{ "agent_trace": { "codex_logs_available_local_only": true, "raw_logs_published": false }, "collected_at": "2026-05-18T13:25:16.593404+00:00", "docker_cpus": 16, "docker_memory": "30g", "eval": { "error_code": null, "failed_tests": [ "eval.tests.test_basic.test_metadata_lines_ignored", "tests.test_harvest.test_fixture_top_view[invalid-lines]", "tests.test_harvest.test_fixture_top_view[recursive]", "tests.test_harvest.test_sorted_flag", "tests.test_harvest.test_recursive_stacks", "tests.test_harvest.test_invalid_lines_handling", "tests.test_harvest.test_fixture_flamegraph_view[ignore-metadata-lines]", "tests.test_harvest.test_search_functionality", "tests.test_harvest.test_fixture_top_view[py-spy-simple]", "tests.test_integration.test_sorted_flag_initial_state", "tests.test_data_processing.test_empty_file_handling", "tests.test_integration.test_navigation_up_at_root_stays_at_root", "tests.test_integration.test_zoom_then_search_then_unzoom", "tests.test_integration.test_table_view_navigation_at_last_row", "tests.test_data_processing.test_parent_counts_include_all_children", "tests.test_integration.test_table_sort_by_own_then_navigate", "tests.test_integration.test_reset_button_clears_state", "tests.test_harvest.test_py_spy_simple_specific_values", "tests.test_harvest.test_fixture_top_view[ignore-metadata-lines]", "tests.test_harvest.test_ignore_metadata_lines", "tests.test_integration.test_sorted_debug_flags_combined", "tests.test_integration.test_search_no_matches_shows_message", "tests.test_integration.test_table_view_page_navigation", "tests.test_parsing.test_file_input_deep_nested_stack", "tests.test_parsing.test_whitespace_only_lines_ignored", "tests.test_parsing.test_very_long_function_names", "tests.test_parsing.test_large_count_values", "tests.test_integration.test_search_next_previous_navigation", "tests.test_parsing.test_invalid_lines_skipped_gracefully", "tests.test_parsing.test_metadata_lines_ignored", "tests.test_parsing.test_sorted_flag_changes_order", "tests.test_parsing.test_empty_file_handling", "tests.test_integration.test_navigation_left_with_no_left_sibling", "tests.test_parsing.test_special_characters_in_function_names", "tests.test_parsing.test_file_input_simple_stack", "tests.test_integration.test_zoom_multiple_levels_deep", "tests.test_parsing.test_wide_tree_many_siblings", "tests.test_integration.test_freeze_mode_toggle", "tests.test_parsing.test_navigation_cursor_movement", "tests.test_parsing.test_repeated_stack_prefix_aggregation", "tests.test_parsing.test_zero_count_accepted", "tests.test_performance.test_mixed_depth_stacks", "tests.test_performance.test_deep_and_wide_combined", "tests.test_performance.test_deep_stack_250_levels_startup", "tests.test_tui_additional.test_table_enter_searches_selected_row", "tests.test_performance.test_wide_tree_1000_siblings_startup", "tests.test_performance.test_many_unique_stacks_10000_lines", "tests.test_tui_additional.test_table_sort_by_own", "tests.test_performance.test_large_sample_counts_billions", "tests.test_performance.test_table_view_with_thousands_of_rows", "tests.test_parsing.test_debug_flag_shows_timing", "tests.test_performance.test_sorted_flag_with_large_input", "tests.test_parsing.test_single_level_no_nesting", "tests.test_tui_display.test_zoom_functionality", "tests.test_tui_additional.test_search_escape_cancels_without_applying", "tests.test_performance.test_zoom_into_heavily_nested_area", "tests.test_performance.test_search_on_large_input_many_matches", "tests.test_tui_additional.test_debug_mode_toggle", "tests.test_performance.test_deep_navigation_traverse_entire_stack", "tests.test_tui_additional.test_table_page_down_and_up", "tests.test_tui_additional.test_invalid_regex_error_message", "tests.test_tui_additional.test_quit_with_ctrl_c", "tests.test_tui_additional.test_table_sort_by_total", "tests.test_tui_additional.test_search_with_no_matches", "tests.test_tui_display.test_search_typing", "tests.test_tui_additional.test_empty_search_clears_pattern", "tests.test_tui_display.test_initial_display_py_spy_simple", "tests.test_tui_display.test_navigation_and_selection", "tests.test_tui_display.test_sorted_flag", "tests.test_tui_display.test_table_view", "tests.test_tui_display.test_search_results", "tests.test_tui_display.test_hjkl_navigation", "tests.test_tui_display.test_search_mode_empty", "tests.test_tui_display.test_debug_flag", "tests.test_tui_navigation.test_navigation_down_j", "tests.test_tui_additional.test_navigation_up_at_boundary", "tests.test_tui_display.test_search_next_result", "tests.test_tui_display.test_reset_after_zoom", "tests.test_tui_navigation.test_navigation_up_k", "tests.test_tui_navigation.test_initial_screen" ], "failed_tests_omitted": 42, "public_eval_path": "eval.json", "summary_path": "eval-summary.json", "test_branch_errors": {}, "test_records": 311, "warnings": [] }, "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "instance_id": "ys-l__flamelens.0b4dc33", "metrics": { "almost_resolved": "False", "cached_input_tokens": "3006464", "calls": "56", "created_at": "2026-05-17T16:55:46.948639+00:00", "docker_cpus": "16", "docker_memory": "30g", "error_code": "", "estimated_cost_usd": "3.3074", "evaluator_problem": "False", "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "input_tokens": "3160048", "instance_id": "ys-l__flamelens.0b4dc33", "long_context_calls": "0", "model": "gpt-5.5", "n_resolved_tests": "130", "n_system_errors": "0", "n_tests": "224", "n_warnings": "0", "output_tokens": "34543", "paper_compliant": "False", "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5", "reasoning_effort": "xhigh", "reasoning_output_tokens": "9565", "resolved": "False", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ys-l-flamelens", "run_version": "20260517T094700Z", "score": "0.5803571428571429", "session_ended_at": "2026-05-17T17:14:59.681Z", "session_started_at": "2026-05-17T16:55:50.379Z", "test_branch_errors": "{}", "total_tokens": "3194591", "wall_clock_seconds": "1149" }, "model": "gpt-5.5", "package": { "contents": [ ".", "./compile.sh", "./executable", "./main.c" ], "contents_omitted": 0, "submission_available_local_only": true }, "paper_compliant": false, "paper_mode": false, "reasoning_effort": "xhigh", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ys-l-flamelens", "run_version": "20260517T094700Z", "usage_audit": { "available_local_artifact": true, "public_path": "usage-audit.json" } }