@agentv/core 0.9.0 → 0.10.1

package/dist/index.js

@@ -9,7 +9,7 @@ import {
   readTextFile,
   resolveFileReference,
   resolveTargetDefinition
-} from "./chunk-SNTZFB24.js";
+} from "./chunk-YQBJAT5I.js";
 
 // src/evaluation/types.ts
 var TEST_MESSAGE_ROLE_VALUES = ["system", "user", "assistant", "tool"];
@@ -73,6 +73,33 @@ var ANSI_YELLOW = "\x1B[33m";
 var ANSI_RESET = "\x1B[0m";
 var SCHEMA_EVAL_V2 = "agentv-eval-v2";
 var SCHEMA_CONFIG_V2 = "agentv-config-v2";
+async function readTestSuiteMetadata(testFilePath) {
+  try {
+    const absolutePath = path.resolve(testFilePath);
+    const content = await readFile(absolutePath, "utf8");
+    const parsed = parse(content);
+    if (!isJsonObject(parsed)) {
+      return {};
+    }
+    return { target: extractTargetFromSuite(parsed) };
+  } catch {
+    return {};
+  }
+}
+function extractTargetFromSuite(suite) {
+  const execution = suite.execution;
+  if (execution && typeof execution === "object" && !Array.isArray(execution)) {
+    const executionTarget = execution.target;
+    if (typeof executionTarget === "string" && executionTarget.trim().length > 0) {
+      return executionTarget.trim();
+    }
+  }
+  const targetValue = suite.target;
+  if (typeof targetValue === "string" && targetValue.trim().length > 0) {
+    return targetValue.trim();
+  }
+  return void 0;
+}
 async function loadConfig(evalFilePath, repoRoot) {
   const directories = buildDirectoryChain(evalFilePath, repoRoot);
   for (const directory of directories) {
@@ -249,6 +276,8 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
     throw new Error(`Invalid test file format: ${evalFilePath} - missing 'evalcases' field`);
   }
   const globalEvaluator = coerceEvaluator(suite.evaluator, "global") ?? "llm_judge";
+  const globalExecution = isJsonObject(suite.execution) ? suite.execution : void 0;
+  const globalTarget = asString(globalExecution?.target) ?? asString(suite.target);
   const results = [];
   for (const rawEvalcase of rawTestcases) {
     if (!isJsonObject(rawEvalcase)) {
@@ -268,14 +297,11 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       logWarning(`Skipping incomplete eval case: ${id ?? "unknown"}`);
       continue;
     }
-    if (!Array.isArray(expectedMessagesValue)) {
-      logWarning(`Eval case '${id}' missing expected_messages array`);
-      continue;
-    }
+    const hasExpectedMessages = Array.isArray(expectedMessagesValue) && expectedMessagesValue.length > 0;
     const inputMessages = inputMessagesValue.filter((msg) => isTestMessage(msg));
-    const expectedMessages = expectedMessagesValue.filter((msg) => isTestMessage(msg));
-    if (expectedMessages.length === 0) {
-      logWarning(`No expected message found for eval case: ${id}`);
+    const expectedMessages = hasExpectedMessages ? expectedMessagesValue.filter((msg) => isTestMessage(msg)) : [];
+    if (hasExpectedMessages && expectedMessages.length === 0) {
+      logWarning(`No valid expected message found for eval case: ${id}`);
       continue;
     }
     if (expectedMessages.length > 1) {
@@ -293,20 +319,20 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       messageType: "input",
       verbose
     });
-    const outputSegments = await processMessages({
+    const outputSegments = hasExpectedMessages ? await processMessages({
       messages: expectedMessages,
       searchRoots,
       repoRootPath,
       guidelinePatterns,
       messageType: "output",
       verbose
-    });
+    }) : [];
     const codeSnippets = extractCodeBlocks(inputSegments);
     const expectedContent = expectedMessages[0]?.content;
-    const referenceAnswer = await resolveAssistantContent(expectedContent, searchRoots, verbose);
+    const referenceAnswer = expectedContent ? await resolveAssistantContent(expectedContent, searchRoots, verbose) : "";
     const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
     const evalCaseEvaluatorKind = coerceEvaluator(evalcase.evaluator, id) ?? globalEvaluator;
-    const evaluators = await parseEvaluators(evalcase, searchRoots, id ?? "unknown");
+    const evaluators = await parseEvaluators(evalcase, globalExecution, searchRoots, id ?? "unknown");
     const userFilePaths = [];
     for (const segment of inputSegments) {
       if (segment.type === "file" && typeof segment.resolvedPath === "string") {
@@ -322,6 +348,7 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       dataset: datasetName,
       conversation_id: conversationId,
       question,
+      input_messages: inputMessages,
       input_segments: inputSegments,
       output_segments: outputSegments,
       reference_answer: referenceAnswer,
@@ -349,6 +376,54 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
   }
   return results;
 }
+function needsRoleMarkers(messages, processedSegmentsByMessage) {
+  if (messages.some((msg) => msg.role === "assistant" || msg.role === "tool")) {
+    return true;
+  }
+  let messagesWithContent = 0;
+  for (const segments of processedSegmentsByMessage) {
+    if (hasVisibleContent(segments)) {
+      messagesWithContent++;
+    }
+  }
+  return messagesWithContent > 1;
+}
+function hasVisibleContent(segments) {
+  return segments.some((segment) => {
+    const type = asString(segment.type);
+    if (type === "text") {
+      const value = asString(segment.value);
+      return value !== void 0 && value.trim().length > 0;
+    }
+    if (type === "guideline_ref") {
+      return false;
+    }
+    if (type === "file") {
+      const text = asString(segment.text);
+      return text !== void 0 && text.trim().length > 0;
+    }
+    return false;
+  });
+}
+function formatSegment(segment) {
+  const type = asString(segment.type);
+  if (type === "text") {
+    return asString(segment.value);
+  }
+  if (type === "guideline_ref") {
+    const refPath = asString(segment.path);
+    return refPath ? `<Attached: ${refPath}>` : void 0;
+  }
+  if (type === "file") {
+    const text = asString(segment.text);
+    const filePath = asString(segment.path);
+    if (text && filePath) {
+      return `=== ${filePath} ===
+${text}`;
+    }
+  }
+  return void 0;
+}
 async function buildPromptInputs(testCase) {
   const guidelineContents = [];
   for (const rawPath of testCase.guideline_paths) {
@@ -365,36 +440,168 @@ ${content}`);
       logWarning(`Could not read guideline file ${absolutePath}: ${error.message}`);
     }
   }
-  const questionParts = [];
+  const guidelines = guidelineContents.map((part) => part.trim()).filter((part) => part.length > 0).join("\n\n");
+  const segmentsByMessage = [];
+  const fileContentsByPath = /* @__PURE__ */ new Map();
   for (const segment of testCase.input_segments) {
-    const typeValue = segment.type;
-    if (typeof typeValue === "string" && typeValue === "file") {
-      const pathValue = segment.path;
-      const textValue = segment.text;
-      const label = typeof pathValue === "string" ? pathValue : "file";
-      const body = typeof textValue === "string" ? textValue : "";
-      questionParts.push(`=== ${label} ===
-${body}`);
-      continue;
+    if (segment.type === "file" && typeof segment.path === "string" && typeof segment.text === "string") {
+      fileContentsByPath.set(segment.path, segment.text);
     }
-    if (typeof typeValue === "string" && typeValue === "text") {
-      const value = segment.value;
-      if (typeof value === "string") {
-        questionParts.push(value);
+  }
+  for (const message of testCase.input_messages) {
+    const messageSegments = [];
+    if (typeof message.content === "string") {
+      if (message.content.trim().length > 0) {
+        messageSegments.push({ type: "text", value: message.content });
+      }
+    } else if (Array.isArray(message.content)) {
+      for (const segment of message.content) {
+        if (typeof segment === "string") {
+          if (segment.trim().length > 0) {
+            messageSegments.push({ type: "text", value: segment });
+          }
+        } else if (isJsonObject(segment)) {
+          const type = asString(segment.type);
+          if (type === "file") {
+            const value = asString(segment.value);
+            if (!value) continue;
+            if (testCase.guideline_patterns && isGuidelineFile(value, testCase.guideline_patterns)) {
+              messageSegments.push({ type: "guideline_ref", path: value });
+              continue;
+            }
+            const fileText = fileContentsByPath.get(value);
+            if (fileText !== void 0) {
+              messageSegments.push({ type: "file", text: fileText, path: value });
+            }
+          } else if (type === "text") {
+            const textValue = asString(segment.value);
+            if (textValue && textValue.trim().length > 0) {
+              messageSegments.push({ type: "text", value: textValue });
+            }
+          }
+        }
       }
-      continue;
     }
-    const genericValue = segment.value;
-    if (typeof genericValue === "string") {
-      questionParts.push(genericValue);
+    segmentsByMessage.push(messageSegments);
+  }
+  const useRoleMarkers = needsRoleMarkers(testCase.input_messages, segmentsByMessage);
+  let question;
+  if (useRoleMarkers) {
+    const messageParts = [];
+    for (let i = 0; i < testCase.input_messages.length; i++) {
+      const message = testCase.input_messages[i];
+      const segments = segmentsByMessage[i];
+      if (!hasVisibleContent(segments)) {
+        continue;
+      }
+      const roleLabel = message.role.charAt(0).toUpperCase() + message.role.slice(1);
+      const contentParts = [];
+      for (const segment of segments) {
+        const formattedContent = formatSegment(segment);
+        if (formattedContent) {
+          contentParts.push(formattedContent);
+        }
+      }
+      if (contentParts.length > 0) {
+        const messageContent = contentParts.join("\n");
+        messageParts.push(`@[${roleLabel}]:
+${messageContent}`);
+      }
+    }
+    question = messageParts.join("\n\n");
+  } else {
+    const questionParts = [];
+    for (const segment of testCase.input_segments) {
+      const formattedContent = formatSegment(segment);
+      if (formattedContent) {
+        questionParts.push(formattedContent);
+      }
     }
+    if (testCase.code_snippets.length > 0) {
+      questionParts.push(testCase.code_snippets.join("\n"));
+    }
+    question = questionParts.map((part) => part.trim()).filter((part) => part.length > 0).join("\n\n");
   }
-  if (testCase.code_snippets.length > 0) {
-    questionParts.push(testCase.code_snippets.join("\n"));
+  const chatPrompt = useRoleMarkers ? buildChatPromptFromSegments({
+    messages: testCase.input_messages,
+    segmentsByMessage,
+    guidelinePatterns: testCase.guideline_patterns,
+    guidelineContent: guidelines
+  }) : void 0;
+  return { question, guidelines, chatPrompt };
+}
+function buildChatPromptFromSegments(options) {
+  const { messages, segmentsByMessage, guidelinePatterns, guidelineContent, systemPrompt } = options;
+  if (messages.length === 0) {
+    return void 0;
   }
-  const question = questionParts.map((part) => part.trim()).filter((part) => part.length > 0).join("\n\n");
-  const guidelines = guidelineContents.map((part) => part.trim()).filter((part) => part.length > 0).join("\n\n");
-  return { question, guidelines };
+  const systemSegments = [];
+  if (systemPrompt && systemPrompt.trim().length > 0) {
+    systemSegments.push(systemPrompt.trim());
+  }
+  if (guidelineContent && guidelineContent.trim().length > 0) {
+    systemSegments.push(`[[ ## Guidelines ## ]]
+
+${guidelineContent.trim()}`);
+  }
+  let startIndex = 0;
+  while (startIndex < messages.length && messages[startIndex].role === "system") {
+    const segments = segmentsByMessage[startIndex];
+    const contentParts = [];
+    for (const segment of segments) {
+      const formatted = formatSegment(segment);
+      if (formatted) {
+        contentParts.push(formatted);
+      }
+    }
+    if (contentParts.length > 0) {
+      systemSegments.push(contentParts.join("\n"));
+    }
+    startIndex += 1;
+  }
+  const chatPrompt = [];
+  if (systemSegments.length > 0) {
+    chatPrompt.push({
+      role: "system",
+      content: systemSegments.join("\n\n")
+    });
+  }
+  for (let i = startIndex; i < messages.length; i++) {
+    const message = messages[i];
+    const segments = segmentsByMessage[i];
+    const contentParts = [];
+    let role = message.role;
+    let name;
+    if (role === "system") {
+      role = "assistant";
+      contentParts.push("@[System]:");
+    } else if (role === "tool") {
+      role = "function";
+      name = "tool";
+    }
+    for (const segment of segments) {
+      if (segment.type === "guideline_ref") {
+        continue;
+      }
+      const formatted = formatSegment(segment);
+      if (formatted) {
+        const isGuidelineRef = segment.type === "file" && typeof segment.path === "string" && guidelinePatterns && isGuidelineFile(segment.path, guidelinePatterns);
+        if (isGuidelineRef) {
+          continue;
+        }
+        contentParts.push(formatted);
+      }
+    }
+    if (contentParts.length === 0) {
+      continue;
+    }
+    chatPrompt.push({
+      role,
+      content: contentParts.join("\n"),
+      ...name ? { name } : {}
+    });
+  }
+  return chatPrompt.length > 0 ? chatPrompt : void 0;
 }
 async function fileExists2(absolutePath) {
   try {
@@ -492,9 +699,9 @@ async function resolveAssistantContent(content, searchRoots, verbose) {
   }
   return parts.join(" ");
 }
-async function parseEvaluators(rawEvalCase, searchRoots, evalId) {
+async function parseEvaluators(rawEvalCase, globalExecution, searchRoots, evalId) {
   const execution = rawEvalCase.execution;
-  const candidateEvaluators = isJsonObject(execution) ? execution.evaluators ?? rawEvalCase.evaluators : rawEvalCase.evaluators;
+  const candidateEvaluators = isJsonObject(execution) ? execution.evaluators ?? rawEvalCase.evaluators : rawEvalCase.evaluators ?? globalExecution?.evaluators;
   if (candidateEvaluators === void 0) {
     return void 0;
   }
@@ -532,6 +739,8 @@ async function parseEvaluators(rawEvalCase, searchRoots, evalId) {
             resolved.attempted.length > 0 ? resolved.attempted.map((attempt) => `  Tried: ${attempt}`) : void 0
           );
         }
+      } else {
+        resolvedCwd = searchRoots[0];
       }
       evaluators.push({
         name,
@@ -560,8 +769,7 @@ async function parseEvaluators(rawEvalCase, searchRoots, evalId) {
       name,
       type: "llm_judge",
       prompt,
-      promptPath,
-      model
+      promptPath
     });
   }
   return evaluators.length > 0 ? evaluators : void 0;
@@ -591,21 +799,14 @@ import { AxAI } from "@ax-llm/ax";
 var DEFAULT_SYSTEM_PROMPT = "You are a careful assistant. Follow all provided instructions and do not fabricate results.";
 function buildChatPrompt(request) {
   if (request.chatPrompt) {
-    return request.chatPrompt;
-  }
-  const systemSegments = [];
-  const metadataSystemPrompt = typeof request.metadata?.systemPrompt === "string" ? request.metadata.systemPrompt : void 0;
-  if (metadataSystemPrompt && metadataSystemPrompt.trim().length > 0) {
-    systemSegments.push(metadataSystemPrompt.trim());
-  } else {
-    systemSegments.push(DEFAULT_SYSTEM_PROMPT);
-  }
-  if (request.guidelines && request.guidelines.trim().length > 0) {
-    systemSegments.push(`[[ ## Guidelines ## ]]
-
-${request.guidelines.trim()}`);
+    const hasSystemMessage = request.chatPrompt.some((message) => message.role === "system");
+    if (hasSystemMessage) {
+      return request.chatPrompt;
+    }
+    const systemContent2 = resolveSystemContent(request);
+    return [{ role: "system", content: systemContent2 }, ...request.chatPrompt];
   }
-  const systemContent = systemSegments.join("\n\n");
+  const systemContent = resolveSystemContent(request);
   const userContent = request.question.trim();
   const prompt = [
     {
@@ -619,6 +820,21 @@ ${request.guidelines.trim()}`);
   ];
   return prompt;
 }
+function resolveSystemContent(request) {
+  const systemSegments = [];
+  const metadataSystemPrompt = typeof request.metadata?.systemPrompt === "string" ? request.metadata.systemPrompt : void 0;
+  if (metadataSystemPrompt && metadataSystemPrompt.trim().length > 0) {
+    systemSegments.push(metadataSystemPrompt.trim());
+  } else {
+    systemSegments.push(DEFAULT_SYSTEM_PROMPT);
+  }
+  if (request.guidelines && request.guidelines.trim().length > 0) {
+    systemSegments.push(`[[ ## Guidelines ## ]]
+
+${request.guidelines.trim()}`);
+  }
+  return systemSegments.join("\n\n");
+}
 function extractModelConfig(request, defaults) {
   const temperature = request.temperature ?? defaults.temperature;
   const maxTokens = request.maxOutputTokens ?? defaults.maxOutputTokens;
@@ -2330,24 +2546,23 @@ var LlmJudgeEvaluator = class {
     return this.evaluateWithPrompt(context, judgeProvider);
   }
   async evaluateWithPrompt(context, judgeProvider) {
-    let prompt = buildQualityPrompt(context.evalCase, context.candidate);
-    let systemPrompt = context.systemPrompt ?? this.customPrompt ?? QUALITY_SYSTEM_PROMPT;
+    const hasReferenceAnswer = hasNonEmptyReferenceAnswer(context.evalCase);
+    const formattedQuestion = context.promptInputs.question && context.promptInputs.question.trim().length > 0 ? context.promptInputs.question : context.evalCase.question;
+    let prompt = buildQualityPrompt(context.evalCase, context.candidate, formattedQuestion);
+    let systemPrompt = context.systemPrompt ?? this.customPrompt ?? buildSystemPrompt(hasReferenceAnswer);
     if (systemPrompt && hasTemplateVariables(systemPrompt)) {
       const variables = {
         input_messages: JSON.stringify(context.evalCase.input_segments, null, 2),
         output_messages: JSON.stringify(context.evalCase.output_segments, null, 2),
         candidate_answer: context.candidate,
-        reference_answer: context.evalCase.reference_answer,
+        reference_answer: context.evalCase.reference_answer ?? "",
         expected_outcome: context.evalCase.expected_outcome,
-        question: context.evalCase.question
+        question: formattedQuestion
       };
       prompt = substituteVariables(systemPrompt, variables);
-      systemPrompt = QUALITY_SYSTEM_PROMPT;
+      systemPrompt = buildSystemPrompt(hasReferenceAnswer);
     }
-    const metadata = {
-      ...systemPrompt !== void 0 ? { systemPrompt } : {},
-      ...context.judgeModel !== void 0 ? { model: context.judgeModel } : {}
-    };
+    const metadata = systemPrompt !== void 0 ? { systemPrompt } : {};
     const response = await judgeProvider.invoke({
       question: prompt,
       metadata,
@@ -2367,8 +2582,7 @@ var LlmJudgeEvaluator = class {
       provider: judgeProvider.id,
       prompt,
       target: context.target.name,
-      ...systemPrompt !== void 0 ? { systemPrompt } : {},
-      ...context.judgeModel !== void 0 ? { model: context.judgeModel } : {}
+      ...systemPrompt !== void 0 && { systemPrompt }
     };
     return {
       score,
@@ -2380,38 +2594,51 @@ var LlmJudgeEvaluator = class {
     };
   }
 };
-var QUALITY_SYSTEM_PROMPT = [
-  "You are an expert evaluator. Your goal is to grade the candidate_answer based on how well it achieves the expected_outcome for the original task.",
-  "",
-  "Use the reference_answer as a gold standard for a high-quality response. The candidate_answer does not need to match it verbatim, but it should capture the key points and follow the same spirit.",
-  "",
-  "Be concise and focused in your evaluation. Provide succinct, specific feedback rather than verbose explanations.",
-  "",
-  "You must respond with a single JSON object matching this schema:",
-  "",
-  "{",
-  '  "score": <number between 0.0 and 1.0>,',
-  '  "hits": [<array of strings, max 4 items, brief specific achievements>],',
-  '  "misses": [<array of strings, max 4 items, brief specific failures or omissions, empty if none>],',
-  '  "reasoning": "<string, concise explanation for the score, 1-2 sentences max>"',
-  "}"
-].join("\n");
-function buildQualityPrompt(evalCase, candidate) {
+function buildSystemPrompt(hasReferenceAnswer) {
+  const basePrompt = [
+    "You are an expert evaluator. Your goal is to grade the candidate_answer based on how well it achieves the expected_outcome for the original task.",
+    ""
+  ];
+  if (hasReferenceAnswer) {
+    basePrompt.push(
+      "Use the reference_answer as a gold standard for a high-quality response. The candidate_answer does not need to match it verbatim, but should capture the key points and follow the same spirit.",
+      ""
+    );
+  }
+  basePrompt.push(
+    "Be concise and focused in your evaluation. Provide succinct, specific feedback rather than verbose explanations.",
+    "",
+    "You must respond with a single JSON object matching this schema:",
+    "",
+    "{",
+    '  "score": <number between 0.0 and 1.0>,',
+    '  "hits": [<array of strings, max 4 items, brief specific achievements>],',
+    '  "misses": [<array of strings, max 4 items, brief specific failures or omissions, empty if none>],',
+    '  "reasoning": "<string, concise explanation for the score, 1-2 sentences max>"',
+    "}"
+  );
+  return basePrompt.join("\n");
+}
+function buildQualityPrompt(evalCase, candidate, question) {
   const parts = [
     "[[ ## expected_outcome ## ]]",
     evalCase.expected_outcome.trim(),
     "",
     "[[ ## question ## ]]",
-    evalCase.question.trim(),
-    "",
-    "[[ ## reference_answer ## ]]",
-    evalCase.reference_answer.trim(),
-    "",
-    "[[ ## candidate_answer ## ]]",
-    candidate.trim(),
-    "",
-    "Respond with a single JSON object matching the schema described in the system prompt."
+    question.trim(),
+    ""
   ];
+  if (hasNonEmptyReferenceAnswer(evalCase)) {
+    parts.push(
+      "[[ ## reference_answer ## ]]",
+      evalCase.reference_answer.trim(),
+      ""
+    );
+  }
+  parts.push(
+    "[[ ## candidate_answer ## ]]",
+    candidate.trim()
+  );
   return parts.join("\n");
 }
 function clampScore(value) {
@@ -2494,6 +2721,9 @@ function extractJsonBlob(text) {
 function isNonEmptyString(value) {
   return typeof value === "string" && value.trim().length > 0;
 }
+function hasNonEmptyReferenceAnswer(evalCase) {
+  return evalCase.reference_answer !== void 0 && evalCase.reference_answer.trim().length > 0;
+}
 var CodeEvaluator = class {
   kind = "code";
   script;
@@ -3152,11 +3382,27 @@ async function evaluateCandidate(options) {
     agentTimeoutMs
   });
   const completedAt = nowFn();
-  const rawRequest = {
-    question: promptInputs.question,
-    ...isAgentProvider(provider) ? {} : { guidelines: promptInputs.guidelines },
-    guideline_paths: evalCase.guideline_paths
-  };
+  let agentProviderRequest;
+  let lmProviderRequest;
+  if (isAgentProvider(provider)) {
+    agentProviderRequest = {
+      question: promptInputs.question,
+      guideline_paths: evalCase.guideline_paths
+    };
+  } else {
+    if (promptInputs.chatPrompt) {
+      lmProviderRequest = {
+        chat_prompt: promptInputs.chatPrompt,
+        guideline_paths: evalCase.guideline_paths
+      };
+    } else {
+      lmProviderRequest = {
+        question: promptInputs.question,
+        guidelines: promptInputs.guidelines,
+        guideline_paths: evalCase.guideline_paths
+      };
+    }
+  }
   return {
     eval_id: evalCase.id,
     dataset: evalCase.dataset,
@@ -3170,7 +3416,8 @@ async function evaluateCandidate(options) {
     timestamp: completedAt.toISOString(),
     reasoning: score.reasoning,
     raw_aspects: score.rawAspects,
-    raw_request: rawRequest,
+    agent_provider_request: agentProviderRequest,
+    lm_provider_request: lmProviderRequest,
     evaluator_raw_request: evaluatorResults ? void 0 : score.evaluatorRawRequest,
     evaluator_results: evaluatorResults
   };
@@ -3329,8 +3576,7 @@ async function runLlmJudgeEvaluator(options) {
     now,
     judgeProvider,
     systemPrompt: customPrompt,
-    evaluator: config,
-    judgeModel: config.model
+    evaluator: config
   });
 }
 async function resolveCustomPrompt(config) {
@@ -3399,6 +3645,7 @@ async function invokeProvider(provider, options) {
       question: promptInputs.question,
       guidelines: promptInputs.guidelines,
       guideline_patterns: evalCase.guideline_patterns,
+      chatPrompt: promptInputs.chatPrompt,
       inputFiles: evalCase.file_paths,
       evalCaseId: evalCase.id,
       attempt,
@@ -3415,12 +3662,30 @@ async function invokeProvider(provider, options) {
 }
 function buildErrorResult(evalCase, targetName, timestamp, error, promptInputs, provider) {
   const message = error instanceof Error ? error.message : String(error);
-  const rawRequest = {
-    question: promptInputs.question,
-    ...isAgentProvider(provider) ? {} : { guidelines: promptInputs.guidelines },
-    guideline_paths: evalCase.guideline_paths,
-    error: message
-  };
+  let agentProviderRequest;
+  let lmProviderRequest;
+  if (isAgentProvider(provider)) {
+    agentProviderRequest = {
+      question: promptInputs.question,
+      guideline_paths: evalCase.guideline_paths,
+      error: message
+    };
+  } else {
+    if (promptInputs.chatPrompt) {
+      lmProviderRequest = {
+        chat_prompt: promptInputs.chatPrompt,
+        guideline_paths: evalCase.guideline_paths,
+        error: message
+      };
+    } else {
+      lmProviderRequest = {
+        question: promptInputs.question,
+        guidelines: promptInputs.guidelines,
+        guideline_paths: evalCase.guideline_paths,
+        error: message
+      };
+    }
+  }
   return {
     eval_id: evalCase.id,
     dataset: evalCase.dataset,
@@ -3433,7 +3698,8 @@ function buildErrorResult(evalCase, targetName, timestamp, error, promptInputs,
     target: targetName,
     timestamp: timestamp.toISOString(),
     raw_aspects: [],
-    raw_request: rawRequest,
+    agent_provider_request: agentProviderRequest,
+    lm_provider_request: lmProviderRequest,
     error: message
   };
 }
@@ -3445,6 +3711,9 @@ function createCacheKey(provider, target, evalCase, promptInputs) {
   hash.update(promptInputs.question);
   hash.update(promptInputs.guidelines);
   hash.update(promptInputs.systemMessage ?? "");
+  if (promptInputs.chatPrompt) {
+    hash.update(JSON.stringify(promptInputs.chatPrompt));
+  }
   return hash.digest("hex");
 }
 function isTimeoutLike(error) {
@@ -3492,6 +3761,7 @@ export {
   loadEvalCases,
   normalizeLineEndings,
   readTargetDefinitions,
+  readTestSuiteMetadata,
   readTextFile,
   resolveAndCreateProvider,
   resolveFileReference,