agentv 0.7.5 → 0.10.0

package/dist/{chunk-J3LVKRRT.js → chunk-J5HK75TC.js}

@@ -590,7 +590,7 @@ import fg from "fast-glob";
 import { stat as stat3 } from "node:fs/promises";
 import path15 from "node:path";
 
-// ../../packages/core/dist/chunk-7XM7HYRS.js
+// ../../packages/core/dist/chunk-YQBJAT5I.js
 import { constants } from "node:fs";
 import { access, readFile } from "node:fs/promises";
 import path from "node:path";
@@ -4636,7 +4636,7 @@ var coerce = {
 };
 var NEVER = INVALID;
 
-// ../../packages/core/dist/chunk-7XM7HYRS.js
+// ../../packages/core/dist/chunk-YQBJAT5I.js
 async function fileExists(filePath) {
   try {
     await access(filePath, constants.F_OK);
@@ -4747,10 +4747,9 @@ var CLI_PLACEHOLDERS = /* @__PURE__ */ new Set(["PROMPT", "GUIDELINES", "EVAL_ID
 var BASE_TARGET_SCHEMA = external_exports.object({
   name: external_exports.string().min(1, "target name is required"),
   provider: external_exports.string().min(1, "provider is required"),
-  settings: external_exports.record(external_exports.unknown()).optional(),
   judge_target: external_exports.string().optional(),
   workers: external_exports.number().int().min(1).optional()
-});
+}).passthrough();
 var DEFAULT_AZURE_API_VERSION = "2024-10-01-preview";
 function normalizeAzureApiVersion(value) {
   if (!value) {
@@ -4763,11 +4762,43 @@ 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,
+    `${target.name} retry status codes`
+  );
+  if (maxRetries === void 0 && initialDelayMs === void 0 && maxDelayMs === void 0 && backoffFactor === void 0 && retryableStatusCodes === void 0) {
+    return void 0;
+  }
+  return {
+    maxRetries,
+    initialDelayMs,
+    maxDelayMs,
+    backoffFactor,
+    retryableStatusCodes
+  };
+}
 function resolveTargetDefinition(definition, env = process.env) {
   const parsed = BASE_TARGET_SCHEMA.parse(definition);
   const provider = parsed.provider.toLowerCase();
   const providerBatching = resolveOptionalBoolean(
-    parsed.settings?.provider_batching ?? parsed.settings?.providerBatching
+    parsed.provider_batching ?? parsed.providerBatching
   );
   switch (provider) {
     case "azure":
@@ -4843,13 +4874,12 @@ function resolveTargetDefinition(definition, env = process.env) {
   }
 }
 function resolveAzureConfig(target, env) {
-  const settings = target.settings ?? {};
-  const endpointSource = settings.endpoint ?? settings.resource ?? settings.resourceName;
-  const apiKeySource = settings.api_key ?? settings.apiKey;
-  const deploymentSource = settings.deployment ?? settings.deploymentName ?? settings.model;
-  const versionSource = settings.version ?? settings.api_version;
-  const temperatureSource = settings.temperature;
-  const maxTokensSource = settings.max_output_tokens ?? settings.maxTokens;
+  const endpointSource = target.endpoint ?? target.resource ?? target.resourceName;
+  const apiKeySource = target.api_key ?? target.apiKey;
+  const deploymentSource = target.deployment ?? target.deploymentName ?? target.model;
+  const versionSource = target.version ?? target.api_version;
+  const temperatureSource = target.temperature;
+  const maxTokensSource = target.max_output_tokens ?? target.maxTokens;
   const resourceName = resolveString(endpointSource, env, `${target.name} endpoint`);
   const apiKey = resolveString(apiKeySource, env, `${target.name} api key`);
   const deploymentName = resolveString(deploymentSource, env, `${target.name} deployment`);
@@ -4861,58 +4891,61 @@ 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, `${target.name} codex executable`, {
     allowLiteral: true,
     optionalEnv: true
@@ -4951,21 +4984,19 @@ function normalizeCodexLogFormat(value) {
   throw new Error("codex log format must be 'summary' or 'json'");
 }
 function resolveMockConfig(target) {
-  const settings = target.settings ?? {};
-  const response = typeof settings.response === "string" ? settings.response : void 0;
+  const response = typeof target.response === "string" ? target.response : void 0;
   return { response };
 }
 function resolveVSCodeConfig(target, env, insiders) {
-  const settings = target.settings ?? {};
-  const workspaceTemplateEnvVar = resolveOptionalLiteralString(settings.workspace_template ?? settings.workspaceTemplate);
+  const workspaceTemplateEnvVar = resolveOptionalLiteralString(target.workspace_template ?? target.workspaceTemplate);
   const workspaceTemplate = workspaceTemplateEnvVar ? resolveOptionalString(workspaceTemplateEnvVar, env, `${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 {
@@ -4980,18 +5011,16 @@ 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,
@@ -5003,29 +5032,10 @@ function resolveCliConfig(target, env) {
     commandTemplate,
     filesFormat,
     cwd,
-    env: envOverrides,
     timeoutMs,
     healthcheck
   };
 }
-function resolveEnvOverrides(source2, env, targetName) {
-  if (source2 === void 0 || source2 === null) {
-    return void 0;
-  }
-  if (typeof source2 !== "object" || Array.isArray(source2)) {
-    throw new Error(`${targetName} env overrides must be an object map of strings`);
-  }
-  const entries = Object.entries(source2);
-  const resolved = {};
-  for (const [key2, value] of entries) {
-    if (typeof value !== "string") {
-      throw new Error(`${targetName} env override '${key2}' must be a string`);
-    }
-    const resolvedValue = resolveString(value, env, `${targetName} env override '${key2}'`);
-    resolved[key2] = resolvedValue;
-  }
-  return Object.keys(resolved).length > 0 ? resolved : void 0;
-}
 function resolveTimeoutMs(source2, description) {
   const seconds = resolveOptionalNumber(source2, `${description} (seconds)`);
   if (seconds === void 0) {
@@ -5221,6 +5231,26 @@ function resolveOptionalStringArray(source2, env, description) {
   }
   return resolved.length > 0 ? resolved : void 0;
 }
+function resolveOptionalNumberArray(source2, description) {
+  if (source2 === void 0 || source2 === null) {
+    return void 0;
+  }
+  if (!Array.isArray(source2)) {
+    throw new Error(`${description} must be an array of numbers`);
+  }
+  if (source2.length === 0) {
+    return void 0;
+  }
+  const resolved = [];
+  for (let i6 = 0; i6 < source2.length; i6++) {
+    const item = source2[i6];
+    if (typeof item !== "number" || !Number.isFinite(item)) {
+      throw new Error(`${description}[${i6}] must be a number`);
+    }
+    resolved.push(item);
+  }
+  return resolved.length > 0 ? resolved : void 0;
+}
 var AGENT_PROVIDER_KINDS = [
   "codex",
   "vscode",
@@ -5252,7 +5282,7 @@ var PROVIDER_ALIASES = [
   "vertex"
   // legacy/future support
 ];
-var TARGETS_SCHEMA_V2 = "agentv-targets-v2.1";
+var TARGETS_SCHEMA_V2 = "agentv-targets-v2.2";
 function isAgentProvider(provider) {
   return provider ? AGENT_PROVIDER_KINDS.includes(provider.kind) : false;
 }
@@ -11917,14 +11947,11 @@ Please add '$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: ${id}`);
       continue;
     }
     if (expectedMessages.length > 1) {
@@ -11942,17 +11969,17 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       messageType: "input",
       verbose
     });
-    const outputSegments = await processMessages({
+    const outputSegments = hasExpectedMessages ? await processMessages({
       messages: expectedMessages,
       searchRoots,
       repoRootPath,
       guidelinePatterns,
       messageType: "output",
       verbose
-    });
+    }) : [];
     const codeSnippets = extractCodeBlocks(inputSegments);
     const expectedContent = expectedMessages[0]?.content;
-    const referenceAnswer = await resolveAssistantContent(expectedContent, searchRoots, verbose);
+    const referenceAnswer = expectedContent ? await resolveAssistantContent(expectedContent, searchRoots, verbose) : "";
     const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
     const evalCaseEvaluatorKind = coerceEvaluator(evalcase.evaluator, id) ?? globalEvaluator;
     const evaluators = await parseEvaluators(evalcase, searchRoots, id ?? "unknown");
@@ -11971,6 +11998,7 @@ Please add '$schema: ${SCHEMA_EVAL_V2}' at the top of the file.`;
       dataset: datasetName,
       conversation_id: conversationId,
       question,
+      input_messages: inputMessages,
       input_segments: inputSegments,
       output_segments: outputSegments,
       reference_answer: referenceAnswer,
@@ -11998,6 +12026,54 @@ Please add '$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) {
@@ -12014,36 +12090,168 @@ ${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 i6 = 0; i6 < testCase.input_messages.length; i6++) {
+      const message = testCase.input_messages[i6];
+      const segments = segmentsByMessage[i6];
+      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 i6 = startIndex; i6 < messages.length; i6++) {
+    const message = messages[i6];
+    const segments = segmentsByMessage[i6];
+    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 {
@@ -12237,21 +12445,14 @@ ${detailBlock}${ANSI_RESET}`);
 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 = [
     {
@@ -12265,6 +12466,21 @@ ${request.guidelines.trim()}`);
   ];
   return prompt;
 }
+function resolveSystemContent(request) {
+  const systemSegments = [];
+  const metadataSystemPrompt = typeof request.metadata?.systemPrompt === "string" ? request.metadata.systemPrompt : void 0;
+  if (metadataSystemPrompt && metadataSystemPrompt.trim().length > 0) {
+    systemSegments.push(metadataSystemPrompt.trim());
+  } else {
+    systemSegments.push(DEFAULT_SYSTEM_PROMPT);
+  }
+  if (request.guidelines && request.guidelines.trim().length > 0) {
+    systemSegments.push(`[[ ## Guidelines ## ]]
+
+${request.guidelines.trim()}`);
+  }
+  return systemSegments.join("\n\n");
+}
 function extractModelConfig(request, defaults) {
   const temperature = request.temperature ?? defaults.temperature;
   const maxTokens = request.maxOutputTokens ?? defaults.maxOutputTokens;
@@ -12308,6 +12524,67 @@ 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 sleep2(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 sleep2(delay);
+      if (signal?.aborted) {
+        throw new Error(`Request aborted: ${signal.reason ?? "Unknown reason"}`);
+      }
+    }
+  }
+  throw lastError;
+}
 var AzureProvider = class {
   constructor(targetName, config) {
     this.config = config;
@@ -12317,6 +12594,7 @@ var AzureProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = Wn.create({
       name: "azure-openai",
       apiKey: config.apiKey,
@@ -12333,16 +12611,21 @@ 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));
   }
@@ -12360,6 +12643,7 @@ var AnthropicProvider = class {
       maxOutputTokens: config.maxOutputTokens,
       thinkingBudget: config.thinkingBudget
     };
+    this.retryConfig = config.retry;
     this.ai = Wn.create({
       name: "anthropic",
       apiKey: config.apiKey
@@ -12370,16 +12654,21 @@ 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));
   }
@@ -12396,6 +12685,7 @@ var GeminiProvider = class {
       temperature: config.temperature,
       maxOutputTokens: config.maxOutputTokens
     };
+    this.retryConfig = config.retry;
     this.ai = Wn.create({
       name: "google-gemini",
       apiKey: config.apiKey
@@ -12406,16 +12696,21 @@ 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));
   }
@@ -12478,10 +12773,9 @@ 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
     });
@@ -12570,10 +12864,9 @@ 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
     });
@@ -13771,20 +14064,13 @@ 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 {
@@ -13855,19 +14141,21 @@ var LlmJudgeEvaluator = class {
     return this.evaluateWithPrompt(context2, judgeProvider);
   }
   async evaluateWithPrompt(context2, judgeProvider) {
-    let prompt = buildQualityPrompt(context2.evalCase, context2.candidate);
-    let systemPrompt = context2.systemPrompt ?? this.customPrompt ?? QUALITY_SYSTEM_PROMPT;
+    const hasReferenceAnswer = hasNonEmptyReferenceAnswer(context2.evalCase);
+    const formattedQuestion = context2.promptInputs.question && context2.promptInputs.question.trim().length > 0 ? context2.promptInputs.question : context2.evalCase.question;
+    let prompt = buildQualityPrompt(context2.evalCase, context2.candidate, formattedQuestion);
+    let systemPrompt = context2.systemPrompt ?? this.customPrompt ?? buildSystemPrompt(hasReferenceAnswer);
     if (systemPrompt && hasTemplateVariables(systemPrompt)) {
       const variables = {
         input_messages: JSON.stringify(context2.evalCase.input_segments, null, 2),
         output_messages: JSON.stringify(context2.evalCase.output_segments, null, 2),
         candidate_answer: context2.candidate,
-        reference_answer: context2.evalCase.reference_answer,
+        reference_answer: context2.evalCase.reference_answer ?? "",
         expected_outcome: context2.evalCase.expected_outcome,
-        question: context2.evalCase.question
+        question: formattedQuestion
       };
       prompt = substituteVariables(systemPrompt, variables);
-      systemPrompt = QUALITY_SYSTEM_PROMPT;
+      systemPrompt = buildSystemPrompt(hasReferenceAnswer);
     }
     const metadata = {
       ...systemPrompt !== void 0 ? { systemPrompt } : {},
@@ -13905,38 +14193,51 @@ 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) {
@@ -14019,6 +14320,9 @@ 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;
@@ -14405,10 +14709,11 @@ 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) {
@@ -14665,11 +14970,27 @@ 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,
@@ -14683,7 +15004,8 @@ 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
   };
@@ -14912,6 +15234,7 @@ 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,
@@ -14928,12 +15251,30 @@ 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,
@@ -14946,7 +15287,9 @@ 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) {
@@ -14957,6 +15300,9 @@ 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) {
@@ -15384,8 +15730,6 @@ import { stripVTControlCharacters } from "node:util";
 var ESC = "\x1B[";
 var CLEAR_LINE = `${ESC}K`;
 var MOVE_CURSOR_UP = `${ESC}1A`;
-var SYNC_START = `${ESC}?2026h`;
-var SYNC_END = `${ESC}?2026l`;
 var ProgressDisplay = class {
   workers = /* @__PURE__ */ new Map();
   maxWorkers;
@@ -15624,6 +15968,8 @@ function buildHistogram(values) {
 function calculateEvaluationSummary(results) {
   const scores = results.map((result) => result.score);
   const total = results.length;
+  const errors = results.filter((result) => result.error !== void 0).map((result) => ({ evalId: result.eval_id, error: result.error }));
+  const errorCount = errors.length;
   if (total === 0) {
     return {
       total: 0,
@@ -15634,7 +15980,9 @@ function calculateEvaluationSummary(results) {
       standardDeviation: void 0,
       histogram: buildHistogram([]),
       topResults: [],
-      bottomResults: []
+      bottomResults: [],
+      errorCount: 0,
+      errors: []
     };
   }
   const mean = computeMean(scores);
@@ -15655,7 +16003,9 @@ function calculateEvaluationSummary(results) {
     standardDeviation,
     histogram,
     topResults,
-    bottomResults
+    bottomResults,
+    errorCount,
+    errors
   };
 }
 function formatScore(value) {
@@ -15666,10 +16016,25 @@ function formatEvaluationSummary(summary) {
     return "\nNo results to summarize";
   }
   const lines = [];
+  if (summary.errorCount > 0) {
+    lines.push("\n==================================================");
+    lines.push("ERRORS");
+    lines.push("==================================================");
+    summary.errors.forEach((error) => {
+      lines.push(`
+\u274C ${error.evalId}`);
+      lines.push(`   ${error.error}`);
+    });
+    lines.push("");
+  }
   lines.push("\n==================================================");
   lines.push("EVALUATION SUMMARY");
   lines.push("==================================================");
   lines.push(`Total eval cases: ${summary.total}`);
+  if (summary.errorCount > 0) {
+    lines.push(`Failed: ${summary.errorCount}`);
+    lines.push(`Passed: ${summary.total - summary.errorCount}`);
+  }
   lines.push(`Mean score: ${formatScore(summary.mean)}`);
   lines.push(`Median score: ${formatScore(summary.median)}`);
   lines.push(`Min score: ${formatScore(summary.min)}`);
@@ -15708,7 +16073,7 @@ import { readFile as readFile5 } from "node:fs/promises";
 import path33 from "node:path";
 import { parse as parse5 } from "yaml";
 var SCHEMA_EVAL_V22 = "agentv-eval-v2";
-var SCHEMA_TARGETS_V2 = "agentv-targets-v2.1";
+var SCHEMA_TARGETS_V2 = "agentv-targets-v2.2";
 var SCHEMA_CONFIG_V22 = "agentv-config-v2";
 async function detectFileType(filePath) {
   try {
@@ -15840,14 +16205,14 @@ async function validateEvalFile(filePath) {
       validateMessages(inputMessages, `${location}.input_messages`, absolutePath, errors);
     }
     const expectedMessages = evalCase["expected_messages"];
-    if (!Array.isArray(expectedMessages)) {
+    if (expectedMessages !== void 0 && !Array.isArray(expectedMessages)) {
       errors.push({
         severity: "error",
         filePath: absolutePath,
         location: `${location}.expected_messages`,
-        message: "Missing or invalid 'expected_messages' field (must be an array)"
+        message: "Invalid 'expected_messages' field (must be an array if provided)"
       });
-    } else {
+    } else if (Array.isArray(expectedMessages)) {
       validateMessages(expectedMessages, `${location}.expected_messages`, absolutePath, errors);
     }
   }
@@ -15883,11 +16248,13 @@ function validateMessages(messages, location, filePath, errors) {
     }
     const content = message["content"];
     if (typeof content === "string") {
+      validateContentForRoleMarkers(content, `${msgLocation}.content`, filePath, errors);
     } else if (Array.isArray(content)) {
       for (let j2 = 0; j2 < content.length; j2++) {
         const contentItem = content[j2];
         const contentLocation = `${msgLocation}.content[${j2}]`;
         if (typeof contentItem === "string") {
+          validateContentForRoleMarkers(contentItem, contentLocation, filePath, errors);
         } else if (isObject(contentItem)) {
           const type = contentItem["type"];
           if (typeof type !== "string") {
@@ -15907,6 +16274,8 @@ function validateMessages(messages, location, filePath, errors) {
                 location: `${contentLocation}.value`,
                 message: "Content with type 'text' must have a 'value' field"
               });
+            } else {
+              validateContentForRoleMarkers(value, `${contentLocation}.value`, filePath, errors);
             }
           }
         } else {
@@ -15928,6 +16297,19 @@ function validateMessages(messages, location, filePath, errors) {
     }
   }
 }
+function validateContentForRoleMarkers(content, location, filePath, errors) {
+  const markers = ["@[System]:", "@[User]:", "@[Assistant]:", "@[Tool]:"];
+  for (const marker of markers) {
+    if (content.toLowerCase().includes(marker.toLowerCase())) {
+      errors.push({
+        severity: "warning",
+        filePath,
+        location,
+        message: `Content contains potential role marker '${marker}'. This may confuse agentic providers or cause prompt injection.`
+      });
+    }
+  }
+}
 function isObject2(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -15935,8 +16317,21 @@ var COMMON_SETTINGS = /* @__PURE__ */ new Set([
   "provider_batching",
   "providerBatching"
 ]);
+var RETRY_SETTINGS = /* @__PURE__ */ new Set([
+  "max_retries",
+  "maxRetries",
+  "retry_initial_delay_ms",
+  "retryInitialDelayMs",
+  "retry_max_delay_ms",
+  "retryMaxDelayMs",
+  "retry_backoff_factor",
+  "retryBackoffFactor",
+  "retry_status_codes",
+  "retryStatusCodes"
+]);
 var AZURE_SETTINGS = /* @__PURE__ */ new Set([
   ...COMMON_SETTINGS,
+  ...RETRY_SETTINGS,
   "endpoint",
   "resource",
   "resourceName",
@@ -15953,6 +16348,7 @@ var AZURE_SETTINGS = /* @__PURE__ */ new Set([
 ]);
 var ANTHROPIC_SETTINGS = /* @__PURE__ */ new Set([
   ...COMMON_SETTINGS,
+  ...RETRY_SETTINGS,
   "api_key",
   "apiKey",
   "model",
@@ -15966,6 +16362,7 @@ var ANTHROPIC_SETTINGS = /* @__PURE__ */ new Set([
 ]);
 var GEMINI_SETTINGS = /* @__PURE__ */ new Set([
   ...COMMON_SETTINGS,
+  ...RETRY_SETTINGS,
   "api_key",
   "apiKey",
   "model",
@@ -16053,13 +16450,14 @@ function getKnownSettings(provider) {
       return null;
   }
 }
-function validateUnknownSettings(settings, provider, absolutePath, location, errors) {
+function validateUnknownSettings(target, provider, absolutePath, location, errors) {
   const knownSettings = getKnownSettings(provider);
   if (!knownSettings) {
     return;
   }
-  for (const key2 of Object.keys(settings)) {
-    if (!knownSettings.has(key2)) {
+  const baseFields = /* @__PURE__ */ new Set(["name", "provider", "judge_target", "workers", "$schema", "targets"]);
+  for (const key2 of Object.keys(target)) {
+    if (!baseFields.has(key2) && !knownSettings.has(key2)) {
       errors.push({
         severity: "warning",
         filePath: absolutePath,
@@ -16089,17 +16487,8 @@ async function validateTargetsFile(filePath) {
       errors
     };
   }
-  function validateCliSettings(settings, absolutePath2, location, errors2) {
-    if (!isObject2(settings)) {
-      errors2.push({
-        severity: "error",
-        filePath: absolutePath2,
-        location,
-        message: "CLI provider requires a 'settings' object"
-      });
-      return;
-    }
-    const commandTemplate = settings["command_template"] ?? settings["commandTemplate"];
+  function validateCliSettings(target, absolutePath2, location, errors2) {
+    const commandTemplate = target["command_template"] ?? target["commandTemplate"];
     if (typeof commandTemplate !== "string" || commandTemplate.trim().length === 0) {
       errors2.push({
         severity: "error",
@@ -16110,7 +16499,7 @@ async function validateTargetsFile(filePath) {
     } else {
       recordUnknownPlaceholders(commandTemplate, absolutePath2, `${location}.commandTemplate`, errors2);
     }
-    const attachmentsFormat = settings["attachments_format"] ?? settings["attachmentsFormat"];
+    const attachmentsFormat = target["attachments_format"] ?? target["attachmentsFormat"];
     if (attachmentsFormat !== void 0 && typeof attachmentsFormat !== "string") {
       errors2.push({
         severity: "error",
@@ -16119,7 +16508,7 @@ async function validateTargetsFile(filePath) {
         message: "'attachmentsFormat' must be a string when provided"
       });
     }
-    const filesFormat = settings["files_format"] ?? settings["filesFormat"];
+    const filesFormat = target["files_format"] ?? target["filesFormat"];
     if (filesFormat !== void 0 && typeof filesFormat !== "string") {
       errors2.push({
         severity: "error",
@@ -16128,7 +16517,7 @@ async function validateTargetsFile(filePath) {
         message: "'filesFormat' must be a string when provided"
       });
     }
-    const cwd = settings["cwd"];
+    const cwd = target["cwd"];
     if (cwd !== void 0 && typeof cwd !== "string") {
       errors2.push({
         severity: "error",
@@ -16137,7 +16526,7 @@ async function validateTargetsFile(filePath) {
         message: "'cwd' must be a string when provided"
       });
     }
-    const timeoutSeconds = settings["timeout_seconds"] ?? settings["timeoutSeconds"];
+    const timeoutSeconds = target["timeout_seconds"] ?? target["timeoutSeconds"];
     if (timeoutSeconds !== void 0) {
       const numericTimeout = Number(timeoutSeconds);
       if (!Number.isFinite(numericTimeout) || numericTimeout <= 0) {
@@ -16149,29 +16538,7 @@ async function validateTargetsFile(filePath) {
         });
       }
     }
-    const envOverrides = settings["env"];
-    if (envOverrides !== void 0) {
-      if (!isObject2(envOverrides)) {
-        errors2.push({
-          severity: "error",
-          filePath: absolutePath2,
-          location: `${location}.env`,
-          message: "'env' must be an object with string values"
-        });
-      } else {
-        for (const [key2, value] of Object.entries(envOverrides)) {
-          if (typeof value !== "string" || value.trim().length === 0) {
-            errors2.push({
-              severity: "error",
-              filePath: absolutePath2,
-              location: `${location}.env.${key2}`,
-              message: `Environment override '${key2}' must be a non-empty string`
-            });
-          }
-        }
-      }
-    }
-    const healthcheck = settings["healthcheck"];
+    const healthcheck = target["healthcheck"];
     if (healthcheck !== void 0) {
       validateCliHealthcheck(healthcheck, absolutePath2, `${location}.healthcheck`, errors2);
     }
@@ -16342,20 +16709,11 @@ async function validateTargetsFile(filePath) {
         message: `Unknown provider '${provider}'. Known providers: ${knownProviders.join(", ")}`
       });
     }
-    const settings = target["settings"];
-    if (providerValue !== "cli" && settings !== void 0 && !isObject2(settings)) {
-      errors.push({
-        severity: "error",
-        filePath: absolutePath,
-        location: `${location}.settings`,
-        message: "Invalid 'settings' field (must be an object)"
-      });
-    }
     if (providerValue === "cli") {
-      validateCliSettings(settings, absolutePath, `${location}.settings`, errors);
+      validateCliSettings(target, absolutePath, location, errors);
     }
-    if (settings !== void 0 && isObject2(settings) && typeof provider === "string") {
-      validateUnknownSettings(settings, provider, absolutePath, `${location}.settings`, errors);
+    if (typeof provider === "string") {
+      validateUnknownSettings(target, provider, absolutePath, location, errors);
     }
     const judgeTarget = target["judge_target"];
     if (judgeTarget !== void 0 && typeof judgeTarget !== "string") {
@@ -17566,4 +17924,4 @@ export {
   createProgram,
   runCli
 };
-//# sourceMappingURL=chunk-J3LVKRRT.js.map
+//# sourceMappingURL=chunk-J5HK75TC.js.map