@@ -11,11 +11,11 @@ import pLimit from 'p-limit'
1111
1212import { resetRepoToCommit } from '../scaffolding'
1313import { createInitialSessionState } from '../test-setup'
14+ import { judgeEvalRun } from './judge-git-eval'
1415import { ClaudeRunner } from './runners/claude'
1516import { CodebuffRunner } from './runners/codebuff'
1617import { extractRepoNameFromUrl , setupTestRepo } from './setup-test-repo'
1718import { AgentDecisionSchema } from './types'
18- import { judgeEvalRun } from './judge-git-eval'
1919
2020import type { AgentStep } from '../scaffolding'
2121import type { Runner } from './runners/runner'
@@ -28,8 +28,8 @@ import type {
2828 FullEvalLog ,
2929 EvalData ,
3030} from './types'
31- import type { z } from 'zod/v4'
3231import type { ChildProcess } from 'child_process'
32+ import type { z } from 'zod/v4'
3333
3434disableLiveUserInputCheck ( )
3535
@@ -487,7 +487,12 @@ export async function runGitEvals(
487487 `Completed eval for commit ${ testRepoName } - ${ evalCommit . spec . split ( '\n' ) [ 0 ] } ` ,
488488 )
489489 if ( ! logToStdout ) {
490- console . log ( `${ JSON . stringify ( message . result , null , 2 ) } ` )
490+ const finalResult = message . result
491+ for ( const cbTrace of finalResult . trace ) {
492+ delete ( cbTrace as any ) . steps
493+ }
494+ delete ( finalResult . eval_commit as any ) . fileStates
495+ console . log ( `${ JSON . stringify ( finalResult , null , 2 ) } ` )
491496 }
492497 resolve ( message . result )
493498 } else if ( message . type === 'error' ) {
0 commit comments