sbx-sandbox (monorepo)

A bun workspace: the sbx-sandbox library, the agent-runner / catalog / bench layers built on it, and runnable examples.

Path	What
packages/sbx-sandbox	The sandbox library — a drop-in @vercel/sandbox-compatible SDK backed by the sbx CLI (Docker Sandboxes) or local Docker.
packages/agent-runner	Run a Claude Code agent task via ai + ai-sdk-provider-claude-code, optionally jailed in an sbx sandbox, with guaranteed cleanup.
packages/catalog	Declare user profiles, stacks, and app specs/prompts in YAML; compose into agent tasks.
packages/bench	Run the catalog matrix (subset + resume) via agent-runner, measure each app (build/tsc/eslint, browser smoke, optional LLM judge), and emit an HTML report — exposed as the bench CLI.
examples/	Runnable demos: ai-agent (lib + ai SDK) and agent-runner (jailed agent run).
ANALYSIS.md	The agent-runner refactor analysis + the provider limitations it surfaced.

Setup

bun install        # one install at the root; links the workspace
bun run build      # builds packages/* → dist (delegates via --filter)
bun run test       # the sandbox library's vitest suite (Docker + fake-sbx)
bun run typecheck

Requires Bun and either the sbx CLI or Docker. Per-package details live in each package's README.

Run the benchmark

bun run bench -- --help                 # all options
bun run bench -- --dry-run              # show the matrix the filters select
bun run bench -- --jail --stacks vike-manual --profiles dev   # one cell, jailed in sbx
bun run bench -- --jail --resume missing                             # continue an interrupted run
bun run bench -- --report-only          # regenerate the HTML report from saved records

Declare your own profiles / stacks / specs in YAML and pass --config <dir> (format in packages/catalog). Records land in results/<spec>/, the report at results/<spec>.html. Full reference: packages/bench.