All notable changes to InferNode are documented in this file.
First public release of InferNode, a 64-bit fork of Inferno OS for AI agents.
- Veltro AI Agent — built-in conversational agent with tool use, delegation, and namespace-isolated sub-agents
- Capability-based namespace isolation — each agent sees only explicitly granted resources; formally verified with TLA+, SPIN, and CBMC
- Post-quantum cryptography — preliminary ML-KEM (FIPS 203) and ML-DSA (FIPS 204) implementations
- ARM64 JIT compiler — native JIT for Apple Silicon and Linux ARM64 (e.g. NVIDIA Jetson)
- Xenith text environment — Acme-inspired editor with markdown, PDF, image, and Mermaid rendering
- Lucifer GUI — presentation zone with app hosting, tab management, and context/namespace browser
- Secstore persistence — encrypted key storage with PAK authentication
- Ollama/OpenAI-compatible backend — local LLM support alongside Anthropic API
- 9P everywhere — LLM, speech, tools, wallet, and UI all exposed as 9P filesystems
| Platform | Architecture | GUI |
|---|---|---|
| macOS | ARM64 (Apple Silicon) | SDL3 |
| Linux | AMD64 | Headless (SDL3 optional via build script) |
| Linux | ARM64 | Headless (SDL3 optional via build script) |
- Secstore key loading can intermittently fail on cold boot due to trfs cache timing; a warmup workaround is in place
- Models that do not support tool use (e.g. llama2) return empty responses; use a tool-capable model (llama3.2, qwen2.5, mistral, etc.)
- The presentation rendering architecture is tightly coupled to the lucipres window; a refactor to a separate wmclient app is planned (see docs/TODO-LUCIPRES-ARCHITECTURE.md)