The LLM tier

Optional: a local model authors each section from what the prose says.

./scripts/fetch-bonsai.sh                              # ~5.2 GB model, once
./scripts/fetch-assets.sh                              # the CC0 3D asset pack (optional)
cargo run --features llm-metal --                      # authored live, in-app
cargo run --features llm-metal -- notes.md --generate  # headless: author all, exit

With the feature and a model present, a background worker authors every section that has no cached output, and each region upgrades in place as its result lands. Results live in notes.md → notes.world.json — a lockfile keyed by each section's content hash: unchanged sections are never regenerated, and a .md plus its sidecar renders identically on any machine, even in the default (no-llm) build, which still applies cached output.

Per section, the tier chain is:

1. Agent build. The model composes the place through tool calls — primitives, lights, and real assets from the shared CC0 Poly Haven pack — emitting localgpt-world-types entities. This is what runs for most sections.
2. Recipe — one styled JSON (palette, landmark kind, props) — if the agent session yields nothing.
3. Rule draft — always the fallback, so the app is never broken by a missing or misbehaving model.

A ```world fence in a section is an exact override that needs no LLM at all: a JSON array of world entities in platform-local coordinates (ids optional). Edit only the fence and the cache re-keys.

Every model-authored value is clamped on the Rust side (position, colours, light intensity, entity count), and any failure keeps the rule-based draft.

llm-metal is the macOS GPU path (the 5 GB Q4_K_M needs it to fit in memory); llm alone expects a smaller GGUF. The model lives in ~/.local/share/localgpt/models/llm/ (set $LOCALGPT_LLM_DIR to move it), the folder Verse and Gen read too, so one download serves all three. Any standard GGUF plus a matching tokenizer.json dropped there is picked up. $LOCALGPT_MD_LLM points MD alone somewhere else, and the older assets/llm/ folders are still searched.

The asset pack

place_asset and scatter_field place real models from a 171-model CC0 Poly Haven pack (the same one Verse uses), grouped by semantic kind (rock, tree, lamp, …) so the model names a kind and the app picks and varies the concrete model. The pack resolves from $LOCALGPT_MD_ASSETS → assets/ → the sibling localgpt-verse-assets checkout, and scripts/fetch-assets.sh copies it locally (about 520 MB). Without a pack the agent tier still works, with primitives only. Placed GLBs render in the app, and --export *.html copies every referenced model beside the page so the web viewer can load them.

What gets cached

The sidecar holds two kinds of output per section-hash: builds (the agent's entities, plus the model's closing description and which model authored them) and recipes (the styled JSON). It's pruned to live sections on save and written atomically, so a half-written file never replaces a good one. Version 2 files (builds) supersede version 1 (recipes only); the loader warns and ignores older versions — regenerate with --generate.