@agentv/core 0.7.5 → 0.10.0

package/dist/index.cjs

@@ -434,14 +434,11 @@ 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: ${id}`);
       continue;
     }
     if (expectedMessages.length > 1) {
@@ -459,17 +456,17 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       messageType: "input",
       verbose
     });
-    const outputSegments = await processMessages({
+    const outputSegments = hasExpectedMessages ? await processMessages({
       messages: expectedMessages,
       searchRoots,
       repoRootPath,
       guidelinePatterns,
       messageType: "output",
       verbose
-    });
+    }) : [];
     const codeSnippets = extractCodeBlocks(inputSegments);
     const expectedContent = expectedMessages[0]?.content;
-    const referenceAnswer = await resolveAssistantContent(expectedContent, searchRoots, verbose);
+    const referenceAnswer = expectedContent ? await resolveAssistantContent(expectedContent, searchRoots, verbose) : "";
     const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
     const evalCaseEvaluatorKind = coerceEvaluator(evalcase.evaluator, id) ?? globalEvaluator;
     const evaluators = await parseEvaluators(evalcase, searchRoots, id ?? "unknown");
@@ -488,6 +485,7 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       dataset: datasetName,
       conversation_id: conversationId,
       question,
+      input_messages: inputMessages,
       input_segments: inputSegments,
       output_segments: outputSegments,
       reference_answer: referenceAnswer,
@@ -515,6 +513,54 @@ Please add '$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) {
@@ -531,36 +577,168 @@ ${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}`);
       }
-      continue;
     }
-    const genericValue = segment.value;
-    if (typeof genericValue === "string") {
-      questionParts.push(genericValue);
+    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 {
@@ -757,21 +935,14 @@ var import_ax = require("@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 = [
     {
@@ -785,6 +956,21 @@ ${request.guidelines.trim()}`);
   ];
   return prompt;
 }
+function resolveSystemContent(request) {
+  const systemSegments = [];
+  const metadataSystemPrompt = typeof request.metadata?.systemPrompt === "string" ? request.metadata.systemPrompt : void 0;
+  if (metadataSystemPrompt && metadataSystemPrompt.trim().length > 0) {
+    systemSegments.push(metadataSystemPrompt.trim());
+  } else {
+    systemSegments.push(DEFAULT_SYSTEM_PROMPT);
+  }
+  if (request.guidelines && request.guidelines.trim().length > 0) {
+    systemSegments.push(`[[ ## Guidelines ## ]]
+
+${request.guidelines.trim()}`);
+  }
+  return systemSegments.join("\n\n");
+}
 function extractModelConfig(request, defaults) {
   const temperature = request.temperature ?? defaults.temperature;
   const maxTokens = request.maxOutputTokens ?? defaults.maxOutputTokens;
@@ -828,6 +1014,67 @@ 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;
@@ -837,6 +1084,7 @@ var AzureProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = import_ax.AxAI.create({
       name: "azure-openai",
       apiKey: config.apiKey,
@@ -853,16 +1101,21 @@ 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));
   }
@@ -880,6 +1133,7 @@ var AnthropicProvider = class {
       maxOutputTokens: config.maxOutputTokens,
       thinkingBudget: config.thinkingBudget
     };
+    this.retryConfig = config.retry;
     this.ai = import_ax.AxAI.create({
       name: "anthropic",
       apiKey: config.apiKey
@@ -890,16 +1144,21 @@ 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));
   }
@@ -916,6 +1175,7 @@ var GeminiProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = import_ax.AxAI.create({
       name: "google-gemini",
       apiKey: config.apiKey
@@ -926,16 +1186,21 @@ 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));
   }
@@ -1005,10 +1270,9 @@ 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
     });
@@ -1097,10 +1361,9 @@ 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
     });
@@ -2051,10 +2314,9 @@ var CLI_PLACEHOLDERS = /* @__PURE__ */ new Set(["PROMPT", "GUIDELINES", "EVAL_ID
 var BASE_TARGET_SCHEMA = import_zod.z.object({
   name: import_zod.z.string().min(1, "target name is required"),
   provider: import_zod.z.string().min(1, "provider is required"),
-  settings: import_zod.z.record(import_zod.z.unknown()).optional(),
   judge_target: import_zod.z.string().optional(),
   workers: import_zod.z.number().int().min(1).optional()
-});
+}).passthrough();
 var DEFAULT_AZURE_API_VERSION = "2024-10-01-preview";
 function normalizeAzureApiVersion(value) {
   if (!value) {
@@ -2067,11 +2329,43 @@ function normalizeAzureApiVersion(value) {
   const withoutPrefix = trimmed.replace(/^api[-_]?version\s*=\s*/i, "").trim();
   return withoutPrefix.length > 0 ? withoutPrefix : DEFAULT_AZURE_API_VERSION;
 }
+function resolveRetryConfig(target) {
+  const maxRetries = resolveOptionalNumber(
+    target.max_retries ?? target.maxRetries,
+    `${target.name} max retries`
+  );
+  const initialDelayMs = resolveOptionalNumber(
+    target.retry_initial_delay_ms ?? target.retryInitialDelayMs,
+    `${target.name} retry initial delay`
+  );
+  const maxDelayMs = resolveOptionalNumber(
+    target.retry_max_delay_ms ?? target.retryMaxDelayMs,
+    `${target.name} retry max delay`
+  );
+  const backoffFactor = resolveOptionalNumber(
+    target.retry_backoff_factor ?? target.retryBackoffFactor,
+    `${target.name} retry backoff factor`
+  );
+  const retryableStatusCodes = resolveOptionalNumberArray(
+    target.retry_status_codes ?? target.retryStatusCodes,
+    `${target.name} retry status codes`
+  );
+  if (maxRetries === void 0 && initialDelayMs === void 0 && maxDelayMs === void 0 && backoffFactor === void 0 && retryableStatusCodes === void 0) {
+    return void 0;
+  }
+  return {
+    maxRetries,
+    initialDelayMs,
+    maxDelayMs,
+    backoffFactor,
+    retryableStatusCodes
+  };
+}
 function resolveTargetDefinition(definition, env = process.env) {
   const parsed = BASE_TARGET_SCHEMA.parse(definition);
   const provider = parsed.provider.toLowerCase();
   const providerBatching = resolveOptionalBoolean(
-    parsed.settings?.provider_batching ?? parsed.settings?.providerBatching
+    parsed.provider_batching ?? parsed.providerBatching
   );
   switch (provider) {
     case "azure":
@@ -2147,13 +2441,12 @@ function resolveTargetDefinition(definition, env = process.env) {
   }
 }
 function resolveAzureConfig(target, env) {
-  const settings = target.settings ?? {};
-  const endpointSource = settings.endpoint ?? settings.resource ?? settings.resourceName;
-  const apiKeySource = settings.api_key ?? settings.apiKey;
-  const deploymentSource = settings.deployment ?? settings.deploymentName ?? settings.model;
-  const versionSource = settings.version ?? settings.api_version;
-  const temperatureSource = settings.temperature;
-  const maxTokensSource = settings.max_output_tokens ?? settings.maxTokens;
+  const endpointSource = target.endpoint ?? target.resource ?? target.resourceName;
+  const apiKeySource = target.api_key ?? target.apiKey;
+  const deploymentSource = target.deployment ?? target.deploymentName ?? target.model;
+  const versionSource = target.version ?? target.api_version;
+  const temperatureSource = target.temperature;
+  const maxTokensSource = target.max_output_tokens ?? target.maxTokens;
   const resourceName = resolveString(endpointSource, env, `${target.name} endpoint`);
   const apiKey = resolveString(apiKeySource, env, `${target.name} api key`);
   const deploymentName = resolveString(deploymentSource, env, `${target.name} deployment`);
@@ -2165,58 +2458,61 @@ function resolveAzureConfig(target, env) {
     maxTokensSource,
     `${target.name} max output tokens`
   );
+  const retry = resolveRetryConfig(target);
   return {
     resourceName,
     deploymentName,
     apiKey,
     version,
     temperature,
-    maxOutputTokens
+    maxOutputTokens,
+    retry
   };
 }
 function resolveAnthropicConfig(target, env) {
-  const settings = target.settings ?? {};
-  const apiKeySource = settings.api_key ?? settings.apiKey;
-  const modelSource = settings.model ?? settings.deployment ?? settings.variant;
-  const temperatureSource = settings.temperature;
-  const maxTokensSource = settings.max_output_tokens ?? settings.maxTokens;
-  const thinkingBudgetSource = settings.thinking_budget ?? settings.thinkingBudget;
+  const apiKeySource = target.api_key ?? target.apiKey;
+  const modelSource = target.model ?? target.deployment ?? target.variant;
+  const temperatureSource = target.temperature;
+  const maxTokensSource = target.max_output_tokens ?? target.maxTokens;
+  const thinkingBudgetSource = target.thinking_budget ?? target.thinkingBudget;
   const apiKey = resolveString(apiKeySource, env, `${target.name} Anthropic api key`);
   const model = resolveString(modelSource, env, `${target.name} Anthropic model`);
+  const retry = resolveRetryConfig(target);
   return {
     apiKey,
     model,
     temperature: resolveOptionalNumber(temperatureSource, `${target.name} temperature`),
     maxOutputTokens: resolveOptionalNumber(maxTokensSource, `${target.name} max output tokens`),
-    thinkingBudget: resolveOptionalNumber(thinkingBudgetSource, `${target.name} thinking budget`)
+    thinkingBudget: resolveOptionalNumber(thinkingBudgetSource, `${target.name} thinking budget`),
+    retry
   };
 }
 function resolveGeminiConfig(target, env) {
-  const settings = target.settings ?? {};
-  const apiKeySource = settings.api_key ?? settings.apiKey;
-  const modelSource = settings.model ?? settings.deployment ?? settings.variant;
-  const temperatureSource = settings.temperature;
-  const maxTokensSource = settings.max_output_tokens ?? settings.maxTokens;
+  const apiKeySource = target.api_key ?? target.apiKey;
+  const modelSource = target.model ?? target.deployment ?? target.variant;
+  const temperatureSource = target.temperature;
+  const maxTokensSource = target.max_output_tokens ?? target.maxTokens;
   const apiKey = resolveString(apiKeySource, env, `${target.name} Google API key`);
   const model = resolveOptionalString(modelSource, env, `${target.name} Gemini model`, {
     allowLiteral: true,
     optionalEnv: true
   }) ?? "gemini-2.5-flash";
+  const retry = resolveRetryConfig(target);
   return {
     apiKey,
     model,
     temperature: resolveOptionalNumber(temperatureSource, `${target.name} temperature`),
-    maxOutputTokens: resolveOptionalNumber(maxTokensSource, `${target.name} max output tokens`)
+    maxOutputTokens: resolveOptionalNumber(maxTokensSource, `${target.name} max output tokens`),
+    retry
   };
 }
 function resolveCodexConfig(target, env) {
-  const settings = target.settings ?? {};
-  const executableSource = settings.executable ?? settings.command ?? settings.binary;
-  const argsSource = settings.args ?? settings.arguments;
-  const cwdSource = settings.cwd;
-  const timeoutSource = settings.timeout_seconds ?? settings.timeoutSeconds;
-  const logDirSource = settings.log_dir ?? settings.logDir ?? settings.log_directory ?? settings.logDirectory;
-  const logFormatSource = settings.log_format ?? settings.logFormat ?? settings.log_output_format ?? settings.logOutputFormat ?? env.AGENTV_CODEX_LOG_FORMAT;
+  const executableSource = target.executable ?? target.command ?? target.binary;
+  const argsSource = target.args ?? target.arguments;
+  const cwdSource = target.cwd;
+  const timeoutSource = target.timeout_seconds ?? target.timeoutSeconds;
+  const logDirSource = target.log_dir ?? target.logDir ?? target.log_directory ?? target.logDirectory;
+  const logFormatSource = target.log_format ?? target.logFormat ?? target.log_output_format ?? target.logOutputFormat ?? env.AGENTV_CODEX_LOG_FORMAT;
   const executable = resolveOptionalString(executableSource, env, `${target.name} codex executable`, {
     allowLiteral: true,
     optionalEnv: true
@@ -2255,21 +2551,19 @@ function normalizeCodexLogFormat(value) {
   throw new Error("codex log format must be 'summary' or 'json'");
 }
 function resolveMockConfig(target) {
-  const settings = target.settings ?? {};
-  const response = typeof settings.response === "string" ? settings.response : void 0;
+  const response = typeof target.response === "string" ? target.response : void 0;
   return { response };
 }
 function resolveVSCodeConfig(target, env, insiders) {
-  const settings = target.settings ?? {};
-  const workspaceTemplateEnvVar = resolveOptionalLiteralString(settings.workspace_template ?? settings.workspaceTemplate);
+  const workspaceTemplateEnvVar = resolveOptionalLiteralString(target.workspace_template ?? target.workspaceTemplate);
   const workspaceTemplate = workspaceTemplateEnvVar ? resolveOptionalString(workspaceTemplateEnvVar, env, `${target.name} workspace template path`, {
     allowLiteral: false,
     optionalEnv: true
   }) : void 0;
-  const commandSource = settings.vscode_cmd ?? settings.command;
-  const waitSource = settings.wait;
-  const dryRunSource = settings.dry_run ?? settings.dryRun;
-  const subagentRootSource = settings.subagent_root ?? settings.subagentRoot;
+  const commandSource = target.vscode_cmd ?? target.command;
+  const waitSource = target.wait;
+  const dryRunSource = target.dry_run ?? target.dryRun;
+  const subagentRootSource = target.subagent_root ?? target.subagentRoot;
   const defaultCommand = insiders ? "code-insiders" : "code";
   const command = resolveOptionalLiteralString(commandSource) ?? defaultCommand;
   return {
@@ -2284,18 +2578,16 @@ function resolveVSCodeConfig(target, env, insiders) {
   };
 }
 function resolveCliConfig(target, env) {
-  const settings = target.settings ?? {};
-  const commandTemplateSource = settings.command_template ?? settings.commandTemplate;
+  const commandTemplateSource = target.command_template ?? target.commandTemplate;
   const filesFormat = resolveOptionalLiteralString(
-    settings.files_format ?? settings.filesFormat ?? settings.attachments_format ?? settings.attachmentsFormat
+    target.files_format ?? target.filesFormat ?? target.attachments_format ?? target.attachmentsFormat
   );
-  const cwd = resolveOptionalString(settings.cwd, env, `${target.name} working directory`, {
+  const cwd = resolveOptionalString(target.cwd, env, `${target.name} working directory`, {
     allowLiteral: true,
     optionalEnv: true
   });
-  const envOverrides = resolveEnvOverrides(settings.env, env, target.name);
-  const timeoutMs = resolveTimeoutMs(settings.timeout_seconds ?? settings.timeoutSeconds, `${target.name} timeout`);
-  const healthcheck = resolveCliHealthcheck(settings.healthcheck, env, target.name);
+  const timeoutMs = resolveTimeoutMs(target.timeout_seconds ?? target.timeoutSeconds, `${target.name} timeout`);
+  const healthcheck = resolveCliHealthcheck(target.healthcheck, env, target.name);
   const commandTemplate = resolveString(
     commandTemplateSource,
     env,
@@ -2307,29 +2599,10 @@ function resolveCliConfig(target, env) {
     commandTemplate,
     filesFormat,
     cwd,
-    env: envOverrides,
     timeoutMs,
     healthcheck
   };
 }
-function resolveEnvOverrides(source, env, targetName) {
-  if (source === void 0 || source === null) {
-    return void 0;
-  }
-  if (typeof source !== "object" || Array.isArray(source)) {
-    throw new Error(`${targetName} env overrides must be an object map of strings`);
-  }
-  const entries = Object.entries(source);
-  const resolved = {};
-  for (const [key, value] of entries) {
-    if (typeof value !== "string") {
-      throw new Error(`${targetName} env override '${key}' must be a string`);
-    }
-    const resolvedValue = resolveString(value, env, `${targetName} env override '${key}'`);
-    resolved[key] = resolvedValue;
-  }
-  return Object.keys(resolved).length > 0 ? resolved : void 0;
-}
 function resolveTimeoutMs(source, description) {
   const seconds = resolveOptionalNumber(source, `${description} (seconds)`);
   if (seconds === void 0) {
@@ -2525,6 +2798,26 @@ function resolveOptionalStringArray(source, env, description) {
   }
   return resolved.length > 0 ? resolved : void 0;
 }
+function resolveOptionalNumberArray(source, description) {
+  if (source === void 0 || source === null) {
+    return void 0;
+  }
+  if (!Array.isArray(source)) {
+    throw new Error(`${description} must be an array of numbers`);
+  }
+  if (source.length === 0) {
+    return void 0;
+  }
+  const resolved = [];
+  for (let i = 0; i < source.length; i++) {
+    const item = source[i];
+    if (typeof item !== "number" || !Number.isFinite(item)) {
+      throw new Error(`${description}[${i}] must be a number`);
+    }
+    resolved.push(item);
+  }
+  return resolved.length > 0 ? resolved : void 0;
+}
 
 // src/evaluation/providers/vscode.ts
 var import_node_path6 = __toESM(require("path"), 1);
@@ -2784,7 +3077,7 @@ var AGENT_PROVIDER_KINDS = [
   "vscode",
   "vscode-insiders"
 ];
-var TARGETS_SCHEMA_V2 = "agentv-targets-v2.1";
+var TARGETS_SCHEMA_V2 = "agentv-targets-v2.2";
 function isAgentProvider(provider) {
   return provider ? AGENT_PROVIDER_KINDS.includes(provider.kind) : false;
 }
@@ -2827,20 +3120,13 @@ 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 {
@@ -2920,19 +3206,21 @@ 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 } : {},
@@ -2970,38 +3258,51 @@ 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) {
@@ -3084,6 +3385,9 @@ 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;
@@ -3481,10 +3785,11 @@ 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) {
@@ -3741,11 +4046,27 @@ 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,
@@ -3759,7 +4080,8 @@ 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
   };
@@ -3988,6 +4310,7 @@ 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,
@@ -4004,12 +4327,30 @@ 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,
@@ -4022,7 +4363,9 @@ 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) {
@@ -4033,6 +4376,9 @@ 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) {