Optimize AI token usage: ~61% reduction in per-step context
- System prompt: 19.5KB → 6.3KB (67% smaller), removed redundant
rules, condensed spacing/typography/layout docs
- Tool tiers: CORE_TOOLS (34) vs EXTENDED_TOOLS (64), AI chat uses
core only (~2.8K tokens vs ~6.5K). ALL_TOOLS for MCP/CLI unchanged.
- <Icon> element in render JSX: <Icon name="lucide:heart" size={20}
color="#FFF" /> fetches and renders inline. Eliminates the
search→fetch→insert×N→delete×N chain (~8 tool calls per design).
- search_icons default limit 20 → 5 (~300 fewer tokens per query)
- get_page_tree returns compact tree (id, type, name, w, h) instead
of full recursive nodeToResult dump
- eval returns { ok: true } instead of null, saves a follow-up
describe call
- renderTree/renderJSX/renderTreeNode now async to support Icon fetch
1454afc036