ensure_ascii=False, sort_keys=True, separators=(",", ":")
    )
    updated_manifest["snapshot"]["digest"] = hashlib.sha256(modules_json.encode()).hexdigest()

    # This fixture stands in for a newer pack whose own manifest verification
    # succeeded; the behavior under test is the active run's compatibility pin.
    monkeypatch.setattr(checkpoint_module, "load_manifest", lambda: updated_manifest)
    monkeypatch.setattr(checkpoint_module, "verify", lambda _manifest: None)

    with pytest.raises(ValueError, match="pack changed since this run"):
        load_checkpoint(workspace, "run-pack-update-01")
    assert checkpoint_path.read_bytes() == original_checkpoint