agent-job-info

What problem does it solve?

Tracking the status of distributed ML/HPC experiments requires manually checking multiple disconnected sources including HPC scheduler queues, W&B dashboards, log files, checkpoint directories, and CSV result ledgers, leading to missed failures, stale progress, and wasted time hunting for critical experiment information.

Core Features & Use Cases

• Cross-Source Status Aggregation: Pulls real-time job status from scheduler systems (Slurm for LUMI/Snellius, process/tmux checks for Brev/AutoDL, RunPod API), W&B runs, workspace state files, CSV ledgers, checkpoint directories, and log files into a single unified view.
• Backend-Agnostic Monitoring: Natively supports all DeepResearch sandbox backends (LUMI, Snellius, Brev, AutoDL, RunPod, local) with platform-specific handling for each backend's unique status query requirements and access constraints.
• Standardized Health Reporting: Generates compact, structured health tables with clear status labels (healthy, failed, blocked, stale) and flags missing artifacts or metrics, with optional student-safe redacted versions for sharing with non-operator audiences.
• Use Case: You have a multi-day training run on LUMI that you have not checked recently; use this skill to quickly confirm if it is still running, what the latest computed metrics (such as FID or loss) are, if checkpoints were saved successfully, and if any errors appear in the logs.

Quick Start

Use the agent-job-info skill to generate a full health report for my latest training experiment running on LUMI.