{ "agent_trace": { "codex_logs_available_local_only": true, "raw_logs_published": false }, "collected_at": "2026-05-18T13:25:15.144583+00:00", "docker_cpus": 16, "docker_memory": "30g", "eval": { "error_code": null, "failed_tests": [ "eval.tests.test_wrapcheck.test_config_ignore_package_globs", "eval.tests.test_wrapcheck.test_config_invalid_yaml", "eval.tests.test_wrapcheck.test_json_output_format", "eval.tests.test_wrapcheck.test_analyze_package_with_issues", "eval.tests.test_wrapcheck.test_detect_unwrapped_json_error", "eval.tests.test_wrapcheck.test_config_ignore_sig_regexps", "eval.tests.test_wrapcheck.test_config_ignore_sigs", "eval.tests.test_wrapcheck.test_version_flag_full", "eval.tests.test_wrapcheck.test_debug_flag", "eval.tests.test_wrapcheck.test_deprecated_all_flag", "eval.tests.test_wrapcheck.test_cpuprofile_flag", "eval.tests.test_wrapcheck.test_memprofile_flag", "eval.tests.test_wrapcheck.test_deprecated_v_flag", "eval.tests.test_wrapcheck.test_deprecated_tags_flag", "eval.tests.test_wrapcheck.test_context_lines_negative", "eval.tests.test_wrapcheck.test_deprecated_source_flag", "eval.tests.test_wrapcheck.test_config_extra_ignore_sigs", "eval.tests.test_wrapcheck.test_context_lines_output", "eval.tests.test_wrapcheck.test_default_output_format", "eval.tests.test_wrapcheck.test_config_ignore_sig_regexps_fail", "eval.tests.test_wrapcheck.test_trace_flag", "eval.tests.test_wrapcheck.test_syntax_error_in_go_file", "eval.tests.test_wrapcheck.test_recursive_package_scan", "eval.tests.test_wrapcheck.test_multiple_packages", "eval.tests.test_wrapcheck.test_context_lines_zero", "eval.tests.test_wrapcheck.test_context_lines_large", "eval.tests.test_wrapcheck.test_nested_function_calls", "eval.tests.test_wrapcheck.test_multiple_errors_in_file", "eval.tests.test_wrapcheck.test_analyze_current_directory", "tests.test_analysis_core.test_interface_method_requires_wrapping", "tests.test_analysis_core.test_anonymous_function_in_func_decl_still_checked", "tests.test_analysis_core.test_struct_method_call_detects_internal_external_call", "tests.test_analysis_core.test_chained_method_calls_unwrapped", "tests.test_analysis_core.test_error_from_type_assertion_unwrapped", "tests.test_analysis_core.test_goroutine_anonymous_func_checked", "tests.test_cli_flags.test_version_flag_full", "tests.test_cli_flags.test_context_lines_large", "tests.test_cli_flags.test_cpuprofile_creates_file", "tests.test_cli_flags.test_memprofile_creates_file", "tests.test_cli_flags.test_trace_creates_file", "tests.test_cli_flags.test_multiple_profile_flags", "tests.test_cli_flags.test_cpuprofile_with_invalid_path", "tests.test_cli_flags.test_json_empty_package", "tests.test_cli_flags.test_trace_flag_with_json", "tests.test_cli_flags.test_debug_flag_output", "tests.test_cli_flags.test_version_output_format", "tests.test_cli_flags.test_multiple_profiling_files_json_mode", "tests.test_config.test_ignoresigregexps_pattern_matching", "tests.test_config.test_ignoresigregexps_invalid_regex_errors", "tests.test_config.test_reportinternalerrors_false_only_external", "tests.test_config.test_reportinternalerrors_true_reports_internal", "tests.test_config.test_ignoreinterfaceregexps_matches_interface_name", "tests.test_config.test_ignoreinterfaceregexps_multiple_patterns", "tests.test_config.test_config_invalid_yaml_syntax_error", "tests.test_config.test_config_multiple_regex_patterns_all_checked", "tests.test_config.test_ignorepackageglobs_exact_match_not_prefix", "tests.test_config.test_reportinternalerrors_default_false", "tests.test_config.test_ignoresigregexps_complex_pattern", "tests.test_config.test_ignoreinterfaceregexps_invalid_regex_error", "tests.test_config.test_ignorepackageglobs_invalid_glob_error", "tests.test_edge_cases.test_syntax_error_in_package", "tests.test_edge_cases.test_type_error_in_package", "tests.test_edge_cases.test_missing_import_undefined_identifier", "tests.test_edge_cases.test_empty_package_no_functions", "tests.test_edge_cases.test_package_without_gomod", "tests.test_edge_cases.test_package_with_no_error_returning_functions", "tests.test_edge_cases.test_large_file_many_functions", "tests.test_edge_cases.test_deeply_nested_anonymous_functions", "tests.test_edge_cases.test_multiple_return_statements_in_function", "tests.test_edge_cases.test_named_return_parameters", "tests.test_edge_cases.test_unicode_filenames", "tests.test_edge_cases.test_package_pattern_recursive", "tests.test_edge_cases.test_variadic_function", "tests.test_edge_cases.test_json_output_format", "tests.test_edge_cases.test_invalid_yaml_config_parse_error", "tests.test_edge_cases.test_invalid_regex_in_config", "tests.test_edge_cases.test_relative_path_current_directory", "tests.test_edge_cases.test_switch_statement_multiple_cases", "tests.test_edge_cases.test_multiple_package_arguments", "tests.test_edge_cases.test_errors_new_with_empty_message" ], "failed_tests_omitted": 249, "public_eval_path": "eval.json", "summary_path": "eval-summary.json", "test_branch_errors": {}, "test_records": 669, "warnings": [] }, "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "instance_id": "tomarrell__wrapcheck.c058da1", "metrics": { "almost_resolved": "False", "cached_input_tokens": "3705216", "calls": "85", "created_at": "2026-05-17T16:10:26.225992+00:00", "docker_cpus": "16", "docker_memory": "30g", "error_code": "", "estimated_cost_usd": "4.1974", "evaluator_problem": "False", "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "input_tokens": "3899552", "instance_id": "tomarrell__wrapcheck.c058da1", "long_context_calls": "0", "model": "gpt-5.5", "n_resolved_tests": "202", "n_system_errors": "0", "n_tests": "480", "n_warnings": "0", "output_tokens": "45769", "paper_compliant": "False", "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5", "reasoning_effort": "xhigh", "reasoning_output_tokens": "15315", "resolved": "False", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-tomarrell-wrapcheck", "run_version": "20260517T094700Z", "score": "0.42083333333333334", "session_ended_at": "2026-05-17T16:23:46.648Z", "session_started_at": "2026-05-17T16:10:29.504Z", "test_branch_errors": "{}", "total_tokens": "3945321", "wall_clock_seconds": "797" }, "model": "gpt-5.5", "package": { "contents": [ ".", "./compile.sh", "./executable", "./wrapcheck.py" ], "contents_omitted": 0, "submission_available_local_only": true }, "paper_compliant": false, "paper_mode": false, "reasoning_effort": "xhigh", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-tomarrell-wrapcheck", "run_version": "20260517T094700Z", "usage_audit": { "available_local_artifact": true, "public_path": "usage-audit.json" } }