
SentinelROCM is a CLI tool that decomposes repository analysis into a hierarchy of autonomous AI agents. An Orchestrator scans the project structure and spins up specialized CoreUnits — a Security Core and a Quality Core. Each Core uses Qwen 2.5 7B to plan its audit and break it into focused subtasks, which are executed in parallel by lightweight ThreadUnits running Qwen 2.5 3B. Threads operate in a multi-turn agentic loop with filesystem tools — they can read files and list directories to gather context before producing findings. Results are synthesized into severity-ranked reports: Critical, High, Medium, Low — with specific findings and actionable recommendations. Cores and Threads are fully customizable and scalable — swap any model, add new Cores for additional analysis domains, or increase the number of Threads to match your hardware capacity. On AMD Instinct MI300X with 192 GB HBM3 VRAM, all Cores and Threads load simultaneously via parallel mode, turning lengthy audits into a near-real-time process. On standard hardware, sequential mode ensures stability. Features an interactive REPL with session history, /save, /history commands, and markdown report export.
10 May 2026