benchy

The home of benchy.run — a personal benchmark for coding agents. Capture a real coding session once, replay it against every agent you use, and get one blind-judged composite score back.

This repo holds both halves of the product:

• cli/ — the benchy Go CLI: prospective session capture, the portability filter, benchy run/results, and the blind judge. See cli/README.md.
• root (src/, public/, wrangler.jsonc, schema.sql) — the benchy.run Cloudflare Worker: landing page, /install, the leaderboard API (D1), and CLI binary distribution from R2 (/dl/, /api/version).

Develop

# CLI
cd cli && go build ./... && go test ./...

# Site (benchy.run)
npx wrangler@latest dev        # local
npx wrangler@latest deploy     # prod

Release the CLI

Distribution is via benchy.run/R2 (not GitHub releases):

1. Tag the version (git tag v0.6.0 && git push origin v0.6.0).
2. Build benchy-<os>-<arch> for darwin/linux × amd64/arm64 with -ldflags "-s -w -X github.com/option-ai/benchy/cli/cmd.version=<tag>".
3. wrangler r2 object put benchy-dl/<binary> for each, plus a latest.json ({"version":"<tag>"}).
4. wrangler deploy the Worker.

benchy update and benchy.run/install both resolve benchy.run.