What problem does it solve? Running plugin benchmark evals locally in WezTerm panes limits parallelism, requires manual verification, and leaves no permanent deploy URLs. This Skill moves eval scenarios into remote Vercel Sandboxes (ephemeral Firecracker microVMs) so benchmarks run in parallel with automated browser verification, deployment, and structured scoring. ## Core Features & Use Cases - 3-Phase Eval Pipeline: Phase 1 builds the app with Claude Code, Phase 2 uses agent-browser to walk through user stories and fix failures, Phase 3 links to the vercel-labs team and runs vercel deploy with up to 3 build-error retries. - Skill Coverage Tracking: Extracts claimed/seen skill artifacts across all phases and produces coverage reports comparing expected vs. actual skill injections. - Haiku Structured Scoring: Each phase is scored via claude -p --json-schema --model haiku, returning structured JSON for build completeness, per-story pass/fail, and deploy success. - Use Case: Generate a JSON file of app scenarios (slug, prompt, expectedSkills, userStories), then run the eval runner to get per-scenario result.json files, public sandbox URLs, permanent *.vercel.app deploy URLs, and a markdown coverage report. ## Quick Start Run the benchmark-sandbox eval with a scenarios JSON file using bun run .claude/skills/benchmark-sandbox/run-eval.ts --scenarios-file /tmp/my-scenarios.json.
