training-monitor

Team role template for reviewing monitoring reports. Checks reasoning PROCESS and logical coherence. Spot-checks the load-bearing claim for ground truth. Catches lazy, shallow, or logically inconsistent analysis.

Heuristics for monitoring multi-GPU and multi-process distributed training. Common patterns, NCCL diagnostics, known failure modes. Reference knowledge, not rules.

Heuristics for monitoring GRPO, PPO, and other RL training. Common patterns, typical indicators, known failure modes. Reference knowledge, not rules.

Heuristics for monitoring training jobs on Kubernetes. Common patterns, pod anomalies, scheduling failures, escalation ladder. Reference knowledge, not rules.

Interactive setup wizard for the training-monitor plugin. Goal-driven — the agent determines which dependencies are needed by checking project context and asking the user only when the context is ambiguous. Installs missing dependencies and reports available capabilities.

Prediction-first monitoring for ML/DL training jobs. Single-agent execution with reviewer sub-agent. Derives judgment criteria from training artifacts, not hardcoded rules.