@agentv/core 0.7.5 → 0.10.0

package/dist/index.js

@@ -9,7 +9,7 @@ import {
   readTextFile,
   resolveFileReference,
   resolveTargetDefinition
-} from "./chunk-7XM7HYRS.js";
+} from "./chunk-YQBJAT5I.js";
 
 // src/evaluation/types.ts
 var TEST_MESSAGE_ROLE_VALUES = ["system", "user", "assistant", "tool"];
@@ -268,14 +268,11 @@ 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: ${id}`);
       continue;
     }
     if (expectedMessages.length > 1) {
@@ -293,17 +290,17 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       messageType: "input",
       verbose
     });
-    const outputSegments = await processMessages({
+    const outputSegments = hasExpectedMessages ? await processMessages({
       messages: expectedMessages,
       searchRoots,
       repoRootPath,
       guidelinePatterns,
       messageType: "output",
       verbose
-    });
+    }) : [];
     const codeSnippets = extractCodeBlocks(inputSegments);
     const expectedContent = expectedMessages[0]?.content;
-    const referenceAnswer = await resolveAssistantContent(expectedContent, searchRoots, verbose);
+    const referenceAnswer = expectedContent ? await resolveAssistantContent(expectedContent, searchRoots, verbose) : "";
     const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
     const evalCaseEvaluatorKind = coerceEvaluator(evalcase.evaluator, id) ?? globalEvaluator;
     const evaluators = await parseEvaluators(evalcase, searchRoots, id ?? "unknown");
@@ -322,6 +319,7 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       dataset: datasetName,
       conversation_id: conversationId,
       question,
+      input_messages: inputMessages,
       input_segments: inputSegments,
       output_segments: outputSegments,
       reference_answer: referenceAnswer,
@@ -349,6 +347,54 @@ Please add '$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 +411,168 @@ ${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 });
+            }
+          }
+        }
+      }
+    }
+    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");
+  }
+  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 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);
       }
-      continue;
     }
-    const genericValue = segment.value;
-    if (typeof genericValue === "string") {
-      questionParts.push(genericValue);
+    if (contentParts.length > 0) {
+      systemSegments.push(contentParts.join("\n"));
     }
+    startIndex += 1;
   }
-  if (testCase.code_snippets.length > 0) {
-    questionParts.push(testCase.code_snippets.join("\n"));
+  const chatPrompt = [];
+  if (systemSegments.length > 0) {
+    chatPrompt.push({
+      role: "system",
+      content: systemSegments.join("\n\n")
+    });
   }
-  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 };
+  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 {
@@ -591,21 +769,14 @@ 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 +790,21 @@ ${request.guidelines.trim()}`);
   ];
   return prompt;
 }
+function resolveSystemContent(request) {
+  const systemSegments = [];
+  const metadataSystemPrompt = typeof request.metadata?.systemPrompt === "string" ? request.metadata.systemPrompt : void 0;
+  if (metadataSystemPrompt && metadataSystemPrompt.trim().length > 0) {
+    systemSegments.push(metadataSystemPrompt.trim());
+  } else {
+    systemSegments.push(DEFAULT_SYSTEM_PROMPT);
+  }
+  if (request.guidelines && request.guidelines.trim().length > 0) {
+    systemSegments.push(`[[ ## Guidelines ## ]]
+
+${request.guidelines.trim()}`);
+  }
+  return systemSegments.join("\n\n");
+}
 function extractModelConfig(request, defaults) {
   const temperature = request.temperature ?? defaults.temperature;
   const maxTokens = request.maxOutputTokens ?? defaults.maxOutputTokens;
@@ -662,6 +848,67 @@ function ensureChatResponse(result) {
   }
   return result;
 }
+function isRetryableError(error, retryableStatusCodes) {
+  if (!error || typeof error !== "object") {
+    return false;
+  }
+  if ("status" in error && typeof error.status === "number") {
+    return retryableStatusCodes.includes(error.status);
+  }
+  if ("message" in error && typeof error.message === "string") {
+    const match = error.message.match(/HTTP (\d{3})/);
+    if (match) {
+      const status = Number.parseInt(match[1], 10);
+      return retryableStatusCodes.includes(status);
+    }
+  }
+  if ("name" in error && error.name === "AxAIServiceNetworkError") {
+    return true;
+  }
+  return false;
+}
+function calculateRetryDelay(attempt, config) {
+  const delay = Math.min(
+    config.maxDelayMs,
+    config.initialDelayMs * config.backoffFactor ** attempt
+  );
+  return delay * (0.75 + Math.random() * 0.5);
+}
+async function sleep(ms) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+async function withRetry(fn, retryConfig, signal) {
+  const config = {
+    maxRetries: retryConfig?.maxRetries ?? 3,
+    initialDelayMs: retryConfig?.initialDelayMs ?? 1e3,
+    maxDelayMs: retryConfig?.maxDelayMs ?? 6e4,
+    backoffFactor: retryConfig?.backoffFactor ?? 2,
+    retryableStatusCodes: retryConfig?.retryableStatusCodes ?? [500, 408, 429, 502, 503, 504]
+  };
+  let lastError;
+  for (let attempt = 0; attempt <= config.maxRetries; attempt++) {
+    if (signal?.aborted) {
+      throw new Error(`Request aborted: ${signal.reason ?? "Unknown reason"}`);
+    }
+    try {
+      return await fn();
+    } catch (error) {
+      lastError = error;
+      if (attempt >= config.maxRetries) {
+        break;
+      }
+      if (!isRetryableError(error, config.retryableStatusCodes)) {
+        throw error;
+      }
+      const delay = calculateRetryDelay(attempt, config);
+      await sleep(delay);
+      if (signal?.aborted) {
+        throw new Error(`Request aborted: ${signal.reason ?? "Unknown reason"}`);
+      }
+    }
+  }
+  throw lastError;
+}
 var AzureProvider = class {
   constructor(targetName, config) {
     this.config = config;
@@ -671,6 +918,7 @@ var AzureProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = AxAI.create({
       name: "azure-openai",
       apiKey: config.apiKey,
@@ -687,16 +935,21 @@ var AzureProvider = class {
   targetName;
   ai;
   defaults;
+  retryConfig;
   async invoke(request) {
     const chatPrompt = buildChatPrompt(request);
     const modelConfig = extractModelConfig(request, this.defaults);
-    const response = await this.ai.chat(
-      {
-        chatPrompt,
-        model: this.config.deploymentName,
-        ...modelConfig ? { modelConfig } : {}
-      },
-      request.signal ? { abortSignal: request.signal } : void 0
+    const response = await withRetry(
+      async () => await this.ai.chat(
+        {
+          chatPrompt,
+          model: this.config.deploymentName,
+          ...modelConfig ? { modelConfig } : {}
+        },
+        request.signal ? { abortSignal: request.signal } : void 0
+      ),
+      this.retryConfig,
+      request.signal
     );
     return mapResponse(ensureChatResponse(response));
   }
@@ -714,6 +967,7 @@ var AnthropicProvider = class {
       maxOutputTokens: config.maxOutputTokens,
       thinkingBudget: config.thinkingBudget
     };
+    this.retryConfig = config.retry;
     this.ai = AxAI.create({
       name: "anthropic",
       apiKey: config.apiKey
@@ -724,16 +978,21 @@ var AnthropicProvider = class {
   targetName;
   ai;
   defaults;
+  retryConfig;
   async invoke(request) {
     const chatPrompt = buildChatPrompt(request);
     const modelConfig = extractModelConfig(request, this.defaults);
-    const response = await this.ai.chat(
-      {
-        chatPrompt,
-        model: this.config.model,
-        ...modelConfig ? { modelConfig } : {}
-      },
-      request.signal ? { abortSignal: request.signal } : void 0
+    const response = await withRetry(
+      async () => await this.ai.chat(
+        {
+          chatPrompt,
+          model: this.config.model,
+          ...modelConfig ? { modelConfig } : {}
+        },
+        request.signal ? { abortSignal: request.signal } : void 0
+      ),
+      this.retryConfig,
+      request.signal
     );
     return mapResponse(ensureChatResponse(response));
   }
@@ -750,6 +1009,7 @@ var GeminiProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = AxAI.create({
       name: "google-gemini",
       apiKey: config.apiKey
@@ -760,16 +1020,21 @@ var GeminiProvider = class {
   targetName;
   ai;
   defaults;
+  retryConfig;
   async invoke(request) {
     const chatPrompt = buildChatPrompt(request);
     const modelConfig = extractModelConfig(request, this.defaults);
-    const response = await this.ai.chat(
-      {
-        chatPrompt,
-        model: this.config.model,
-        ...modelConfig ? { modelConfig } : {}
-      },
-      request.signal ? { abortSignal: request.signal } : void 0
+    const response = await withRetry(
+      async () => await this.ai.chat(
+        {
+          chatPrompt,
+          model: this.config.model,
+          ...modelConfig ? { modelConfig } : {}
+        },
+        request.signal ? { abortSignal: request.signal } : void 0
+      ),
+      this.retryConfig,
+      request.signal
     );
     return mapResponse(ensureChatResponse(response));
   }
@@ -839,10 +1104,9 @@ var CliProvider = class {
     const outputFilePath = generateOutputFilePath(request.evalCaseId);
     const templateValues = buildTemplateValues(request, this.config, outputFilePath);
     const renderedCommand = renderTemplate(this.config.commandTemplate, templateValues);
-    const env = this.config.env ? { ...process.env, ...this.config.env } : process.env;
     const result = await this.runCommand(renderedCommand, {
       cwd: this.config.cwd,
-      env,
+      env: process.env,
       timeoutMs: this.config.timeoutMs,
       signal: request.signal
     });
@@ -931,10 +1195,9 @@ var CliProvider = class {
         generateOutputFilePath("healthcheck")
       )
     );
-    const env = this.config.env ? { ...process.env, ...this.config.env } : process.env;
     const result = await this.runCommand(renderedCommand, {
       cwd: healthcheck.cwd ?? this.config.cwd,
-      env,
+      env: process.env,
       timeoutMs,
       signal
     });
@@ -2167,20 +2430,13 @@ function assertTargetDefinition(value, index, filePath) {
   }
   const name = value.name;
   const provider = value.provider;
-  const settings = value.settings;
-  const judgeTarget = value.judge_target;
   if (typeof name !== "string" || name.trim().length === 0) {
     throw new Error(`targets.yaml entry at index ${index} in ${filePath} is missing a valid 'name'`);
   }
   if (typeof provider !== "string" || provider.trim().length === 0) {
     throw new Error(`targets.yaml entry '${name}' in ${filePath} is missing a valid 'provider'`);
   }
-  return {
-    name,
-    provider,
-    settings: isRecord(settings) ? settings : void 0,
-    judge_target: typeof judgeTarget === "string" ? judgeTarget : void 0
-  };
+  return value;
 }
 async function fileExists3(filePath) {
   try {
@@ -2260,19 +2516,21 @@ 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 } : {},
@@ -2310,38 +2568,51 @@ 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) {
@@ -2424,6 +2695,9 @@ 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;
@@ -2821,10 +3095,11 @@ async function runEvaluation(options) {
           await onProgress({
             workerId,
             evalId: evalCase.id,
-            status: "completed",
+            status: result.error ? "failed" : "completed",
             startedAt: 0,
             // Not used for completed status
-            completedAt: Date.now()
+            completedAt: Date.now(),
+            error: result.error
           });
         }
         if (onResult) {
@@ -3081,11 +3356,27 @@ 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,
@@ -3099,7 +3390,8 @@ 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
   };
@@ -3328,6 +3620,7 @@ 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,
@@ -3344,12 +3637,30 @@ 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,
@@ -3362,7 +3673,9 @@ 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
   };
 }
 function createCacheKey(provider, target, evalCase, promptInputs) {
@@ -3373,6 +3686,9 @@ 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) {