Agent Eval Lab first searches for agent/trajectory.json. Some Harbor versions can finish a trial but fail to convert the trajectory to ATIF.
When no ATIF file exists, Agent Eval Lab searches the native session directory for a supported JSONL file.
Keep the native session files until you build the report. Check provenance.format in the report to identify the adapter.
Confirm that Codex works in the WSL distribution that runs Harbor. Use -CodexAuthJson to pass the WSL path to the authentication file.
Do not copy the authentication file into the task or repository.
The example runner uses examples/docker-public-config/config.json. This configuration does not use a host credential helper for public image pulls.
If Docker Desktop reports concurrent helper errors, run the trials one at a time.
Do not add private registry credentials to the example configuration.
Harbor runs in WSL in this example. The PowerShell runner converts C:\work\repo to /mnt/c/work/repo.
The runner rejects a path that does not have a local drive letter.
The source event has no explicit result status. The analyzer does not infer a status from result text.
Check the producer export. If the format has a documented status field, add the field to its adapter.
Not every producer records timestamps, tool durations, tokens, or costs. The analyzer reports only available fields. It does not estimate missing values.
The public task is a teaching fixture. Its verifier is public.
For an active evaluation, keep the verifier in a private repository. Make it available only during the verifier phase.