评测与安全01
ronish645/minilang-compiler · Can LLMs learn a programming language they've never seen? A from-scratch compiler for MiniLang, used as an uncontaminated benchmark where Claude, GPT and Qwen learn it from the spec and debug through a compiler-as-tool agent loop. MCP server included.
MiniLang 是为评测「LLM 能否只凭规范学会一门从未见过的编程语言」而设计的小型 C/JavaScript 风格语言,配套从零用 Python 编写的编译器(词法分析、递归下降解析、渐进类型检查、三地址码生成与带资源限制的沙箱解释器)、支持 Claude/GPT/本地 Qwen 的编译器即工具智能体评测基准,以及把编译器暴露给 AI 助手的 MCP 服务器。
AI 编程02
oooscoos/Benzi · Benzi is a compiler-backed coding agent that compiles your codebase into a resolved, queryable map of calls, data flow, and control flow first; then it explains, runs, edits and verifies your code using it. Includes a runtime tracer that settles what static analysis can't. The Benzi compiler also works as an MCP server.
Benzi 是一个由编译器驱动的编程智能体(coding agent):它先用基于 tree-sitter 的编译器解析整个代码库,构建出可查询的调用关系、数据流与控制流映射,再基于该索引来讲解、运行、修改并验证代码。附带运行期追踪器,用于补足静态分析无法确定的部分。该编译器同时可作为 MCP 服务器使用,支持多语言与多模型接入,官方称在 SWE-bench Verified 上取得 391/500(78.2%)的结果。