loobster

Loobster

██╗                 ██████╗  ██████╗  ███████╗ ████████╗ ███████╗ ██████╗
██║                ██╔═══██╗ ██╔══██╗ ██╔════╝ ╚══██╔══╝ ██╔════╝ ██╔══██╗
██║       ██████╗  ██║   ██║ ██████╔╝ ███████╗    ██║    █████╗   ██████╔╝
██║      ██╔═══██╗ ██║   ██║ ██╔══██╗ ╚════██║    ██║    ██╔══╝   ██╔══██╗
███████╗ ╚██████╔╝ ╚██████╔╝ ██████╔╝ ███████║    ██║    ███████╗ ██║  ██║
╚══════╝  ╚═════╝   ╚═════╝  ╚═════╝  ╚══════╝    ╚═╝    ╚══════╝ ╚═╝  ╚═╝
          AI that plans, builds, verifies & secures every change — solo or team

plan → build → verify → secure · goal-loops · shared signals · configurable compliance · independent verification · local-first

🦞 Loobster — loop + lobster. Run bin/loobster.sh for the animated red version of the mascot.

📖 Docs: nilswidal.github.io/loobster — a team-focused docs site (source in docs/, served via GitHub Pages; see docs/DEPLOY.md).

Loobster is a cross-tool loop harness — it runs in Claude Code and Codex (and any agent that reads AGENTS.md / .agents/skills) — that turns AI-assisted development into a repeatable, reviewable, secure loop. It right-sizes each task, plans before it builds, can run autonomously between approval gates, and proves its work with an independent verifier instead of trusting itself. It coordinates whole teams through a shared signals hub, and runs the compliance frameworks you choose against every diff.

Under the hood it follows the RePPITS method — Research, Propose, Plan, Implement, Test, Secure — created by Mihail Eric (the RePPIT framework, from the creator of Stanford's first AI software engineering course). Healthcare (HIPAA/HITRUST) is one aspect, alongside ISO 27001 and SOC 2 — enable only what your repo needs.

Why loop engineering

AI-assisted development is shifting from one-shot prompts to durable, autonomous loops — the loop, not the prompt, is becoming the unit of work. An agent that plans, builds, tests, and retries against a goal does far more than one that answers once. The catch: an unsupervised loop is only as trustworthy as its guardrails. Left alone, a loop will happily rubber-stamp its own output, drift from scope, or burn the context window.

Loobster is the harness that makes a loop safe to let run:

• Risk-tiered gates — friction scales with blast radius; sensitive changes never auto-advance.
• Bounded convergence — the Implement→Test→Secure fix loop caps at 3 iterations, then escalates to a human (no infinite loops, no silent commits past failures).
• Never self-verify — every check runs in a separate verifier agent, so a loop can't grade its own work.
• Signals — independent loops (and people) coordinate through a shared, mergeable channel.
• Compliance + token discipline — gates run against the diff; context stays lean enough that long loops stay affordable.

It stands on two foundations: RePPIT (Mihail Eric) gives the phase structure a loop iterates over, and headroom (Tejas Chopra) gives the token economics that make long loops viable. Loobster wires both into the agent's control loop.

At a glance