{ "agent_trace": { "codex_logs_available_local_only": true, "raw_logs_published": false }, "collected_at": "2026-05-18T13:24:55.806443+00:00", "docker_cpus": 16, "docker_memory": "30g", "eval": { "error_code": null, "failed_tests": [ "eval.tests.test_size_filtering.test_depth_zero", "eval.tests.test_error_handling.test_invalid_regex_pattern", "eval.tests.test_error_handling.test_nonexistent_ignore_file", "tests.test_cli.test_full_paths_flag", "tests.test_cli.test_depth_zero", "tests.test_cli.test_conflicting_file_types_depth", "tests.test_cli.test_combined_ignore_hidden_only_file", "tests.test_cli.test_output_format_tib", "tests.test_cli.test_only_directories_flag", "tests.test_cli.test_invalid_stack_size", "tests.test_cli.test_no_percent_bars_flag", "tests.test_cli.test_terminal_width_override", "tests.test_cli.test_screen_reader_mode", "tests.test_cli.test_depth_with_number_of_lines", "tests.test_config_file.test_invalid_config_file_shows_error_continues_with_defaults", "tests.test_cli.test_invalid_threads_value", "tests.test_cli.test_no_colors_flag", "tests.test_coverage_push.test_invalid_regex_invert_filter_exits_with_error", "tests.test_coverage_push.test_invalid_regex_filter_exits_with_error", "tests.test_cli.test_multiple_paths", "tests.test_cli.test_output_format_si", "tests.test_cli.test_reverse_flag", "tests.test_cli.test_only_files_flag", "tests.test_cli.test_combined_apparent_size_si", "tests.test_cli.test_output_format_mebibytes", "tests.test_cli.test_skip_total_flag", "tests.test_cli.test_output_format_kilobytes", "tests.test_cli.test_conflicting_file_types_only_dir", "tests.test_cli.test_conflicting_only_dir_only_file", "tests.test_cli.test_output_format_kib", "tests.test_config_file.test_config_file_precedence_dot_dust_over_xdg", "tests.test_coverage_push.test_filter_with_modified_time_exact_day", "tests.test_coverage_push.test_file_types_with_filetime_modified", "tests.test_coverage_push.test_ignore_hidden_files", "tests.test_coverage_push.test_multiple_nonexistent_files_exits_with_error", "tests.test_cli.test_depth_limit", "tests.test_cli.test_combined_screen_reader_full_paths", "tests.test_cli.test_conflicting_filter_invert_filter", "tests.test_cli.test_number_of_lines_limit", "tests.test_cli.test_output_format_gigabytes", "tests.test_cli.test_output_format_terabytes", "tests.test_cli.test_bars_on_right_flag", "tests.test_cli.test_output_format_bytes", "tests.test_cli.test_output_format_mb", "tests.test_coverage_push.test_filter_with_accessed_time_comparison", "tests.test_coverage_push.test_empty_stdin_files_from_defaults_to_current_directory", "tests.test_cli.test_invert_filter_regex", "tests.test_coverage_push.test_custom_terminal_width", "tests.test_config_file.test_config_min_size_filtering", "tests.test_coverage_push.test_ignore_specific_directory", "tests.test_coverage_push.test_filetime_accessed_mode", "tests.test_cli.test_conflicting_file_types_filter", "tests.test_cli.test_ignore_hidden_flag", "tests.test_coverage_push.test_filter_with_changed_time_comparison", "tests.test_cli.test_combined_reverse_depth_nobars", "tests.test_coverage_push.test_full_paths_no_shortening", "tests.test_coverage_push.test_files_from_nonexistent_file_shows_error_and_defaults", "tests.test_coverage_push.test_output_format_kilobytes", "tests.test_cli.test_apparent_size_flag", "tests.test_cli.test_threads_single", "tests.test_coverage_push.test_ignore_patterns_from_file", "tests.test_cli.test_output_format_gib", "tests.test_dir_walker.test_very_deep_directory_nesting", "tests.test_coverage_push.test_limit_filesystem_boundary", "tests.test_dir_walker.test_symlink_outside_scan_directory", "tests.test_edge_cases.test_depth_limit", "tests.test_coverage_push.test_dereference_symlinks_follows_and_reports_broken", "tests.test_coverage_push.test_file_types_with_more_types_than_display_limit", "tests.test_coverage_push.test_skip_total_row", "tests.test_dir_walker.test_thread_count_produces_deterministic_output", "tests.test_coverage_push.test_null_terminated_paths_from_stdin", "tests.test_edge_cases.test_filetype_aggregation_others_node", "tests.test_coverage_push.test_min_size_filter", "tests.test_coverage_push.test_output_format_si_prefix", "tests.test_edge_cases.test_full_paths_display", "tests.test_edge_cases.test_terminal_width_extremely_wide", "tests.test_dir_walker.test_mixed_absolute_and_relative_symlinks", "tests.test_dir_walker.test_symlinks_not_followed_by_default", "tests.test_edge_cases.test_terminal_width_panic_too_narrow", "tests.test_edge_cases.test_reverse_sort" ], "failed_tests_omitted": 120, "public_eval_path": "eval.json", "summary_path": "eval-summary.json", "test_branch_errors": {}, "test_records": 965, "warnings": [] }, "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "instance_id": "bootandy__dust.62bf1e1", "metrics": { "almost_resolved": "False", "cached_input_tokens": "8474240", "calls": "107", "created_at": "2026-05-17T10:05:06.157916+00:00", "docker_cpus": "16", "docker_memory": "30g", "error_code": "", "estimated_cost_usd": "6.8461", "evaluator_problem": "False", "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "input_tokens": "8628053", "instance_id": "bootandy__dust.62bf1e1", "long_context_calls": "0", "model": "gpt-5.5", "n_resolved_tests": "440", "n_system_errors": "0", "n_tests": "584", "n_warnings": "0", "output_tokens": "61330", "paper_compliant": "False", "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5", "reasoning_effort": "xhigh", "reasoning_output_tokens": "26901", "resolved": "False", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-bootandy-dust", "run_version": "", "score": "0.7534246575342466", "session_ended_at": "2026-05-17T10:26:11.192Z", "session_started_at": "2026-05-17T10:05:08.952Z", "test_branch_errors": "{}", "total_tokens": "8689383", "wall_clock_seconds": "1262" }, "model": "gpt-5.5", "package": { "contents": [ ".", "./__pycache__", "./__pycache__/dust.cpython-312.pyc", "./compile.sh", "./dust.py", "./executable" ], "contents_omitted": 0, "submission_available_local_only": true }, "paper_compliant": false, "paper_mode": false, "reasoning_effort": "xhigh", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-bootandy-dust", "run_version": "", "usage_audit": { "available_local_artifact": true, "public_path": "usage-audit.json" } }