Two new tools: foco (kanban+pomodoro) and paper-scan (new branches)

Date: 2026-09-01 Project: astro-starter-template + new repo [usuario]/foco

1. foco — kanban + pomodoro local-first (repo aparte)

gh: [usuario]/foco (public). Zero deps, localStorage, run npx serve -l 4000.
3 boards = the three clocks (carrera/práctica/investigación), lists Backlog/Hoy/En curso/Bloqueado/Hecho, DnD cards.
Slot modes → pomodoro durations (trans 90, lateral 75, reactivo 30, exploración 45, pomodoro 25).
Completing a pomodoro forces the three-line close (Produced/Blocked by/ Next) → per-card evidence; “exportar diario” → Markdown ready to paste into study/discovery-log.md.

npm run paper:scan -- <arxiv-id|query> in the astro repo.
arXiv fetch → local LLM concept extraction (Ollama, gemma4 — qwen3.6 is 24GB and does NOT fit 17GB RAM; use gemma4 ~9.6GB) → diff vs graph (fuzzy title match, parens stripped) → NEW concepts as stubs in _stubs/papers/ + branch proposal (section from tags, or flag “posible rama nueva”) + report in reports/paper-scan-.md.
Extraction cached in data/paper-cache/; fallback rule-based if Ollama off.
Adopting stubs: review → move to content/nodes/ → npm run build (integrity).

Big models on 16-17GB machines thrash and the API call hangs (no timeout in the fetch — generator can stall for many minutes). Keep to ≤10GB models.
LLM output is Spanish despite English prompts sometimes — acceptable; matching is language-agnostic (title match).
LLM “kind” values must be validated against the 8 allowed kinds.

6 concepts extracted → 1 covered (large-language-models) + 5 new stubs (virtual-context-management, memgpt, etc.), branch NLP & Transformers.