Eval Model Compare

Models tested by default: claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-6 (cheapest to most capable)

Eval command: tessl eval run <path/to/tile> --agent=...

Phase 1: Identify the tile

1.1 Find the tile

Look for a tile.json in the current directory or a parent/sibling directory. Exclude .tessl/ cache directories:

find . -name "tile.json" -not -path "*/node_modules/*" -not -path "*/.tessl/*" 2>/dev/null | head -10

If the user provides a path inside a .tessl/tiles/ directory (an installed tile cache), stop and warn them: that path is Tessl's local install cache — running evals from there won't work and changes would be overwritten on the next tessl install. Offer two options: point to the original tile source, or copy the tile out of .tessl/tiles/ to a new location (cp -r .tessl/tiles/<workspace>/<tile> ./<tile>).