OpenJet v0.4 adds multi-step agent

This is a solid answer to the local-agent setup problem: less model heroics, more making the whole stack usable on real machines.

• –Auto-detecting hardware and provisioning `llama.cpp` from scratch cuts out the annoying setup tax that keeps most people on hosted tools
• –Support for RTX, Mac, Jetson, and CPU-only systems makes it unusually broad for a local-first coding agent
• –Persistent per-project and system memory moves it closer to a real agent harness than a disposable chat UI
• –The new UI and smaller-GPU model improvements matter because most local LLM users are constrained by memory, not ambition
• –Permission gates before shell and write actions make the workflow safer without turning it into a toy