@wix/evalforge-evaluator 0.53.0 → 0.55.0

package/build/index.js

@@ -24,7 +24,7 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
 ));
 
 // src/index.ts
-var import_evalforge_types5 = require("@wix/evalforge-types");
+var import_evalforge_types6 = require("@wix/evalforge-types");
 
 // src/config.ts
 function loadConfig() {
@@ -349,7 +349,7 @@ async function fetchEvaluationData(api, projectId2, evalRunId2) {
 }
 
 // src/run-scenario/index.ts
-var import_evalforge_types3 = require("@wix/evalforge-types");
+var import_evalforge_types4 = require("@wix/evalforge-types");
 var import_eval_assertions = require("@wix/eval-assertions");
 
 // src/run-scenario/environment.ts
@@ -6337,6 +6337,9 @@ function getAdapter(runCommand) {
   return adapter;
 }
 
+// src/run-scenario/agents/claude-code/claude-code-adapter.ts
+var import_evalforge_types3 = require("@wix/evalforge-types");
+
 // src/run-scenario/agents/claude-code/execute.ts
 var import_evalforge_types2 = require("@wix/evalforge-types");
 var import_crypto = require("crypto");
@@ -7310,10 +7313,11 @@ var ClaudeCodeAdapter = class {
       aiGatewayHeaders,
       traceContext
     } = context;
+    const modelForSdk = modelConfig?.model ? import_evalforge_types3.AVAILABLE_MODELS_MAP[modelConfig.model]?.providerModelId ?? modelConfig.model : void 0;
     const options = {
       cwd,
       systemPrompt: skill.skillMd,
-      model: modelConfig?.model,
+      model: modelForSdk,
       temperature: modelConfig?.temperature,
       maxTokens: modelConfig?.maxTokens,
       aiGatewayUrl,
@@ -8185,10 +8189,10 @@ async function runScenario(config, evalRunId2, scenario, target, template, resol
     assertionContext
   ) : [];
   const passed = assertionResults.filter(
-    (r) => r.status === import_evalforge_types3.AssertionResultStatus.PASSED
+    (r) => r.status === import_evalforge_types4.AssertionResultStatus.PASSED
   ).length;
   const failed = assertionResults.filter(
-    (r) => r.status === import_evalforge_types3.AssertionResultStatus.FAILED
+    (r) => r.status === import_evalforge_types4.AssertionResultStatus.FAILED
   ).length;
   const total = assertionResults.length;
   const passRate = total > 0 ? Math.round(passed / total * 100) : 100;
@@ -8202,7 +8206,7 @@ async function runScenario(config, evalRunId2, scenario, target, template, resol
 }
 
 // src/error-reporter.ts
-var import_evalforge_types4 = require("@wix/evalforge-types");
+var import_evalforge_types5 = require("@wix/evalforge-types");
 function formatError(error, phase, context) {
   const timestamp = (/* @__PURE__ */ new Date()).toISOString();
   if (error instanceof Error) {
@@ -8451,7 +8455,7 @@ async function runEvaluation(projectId2, evalRunId2) {
   };
   try {
     await api.updateEvalRun(projectId2, evalRunId2, {
-      status: import_evalforge_types5.EvalStatus.COMPLETED,
+      status: import_evalforge_types6.EvalStatus.COMPLETED,
       completedAt: (/* @__PURE__ */ new Date()).toISOString()
     });
   } catch (updateErr) {
@@ -8492,7 +8496,7 @@ runEvaluation(projectId, evalRunId).then(() => {
       authToken: config.authToken
     });
     await api.updateEvalRun(projectId, evalRunId, {
-      status: import_evalforge_types5.EvalStatus.FAILED,
+      status: import_evalforge_types6.EvalStatus.FAILED,
       completedAt: (/* @__PURE__ */ new Date()).toISOString(),
       jobError,
       jobStatus: "FAILED"
@@ -8515,7 +8519,7 @@ runEvaluation(projectId, evalRunId).then(() => {
           authToken
         });
         await api.updateEvalRun(projectId, evalRunId, {
-          status: import_evalforge_types5.EvalStatus.FAILED,
+          status: import_evalforge_types6.EvalStatus.FAILED,
           completedAt: (/* @__PURE__ */ new Date()).toISOString(),
           jobError: `Config load failed, then: ${jobError}`,
           jobStatus: "FAILED"