{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:07.476911+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "tests.test_basic_invocation.test_completion_bash",
      "tests.test_tui_advanced.test_tui_half_page_down",
      "tests.test_tui_advanced.test_tui_screen_sync",
      "tests.test_config_detailed.test_force_screen_with_content",
      "tests.test_additional_main_coverage.test_argsToFiles_with_glob_pattern",
      "tests.test_coverage_target.test_exec_mode_simple_command",
      "tests.test_coverage_target.test_ruler_absolute",
      "tests.test_coverage_target_fixed.test_exec_mode_echo",
      "tests.test_final_coverage_push.test_multiple_glob_patterns",
      "tests.test_file_glob_patterns.test_glob_pattern_matching",
      "tests.test_final_coverage_push.test_glob_pattern_with_brackets",
      "tests.test_edge_cases_final.test_skip_lines_more_than_content",
      "tests.test_more_cli_options.test_all_flags_combination_1",
      "tests.test_more_cli_options.test_notify_eof_various",
      "tests.test_final_coverage_push_70.test_ruler_with_file",
      "tests.test_stdin_and_edge_cases.test_notify_eof_with_file",
      "tests.test_coverage_target.test_exec_mode_ps_command",
      "tests.test_coverage_target.test_ruler_relative",
      "tests.test_coverage_target_fixed.test_ruler_with_file",
      "tests.test_final_coverage_push.test_multiple_compressed_files",
      "tests.test_tui_actions.test_tui_toggle_mouse",
      "tests.test_more_cli_options.test_all_flags_combination_3",
      "tests.test_more_cli_options.test_ruler_mode_0",
      "tests.test_tui_actions.test_tui_ctrl_u_half_page_backward",
      "tests.test_more_cli_options.test_caption_with_content",
      "tests.test_tui_comprehensive.test_comprehensive_mark_operations",
      "tests.test_coverage_target.test_ruler_none",
      "tests.test_coverage_target_fixed.test_notify_eof_with_file",
      "tests.test_coverage_target.test_notify_eof_flag",
      "tests.test_more_cli_options.test_ruler_mode_1",
      "tests.test_tui_actions.test_tui_horizontal_scroll_right",
      "tests.test_tui_actions.test_tui_navigation_home",
      "tests.test_tui_comprehensive.test_plain_and_raw_toggles",
      "tests.test_tui_comprehensive.test_sidebar_mark_list",
      "tests.test_tui_comprehensive.test_alternate_rows_with_navigation",
      "tests.test_coverage_target_fixed.test_force_screen_with_file",
      "tests.test_tui_comprehensive.test_comprehensive_file_navigation",
      "tests.test_tui_actions.test_tui_toggle_ruler",
      "tests.test_more_cli_options.test_ruler_mode_2",
      "tests.test_tui_actions.test_tui_down_arrow",
      "tests.test_tui_comprehensive.test_alt_scroll_movements",
      "tests.test_tui_comprehensive.test_with_different_file_types",
      "tests.test_tui_flows.test_flow_horizontal_nav_extensive",
      "tests.test_tui_flows.test_flow_column_mode_full",
      "tests.test_tui_actions.test_tui_navigation_end",
      "tests.test_tui_comprehensive.test_mouse_toggle",
      "tests.test_tui_comprehensive.test_sidebar_doc_list",
      "tests.test_tui_flows.test_flow_ruler_toggles",
      "tests.test_tui_flows.test_flow_cancel_operations",
      "tests.test_tui_flows.test_flow_mouse_toggle",
      "tests.test_tui_actions.test_tui_toggle_sidebar_help",
      "tests.test_tui_actions.test_tui_page_down",
      "tests.test_tui_comprehensive.test_sidebar_scrolling",
      "tests.test_tui_actions.test_tui_up_arrow",
      "tests.test_tui_flows.test_flow_shift_navigation",
      "tests.test_tui_comprehensive.test_comprehensive_column_operations",
      "tests.test_tui_comprehensive.test_with_csv_file_operations",
      "tests.test_tui_actions.test_tui_horizontal_scroll_left",
      "tests.test_tui_flows.test_flow_large_file_extensive",
      "tests.test_tui_flows.test_flow_navigation_sequence",
      "tests.test_tui_flows.test_flow_status_line_toggle",
      "tests.test_tui_flows.test_flow_screen_sync",
      "tests.test_tui_actions.test_tui_toggle_sidebar_mark_list",
      "tests.test_tui_flows.test_flow_alternate_rows_navigation",
      "tests.test_tui_comprehensive.test_line_number_with_navigation",
      "tests.test_tui_actions.test_tui_page_up",
      "tests.test_tui_comprehensive.test_cancel_operations",
      "tests.test_tui_interactions.test_tui_page_up",
      "tests.test_tui_flows.test_flow_wrap_and_scroll",
      "tests.test_tui_flows.test_flow_alt_scroll",
      "tests.test_tui_comprehensive.test_comprehensive_sidebar_operations",
      "tests.test_tui_comprehensive.test_horizontal_scrolling_extensive",
      "tests.test_tui_flows.test_flow_plain_and_raw",
      "tests.test_tui_flows.test_flow_sidebar_help_nav",
      "tests.test_tui_flows.test_flow_mark_remove",
      "tests.test_tui_actions.test_tui_toggle_sidebar_doc_list",
      "tests.test_tui_flows.test_flow_line_numbers_navigation",
      "tests.test_tui_comprehensive.test_shift_navigation",
      "tests.test_tui_interactions.test_tui_with_delimiter",
      "tests.test_tui_flows.test_flow_column_align_operations"
    ],
    "failed_tests_omitted": 396,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 2447,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "noborus__ov.b96c2ba",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "5622400",
    "calls": "87",
    "created_at": "2026-05-17T13:29:01.373743+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "5.4100",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "5864992",
    "instance_id": "noborus__ov.b96c2ba",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "1465",
    "n_system_errors": "0",
    "n_tests": "1854",
    "n_warnings": "0",
    "output_tokens": "46194",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "12703",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-noborus-ov",
    "run_version": "20260517T094700Z",
    "score": "0.7901833872707659",
    "session_ended_at": "2026-05-17T13:43:10.818Z",
    "session_started_at": "2026-05-17T13:29:04.445Z",
    "test_branch_errors": "{}",
    "total_tokens": "5911186",
    "wall_clock_seconds": "846"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./ov-disable-default.yaml",
      "./ov-less.yaml",
      "./ov.py",
      "./ov.yaml"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-noborus-ov",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}