@wix/evalforge-evaluator 0.96.0 → 0.97.0

package/build/index.mjs

@@ -680,6 +680,7 @@ import { AgentRunCommand } from "@wix/evalforge-types";
 // src/run-scenario/agents/claude-code/execute.ts
 import {
   ClaudeModel,
+  DEFAULT_EVALUATOR_SYSTEM_PROMPT,
   LLMStepType,
   LiveTraceEventType,
   TRACE_EVENT_PREFIX
@@ -1128,6 +1129,20 @@ async function executeWithClaudeCode(skills, scenario, options) {
     permissionMode: "default",
     canUseTool
   };
+  if (options.systemPrompt === null || options.systemPrompt === "") {
+  } else if (options.systemPrompt != null) {
+    queryOptions.systemPrompt = {
+      type: "preset",
+      preset: "claude_code",
+      append: options.systemPrompt
+    };
+  } else {
+    queryOptions.systemPrompt = {
+      type: "preset",
+      preset: "claude_code",
+      append: DEFAULT_EVALUATOR_SYSTEM_PROMPT
+    };
+  }
   if (options.temperature !== void 0) {
     queryOptions.temperature = options.temperature;
   }
@@ -1148,6 +1163,7 @@ async function executeWithClaudeCode(skills, scenario, options) {
     "[SDK-DEBUG]   canUseTool:",
     queryOptions.canUseTool ? "custom handler (auto-allow)" : "not set"
   );
+  console.log("[SDK-DEBUG]   systemPrompt:", queryOptions.systemPrompt);
   console.log("[SDK-DEBUG]   settingSources:", queryOptions.settingSources);
   console.log("[SDK-DEBUG]   allowedTools:", queryOptions.allowedTools);
   console.log("[SDK-DEBUG] Calling SDK query()...");
@@ -1245,20 +1261,8 @@ async function executeWithClaudeCode(skills, scenario, options) {
       }, HEARTBEAT_INTERVAL_MS);
     }
     const sdkPromise = (async () => {
-      const evaluatorPromptSuffix = `
-
-IMPORTANT: This is an automated evaluation run. Follow these guidelines:
-1. Execute the requested changes immediately without asking for confirmation.
-2. Do NOT ask "would you like me to proceed?" or similar questions.
-3. Do NOT use the Task tool to delegate simple operations - do them directly yourself.
-4. Keep your approach simple and direct - avoid excessive planning.
-5. Make targeted edits using Read and Edit tools rather than exploring the entire codebase.
-6. If you encounter an error, fix it directly rather than starting over.
-7. Your project root is the current working directory. Always create and modify source code files relative to the project root, NOT inside .claude/skills/ directories.
-8. Before finishing, run the project's package manager install command (e.g. \`npm install\`, \`yarn install\`, or \`pnpm install\` depending on the lockfile present) to ensure all dependencies are installed and the project is ready to build.`;
-      const fullPrompt = scenario.triggerPrompt + evaluatorPromptSuffix;
       for await (const message of query({
-        prompt: fullPrompt,
+        prompt: scenario.triggerPrompt,
         options: queryOptions
       })) {
         messageCount++;
@@ -1769,7 +1773,8 @@ var ClaudeCodeAdapter = class {
       traceContext,
       mcps,
       subAgents,
-      rules
+      rules,
+      systemPrompt
     } = context;
     const modelForSdk = modelConfig?.model;
     const options = {
@@ -1782,7 +1787,8 @@ var ClaudeCodeAdapter = class {
       traceContext,
       mcps,
       subAgents,
-      rules
+      rules,
+      systemPrompt
     };
     const { result, llmTrace } = await executeWithClaudeCode(
       skills,
@@ -2565,7 +2571,8 @@ async function runAgentWithContext(config, evalRunId2, scenario, evalData, workD
     },
     mcps: evalData.mcps.length > 0 ? evalData.mcps : void 0,
     subAgents: evalData.subAgents.length > 0 ? evalData.subAgents : void 0,
-    rules: evalData.rules?.length > 0 ? evalData.rules : void 0
+    rules: evalData.rules?.length > 0 ? evalData.rules : void 0,
+    systemPrompt: agent?.systemPrompt
   };
   const { outputText, durationMs, llmTrace } = await adapter.execute(executionContext);
   const completedAt = (/* @__PURE__ */ new Date()).toISOString();