Skip to main content
Follow a run while it goes, find out why tasks failed, and look up earlier runs. Every command is in the valkyrie run reference. To do the same from Python, see Manage runs with Python. In the commands on this page, replace RUN_ID with the run’s ID, which valkyrie run start prints, and TASK_ID with a task’s ID.

​
Follow a run

This prints the run’s details, then its progress until the run finishes. Press Control+C to stop watching; the run keeps going. Without --connect, run fetch prints one snapshot, and --format json prints it as JSON. A run ends as FINISHED, ERROR, or STOPPED. The command exits with 0 whichever way the run ends, so check the status rather than the exit code.

​
Follow a run from a script

Each line is a JSON record with an event field. The first record is a snapshot, then update records follow, and the last is one of complete, error, stopped, disconnect, or interrupted. If the stream ends without a final record, the command exits with 1. In these records, finished_tasks counts tasks that finished, failed, or stopped.

​
Check several runs at once

Replace RUN_ID_1 and RUN_ID_2 with the runs’ IDs. The output keeps the order you gave. IDs that don’t exist or that you can’t see are listed under missing_run_ids, and the command then exits with 1.

​
Find failed tasks

This shows the run’s own errors and each failed task’s error. To look closer:
For scripts, add --format json to any of these.

​
Check how long a task took

The output includes the task’s exit reason and its time breakdown: See Task time limits for what counts toward each limit.

​
Read logs

Narrow the logs to one task, to text, or to a time range. Times need a time zone, and logs at exactly those times are included:
To watch new log lines arrive, add --follow with --task-id. For scripts, add --format jsonl.

​
Search your run history

Filter by --agent-name, --benchmark-name, --model, --dataset, -l for the label, --status, and --started-by, and sort with --order-by asc or desc. Page through the list with h and l, and quit with q. To print every match as one JSON document:
JSON output leaves out agent secrets, parameters, and raw error messages. Missing values and scores that aren’t finite numbers are null, and times are UTC in ISO 8601. valkyrie run filter-options lists the agents, benchmarks, models, and other values your organization’s earlier runs used, which are the values the filters accept.

​
Inspect the priority queue

On a Tracker that queues tasks in shared sandbox pools, see your organization’s waiting and active tasks:
Waiting tasks are ordered by priority, then by when they joined the queue. Positions count every task in the pool, including other organizations’ tasks, which aren’t listed. The command only reads the queue. Each list shows up to 100 entries. When there are more, the output prints the offset of the next page. Pass --waiting-offset and --active-offset to page through each list separately:
The queue changes between calls, so a task can appear on two pages or on none.

​
Browse runs in the terminal

valkyrie ui lists your recent runs with live status. It only reads: stop or retry a run with valkyrie run. To try it on sample data, add --fake.