Every tool call meets the gate.
The firewall parses each call into its effects and tiers it: routine work is allowed, riskier calls ask you, critical ones are denied, and every decision goes to the audit record. See the autonomy gate.
The pi agent system · public beta
An opinionated distribution of the pi coding agent. A permission gate for tool calls, governed delegation to sub-agents, verification, durable memory and bounded recovery, chosen together as profiles. One pi package, installed once.
Guard rails, not a sandbox.
01 · Install
Needs Node.js 22.19 or newer, git and pi 0.85.1 or newer.
$ npm install -g @earendil-works/pi-coding-agent
$ pi install git:github.com/SATUNIX/luna@v0.3.0-beta.3
$ pi # first start: the balanced profile is applied; /reload when asked # no models listed? sign in with /login
The repository is public, so no credentials are involved. Other channels, the scripted installer, uninstalling and troubleshooting are in Installation and Getting started.
| Channel | Command | For |
|---|---|---|
| Release | pi install git:github.com/SATUNIX/luna@v0.3.0-beta.3 | Normal use. The tag is the release. |
| Notes | GitHub Releases | Release notes and source archives for each tag. |
| Installer | node packages/core/install.ts --channel latest --profile balanced | The profile chosen up front, or a scripted install, from a clone. |
| Checkout | node packages/core/install.ts --profile balanced --yes | Developing Luna: pi loads the checkout in place. |
| Upgrade | /update | Inside pi: what pi, Luna and packages can update to. |
02 · How it works
Three rules hold in every profile that loads them.
The firewall parses each call into its effects and tiers it: routine work is allowed, riskier calls ask you, critical ones are denied, and every decision goes to the audit record. See the autonomy gate.
A lead agent hands sub-agents a bounded question and integrates what comes back. The delegation guard decides what each child loads, reserves its budget before it starts, and checks inside the child that its protections loaded. Repeated multi-step work runs as a workflow, where code rather than the model decides the next step. See agent orchestration.
/verify and the agent's verify_completion tool run the project's check command and an isolated, read-only reviewer against the definition of done, and record the verdict. A review that did not run never reads as a pass. See concepts.
03 · Profiles
A profile chooses which extensions, skills and prompts load, and the firewall mode. Switch with /profile.
| Profile | For |
|---|---|
| balanced | Everyday work with verification and context helpers (the default). |
| long-horizon | Multi-hour work and unattended runs in a container: goals, task graph, branches, delegation. |
| self-improving | Research use: adds memory services, dual review and skill mining (experimental). |
| engagement | Authorised security engagements: strict firewall policy, scope and rules of engagement. |
| lite | Quick tasks, small local models and short contexts. |
What each profile loads is generated in the capability matrix.
04 · Documentation
The documentation describes what the code does now, and names the test that pins each claim.
05 · Status
Expect rough edges. Read Beta status for what is and is not covered, report problems in the issue tracker, and report vulnerabilities privately as SECURITY.md describes.