Tweaks to eval output: add github link to commit, print spec, quiet some git operations

Author: jahooma

evals/git-evals/pick-commits.ts

@@ -295,7 +295,7 @@ function basicFilter(commits: CommitInfo[]): CommitInfo[] {
   })
 }
 
-function createGithubUrl(repoUrl: string, sha: string): string {
+export function createGithubUrl(repoUrl: string, sha: string): string {
   // Convert repo URL to GitHub commit URL
   let baseUrl = repoUrl
   if (baseUrl.endsWith('.git')) {

evals/git-evals/run-git-evals.ts

@@ -13,6 +13,7 @@ import pLimit from 'p-limit'
 import { resetRepoToCommit } from '../scaffolding'
 import { createInitialSessionState } from '../test-setup'
 import { judgeEvalRun } from './judge-git-eval'
+import { createGithubUrl } from './pick-commits'
 import { ClaudeRunner } from './runners/claude'
 import { CodebuffRunner } from './runners/codebuff'
 import { extractRepoNameFromUrl, setupTestRepo } from './setup-test-repo'
@@ -505,6 +506,19 @@ export async function runGitEvals(
                   console.log(
                     `Completed eval for commit ${testRepoName} - ${evalCommit.spec.split('\n')[0]}`,
                   )
+
+                  // Repeat spec before printing judge ruling
+                  console.log(`Spec for commit ${testRepoName}:`)
+                  console.log(evalCommit.spec)
+
+                  // Print GitHub commit link after judging results
+                  try {
+                    const githubUrl = createGithubUrl(repoUrl, evalCommit.sha)
+                    console.log(`GitHub commit: ${githubUrl}`)
+                  } catch (error) {
+                    // Ignore errors if URL generation fails (e.g., non-GitHub repos)
+                  }
+
                   if (!logToStdout) {
                     const finalResult = cloneDeep(message.result)
                     for (const cbTrace of finalResult.trace) {

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

@@ -12,6 +12,7 @@ import { Command, Flags } from '@oclif/core'
 
 import { createFileReadingMock } from '../scaffolding'
 import { setupTestEnvironmentVariables } from '../test-setup'
+import { createGithubUrl } from './pick-commits'
 import { runSingleEval } from './run-git-evals'
 import { extractRepoNameFromUrl, setupTestRepo } from './setup-test-repo'
 
@@ -203,6 +204,18 @@ async function runSingleEvalTask(options: {
     const duration = Date.now() - startTime
     console.log(`✅ Evaluation completed in ${(duration / 1000).toFixed(1)}s`)
 
+    // Repeat spec before printing judge ruling
+    console.log(`Spec for commit:`)
+    console.log(evalCommit.spec)
+
+    // Print GitHub commit link after judging results
+    try {
+      const githubUrl = createGithubUrl(evalData.repoUrl, evalCommit.sha)
+      console.log(`🔗 GitHub commit: ${githubUrl}`)
+    } catch (error) {
+      // Ignore errors if URL generation fails (e.g., non-GitHub repos)
+    }
+
     // Display results
     if (result.error) {
       console.log(`❌ Error occurred: ${result.error}`)

evals/git-evals/setup-test-repo.ts

@@ -135,15 +135,19 @@ export async function setupTestRepo(
       // Local development or public repos
       console.log(`Local environment detected - cloning from: ${repoUrl}`)
 
-      execFileSync('git', ['clone', '--no-checkout', repoUrl, repoDir], {
-        timeout: 120_000, // 2 minute timeout for cloning
-        stdio: 'inherit',
-      })
-      execFileSync('git', ['fetch', 'origin', commitSha], {
+      execFileSync(
+        'git',
+        ['clone', '--no-checkout', '--quiet', repoUrl, repoDir],
+        {
+          timeout: 120_000, // 2 minute timeout for cloning
+          stdio: 'inherit',
+        },
+      )
+      execFileSync('git', ['fetch', 'origin', '--quiet', commitSha], {
         cwd: repoDir,
         stdio: 'inherit',
       })
-      execFileSync('git', ['checkout', commitSha], {
+      execFileSync('git', ['checkout', '--quiet', commitSha], {
         cwd: repoDir,
         stdio: 'inherit',
       })