@wix/evalforge-evaluator 0.129.0 → 0.130.0

package/build/index.js

@@ -2327,7 +2327,12 @@ async function buildOpenCodeEnv(options) {
 // src/run-scenario/agents/opencode/build-trace.ts
 var import_evalforge_types7 = require("@wix/evalforge-types");
 var import_crypto2 = require("crypto");
+function toCanonicalModelId(modelId) {
+  const slashIndex = modelId.indexOf("/");
+  return slashIndex > 0 ? modelId.slice(slashIndex + 1) : modelId;
+}
 function buildLLMTrace(timestampedEvents, totalDurationMs, model, provider, executionStartTime) {
+  const canonicalModel = toCanonicalModelId(model);
   const turns = [];
   let current = {
     textParts: [],
@@ -2377,7 +2382,7 @@ function buildLLMTrace(timestampedEvents, totalDurationMs, model, provider, exec
     const stepOutputTokens = sf?.tokens.output ?? 0;
     const stepCost = sf?.cost ?? 0;
     const finishReason = sf?.reason ?? "unknown";
-    const stepModel = sf?.modelID || model;
+    const stepModel = toCanonicalModelId(sf?.modelID || model);
     const stepProvider = sf?.providerID || provider;
     const turnEndMs = turn.receivedAt ?? executionStartMs + totalDurationMs;
     const prevEndMs = turnIndex > 0 ? turns[turnIndex - 1].receivedAt ?? executionStartMs : executionStartMs;
@@ -2528,7 +2533,7 @@ function buildLLMTrace(timestampedEvents, totalDurationMs, model, provider, exec
     entry.costUsd += step.costUsd;
     stepTypeBreakdown[step.type] = entry;
   }
-  const modelUsed = allSteps[0]?.model || model;
+  const modelUsed = allSteps[0]?.model || canonicalModel;
   const summary = {
     totalSteps: allSteps.length,
     totalTurns: turns.length,
@@ -2720,6 +2725,16 @@ async function prepareOpenCodeEnvironment(cwd, skills, options) {
       `Failed to write skills to filesystem: ${writeError instanceof Error ? writeError.message : String(writeError)}`
     );
   }
+  let systemPrompt;
+  if (options.systemPrompt === null || options.systemPrompt === "") {
+  } else if (options.systemPrompt != null) {
+    systemPrompt = options.systemPrompt;
+  } else {
+    systemPrompt = import_evalforge_types8.DEFAULT_EVALUATOR_SYSTEM_PROMPT;
+  }
+  if (systemPrompt) {
+    await writeSystemPromptRule(cwd, systemPrompt);
+  }
 }
 async function writeSystemPromptRule(cwd, systemPrompt) {
   const rulesDir = (0, import_path10.join)(cwd, ".opencode", "rules");
@@ -3053,16 +3068,6 @@ async function executeWithOpenCode(skills, scenario, options) {
       traceContext.authToken
     );
   }
-  let systemPrompt;
-  if (options.systemPrompt === null || options.systemPrompt === "") {
-  } else if (options.systemPrompt != null) {
-    systemPrompt = options.systemPrompt;
-  } else {
-    systemPrompt = import_evalforge_types8.DEFAULT_EVALUATOR_SYSTEM_PROMPT;
-  }
-  if (systemPrompt) {
-    await writeSystemPromptRule(options.cwd, systemPrompt);
-  }
   const baseArgs = [
     "run",
     "--format",
@@ -3240,7 +3245,8 @@ var OpenCodeAdapter = class {
     await prepareOpenCodeEnvironment(context.cwd, context.skills, {
       mcps: context.mcps,
       subAgents: context.subAgents,
-      rules: context.rules
+      rules: context.rules,
+      systemPrompt: context.systemPrompt
     });
   }
   async execute(context) {