Build neva-ai as a separate repo and implement the control plane in Nevalang.
- Parse CLI args into typed
Command. - Build normalized user message from command + optional prompt.
- Load tool definitions.
- Call LLM with
tools + prompt. - Execute selected tool by calling a typed runtime bridge (
pkg/cliadapter later). - Print result.
Nevalang runtime APIs for full LLM/tool-calling are not complete yet.
To keep momentum, MVP uses explicit #extern contracts and implements orchestration now.
- robust CLI parsing (flags, quoting, escaping)
- HTTP client with headers/auth + JSON encode/decode helpers
- tool-call schema marshalling
- interactive chat loop with state/history
- runtime bridge for
pkg/clifunction calls
These are intentionally modeled as explicit extern boundaries in flows/main.neva.