Pavilion-devs

GitHub profile1 plugin in the catalog

Plugins by Pavilion-devs

1 plugins by Pavilion-devs in the catalog, most-starred first.

Pavilion-devsnightshiftHermes takes the night shift on your ML training runs: Nightshift watches a run's metrics (stdout, JSONL, CSV or TensorBoard), catches NaN, loss spikes, divergence, out-of-memory crashes and stalls, recovers the run within a playbook you approved (roll back to the last healthy checkpoint with a lower learning rate, or restart with half the batch), and pushes the chart to your phone. Desktop pane, dashboard tab, /nightshift, hermes nightshift and a morning-report blueprint. Zero runtime dependencies. Disclosure — the agent can only propose a training run; after your explicit approval (/nightshift approve in chat or the Approve button; there is no CLI approve) it launches the approved command as a long-running background process from a background thread that starts with Hermes, with a minimal environment (no Hermes keys unless you name them in env_passthrough, shown on the approval card), and may stop or restart that process group, re-rendering only the approved command with the playbook's parameters (lr, batch size, grad accumulation, checkpoint path) within an intervention budget; it signals only process groups it started (start-time checked, never a run it cannot verify), reads the log, metric and checkpoint paths you or the agent give it, runs nvidia-smi and ps, stores run data in its plugin data folder and alert charts in Hermes's image cache, makes capped LLM calls on your configured model with a redacted log tail on critical incidents, and sends alerts through Hermes's own `hermes send` to your messaging platform; no network calls of its own.Communitytools