{
  "agent_trace": {
    "codex_logs_available_local_only": true,
    "raw_logs_published": false
  },
  "collected_at": "2026-05-18T13:25:12.410248+00:00",
  "docker_cpus": 16,
  "docker_memory": "30g",
  "eval": {
    "error_code": null,
    "failed_tests": [
      "eval.tests.test_comprehensive_generation.test_complex_markdown_features",
      "eval.tests.test_comprehensive_generation.test_full_blog_with_all_features",
      "eval.tests.test_configuration.test_default_author_config",
      "eval.tests.test_content_types.test_multiple_authors",
      "eval.tests.test_shortcodes.test_shortcode_authors",
      "eval.tests.test_additional_features.test_custom_css_js_files",
      "eval.tests.test_additional_features.test_post_with_special_characters_in_title",
      "eval.tests.test_additional_features.test_long_post_title",
      "eval.tests.test_content_handling.test_page_without_date",
      "eval.tests.test_content_handling.test_frontmatter_parsing",
      "eval.tests.test_content_handling.test_draft_post_handling",
      "eval.tests.test_markdown_features.test_shortcodes_listing",
      "eval.tests.test_markdown_features.test_table_of_contents",
      "eval.tests.test_markdown_features.test_backlinks",
      "eval.tests.test_output_generation.test_pagination",
      "eval.tests.test_output_generation.test_next_prev_links",
      "eval.tests.test_additional_features.test_empty_frontmatter_field",
      "eval.tests.test_site_generation.test_basic_site_generation",
      "eval.tests.test_cli_operations.test_init_templates_in_nonexistent_directory_fails",
      "eval.tests.test_content_edge_cases.test_pinned_field_in_frontmatter",
      "eval.tests.test_content_edge_cases.test_series_metadata_creates_series_page",
      "eval.tests.test_error_handling.test_frontmatter_invalid_date_format",
      "eval.tests.test_config_edge_cases.test_theme_path_resolution_with_valid_theme",
      "eval.tests.test_config_edge_cases.test_generate_config_file_already_exists",
      "eval.tests.test_content_edge_cases.test_banner_image_auto_detection_from_file",
      "eval.tests.test_error_handling.test_input_is_file_not_directory",
      "eval.tests.test_cli_operations.test_init_templates_fails_if_directory_exists",
      "eval.tests.test_cli_operations.test_generate_config_creates_default_config",
      "eval.tests.test_cli_operations.test_generate_config_skips_if_exists",
      "eval.tests.test_error_handling.test_markdown_with_html_content",
      "eval.tests.test_feeds.test_rss_feed_with_all_draft_content",
      "eval.tests.test_feeds.test_rss_feed_per_tag",
      "eval.tests.test_gallery.test_gallery_wildcard_description",
      "eval.tests.test_feeds.test_rss_feed_unicode_and_special_characters",
      "eval.tests.test_config_edge_cases.test_theme_folder_does_not_exist",
      "eval.tests.test_content_edge_cases.test_card_image_with_slug_extension_only",
      "eval.tests.test_cli_operations.test_start_theme_with_same_name_twice",
      "eval.tests.test_error_handling.test_malformed_yaml_config",
      "eval.tests.test_content_edge_cases.test_banner_image_from_extra_metadata",
      "eval.tests.test_content_edge_cases.test_author_singular_field",
      "eval.tests.test_content_edge_cases.test_date_extraction_from_stream_prefixed_filename",
      "eval.tests.test_cli_operations.test_new_without_input_folder_fails",
      "eval.tests.test_cli_operations.test_start_theme_creates_complete_theme_structure",
      "eval.tests.test_config_edge_cases.test_cli_override_all_config_options",
      "eval.tests.test_feeds.test_rss_feed_url_without_protocol_http",
      "eval.tests.test_gallery.test_gallery_partial_match_description",
      "eval.tests.test_feeds.test_rss_feed_multiple_authors_first_only",
      "eval.tests.test_feeds.test_rss_feed_limit_15_items",
      "eval.tests.test_gallery.test_gallery_regex_description",
      "eval.tests.test_feeds.test_json_feed_unicode_and_special_characters",
      "eval.tests.test_config_edge_cases.test_cli_override_colorscheme_creates_extra_field",
      "eval.tests.test_error_handling.test_config_wrong_type_for_field",
      "eval.tests.test_content_edge_cases.test_card_image_fallback_to_inline_image",
      "eval.tests.test_content_edge_cases.test_authors_array_field",
      "eval.tests.test_content_edge_cases.test_card_image_auto_detection_from_media_folder",
      "eval.tests.test_feeds.test_json_feed_author_without_links",
      "eval.tests.test_config_edge_cases.test_cli_override_name_url_https_footer_language_pagination",
      "eval.tests.test_feeds.test_rss_feed_channel_image",
      "eval.tests.test_cli_operations.test_generate_config_with_cli_overrides",
      "eval.tests.test_gallery.test_gallery_no_images",
      "eval.tests.test_gallery.test_gallery_corrupted_image_file",
      "eval.tests.test_feeds.test_json_feed_author_metadata",
      "eval.tests.test_gallery.test_gallery_special_characters_in_filenames",
      "eval.tests.test_feeds.test_json_feed_limit_15_items",
      "eval.tests.test_gallery.test_gallery_multiple_image_formats",
      "eval.tests.test_gallery.test_basic_gallery_generation",
      "eval.tests.test_config_edge_cases.test_cli_override_path_options",
      "eval.tests.test_error_handling.test_frontmatter_malformed_yaml",
      "eval.tests.test_content_edge_cases.test_banner_image_from_frontmatter",
      "eval.tests.test_content_edge_cases.test_card_image_with_banner_fallback",
      "eval.tests.test_content_edge_cases.test_authors_comma_separated_string",
      "eval.tests.test_config_edge_cases.test_cli_invalid_image_provider_warning",
      "eval.tests.test_cli_operations.test_init_site_creates_specific_markdown_fragments",
      "eval.tests.test_feeds.test_json_feed_author_with_avatar_but_no_url",
      "eval.tests.test_feeds.test_rss_feed_multiple_categories",
      "eval.tests.test_gallery.test_gallery_description_precedence",
      "eval.tests.test_gallery.test_gallery_invalid_regex_in_config",
      "eval.tests.test_feeds.test_json_feed_per_tag",
      "eval.tests.test_gallery.test_gallery_description_exact_match_precedence",
      "eval.tests.test_gallery.test_gallery_cover_nonexistent_uses_first"
    ],
    "failed_tests_omitted": 199,
    "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": "rochacbruno__marmite.7d4bc2d",
  "metrics": {
    "almost_resolved": "False",
    "cached_input_tokens": "7023360",
    "calls": "94",
    "created_at": "2026-05-17T14:37:19.470783+00:00",
    "docker_cpus": "16",
    "docker_memory": "30g",
    "error_code": "",
    "estimated_cost_usd": "6.1356",
    "evaluator_problem": "False",
    "host_machine": "x86_64",
    "host_system": "Linux",
    "inference_mode": "mini-swe-compatible-nointernet",
    "input_tokens": "7199855",
    "instance_id": "rochacbruno__marmite.7d4bc2d",
    "long_context_calls": "0",
    "model": "gpt-5.5",
    "n_resolved_tests": "434",
    "n_system_errors": "0",
    "n_tests": "668",
    "n_warnings": "0",
    "output_tokens": "58048",
    "paper_compliant": "False",
    "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5",
    "reasoning_effort": "xhigh",
    "reasoning_output_tokens": "13268",
    "resolved": "False",
    "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-rochacbruno-marmite",
    "run_version": "20260517T094700Z",
    "score": "0.6497005988023952",
    "session_ended_at": "2026-05-17T14:56:00.986Z",
    "session_started_at": "2026-05-17T14:37:22.573Z",
    "test_branch_errors": "{}",
    "total_tokens": "7257903",
    "wall_clock_seconds": "1118"
  },
  "model": "gpt-5.5",
  "package": {
    "contents": [
      ".",
      "./compile.sh",
      "./executable",
      "./marmite.py"
    ],
    "contents_omitted": 0,
    "submission_available_local_only": true
  },
  "paper_compliant": false,
  "paper_mode": false,
  "reasoning_effort": "xhigh",
  "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-rochacbruno-marmite",
  "run_version": "20260517T094700Z",
  "usage_audit": {
    "available_local_artifact": true,
    "public_path": "usage-audit.json"
  }
}
