Hurricane Sentinel

Plain English

What Sentinel does

Sentinel is a local harness that runs AI agents under real enforcement and human approval. It runs each agent as a Leash-leashed user with model calls routed through Muzzle, keeps durable runs that pause for your approval before anything risky, and lets you build a roster of persistent specialist agents — each with its own memory — that a main agent can delegate to.

• Runs agents as a dedicated OS user (Leash) with model traffic inspected by Muzzle — enforcement is real, not advisory.
• Durable, resumable runs: a flagged action pauses, waits for your approval, then resumes exactly where it left off.
• Persistent agents — each with its own soul, skills, model, and memory bank that persists across runs — that a main agent delegates to, or spawns a temporary specialist when none fits.
• An interactive chat shell (sentinel chat): live step-by-step progress, inline yes/no approvals, ↑/↓ history, and menus for / commands, @ specialists, and $ skills.
• Saved workflows: named multi-step pipelines where each step runs an agent and outputs thread into the next — run from the CLI, the chat, or by an agent.
• Built for small local models: tool-use scaffolding, forgiving tool-call parsing, self-correcting validation, and automatic context compaction.