@wix/evalforge-evaluator 0.21.0 → 0.22.0

package/build/index.js

@@ -6285,6 +6285,9 @@ async function executeWithClaudeCode(skill, scenario, options) {
     })
   );
   let messageCount = 0;
+  const canUseTool = async () => {
+    return { behavior: "allow" };
+  };
   const queryOptions = {
     env: sdkEnv,
     cwd: options.cwd,
@@ -6294,8 +6297,10 @@ async function executeWithClaudeCode(skill, scenario, options) {
     maxTurns,
     maxThinkingTokens: options.maxThinkingTokens,
     mcpServers: options.mcpServers,
-    permissionMode: "bypassPermissions",
-    allowDangerouslySkipPermissions: true
+    // Use 'default' permission mode with custom canUseTool handler
+    // instead of 'bypassPermissions' which fails on root
+    permissionMode: "default",
+    canUseTool
   };
   if (options.temperature !== void 0) {
     queryOptions.temperature = options.temperature;
@@ -6314,8 +6319,8 @@ async function executeWithClaudeCode(skill, scenario, options) {
   console.log("[SDK-DEBUG]   maxTokens:", queryOptions.maxTokens);
   console.log("[SDK-DEBUG]   permissionMode:", queryOptions.permissionMode);
   console.log(
-    "[SDK-DEBUG]   allowDangerouslySkipPermissions:",
-    queryOptions.allowDangerouslySkipPermissions
+    "[SDK-DEBUG]   canUseTool:",
+    queryOptions.canUseTool ? "custom handler (auto-allow)" : "not set"
   );
   console.log("[SDK-DEBUG]   settingSources:", queryOptions.settingSources);
   console.log("[SDK-DEBUG]   allowedTools:", queryOptions.allowedTools);
@@ -7440,8 +7445,8 @@ async function testClaudeDirectExecution(config) {
     15e3
   );
   const fullCmdResult = await runAndLog(
-    "claude -p (full SDK style)",
-    `${envExports} && "${claudePath}" -p "Say hello" --output-format json --dangerously-skip-permissions 2>&1`,
+    "claude -p (with json output)",
+    `${envExports} && "${claudePath}" -p "Say hello" --output-format json 2>&1`,
     45e3
   );
   const simpleCmdResult = await runAndLog(