Evaluate model-generated code execution with SandboxFusion
Use SandboxFusion to run and judge LLM-generated code in controlled sandboxes across many languages and benchmark-style evaluation tasks.
npx skills add agentskillexchange/skills --skill evaluate-model-generated-code-execution-with-sandboxfusion
SandboxFusion is ByteDance’s open-source code sandbox for running and judging code generated by LLMs. It supports code execution across Python, C++, Go, Java, Node.js, TypeScript, Rust, Bash, CUDA, and other runtimes, plus online-judge style evaluation datasets such as HumanEval, MBPP, MBXP, CRUXEval, and related coding benchmarks.\n\nUse this skill when an agent needs to evaluate generated code rather than merely produce it: submit code or benchmark tasks, execute outputs in the sandbox service, collect run results, compare them against expected behavior, and feed failures back into the evaluation or repair loop. The scope boundary is controlled code execution and judging for LLM/code-agent evaluation, not a generic sandbox product, deployment platform, or low-level library listing.
What this skill actually does
Inputs and prerequisites: Docker, or conda and Poetry for manual installation.
Setup notes: Build the Docker images from scripts/Dockerfile.base and scripts/Dockerfile.server, then run the sandbox service with make run-online; manual setup uses Python 3.12, conda, Poetry, and runtime-specific language installers.
Source and verification boundary: use https://bytedance.github.io/SandboxFusion/ as the canonical reference before running the workflow; keep commands, API calls, CLI usage, and generated outputs reviewable against that upstream source.
Framework fit: publish this as a Multi-Framework workflow only when the operator can invoke the documented toolchain directly, rather than treating the upstream project as a generic product listing.