agentv 1.2.0 → 1.5.0

package/dist/{chunk-IVIT4U6S.js → chunk-3RYQPI4H.js}

@@ -141,30 +141,14 @@ var require_dist = __commonJS({
 });
 
 // src/index.ts
-import { readFileSync as readFileSync2 } from "node:fs";
+import { readFileSync as readFileSync3 } from "node:fs";
 import { binary, run, subcommands as subcommands2 } from "cmd-ts";
 
-// src/commands/eval/index.ts
-import { stat as stat4 } from "node:fs/promises";
-import path20 from "node:path";
-import {
-  command,
-  flag,
-  number as number4,
-  option,
-  optional as optional2,
-  restPositionals,
-  string as string4
-} from "cmd-ts";
-import fg from "fast-glob";
-
-// src/commands/eval/run-eval.ts
-import { constants as constants6 } from "node:fs";
-import { access as access6, mkdir as mkdir7 } from "node:fs/promises";
-import path19 from "node:path";
-import { pathToFileURL } from "node:url";
+// src/commands/convert/index.ts
+import { readFileSync, writeFileSync } from "node:fs";
+import path14 from "node:path";
 
-// ../../packages/core/dist/chunk-V3JCB3HI.js
+// ../../packages/core/dist/chunk-KPHTMTZ3.js
 import { constants } from "node:fs";
 import { access, readFile } from "node:fs/promises";
 import path from "node:path";
@@ -648,8 +632,8 @@ function getErrorMap() {
 
 // ../../node_modules/.bun/zod@3.25.76/node_modules/zod/v3/helpers/parseUtil.js
 var makeIssue = (params) => {
-  const { data, path: path27, errorMaps, issueData } = params;
-  const fullPath = [...path27, ...issueData.path || []];
+  const { data, path: path28, errorMaps, issueData } = params;
+  const fullPath = [...path28, ...issueData.path || []];
   const fullIssue = {
     ...issueData,
     path: fullPath
@@ -765,11 +749,11 @@ var errorUtil;
 
 // ../../node_modules/.bun/zod@3.25.76/node_modules/zod/v3/types.js
 var ParseInputLazyPath = class {
-  constructor(parent, value, path27, key2) {
+  constructor(parent, value, path28, key2) {
     this._cachedPath = [];
     this.parent = parent;
     this.data = value;
-    this._path = path27;
+    this._path = path28;
     this._key = key2;
   }
   get path() {
@@ -1049,8 +1033,8 @@ var ZodType = class {
   promise() {
     return ZodPromise.create(this, this._def);
   }
-  or(option4) {
-    return ZodUnion.create([this, option4], this._def);
+  or(option5) {
+    return ZodUnion.create([this, option5], this._def);
   }
   and(incoming) {
     return ZodIntersection.create(this, incoming, this._def);
@@ -2900,7 +2884,7 @@ var ZodUnion = class extends ZodType {
       return INVALID;
     }
     if (ctx.common.async) {
-      return Promise.all(options.map(async (option4) => {
+      return Promise.all(options.map(async (option5) => {
         const childCtx = {
           ...ctx,
           common: {
@@ -2910,7 +2894,7 @@ var ZodUnion = class extends ZodType {
           parent: null
         };
         return {
-          result: await option4._parseAsync({
+          result: await option5._parseAsync({
             data: ctx.data,
             path: ctx.path,
             parent: childCtx
@@ -2921,7 +2905,7 @@ var ZodUnion = class extends ZodType {
     } else {
       let dirty = void 0;
       const issues = [];
-      for (const option4 of options) {
+      for (const option5 of options) {
         const childCtx = {
           ...ctx,
           common: {
@@ -2930,7 +2914,7 @@ var ZodUnion = class extends ZodType {
           },
           parent: null
         };
-        const result = option4._parseSync({
+        const result = option5._parseSync({
           data: ctx.data,
           path: ctx.path,
           parent: childCtx
@@ -3011,8 +2995,8 @@ var ZodDiscriminatedUnion = class _ZodDiscriminatedUnion extends ZodType {
     }
     const discriminator = this.discriminator;
     const discriminatorValue = ctx.data[discriminator];
-    const option4 = this.optionsMap.get(discriminatorValue);
-    if (!option4) {
+    const option5 = this.optionsMap.get(discriminatorValue);
+    if (!option5) {
       addIssueToContext(ctx, {
         code: ZodIssueCode.invalid_union_discriminator,
         options: Array.from(this.optionsMap.keys()),
@@ -3021,13 +3005,13 @@ var ZodDiscriminatedUnion = class _ZodDiscriminatedUnion extends ZodType {
       return INVALID;
     }
     if (ctx.common.async) {
-      return option4._parseAsync({
+      return option5._parseAsync({
         data: ctx.data,
         path: ctx.path,
         parent: ctx
       });
     } else {
-      return option4._parseSync({
+      return option5._parseSync({
         data: ctx.data,
         path: ctx.path,
         parent: ctx
@@ -4211,7 +4195,7 @@ var coerce = {
 };
 var NEVER = INVALID;
 
-// ../../packages/core/dist/chunk-V3JCB3HI.js
+// ../../packages/core/dist/chunk-KPHTMTZ3.js
 async function fileExists(filePath) {
   try {
     await access(filePath, constants.F_OK);
@@ -4227,10 +4211,6 @@ async function readTextFile(filePath) {
   const content = await readFile(filePath, "utf8");
   return normalizeLineEndings(content);
 }
-async function readJsonFile(filePath) {
-  const content = await readFile(filePath, "utf8");
-  return JSON.parse(content);
-}
 async function findGitRoot(startPath) {
   let currentDir = path.dirname(path.resolve(startPath));
   const root2 = path.parse(currentDir).root;
@@ -4574,8 +4554,7 @@ function normalizeCodexLogFormat(value) {
 }
 function resolveMockConfig(target) {
   const response = typeof target.response === "string" ? target.response : void 0;
-  const trace2 = Array.isArray(target.trace) ? target.trace : void 0;
-  return { response, trace: trace2 };
+  return { response };
 }
 function resolveVSCodeConfig(target, env, insiders) {
   const workspaceTemplateEnvVar = resolveOptionalLiteralString(
@@ -4595,9 +4574,9 @@ function resolveVSCodeConfig(target, env, insiders) {
   const dryRunSource = target.dry_run ?? target.dryRun;
   const subagentRootSource = target.subagent_root ?? target.subagentRoot;
   const defaultCommand = insiders ? "code-insiders" : "code";
-  const command5 = resolveOptionalLiteralString(commandSource) ?? defaultCommand;
+  const command6 = resolveOptionalLiteralString(commandSource) ?? defaultCommand;
   return {
-    command: command5,
+    command: command6,
     waitForResponse: resolveOptionalBoolean(waitSource) ?? true,
     dryRun: resolveOptionalBoolean(dryRunSource) ?? false,
     subagentRoot: resolveOptionalString(subagentRootSource, env, `${target.name} subagent root`, {
@@ -4612,10 +4591,17 @@ function resolveCliConfig(target, env, evalFilePath) {
   const filesFormat = resolveOptionalLiteralString(
     target.files_format ?? target.filesFormat ?? target.attachments_format ?? target.attachmentsFormat
   );
+  const verbose = resolveOptionalBoolean(target.verbose ?? target.cli_verbose ?? target.cliVerbose);
+  const keepTempFiles = resolveOptionalBoolean(
+    target.keep_temp_files ?? target.keepTempFiles ?? target.keep_output_files ?? target.keepOutputFiles
+  );
   let cwd = resolveOptionalString(target.cwd, env, `${target.name} working directory`, {
     allowLiteral: true,
     optionalEnv: true
   });
+  if (cwd && evalFilePath && !path2.isAbsolute(cwd)) {
+    cwd = path2.resolve(path2.dirname(path2.resolve(evalFilePath)), cwd);
+  }
   if (!cwd && evalFilePath) {
     cwd = path2.dirname(path2.resolve(evalFilePath));
   }
@@ -4623,7 +4609,7 @@ function resolveCliConfig(target, env, evalFilePath) {
     target.timeout_seconds ?? target.timeoutSeconds,
     `${target.name} timeout`
   );
-  const healthcheck = resolveCliHealthcheck(target.healthcheck, env, target.name);
+  const healthcheck = resolveCliHealthcheck(target.healthcheck, env, target.name, evalFilePath);
   const commandTemplate = resolveString(
     commandTemplateSource,
     env,
@@ -4636,7 +4622,9 @@ function resolveCliConfig(target, env, evalFilePath) {
     filesFormat,
     cwd,
     timeoutMs,
-    healthcheck
+    healthcheck,
+    verbose,
+    keepTempFiles
   };
 }
 function resolveTimeoutMs(source2, description) {
@@ -4649,7 +4637,7 @@ function resolveTimeoutMs(source2, description) {
   }
   return Math.floor(seconds * 1e3);
 }
-function resolveCliHealthcheck(source2, env, targetName) {
+function resolveCliHealthcheck(source2, env, targetName, evalFilePath) {
   if (source2 === void 0 || source2 === null) {
     return void 0;
   }
@@ -4682,11 +4670,12 @@ function resolveCliHealthcheck(source2, env, targetName) {
       allowLiteral: true,
       optionalEnv: true
     });
+    const resolvedCwd = cwd && evalFilePath && !path2.isAbsolute(cwd) ? path2.resolve(path2.dirname(path2.resolve(evalFilePath)), cwd) : cwd;
     return {
       type: "command",
       commandTemplate,
       timeoutMs,
-      cwd
+      cwd: resolvedCwd
     };
   }
   throw new Error(`${targetName} healthcheck type must be 'http' or 'command'`);
@@ -4885,6 +4874,21 @@ var PROVIDER_ALIASES = [
   "vertex"
   // legacy/future support
 ];
+function extractLastAssistantContent(messages) {
+  if (!messages || messages.length === 0) {
+    return "";
+  }
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (msg.role === "assistant" && msg.content !== void 0) {
+      if (typeof msg.content === "string") {
+        return msg.content;
+      }
+      return JSON.stringify(msg.content);
+    }
+  }
+  return "";
+}
 function isAgentProvider(provider) {
   return provider ? AGENT_PROVIDER_KINDS.includes(provider.kind) : false;
 }
@@ -5995,10 +5999,10 @@ function assignProp(target, prop, value) {
     configurable: true
   });
 }
-function getElementAtPath(obj, path27) {
-  if (!path27)
+function getElementAtPath(obj, path28) {
+  if (!path28)
     return obj;
-  return path27.reduce((acc, key2) => acc?.[key2], obj);
+  return path28.reduce((acc, key2) => acc?.[key2], obj);
 }
 function promiseAllObject(promisesObj) {
   const keys = Object.keys(promisesObj);
@@ -6318,11 +6322,11 @@ function aborted(x, startIndex = 0) {
   }
   return false;
 }
-function prefixIssues(path27, issues) {
+function prefixIssues(path28, issues) {
   return issues.map((iss) => {
     var _a17;
     (_a17 = iss).path ?? (_a17.path = []);
-    iss.path.unshift(path27);
+    iss.path.unshift(path28);
     return iss;
   });
 }
@@ -6459,7 +6463,7 @@ function treeifyError(error40, _mapper) {
     return issue2.message;
   };
   const result = { errors: [] };
-  const processError = (error41, path27 = []) => {
+  const processError = (error41, path28 = []) => {
     var _a17, _b8;
     for (const issue2 of error41.issues) {
       if (issue2.code === "invalid_union" && issue2.errors.length) {
@@ -6469,7 +6473,7 @@ function treeifyError(error40, _mapper) {
       } else if (issue2.code === "invalid_element") {
         processError({ issues: issue2.issues }, issue2.path);
       } else {
-        const fullpath = [...path27, ...issue2.path];
+        const fullpath = [...path28, ...issue2.path];
         if (fullpath.length === 0) {
           result.errors.push(mapper(issue2));
           continue;
@@ -6499,9 +6503,9 @@ function treeifyError(error40, _mapper) {
   processError(error40);
   return result;
 }
-function toDotPath(path27) {
+function toDotPath(path28) {
   const segs = [];
-  for (const seg of path27) {
+  for (const seg of path28) {
     if (typeof seg === "number")
       segs.push(`[${seg}]`);
     else if (typeof seg === "symbol")
@@ -8100,7 +8104,7 @@ var $ZodUnion = /* @__PURE__ */ $constructor("$ZodUnion", (inst, def) => {
   defineLazy(inst._zod, "optout", () => def.options.some((o) => o._zod.optout === "optional") ? "optional" : void 0);
   defineLazy(inst._zod, "values", () => {
     if (def.options.every((o) => o._zod.values)) {
-      return new Set(def.options.flatMap((option4) => Array.from(option4._zod.values)));
+      return new Set(def.options.flatMap((option5) => Array.from(option5._zod.values)));
     }
     return void 0;
   });
@@ -8114,8 +8118,8 @@ var $ZodUnion = /* @__PURE__ */ $constructor("$ZodUnion", (inst, def) => {
   inst._zod.parse = (payload, ctx) => {
     let async = false;
     const results = [];
-    for (const option4 of def.options) {
-      const result = option4._zod.run({
+    for (const option5 of def.options) {
+      const result = option5._zod.run({
         value: payload.value,
         issues: []
       }, ctx);
@@ -8140,10 +8144,10 @@ var $ZodDiscriminatedUnion = /* @__PURE__ */ $constructor("$ZodDiscriminatedUnio
   const _super = inst._zod.parse;
   defineLazy(inst._zod, "propValues", () => {
     const propValues = {};
-    for (const option4 of def.options) {
-      const pv = option4._zod.propValues;
+    for (const option5 of def.options) {
+      const pv = option5._zod.propValues;
       if (!pv || Object.keys(pv).length === 0)
-        throw new Error(`Invalid discriminated union option at index "${def.options.indexOf(option4)}"`);
+        throw new Error(`Invalid discriminated union option at index "${def.options.indexOf(option5)}"`);
       for (const [k, v] of Object.entries(pv)) {
         if (!propValues[k])
           propValues[k] = /* @__PURE__ */ new Set();
@@ -15347,8 +15351,8 @@ function isTransforming(_schema, _ctx) {
       return false;
     }
     case "union": {
-      for (const option4 of def.options) {
-        if (isTransforming(option4, ctx))
+      for (const option5 of def.options) {
+        if (isTransforming(option5, ctx))
           return true;
       }
       return false;
@@ -26054,14 +26058,14 @@ function createAzure(options = {}) {
     description: "Azure OpenAI resource name"
   });
   const apiVersion = (_a17 = options.apiVersion) != null ? _a17 : "v1";
-  const url2 = ({ path: path27, modelId }) => {
+  const url2 = ({ path: path28, modelId }) => {
     var _a24;
     const baseUrlPrefix = (_a24 = options.baseURL) != null ? _a24 : `https://${getResourceName()}.openai.azure.com/openai`;
     let fullUrl;
     if (options.useDeploymentBasedUrls) {
-      fullUrl = new URL(`${baseUrlPrefix}/deployments/${modelId}${path27}`);
+      fullUrl = new URL(`${baseUrlPrefix}/deployments/${modelId}${path28}`);
     } else {
-      fullUrl = new URL(`${baseUrlPrefix}/v1${path27}`);
+      fullUrl = new URL(`${baseUrlPrefix}/v1${path28}`);
     }
     fullUrl.searchParams.set("api-version", apiVersion);
     return fullUrl.toString();
@@ -34589,33 +34593,22 @@ var EVALUATOR_KIND_SET = new Set(EVALUATOR_KIND_VALUES);
 function isEvaluatorKind(value) {
   return typeof value === "string" && EVALUATOR_KIND_SET.has(value);
 }
-function isTraceEventType(value) {
-  return typeof value === "string" && ["model_step", "tool_call", "tool_result", "message", "error"].includes(value);
-}
-function isTraceEvent(value) {
-  if (typeof value !== "object" || value === null) {
-    return false;
-  }
-  const candidate = value;
-  return isTraceEventType(candidate.type) && typeof candidate.timestamp === "string";
-}
-function computeTraceSummary(trace2) {
+function computeTraceSummary(messages) {
   const toolCallCounts = {};
-  let errorCount = 0;
-  for (const event of trace2) {
-    if (event.type === "tool_call" && event.name) {
-      toolCallCounts[event.name] = (toolCallCounts[event.name] ?? 0) + 1;
-    }
-    if (event.type === "error") {
-      errorCount++;
+  let totalToolCalls = 0;
+  for (const message of messages) {
+    if (!message.toolCalls) continue;
+    for (const toolCall of message.toolCalls) {
+      toolCallCounts[toolCall.tool] = (toolCallCounts[toolCall.tool] ?? 0) + 1;
+      totalToolCalls++;
     }
   }
   const toolNames = Object.keys(toolCallCounts).sort();
   return {
-    eventCount: trace2.length,
+    eventCount: totalToolCalls,
     toolNames,
     toolCallsByName: toolCallCounts,
-    errorCount
+    errorCount: 0
   };
 }
 function extractCodeBlocks(segments) {
@@ -34863,7 +34856,8 @@ var TEMPLATE_VARIABLES = {
   QUESTION: "question",
   EXPECTED_OUTCOME: "expected_outcome",
   REFERENCE_ANSWER: "reference_answer",
-  INPUT_MESSAGES: "input_messages"
+  INPUT_MESSAGES: "input_messages",
+  OUTPUT_MESSAGES: "output_messages"
 };
 var VALID_TEMPLATE_VARIABLES = new Set(Object.values(TEMPLATE_VARIABLES));
 var REQUIRED_TEMPLATE_VARIABLES = /* @__PURE__ */ new Set([
@@ -35253,6 +35247,17 @@ async function processMessages(options) {
       }
       continue;
     }
+    if (isJsonObject(content)) {
+      const rendered = JSON.stringify(content, null, 2);
+      segments.push({ type: "text", value: rendered });
+      if (textParts) {
+        textParts.push(rendered);
+      }
+      continue;
+    }
+    if (!Array.isArray(content)) {
+      continue;
+    }
     for (const rawSegment of content) {
       if (!isJsonObject(rawSegment)) {
         continue;
@@ -35475,6 +35480,11 @@ async function buildPromptInputs(testCase, mode = "lm") {
           }
         }
       }
+    } else if (isJsonObject(message.content)) {
+      const rendered = JSON.stringify(message.content, null, 2);
+      if (rendered.trim().length > 0) {
+        messageSegments.push({ type: "text", value: rendered });
+      }
     }
     segmentsByMessage.push(messageSegments);
   }
@@ -35716,16 +35726,16 @@ async function loadEvalCases(evalFilePath, repoRoot, options) {
     }) : [];
     const codeSnippets = extractCodeBlocks(inputSegments);
     let referenceAnswer = "";
-    if (outputSegments.length > 1) {
-      referenceAnswer = JSON.stringify(outputSegments, null, 2);
-    } else if (outputSegments.length === 1) {
-      const singleMessage = outputSegments[0];
-      if (typeof singleMessage.content === "string") {
-        referenceAnswer = singleMessage.content;
-      } else if (singleMessage.content) {
-        referenceAnswer = JSON.stringify(singleMessage, null, 2);
-      } else if (singleMessage.tool_calls) {
-        referenceAnswer = JSON.stringify(singleMessage, null, 2);
+    if (outputSegments.length > 0) {
+      const lastMessage = outputSegments[outputSegments.length - 1];
+      const content = lastMessage.content;
+      const toolCalls = lastMessage.tool_calls;
+      if (typeof content === "string") {
+        referenceAnswer = content;
+      } else if (content !== void 0 && content !== null) {
+        referenceAnswer = JSON.stringify(content, null, 2);
+      } else if (toolCalls !== void 0 && toolCalls !== null) {
+        referenceAnswer = JSON.stringify(toolCalls, null, 2);
       }
     }
     const question = inputTextParts.map((part) => part.trim()).filter((part) => part.length > 0).join(" ");
@@ -36047,11 +36057,11 @@ async function invokeModel(options) {
   return mapResponse(result);
 }
 function mapResponse(result) {
+  const content = result.text ?? "";
   return {
-    text: result.text ?? "",
-    reasoning: result.reasoningText ?? void 0,
     raw: result,
-    usage: toJsonObject(result.totalUsage ?? result.usage)
+    usage: toJsonObject(result.totalUsage ?? result.usage),
+    outputMessages: [{ role: "assistant", content }]
   };
 }
 function toJsonObject(value) {
@@ -36158,7 +36168,7 @@ async function withRetry(fn, retryConfig, signal) {
 }
 var execAsync2 = promisify2(execWithCallback);
 var DEFAULT_MAX_BUFFER = 10 * 1024 * 1024;
-async function defaultCommandRunner(command5, options) {
+async function defaultCommandRunner(command6, options) {
   const execOptions = {
     cwd: options.cwd,
     env: options.env,
@@ -36168,7 +36178,7 @@ async function defaultCommandRunner(command5, options) {
     shell: process.platform === "win32" ? "powershell.exe" : void 0
   };
   try {
-    const { stdout, stderr } = await execAsync2(command5, execOptions);
+    const { stdout, stderr } = await execAsync2(command6, execOptions);
     return {
       stdout,
       stderr,
@@ -36193,10 +36203,11 @@ var CliProvider = class {
   id;
   kind = "cli";
   targetName;
-  supportsBatch = false;
+  supportsBatch = true;
   config;
   runCommand;
   verbose;
+  keepTempFiles;
   healthcheckPromise;
   constructor(targetName, config2, runner = defaultCommandRunner) {
     this.targetName = targetName;
@@ -36204,6 +36215,7 @@ var CliProvider = class {
     this.config = config2;
     this.runCommand = runner;
     this.verbose = config2.verbose ?? false;
+    this.keepTempFiles = config2.keepTempFiles ?? false;
   }
   async invoke(request) {
     if (request.signal?.aborted) {
@@ -36213,6 +36225,11 @@ var CliProvider = class {
     const outputFilePath = generateOutputFilePath(request.evalCaseId);
     const templateValues = buildTemplateValues(request, this.config, outputFilePath);
     const renderedCommand = renderTemplate2(this.config.commandTemplate, templateValues);
+    if (this.verbose) {
+      console.log(
+        `[cli-provider:${this.targetName}] cwd=${this.config.cwd ?? ""} command=${renderedCommand}`
+      );
+    }
     const result = await this.runCommand(renderedCommand, {
       cwd: this.config.cwd,
       env: process.env,
@@ -36236,8 +36253,7 @@ var CliProvider = class {
     const responseContent = await this.readAndCleanupOutputFile(outputFilePath);
     const parsed = this.parseOutputContent(responseContent);
     return {
-      text: parsed.text,
-      trace: parsed.trace,
+      outputMessages: parsed.outputMessages,
       raw: {
         command: renderedCommand,
         stderr: result.stderr,
@@ -36247,30 +36263,225 @@ var CliProvider = class {
       }
     };
   }
+  async invokeBatch(requests) {
+    if (requests.length === 0) {
+      return [];
+    }
+    for (const request of requests) {
+      if (request.signal?.aborted) {
+        throw new Error("CLI provider batch request was aborted before execution");
+      }
+    }
+    const controller = new AbortController();
+    for (const request of requests) {
+      request.signal?.addEventListener("abort", () => controller.abort(), { once: true });
+    }
+    await this.ensureHealthy(controller.signal);
+    const outputFilePath = generateOutputFilePath("batch", ".jsonl");
+    const batchInputFiles = [];
+    for (const request of requests) {
+      if (request.inputFiles && request.inputFiles.length > 0) {
+        batchInputFiles.push(...request.inputFiles);
+      }
+    }
+    const templateValues = buildTemplateValues(
+      {
+        question: "",
+        guidelines: "",
+        inputFiles: batchInputFiles,
+        evalCaseId: "batch",
+        attempt: 0
+      },
+      this.config,
+      outputFilePath
+    );
+    const renderedCommand = renderTemplate2(this.config.commandTemplate, templateValues);
+    if (this.verbose) {
+      console.log(
+        `[cli-provider:${this.targetName}] (batch size=${requests.length}) cwd=${this.config.cwd ?? ""} command=${renderedCommand}`
+      );
+    }
+    const result = await this.runCommand(renderedCommand, {
+      cwd: this.config.cwd,
+      env: process.env,
+      timeoutMs: this.config.timeoutMs,
+      signal: controller.signal
+    });
+    if (result.failed || (result.exitCode ?? 0) !== 0) {
+      if (controller.signal.aborted) {
+        throw new Error("CLI provider request was aborted");
+      }
+      if (result.timedOut) {
+        throw new Error(
+          `CLI provider timed out${formatTimeoutSuffix(this.config.timeoutMs ?? void 0)}`
+        );
+      }
+      const codeText = result.exitCode !== null ? result.exitCode : "unknown";
+      const detail = result.stderr.trim() || result.stdout.trim();
+      const message = detail ? `${detail} (exit code ${codeText})` : `CLI exited with code ${codeText}`;
+      throw new Error(message);
+    }
+    const responseContent = await this.readAndCleanupOutputFile(outputFilePath);
+    const recordsById = this.parseJsonlBatchOutput(responseContent);
+    const requestedIds = requests.map((request) => request.evalCaseId).filter((id) => typeof id === "string" && id.trim().length > 0);
+    const missingIds = requestedIds.filter((id) => !recordsById.has(id));
+    if (missingIds.length > 0) {
+      throw new Error(`CLI batch output missing ids: ${missingIds.join(", ")}`);
+    }
+    const responses = requests.map((request) => {
+      const evalCaseId = request.evalCaseId;
+      if (!evalCaseId) {
+        return {
+          outputMessages: [],
+          raw: {
+            command: renderedCommand,
+            stderr: result.stderr,
+            exitCode: result.exitCode ?? 0,
+            cwd: this.config.cwd,
+            outputFile: outputFilePath
+          }
+        };
+      }
+      const parsed = recordsById.get(evalCaseId);
+      if (!parsed) {
+        return {
+          outputMessages: [],
+          raw: {
+            command: renderedCommand,
+            stderr: result.stderr,
+            exitCode: result.exitCode ?? 0,
+            cwd: this.config.cwd,
+            outputFile: outputFilePath
+          }
+        };
+      }
+      return {
+        outputMessages: parsed.outputMessages,
+        raw: {
+          command: renderedCommand,
+          stderr: result.stderr,
+          exitCode: result.exitCode ?? 0,
+          cwd: this.config.cwd,
+          outputFile: outputFilePath,
+          recordId: evalCaseId
+        }
+      };
+    });
+    return responses;
+  }
   /**
    * Parse output content from CLI.
-   * If the content is valid JSON with a 'text' field, extract text and optional trace.
-   * Otherwise, treat the entire content as plain text.
+   * If the content is valid JSON with 'output_messages' or 'text' field, extract them.
+   * If only 'text' is provided, wrap it in outputMessages.
+   * Otherwise, treat the entire content as plain text wrapped in outputMessages.
    */
   parseOutputContent(content) {
     try {
       const parsed = JSON.parse(content);
-      if (typeof parsed === "object" && parsed !== null && "text" in parsed) {
+      if (typeof parsed === "object" && parsed !== null) {
         const obj = parsed;
-        const text2 = typeof obj.text === "string" ? obj.text : String(obj.text);
-        const trace2 = this.parseTrace(obj.trace);
-        return { text: text2, trace: trace2 };
+        const outputMessages = this.parseOutputMessages(obj.output_messages);
+        if (outputMessages && outputMessages.length > 0) {
+          return { outputMessages };
+        }
+        if ("text" in obj) {
+          const text2 = typeof obj.text === "string" ? obj.text : String(obj.text);
+          return { outputMessages: [{ role: "assistant", content: text2 }] };
+        }
       }
     } catch {
     }
-    return { text: content };
+    return { outputMessages: [{ role: "assistant", content }] };
   }
-  parseTrace(trace2) {
-    if (!Array.isArray(trace2)) {
+  /**
+   * Parse output_messages from JSONL (snake_case) and convert to OutputMessage[] (camelCase).
+   */
+  parseOutputMessages(outputMessages) {
+    if (!Array.isArray(outputMessages)) {
       return void 0;
     }
-    const validEvents = trace2.filter(isTraceEvent);
-    return validEvents.length > 0 ? validEvents : void 0;
+    const messages = [];
+    for (const msg of outputMessages) {
+      if (typeof msg !== "object" || msg === null) {
+        continue;
+      }
+      const rawMsg = msg;
+      if (typeof rawMsg.role !== "string") {
+        continue;
+      }
+      const message = {
+        role: rawMsg.role,
+        name: typeof rawMsg.name === "string" ? rawMsg.name : void 0,
+        content: rawMsg.content,
+        toolCalls: this.parseToolCalls(rawMsg.tool_calls),
+        timestamp: typeof rawMsg.timestamp === "string" ? rawMsg.timestamp : void 0,
+        metadata: typeof rawMsg.metadata === "object" && rawMsg.metadata !== null ? rawMsg.metadata : void 0
+      };
+      messages.push(message);
+    }
+    return messages.length > 0 ? messages : void 0;
+  }
+  /**
+   * Parse tool_calls from JSONL (snake_case) and convert to ToolCall[] format.
+   */
+  parseToolCalls(toolCalls) {
+    if (!Array.isArray(toolCalls)) {
+      return void 0;
+    }
+    const calls = [];
+    for (const call of toolCalls) {
+      if (typeof call !== "object" || call === null) {
+        continue;
+      }
+      const rawCall = call;
+      if (typeof rawCall.tool !== "string") {
+        continue;
+      }
+      calls.push({
+        tool: rawCall.tool,
+        input: rawCall.input,
+        output: rawCall.output,
+        id: typeof rawCall.id === "string" ? rawCall.id : void 0,
+        timestamp: typeof rawCall.timestamp === "string" ? rawCall.timestamp : void 0
+      });
+    }
+    return calls.length > 0 ? calls : void 0;
+  }
+  parseJsonlBatchOutput(content) {
+    const records = /* @__PURE__ */ new Map();
+    const lines = content.split(/\r?\n/).map((line2) => line2.trim()).filter((line2) => line2.length > 0);
+    for (const line2 of lines) {
+      let parsed;
+      try {
+        parsed = JSON.parse(line2);
+      } catch (error40) {
+        const reason = error40 instanceof Error ? error40.message : String(error40);
+        throw new Error(`CLI batch output contains invalid JSONL line: ${reason}`);
+      }
+      if (typeof parsed !== "object" || parsed === null) {
+        throw new Error("CLI batch output JSONL line must be an object");
+      }
+      const obj = parsed;
+      const id = typeof obj.id === "string" ? obj.id : void 0;
+      if (!id || id.trim().length === 0) {
+        throw new Error("CLI batch output JSONL line missing required string field: id");
+      }
+      if (records.has(id)) {
+        throw new Error(`CLI batch output contains duplicate id: ${id}`);
+      }
+      const parsedOutputMessages = this.parseOutputMessages(obj.output_messages);
+      let outputMessages;
+      if (parsedOutputMessages && parsedOutputMessages.length > 0) {
+        outputMessages = parsedOutputMessages;
+      } else {
+        const text2 = typeof obj.text === "string" ? obj.text : obj.text === void 0 ? "" : JSON.stringify(obj.text);
+        outputMessages = text2 ? [{ role: "assistant", content: text2 }] : [];
+      }
+      records.set(id, {
+        outputMessages
+      });
+    }
+    return records;
   }
   async readAndCleanupOutputFile(filePath) {
     try {
@@ -36280,8 +36491,10 @@ var CliProvider = class {
       const errorMsg = error40 instanceof Error ? error40.message : String(error40);
       throw new Error(`Failed to read output file '${filePath}': ${errorMsg}`);
     } finally {
-      await fs.unlink(filePath).catch(() => {
-      });
+      if (!this.keepTempFiles) {
+        await fs.unlink(filePath).catch(() => {
+        });
+      }
     }
   }
   async ensureHealthy(signal) {
@@ -36333,7 +36546,7 @@ var CliProvider = class {
     );
     if (this.verbose) {
       console.log(
-        `[cli-provider:${this.targetName}] (healthcheck) CLI_EVALS_DIR=${process.env.CLI_EVALS_DIR ?? ""} cwd=${healthcheck.cwd ?? this.config.cwd ?? ""} command=${renderedCommand}`
+        `[cli-provider:${this.targetName}] (healthcheck) cwd=${healthcheck.cwd ?? this.config.cwd ?? ""} command=${renderedCommand}`
       );
     }
     const result = await this.runCommand(renderedCommand, {
@@ -36401,11 +36614,11 @@ function shellEscape(value) {
   }
   return `'${value.replace(/'/g, `'"'"'`)}'`;
 }
-function generateOutputFilePath(evalCaseId) {
+function generateOutputFilePath(evalCaseId, extension = ".json") {
   const safeEvalId = evalCaseId || "unknown";
   const timestamp = Date.now();
   const random = Math.random().toString(36).substring(2, 9);
-  return path72.join(os2.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}.json`);
+  return path72.join(os2.tmpdir(), `agentv-${safeEvalId}-${timestamp}-${random}${extension}`);
 }
 function formatTimeoutSuffix(timeoutMs) {
   if (!timeoutMs || timeoutMs <= 0) {
@@ -36601,7 +36814,6 @@ var CodexProvider = class {
       const parsed = parseCodexJson(result.stdout);
       const assistantText = extractAssistantText(parsed);
       return {
-        text: assistantText,
         raw: {
           response: parsed,
           stdout: result.stdout,
@@ -36613,7 +36825,8 @@ var CodexProvider = class {
           workspace: workspaceRoot,
           inputFiles,
           logFile: logger?.filePath
-        }
+        },
+        outputMessages: [{ role: "assistant", content: assistantText }]
       };
     } finally {
       await logger?.close();
@@ -37233,7 +37446,6 @@ var MockProvider = class {
   delayMs;
   delayMinMs;
   delayMaxMs;
-  trace;
   constructor(targetName, config2) {
     this.id = `mock:${targetName}`;
     this.targetName = targetName;
@@ -37241,7 +37453,6 @@ var MockProvider = class {
     this.delayMs = config2.delayMs ?? 0;
     this.delayMinMs = config2.delayMinMs ?? 0;
     this.delayMaxMs = config2.delayMaxMs ?? 0;
-    this.trace = config2.trace;
   }
   async invoke(request) {
     const delay2 = this.calculateDelay();
@@ -37249,12 +37460,11 @@ var MockProvider = class {
       await new Promise((resolve2) => setTimeout(resolve2, delay2));
     }
     return {
-      text: this.cannedResponse,
+      outputMessages: [{ role: "assistant", content: this.cannedResponse }],
       raw: {
         question: request.question,
         guidelines: request.guidelines
-      },
-      trace: this.trace
+      }
     };
   }
   calculateDelay() {
@@ -37334,7 +37544,7 @@ var VSCodeProvider = class {
     }
     if (this.config.dryRun) {
       return {
-        text: "",
+        outputMessages: [],
         raw: {
           session,
           inputFiles
@@ -37343,7 +37553,7 @@ var VSCodeProvider = class {
     }
     const responseText = await readTextFile(session.responseFile);
     return {
-      text: responseText,
+      outputMessages: [{ role: "assistant", content: responseText }],
       raw: {
         session,
         inputFiles
@@ -37381,7 +37591,7 @@ var VSCodeProvider = class {
     }
     if (this.config.dryRun) {
       return normalizedRequests.map(({ inputFiles }) => ({
-        text: "",
+        outputMessages: [],
         raw: {
           session,
           inputFiles,
@@ -37398,7 +37608,7 @@ var VSCodeProvider = class {
     for (const [index, responseFile] of session.responseFiles.entries()) {
       const responseText = await readTextFile(responseFile);
       responses.push({
-        text: responseText,
+        outputMessages: [{ role: "assistant", content: responseText }],
         raw: {
           session,
           inputFiles: normalizedRequests[index]?.inputFiles,
@@ -37686,6 +37896,7 @@ var LlmJudgeEvaluator = class {
         null,
         2
       ),
+      [TEMPLATE_VARIABLES.OUTPUT_MESSAGES]: JSON.stringify(context.outputMessages ?? [], null, 2),
       [TEMPLATE_VARIABLES.CANDIDATE_ANSWER]: context.candidate.trim(),
       [TEMPLATE_VARIABLES.REFERENCE_ANSWER]: (context.evalCase.reference_answer ?? "").trim(),
       [TEMPLATE_VARIABLES.EXPECTED_OUTCOME]: context.evalCase.expected_outcome.trim(),
@@ -37710,7 +37921,7 @@ var LlmJudgeEvaluator = class {
       const score = clampScore(data.score);
       const hits = Array.isArray(data.hits) ? data.hits.filter(isNonEmptyString).slice(0, 4) : [];
       const misses = Array.isArray(data.misses) ? data.misses.filter(isNonEmptyString).slice(0, 4) : [];
-      const reasoning = data.reasoning ?? providerResponse?.reasoning;
+      const reasoning = data.reasoning;
       const expectedAspectCount = Math.max(hits.length + misses.length, 1);
       return {
         score,
@@ -37812,7 +38023,9 @@ var LlmJudgeEvaluator = class {
           maxOutputTokens: this.maxOutputTokens,
           temperature: this.temperature
         });
-        const data = schema.parse(parseJsonFromText(response.text ?? ""));
+        const data = schema.parse(
+          parseJsonFromText(extractLastAssistantContent(response.outputMessages))
+        );
         return { data, providerResponse: response };
       } catch (e) {
         lastError = e instanceof Error ? e : new Error(String(e));
@@ -37895,15 +38108,16 @@ var CodeEvaluator = class {
       {
         question: context.evalCase.question,
         expected_outcome: context.evalCase.expected_outcome,
+        expected_messages: context.evalCase.expected_messages,
         reference_answer: context.evalCase.reference_answer,
         candidate_answer: context.candidate,
+        output_messages: context.outputMessages ?? null,
         guideline_files: context.evalCase.guideline_paths,
         input_files: context.evalCase.file_paths.filter(
           (path132) => !context.evalCase.guideline_paths.includes(path132)
         ),
         input_messages: context.evalCase.input_messages,
-        candidate_trace_file: context.candidateTraceRef ?? null,
-        candidate_trace_summary: context.candidateTraceSummary ?? null
+        candidate_trace_summary: context.traceSummary ?? null
       },
       null,
       2
@@ -38030,8 +38244,19 @@ var ToolTrajectoryEvaluator = class {
     this.config = options.config;
   }
   evaluate(context) {
-    const { candidateTrace, candidateTraceSummary } = context;
-    if (!candidateTrace || !candidateTraceSummary) {
+    const { outputMessages, traceSummary } = context;
+    const toolCalls = this.extractToolCallsFromMessages(outputMessages);
+    if (toolCalls.length === 0 && !traceSummary) {
+      return {
+        score: 0,
+        verdict: "fail",
+        hits: [],
+        misses: ["No trace available for evaluation"],
+        expectedAspectCount: 1
+      };
+    }
+    const summary = toolCalls.length > 0 ? this.buildSummary(toolCalls) : traceSummary;
+    if (!summary) {
       return {
         score: 0,
         verdict: "fail",
@@ -38042,11 +38267,11 @@ var ToolTrajectoryEvaluator = class {
     }
     switch (this.config.mode) {
       case "any_order":
-        return this.evaluateAnyOrder(candidateTraceSummary);
+        return this.evaluateAnyOrder(summary);
       case "in_order":
-        return this.evaluateInOrder(candidateTrace);
+        return this.evaluateInOrder(toolCalls);
       case "exact":
-        return this.evaluateExact(candidateTrace);
+        return this.evaluateExact(toolCalls);
       default:
         return {
           score: 0,
@@ -38057,6 +38282,39 @@ var ToolTrajectoryEvaluator = class {
         };
     }
   }
+  /**
+   * Extract tool calls from output messages.
+   */
+  extractToolCallsFromMessages(messages) {
+    if (!messages) {
+      return [];
+    }
+    const toolCalls = [];
+    for (const message of messages) {
+      if (message.toolCalls) {
+        for (const call of message.toolCalls) {
+          toolCalls.push({ name: call.tool });
+        }
+      }
+    }
+    return toolCalls;
+  }
+  /**
+   * Build a summary from extracted tool calls.
+   */
+  buildSummary(toolCalls) {
+    const toolCallsByName = {};
+    for (const call of toolCalls) {
+      toolCallsByName[call.name] = (toolCallsByName[call.name] ?? 0) + 1;
+    }
+    const toolNames = Object.keys(toolCallsByName).sort();
+    return {
+      eventCount: toolCalls.length,
+      toolNames,
+      toolCallsByName,
+      errorCount: 0
+    };
+  }
   evaluateAnyOrder(summary) {
     const minimums = this.config.minimums ?? {};
     const toolNames = Object.keys(minimums);
@@ -38089,7 +38347,7 @@ var ToolTrajectoryEvaluator = class {
       expectedAspectCount: toolNames.length
     };
   }
-  evaluateInOrder(trace2) {
+  evaluateInOrder(toolCalls) {
     const expected = this.config.expected ?? [];
     if (expected.length === 0) {
       return {
@@ -38100,15 +38358,14 @@ var ToolTrajectoryEvaluator = class {
         expectedAspectCount: 0
       };
     }
-    const actualToolCalls = trace2.filter((e) => e.type === "tool_call" && e.name);
     const hits = [];
     const misses = [];
     let actualIndex = 0;
     for (let i = 0; i < expected.length; i++) {
       const expectedTool = expected[i].tool;
       let found = false;
-      while (actualIndex < actualToolCalls.length) {
-        if (actualToolCalls[actualIndex].name === expectedTool) {
+      while (actualIndex < toolCalls.length) {
+        if (toolCalls[actualIndex].name === expectedTool) {
           hits.push(`Found ${expectedTool} at position ${actualIndex}`);
           actualIndex++;
           found = true;
@@ -38129,7 +38386,7 @@ var ToolTrajectoryEvaluator = class {
       expectedAspectCount: expected.length
     };
   }
-  evaluateExact(trace2) {
+  evaluateExact(toolCalls) {
     const expected = this.config.expected ?? [];
     if (expected.length === 0) {
       return {
@@ -38140,16 +38397,15 @@ var ToolTrajectoryEvaluator = class {
         expectedAspectCount: 0
       };
     }
-    const actualToolCalls = trace2.filter((e) => e.type === "tool_call" && e.name);
     const hits = [];
     const misses = [];
-    if (actualToolCalls.length !== expected.length) {
-      misses.push(`Expected ${expected.length} tool calls, got ${actualToolCalls.length}`);
+    if (toolCalls.length !== expected.length) {
+      misses.push(`Expected ${expected.length} tool calls, got ${toolCalls.length}`);
     }
-    const checkLength = Math.min(expected.length, actualToolCalls.length);
+    const checkLength = Math.min(expected.length, toolCalls.length);
     for (let i = 0; i < checkLength; i++) {
       const expectedTool = expected[i].tool;
-      const actualTool = actualToolCalls[i].name;
+      const actualTool = toolCalls[i].name;
       if (actualTool === expectedTool) {
         hits.push(`Position ${i}: ${expectedTool} \u2713`);
       } else {
@@ -38363,11 +38619,13 @@ var CompositeEvaluator = class {
         evalCaseId: context.evalCase.id,
         attempt: context.attempt
       });
-      const data = freeformEvaluationSchema.parse(parseJsonFromText(response.text ?? ""));
+      const data = freeformEvaluationSchema.parse(
+        parseJsonFromText(extractLastAssistantContent(response.outputMessages))
+      );
       const score = clampScore(data.score);
       const hits = Array.isArray(data.hits) ? data.hits.filter(isNonEmptyString).slice(0, 4) : [];
       const misses = Array.isArray(data.misses) ? data.misses.filter(isNonEmptyString).slice(0, 4) : [];
-      const reasoning = data.reasoning ?? response.reasoning;
+      const reasoning = data.reasoning;
       return {
         score,
         verdict: scoreToVerdict(score),
@@ -38779,11 +39037,14 @@ async function runBatchEvaluation(options) {
     const evalCase = evalCases[i];
     const promptInputs = promptInputsList[i];
     const providerResponse = batchResponse[i];
+    const outputMessages = providerResponse.outputMessages;
+    const traceSummary = outputMessages ? computeTraceSummary(outputMessages) : void 0;
+    const candidate = extractLastAssistantContent(outputMessages);
     let result;
     try {
       result = await evaluateCandidate({
         evalCase,
-        candidate: providerResponse.text ?? "",
+        candidate,
         target,
         provider,
         evaluators: evaluatorRegistry,
@@ -38791,7 +39052,9 @@ async function runBatchEvaluation(options) {
         nowFn,
         attempt: 0,
         judgeProvider: await resolveJudgeProvider(target),
-        agentTimeoutMs
+        agentTimeoutMs,
+        outputMessages,
+        traceSummary
       });
     } catch (error40) {
       const errorResult = buildErrorResult(
@@ -38895,21 +39158,13 @@ async function runEvalCase(options) {
   if (cacheKey && cache && !cachedResponse) {
     await cache.set(cacheKey, providerResponse);
   }
-  let candidateTrace = providerResponse.trace;
-  if (!candidateTrace && providerResponse.traceRef) {
-    try {
-      const rawTrace = await readJsonFile(providerResponse.traceRef);
-      if (Array.isArray(rawTrace) && rawTrace.every(isTraceEvent)) {
-        candidateTrace = rawTrace;
-      }
-    } catch {
-    }
-  }
-  const candidateTraceSummary = candidateTrace ? computeTraceSummary(candidateTrace) : void 0;
+  const outputMessages = providerResponse.outputMessages;
+  const traceSummary = outputMessages ? computeTraceSummary(outputMessages) : void 0;
+  const candidate = extractLastAssistantContent(outputMessages);
   try {
     return await evaluateCandidate({
       evalCase,
-      candidate: providerResponse.text ?? "",
+      candidate,
       target,
       provider,
       evaluators,
@@ -38918,9 +39173,8 @@ async function runEvalCase(options) {
       attempt,
       judgeProvider,
       agentTimeoutMs,
-      candidateTrace,
-      candidateTraceRef: providerResponse.traceRef,
-      candidateTraceSummary
+      outputMessages,
+      traceSummary
     });
   } catch (error40) {
     return buildErrorResult(evalCase, target.name, nowFn(), error40, promptInputs, provider);
@@ -38938,9 +39192,8 @@ async function evaluateCandidate(options) {
     attempt,
     judgeProvider,
     agentTimeoutMs,
-    candidateTrace,
-    candidateTraceRef,
-    candidateTraceSummary
+    outputMessages,
+    traceSummary
   } = options;
   const gradeTimestamp = nowFn();
   const { score, evaluatorResults } = await runEvaluatorsForCase({
@@ -38954,9 +39207,8 @@ async function evaluateCandidate(options) {
     now: gradeTimestamp,
     judgeProvider,
     agentTimeoutMs,
-    candidateTrace,
-    candidateTraceRef,
-    candidateTraceSummary
+    outputMessages,
+    traceSummary
   });
   const completedAt = nowFn();
   let agentProviderRequest;
@@ -38994,7 +39246,7 @@ async function evaluateCandidate(options) {
     lm_provider_request: lmProviderRequest,
     evaluator_provider_request: evaluatorResults ? void 0 : score.evaluatorRawRequest,
     evaluator_results: evaluatorResults,
-    trace_summary: candidateTraceSummary
+    trace_summary: traceSummary
   };
 }
 async function runEvaluatorsForCase(options) {
@@ -39009,9 +39261,8 @@ async function runEvaluatorsForCase(options) {
     now,
     judgeProvider,
     agentTimeoutMs,
-    candidateTrace,
-    candidateTraceRef,
-    candidateTraceSummary
+    outputMessages,
+    traceSummary
   } = options;
   if (evalCase.evaluators && evalCase.evaluators.length > 0) {
     return runEvaluatorList({
@@ -39026,9 +39277,8 @@ async function runEvaluatorsForCase(options) {
       now,
       judgeProvider,
       agentTimeoutMs,
-      candidateTrace,
-      candidateTraceRef,
-      candidateTraceSummary
+      outputMessages,
+      traceSummary
     });
   }
   const evaluatorKind = evalCase.evaluator ?? "llm_judge";
@@ -39045,9 +39295,8 @@ async function runEvaluatorsForCase(options) {
     promptInputs,
     now,
     judgeProvider,
-    candidateTrace,
-    candidateTraceRef,
-    candidateTraceSummary
+    outputMessages,
+    traceSummary
   });
   return { score };
 }
@@ -39064,9 +39313,8 @@ async function runEvaluatorList(options) {
     now,
     judgeProvider,
     agentTimeoutMs,
-    candidateTrace,
-    candidateTraceRef,
-    candidateTraceSummary
+    outputMessages,
+    traceSummary
   } = options;
   const scored = [];
   const evaluatorResults = [];
@@ -39113,8 +39361,8 @@ async function runEvaluatorList(options) {
           attempt,
           promptInputs,
           now,
-          candidateTraceRef,
-          candidateTraceSummary
+          outputMessages,
+          traceSummary
         });
         const weight = evaluator.weight ?? 1;
         scored.push({ score: score2, name: evaluator.name, type: "code_judge", weight });
@@ -39200,9 +39448,8 @@ async function runEvaluatorList(options) {
           attempt,
           promptInputs,
           now,
-          candidateTrace,
-          candidateTraceRef,
-          candidateTraceSummary
+          outputMessages,
+          traceSummary
         });
         const weight = evaluator.weight ?? 1;
         scored.push({ score: score2, name: evaluator.name, type: evaluator.type, weight });
@@ -39562,16 +39809,90 @@ function buildPrompt(expectedOutcome, question, referenceAnswer) {
   return parts.join("\n");
 }
 
+// src/commands/convert/index.ts
+import { command, option, optional as optional2, positional, string as string4 } from "cmd-ts";
+import { stringify as stringifyYaml } from "yaml";
+function convertJsonlToYaml(inputPath, outputPath) {
+  const content = readFileSync(inputPath, "utf8");
+  const lines = content.trim().split("\n").filter((line2) => line2.trim());
+  let yamlOutput = "";
+  let isFirst = true;
+  for (const line2 of lines) {
+    const record2 = JSON.parse(line2);
+    const yamlDoc = stringifyYaml(record2, {
+      indent: 2,
+      lineWidth: 0
+    });
+    const normalizedYaml = normalizeLineEndings(yamlDoc);
+    const separator = isFirst ? "---\n" : "\n---\n";
+    isFirst = false;
+    yamlOutput += separator + normalizedYaml;
+  }
+  writeFileSync(outputPath, yamlOutput);
+  return lines.length;
+}
+var convertCommand = command({
+  name: "convert",
+  description: "Convert evaluation results from JSONL to YAML format",
+  args: {
+    input: positional({
+      type: string4,
+      displayName: "input",
+      description: "Path to input JSONL file"
+    }),
+    out: option({
+      type: optional2(string4),
+      long: "out",
+      short: "o",
+      description: "Output file path (defaults to input path with .yaml extension)"
+    })
+  },
+  handler: async ({ input, out }) => {
+    if (!input.endsWith(".jsonl")) {
+      console.error("Error: Input file must be a .jsonl file");
+      process.exit(1);
+    }
+    const outputPath = out ?? input.replace(/\.jsonl$/, ".yaml");
+    try {
+      const count = convertJsonlToYaml(input, outputPath);
+      console.log(`Converted ${count} records to ${path14.resolve(outputPath)}`);
+    } catch (error40) {
+      console.error(`Error: ${error40.message}`);
+      process.exit(1);
+    }
+  }
+});
+
+// src/commands/eval/index.ts
+import { stat as stat4 } from "node:fs/promises";
+import path21 from "node:path";
+import {
+  command as command2,
+  flag,
+  number as number4,
+  option as option2,
+  optional as optional3,
+  restPositionals,
+  string as string5
+} from "cmd-ts";
+import fg from "fast-glob";
+
+// src/commands/eval/run-eval.ts
+import { constants as constants6 } from "node:fs";
+import { access as access6, mkdir as mkdir7 } from "node:fs/promises";
+import path20 from "node:path";
+import { pathToFileURL } from "node:url";
+
 // src/commands/eval/env.ts
 import { constants as constants4 } from "node:fs";
 import { access as access4 } from "node:fs/promises";
-import path14 from "node:path";
+import path15 from "node:path";
 import { config as loadDotenv } from "dotenv";
 function uniqueDirs(directories) {
   const seen = /* @__PURE__ */ new Set();
   const result = [];
   for (const dir of directories) {
-    const absolute = path14.resolve(dir);
+    const absolute = path15.resolve(dir);
     if (seen.has(absolute)) {
       continue;
     }
@@ -39590,14 +39911,14 @@ async function fileExists4(filePath) {
 }
 function collectAncestorDirectories(start, boundary) {
   const directories = [];
-  const boundaryDir = path14.resolve(boundary);
-  let current = path14.resolve(start);
+  const boundaryDir = path15.resolve(boundary);
+  let current = path15.resolve(start);
   while (current !== void 0) {
     directories.push(current);
     if (current === boundaryDir) {
       break;
     }
-    const parent = path14.dirname(current);
+    const parent = path15.dirname(current);
     if (parent === current) {
       break;
     }
@@ -39607,12 +39928,12 @@ function collectAncestorDirectories(start, boundary) {
 }
 async function loadEnvFromHierarchy(options) {
   const { testFilePath, repoRoot, verbose } = options;
-  const testDir = path14.dirname(path14.resolve(testFilePath));
+  const testDir = path15.dirname(path15.resolve(testFilePath));
   const cwd = process.cwd();
   const searchDirs = uniqueDirs([...collectAncestorDirectories(testDir, repoRoot), repoRoot, cwd]);
   const envFiles = [];
   for (const dir of searchDirs) {
-    const candidate = path14.join(dir, ".env");
+    const candidate = path15.join(dir, ".env");
     if (await fileExists4(candidate)) {
       envFiles.push(candidate);
     }
@@ -39636,7 +39957,7 @@ async function loadEnvFromHierarchy(options) {
 // src/commands/eval/jsonl-writer.ts
 import { createWriteStream as createWriteStream2 } from "node:fs";
 import { mkdir as mkdir5 } from "node:fs/promises";
-import path15 from "node:path";
+import path16 from "node:path";
 import { finished } from "node:stream/promises";
 
 // ../../node_modules/.bun/async-mutex@0.5.0/node_modules/async-mutex/index.mjs
@@ -39854,7 +40175,7 @@ var JsonlWriter = class _JsonlWriter {
     this.stream = stream;
   }
   static async open(filePath) {
-    await mkdir5(path15.dirname(filePath), { recursive: true });
+    await mkdir5(path16.dirname(filePath), { recursive: true });
     const stream = createWriteStream2(filePath, { flags: "w", encoding: "utf8" });
     return new _JsonlWriter(stream);
   }
@@ -39886,9 +40207,9 @@ var JsonlWriter = class _JsonlWriter {
 // src/commands/eval/yaml-writer.ts
 import { createWriteStream as createWriteStream3 } from "node:fs";
 import { mkdir as mkdir6 } from "node:fs/promises";
-import path16 from "node:path";
+import path17 from "node:path";
 import { finished as finished2 } from "node:stream/promises";
-import { stringify as stringifyYaml } from "yaml";
+import { stringify as stringifyYaml2 } from "yaml";
 var YamlWriter = class _YamlWriter {
   stream;
   mutex = new Mutex();
@@ -39898,7 +40219,7 @@ var YamlWriter = class _YamlWriter {
     this.stream = stream;
   }
   static async open(filePath) {
-    await mkdir6(path16.dirname(filePath), { recursive: true });
+    await mkdir6(path17.dirname(filePath), { recursive: true });
     const stream = createWriteStream3(filePath, { flags: "w", encoding: "utf8" });
     return new _YamlWriter(stream);
   }
@@ -39907,7 +40228,7 @@ var YamlWriter = class _YamlWriter {
       if (this.closed) {
         throw new Error("Cannot write to closed YAML writer");
       }
-      const yamlDoc = stringifyYaml(record2, {
+      const yamlDoc = stringifyYaml2(record2, {
         indent: 2,
         lineWidth: 0
         // Disable line wrapping
@@ -39963,196 +40284,86 @@ function getDefaultExtension(format) {
 }
 
 // src/commands/eval/progress-display.ts
-import { stripVTControlCharacters } from "node:util";
-var ESC = "\x1B[";
-var CLEAR_LINE = `${ESC}K`;
-var MOVE_CURSOR_UP = `${ESC}1A`;
 var ProgressDisplay = class {
   workers = /* @__PURE__ */ new Map();
-  maxWorkers;
   totalTests = 0;
   completedTests = 0;
-  renderTimer;
-  renderScheduled = false;
-  isInteractive;
   logPaths = [];
   logPathSet = /* @__PURE__ */ new Set();
   hasPrintedLogHeader = false;
-  windowHeight = 0;
   started = false;
   finished = false;
-  constructor(maxWorkers) {
-    this.maxWorkers = maxWorkers;
-    this.isInteractive = process.stdout.isTTY && !process.env.CI;
+  verbose;
+  constructor(_maxWorkers, options) {
+    this.verbose = options?.verbose ?? false;
   }
   isInteractiveMode() {
-    return this.isInteractive;
+    return false;
   }
   start() {
     this.started = true;
     this.finished = false;
-    if (this.isInteractive) {
-      this.write("\n");
-      this.renderTimer = setInterval(() => {
-        this.scheduleRender();
-      }, 1e3);
-      this.renderTimer.unref?.();
-    }
   }
   setTotalTests(count) {
     this.totalTests = count;
   }
   updateWorker(progress) {
+    const previous = this.workers.get(progress.workerId);
     this.workers.set(progress.workerId, progress);
     if (progress.status === "completed" || progress.status === "failed") {
       this.completedTests++;
     }
-    if (this.isInteractive) {
-      this.scheduleRender();
-    } else {
-      const targetSuffix = progress.targetLabel ? ` | ${progress.targetLabel}` : "";
-      if (progress.status === "completed") {
-        console.log(`\u2713 Eval ${progress.evalId}${targetSuffix} completed`);
-      } else if (progress.status === "failed") {
+    const targetSuffix = progress.targetLabel ? ` | ${progress.targetLabel}` : "";
+    const countPrefix = `${this.completedTests}/${this.totalTests}`;
+    switch (progress.status) {
+      case "pending":
+        if (this.verbose && !previous) {
+          console.log(`${countPrefix}   \u23F3 ${progress.evalId}${targetSuffix}`);
+        }
+        break;
+      case "running":
+        if (!previous || previous.status === "pending") {
+          console.log(`${countPrefix}   \u{1F504} ${progress.evalId}${targetSuffix}`);
+        }
+        break;
+      case "completed":
+        console.log(`${countPrefix}   \u2705 ${progress.evalId}${targetSuffix}`);
+        break;
+      case "failed":
         console.log(
-          `\u2717 Eval ${progress.evalId}${targetSuffix} failed${progress.error ? `: ${progress.error}` : ""}`
+          `${countPrefix}   \u274C ${progress.evalId}${targetSuffix}${progress.error ? `: ${progress.error}` : ""}`
         );
-      }
+        break;
     }
   }
   addLogPaths(paths) {
     const newPaths = [];
-    for (const path27 of paths) {
-      if (this.logPathSet.has(path27)) {
+    for (const path28 of paths) {
+      if (this.logPathSet.has(path28)) {
         continue;
       }
-      this.logPathSet.add(path27);
-      newPaths.push(path27);
+      this.logPathSet.add(path28);
+      newPaths.push(path28);
     }
     if (newPaths.length === 0) {
       return;
     }
     this.logPaths.push(...newPaths);
-    if (this.isInteractive) {
-      this.scheduleRender();
-      return;
-    }
     if (!this.hasPrintedLogHeader) {
       console.log("");
       console.log("Codex CLI logs:");
       this.hasPrintedLogHeader = true;
     }
     const startIndex = this.logPaths.length - newPaths.length;
-    newPaths.forEach((path27, offset) => {
-      console.log(`${startIndex + offset + 1}. ${path27}`);
+    newPaths.forEach((path28, offset) => {
+      console.log(`${startIndex + offset + 1}. ${path28}`);
     });
   }
-  scheduleRender() {
-    if (this.renderScheduled || this.finished) {
-      return;
-    }
-    this.renderScheduled = true;
-    setTimeout(() => {
-      this.renderScheduled = false;
-      this.render();
-    }, 100);
-  }
-  write(content) {
-    process.stdout.write(content);
-  }
-  clearWindow() {
-    if (this.windowHeight === 0) {
-      return;
-    }
-    this.write(`\r${CLEAR_LINE}`);
-    for (let i = 1; i < this.windowHeight; i++) {
-      this.write(`${MOVE_CURSOR_UP}\r${CLEAR_LINE}`);
-    }
-    this.windowHeight = 0;
-  }
-  getRenderedRowCount(rows) {
-    const columns = process.stdout.columns || 80;
-    let count = 0;
-    for (const row of rows) {
-      const text2 = stripVTControlCharacters(row);
-      count += Math.max(1, Math.ceil(text2.length / columns));
-    }
-    return count;
-  }
-  render() {
-    if (!this.isInteractive || !this.started || this.finished) {
-      return;
-    }
-    const lines = [];
-    const sortedWorkers = Array.from(this.workers.values()).sort((a, b) => a.workerId - b.workerId);
-    for (const worker of sortedWorkers) {
-      const line2 = this.formatWorkerLine(worker);
-      lines.push(line2);
-    }
-    if (this.logPaths.length > 0) {
-      lines.push("");
-      lines.push("Codex CLI logs:");
-      this.logPaths.forEach((path27, index) => {
-        lines.push(`${index + 1}. ${path27}`);
-      });
-    }
-    const rowCount = this.getRenderedRowCount(lines);
-    this.clearWindow();
-    if (lines.length > 0) {
-      this.write(lines.join("\n"));
-    }
-    this.windowHeight = rowCount;
-  }
-  formatWorkerLine(worker) {
-    const workerLabel = `${worker.workerId}.`.padEnd(4);
-    const statusIcon = this.getStatusIcon(worker.status);
-    const targetLabel = worker.targetLabel ? `  | ${worker.targetLabel}` : "";
-    const columns = process.stdout.columns || 80;
-    const maxLineLength = Math.max(40, columns - 4);
-    const reservedLength = workerLabel.length + statusIcon.length + targetLabel.length + 4;
-    const availableLabelLength = Math.max(15, maxLineLength - reservedLength);
-    let testLabel = worker.evalId;
-    if (testLabel.length > availableLabelLength) {
-      testLabel = `${testLabel.substring(0, Math.max(0, availableLabelLength - 3))}...`;
-    }
-    return `${workerLabel} ${statusIcon} ${testLabel}${targetLabel}`;
-  }
-  getStatusIcon(status) {
-    switch (status) {
-      case "pending":
-        return "\u23F3";
-      case "running":
-        return "\u{1F504}";
-      case "completed":
-        return "\u2705";
-      case "failed":
-        return "\u274C";
-      default:
-        return "  ";
-    }
-  }
   finish() {
-    if (this.renderTimer) {
-      clearInterval(this.renderTimer);
-      this.renderTimer = void 0;
-    }
     this.finished = true;
-    if (this.isInteractive && this.started) {
-      this.clearWindow();
-      const sortedWorkers = Array.from(this.workers.values()).sort(
-        (a, b) => a.workerId - b.workerId
-      );
-      for (const worker of sortedWorkers) {
-        this.write(`${this.formatWorkerLine(worker)}
-`);
-      }
-      this.write("\n");
-    }
+    console.log("");
   }
   clear() {
-    if (this.isInteractive) {
-      this.clearWindow();
-    }
   }
 };
 
@@ -40300,7 +40511,7 @@ function formatEvaluationSummary(summary) {
 
 // ../../packages/core/dist/evaluation/validation/index.js
 import { readFile as readFile7 } from "node:fs/promises";
-import path17 from "node:path";
+import path18 from "node:path";
 import { parse as parse6 } from "yaml";
 import { readFile as readFile23 } from "node:fs/promises";
 import path23 from "node:path";
@@ -40343,8 +40554,8 @@ async function detectFileType(filePath) {
   }
 }
 function inferFileTypeFromPath(filePath) {
-  const normalized = path17.normalize(filePath).replace(/\\/g, "/");
-  const basename = path17.basename(filePath);
+  const normalized = path18.normalize(filePath).replace(/\\/g, "/");
+  const basename = path18.basename(filePath);
   if (normalized.includes("/.agentv/")) {
     if (basename === "config.yaml" || basename === "config.yml") {
       return "config";
@@ -40656,6 +40867,9 @@ var CLI_SETTINGS = /* @__PURE__ */ new Set([
   ...COMMON_SETTINGS,
   "command_template",
   "commandTemplate",
+  "verbose",
+  "cli_verbose",
+  "cliVerbose",
   "files_format",
   "filesFormat",
   "attachments_format",
@@ -40664,7 +40878,11 @@ var CLI_SETTINGS = /* @__PURE__ */ new Set([
   "env",
   "timeout_seconds",
   "timeoutSeconds",
-  "healthcheck"
+  "healthcheck",
+  "keep_temp_files",
+  "keepTempFiles",
+  "keep_output_files",
+  "keepOutputFiles"
 ]);
 function getKnownSettings(provider) {
   const normalizedProvider = provider.toLowerCase();
@@ -40789,6 +41007,15 @@ async function validateTargetsFile(filePath) {
     if (healthcheck !== void 0) {
       validateCliHealthcheck(healthcheck, absolutePath2, `${location}.healthcheck`, errors2);
     }
+    const verbose = target.verbose ?? target.cli_verbose ?? target.cliVerbose;
+    if (verbose !== void 0 && typeof verbose !== "boolean") {
+      errors2.push({
+        severity: "error",
+        filePath: absolutePath2,
+        location: `${location}.verbose`,
+        message: "'verbose' must be a boolean when provided"
+      });
+    }
   }
   function validateCliHealthcheck(healthcheck, absolutePath2, location, errors2) {
     if (!isObject22(healthcheck)) {
@@ -41173,12 +41400,12 @@ async function validateMessagesFileRefs(messages, location, searchRoots, filePat
 // src/utils/targets.ts
 import { constants as constants5 } from "node:fs";
 import { access as access5 } from "node:fs/promises";
-import path18 from "node:path";
+import path19 from "node:path";
 var TARGET_FILE_CANDIDATES = [
   "targets.yaml",
   "targets.yml",
-  path18.join(".agentv", "targets.yaml"),
-  path18.join(".agentv", "targets.yml")
+  path19.join(".agentv", "targets.yaml"),
+  path19.join(".agentv", "targets.yml")
 ];
 async function fileExists5(filePath) {
   try {
@@ -41191,12 +41418,12 @@ async function fileExists5(filePath) {
 async function discoverTargetsFile(options) {
   const { explicitPath, testFilePath, repoRoot, cwd } = options;
   if (explicitPath) {
-    const resolvedExplicit = path18.resolve(explicitPath);
+    const resolvedExplicit = path19.resolve(explicitPath);
     if (await fileExists5(resolvedExplicit)) {
       return resolvedExplicit;
     }
     for (const candidate of TARGET_FILE_CANDIDATES) {
-      const nested = path18.join(resolvedExplicit, candidate);
+      const nested = path19.join(resolvedExplicit, candidate);
       if (await fileExists5(nested)) {
         return nested;
       }
@@ -41204,13 +41431,13 @@ async function discoverTargetsFile(options) {
     throw new Error(`targets.yaml not found at provided path: ${resolvedExplicit}`);
   }
   const directories = [...buildDirectoryChain(testFilePath, repoRoot)];
-  const resolvedCwd = path18.resolve(cwd);
+  const resolvedCwd = path19.resolve(cwd);
   if (!directories.includes(resolvedCwd)) {
     directories.push(resolvedCwd);
   }
   for (const directory of directories) {
     for (const candidate of TARGET_FILE_CANDIDATES) {
-      const fullPath = path18.join(directory, candidate);
+      const fullPath = path19.join(directory, candidate);
       if (await fileExists5(fullPath)) {
         return fullPath;
       }
@@ -41389,15 +41616,15 @@ async function ensureFileExists(filePath, description) {
   }
 }
 async function findRepoRoot(start) {
-  const fallback = path19.resolve(start);
+  const fallback = path20.resolve(start);
   let current = fallback;
   while (current !== void 0) {
-    const candidate = path19.join(current, ".git");
+    const candidate = path20.join(current, ".git");
     try {
       await access6(candidate, constants6.F_OK);
       return current;
     } catch {
-      const parent = path19.dirname(current);
+      const parent = path20.dirname(current);
       if (parent === current) {
         break;
       }
@@ -41410,16 +41637,16 @@ function buildDefaultOutputPath(cwd, format) {
   const timestamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
   const baseName = "eval";
   const extension = getDefaultExtension(format);
-  return path19.join(cwd, ".agentv", "results", `${baseName}_${timestamp}${extension}`);
+  return path20.join(cwd, ".agentv", "results", `${baseName}_${timestamp}${extension}`);
 }
-function resolvePromptDirectory(option4, cwd) {
-  if (option4 === void 0) {
+function resolvePromptDirectory(option5, cwd) {
+  if (option5 === void 0) {
     return void 0;
   }
-  if (typeof option4 === "string" && option4.trim().length > 0) {
-    return path19.resolve(cwd, option4);
+  if (typeof option5 === "string" && option5.trim().length > 0) {
+    return path20.resolve(cwd, option5);
   }
-  return path19.join(cwd, ".agentv", "prompts");
+  return path20.join(cwd, ".agentv", "prompts");
 }
 function createEvaluationCache() {
   const store = /* @__PURE__ */ new Map();
@@ -41432,8 +41659,8 @@ function createEvaluationCache() {
     }
   };
 }
-function createProgressReporter(maxWorkers) {
-  const display = new ProgressDisplay(maxWorkers);
+function createProgressReporter(maxWorkers, options) {
+  const display = new ProgressDisplay(maxWorkers, options);
   return {
     isInteractive: display.isInteractiveMode(),
     start: () => display.start(),
@@ -41444,7 +41671,7 @@ function createProgressReporter(maxWorkers) {
   };
 }
 function makeEvalKey(testFilePath, evalId) {
-  return `${path19.resolve(testFilePath)}::${evalId}`;
+  return `${path20.resolve(testFilePath)}::${evalId}`;
 }
 function createDisplayIdTracker() {
   const map2 = /* @__PURE__ */ new Map();
@@ -41461,6 +41688,22 @@ function createDisplayIdTracker() {
     }
   };
 }
+function applyVerboseOverride(selection, cliVerbose) {
+  const { resolvedTarget } = selection;
+  if (resolvedTarget.kind !== "cli") {
+    return selection;
+  }
+  return {
+    ...selection,
+    resolvedTarget: {
+      ...resolvedTarget,
+      config: {
+        ...resolvedTarget.config,
+        verbose: cliVerbose
+      }
+    }
+  };
+}
 async function prepareFileMetadata(params) {
   const { testFilePath, repoRoot, cwd, options } = params;
   await ensureFileExists(testFilePath, "Test file");
@@ -41520,7 +41763,7 @@ async function runSingleEvalFile(params) {
     evalCases
   } = params;
   await ensureFileExists(testFilePath, "Test file");
-  const resolvedTargetSelection = selection;
+  const resolvedTargetSelection = applyVerboseOverride(selection, options.verbose);
   const providerLabel = options.dryRun ? `${resolvedTargetSelection.resolvedTarget.kind} (dry-run)` : resolvedTargetSelection.resolvedTarget.kind;
   const targetMessage = options.verbose ? `Using target (${resolvedTargetSelection.targetSource}): ${resolvedTargetSelection.targetName} [provider=${providerLabel}] via ${resolvedTargetSelection.targetsFilePath}` : `Using target: ${inlineTargetLabel}`;
   if (!progressReporter.isInteractive || options.verbose) {
@@ -41600,7 +41843,7 @@ async function runEvalCommand(input) {
   if (options.verbose) {
     console.log(`Repository root: ${repoRoot}`);
   }
-  const outputPath = options.outPath ? path19.resolve(options.outPath) : buildDefaultOutputPath(cwd, options.format);
+  const outputPath = options.outPath ? path20.resolve(options.outPath) : buildDefaultOutputPath(cwd, options.format);
   console.log(`Output path: ${outputPath}`);
   const outputWriter = await createOutputWriter(outputPath, options.format);
   const cache = options.cache ? createEvaluationCache() : void 0;
@@ -41608,7 +41851,7 @@ async function runEvalCommand(input) {
   const allResults = [];
   let lastPromptDumpDir;
   const seenEvalCases = /* @__PURE__ */ new Set();
-  const resolvedTestFiles = input.testFiles.map((file2) => path19.resolve(file2));
+  const resolvedTestFiles = input.testFiles.map((file2) => path20.resolve(file2));
   const displayIdTracker = createDisplayIdTracker();
   const totalWorkers = options.workers ?? DEFAULT_WORKERS;
   const fileConcurrency = Math.min(
@@ -41633,7 +41876,7 @@ async function runEvalCommand(input) {
   if (totalEvalCount === 0) {
     throw new Error("No eval cases matched the provided filters.");
   }
-  const progressReporter = createProgressReporter(totalWorkers);
+  const progressReporter = createProgressReporter(totalWorkers, { verbose: options.verbose });
   progressReporter.start();
   progressReporter.setTotal(totalEvalCount);
   const seenCodexLogPaths = /* @__PURE__ */ new Set();
@@ -41704,7 +41947,7 @@ async function resolveEvaluationRunner() {
   if (!overridePath) {
     return runEvaluation;
   }
-  const resolved = path19.isAbsolute(overridePath) ? overridePath : path19.resolve(process.cwd(), overridePath);
+  const resolved = path20.isAbsolute(overridePath) ? overridePath : path20.resolve(process.cwd(), overridePath);
   const moduleUrl = pathToFileURL(resolved).href;
   const mod = await import(moduleUrl);
   const candidate = mod.runEvaluation;
@@ -41717,44 +41960,44 @@ async function resolveEvaluationRunner() {
 }
 
 // src/commands/eval/index.ts
-var evalCommand = command({
+var evalCommand = command2({
   name: "eval",
   description: "Run eval suites and report results",
   args: {
     evalPaths: restPositionals({
-      type: string4,
+      type: string5,
       displayName: "eval-paths",
       description: "Path(s) or glob(s) to evaluation .yaml file(s)"
     }),
-    target: option({
-      type: string4,
+    target: option2({
+      type: string5,
       long: "target",
       description: "Override target name from targets.yaml",
       defaultValue: () => "default"
     }),
-    targets: option({
-      type: optional2(string4),
+    targets: option2({
+      type: optional3(string5),
       long: "targets",
       description: "Path to targets.yaml (overrides discovery)"
     }),
-    evalId: option({
-      type: optional2(string4),
+    evalId: option2({
+      type: optional3(string5),
       long: "eval-id",
       description: "Run only the eval case with this identifier"
     }),
-    workers: option({
+    workers: option2({
       type: number4,
       long: "workers",
       description: "Number of parallel workers (default: 3, max: 50). Can also be set per-target in targets.yaml",
       defaultValue: () => 3
     }),
-    out: option({
-      type: optional2(string4),
+    out: option2({
+      type: optional3(string5),
       long: "out",
       description: "Write results to the specified path"
     }),
-    outputFormat: option({
-      type: string4,
+    outputFormat: option2({
+      type: string5,
       long: "output-format",
       description: "Output format: 'jsonl' or 'yaml' (default: jsonl)",
       defaultValue: () => "jsonl"
@@ -41763,31 +42006,31 @@ var evalCommand = command({
       long: "dry-run",
       description: "Use mock provider responses instead of real LLM calls"
     }),
-    dryRunDelay: option({
+    dryRunDelay: option2({
       type: number4,
       long: "dry-run-delay",
       description: "Fixed delay in milliseconds for dry-run mode (overridden by delay range if specified)",
       defaultValue: () => 0
     }),
-    dryRunDelayMin: option({
+    dryRunDelayMin: option2({
       type: number4,
       long: "dry-run-delay-min",
       description: "Minimum delay in milliseconds for dry-run mode (requires --dry-run-delay-max)",
       defaultValue: () => 0
     }),
-    dryRunDelayMax: option({
+    dryRunDelayMax: option2({
       type: number4,
       long: "dry-run-delay-max",
       description: "Maximum delay in milliseconds for dry-run mode (requires --dry-run-delay-min)",
       defaultValue: () => 0
     }),
-    agentTimeout: option({
+    agentTimeout: option2({
       type: number4,
       long: "agent-timeout",
       description: "Timeout in seconds for provider responses (default: 120)",
       defaultValue: () => 120
     }),
-    maxRetries: option({
+    maxRetries: option2({
       type: number4,
       long: "max-retries",
       description: "Retry count for timeout recoveries (default: 2)",
@@ -41801,8 +42044,8 @@ var evalCommand = command({
       long: "verbose",
       description: "Enable verbose logging"
     }),
-    dumpPrompts: option({
-      type: optional2(string4),
+    dumpPrompts: option2({
+      type: optional3(string5),
       long: "dump-prompts",
       description: "Directory path for persisting prompt payloads for debugging"
     }),
@@ -41848,7 +42091,7 @@ async function resolveEvalPaths(evalPaths, cwd) {
   const unmatched = [];
   const results = /* @__PURE__ */ new Set();
   for (const pattern of normalizedInputs) {
-    const candidatePath = path20.isAbsolute(pattern) ? path20.normalize(pattern) : path20.resolve(cwd, pattern);
+    const candidatePath = path21.isAbsolute(pattern) ? path21.normalize(pattern) : path21.resolve(cwd, pattern);
     try {
       const stats = await stat4(candidatePath);
       if (stats.isFile() && /\.ya?ml$/i.test(candidatePath)) {
@@ -41872,7 +42115,7 @@ async function resolveEvalPaths(evalPaths, cwd) {
       continue;
     }
     for (const filePath of yamlMatches) {
-      results.add(path20.normalize(filePath));
+      results.add(path21.normalize(filePath));
     }
   }
   if (unmatched.length > 0) {
@@ -41888,11 +42131,11 @@ async function resolveEvalPaths(evalPaths, cwd) {
 }
 
 // src/commands/generate/index.ts
-import { command as command2, flag as flag2, option as option2, optional as optional3, positional as positional2, string as string5, subcommands } from "cmd-ts";
+import { command as command3, flag as flag2, option as option3, optional as optional4, positional as positional3, string as string6, subcommands } from "cmd-ts";
 
 // src/commands/generate/rubrics.ts
 import { readFile as readFile8, writeFile as writeFile6 } from "node:fs/promises";
-import path21 from "node:path";
+import path24 from "node:path";
 import { pathToFileURL as pathToFileURL2 } from "node:url";
 import { isMap, isSeq, parseDocument } from "yaml";
 function isJsonObject3(value) {
@@ -41904,7 +42147,7 @@ function asString6(value) {
 async function loadRubricGenerator() {
   const customGenerator = process.env.AGENTEVO_CLI_RUBRIC_GENERATOR;
   if (customGenerator) {
-    const generatorPath = path21.resolve(customGenerator);
+    const generatorPath = path24.resolve(customGenerator);
     const generatorUrl = pathToFileURL2(generatorPath).href;
     const module = await import(generatorUrl);
     return module.generateRubrics;
@@ -41914,7 +42157,7 @@ async function loadRubricGenerator() {
 async function generateRubricsCommand(options) {
   const { file: file2, target: targetOverride, verbose } = options;
   console.log(`Generating rubrics for: ${file2}`);
-  const absolutePath = path21.resolve(file2);
+  const absolutePath = path24.resolve(file2);
   const content = await readFile8(absolutePath, "utf8");
   const doc = parseDocument(content);
   const parsed = doc.toJSON();
@@ -42031,17 +42274,17 @@ function extractQuestion(evalCase) {
 }
 
 // src/commands/generate/index.ts
-var rubricsCommand = command2({
+var rubricsCommand = command3({
   name: "rubrics",
   description: "Generate rubrics from expected_outcome in YAML eval file",
   args: {
-    file: positional2({
-      type: string5,
+    file: positional3({
+      type: string6,
       displayName: "file",
       description: "Path to YAML eval file"
     }),
-    target: option2({
-      type: optional3(string5),
+    target: option3({
+      type: optional4(string6),
       long: "target",
       short: "t",
       description: "Override target for rubric generation (default: file target or openai:gpt-4o)"
@@ -42074,14 +42317,14 @@ var generateCommand = subcommands({
 });
 
 // src/commands/init/index.ts
-import { existsSync, mkdirSync, writeFileSync } from "node:fs";
-import path25 from "node:path";
+import { existsSync, mkdirSync, writeFileSync as writeFileSync2 } from "node:fs";
+import path26 from "node:path";
 import * as readline from "node:readline/promises";
-import { command as command3, option as option3, optional as optional4, string as string6 } from "cmd-ts";
+import { command as command4, option as option4, optional as optional5, string as string7 } from "cmd-ts";
 
 // src/templates/index.ts
-import { readFileSync, readdirSync, statSync } from "node:fs";
-import path24 from "node:path";
+import { readFileSync as readFileSync2, readdirSync, statSync } from "node:fs";
+import path25 from "node:path";
 import { fileURLToPath } from "node:url";
 function getGithubTemplates() {
   return getTemplatesFromDir(".github");
@@ -42093,12 +42336,12 @@ function getClaudeTemplates() {
   return getTemplatesFromDir(".claude");
 }
 function getTemplatesFromDir(subdir) {
-  const currentDir = path24.dirname(fileURLToPath(import.meta.url));
+  const currentDir = path25.dirname(fileURLToPath(import.meta.url));
   let templatesDir;
-  if (currentDir.includes(`${path24.sep}dist`)) {
-    templatesDir = path24.join(currentDir, "templates", subdir);
+  if (currentDir.includes(`${path25.sep}dist`)) {
+    templatesDir = path25.join(currentDir, "templates", subdir);
   } else {
-    templatesDir = path24.join(currentDir, subdir);
+    templatesDir = path25.join(currentDir, subdir);
   }
   return readTemplatesRecursively(templatesDir, "");
 }
@@ -42106,15 +42349,15 @@ function readTemplatesRecursively(dir, relativePath) {
   const templates = [];
   const entries = readdirSync(dir);
   for (const entry of entries) {
-    const fullPath = path24.join(dir, entry);
+    const fullPath = path25.join(dir, entry);
     const stat6 = statSync(fullPath);
-    const entryRelativePath = relativePath ? path24.join(relativePath, entry) : entry;
+    const entryRelativePath = relativePath ? path25.join(relativePath, entry) : entry;
     if (stat6.isDirectory()) {
       templates.push(...readTemplatesRecursively(fullPath, entryRelativePath));
     } else {
-      const content = readFileSync(fullPath, "utf-8");
+      const content = readFileSync2(fullPath, "utf-8");
       templates.push({
-        path: entryRelativePath.split(path24.sep).join("/"),
+        path: entryRelativePath.split(path25.sep).join("/"),
         // Normalize to forward slashes
         content
       });
@@ -42137,10 +42380,10 @@ async function promptYesNo(message) {
   }
 }
 async function initCommand(options = {}) {
-  const targetPath = path25.resolve(options.targetPath ?? ".");
-  const githubDir = path25.join(targetPath, ".github");
-  const agentvDir = path25.join(targetPath, ".agentv");
-  const claudeDir = path25.join(targetPath, ".claude");
+  const targetPath = path26.resolve(options.targetPath ?? ".");
+  const githubDir = path26.join(targetPath, ".github");
+  const agentvDir = path26.join(targetPath, ".agentv");
+  const claudeDir = path26.join(targetPath, ".claude");
   const githubTemplates = getGithubTemplates();
   const agentvTemplates = getAgentvTemplates();
   const claudeTemplates = getClaudeTemplates();
@@ -42148,32 +42391,32 @@ async function initCommand(options = {}) {
   const otherAgentvTemplates = agentvTemplates.filter((t) => t.path !== ".env.template");
   const existingFiles = [];
   if (envTemplate) {
-    const envFilePath = path25.join(targetPath, ".env.template");
+    const envFilePath = path26.join(targetPath, ".env.template");
     if (existsSync(envFilePath)) {
       existingFiles.push(".env.template");
     }
   }
   if (existsSync(githubDir)) {
     for (const template of githubTemplates) {
-      const targetFilePath = path25.join(githubDir, template.path);
+      const targetFilePath = path26.join(githubDir, template.path);
       if (existsSync(targetFilePath)) {
-        existingFiles.push(path25.relative(targetPath, targetFilePath));
+        existingFiles.push(path26.relative(targetPath, targetFilePath));
       }
     }
   }
   if (existsSync(agentvDir)) {
     for (const template of otherAgentvTemplates) {
-      const targetFilePath = path25.join(agentvDir, template.path);
+      const targetFilePath = path26.join(agentvDir, template.path);
       if (existsSync(targetFilePath)) {
-        existingFiles.push(path25.relative(targetPath, targetFilePath));
+        existingFiles.push(path26.relative(targetPath, targetFilePath));
       }
     }
   }
   if (existsSync(claudeDir)) {
     for (const template of claudeTemplates) {
-      const targetFilePath = path25.join(claudeDir, template.path);
+      const targetFilePath = path26.join(claudeDir, template.path);
       if (existsSync(targetFilePath)) {
-        existingFiles.push(path25.relative(targetPath, targetFilePath));
+        existingFiles.push(path26.relative(targetPath, targetFilePath));
       }
     }
   }
@@ -42200,36 +42443,36 @@ async function initCommand(options = {}) {
     mkdirSync(claudeDir, { recursive: true });
   }
   if (envTemplate) {
-    const envFilePath = path25.join(targetPath, ".env.template");
-    writeFileSync(envFilePath, envTemplate.content, "utf-8");
+    const envFilePath = path26.join(targetPath, ".env.template");
+    writeFileSync2(envFilePath, envTemplate.content, "utf-8");
     console.log("Created .env.template");
   }
   for (const template of githubTemplates) {
-    const targetFilePath = path25.join(githubDir, template.path);
-    const targetDirPath = path25.dirname(targetFilePath);
+    const targetFilePath = path26.join(githubDir, template.path);
+    const targetDirPath = path26.dirname(targetFilePath);
     if (!existsSync(targetDirPath)) {
       mkdirSync(targetDirPath, { recursive: true });
     }
-    writeFileSync(targetFilePath, template.content, "utf-8");
-    console.log(`Created ${path25.relative(targetPath, targetFilePath)}`);
+    writeFileSync2(targetFilePath, template.content, "utf-8");
+    console.log(`Created ${path26.relative(targetPath, targetFilePath)}`);
   }
   for (const template of otherAgentvTemplates) {
-    const targetFilePath = path25.join(agentvDir, template.path);
-    const targetDirPath = path25.dirname(targetFilePath);
+    const targetFilePath = path26.join(agentvDir, template.path);
+    const targetDirPath = path26.dirname(targetFilePath);
     if (!existsSync(targetDirPath)) {
       mkdirSync(targetDirPath, { recursive: true });
     }
-    writeFileSync(targetFilePath, template.content, "utf-8");
-    console.log(`Created ${path25.relative(targetPath, targetFilePath)}`);
+    writeFileSync2(targetFilePath, template.content, "utf-8");
+    console.log(`Created ${path26.relative(targetPath, targetFilePath)}`);
   }
   for (const template of claudeTemplates) {
-    const targetFilePath = path25.join(claudeDir, template.path);
-    const targetDirPath = path25.dirname(targetFilePath);
+    const targetFilePath = path26.join(claudeDir, template.path);
+    const targetDirPath = path26.dirname(targetFilePath);
     if (!existsSync(targetDirPath)) {
       mkdirSync(targetDirPath, { recursive: true });
     }
-    writeFileSync(targetFilePath, template.content, "utf-8");
-    console.log(`Created ${path25.relative(targetPath, targetFilePath)}`);
+    writeFileSync2(targetFilePath, template.content, "utf-8");
+    console.log(`Created ${path26.relative(targetPath, targetFilePath)}`);
   }
   console.log("\nAgentV initialized successfully!");
   console.log("\nFiles installed to root:");
@@ -42237,17 +42480,17 @@ async function initCommand(options = {}) {
     console.log("  - .env.template");
   }
   console.log(`
-Files installed to ${path25.relative(targetPath, githubDir)}:`);
+Files installed to ${path26.relative(targetPath, githubDir)}:`);
   for (const t of githubTemplates) {
     console.log(`  - ${t.path}`);
   }
   console.log(`
-Files installed to ${path25.relative(targetPath, agentvDir)}:`);
+Files installed to ${path26.relative(targetPath, agentvDir)}:`);
   for (const t of otherAgentvTemplates) {
     console.log(`  - ${t.path}`);
   }
   console.log(`
-Files installed to ${path25.relative(targetPath, claudeDir)}:`);
+Files installed to ${path26.relative(targetPath, claudeDir)}:`);
   for (const t of claudeTemplates) {
     console.log(`  - ${t.path}`);
   }
@@ -42256,12 +42499,12 @@ Files installed to ${path25.relative(targetPath, claudeDir)}:`);
   console.log("  2. Configure targets in .agentv/targets.yaml");
   console.log("  3. Create eval files using the schema and prompt templates");
 }
-var initCmdTsCommand = command3({
+var initCmdTsCommand = command4({
   name: "init",
   description: "Initialize AgentV in your project (installs prompt templates and schema to .github)",
   args: {
-    path: option3({
-      type: optional4(string6),
+    path: option4({
+      type: optional5(string7),
       long: "path",
       description: "Target directory for initialization (default: current directory)"
     })
@@ -42277,7 +42520,7 @@ var initCmdTsCommand = command3({
 });
 
 // src/commands/validate/index.ts
-import { command as command4, restPositionals as restPositionals2, string as string7 } from "cmd-ts";
+import { command as command5, restPositionals as restPositionals2, string as string8 } from "cmd-ts";
 
 // src/commands/validate/format-output.ts
 var ANSI_RED3 = "\x1B[31m";
@@ -42362,7 +42605,7 @@ function isTTY2() {
 // src/commands/validate/validate-files.ts
 import { constants as constants7 } from "node:fs";
 import { access as access7, readdir as readdir3, stat as stat5 } from "node:fs/promises";
-import path26 from "node:path";
+import path27 from "node:path";
 async function validateFiles(paths) {
   const filePaths = await expandPaths(paths);
   const results = [];
@@ -42380,7 +42623,7 @@ async function validateFiles(paths) {
   };
 }
 async function validateSingleFile(filePath) {
-  const absolutePath = path26.resolve(filePath);
+  const absolutePath = path27.resolve(filePath);
   const fileType = await detectFileType(absolutePath);
   let result;
   if (fileType === "eval") {
@@ -42405,7 +42648,7 @@ async function validateSingleFile(filePath) {
 async function expandPaths(paths) {
   const expanded = [];
   for (const inputPath of paths) {
-    const absolutePath = path26.resolve(inputPath);
+    const absolutePath = path27.resolve(inputPath);
     try {
       await access7(absolutePath, constants7.F_OK);
     } catch {
@@ -42429,7 +42672,7 @@ async function findYamlFiles(dirPath) {
   try {
     const entries = await readdir3(dirPath, { withFileTypes: true });
     for (const entry of entries) {
-      const fullPath = path26.join(dirPath, entry.name);
+      const fullPath = path27.join(dirPath, entry.name);
       if (entry.isDirectory()) {
         if (entry.name === "node_modules" || entry.name.startsWith(".")) {
           continue;
@@ -42446,7 +42689,7 @@ async function findYamlFiles(dirPath) {
   return results;
 }
 function isYamlFile(filePath) {
-  const ext = path26.extname(filePath).toLowerCase();
+  const ext = path27.extname(filePath).toLowerCase();
   return ext === ".yaml" || ext === ".yml";
 }
 
@@ -42463,12 +42706,12 @@ async function runValidateCommand(paths) {
     process.exit(1);
   }
 }
-var validateCommand = command4({
+var validateCommand = command5({
   name: "validate",
   description: "Validate AgentV eval and targets YAML files",
   args: {
     paths: restPositionals2({
-      type: string7,
+      type: string8,
       displayName: "paths",
       description: "Files or directories to validate"
     })
@@ -42484,16 +42727,17 @@ var validateCommand = command4({
 });
 
 // src/index.ts
-var packageJson = JSON.parse(readFileSync2(new URL("../package.json", import.meta.url), "utf8"));
+var packageJson = JSON.parse(readFileSync3(new URL("../package.json", import.meta.url), "utf8"));
 var app = subcommands2({
   name: "agentv",
   description: "AgentV CLI",
   version: packageJson.version,
   cmds: {
+    convert: convertCommand,
     eval: evalCommand,
-    validate: validateCommand,
     generate: generateCommand,
-    init: initCmdTsCommand
+    init: initCmdTsCommand,
+    validate: validateCommand
   }
 });
 async function runCli(argv = process.argv) {
@@ -42504,4 +42748,4 @@ export {
   app,
   runCli
 };
-//# sourceMappingURL=chunk-IVIT4U6S.js.map
+//# sourceMappingURL=chunk-3RYQPI4H.js.map