Model routing & tokens

Auto model routing

Model	Role	Relative cost/token
Opus	Plans, architecture, final verification — never writes production code	15×
Sonnet	Builds, refactors, tests; design-scout & learner	3×
Haiku	Summaries, digests, mechanical edits	1×

You never switch models manually — each agent is pinned to its tier, and the loop engine decides which phase runs where. Deep thought is rented by the minute; Atelier only rents it for decisions.

The knowledge graph

workspace/atelier.db holds rules, plans, criteria, registry, and the framework map. Instead of re-reading giant markdown files every turn, agents ask:

store.py map animation   →  skills/living-graphics/SKILL.md
store.py map bug         →  skills/principal-mind/SKILL.md
store.py rules design    →  the active design rules, pipe-row lean
store.py registry component → shadcn/ui, Aceternity, ReactBits…

Rows come back TOON-style (pipe-delimited), a fraction of JSON’s token cost. Skill descriptions stay always-on (~1.7k tokens total); full bodies load only when the map says they’re relevant.

Why this matters

The combination — right model per phase, right law per topic, lean rows over fat files — is what lets long builds run without context bloat or $50-of-Opus-per-refactor bills.