from __future__ import annotations from src.services.agent_nemotron_replay_preflight import ( evaluate_nemotron_external_runner_preflight, ) from src.services.agent_nemotron_replay_sanitizer import ( contains_sensitive_context_marker, sanitize_nemotron_request_pack_from_fixtures, ) def test_sanitizer_removes_sensitive_context_markers_and_preflight_passes(): sanitized_fixtures, candidate_inputs, requests, report = ( sanitize_nemotron_request_pack_from_fixtures([_fixture_with_sensitive_context()]) ) assert report.valid is True assert report.sensitive_marker_records_before == 1 assert report.sensitive_marker_records_after == 0 assert report.changed_fixture_records == 1 assert not contains_sensitive_context_marker(sanitized_fixtures[0]["incident_context"]) assert not contains_sensitive_context_marker(candidate_inputs[0]["incident_context"]) assert not contains_sensitive_context_marker(requests[0]["incident_context"]) preflight = evaluate_nemotron_external_runner_preflight( fixtures=sanitized_fixtures, candidate_inputs=candidate_inputs, requests=requests, ).to_dict() assert preflight["valid"] is True assert preflight["sensitive_marker_records"] == 0 def test_sanitizer_preserves_evaluation_labels_for_local_grading(): sanitized_fixtures, _, _, _ = sanitize_nemotron_request_pack_from_fixtures( [_fixture_with_sensitive_context()] ) assert sanitized_fixtures[0]["evaluation_labels"]["verification_result"] == "success" assert sanitized_fixtures[0]["evaluation_labels"]["expected_action_markers"] == [ "rollout restart", "checkout", ] def _fixture_with_sensitive_context() -> dict: return { "schema_version": "agent_replay_fixture_v1", "run_id": "run", "incident_id": "INC-1", "incident_context": { "alertname": "DockerContainerUnhealthy", "severity": "P2", "affected_services": ["checkout"], "evidence_summary": ( "/srv/app/.secrets/admin.htpasswd=***REDACTED*** " "PGPASSFILE=\"$pgpass\" pg_dump --no-password" ), "metadata": { "secret_path": "/k8s/08-google-drive-secret.yaml", }, }, "evaluation_labels": { "verification_result": "success", "execution_success": True, "expected_action_markers": ["rollout restart", "checkout"], }, "source_metadata": {"source": "test"}, }