Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

1.3 KB
Raw Blame

⌘B's router, measured on a real desk

How ⌘B's routing was hill-climbed (docs/design/2026-10-09-auto-router.md), and how to do it again. Nothing here sends anything: the router only decides. Each decision costs about $0.0002 of the OpenRouter key the router uses.

  1. Snapshot the desk, from cli/: node scripts/route-eval/collect.mjs ~/route-eval/desk.json. It holds your sessions' names, last prompts and summaries: keep it out of the repository.
  2. Write cases against it, in English, as a person really types: [{"text": "...", "want": [index, ...] or ["new"], "last": index (optional), "why": "..."}]. want is the desk index (or indexes, when two are equally right) the message belongs to, or "new". last is where the previous ⌘B message went, for a follow-up. Have someone who did not write the router write some of them blind, keep a set held out until the end, and have a red team write messages to break it: sets written while reading the sessions are too easy.
  3. Run them through the router as the desktop calls it: npx tsx scripts/route-eval/eval.mts ~/route-eval/desk.json ~/route-eval/cases.json 3. A WRONG send (a message typed into the wrong conversation) is the failure that matters; a miss (new work where a session should have taken it) costs a harness.