perf: let agents finish Figma tasks in fewer calls

[silships]

What changes

Agents that drive figma-cli now finish a Figma task in fewer calls. Every call is a full agent turn so fewer calls means less time and fewer tokens for the user.

New in render and render-batch

• --page "Name" renders on that page and creates it when missing
• the output lists the structure that was built (sizes and layout and padding and bound variables and which component each instance uses) so the agent does not read it back
• <Instance component="Button" variant="size=large" text="Save" /> finds a component on any page and picks the variant and relabels it
• effectStyle on frames and textStyle on text
• pt pr pb pl now work on the outermost frame too (they were ignored before)
• render-batch --variant-set Button turns frames named prop=value, prop=value into one component set

New in eval

• helpers $page $var $bind $style $component $instance $fontSafe $describe
• output to a pipe is compact JSON and capped at 20,000 characters. A terminal still gets indented JSON

Missing fonts
Text in a component whose font is not installed could not be relabelled and Figma even refused to move the instance into a frame. Those fonts are now swapped to Inter automatically and reported as a note: line.

Agent rules
The rules init-agent writes live in src/lib/agent-rules.js inside a hashed marker block. connect refreshes an outdated copy in the project. Text the user wrote around the block stays untouched.

Fixes

• instantiate bounds every lookup to 5 seconds so a library key that was never published no longer hangs the CLI
• the roundtrip test builds on its own page and removes it so it no longer leaves fixtures in the connected file

Behaviour change

One visible behaviour change: piped eval output has no indentation anymore. JSON.parse reads it the same.

🤖 Generated with Claude Code