AION

Technical position paper + reference implementation scaffolding for “AION: An AI-Native Operating System for Context-Driven Multimodal Interaction and Semantic Memory”.

Repository layout

• paper/: manuscript (main.tex), bibliography (references.bib), figures/assets.
• docs/: v0.1 specification + PoC roadmap.
• crates/: Rust workspace crates (event log, policy, memory, core).
• apps/aion-tui/: TUI/CLI runner (--backend mock|llamacpp, --ui cli|genui).
• configs/: pinned model lockfiles (artifact paths/hashes + optional HF source revisions).
• tools/: offline toolchain helpers (HF downloads, artifact prep).

Build (paper)

cd paper
latexmk -pdf main.tex

Clean auxiliary files:

cd paper
latexmk -c

Build (Rust workspace)

Unit tests (no model artifacts required):

cargo test --workspace

Run the TUI app (logs to .aion/demo/events.jsonl by default):

cargo run -p aion-tui

Run the Phase 2 daemon (aiond) + active-mode CLI:

# terminal 1
cargo run -p aiond -F aion-llm-llamacpp/llamacpp -- --backend llamacpp --model <path.gguf>

# terminal 2
cargo run -p aion-tui -- --mode active --sock .aion-data/daemon/aiond.sock

Run with the in-process llama.cpp backend (requires third_party/llama.cpp built; see docs/qwen3-macos.md):

cargo run -p aion-tui -F aion-llm-llamacpp/llamacpp -- \
  --backend llamacpp --model <path.gguf> --ui cli

Specs / Roadmap

• docs/spec-v0.1.md
• docs/poc-next-steps.md
• docs/qwen3-macos.md (Qwen3 bring‑up + performance flags for macOS + Metal)
• docs/phase2-active.md (Active mode + daemon bring-up)
• tools/qwen3_macos_toolchain.sh (offline helper to download/convert/quantize Qwen3)

Structured LLM Outputs (Agent)

crates/aion-agent expects tool decisions as a single JSON object terminated with <END_ACTION> and validates semantics (read-only shell allowlist, safe relative paths, MCP tool+arg checks) before executing.

License

No license file is currently provided.