cool, i'll check out termalin!

yes indeed! do you have a setup for this? is it something similar?

It's strictly read-only, no write access to the Pi at all, and that's staying true. it now logs the actual PromQL queries and raw values behind every report, so a claim like a queue-wait spike is checkable, not just asserted.

Cool, I’ll check it out!

Good call! right now agent-monitor only sees request-level metrics, nothing about what's happening inside Ollama itself. Piping its logs into the Loki setup from the hardware-baseline episode would catch failures that never show up as a metric. Adding that to the list.

Any experience adding a webhook to Ollama? Any pointers in how to set it up? i can watch requests queue and drain in real time, watching it in grafana

Hardware's covered by the Grafana stack (Prometheus/Loki/Tempo) already on the Pi. For agent-level, still deciding. Considered MLflow or something simpler like watching the GPU at infra layer and some small queue in front of the models counting awaits and completes. What's worked for you?