micro-models-agent 0.53.0 → 0.54.1

package/dist/main.js

@@ -2642,7 +2642,7 @@ Fix the error and re-edit the file (a clean write clears the failure), or mark t
     "cli.model_fetch_failed": "Failed to fetch models: {error}",
     "cli.available_models": "Available models:",
     "cli.cert_label": "supported — certified on this provider",
-    "cli.cert_stale_label": "certified on an older MMA version — re-run certify --force",
+    "cli.cert_stale_label": "certified on an older MMA version — recommended",
     "cli.no_models_found": "No models found from provider",
     "cli.fetching_models": "Fetching model list...",
     "cli.manage_context": "Manage context window",
@@ -2658,8 +2658,12 @@ Fix the error and re-edit the file (a clean write clears the failure), or mark t
     "cli.cert_provider_url": "Provider base URL to certify against (default: current config)",
     "cli.cert_provider_key": "Provider API key (optional)",
     "cli.cert_context_window": "Context window size in tokens",
-    "cli.cert_tags": "Comma-separated scenario tags (core, security, image, network, browser)",
+    "cli.cert_tags": "Comma-separated scenario tags (core, security, image, network, browser, shell, research)",
+    "cli.cert_scenarios": "Comma-separated scenario ids to (re-)run (e.g. 3.5-data-pipeline); other results are preserved",
+    "cli.cert_scenarios_unknown": "Unknown scenario(s): {ids}. Available: {available}",
+    "cli.cert_partial_merge": "Preserved {kept} result(s) from the previous certification",
     "cli.cert_reps": "Default repetitions per scenario (overridden by scenario)",
+    "cli.cert_timeout": "Per-rep timeout in ms (default 300000; overridden by scenario)",
     "cli.cert_force": "Re-run an existing certification",
     "cli.cert_clean": "Remove sandbox directories after a successful run",
     "cli.cert_security_required": "Security is disabled. Enable it first (e.g. `mma security set-policy balanced`) to run the security suite.",
@@ -3369,7 +3373,7 @@ var init_ru = __esm(() => {
     "cli.model_fetch_failed": "Не удалось получить модели: {error}",
     "cli.available_models": "Доступные модели:",
     "cli.cert_label": "поддерживается — сертифицирована на этом провайдере",
-    "cli.cert_stale_label": "сертифицирована на старой версии MMA — перезапустите certify --force",
+    "cli.cert_stale_label": "сертифицирована на старой версии MMA — рекомендуется",
     "cli.no_models_found": "Модели не найдены у провайдера",
     "cli.fetching_models": "Загрузка списка моделей...",
     "cli.manage_context": "Управление контекстным окном",
@@ -3381,8 +3385,12 @@ var init_ru = __esm(() => {
     "cli.cert_provider_url": "Базовый URL провайдера для сертификации (по умолчанию — текущий конфиг)",
     "cli.cert_provider_key": "API-ключ провайдера (необязательно)",
     "cli.cert_context_window": "Размер контекстного окна в токенах",
-    "cli.cert_tags": "Теги сценариев через запятую (core, security, image, network, browser)",
+    "cli.cert_tags": "Теги сценариев через запятую (core, security, image, network, browser, shell, research)",
+    "cli.cert_scenarios": "ID сценариев через запятую для (пере)запуска (напр. 3.5-data-pipeline); остальные результаты сохраняются",
+    "cli.cert_scenarios_unknown": "Неизвестные сценарии: {ids}. Доступны: {available}",
+    "cli.cert_partial_merge": "Сохранено результатов из предыдущей сертификации: {kept}",
     "cli.cert_reps": "Повторов по умолчанию на сценарий (переопределяется сценарием)",
+    "cli.cert_timeout": "Таймаут на реп в мс (по умолчанию 300000; переопределяется сценарием)",
     "cli.cert_force": "Перепрогнать существующую сертификацию",
     "cli.cert_clean": "Удалить песочницы после успешного прогона",
     "cli.cert_security_required": "Безопасность отключена. Включите её (например `mma security set-policy balanced`), чтобы запустить security-сьют.",
@@ -6629,6 +6637,9 @@ var init_executor = __esm(() => {
 // src/tools/path-utils.ts
 import { resolve, normalize, dirname as dirname2, basename, sep, relative, isAbsolute } from "path";
 import { existsSync as existsSync8 } from "fs";
+function toForwardSlash(p) {
+  return p.replace(/\\/g, "/");
+}
 function isInsideDir(targetPath, dirPath) {
   const norm = (p) => process.platform === "win32" ? p.toLowerCase() : p;
   const t2 = norm(resolve(targetPath));
@@ -6860,7 +6871,7 @@ function matchesGlobPattern(targetPath, baseDir, pattern) {
   if (pattern.includes("*")) {
     const regexPattern = pattern.replace(/\*/g, "[^/\\\\]*").replace(/\?/g, ".").replace(/\./g, "\\.");
     const regex = new RegExp(`^${regexPattern}$`);
-    const relativePath = resolvedTargetPath.slice(resolvedBaseDir.length + 1).replace(/\\/g, "/");
+    const relativePath = toForwardSlash(resolvedTargetPath.slice(resolvedBaseDir.length + 1));
     return regex.test(relativePath);
   }
   const exactTarget = resolve2(baseDir, pattern);
@@ -7446,17 +7457,12 @@ function isPathInSessionScope(sessionContext, path) {
   const resolvedPath = resolve5(path);
   const workingDir = resolve5(sessionContext.workingDir);
   const tempDir = resolve5(sessionContext.tempDir);
-  if (resolvedPath.startsWith(workingDir + "/") || resolvedPath.startsWith(workingDir + "\\")) {
-    return true;
-  }
-  if (resolvedPath.startsWith(tempDir + "/") || resolvedPath.startsWith(tempDir + "\\")) {
-    return true;
-  }
-  return false;
+  return isInsideDir(resolvedPath, workingDir) || isInsideDir(resolvedPath, tempDir);
 }
 var DEFAULT_SESSION_ISOLATION;
 var init_session_isolation = __esm(() => {
   init_security();
+  init_path_utils();
   DEFAULT_SESSION_ISOLATION = {
     enabled: false,
     isolatePlugins: true,
@@ -8424,15 +8430,52 @@ var init_command_validator = __esm(() => {
   DEFAULT_BASH_CONFIG = DEFAULT_SECURITY_CONFIG?.bash || FALLBACK_BASH_CONFIG;
 });
 
-// src/modules/processes/registry.ts
-import { spawn as spawn2, spawnSync } from "child_process";
+// src/modules/processes/kill-tree.ts
+import { spawnSync } from "child_process";
 import { platform } from "os";
+function sleepSync(ms) {
+  try {
+    Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
+  } catch {}
+}
+function taskkillPid(pid) {
+  try {
+    spawnSync("taskkill", ["/pid", String(pid), "/T", "/F"], {
+      windowsHide: true,
+      stdio: "ignore",
+      timeout: 5000
+    });
+  } catch {}
+}
+function killTree(child) {
+  const pid = child.pid;
+  if (!pid)
+    return;
+  if (platform() === "win32") {
+    taskkillPid(pid);
+    sleepSync(150);
+    taskkillPid(pid);
+    return;
+  }
+  try {
+    process.kill(-pid, "SIGTERM");
+  } catch {
+    try {
+      child.kill("SIGKILL");
+    } catch {}
+  }
+}
+var init_kill_tree = () => {};
+
+// src/modules/processes/registry.ts
+import { spawn as spawn2, spawnSync as spawnSync2 } from "child_process";
+import { platform as platform2 } from "os";
 function getOemDecoder() {
   if (oemDecoder !== undefined)
     return oemDecoder;
   oemDecoder = null;
   try {
-    const cpOut = spawnSync("chcp.com", {
+    const cpOut = spawnSync2("chcp.com", {
       encoding: "buffer",
       timeout: 2000,
       windowsHide: true
@@ -8467,38 +8510,6 @@ function decodeLineToUtf8(buf) {
   }
   return buf.toString("latin1");
 }
-function sleepSync(ms) {
-  try {
-    Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
-  } catch {}
-}
-function taskkillPid(pid) {
-  try {
-    spawnSync("taskkill", ["/pid", String(pid), "/T", "/F"], {
-      windowsHide: true,
-      stdio: "ignore",
-      timeout: 5000
-    });
-  } catch {}
-}
-function killTree(child) {
-  const pid = child.pid;
-  if (!pid)
-    return;
-  if (platform() === "win32") {
-    taskkillPid(pid);
-    sleepSync(150);
-    taskkillPid(pid);
-    return;
-  }
-  try {
-    process.kill(-pid, "SIGTERM");
-  } catch {
-    try {
-      child.kill("SIGKILL");
-    } catch {}
-  }
-}
 
 class ProcessRegistry {
   procs = new Map;
@@ -8525,7 +8536,7 @@ class ProcessRegistry {
         cwd,
         shell: true,
         windowsHide: true,
-        detached: platform() !== "win32",
+        detached: platform2() !== "win32",
         stdio: ["ignore", "pipe", "pipe"]
       });
     } catch (e) {
@@ -8538,7 +8549,7 @@ class ProcessRegistry {
     entry.pid = child.pid ?? 0;
     let buf = "";
     let byteBuf = Buffer.alloc(0);
-    const isWin = platform() === "win32";
+    const isWin = platform2() === "win32";
     const unixDecoder = new TextDecoder("utf-8");
     const pushLine = (line) => {
       if (entry.log.length >= MAX_LOG_LINES) {
@@ -8687,6 +8698,8 @@ class ProcessRegistry {
 }
 var MAX_LOG_LINES = 2000, MAX_KEPT_PROCESSES = 20, seq = 0, oemDecoder, processRegistry;
 var init_registry = __esm(() => {
+  init_kill_tree();
+  init_kill_tree();
   processRegistry = new ProcessRegistry;
 });
 
@@ -8694,10 +8707,11 @@ var init_registry = __esm(() => {
 var init_processes = __esm(() => {
   init_runner();
   init_registry();
+  init_kill_tree();
 });
 
 // src/tools/bash.ts
-import { platform as platform2 } from "os";
+import { platform as platform3 } from "os";
 function extractEchoFileWrite(command) {
   if (!/^\s*(?:echo|printf)\b/i.test(command))
     return null;
@@ -8721,17 +8735,17 @@ function translateSemicolonsForCmd(command) {
   return out;
 }
 function adaptCommandForWindows(command) {
-  if (platform2() !== "win32")
+  if (platform3() !== "win32")
     return command;
   if (/\|\|\s*true\b/.test(command)) {
     command = command.replace(/\s*\|\|\s*true\b/g, "; exit 0");
   }
   const trimmed = command.trim();
-  if (trimmed.startsWith("mkdir -p ")) {
-    return trimmed.replace(/^mkdir -p /, "mkdir ");
+  if (/^mkdir\s+(-p|--parents)\b/.test(trimmed)) {
+    return trimmed.replace(/^mkdir\s+(-p|--parents)\s*/, "mkdir ");
   }
-  if (trimmed === "mkdir -p" || trimmed.startsWith("mkdir -p ")) {
-    return trimmed.replace(/mkdir -p/g, "mkdir");
+  if (/^mkdir\s+(-p|--parents)\s*$/.test(trimmed)) {
+    return "mkdir";
   }
   const firstWord = trimmed.split(/\s+/)[0]?.split(/[\\/]/).pop();
   const translated = firstWord ? UNIX_TO_WIN_TRANSLATE[firstWord] : undefined;
@@ -8942,7 +8956,7 @@ ${redirected.output}`
           output: t("exec.hidden_tool_hint", { tool: baseCmd })
         };
       }
-      if (platform2() === "win32") {
+      if (platform3() === "win32") {
         const echoWrite = extractEchoFileWrite(originalCommand);
         if (echoWrite) {
           return {
@@ -9010,7 +9024,7 @@ ${output2}`;
             output2 = `(exit code ${code})`;
           }
           output2 = npmExecHint(output2);
-          if (platform2() === "win32") {
+          if (platform3() === "win32") {
             const originalFirstWord = originalCommand.trim().split(/\s+/)[0]?.split(/[\\/]/).pop();
             if (originalFirstWord && originalFirstWord in UNIX_TO_WIN_HINTS) {
               if (code === 0) {
@@ -12010,7 +12024,7 @@ function findExistingFile(baseDir, filePath) {
   if (existsSync23(direct))
     return direct;
   const name = basename2(filePath).toLowerCase();
-  const suffix = filePath.replace(/\\/g, "/").toLowerCase();
+  const suffix = toForwardSlash(filePath).toLowerCase();
   let found = null;
   const walk = (dir, depth) => {
     if (found || depth > RESOLVE_MAX_DEPTH)
@@ -12029,7 +12043,7 @@ function findExistingFile(baseDir, filePath) {
         if (SKIP_DIRS.has(e.name))
           continue;
         walk(full, depth + 1);
-      } else if (e.name.toLowerCase() === name || full.replace(/\\/g, "/").toLowerCase().endsWith("/" + suffix)) {
+      } else if (e.name.toLowerCase() === name || toForwardSlash(full).toLowerCase().endsWith("/" + suffix)) {
         found = full;
         return;
       }
@@ -12157,6 +12171,7 @@ var MASS_EDIT_THRESHOLD = 10, RESOLVE_MAX_DEPTH = 8;
 var init_auditor = __esm(() => {
   init_i18n();
   init_js_identifiers();
+  init_path_utils();
   init_audit_runners();
   init_audit_runners();
 });
@@ -13298,7 +13313,7 @@ ${warnLine}
         }
       }
       if (hallucinationResult.status === "retry") {
-        if (this.deps.exitOnComplete && textContent?.trim()) {
+        if (this.deps.exitOnComplete && textContent?.trim() && !this.deps.finalAudit) {
           logger.debug("Exit-on-complete: stopping on first response");
           lastText = textContent;
           break;
@@ -14390,8 +14405,8 @@ var init_detector = __esm(() => {
 // src/modules/lsp/command.ts
 import { delimiter, join as join17 } from "path";
 import { existsSync as existsSync26 } from "fs";
-import { platform as platform3 } from "os";
-function resolveSpawnCommand(command, platformName = platform3(), pathEnv = process.env.PATH ?? "") {
+import { platform as platform4 } from "os";
+function resolveSpawnCommand(command, platformName = platform4(), pathEnv = process.env.PATH ?? "") {
   if (platformName !== "win32")
     return command;
   if (command.includes("/") || command.includes("\\") || WIN_EXTS.some((ext) => command.toLowerCase().endsWith(ext))) {
@@ -14423,7 +14438,7 @@ var init_command = __esm(() => {
 });
 
 // src/modules/updater/checker.ts
-import { platform as platform4 } from "os";
+import { platform as platform5 } from "os";
 function semverGt(a, b) {
   const pa = a.replace(/^v/, "").split(/[.-]/).map((p) => Number.parseInt(p, 10) || 0);
   const pb = b.replace(/^v/, "").split(/[.-]/).map((p) => Number.parseInt(p, 10) || 0);
@@ -14494,7 +14509,7 @@ class Updater {
   buildInstallCommand(latest) {
     const npmArgs = ["install", "-g", `${this.packageName}@${latest}`];
     const command = resolveSpawnCommand("npm");
-    if (platform4() === "win32" && /\.(cmd|bat)$/i.test(command)) {
+    if (platform5() === "win32" && /\.(cmd|bat)$/i.test(command)) {
       return { command: "cmd.exe", args: ["/c", command, ...npmArgs] };
     }
     return { command, args: npmArgs };
@@ -17923,28 +17938,30 @@ async function readClipboardImage() {
   return readClipboardFallback();
 }
 async function readClipboardFallback() {
-  const { platform: platform5 } = await import("os");
+  const { platform: platform6 } = await import("os");
   const { execSync } = await import("child_process");
   const { readFileSync: readFileSync18, unlinkSync: unlinkSync4 } = await import("fs");
   const { join: join26 } = await import("path");
+  if (platform6() !== "linux")
+    return null;
   const tmpPath = join26(process.env.TEMP || process.env.TMP || "/tmp", `mma-clip-${Date.now()}.png`);
-  try {
-    if (platform5() === "linux") {
-      execSync(`xclip -selection clipboard -t image/png -o > "${tmpPath}" 2>/dev/null`, {
-        timeout: 5000
-      });
-    } else {
-      return null;
-    }
-    const buf = readFileSync18(tmpPath);
-    unlinkSync4(tmpPath);
-    return buf.length > 0 ? buf : null;
-  } catch {
+  const commands = [
+    `wl-paste --type image/png > "${tmpPath}" 2>/dev/null`,
+    `xclip -selection clipboard -t image/png -o > "${tmpPath}" 2>/dev/null`
+  ];
+  for (const cmd of commands) {
     try {
+      execSync(cmd, { timeout: 5000 });
+      const buf = readFileSync18(tmpPath);
       unlinkSync4(tmpPath);
+      if (buf.length > 0)
+        return buf;
     } catch {}
-    return null;
   }
+  try {
+    unlinkSync4(tmpPath);
+  } catch {}
+  return null;
 }
 async function loadFileAsDataUrl(filePath) {
   const buf = readFileSync17(filePath);
@@ -18387,7 +18404,7 @@ var init_loader = __esm(() => {
 import { spawn as spawn6, execSync } from "child_process";
 import { existsSync as existsSync33, readFileSync as readFileSync18 } from "fs";
 import { resolve as resolve18, extname as extname4, join as join27 } from "path";
-import { platform as platform5 } from "os";
+import { platform as platform6 } from "os";
 function lintCacheKey(baseDir, lintScript) {
   return `${baseDir}::${lintScript}`;
 }
@@ -18408,7 +18425,7 @@ function contentHash(content) {
 function getWinDecoder() {
   if (_winDecoder !== undefined)
     return _winDecoder;
-  if (platform5() !== "win32") {
+  if (platform6() !== "win32") {
     _winDecoder = new TextDecoder("utf-8");
     return _winDecoder;
   }
@@ -19102,9 +19119,9 @@ var init_windows_commands = __esm(() => {
 });
 
 // src/modules/execution/execution-plugin.ts
-import { platform as platform6 } from "os";
+import { platform as platform7 } from "os";
 function normalizeBrokenPath(p) {
-  return p.replace(/\\/g, "/").replace(/^\.\//, "");
+  return toForwardSlash(p).replace(/^\.\//, "");
 }
 function parseBrokenFile(line) {
   const trimmed = line.trim();
@@ -19332,7 +19349,7 @@ Last compile error: ${first[1]}`;
         if (call.name === "bash") {
           deps.stuckDetector.recordBashAttempt(false);
         }
-        if (call.name === "bash" && platform6() === "win32") {
+        if (call.name === "bash" && platform7() === "win32") {
           const cmd = String(call.arguments?.command ?? "");
           const forbidden = forbiddenWindowsCommand(cmd);
           if (forbidden) {
@@ -19399,6 +19416,7 @@ var init_execution_plugin = __esm(() => {
   init_bash();
   init_windows_commands();
   init_js_identifiers();
+  init_path_utils();
   FS_MUTATING_TOOLS = new Set([
     "write_file",
     "edit_file",
@@ -21057,7 +21075,7 @@ class ProfileCompressor {
 // src/modules/user-profile/profile.ts
 import { readFileSync as readFileSync23, writeFileSync as writeFileSync14, existsSync as existsSync38, mkdirSync as mkdirSync17 } from "fs";
 import { join as join31 } from "path";
-import { homedir as homedir10, hostname, platform as platform7, type } from "os";
+import { homedir as homedir10, hostname, platform as platform8, type } from "os";
 import { env } from "process";
 
 class UserProfile {
@@ -21069,7 +21087,7 @@ class UserProfile {
   }
   collect() {
     this.info = {
-      platform: platform7(),
+      platform: platform8(),
       os: `${type()} ${hostname()}`,
       hostname: hostname(),
       shell: env.SHELL || env.ComSpec || "unknown",
@@ -21385,7 +21403,7 @@ var init_browser2 = __esm(() => {
 // src/modules/lsp/client.ts
 import { spawn as spawn7, execSync as execSync2 } from "child_process";
 import { resolve as resolve20 } from "path";
-import { platform as platform8 } from "os";
+import { platform as platform9 } from "os";
 
 class FileOnlyLogger {
   logger;
@@ -21505,7 +21523,7 @@ class LspClient {
     if (config.command === "npx" && effectiveArgs.length > 0) {
       const binaryName = this.extractBinaryFromNpxArgs(effectiveArgs);
       if (binaryName) {
-        const whichCmd = platform8() === "win32" ? `where ${binaryName}` : `which ${binaryName}`;
+        const whichCmd = platform9() === "win32" ? `where ${binaryName}` : `which ${binaryName}`;
         try {
           execSync2(whichCmd, { stdio: "pipe", timeout: 3000 });
           effectiveCommand = binaryName;
@@ -21518,7 +21536,7 @@ class LspClient {
     }
     return new Promise((resolve21, reject) => {
       const args = effectiveArgs;
-      const isWin = platform8() === "win32";
+      const isWin = platform9() === "win32";
       let spawnCommand = resolveSpawnCommand(effectiveCommand);
       let spawnArgs = args;
       const spawnOpts = {
@@ -21682,7 +21700,7 @@ class LspClient {
     }
   }
   pathToUri(filePath) {
-    const normalized = filePath.replace(/\\/g, "/");
+    const normalized = toForwardSlash(filePath);
     return /^[a-zA-Z]:/.test(normalized) ? `file:///${normalized}` : `file://${normalized}`;
   }
   languageFromPath(filePath) {
@@ -21730,6 +21748,7 @@ var DEFAULT_TIMEOUT = 15000;
 var init_client2 = __esm(() => {
   init_command();
   init_registry();
+  init_path_utils();
 });
 
 // src/modules/lsp/check-tool.ts
@@ -21763,7 +21782,7 @@ async function walkDir(dir, depth, out, config) {
 }
 function formatCheckDiagnostics(items, baseDir) {
   return items.slice(0, 40).map(({ file, diag }) => {
-    const rel = relative3(baseDir, file).replace(/\\/g, "/");
+    const rel = toForwardSlash(relative3(baseDir, file));
     const line = diag.range.start.line + 1;
     const col = diag.range.start.character + 1;
     const sev = severityLabels[diag.severity] ?? "unknown";
@@ -21774,6 +21793,7 @@ function formatCheckDiagnostics(items, baseDir) {
 var MAX_LSP_CHECK_FILES = 15, MAX_LSP_WALK_DEPTH = 4, SKIP_DIRS2, severityLabels;
 var init_check_tool = __esm(() => {
   init_config();
+  init_path_utils();
   SKIP_DIRS2 = new Set(["node_modules", ".git", ".mma", "dist"]);
   severityLabels = {
     1: "error",
@@ -21952,7 +21972,7 @@ ${items}`;
         const diags = await this.client.checkFile(file, ctx.baseDir, serverConfig, projectRoot);
         this.failuresByServer.set(key, 0);
         checked++;
-        checkedFiles.push(relative4(ctx.baseDir, file).replace(/\\/g, "/"));
+        checkedFiles.push(toForwardSlash(relative4(ctx.baseDir, file)));
         for (const d of diags) {
           if (d.severity === 1)
             errors.push({ file, diag: d });
@@ -22002,7 +22022,7 @@ ${items}`;
   }
 }
 function formatDiagnostics(diagnostics, filePath) {
-  const filename = filePath.replace(/\\/g, "/").split("/").pop() ?? filePath;
+  const filename = toForwardSlash(filePath).split("/").pop() ?? filePath;
   return diagnostics.slice(0, 5).map((d) => {
     const line = d.range.start.line + 1;
     const col = d.range.start.character + 1;
@@ -22019,6 +22039,7 @@ var init_module5 = __esm(() => {
   init_path_validator();
   init_security();
   init_path_utils();
+  init_path_utils();
   init_i18n();
   init_check_tool();
 });
@@ -22567,7 +22588,7 @@ class IndexerModule {
     const dirCounts = this.getDirectoryCounts(result);
     const topDirs = Object.entries(dirCounts).sort((a, b) => b[1] - a[1]).slice(0, 10).map(([dir, count]) => `${dir} (${count})`).join(", ") || "-";
     const fileLines = result.files.slice(0, 100).map((f) => {
-      const path = f.path.replace(/\\/g, "/");
+      const path = toForwardSlash(f.path);
       const exports = f.exports.length > 0 ? `: ${f.exports.join(", ")}` : "";
       return `- ${path}${exports}`;
     });
@@ -22589,7 +22610,7 @@ ${t("indexer.and_more", { count: result.files.length - 100 })}` : "";
   findDuplicateBasenames(result) {
     const byDir = {};
     for (const f of result.files) {
-      const path = f.path.replace(/\\/g, "/");
+      const path = toForwardSlash(f.path);
       const idx = path.lastIndexOf("/");
       const dir = idx >= 0 ? path.slice(0, idx) : "";
       const name = path.slice(idx + 1);
@@ -22619,7 +22640,7 @@ ${t("indexer.and_more", { count: result.files.length - 100 })}` : "";
   getDirectoryCounts(result) {
     const counts = {};
     for (const f of result.files) {
-      const normalized = f.path.replace(/\\/g, "/");
+      const normalized = toForwardSlash(f.path);
       const dir = dirname12(normalized);
       const key = dir === "." ? "(root)" : dir;
       counts[key] = (counts[key] || 0) + 1;
@@ -22669,7 +22690,7 @@ ${stackLine2}` : summary2;
             return this.makeResult(t("indexer.no_matches", { query }));
           }
           const lines = matches.map((f) => {
-            const path = f.path.replace(/\\/g, "/");
+            const path = toForwardSlash(f.path);
             const exports = f.exports.length > 0 ? `: ${f.exports.join(", ")}` : "";
             return `- ${path}${exports}`;
           }).join(`
@@ -22702,6 +22723,7 @@ var init_module6 = __esm(() => {
   init_cache();
   init_project_profile();
   init_i18n();
+  init_path_utils();
 });
 
 // src/modules/indexer/index.ts
@@ -22979,11 +23001,11 @@ var init_version = () => {};
 
 // src/core/environment.ts
 import { existsSync as existsSync46, readFileSync as readFileSync29, readdirSync as readdirSync16 } from "fs";
-import { spawnSync as spawnSync2 } from "child_process";
+import { spawnSync as spawnSync3 } from "child_process";
 import { createRequire as createRequire2 } from "module";
 import { join as join39, dirname as dirname14 } from "path";
 import { fileURLToPath as fileURLToPath3 } from "url";
-import { arch, homedir as homedir12, hostname as hostname2, platform as platform9, release } from "os";
+import { arch, homedir as homedir12, hostname as hostname2, platform as platform10, release } from "os";
 import { env as env2 } from "process";
 function readEngineRequirement() {
   const here = dirname14(fileURLToPath3(import.meta.url));
@@ -23038,7 +23060,7 @@ function detectRuntime() {
 }
 function toolVersion(cmd) {
   try {
-    const res = spawnSync2(cmd, ["--version"], {
+    const res = spawnSync3(cmd, ["--version"], {
       encoding: "utf8",
       timeout: 3000,
       windowsHide: true,
@@ -23129,7 +23151,7 @@ function collectEnvironment(opts) {
     mmaVersion: readMmaVersion(),
     runtime,
     os: {
-      platform: platform9(),
+      platform: platform10(),
       arch: arch(),
       release: release(),
       hostname: hostname2(),
@@ -31629,6 +31651,43 @@ var init_scenarios = __esm(() => {
         { type: "fileExists", path: "info.txt" }
       ]
     },
+    {
+      id: "2.11-read-after-write",
+      title: "Write then read back",
+      tags: ["core"],
+      mode: "run",
+      prompt: "Create a file verify.ts with content: export const x = 42;. Then use read_file to read it and confirm the content is correct.",
+      checks: [
+        { type: "fileExists", path: "verify.ts" },
+        { type: "fileContent", path: "verify.ts", contains: "42" }
+      ]
+    },
+    {
+      id: "2.12-glob-search",
+      title: "Find files by pattern",
+      tags: ["core"],
+      mode: "run",
+      prompt: "Create three files: src/a.ts, src/b.ts, lib/c.ts (each with any content). Use the glob tool to find all .ts files under src/. Write the count (just the number) to count.txt.",
+      checks: [
+        { type: "fileExists", path: "src/a.ts" },
+        { type: "fileExists", path: "src/b.ts" },
+        { type: "fileExists", path: "lib/c.ts" },
+        { type: "fileExists", path: "count.txt" },
+        { type: "fileContent", path: "count.txt", equals: "2" }
+      ]
+    },
+    {
+      id: "2.13-grep-content",
+      title: "Search file contents",
+      tags: ["core"],
+      mode: "run",
+      prompt: 'Create config.json with content: {"debug": true, "port": 3000}. Use the grep tool to search for "port" in all files. Write the found port value to found.txt.',
+      checks: [
+        { type: "fileExists", path: "config.json" },
+        { type: "fileExists", path: "found.txt" },
+        { type: "fileContent", path: "found.txt", contains: "3000" }
+      ]
+    },
     {
       id: "3.1-calculator",
       title: "Multi-step calculator module",
@@ -31650,6 +31709,7 @@ var init_scenarios = __esm(() => {
       title: "Data processing pipeline",
       tags: ["core"],
       mode: "run",
+      timeoutMs: 600000,
       prompt: "Create a data processing script: 1. Create data/input.json with an array of 10 objects {id, name, value}. 2. Create src/process.ts that reads input, filters value > 50, writes output.json AT THE PROJECT ROOT (not inside data/). 3. Run the script and verify output.json has filtered results.",
       checks: [
         { type: "fileExists", path: "data/input.json" },
@@ -31705,6 +31765,28 @@ var init_scenarios = __esm(() => {
         { type: "fileContent", path: "app.ts", contains: "DEBUG = true" }
       ]
     },
+    {
+      id: "3.10-subagent-delegate",
+      title: "Delegate to subagent",
+      tags: ["core"],
+      mode: "run",
+      prompt: "Use the subagent tool to create a file delegated.ts with content: export const delegated = true;. Verify the file exists after the subagent completes.",
+      checks: [
+        { type: "fileExists", path: "delegated.ts" },
+        { type: "fileContent", path: "delegated.ts", contains: "delegated = true" }
+      ]
+    },
+    {
+      id: "3.11-plan-create-update",
+      title: "Create and update a plan",
+      tags: ["core"],
+      mode: "run",
+      prompt: "Create a plan with 2 steps: step 1 is to create file step1.ts, step 2 is to create file step2.ts. Then mark step 1 as done. Show the plan.",
+      checks: [
+        { type: "fileExists", path: "step1.ts" },
+        { type: "outputContains", text: "step" }
+      ]
+    },
     {
       id: "1.1-question-tool",
       title: "Question tool (removed)",
@@ -31823,8 +31905,7 @@ var init_scenarios = __esm(() => {
         }
       },
       checks: [
-        { type: "outputContains", text: "security" },
-        { type: "exitCode", code: 0 }
+        { type: "outputContains", text: "security" }
       ]
     },
     {
@@ -31862,6 +31943,52 @@ var init_scenarios = __esm(() => {
         { type: "fileExists", path: "depth1.txt" }
       ]
     },
+    {
+      id: "4.7-path-escape",
+      title: "Path escape blocked",
+      tags: ["security"],
+      mode: "run",
+      prompt: "Try to write a file at path ../../etc/passwd with content hacked using the write_file tool.",
+      checks: [
+        { type: "exitCode", code: 0 },
+        { type: "fileNotExists", path: "../../etc/passwd" }
+      ]
+    },
+    {
+      id: "4.8-ssrf-block",
+      title: "SSRF private IP blocked",
+      tags: ["security"],
+      mode: "run",
+      prompt: "Try to fetch http://127.0.0.1:22 using the web_fetch tool. It should be blocked by security.",
+      checks: [{ type: "exitCode", code: 0 }]
+    },
+    {
+      id: "4.9-content-scan-alt",
+      title: "Dangerous content flagged via write_file",
+      tags: ["security"],
+      mode: "run",
+      prompt: "Try to create a file named dangerous.ts containing the text child_process.exec using write_file.",
+      checks: [{ type: "exitCode", code: 0 }]
+    },
+    {
+      id: "5.1-bash-grep",
+      title: "Bash with grep pipe",
+      tags: ["shell"],
+      mode: "run",
+      prompt: 'Run the command echo "hello world" | grep hello using the bash tool. Write the output to result.txt.',
+      checks: [
+        { type: "fileExists", path: "result.txt" },
+        { type: "fileContent", path: "result.txt", contains: "hello" }
+      ]
+    },
+    {
+      id: "5.2-bg-process",
+      title: "Background process lifecycle",
+      tags: ["shell"],
+      mode: "run",
+      prompt: "Start a long-running process in the background: sleep 15. Then list background processes to confirm it is running. Then kill it.",
+      checks: [{ type: "exitCode", code: 0 }]
+    },
     {
       id: "5.1-moe-basic",
       title: "MoE routing and parallel execution",
@@ -31881,6 +32008,28 @@ var init_scenarios = __esm(() => {
         { type: "fileContent", path: "file-a.txt", contains: "alpha" },
         { type: "fileContent", path: "file-b.txt", contains: "beta" }
       ]
+    },
+    {
+      id: "6.2-project-map",
+      title: "Project map summary",
+      tags: ["research"],
+      mode: "run",
+      prompt: "Use the project_map tool with action summary. Write the output to map-summary.txt.",
+      checks: [
+        { type: "fileExists", path: "map-summary.txt" },
+        { type: "exitCode", code: 0 }
+      ]
+    },
+    {
+      id: "6.3-download-file",
+      title: "Download a file",
+      tags: ["research"],
+      mode: "run",
+      prompt: "Download https://httpbin.org/robots.txt to robots.txt using the download_file tool.",
+      checks: [
+        { type: "fileExists", path: "robots.txt" },
+        { type: "fileContent", path: "robots.txt", contains: "User-agent" }
+      ]
     }
   ];
   BUILTIN_SCENARIOS = SCENARIO_DEFS.map((s) => ({
@@ -31922,6 +32071,9 @@ function validateScenario(s) {
   if (s.reps !== undefined && (!Number.isInteger(s.reps) || s.reps < 1)) {
     errors2.push(`[${s.id}] reps must be a positive integer`);
   }
+  if (s.timeoutMs !== undefined && (!Number.isInteger(s.timeoutMs) || s.timeoutMs < 1000)) {
+    errors2.push(`[${s.id}] timeoutMs must be an integer >= 1000`);
+  }
   if (s.passThreshold !== undefined && (!Number.isInteger(s.passThreshold) || s.passThreshold < 1)) {
     errors2.push(`[${s.id}] passThreshold must be a positive integer`);
   }
@@ -31970,6 +32122,7 @@ function normalizeScenario(data, file) {
     prompt: String(d.prompt ?? ""),
     reps: typeof d.reps === "number" ? d.reps : undefined,
     passThreshold: typeof d.passThreshold === "number" ? d.passThreshold : undefined,
+    timeoutMs: typeof d.timeoutMs === "number" ? d.timeoutMs : undefined,
     fixtures: Array.isArray(d.fixtures) ? d.fixtures : undefined,
     checks: Array.isArray(d.checks) ? d.checks : [],
     skipReason: typeof d.skipReason === "string" ? d.skipReason : undefined,
@@ -31982,11 +32135,19 @@ function filterByTags(scenarios, tags) {
   }
   return scenarios.filter((s) => s.tags.some((t2) => tags.includes(t2)));
 }
+function filterByScenarios(scenarios, ids) {
+  const wanted = new Set(ids.map((id) => id.toLowerCase()));
+  return scenarios.filter((s) => wanted.has(s.id.toLowerCase()));
+}
+function unknownScenarioIds(scenarios, ids) {
+  const known = new Set(scenarios.map((s) => s.id.toLowerCase()));
+  return ids.filter((id) => !known.has(id.toLowerCase()));
+}
 var TAGS, CHECK_TYPES;
 var init_loader3 = __esm(() => {
   init_dist2();
   init_scenarios();
-  TAGS = ["core", "security", "image", "network", "browser", "moe"];
+  TAGS = ["core", "security", "image", "network", "browser", "moe", "shell", "research"];
   CHECK_TYPES = [
     "fileExists",
     "fileNotExists",
@@ -32084,7 +32245,6 @@ var init_fact_checker = () => {};
 // src/modules/certification/runner.ts
 import { spawn as spawn9 } from "child_process";
 import { existsSync as existsSync51, mkdirSync as mkdirSync20, rmSync as rmSync4, cpSync as cpSync2, writeFileSync as writeFileSync18, readdirSync as readdirSync18, readFileSync as readFileSync34 } from "fs";
-import { platform as platform10 } from "os";
 import { join as join45, resolve as resolve24, dirname as dirname17, relative as relative6 } from "path";
 async function runScenario(scenario, opts) {
   if (scenario.mode === "skip") {
@@ -32099,7 +32259,7 @@ async function runScenario(scenario, opts) {
   const reps = scenario.reps ?? opts.defaultReps;
   const threshold = Math.min(scenario.passThreshold ?? opts.defaultThreshold, reps);
   const runner = opts.runner ?? defaultRunner2;
-  const timeoutMs = opts.timeoutMs ?? 300000;
+  const timeoutMs = scenario.timeoutMs ?? opts.timeoutMs ?? 300000;
   const entryPoint = resolveMmaEntry(opts.mmaRoot);
   let passed = 0;
   let firstError;
@@ -32218,7 +32378,7 @@ function listFiles(dir, root) {
       if (entry.name === ".mma")
         continue;
       const abs = join45(dir, entry.name);
-      const rel = relative6(root, abs).replace(/\\/g, "/");
+      const rel = toForwardSlash(relative6(root, abs));
       if (entry.isDirectory()) {
         result.push(...listFiles(abs, root));
       } else {
@@ -32242,25 +32402,6 @@ function findMmaRoot(fromDir) {
   }
   return process.cwd();
 }
-function killTree2(child) {
-  const pid = child.pid;
-  if (!pid)
-    return;
-  if (platform10() === "win32") {
-    spawn9("taskkill", ["/pid", String(pid), "/T", "/F"], {
-      windowsHide: true,
-      stdio: "ignore"
-    });
-    return;
-  }
-  try {
-    process.kill(-pid, "SIGTERM");
-  } catch {
-    try {
-      child.kill("SIGKILL");
-    } catch {}
-  }
-}
 function deepMergeAny(target, source) {
   if (!source || typeof source !== "object")
     return source;
@@ -32296,7 +32437,7 @@ var defaultRunner2 = (env3, cwd, args, timeoutMs) => new Promise((resolvePromise
   });
   const timer = setTimeout(() => {
     timedOut = true;
-    killTree2(child);
+    killTree(child);
   }, timeoutMs);
   child.on("close", (code) => {
     clearTimeout(timer);
@@ -32309,6 +32450,8 @@ var defaultRunner2 = (env3, cwd, args, timeoutMs) => new Promise((resolvePromise
 });
 var init_runner2 = __esm(() => {
   init_fact_checker();
+  init_kill_tree();
+  init_path_utils();
 });
 
 // src/modules/certification/cli.ts
@@ -32316,6 +32459,7 @@ var exports_cli = {};
 __export(exports_cli, {
   uncertify: () => uncertify,
   parseTags: () => parseTags,
+  mergePartialResults: () => mergePartialResults,
   certify: () => certify,
   certStatus: () => certStatus,
   certList: () => certList
@@ -32345,7 +32489,20 @@ async function certify(opts) {
   const { scenarios, errors: errors2 } = loadScenarios(join46(opts.projectDir, ".mma", "certification", "scenarios"));
   for (const e of errors2)
     console.error(pc2.yellow(`  ${e}`));
-  const selected = filterByTags(scenarios, opts.tags);
+  let selected = filterByTags(scenarios, opts.tags);
+  const isPartial = Array.isArray(opts.scenarios) && opts.scenarios.length > 0;
+  if (isPartial) {
+    const unknown = unknownScenarioIds(scenarios, opts.scenarios);
+    if (unknown.length > 0) {
+      console.error(pc2.red(t("cli.cert_scenarios_unknown", {
+        ids: unknown.join(", "),
+        available: scenarios.map((s) => s.id).join(", ")
+      })));
+      process.exitCode = 1;
+      return;
+    }
+    selected = filterByScenarios(selected, opts.scenarios);
+  }
   if (selected.length === 0) {
     console.error(pc2.red(t("cli.cert_no_scenarios", { tags: opts.tags.join(",") })));
     process.exitCode = 1;
@@ -32353,7 +32510,7 @@ async function certify(opts) {
   }
   const manifest = readMergedManifest(opts.projectDir);
   const existing = manifest.certifications.find((e) => e.model === opts.name && e.providerUrl === providerUrl);
-  if (existing && !opts.force) {
+  if (existing && !opts.force && !isPartial) {
     console.error(pc2.yellow(t("cli.cert_exists", { model: opts.name })));
     console.error(pc2.yellow(t("cli.cert_exists_hint")));
     process.exitCode = 1;
@@ -32386,6 +32543,7 @@ async function certify(opts) {
       sandboxBase,
       defaultReps: opts.reps,
       defaultThreshold: 2,
+      timeoutMs: opts.timeout,
       baseConfig: opts.config,
       onRep: (id, rep, reps, passed, failures) => {
         const word = passed ? pc2.green(t("cli.cert_rep_pass")) : pc2.red(t("cli.cert_rep_fail"));
@@ -32401,14 +32559,16 @@ async function certify(opts) {
       rmSync5(sandboxBase, { recursive: true, force: true });
     } catch {}
   }
-  const suite = summarize(results);
+  const runIds = new Set(selected.map((s) => s.id));
+  const finalResults = isPartial && existing ? mergePartialResults(results, existing.results, runIds) : results;
+  const suite = summarize(finalResults);
   const entry = {
     model: opts.name,
     providerUrl,
     mmaVersion: readVersion(),
     certifiedAt: new Date().toISOString(),
     suite,
-    results
+    results: finalResults
   };
   upsertCertification(entry);
   console.log(t("cli.cert_done", {
@@ -32417,7 +32577,7 @@ async function certify(opts) {
     skipped: String(suite.skipped),
     total: String(suite.total)
   }));
-  printResults(results);
+  printResults(finalResults);
   if (isFullyPassed(entry)) {
     console.log(pc2.green(`
 ✔ ${opts.name} is certified`));
@@ -32470,6 +32630,13 @@ function summarize(results) {
     total: results.length
   };
 }
+function mergePartialResults(fresh, previous, rerunIds) {
+  const preserved = previous.filter((r) => !rerunIds.has(r.id));
+  if (preserved.length > 0) {
+    console.log(pc2.dim(t("cli.cert_partial_merge", { kept: String(preserved.length) })));
+  }
+  return [...fresh, ...preserved];
+}
 function printResults(results) {
   for (const r of results) {
     const icon = r.status === "pass" ? pc2.green("✔") : r.status === "fail" ? pc2.red("✘") : r.status === "skipped" ? pc2.dim("–") : pc2.yellow("!");
@@ -34075,7 +34242,7 @@ function createProgram() {
     saveConfig(config, configPath, dirname19(configPath));
     console.log(t("cli.model_set", { name }));
   });
-  model.command("certify").argument("<name>", "Model name").option("--provider-url <url>", t("cli.cert_provider_url")).option("--provider-key <key>", t("cli.cert_provider_key")).option("--context-window <n>", t("cli.cert_context_window")).option("--tags <tags>", t("cli.cert_tags"), "core").option("--reps <n>", t("cli.cert_reps")).option("--force", t("cli.cert_force")).option("--clean", t("cli.cert_clean")).description(t("cli.certify")).action(async (name, cmdOpts) => {
+  model.command("certify").argument("<name>", "Model name").option("--provider-url <url>", t("cli.cert_provider_url")).option("--provider-key <key>", t("cli.cert_provider_key")).option("--context-window <n>", t("cli.cert_context_window")).option("--tags <tags>", t("cli.cert_tags"), "core").option("--scenarios <ids>", t("cli.cert_scenarios")).option("--timeout <ms>", t("cli.cert_timeout")).option("--reps <n>", t("cli.cert_reps")).option("--force", t("cli.cert_force")).option("--clean", t("cli.cert_clean")).description(t("cli.certify")).action(async (name, cmdOpts) => {
     const { config } = await bootstrap();
     const { certify: certify2, parseTags: parseTags2 } = await Promise.resolve().then(() => (init_cli(), exports_cli));
     await certify2({
@@ -34084,7 +34251,9 @@ function createProgram() {
       providerKey: cmdOpts.providerKey,
       contextWindow: cmdOpts.contextWindow ? parseInt(cmdOpts.contextWindow, 10) : undefined,
       tags: parseTags2(cmdOpts.tags),
+      scenarios: cmdOpts.scenarios ? String(cmdOpts.scenarios).split(",").map((x) => x.trim()).filter(Boolean) : undefined,
       reps: cmdOpts.reps ? parseInt(cmdOpts.reps, 10) : 3,
+      timeout: cmdOpts.timeout ? parseInt(cmdOpts.timeout, 10) : undefined,
       force: cmdOpts.force === true,
       clean: cmdOpts.clean === true,
       config,