{ "agent_trace": { "codex_logs_available_local_only": true, "raw_logs_published": false }, "collected_at": "2026-05-18T13:24:59.318042+00:00", "docker_cpus": 16, "docker_memory": "30g", "eval": { "error_code": null, "failed_tests": [ "tests.test_cli_basics.test_env_bore_server_used_when_no_to_flag", "tests.test_cli_basics.test_server_min_port_explicit_1024", "tests.test_cli_basics.test_local_long_host_name", "tests.test_client.test_port_boundary_minimum", "tests.test_client.test_env_server_invalid", "tests.test_client.test_port_boundary_maximum", "tests.test_client.test_remote_port_boundary_maximum", "tests.test_client.test_dns_lookup_failure", "tests.test_client.test_remote_port_boundary_minimum", "tests.test_client.test_localhost_hostname", "tests.test_client.test_unreachable_ip_address", "tests.test_auth.test_auth_wrong_secret_error@server_tests", "tests.test_client.test_very_long_hostname", "tests.test_client.test_invalid_domain_name", "tests.test_client.test_empty_server_address", "tests.test_client.test_whitespace_server_address", "tests.test_auth.test_auth_empty_secret_mismatch@server_tests", "tests.test_auth.test_auth_long_secret@server_tests", "tests.test_auth.test_auth_case_sensitive_secret@server_tests", "tests.test_auth.test_auth_repeated_char_secret@server_tests", "tests.test_auth.test_auth_multiple_clients_different_secrets@server_tests", "tests.test_auth.test_auth_prefix_suffix_mismatch@server_tests", "tests.test_auth.test_auth_rapid_authentication_attempts@server_tests", "tests.test_client_gaps.test_client_handles_wrong_authentication_secret@server_tests", "tests.test_client_gaps.test_client_handles_local_connection_failure@server_tests", "tests.test_edge_cases.test_local_empty_hostname@server_tests", "tests.test_edge_cases.test_local_nonexistent_hostname@server_tests", "tests.test_edge_cases.test_local_255_char_hostname@server_tests", "tests.test_edge_cases.test_local_unicode_hostname@server_tests", "tests.test_edge_cases.test_local_to_arg_with_spaces@server_tests", "tests.test_edge_cases.test_local_to_arg_with_tab@server_tests", "tests.test_harvest.test_mismatched_secret[None-my secret]@server_tests", "tests.test_harvest.test_invalid_address@server_tests", "tests.test_harvest.test_very_long_frame@server_tests", "tests.test_integration.test_client_wrong_secret_to_protected_server@server_tests", "tests.test_integration.test_authentication_mismatched_secrets@server_tests", "tests.test_protocol.test_oversized_frame_rejected@server_tests", "tests.test_protocol.test_wrong_authentication_rejected@server_tests", "tests.test_protocol.test_timeout_waiting_for_authentication@server_tests", "tests.test_protocol.test_timeout_waiting_for_hello@server_tests", "tests.test_protocol.test_heartbeat_messages_sent@server_tests", "tests.test_protocol.test_message_exactly_at_max_length@server_tests", "tests.test_protocol.test_negative_port_number@server_tests", "tests.test_protocol.test_connection_with_unknown_message_type@server_tests", "tests.test_protocol.test_very_large_port_number@server_tests", "tests.test_server.test_server_negative_port_error@server_tests", "tests.test_server_gaps.test_server_sends_heartbeat_after_hello@server_tests", "tests.test_server_gaps.test_server_handles_unexpected_authenticate_message@server_tests", "tests.test_server_gaps.test_server_sends_connection_message_on_incoming_tunnel@server_tests", "tests.test_server_gaps.test_server_accept_message_initiates_forwarding@server_tests", "tests.test_server_gaps.test_server_multiple_incoming_connections_get_unique_uuids@server_tests", "tests.test_server_gaps.test_server_stale_connection_cleanup@server_tests", "tests.test_server_gaps.test_server_handles_random_port_allocation_exhaustion@server_tests", "tests.test_server_gaps.test_server_connection_message_timing@server_tests", "tests.test_server_gaps.test_server_heartbeat_continues_after_tunnel_disconnect@server_tests", "eval.tests.test_args_parsing.test_unexpected_args[args0-needles0]", "eval.tests.test_args_parsing.test_local_missing_option_values[args3-needles3]", "eval.tests.test_cli_io.test_local_invalid_host_name_exit_1_and_shows_name_or_service_not_known", "eval.tests.test_bore.TestEdgeCases.test_local_short_form_options", "eval.tests.test_bore.TestEdgeCases.test_local_very_high_port" ], "failed_tests_omitted": 0, "public_eval_path": "eval.json", "summary_path": "eval-summary.json", "test_branch_errors": {}, "test_records": 450, "warnings": [] }, "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "instance_id": "ekzhang__bore.8e059cd", "metrics": { "almost_resolved": "False", "cached_input_tokens": "3455872", "calls": "77", "created_at": "2026-05-17T10:47:32.804481+00:00", "docker_cpus": "16", "docker_memory": "30g", "error_code": "", "estimated_cost_usd": "3.7228", "evaluator_problem": "False", "host_machine": "x86_64", "host_system": "Linux", "inference_mode": "mini-swe-compatible-nointernet", "input_tokens": "3607477", "instance_id": "ekzhang__bore.8e059cd", "long_context_calls": "0", "model": "gpt-5.5", "n_resolved_tests": "346", "n_system_errors": "0", "n_tests": "394", "n_warnings": "0", "output_tokens": "41228", "paper_compliant": "False", "pricing_source": "https://developers.openai.com/api/docs/models/gpt-5.5", "reasoning_effort": "xhigh", "reasoning_output_tokens": "11442", "resolved": "False", "run_name": "cpx62-miniswecompat-xhigh-20260517T094700Z-ekzhang-bore", "run_version": "", "score": "0.8781725888324873", "session_ended_at": "2026-05-17T11:00:58.693Z", "session_started_at": "2026-05-17T10:47:36.890Z", "test_branch_errors": "{}", "total_tokens": "3648705", "wall_clock_seconds": "801" }, "model": "gpt-5.5", "package": { "contents": [ ".", "./bore_reimpl.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-ekzhang-bore", "run_version": "", "usage_audit": { "available_local_artifact": true, "public_path": "usage-audit.json" } }