{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:24:56.663803+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_edge_cases.test_malformed_json_highlight_lines",
      "eval.tests.test_input_sources.test_from_code_stdin",
      "eval.tests.test_output_destinations.test_output_raw_to_stdout",
      "tests.test_configuration.test_silent_mode_with_first_run_shows_info",
      "tests.test_config_gaps.test_solid_background_color_string",
      "tests.test_configuration.test_custom_config_file_with_config_flag",
      "tests.test_config_gaps.test_dimension_value_invalid_non_string_non_number",
      "tests.test_config_gaps.test_dimension_value_large_number",
      "tests.test_configuration.test_config_with_numeric_string_values",
      "tests.test_edges_utilities.test_margin_x_only",
      "tests.test_error_handling.test_malformed_config_file",
      "tests.test_edges_utilities.test_code_min_width_enforcement",
      "tests.test_configuration.test_silent_mode_with_execute_command",
      "tests.test_config_gaps.test_dimension_value_max_string",
      "tests.test_error_handling.test_raw_highlight_lines_invalid_json",
      "tests.test_from_image.test_from_image_basic_stdin_exact_golden",
      "tests.test_error_handling.test_binary_data_in_file",
      "tests.test_error_handling.test_from_file_with_directory",
      "tests.test_error_handling.test_delete_line_invalid_format",
      "tests.test_configuration.test_silent_mode_still_shows_errors",
      "tests.test_config_gaps.test_gradient_with_exact_boundary_positions",
      "tests.test_config_gaps.test_gradient_with_multiple_stops_and_max_dimensions",
      "tests.test_config_gaps.test_default_scale_factor_when_not_specified",
      "tests.test_configuration.test_config_flag_with_nonexistent_file",
      "tests.test_config_gaps.test_dimension_value_invalid_string_not_max",
      "tests.test_edges_utilities.test_margin_y_only",
      "tests.test_from_image.test_from_image_medium_size_exact_golden",
      "tests.test_error_handling.test_invalid_range_format_multiple_colons",
      "tests.test_error_handling.test_highlight_range_invalid_format",
      "tests.test_configuration.test_silent_flag_position_independence",
      "tests.test_edges_utilities.test_code_width_calculation_with_varied_line_lengths",
      "tests.test_configuration.test_config_with_empty_json_object",
      "tests.test_error_handling.test_invalid_theme_name",
      "tests.test_highlight_ranges.test_relative_highlight_range",
      "tests.test_error_handling.test_add_line_invalid_format",
      "tests.test_configuration.test_config_with_unknown_fields_ignored",
      "tests.test_from_image.test_from_image_no_shadow_exact_golden",
      "tests.test_configuration.test_first_run_creates_default_config",
      "tests.test_config_gaps.test_gradient_with_multiple_intermediate_stops",
      "tests.test_config_gaps.test_config_with_watermark_present",
      "tests.test_configuration.test_config_flag_with_invalid_json",
      "tests.test_from_image.test_from_image_combined_features_exact_golden",
      "tests.test_from_image.test_from_image_with_title_exact_golden",
      "tests.test_edges_utilities.test_code_tab_replacement",
      "tests.test_configuration.test_config_directory_created_if_missing",
      "tests.test_configuration.test_config_absolute_path_with_spaces",
      "tests.test_edges_utilities.test_code_trim_with_no_common_indentation",
      "tests.test_highlight_ranges.test_highlight_custom_color",
      "tests.test_from_image.test_from_image_large_image_exact_golden",
      "tests.test_highlight_ranges.test_absolute_highlight_range_with_code_range",
      "tests.test_highlight_ranges.test_relative_highlight_beyond_range",
      "tests.test_from_image.test_from_image_png_output_format",
      "tests.test_configuration.test_config_partial_fails_appropriately",
      "tests.test_configuration.test_config_file_location_standard_path",
      "tests.test_from_image.test_from_image_with_watermark_exact_golden",
      "tests.test_config_gaps.test_config_without_watermark_uses_default",
      "tests.test_config_gaps.test_config_with_explicit_scale_factor",
      "tests.test_configuration.test_config_flag_with_relative_path",
      "tests.test_from_image.test_from_image_with_background_color_exact_golden",
      "tests.test_configuration.test_config_json_with_trailing_comma",
      "tests.test_edges_utilities.test_code_trim_leading_space",
      "tests.test_edges_utilities.test_margin_zero_values",
      "tests.test_highlight_ranges.test_highlight_range_simple",
      "tests.test_edges_utilities.test_code_trim_with_empty_lines_preserves_blank_lines",
      "tests.test_highlight_ranges.test_raw_highlight_lines_multiple_colors",
      "tests.test_highlight_ranges.test_single_line_file_highlight",
      "tests.test_highlight_ranges.test_highlight_before_code_range",
      "tests.test_from_image.test_from_image_invalid_png_error",
      "tests.test_configuration.test_cli_flag_overrides_config_breadcrumbs",
      "tests.test_config_gaps.test_dimension_value_zero",
      "tests.test_from_image.test_from_image_with_margins_exact_golden",
      "tests.test_highlight_ranges.test_highlight_with_start_line_number",
      "tests.test_config_gaps.test_scale_factor_validation_not_enforced_during_deserialization",
      "tests.test_configuration.test_second_run_uses_existing_config",
      "tests.test_configuration.test_silent_mode_suppresses_success_messages",
      "tests.test_configuration.test_multiple_runs_different_homes_isolated",
      "tests.test_edges_utilities.test_margin_custom_values",
      "tests.test_edges_utilities.test_line_number_length_calculation_large_numbers",
      "tests.test_edges_utilities.test_single_line_number_display",
      "tests.test_highlight_ranges.test_highlight_single_line"
    ],
    "failed_tests_omitted": 217,
    "public_eval_path": "eval.json",
    "summary_path": "eval-summary.json",
    "test_branch_errors": {},
    "test_records": 853,
    "warnings": []
  },
  "host_machine": "x86_64",
  "host_system": "Linux",
  "inference_mode": "mini-swe-compatible-nointernet",
  "instance_id": "codesnap-rs__codesnap.f81e4f3",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "3208192",
    "calls": "36",
    "created_at": "2026-05-17T10:23:03.033740+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "3.3166",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "3357531",
    "instance_id": "codesnap-rs__codesnap.f81e4f3",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "439",
    "n_system_errors": "0",
    "n_tests": "571",
    "n_warnings": "0",
    "output_tokens": "32194",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "10399",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-codesnap-rs-codesnap",
    "run_version": "",
    "score": "0.7688266199649737",
    "session_ended_at": "2026-05-17T10:32:39.545Z",
    "session_started_at": "2026-05-17T10:23:06.320Z",
    "test_branch_errors": "{}",
    "total_tokens": "3389725",
    "wall_clock_seconds": "573"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./codesnap_cli.py",
      "./compile.sh",
      "./executable"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-codesnap-rs-codesnap",
  "run_version": "",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}