Skip to main content
Stream updates until the run completes or disconnects. Full syntax for each command below is in the run fetch, run status, run errors, and run list reference, and in client.runs.stream for Python.

​
Fetch one run

Use a one-time text or machine-readable snapshot when streaming is not needed:
Connected text output shows run identity and metadata before progress updates. JSONL begins with a snapshot record, followed by update records. Recognized termination emits complete, error, stopped, disconnect, or interrupted. Exit code 0 means the CLI handled the tracker response or stream event. It does not mean the benchmark succeeded. Inspect the run status and each JSONL record’s event.

​
Machine-readable output

Unexpected clean JSONL exhaustion emits a final disconnect record and exits nonzero. Transport or malformed-protocol failures can exit nonzero without a final record. In fetch output, metadata_available: false means identity metadata could not load. Batch finished_tasks includes FINISHED, ERROR, and STOPPED tasks.

​
Check several run IDs

Batch status preserves requested order, ignores duplicate IDs, and requests at most 50 IDs at a time. Missing or inaccessible IDs appear in missing_run_ids and make the command exit nonzero after it emits the JSON document.

​
Inspect errors

The command shows stored run and current task error messages. Add --format json to consume them from a script.

​
List runs

Filter the history to find a run:
Filters combine, and the interactive list supports [h] for previous, [l] for next, and [q] to quit. Emit all matching runs as one document with:
Machine-readable output uses a versioned allowlist. It excludes stored agent secrets, kwargs, and raw error messages. Optional identity fields and scores can be null; timestamps are UTC ISO 8601 strings, and non-finite scores become null.