{ "agent_trace": { "codex_logs_available_local_only": true, "raw_logs_published": false }, "collected_at": "2026-05-18T13:25:01.092123+00:00", "docker_cpus": 16, "docker_memory": "30g", "eval": { "error_code": null, "failed_tests": [ "eval.tests.test_basic.TestInputHandling.test_empty_file_path", "eval.tests.test_basic.TestOutputGeneration.test_output_same_as_input_error", "eval.tests.test_error_handling.test_output_file_same_as_input", "eval.tests.test_error_handling.test_empty_wsdl_file", "eval.tests.test_error_handling.test_treat_last_arg_as_wsdl", "eval.tests.test_error_handling.test_invalid_xml_but_looks_like_wsdl", "tests.test_cli.test_invalid_xml_eof_error", "tests.test_cli.test_directory_input_error", "tests.test_cli.test_make_public_flag_false", "tests.test_errors.test_invalid_output_directory", "tests.test_errors.test_output_file_same_as_input", "tests.test_errors.test_empty_file", "tests.test_errors.test_empty_output_filename_creates_directory_error", "tests.test_errors.test_package_name_starts_with_digit", "tests.test_gap_wsdl_xsd.test_xsd_complextype_invalid_boolean_mixed", "tests.test_gap_wsdl_xsd.test_xsd_element_invalid_boolean_nillable", "tests.test_errors.test_no_arguments", "tests.test_harvest.test_unmarshal_vim_wsdl", "tests.test_errors.test_http_500_error", "tests.test_location.test_relative_import_resolution", "tests.test_errors.test_https_with_insecure_flag_bypasses_cert_check", "tests.test_cli.test_output_file_equals_input_error", "tests.test_cli.test_malformed_xml_syntax_error", "tests.test_gap_wsdl_xsd.test_xsd_element_multiple_invalid_booleans", "tests.test_errors.test_unsupported_url_scheme", "tests.test_errors.test_invalid_xml_content", "tests.test_gap_wsdl_xsd.test_wsdl_types_nested_schema_decode_error", "tests.test_errors.test_file_path_exceeds_system_limit", "tests.test_gap_wsdl_xsd.test_xsd_complextype_invalid_boolean_abstract", "tests.test_errors.test_invalid_url_dns_failure", "tests.test_harvest.test_enumerations_generated_correctly", "tests.test_location.test_http_url_404_error", "tests.test_harvest.test_output_file_same_as_input", "tests.test_location.test_nested_parent_directory_reference", "tests.test_errors.test_https_certificate_verification_failure", "tests.test_location.test_path_with_spaces", "tests.test_location.test_symlink_resolution", "tests.test_gap_wsdl_xsd.test_xsd_complextype_both_abstract_and_mixed_invalid", "tests.test_errors.test_invalid_package_name_with_spaces", "tests.test_errors.test_malformed_xml", "tests.test_location.test_epcis_multi_file_import_chain", "tests.test_location.test_current_directory_dot_slash_references", "tests.test_location.test_url_encoded_characters_in_http_path", "tests.test_location.test_file_wsdl_with_http_import", "tests.test_errors.test_http_404_error", "tests.test_location.test_special_characters_in_path", "tests.test_errors.test_http_connection_refused", "tests.test_location.test_https_import_in_wsdl", "tests.test_location.test_absolute_path_import", "tests.test_location.test_http_url_as_main_input", "tests.test_location.test_deeply_nested_parent_directory_traversal", "tests.test_location.test_malformed_url_treated_as_file", "tests.test_location.test_multiple_imports_same_wsdl", "tests.test_location.test_https_url_as_main_input", "tests.test_operations.test_operation_interface_with_multiple_methods", "tests.test_location.test_deeply_nested_relative_path", "tests.test_operations.test_operations_with_faults_generate_error_documentation", "tests.test_operations.test_operation_documentation_preserved", "tests.test_operations.test_empty_string_type_for_missing_io", "tests.test_type_generation.test_basic_xsd_types_generation", "tests.test_operations.test_operation_with_no_input", "tests.test_type_generation.test_complex_content_extension", "tests.test_type_generation.test_reserved_go_keywords_escaping", "tests.test_type_generation.test_unicode_names_in_types", "tests.test_operations.test_operation_with_no_output", "tests.test_type_generation.test_invalid_wsdl_file_error_handling", "tests.test_type_generation.test_enumeration_restriction_generation", "tests.test_type_generation.test_list_type_generation", "tests.test_errors.test_invalid_port_number_in_url", "eval.tests.test_argparse_validation.test_bool_flag_invalid_value_is_error", "eval.tests.test_argparse_validation.test_double_dash_forces_following_token_to_be_positional_not_flag", "eval.tests.test_argparse_validation.test_positional_first_then_flag_treated_as_filename_and_causes_open_error", "eval.tests.test_help_output.test_help_has_usage_line", "eval.tests.test_help_output.test_help_indentation_uses_tabbed_description_lines", "eval.tests.test_help_output.test_help_output_matches_baseline_exactly", "eval.tests.test_error_cases.test_invalid_wsdl_content_exits_nonzero_and_prints_EOF", "eval.tests.test_gowsdl.TestErrorHandling.test_nonexistent_output_directory", "eval.tests.test_gowsdl.TestCommandLineFlags.test_make_public_false", "eval.tests.test_gowsdl_cli.test_help_long_exact_except_usage_path", "eval.tests.test_gowsdl_cli.test_make_public_flag_changes_generated_code" ], "failed_tests_omitted": 3, "public_eval_path": "eval.json", "summary_path": "eval-summary.json", "test_branch_errors": {}, "test_records": 419, "warnings": [] }, "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "instance_id": "hooklift__gowsdl.2a06cec", "metrics": { "almost_resolved": "False", "cached_input_tokens": "5531136", "calls": "64", "created_at": "2026-05-17T11:32:40.776452+00:00", "docker_cpus": "16", "docker_memory": "30g", "error_code": "", "estimated_cost_usd": "5.0868", "evaluator_problem": "False", "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "input_tokens": "5681223", "instance_id": "hooklift__gowsdl.2a06cec", "long_context_calls": "0", "model": "gpt-5.5", "n_resolved_tests": "312", "n_system_errors": "0", "n_tests": "391", "n_warnings": "0", "output_tokens": "52361", "paper_compliant": "False", "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5", "reasoning_effort": "xhigh", "reasoning_output_tokens": "21631", "resolved": "False", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-hooklift-gowsdl", "run_version": "", "score": "0.7979539641943734", "session_ended_at": "2026-05-17T11:49:40.009Z", "session_started_at": "2026-05-17T11:32:43.713Z", "test_branch_errors": "{}", "total_tokens": "5733584", "wall_clock_seconds": "1016" }, "model": "gpt-5.5", "package": { "contents": [ ".", "./compile.sh", "./executable", "./gowsdl_reimpl.py" ], "contents_omitted": 0, "submission_available_local_only": true }, "paper_compliant": false, "paper_mode": false, "reasoning_effort": "xhigh", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-hooklift-gowsdl", "run_version": "", "usage_audit": { "available_local_artifact": true, "public_path": "usage-audit.json" } }