Add inference_snapshot() that pulls loaded models from /api/ps and GPU
stats over SSH, then surface as facts: model name/quant/VRAM, processor
split, TTL countdown, and a hot/idle gauge for the inference GPU. Doge
can now riff on the LLM box too.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New widget queries maeth-inference:11434/api/{version,ps} and renders ollama
status, loaded model (params/quant/size/ctx/TTL/processor split) and totals
across all loaded models. Host CPU/mem/GPU still come over SSH best-effort;
gracefully degrades to "ssh not authorized" when the key is missing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
qwen2.5:1.5b kept dropping "Netflix" when it saw show titles in the facts
because it has no concept of Jellyfin in training. Add an explicit rule:
everything is local self-hosted Jellyfin, never name a commercial streamer.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add 4 new mood variants (humor negro, nihilista cómico, edgelord cariñoso,
fatalista chileno). Reframe the prompt so DogeBot calls the user "Maeth"
(occasionally Francisco) and is allowed to lean into dark/irreverent jokes
without being cruel.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three conky panels (main, dayz, doge) plus all widgets, NUT config templates,
and the systemd timer that drives the qwen2.5:1.5b vibe report on maeth-storage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
