Put back run eval set to initial config

Author: jahooma

evals/git-evals/run-eval-set.ts

@@ -72,7 +72,7 @@ class RunEvalSetCommand extends Command {
     }),
     agent: Flags.string({
       description: 'Codebuff agent id to use',
-      default: 'base-layer', // hi
+      default: 'base-lite',
     }),
     help: Flags.help({ char: 'h' }),
   }
@@ -134,21 +134,21 @@ async function runEvalSet(options: {
       evalDataPath: path.join(__dirname, 'eval-codebuff2.json'),
       outputDir,
     },
-    // {
-    //   name: 'manifold',
-    //   evalDataPath: path.join(__dirname, 'eval-manifold2.json'),
-    //   outputDir,
-    // },
-    // {
-    //   name: 'plane',
-    //   evalDataPath: path.join(__dirname, 'eval-plane.json'),
-    //   outputDir,
-    // },
-    // {
-    //   name: 'saleor',
-    //   evalDataPath: path.join(__dirname, 'eval-saleor.json'),
-    //   outputDir,
-    // },
+    {
+      name: 'manifold',
+      evalDataPath: path.join(__dirname, 'eval-manifold2.json'),
+      outputDir,
+    },
+    {
+      name: 'plane',
+      evalDataPath: path.join(__dirname, 'eval-plane.json'),
+      outputDir,
+    },
+    {
+      name: 'saleor',
+      evalDataPath: path.join(__dirname, 'eval-saleor.json'),
+      outputDir,
+    },
   ]
 
   console.log(`Running ${evalConfigs.length} evaluations:`)

evals/git-evals/run-single-eval-process.ts

@@ -74,7 +74,7 @@ async function main() {
       fingerprintId,
       codingAgent as any,
       agent,
-      true,
+      false,
     )
 
     // Check again after long-running operation

evals/git-evals/run-single-eval.ts

@@ -199,7 +199,7 @@ async function runSingleEvalTask(options: {
       fingerprintId,
       codingAgent,
       agentType,
-      true,
+      false,
     )
 
     const duration = Date.now() - startTime