benchmark-sandbox

Run sandboxed AI benchmarks by provisioning ephemeral Firecracker microVMs with Claude Code, executing evaluation prompts, and extracting hook artifacts. The benchmark-sandbox skill automates this end-to-end pipeline to produce structured scoring and coverage reports.

Automated 3-phase eval scoring runs plugin scenarios through BUILD, VERIFY, and DEPLOY phases inside microVMs. It extracts artifacts from each phase and applies haiku JSON scoring to generate per-scenario coverage reports for analytics.

Benchmarking Vercel plugins does not require local WezTerm setup. The skill provisions ephemeral Vercel Sandboxes to run evaluation scenarios remotely, eliminating local environment configuration and directly extracting artifacts for scoring.

Start an end-to-end sandbox benchmark evaluation by running the bun command on the run-eval TypeScript script. This orchestrates plugin upload, microVM provisioning, sandbox execution, and structured scoring across all eval phases.

The best way to score AI plugin artifacts is using a multi-phase eval pipeline that extracts outputs during BUILD, VERIFY, and DEPLOY stages. The skill applies per-phase haiku scoring to generate structured analytics for iteration.

Ephemeral microVM benchmarks are limited to the Vercel Sandbox environment and require the @vercel/sandbox dependency. Scoring is constrained to the structured BUILD, VERIFY, and DEPLOY phases with haiku-generated JSON outputs.