# Brain — Short Plan

## Repo
- Gitea: `git@git.fnet.cz:lachtan/brain.git`
- Private, branch `main`
- Local path: `/home/nanobot/brain/`
- Inside: standard `wiki/` structure from plugin

## Git
- SSH key: `/home/nanobot/.ssh/id_ed25519` (ed25519)
- Auto commit + push after every nanobot write
- Git history = safety net for bad ingests

## Wiki structure
- `wiki/index.md` — main index (shards at >300 lines)
- `wiki/pages/` — compiled pages (LLM output)
- `wiki/sources/` — raw inputs
- Every page has YAML frontmatter: title, type, tags, created, updated, sources

## Nanobot commands
- `/brain-write` — write raw source or page with auto frontmatter, auto commit+push
- `/brain-search` — fulltext search across markdown files
- `/brain-ingest` — full LLM ingest (raw source → compiled page), auto commit+push
- Keyword "brain/wiki" = write to brain; no keyword = MEMORY.md

## Models
- Ingest: `qwen-3.6-plus-openrouter` (fallback `kimi-k2.6-openrouter`)
- All models cloud via OpenRouter/Ollama gateway
- No local inference for ingest

## Scripts
- From `praneybehl/llm-wiki-plugin`: init, search, lint, stats, graph
- Prepared in Claude Code, integrated into nanobot skill wrapper
- Pure Python 3.10+, stdlib + optional PyYAML

## Scaling
- Page soft cap: 400 lines, hard cap: 800 lines
- Index shards at >300 lines or >150 pages
- BM25 fallback at >300 pages

## Machines
- home pc, office pc — Claude Code GUI
- wood.hell — Claude Code CLI
- this LXC — nanobot host

## Implementation phases
1. SSH key → Gitea, clone repo, init wiki, test git push
2. Nanobot skill: write, search
3. Test ingest with qwen-3.6-plus, implement /brain-ingest
4. Add lint/stats/graph scripts, sync across machines