@agentxjs/devtools 1.9.5-dev → 2.0.0

package/dist/bdd/index.js

@@ -0,0 +1,381 @@
+import {
+  env
+} from "../chunk-S7J75AXG.js";
+import {
+  __require
+} from "../chunk-DGUM43GV.js";
+
+// src/bdd/cucumber.config.ts
+function createCucumberConfig(options) {
+  return {
+    format: options.format ?? ["progress"],
+    formatOptions: { snippetInterface: "async-await" },
+    import: options.import,
+    paths: options.paths,
+    tags: options.tags ?? "not @pending and not @skip",
+    worldParameters: {
+      defaultTimeout: options.timeout ?? 3e4
+    }
+  };
+}
+
+// src/bdd/playwright.ts
+import { chromium } from "@playwright/test";
+var browser = null;
+var page = null;
+async function launchBrowser(options = {}) {
+  if (browser) return browser;
+  const headless = options.headless ?? process.env.CI === "true";
+  browser = await chromium.launch({
+    channel: "chrome",
+    headless,
+    slowMo: options.slowMo
+  });
+  return browser;
+}
+async function getPage() {
+  if (page && !page.isClosed()) return page;
+  const b = await launchBrowser();
+  page = await b.newPage();
+  return page;
+}
+async function resetPage() {
+  try {
+    if (page && !page.isClosed()) {
+      const context = page.context();
+      await context.clearCookies();
+      await page.goto("about:blank");
+    }
+  } catch {
+    page = null;
+  }
+}
+async function closePage() {
+  if (page) {
+    await page.close();
+    page = null;
+  }
+}
+async function closeBrowser() {
+  if (page && !page.isClosed()) {
+    await page.close();
+    page = null;
+  }
+  if (browser) {
+    await browser.close();
+    browser = null;
+  }
+}
+async function waitForUrl(url, timeout = 3e4) {
+  const start = Date.now();
+  while (Date.now() - start < timeout) {
+    try {
+      const response = await fetch(url);
+      if (response.ok) {
+        return true;
+      }
+    } catch {
+    }
+    await new Promise((r) => setTimeout(r, 500));
+  }
+  return false;
+}
+
+// src/bdd/dev-server.ts
+import { spawn } from "child_process";
+var devServer = null;
+async function startDevServer(options) {
+  if (devServer) return;
+  const {
+    cwd,
+    command = "bun",
+    args = ["run", "dev"],
+    port,
+    timeout = 3e4,
+    debug = !!process.env.DEBUG
+  } = options;
+  devServer = spawn(command, args, {
+    cwd,
+    stdio: "pipe",
+    detached: false
+  });
+  if (debug) {
+    devServer.stdout?.on("data", (data) => {
+      console.log("[dev-server]", data.toString());
+    });
+    devServer.stderr?.on("data", (data) => {
+      console.error("[dev-server error]", data.toString());
+    });
+  }
+  const url = `http://localhost:${port}`;
+  const ready = await waitForUrl(url, timeout);
+  if (!ready) {
+    stopDevServer();
+    throw new Error(`Dev server failed to start on port ${port}`);
+  }
+}
+function stopDevServer() {
+  if (devServer) {
+    devServer.kill("SIGTERM");
+    devServer = null;
+  }
+}
+function getDevServer() {
+  return devServer;
+}
+
+// src/bdd/paths.ts
+import { resolve, dirname } from "path";
+import { existsSync, mkdtempSync, mkdirSync } from "fs";
+import { tmpdir } from "os";
+function findMonorepoRoot(startDir = process.cwd()) {
+  let dir = startDir;
+  while (dir !== "/") {
+    const pkgPath = resolve(dir, "package.json");
+    if (existsSync(pkgPath)) {
+      try {
+        const pkg = __require(pkgPath);
+        if (pkg.workspaces || pkg.private === true) {
+          const hasPackages = existsSync(resolve(dir, "packages"));
+          const hasApps = existsSync(resolve(dir, "apps"));
+          if (hasPackages || hasApps) {
+            return dir;
+          }
+        }
+      } catch {
+      }
+    }
+    dir = dirname(dir);
+  }
+  return startDir;
+}
+function getPackageRoot(startDir = process.cwd()) {
+  let dir = startDir;
+  while (dir !== "/") {
+    if (existsSync(resolve(dir, "package.json"))) {
+      return dir;
+    }
+    dir = dirname(dir);
+  }
+  return startDir;
+}
+var _monorepoRoot = null;
+var _packageRoot = null;
+var _tempDir = null;
+function getMonorepoPath() {
+  if (!_monorepoRoot) {
+    _monorepoRoot = findMonorepoRoot();
+  }
+  return _monorepoRoot;
+}
+function getPackagePath() {
+  if (!_packageRoot) {
+    _packageRoot = getPackageRoot();
+  }
+  return _packageRoot;
+}
+function getBddPath() {
+  return resolve(getPackagePath(), "bdd");
+}
+function getFixturesPath(subdir) {
+  const base = resolve(getBddPath(), "fixtures");
+  return subdir ? resolve(base, subdir) : base;
+}
+function getTempPath(prefix = "bdd-") {
+  if (!_tempDir) {
+    _tempDir = mkdtempSync(resolve(tmpdir(), prefix));
+  }
+  return _tempDir;
+}
+function ensureDir(path) {
+  if (!existsSync(path)) {
+    mkdirSync(path, { recursive: true });
+  }
+  return path;
+}
+function resetPaths() {
+  _monorepoRoot = null;
+  _packageRoot = null;
+  _tempDir = null;
+}
+var paths = {
+  monorepo: getMonorepoPath,
+  package: getPackagePath,
+  bdd: getBddPath,
+  fixtures: getFixturesPath,
+  temp: getTempPath,
+  ensure: ensureDir,
+  reset: resetPaths
+};
+
+// src/bdd/agent-ui-tester.ts
+import { execFileSync } from "child_process";
+import { readFileSync } from "fs";
+import { resolve as resolve2, dirname as dirname2 } from "path";
+import { fileURLToPath } from "url";
+var __dirname = dirname2(fileURLToPath(import.meta.url));
+var SKILL_PATH = resolve2(__dirname, "../../../../.claude/skills/agent-browser/SKILL.md");
+function loadSystemPrompt(headed = false) {
+  let skillContent = "";
+  try {
+    skillContent = readFileSync(SKILL_PATH, "utf-8");
+  } catch {
+  }
+  return `You are a UI tester. You test web application scenarios using the agent-browser CLI.
+
+RULES:
+- ONLY use agent-browser commands via Bash tool
+- Use ${headed ? "--headed " : ""}--executable-path "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" for all commands
+- After each navigation or click, run: agent-browser snapshot -i
+- Refs (@e1, @e2) are invalidated after page changes \u2014 always re-snapshot
+- At the end, close the browser with: agent-browser close
+- Output your result as a single line: PASS or FAIL followed by a brief reason
+
+${skillContent ? `AGENT-BROWSER REFERENCE:
+${skillContent}` : ""}`;
+}
+function agentUiTester(prompt, options = {}) {
+  const { model = "haiku", baseUrl, timeout = 3e5, headed = false } = options;
+  const fullPrompt = baseUrl ? `Base URL: ${baseUrl}
+
+${prompt}` : prompt;
+  const systemPrompt = loadSystemPrompt(headed);
+  const cleanEnv = Object.fromEntries(
+    Object.entries(process.env).filter(([k]) => !k.startsWith("CLAUDE"))
+  );
+  try {
+    const output = execFileSync(
+      "claude",
+      [
+        "-p",
+        fullPrompt,
+        "--model",
+        model,
+        "--append-system-prompt",
+        systemPrompt,
+        "--allowedTools",
+        "Bash(agent-browser:*)"
+      ],
+      {
+        encoding: "utf-8",
+        timeout,
+        env: cleanEnv,
+        maxBuffer: 10 * 1024 * 1024
+      }
+    ).trim();
+    const passed = /\*{0,2}PASS\*{0,2}\b/m.test(output);
+    return { passed, output };
+  } catch (error) {
+    const output = error.stdout || error.stderr || error.message || "Unknown error";
+    return { passed: false, output: output.trim() };
+  }
+}
+
+// src/bdd/agent-doc-tester.ts
+import { readFileSync as readFileSync2, existsSync as existsSync2 } from "fs";
+var SYSTEM_PROMPT = `You are a documentation reviewer evaluating documents from the reader's experience.
+
+EVALUATION DIMENSIONS:
+1. Completeness \u2014 All required information is present. Nothing critical is missing.
+2. Logic \u2014 Structure flows naturally. Concepts build on each other without jumps.
+3. Readability \u2014 A newcomer can follow without confusion. No unexplained jargon.
+
+RULES:
+- Read the provided document carefully
+- Evaluate each requirement listed in the prompt against ALL three dimensions
+- Be strict but fair \u2014 the document should genuinely help the reader achieve the stated goal
+- Output your result as a single line: PASS or FAIL followed by a brief reason
+- If FAIL, list which specific requirements are not met and which dimension they violate`;
+async function agentDocTester(options, testerOptions = {}) {
+  const {
+    provider = process.env.AGENTX_PROVIDER || "anthropic",
+    model = env.model,
+    apiKey = env.apiKey || "",
+    baseUrl = env.baseUrl,
+    timeout = 12e4
+  } = testerOptions;
+  const docContents = options.files.map((filePath) => {
+    if (!existsSync2(filePath)) {
+      return `--- ${filePath} ---
+[FILE NOT FOUND]`;
+    }
+    return `--- ${filePath} ---
+${readFileSync2(filePath, "utf-8")}`;
+  }).join("\n\n");
+  const userPrompt = [
+    "Evaluate the following document(s) against the requirements below.",
+    "",
+    "DOCUMENTS:",
+    docContents,
+    "",
+    "REQUIREMENTS:",
+    options.requirements,
+    "",
+    "Evaluate each requirement. Output PASS if all are met, FAIL if any are not."
+  ].join("\n");
+  const moduleName = "agentxjs";
+  const agentxjs = await import(
+    /* @vite-ignore */
+    moduleName
+  );
+  const createAgentX = agentxjs.createAgentX;
+  let agentx = null;
+  try {
+    agentx = await createAgentX({
+      apiKey,
+      provider,
+      model,
+      baseUrl,
+      logLevel: "silent"
+    });
+    await agentx.containers.create("doc-tester");
+    const { record: image } = await agentx.images.create({
+      containerId: "doc-tester",
+      systemPrompt: SYSTEM_PROMPT
+    });
+    const { agentId } = await agentx.agents.create({ imageId: image.imageId });
+    let output = "";
+    agentx.on("text_delta", (e) => {
+      output += e.data.text;
+    });
+    await Promise.race([
+      agentx.sessions.send(agentId, userPrompt),
+      new Promise((_, reject) => setTimeout(() => reject(new Error("Timeout")), timeout))
+    ]);
+    output = output.trim();
+    const passed = /\*{0,2}PASS\*{0,2}\b/m.test(output);
+    return { passed, output };
+  } catch (error) {
+    return { passed: false, output: error.message || "Unknown error" };
+  } finally {
+    if (agentx) {
+      try {
+        await agentx.shutdown();
+      } catch {
+      }
+    }
+  }
+}
+export {
+  agentDocTester,
+  agentUiTester,
+  closeBrowser,
+  closePage,
+  createCucumberConfig,
+  ensureDir,
+  getBddPath,
+  getDevServer,
+  getFixturesPath,
+  getMonorepoPath,
+  getPackagePath,
+  getPage,
+  getTempPath,
+  launchBrowser,
+  paths,
+  resetPage,
+  resetPaths,
+  startDevServer,
+  stopDevServer,
+  waitForUrl
+};
+//# sourceMappingURL=index.js.map

package/dist/bdd/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/bdd/cucumber.config.ts","../../src/bdd/playwright.ts","../../src/bdd/dev-server.ts","../../src/bdd/paths.ts","../../src/bdd/agent-ui-tester.ts","../../src/bdd/agent-doc-tester.ts"],"sourcesContent":["/**\n * Shared Cucumber configuration for BDD tests\n *\n * Usage in project's cucumber.js:\n *\n * ```js\n * import { createCucumberConfig } from \"@agentxjs/devtools/bdd\";\n *\n * export default createCucumberConfig({\n *   paths: [\"bdd/journeys/** /*.feature\"],\n *   import: [\"bdd/steps/** /*.ts\"],\n * });\n * ```\n */\n\nexport interface CucumberConfigOptions {\n  /** Feature file paths */\n  paths: string[];\n  /** Step definition paths */\n  import: string[];\n  /** Tags to filter (default: exclude @pending and @skip) */\n  tags?: string;\n  /** Default timeout in ms (default: 30000) */\n  timeout?: number;\n  /** Format output (default: progress) */\n  format?: string[];\n}\n\nexport function createCucumberConfig(options: CucumberConfigOptions) {\n  return {\n    format: options.format ?? [\"progress\"],\n    formatOptions: { snippetInterface: \"async-await\" },\n    import: options.import,\n    paths: options.paths,\n    tags: options.tags ?? \"not @pending and not @skip\",\n    worldParameters: {\n      defaultTimeout: options.timeout ?? 30000,\n    },\n  };\n}\n","/**\n * Playwright utilities for BDD testing\n *\n * Uses system Chrome to avoid downloading Chromium.\n * Install: PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 bun add -d @playwright/test\n *\n * Browser lifecycle:\n * - Single browser instance for all tests\n * - Single page (tab) reused across scenarios\n * - resetPage() clears state between scenarios\n */\n\nimport { chromium, Browser, Page } from \"@playwright/test\";\n\nexport interface BrowserOptions {\n  headless?: boolean;\n  slowMo?: number;\n}\n\nlet browser: Browser | null = null;\nlet page: Page | null = null;\n\n/**\n * Launch browser using system Chrome (singleton)\n */\nexport async function launchBrowser(options: BrowserOptions = {}): Promise<Browser> {\n  if (browser) return browser;\n\n  const headless = options.headless ?? process.env.CI === \"true\";\n\n  browser = await chromium.launch({\n    channel: \"chrome\",\n    headless,\n    slowMo: options.slowMo,\n  });\n\n  return browser;\n}\n\n/**\n * Get or create a page (singleton, reused across scenarios)\n */\nexport async function getPage(): Promise<Page> {\n  if (page && !page.isClosed()) return page;\n\n  const b = await launchBrowser();\n  page = await b.newPage();\n  return page;\n}\n\n/**\n * Reset page state between scenarios (without closing)\n * Use this instead of closePage() for faster tests\n */\nexport async function resetPage(): Promise<void> {\n  try {\n    if (page && !page.isClosed()) {\n      const context = page.context();\n      await context.clearCookies();\n      await page.goto(\"about:blank\");\n    }\n  } catch {\n    // Browser may have crashed, will recreate on next getPage()\n    page = null;\n  }\n}\n\n/**\n * Close current page\n * @deprecated Use resetPage() for faster tests. Only use closePage() if you need full isolation.\n */\nexport async function closePage(): Promise<void> {\n  if (page) {\n    await page.close();\n    page = null;\n  }\n}\n\n/**\n * Close browser and cleanup\n */\nexport async function closeBrowser(): Promise<void> {\n  if (page && !page.isClosed()) {\n    await page.close();\n    page = null;\n  }\n  if (browser) {\n    await browser.close();\n    browser = null;\n  }\n}\n\n/**\n * Wait for a URL to be accessible\n */\nexport async function waitForUrl(url: string, timeout = 30000): Promise<boolean> {\n  const start = Date.now();\n  while (Date.now() - start < timeout) {\n    try {\n      const response = await fetch(url);\n      if (response.ok) {\n        return true;\n      }\n    } catch {\n      // Not ready yet\n    }\n    await new Promise((r) => setTimeout(r, 500));\n  }\n  return false;\n}\n","/**\n * Dev server utilities for BDD testing\n *\n * Start and stop dev servers during test runs.\n */\n\nimport { spawn, ChildProcess } from \"node:child_process\";\nimport { waitForUrl } from \"./playwright\";\n\nexport interface DevServerOptions {\n  /** Working directory */\n  cwd: string;\n  /** Command to run (default: \"bun\") */\n  command?: string;\n  /** Command arguments (default: [\"run\", \"dev\"]) */\n  args?: string[];\n  /** Port to wait for */\n  port: number;\n  /** Startup timeout in ms (default: 30000) */\n  timeout?: number;\n  /** Show server output (default: false, or true if DEBUG env is set) */\n  debug?: boolean;\n}\n\nlet devServer: ChildProcess | null = null;\n\n/**\n * Start a dev server and wait for it to be ready\n */\nexport async function startDevServer(options: DevServerOptions): Promise<void> {\n  if (devServer) return;\n\n  const {\n    cwd,\n    command = \"bun\",\n    args = [\"run\", \"dev\"],\n    port,\n    timeout = 30000,\n    debug = !!process.env.DEBUG,\n  } = options;\n\n  devServer = spawn(command, args, {\n    cwd,\n    stdio: \"pipe\",\n    detached: false,\n  });\n\n  if (debug) {\n    devServer.stdout?.on(\"data\", (data) => {\n      console.log(\"[dev-server]\", data.toString());\n    });\n\n    devServer.stderr?.on(\"data\", (data) => {\n      console.error(\"[dev-server error]\", data.toString());\n    });\n  }\n\n  const url = `http://localhost:${port}`;\n  const ready = await waitForUrl(url, timeout);\n\n  if (!ready) {\n    stopDevServer();\n    throw new Error(`Dev server failed to start on port ${port}`);\n  }\n}\n\n/**\n * Stop the dev server\n */\nexport function stopDevServer(): void {\n  if (devServer) {\n    devServer.kill(\"SIGTERM\");\n    devServer = null;\n  }\n}\n\n/**\n * Get the dev server process (for advanced use)\n */\nexport function getDevServer(): ChildProcess | null {\n  return devServer;\n}\n","/**\n * Unified path utilities for BDD testing\n *\n * Provides consistent path resolution across all packages.\n */\n\nimport { resolve, dirname } from \"node:path\";\nimport { existsSync, mkdtempSync, mkdirSync } from \"node:fs\";\nimport { tmpdir } from \"node:os\";\n\n// ============================================================================\n// Path Resolution\n// ============================================================================\n\n/**\n * Find the monorepo root by looking for root package.json with workspaces\n */\nexport function findMonorepoRoot(startDir: string = process.cwd()): string {\n  let dir = startDir;\n  while (dir !== \"/\") {\n    const pkgPath = resolve(dir, \"package.json\");\n    if (existsSync(pkgPath)) {\n      try {\n        const pkg = require(pkgPath);\n        if (pkg.workspaces || pkg.private === true) {\n          // Check if it looks like a monorepo root\n          const hasPackages = existsSync(resolve(dir, \"packages\"));\n          const hasApps = existsSync(resolve(dir, \"apps\"));\n          if (hasPackages || hasApps) {\n            return dir;\n          }\n        }\n      } catch {\n        // Ignore errors\n      }\n    }\n    dir = dirname(dir);\n  }\n  return startDir;\n}\n\n/**\n * Get the current package root (where package.json is)\n */\nexport function getPackageRoot(startDir: string = process.cwd()): string {\n  let dir = startDir;\n  while (dir !== \"/\") {\n    if (existsSync(resolve(dir, \"package.json\"))) {\n      return dir;\n    }\n    dir = dirname(dir);\n  }\n  return startDir;\n}\n\n// ============================================================================\n// Standard Paths\n// ============================================================================\n\nlet _monorepoRoot: string | null = null;\nlet _packageRoot: string | null = null;\nlet _tempDir: string | null = null;\n\n/**\n * Monorepo root directory\n */\nexport function getMonorepoPath(): string {\n  if (!_monorepoRoot) {\n    _monorepoRoot = findMonorepoRoot();\n  }\n  return _monorepoRoot;\n}\n\n/**\n * Current package root directory\n */\nexport function getPackagePath(): string {\n  if (!_packageRoot) {\n    _packageRoot = getPackageRoot();\n  }\n  return _packageRoot;\n}\n\n/**\n * BDD directory for current package\n */\nexport function getBddPath(): string {\n  return resolve(getPackagePath(), \"bdd\");\n}\n\n/**\n * Fixtures directory for current package's BDD tests\n */\nexport function getFixturesPath(subdir?: string): string {\n  const base = resolve(getBddPath(), \"fixtures\");\n  return subdir ? resolve(base, subdir) : base;\n}\n\n/**\n * Get or create a temporary directory for tests\n */\nexport function getTempPath(prefix: string = \"bdd-\"): string {\n  if (!_tempDir) {\n    _tempDir = mkdtempSync(resolve(tmpdir(), prefix));\n  }\n  return _tempDir;\n}\n\n/**\n * Ensure a directory exists, creating it if necessary\n */\nexport function ensureDir(path: string): string {\n  if (!existsSync(path)) {\n    mkdirSync(path, { recursive: true });\n  }\n  return path;\n}\n\n/**\n * Reset cached paths (useful for testing)\n */\nexport function resetPaths(): void {\n  _monorepoRoot = null;\n  _packageRoot = null;\n  _tempDir = null;\n}\n\n// ============================================================================\n// Convenience exports\n// ============================================================================\n\nexport const paths = {\n  monorepo: getMonorepoPath,\n  package: getPackagePath,\n  bdd: getBddPath,\n  fixtures: getFixturesPath,\n  temp: getTempPath,\n  ensure: ensureDir,\n  reset: resetPaths,\n};\n","import { execFileSync } from \"node:child_process\";\nimport { readFileSync } from \"node:fs\";\nimport { resolve, dirname } from \"node:path\";\nimport { fileURLToPath } from \"node:url\";\n\nconst __dirname = dirname(fileURLToPath(import.meta.url));\n\nconst SKILL_PATH = resolve(__dirname, \"../../../../.claude/skills/agent-browser/SKILL.md\");\n\nfunction loadSystemPrompt(headed = false): string {\n  let skillContent = \"\";\n  try {\n    skillContent = readFileSync(SKILL_PATH, \"utf-8\");\n  } catch {\n    // Skill file not found, continue without it\n  }\n\n  return `You are a UI tester. You test web application scenarios using the agent-browser CLI.\n\nRULES:\n- ONLY use agent-browser commands via Bash tool\n- Use ${headed ? \"--headed \" : \"\"}--executable-path \"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome\" for all commands\n- After each navigation or click, run: agent-browser snapshot -i\n- Refs (@e1, @e2) are invalidated after page changes — always re-snapshot\n- At the end, close the browser with: agent-browser close\n- Output your result as a single line: PASS or FAIL followed by a brief reason\n\n${skillContent ? `AGENT-BROWSER REFERENCE:\\n${skillContent}` : \"\"}`;\n}\n\nexport interface UiTestResult {\n  passed: boolean;\n  output: string;\n}\n\nexport interface UiTesterOptions {\n  model?: string;\n  baseUrl?: string;\n  timeout?: number;\n  /** Show browser window (default: false) */\n  headed?: boolean;\n}\n\n/**\n * Run a UI test scenario using Claude Code CLI + agent-browser.\n *\n * BDD scripts must run under Node.js (not Bun) to avoid claude CLI auth bug.\n */\nexport function agentUiTester(prompt: string, options: UiTesterOptions = {}): UiTestResult {\n  const { model = \"haiku\", baseUrl, timeout = 300_000, headed = false } = options;\n\n  const fullPrompt = baseUrl ? `Base URL: ${baseUrl}\\n\\n${prompt}` : prompt;\n\n  const systemPrompt = loadSystemPrompt(headed);\n\n  // Filter out CLAUDE* env vars to avoid auth conflicts when spawned from Claude Code\n  const cleanEnv = Object.fromEntries(\n    Object.entries(process.env).filter(([k]) => !k.startsWith(\"CLAUDE\"))\n  );\n\n  try {\n    const output = execFileSync(\n      \"claude\",\n      [\n        \"-p\",\n        fullPrompt,\n        \"--model\",\n        model,\n        \"--append-system-prompt\",\n        systemPrompt,\n        \"--allowedTools\",\n        \"Bash(agent-browser:*)\",\n      ],\n      {\n        encoding: \"utf-8\",\n        timeout,\n        env: cleanEnv,\n        maxBuffer: 10 * 1024 * 1024,\n      }\n    ).trim();\n\n    const passed = /\\*{0,2}PASS\\*{0,2}\\b/m.test(output);\n    return { passed, output };\n  } catch (error: any) {\n    const output = error.stdout || error.stderr || error.message || \"Unknown error\";\n    return { passed: false, output: output.trim() };\n  }\n}\n","import { readFileSync, existsSync } from \"node:fs\";\nimport { env } from \"../env\";\n\nconst SYSTEM_PROMPT = `You are a documentation reviewer evaluating documents from the reader's experience.\n\nEVALUATION DIMENSIONS:\n1. Completeness — All required information is present. Nothing critical is missing.\n2. Logic — Structure flows naturally. Concepts build on each other without jumps.\n3. Readability — A newcomer can follow without confusion. No unexplained jargon.\n\nRULES:\n- Read the provided document carefully\n- Evaluate each requirement listed in the prompt against ALL three dimensions\n- Be strict but fair — the document should genuinely help the reader achieve the stated goal\n- Output your result as a single line: PASS or FAIL followed by a brief reason\n- If FAIL, list which specific requirements are not met and which dimension they violate`;\n\nexport interface DocTestResult {\n  passed: boolean;\n  output: string;\n}\n\nexport interface DocTesterOptions {\n  /** LLM provider (default: \"anthropic\") */\n  provider?: string;\n  /** Model name */\n  model?: string;\n  /** API key (reads from env if not provided) */\n  apiKey?: string;\n  /** Base URL (reads from env if not provided) */\n  baseUrl?: string;\n  /** Timeout in ms */\n  timeout?: number;\n}\n\n/**\n * Evaluate a document against requirements using AgentX.\n *\n * Uses agentxjs local mode — no subprocess, no CLI, no auth issues.\n * Requires `agentxjs` as a peer dependency.\n */\nexport async function agentDocTester(\n  options: {\n    files: string[];\n    requirements: string;\n  },\n  testerOptions: DocTesterOptions = {}\n): Promise<DocTestResult> {\n  const {\n    provider = process.env.AGENTX_PROVIDER || \"anthropic\",\n    model = env.model,\n    apiKey = env.apiKey || \"\",\n    baseUrl = env.baseUrl,\n    timeout = 120_000,\n  } = testerOptions;\n\n  const docContents = options.files\n    .map((filePath) => {\n      if (!existsSync(filePath)) {\n        return `--- ${filePath} ---\\n[FILE NOT FOUND]`;\n      }\n      return `--- ${filePath} ---\\n${readFileSync(filePath, \"utf-8\")}`;\n    })\n    .join(\"\\n\\n\");\n\n  const userPrompt = [\n    \"Evaluate the following document(s) against the requirements below.\",\n    \"\",\n    \"DOCUMENTS:\",\n    docContents,\n    \"\",\n    \"REQUIREMENTS:\",\n    options.requirements,\n    \"\",\n    \"Evaluate each requirement. Output PASS if all are met, FAIL if any are not.\",\n  ].join(\"\\n\");\n\n  // Dynamic import to avoid circular dependency (devtools ↔ agentxjs)\n  // Use variable to prevent TypeScript DTS from resolving the module\n  const moduleName = \"agentxjs\";\n  const agentxjs: any = await import(/* @vite-ignore */ moduleName);\n  const createAgentX: (...args: any[]) => Promise<any> = agentxjs.createAgentX;\n\n  let agentx: any = null;\n\n  try {\n    agentx = await createAgentX({\n      apiKey,\n      provider,\n      model,\n      baseUrl,\n      logLevel: \"silent\",\n    });\n\n    await agentx.containers.create(\"doc-tester\");\n\n    const { record: image } = await agentx.images.create({\n      containerId: \"doc-tester\",\n      systemPrompt: SYSTEM_PROMPT,\n    });\n\n    const { agentId } = await agentx.agents.create({ imageId: image.imageId });\n\n    // Collect response text\n    let output = \"\";\n    agentx.on(\"text_delta\", (e: any) => {\n      output += e.data.text;\n    });\n\n    // Send prompt and wait for completion\n    await Promise.race([\n      agentx.sessions.send(agentId, userPrompt),\n      new Promise((_, reject) => setTimeout(() => reject(new Error(\"Timeout\")), timeout)),\n    ]);\n\n    output = output.trim();\n    const passed = /\\*{0,2}PASS\\*{0,2}\\b/m.test(output);\n    return { passed, output };\n  } catch (error: any) {\n    return { passed: false, output: error.message || \"Unknown error\" };\n  } finally {\n    if (agentx) {\n      try {\n        await agentx.shutdown();\n      } catch {\n        // ignore shutdown errors\n      }\n    }\n  }\n}\n"],"mappings":";;;;;;;;AA4BO,SAAS,qBAAqB,SAAgC;AACnE,SAAO;AAAA,IACL,QAAQ,QAAQ,UAAU,CAAC,UAAU;AAAA,IACrC,eAAe,EAAE,kBAAkB,cAAc;AAAA,IACjD,QAAQ,QAAQ;AAAA,IAChB,OAAO,QAAQ;AAAA,IACf,MAAM,QAAQ,QAAQ;AAAA,IACtB,iBAAiB;AAAA,MACf,gBAAgB,QAAQ,WAAW;AAAA,IACrC;AAAA,EACF;AACF;;;AC3BA,SAAS,gBAA+B;AAOxC,IAAI,UAA0B;AAC9B,IAAI,OAAoB;AAKxB,eAAsB,cAAc,UAA0B,CAAC,GAAqB;AAClF,MAAI,QAAS,QAAO;AAEpB,QAAM,WAAW,QAAQ,YAAY,QAAQ,IAAI,OAAO;AAExD,YAAU,MAAM,SAAS,OAAO;AAAA,IAC9B,SAAS;AAAA,IACT;AAAA,IACA,QAAQ,QAAQ;AAAA,EAClB,CAAC;AAED,SAAO;AACT;AAKA,eAAsB,UAAyB;AAC7C,MAAI,QAAQ,CAAC,KAAK,SAAS,EAAG,QAAO;AAErC,QAAM,IAAI,MAAM,cAAc;AAC9B,SAAO,MAAM,EAAE,QAAQ;AACvB,SAAO;AACT;AAMA,eAAsB,YAA2B;AAC/C,MAAI;AACF,QAAI,QAAQ,CAAC,KAAK,SAAS,GAAG;AAC5B,YAAM,UAAU,KAAK,QAAQ;AAC7B,YAAM,QAAQ,aAAa;AAC3B,YAAM,KAAK,KAAK,aAAa;AAAA,IAC/B;AAAA,EACF,QAAQ;AAEN,WAAO;AAAA,EACT;AACF;AAMA,eAAsB,YAA2B;AAC/C,MAAI,MAAM;AACR,UAAM,KAAK,MAAM;AACjB,WAAO;AAAA,EACT;AACF;AAKA,eAAsB,eAA8B;AAClD,MAAI,QAAQ,CAAC,KAAK,SAAS,GAAG;AAC5B,UAAM,KAAK,MAAM;AACjB,WAAO;AAAA,EACT;AACA,MAAI,SAAS;AACX,UAAM,QAAQ,MAAM;AACpB,cAAU;AAAA,EACZ;AACF;AAKA,eAAsB,WAAW,KAAa,UAAU,KAAyB;AAC/E,QAAM,QAAQ,KAAK,IAAI;AACvB,SAAO,KAAK,IAAI,IAAI,QAAQ,SAAS;AACnC,QAAI;AACF,YAAM,WAAW,MAAM,MAAM,GAAG;AAChC,UAAI,SAAS,IAAI;AACf,eAAO;AAAA,MACT;AAAA,IACF,QAAQ;AAAA,IAER;AACA,UAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,GAAG,CAAC;AAAA,EAC7C;AACA,SAAO;AACT;;;ACvGA,SAAS,aAA2B;AAkBpC,IAAI,YAAiC;AAKrC,eAAsB,eAAe,SAA0C;AAC7E,MAAI,UAAW;AAEf,QAAM;AAAA,IACJ;AAAA,IACA,UAAU;AAAA,IACV,OAAO,CAAC,OAAO,KAAK;AAAA,IACpB;AAAA,IACA,UAAU;AAAA,IACV,QAAQ,CAAC,CAAC,QAAQ,IAAI;AAAA,EACxB,IAAI;AAEJ,cAAY,MAAM,SAAS,MAAM;AAAA,IAC/B;AAAA,IACA,OAAO;AAAA,IACP,UAAU;AAAA,EACZ,CAAC;AAED,MAAI,OAAO;AACT,cAAU,QAAQ,GAAG,QAAQ,CAAC,SAAS;AACrC,cAAQ,IAAI,gBAAgB,KAAK,SAAS,CAAC;AAAA,IAC7C,CAAC;AAED,cAAU,QAAQ,GAAG,QAAQ,CAAC,SAAS;AACrC,cAAQ,MAAM,sBAAsB,KAAK,SAAS,CAAC;AAAA,IACrD,CAAC;AAAA,EACH;AAEA,QAAM,MAAM,oBAAoB,IAAI;AACpC,QAAM,QAAQ,MAAM,WAAW,KAAK,OAAO;AAE3C,MAAI,CAAC,OAAO;AACV,kBAAc;AACd,UAAM,IAAI,MAAM,sCAAsC,IAAI,EAAE;AAAA,EAC9D;AACF;AAKO,SAAS,gBAAsB;AACpC,MAAI,WAAW;AACb,cAAU,KAAK,SAAS;AACxB,gBAAY;AAAA,EACd;AACF;AAKO,SAAS,eAAoC;AAClD,SAAO;AACT;;;AC3EA,SAAS,SAAS,eAAe;AACjC,SAAS,YAAY,aAAa,iBAAiB;AACnD,SAAS,cAAc;AAShB,SAAS,iBAAiB,WAAmB,QAAQ,IAAI,GAAW;AACzE,MAAI,MAAM;AACV,SAAO,QAAQ,KAAK;AAClB,UAAM,UAAU,QAAQ,KAAK,cAAc;AAC3C,QAAI,WAAW,OAAO,GAAG;AACvB,UAAI;AACF,cAAM,MAAM,UAAQ,OAAO;AAC3B,YAAI,IAAI,cAAc,IAAI,YAAY,MAAM;AAE1C,gBAAM,cAAc,WAAW,QAAQ,KAAK,UAAU,CAAC;AACvD,gBAAM,UAAU,WAAW,QAAQ,KAAK,MAAM,CAAC;AAC/C,cAAI,eAAe,SAAS;AAC1B,mBAAO;AAAA,UACT;AAAA,QACF;AAAA,MACF,QAAQ;AAAA,MAER;AAAA,IACF;AACA,UAAM,QAAQ,GAAG;AAAA,EACnB;AACA,SAAO;AACT;AAKO,SAAS,eAAe,WAAmB,QAAQ,IAAI,GAAW;AACvE,MAAI,MAAM;AACV,SAAO,QAAQ,KAAK;AAClB,QAAI,WAAW,QAAQ,KAAK,cAAc,CAAC,GAAG;AAC5C,aAAO;AAAA,IACT;AACA,UAAM,QAAQ,GAAG;AAAA,EACnB;AACA,SAAO;AACT;AAMA,IAAI,gBAA+B;AACnC,IAAI,eAA8B;AAClC,IAAI,WAA0B;AAKvB,SAAS,kBAA0B;AACxC,MAAI,CAAC,eAAe;AAClB,oBAAgB,iBAAiB;AAAA,EACnC;AACA,SAAO;AACT;AAKO,SAAS,iBAAyB;AACvC,MAAI,CAAC,cAAc;AACjB,mBAAe,eAAe;AAAA,EAChC;AACA,SAAO;AACT;AAKO,SAAS,aAAqB;AACnC,SAAO,QAAQ,eAAe,GAAG,KAAK;AACxC;AAKO,SAAS,gBAAgB,QAAyB;AACvD,QAAM,OAAO,QAAQ,WAAW,GAAG,UAAU;AAC7C,SAAO,SAAS,QAAQ,MAAM,MAAM,IAAI;AAC1C;AAKO,SAAS,YAAY,SAAiB,QAAgB;AAC3D,MAAI,CAAC,UAAU;AACb,eAAW,YAAY,QAAQ,OAAO,GAAG,MAAM,CAAC;AAAA,EAClD;AACA,SAAO;AACT;AAKO,SAAS,UAAU,MAAsB;AAC9C,MAAI,CAAC,WAAW,IAAI,GAAG;AACrB,cAAU,MAAM,EAAE,WAAW,KAAK,CAAC;AAAA,EACrC;AACA,SAAO;AACT;AAKO,SAAS,aAAmB;AACjC,kBAAgB;AAChB,iBAAe;AACf,aAAW;AACb;AAMO,IAAM,QAAQ;AAAA,EACnB,UAAU;AAAA,EACV,SAAS;AAAA,EACT,KAAK;AAAA,EACL,UAAU;AAAA,EACV,MAAM;AAAA,EACN,QAAQ;AAAA,EACR,OAAO;AACT;;;AC3IA,SAAS,oBAAoB;AAC7B,SAAS,oBAAoB;AAC7B,SAAS,WAAAA,UAAS,WAAAC,gBAAe;AACjC,SAAS,qBAAqB;AAE9B,IAAM,YAAYA,SAAQ,cAAc,YAAY,GAAG,CAAC;AAExD,IAAM,aAAaD,SAAQ,WAAW,mDAAmD;AAEzF,SAAS,iBAAiB,SAAS,OAAe;AAChD,MAAI,eAAe;AACnB,MAAI;AACF,mBAAe,aAAa,YAAY,OAAO;AAAA,EACjD,QAAQ;AAAA,EAER;AAEA,SAAO;AAAA;AAAA;AAAA;AAAA,QAID,SAAS,cAAc,EAAE;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAM/B,eAAe;AAAA,EAA6B,YAAY,KAAK,EAAE;AACjE;AAoBO,SAAS,cAAc,QAAgB,UAA2B,CAAC,GAAiB;AACzF,QAAM,EAAE,QAAQ,SAAS,SAAS,UAAU,KAAS,SAAS,MAAM,IAAI;AAExE,QAAM,aAAa,UAAU,aAAa,OAAO;AAAA;AAAA,EAAO,MAAM,KAAK;AAEnE,QAAM,eAAe,iBAAiB,MAAM;AAG5C,QAAM,WAAW,OAAO;AAAA,IACtB,OAAO,QAAQ,QAAQ,GAAG,EAAE,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,WAAW,QAAQ,CAAC;AAAA,EACrE;AAEA,MAAI;AACF,UAAM,SAAS;AAAA,MACb;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AAAA,MACA;AAAA,QACE,UAAU;AAAA,QACV;AAAA,QACA,KAAK;AAAA,QACL,WAAW,KAAK,OAAO;AAAA,MACzB;AAAA,IACF,EAAE,KAAK;AAEP,UAAM,SAAS,wBAAwB,KAAK,MAAM;AAClD,WAAO,EAAE,QAAQ,OAAO;AAAA,EAC1B,SAAS,OAAY;AACnB,UAAM,SAAS,MAAM,UAAU,MAAM,UAAU,MAAM,WAAW;AAChE,WAAO,EAAE,QAAQ,OAAO,QAAQ,OAAO,KAAK,EAAE;AAAA,EAChD;AACF;;;ACvFA,SAAS,gBAAAE,eAAc,cAAAC,mBAAkB;AAGzC,IAAM,gBAAgB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAsCtB,eAAsB,eACpB,SAIA,gBAAkC,CAAC,GACX;AACxB,QAAM;AAAA,IACJ,WAAW,QAAQ,IAAI,mBAAmB;AAAA,IAC1C,QAAQ,IAAI;AAAA,IACZ,SAAS,IAAI,UAAU;AAAA,IACvB,UAAU,IAAI;AAAA,IACd,UAAU;AAAA,EACZ,IAAI;AAEJ,QAAM,cAAc,QAAQ,MACzB,IAAI,CAAC,aAAa;AACjB,QAAI,CAACC,YAAW,QAAQ,GAAG;AACzB,aAAO,OAAO,QAAQ;AAAA;AAAA,IACxB;AACA,WAAO,OAAO,QAAQ;AAAA,EAASC,cAAa,UAAU,OAAO,CAAC;AAAA,EAChE,CAAC,EACA,KAAK,MAAM;AAEd,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ;AAAA,IACR;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AAIX,QAAM,aAAa;AACnB,QAAM,WAAgB,MAAM;AAAA;AAAA,IAA0B;AAAA;AACtD,QAAM,eAAiD,SAAS;AAEhE,MAAI,SAAc;AAElB,MAAI;AACF,aAAS,MAAM,aAAa;AAAA,MAC1B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,CAAC;AAED,UAAM,OAAO,WAAW,OAAO,YAAY;AAE3C,UAAM,EAAE,QAAQ,MAAM,IAAI,MAAM,OAAO,OAAO,OAAO;AAAA,MACnD,aAAa;AAAA,MACb,cAAc;AAAA,IAChB,CAAC;AAED,UAAM,EAAE,QAAQ,IAAI,MAAM,OAAO,OAAO,OAAO,EAAE,SAAS,MAAM,QAAQ,CAAC;AAGzE,QAAI,SAAS;AACb,WAAO,GAAG,cAAc,CAAC,MAAW;AAClC,gBAAU,EAAE,KAAK;AAAA,IACnB,CAAC;AAGD,UAAM,QAAQ,KAAK;AAAA,MACjB,OAAO,SAAS,KAAK,SAAS,UAAU;AAAA,MACxC,IAAI,QAAQ,CAAC,GAAG,WAAW,WAAW,MAAM,OAAO,IAAI,MAAM,SAAS,CAAC,GAAG,OAAO,CAAC;AAAA,IACpF,CAAC;AAED,aAAS,OAAO,KAAK;AACrB,UAAM,SAAS,wBAAwB,KAAK,MAAM;AAClD,WAAO,EAAE,QAAQ,OAAO;AAAA,EAC1B,SAAS,OAAY;AACnB,WAAO,EAAE,QAAQ,OAAO,QAAQ,MAAM,WAAW,gBAAgB;AAAA,EACnE,UAAE;AACA,QAAI,QAAQ;AACV,UAAI;AACF,cAAM,OAAO,SAAS;AAAA,MACxB,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AACF;","names":["resolve","dirname","readFileSync","existsSync","existsSync","readFileSync"]}