{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:07.706639+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_coverage_targeted.TestCompressedState.test_compressed_state_saves_and_restores",
      "eval.tests.test_additional_coverage.TestEditFilesModeEdgeCases.test_edit_files_mode_empty_path_fails",
      "eval.tests.test_coverage_targeted.TestEditTomlMode.test_edit_in_toml_mode",
      "eval.tests.test_coverage_boost.TestLogJsonVariants.test_log_json_all",
      "eval.tests.test_additional_coverage.TestOrderByStatusCoverage.test_order_by_status_with_mixed_statuses",
      "eval.tests.test_coverage_targeted.TestCleanCommand.test_clean_successful_only_keeps_failed",
      "eval.tests.test_additional_coverage.TestEditCoverage.test_edit_multiple_tasks",
      "eval.tests.test_coverage_targeted.TestFollowNoTaskId.test_follow_multiple_running_tasks_error",
      "eval.tests.test_coverage_boost.TestUnlimitedParallel.test_group_with_unlimited_parallel",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_label_contains",
      "eval.tests.test_coverage_boost.TestQueryDescendingOrder.test_order_by_command_desc",
      "eval.tests.test_additional_coverage.TestStateOrderByColumns.test_order_by_label_desc",
      "eval.tests.test_coverage_boost.TestUnlimitedParallel.test_unlimited_parallel_many_tasks",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_no_command_fails",
      "eval.tests.test_coverage_targeted.TestTaskManagementEdgeCases.test_enqueue_all_flag",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_with_label",
      "eval.tests.test_additional_coverage.TestDaemonizeFunctionality.test_pueued_daemonize_forks_to_background",
      "eval.tests.test_coverage_targeted.TestEditTomlMode.test_edit_files_mode_with_editor",
      "eval.tests.test_coverage_boost.TestQueryDescendingOrder.test_order_by_end_desc",
      "eval.tests.test_coverage_boost.TestLogJsonVariants.test_log_json_lines",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_status_running",
      "eval.tests.test_coverage_targeted.TestFollowNoTaskId.test_follow_single_running_task_auto_select",
      "eval.tests.test_coverage_boost.TestQueryDatetimeCoverage.test_filter_start_equals_datetime_format",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_command_contains",
      "eval.tests.test_additional_coverage.TestOrderByStatusCoverage.test_order_by_status_with_done_tasks",
      "eval.tests.test_additional_coverage.TestStateOrderByColumns.test_order_by_path_asc",
      "eval.tests.test_additional_coverage.TestEditCoverage.test_edit_running_task_fails",
      "eval.tests.test_coverage_targeted.TestGroupStatusHandling.test_group_status_shows_in_group_list",
      "eval.tests.test_coverage_targeted.TestRemoveCommand.test_remove_running_task_fails",
      "eval.tests.test_daemon_add.TestAddWithDelay.test_add_with_delay_seconds",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_with_working_directory",
      "eval.tests.test_coverage_targeted.TestEnvCommands.test_env_set_on_stashed_task",
      "eval.tests.test_coverage_boost.TestDaemonRestart.test_daemon_state_persistence",
      "eval.tests.test_coverage_targeted.TestCallbackTemplate.test_callback_with_failed_task",
      "eval.tests.test_coverage_targeted.TestEditTomlMode.test_edit_with_nonzero_exit_editor",
      "eval.tests.test_additional_coverage.TestDaemonizeFunctionality.test_pueued_daemonize_with_profile",
      "eval.tests.test_coverage_boost.TestQueryDescendingOrder.test_order_by_label_desc",
      "eval.tests.test_coverage_boost.TestQueryDatetimeCoverage.test_filter_start_with_time_only",
      "eval.tests.test_coverage_targeted.TestFollowNoTaskId.test_follow_with_lines",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_status_stashed",
      "eval.tests.test_additional_coverage.TestOrderByStatusCoverage.test_order_by_enqueue_at_with_queued_tasks",
      "eval.tests.test_coverage_targeted.TestGroupStatusHandling.test_group_json_shows_all_groups",
      "eval.tests.test_additional_coverage.TestStateOrderByColumns.test_order_by_command_asc",
      "eval.tests.test_additional_coverage.TestSendCommand.test_send_input_to_running_task",
      "eval.tests.test_coverage_targeted.TestSwitchCommand.test_switch_two_stashed_tasks",
      "eval.tests.test_daemon_add.TestAddWithDelay.test_add_with_delay_hours",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_with_group",
      "eval.tests.test_coverage_targeted.TestEnvCommands.test_env_unset_on_stashed_task",
      "eval.tests.test_coverage_targeted.TestCallbackTemplate.test_callback_with_template_variables",
      "eval.tests.test_additional_coverage.TestDaemonizeFunctionality.test_pueued_daemonize_with_verbose",
      "eval.tests.test_coverage_targeted.TestLogRemote.test_remote_log_lines_limit",
      "eval.tests.test_additional_coverage.TestSettingsSave.test_daemon_creates_config_in_xdg_dir",
      "eval.tests.test_coverage_targeted.TestShowExpandedAliases.test_show_expanded_aliases_config",
      "eval.tests.test_coverage_boost.TestCliEdgeCases.test_verbose_v",
      "eval.tests.test_coverage_boost.TestQueryDatetimeCoverage.test_filter_start_with_full_time",
      "eval.tests.test_coverage_targeted.TestResetAllGroups.test_reset_all_with_force",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_status_neq",
      "eval.tests.test_coverage_targeted.TestTaskManagementEdgeCases.test_kill_specific_task_with_signal",
      "eval.tests.test_additional_coverage.TestProcessHelperCoverage.test_kill_with_sigkill",
      "eval.tests.test_additional_coverage.TestOrderByStatusCoverage.test_order_by_priority_sorts_correctly",
      "eval.tests.test_coverage_targeted.TestFollowNoTaskId.test_follow_no_running_tasks_error",
      "eval.tests.test_daemon_advanced.TestDependencies.test_task_with_dependency_waits",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_with_priority",
      "eval.tests.test_coverage_targeted.TestEnvCommands.test_env_set_invalid_task_id",
      "eval.tests.test_additional_coverage.TestSettingsSave.test_daemon_reads_config_from_xdg_dir",
      "eval.tests.test_additional_coverage.TestDaemonDirectoryCreation.test_daemon_creates_pueue_directory",
      "eval.tests.test_coverage_targeted.TestLogRemote.test_remote_log_all_tasks",
      "eval.tests.test_coverage_boost.TestCliEdgeCases.test_verbose_vvv",
      "eval.tests.test_coverage_boost.TestQueryDatetimeCoverage.test_enqueue_at_filter_datetime",
      "eval.tests.test_coverage_targeted.TestResetAllGroups.test_reset_nonexistent_group_fails",
      "eval.tests.test_coverage_targeted.TestStatusWithAllColumns.test_status_all_columns",
      "eval.tests.test_coverage_targeted.TestTaskManagementEdgeCases.test_kill_with_numeric_signal",
      "eval.tests.test_additional_coverage.TestProcessHelperCoverage.test_kill_with_sigstop_signal",
      "eval.tests.test_coverage_targeted.TestStatusQueries.test_query_last_limit",
      "eval.tests.test_daemon_advanced.TestPriority.test_priority_in_json",
      "eval.tests.test_additional_coverage.TestOrderByStatusCoverage.test_priority_sort_on_parallel_tasks",
      "eval.tests.test_daemon_advanced.TestDependencies.test_dependency_column_in_status",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_with_dependencies",
      "eval.tests.test_daemon_add.TestAddBasic.test_add_basic",
      "eval.tests.test_coverage_targeted.TestStatusNoColorMode.test_status_no_color"
    ],
    "failed_tests_omitted": 998,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 1223,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "nukesor__pueue.8b9d6fe",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "2300928",
    "calls": "48",
    "created_at": "2026-05-17T13:30:11.798780+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "3.0805",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "2465947",
    "instance_id": "nukesor__pueue.8b9d6fe",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "109",
    "n_system_errors": "0",
    "n_tests": "638",
    "n_warnings": "0",
    "output_tokens": "36833",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "9174",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-nukesor-pueue",
    "run_version": "20260517T094700Z",
    "score": "0.17084639498432602",
    "session_ended_at": "2026-05-17T13:41:57.805Z",
    "session_started_at": "2026-05-17T13:30:14.758Z",
    "test_branch_errors": "{}",
    "total_tokens": "2502780",
    "wall_clock_seconds": "703"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./completion_data",
      "./completion_data/bash",
      "./completion_data/elvish",
      "./completion_data/fish",
      "./completion_data/nushell",
      "./completion_data/power-shell",
      "./completion_data/zsh",
      "./executable",
      "./pueue_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-nukesor-pueue",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}
