Sovereign, in-process autonomous software engineering runtime. Transforms open-weights SLMs (3B to 7B) into verified compilers. Executes completely offline without cloud API tokens, external daemons, or telemetry.
Upstream inference is linked directly into engine memory. Eliminates HTTP/TCP socket latency, unlocking sub-millisecond causal KV-cache rollbacks and token steering.
Code generation is staged in RAM VFS and verified inside Bubblewrap (Linux) or Apple Seatbelt (macOS). Unverified code never touches host storage.
Mathematical guarantee that verified functions and classes are permanently frozen in memory. Erasing working code during refactoring is physically impossible.
Multi-file projects are scheduled topologically according to strict type and symbol dependencies, preventing out-of-order synthesis hallucination loops.
| Runtime & Model | Environment | Makespan | Exit Code | Concurrency Logic | Outcome |
|---|---|---|---|---|---|
| Omni Engine (Local 7B) | Google Colab (T4 GPU) | 37m 58s | Exit 0 | RwLock + Mutex | 100% Passed (Clean) |
| Omni Engine (Local 4B) | Commodity 8-Core CPU | 1h 39m | Exit 0 | Canonical Min/Max Lock | 100% Staged via Gate |
| Aider Local (4B Ollama) | Commodity 8-Core CPU | >5h 24m (Timeout) | kill -9 | Never Reached | 0 files delivered |
| Aider Cloud (Gemini 2.5 Flash) | Cloud Infrastructure | 58 seconds | Exit 101 | sort_unstable | FAILED (Panic test) |