llm-digest

Trigger the Skill from the llm-tools-daily-scout cron job. It searches Reddit, HuggingFace, and GitHub for the last 24-48 hours of LLM developments, writes a dated Markdown digest to ~/homelab/docs/digests/, and commits and pushes it to the homelab repo.

Before writing, the Skill lists llm-digest-*.md files from the last 7 days and reads each Recommendations section. Models, tools, repos, and quantizations already flagged are skipped unless something genuinely new happened, such as a new benchmark or quantization.

It filters for a Framework Desktop with Ryzen AI MAX+ 395 (Strix Halo, gfx1151, 128 GB unified memory), a Raspberry Pi running Hermes Agent, Ollama and Lemonade Server with Vulkan/ROCm backends, llama.cpp, and Tailscale networking. Cloud-only or API-only findings are excluded.

If the push is rejected because the remote has changes, the workflow runs git pull --rebase and pushes again. If no remote is configured, a preflight check detects it and the digest is committed locally only, with a notification that the repo needs a remote.

According to the bundled references, MTP (Multi-Token Prediction) is fastest on Strix Halo's Vulkan path, reaching about 3.1-3.7x baseline on Qwen 3.6 27B and Gemma 4 31B. DFlash is the easier plug-and-play option when no MTP GGUF exists for a model family.