Show HN: VebGen – AI agent that understands projects with AST, not LLM tokens (github.com)

VebGen is an open-source, MIT-licensed autonomous development agent that claims to build full production-ready apps from natural language while never “reading” files into an LLM. Instead it parses projects locally using Abstract Syntax Trees (AST) to create a zero-token code context: 95+ Django constructs (models, views, serializers, URLs, Celery, Channels, etc.) are understood structurally, which yields instant context building, token-cost savings, and stronger local privacy compared with file-dumping approaches. The 800KB codebase (309 tests, 850KB docs) was developed using free-tier APIs, demonstrating a low-cost path to sophisticated agent behavior. The system emphasizes safety and reliability: a multi-layer sandbox with file-system isolation, whitelist-only commands, path-traversal and regex blocklists, and no direct shell access; a Memory Manager with SHA-256 verification, atomic writes, 5 backups and thread-safe saves; and defensive change-management (compile-on-write, snapshot rollback). Autonomy comes from a dual-agent workflow—TARS (planner) and CASE (developer)—with a 3-attempt remediation loop, fuzzy-patch fallback (raising fix success from ~70% to ~92%), circuit-breaker detection for looping, and automated escalation. For AI/ML practitioners this architecture highlights practical advances in cost-efficient local code understanding, safer autonomous code modification, and resilient state persistence—key steps toward trustworthy, production-grade developer agents.