Save Files Locally
To save the Benchmark’s declared files even when evaluation reuses the execution Environment, add this to your configuration:--config, or add --execution-params '{"save_artifacts":true}' to your existing run command. The Benchmark must support artifact collection. Saved files appear under the attempt’s local artifacts/ directory; see the result directory layout.
Saving files does not retain a running sandbox. Use --keep-environment when you need to inspect the Environment itself.
Understand Saving and Restoration
The Environment’s evaluation mode determines the default behavior:
Omitting
save_artifacts or setting it to null uses the mode’s default. Disabling downloads does not skip declared preparation commands or Harness result collection. See evaluation environments for mode configuration.
Customize Submission Files
Only override paths or preparation commands when you know what the Benchmark’s evaluator expects. Both lists replace the Benchmark’s declarations, so include every required original entry in your replacement.
For example, if your Benchmark expects
/app/submission.tar.gz and the agent writes its submission under /app/submission, prepare and save that archive:
artifacts/submission.tar.gz. Fresh evaluation restores it to its original source path.
Use absolute sandbox paths for source and relative local paths for destination. An optional exclude list filters files by pattern. Preparation commands run in the Environment’s default working directory; use absolute paths or an explicit cd for task-specific locations. Missing sources are recorded in the artifact manifest without failing collection.
Limit Preparation and Transfer
Change these limits when submission preparation or transferring large files exceeds the defaults:timeout_seconds, as in the submission example above.
For preparation, the first of the command or group deadlines wins. A batch download and a later restore each start a new transfer clock; at the defaults, each can use 600 seconds. Skipped transfers consume no budget. These limits do not extend the overall evaluation timeout.
Preparation or transfer failure preserves validated files and execution diagnostics, but skips grading and creation of a resumable evaluation checkpoint for that attempt. Check the task’s issues and transfer progress before increasing a limit.
