1023. 1023 "explain the architectural differences",
1024. 1024 substantive_answer,
1025. 1025 Some(&spec),
1026. 1026 &receipts,
1027. 1027 false,
1028. 1028 );
1029. 1029 assert_eq!(blocked, None);
1030. 1030
1031. 1031 // Empty or non-substantive answer -> blocked
1032. 1032 let blocked_empty = p.evaluate_receipts(
1033. 1033 "explain the architectural differences",
1034. 1034 "Okay",
1035. 1035 Some(&spec),
1036. 1036 &receipts,
1037. 1037 false,
1038. 1038 );
1039. 1039 assert!(matches!(
1040. 1040 blocked_empty,
1041. 1041 Some(BlockReason::ExecutionReceiptMissing { .. })
1042. 1042 ));
1043. 1043 }
1044. 1044
1045. 1045 #[test]
1046. 1046 fn test_is_code_path_accurately_classifies_code_vs_doc_paths() {
1047. 1047 assert!(is_code_path("src/main.rs"));
1048. 1048 assert!(is_code_path("backend/app.py"));
1049. 1049 assert!(is_code_path("web/index.ts"));
1050. 1050 assert!(is_code_path("scripts/deploy.sh"));
1051. 1051
1052. 1052 assert!(!is_code_path("README.md"));
1053. 1053 assert!(!is_code_path("docs/architecture.md"));
1054. 1054 assert!(!is_code_path("recipes/sourdough.txt"));
1055. 1055 assert!(!is_code_path("data/analysis.csv"));
1056. 1056 assert!(!is_code_path("notes.org"));
1057. 1057 }
1058. 1058
1059. 1059 #[test]
1060. 1060 fn test_universal_doc_modification_with_verify_not_blocked_on_bash() {
1061. 1061 let p = StopPolicy::default();
1062. 1062 let prompt = "Update README.md to describe the release steps and verify that all links are formatted correctly.";
1063. 1063 let final_text = "Updated README.md with release steps and verified that the Markdown links match the repository structure.";
1064. 1064
1065. 1065 let receipts = ReceiptSummary {
1066. 1066 total_tool_calls: 2,
1067. 1067 successful_tool_calls: 2,
1068. 1068 files_modified: 1,
1069. 1069 code_files_modified: 0,
1070. 1070 doc_files_modified: 1,
1071. 1071 read_or_inspected: 1,
1072. 1072 ..Default::default()
1073. 1073 };
1074. 1074
1075. 1075 // Even though prompt says "verify that", since no code was modified and no code execution was demanded,
1076. 1076 // it must NOT block on non-existent bash commands or stale verification!
1077. 1077 let blocked = p.evaluate_receipts(prompt, final_text, None, &receipts, false);
1078. 1078 assert_eq!(blocked, None);
1079. 1079
1080. 1080 let blocked_stale = p.evaluate_receipts(prompt, final_text, None, &receipts, true);
1081. 1081 assert_eq!(blocked_stale, None);
1082. 1082 }
1083. 1083
1084. 1084 #[test]
1085. 1085 fn test_universal_research_and_lifestyle_with_verify_not_blocked() {
1086. 1086 let p = StopPolicy::default();
1087. 1087 let prompt =
1088. 1088 "Compare the top 3 pour-over drippers and verify that the brew ratios are accurate.";
1089. 1089 let final_text = "Here is a detailed comparison of Hario V60, Kalita Wave, and Chemex. All brew ratios are verified between 1:15 and 1:17 for balanced extraction across light and medium roasts.";
1090. 1090
1091. 1091 let receipts = ReceiptSummary {
1092. 1092 total_tool_calls: 1,
1093. 1093 successful_tool_calls: 1,
1094. 1094 read_or_inspected: 1,
1095. 1095 ..Default::default()
1096. 1096 };
1097. 1097
1098. 1098 let blocked = p.evaluate_receipts(prompt, final_text, None, &receipts, false);
1099. 1099 assert_eq!(blocked, None);
1100. 1100 }
1101. 1101
1102. 1102 #[test]
1103. 1103 fn test_code_modification_with_verify_blocked_without_execution() {
1104. 1104 let p = StopPolicy::default();
1105. 1105 let prompt =
1106. 1106 "Fix the off-by-one bug in quicksort.py and verify that the sort works correctly.";
1107. 1107 let final_text = "I fixed the index in quicksort.py.";
1108. 1108
1109. 1109 let receipts = ReceiptSummary {
1110. 1110 total_tool_calls: 1,
1111. 1111 successful_tool_calls: 1,
1112. 1112 files_modified: 1,
1113. 1113 code_files_modified: 1,
1114. 1114 doc_files_modified: 0,
1115. 1115 ..Default::default()
1116. 1116 };
1117. 1117
1118. 1118 // Code was modified and prompt asks to "verify that" -> must block with VerificationMissing
1119. 1119 let blocked = p.evaluate_receipts(prompt, final_text, None, &receipts, false);
1120. 1120 assert_eq!(blocked, Some(BlockReason::VerificationMissing));
1121. 1121
1122. 1122 // If code files were modified after test ran, stale verification blocks
1123. 1123 let with_bash = ReceiptSummary {
1124. 1124 total_tool_calls: 2,
1125. 1125 successful_tool_calls: 2,
1126. 1126 substantive_bash_calls: 1,
1127. 1127 files_modified: 1,
1128. 1128 code_files_modified: 1,
1129. 1129 ..Default::default()
1130. 1130 };
1131. 1131 let blocked_stale = p.evaluate_receipts(prompt, final_text, None, &with_bash, true);
1132. 1132 assert_eq!(blocked_stale, Some(BlockReason::VerificationStale));
1133. 1133 }
1134. 1134}
1135. 1135