pi-loops

A reusable Pi extension that adds a supervised loop harness around the Pi coding agent.

The agent still works as normal — it thinks, investigates, edits files, runs tools, and writes code. The loop extension owns the workflow around the agent: state, node order, tool scope, steering, evidence, validation, approvals, and completion.

This extension ships with no pre-built loops. It gives you the runtime and tooling to build your own workflows. Add your loops in src/loops/ and the agent will help you design them.

Features

• Generic loop runtime — sequence typed nodes (code, LLM, gate, approval, finalizer)
• Bounded LLM nodes — each node has scoped tools, timeout, model assignment, and steering rules
• Deterministic gates — validate artifacts and evidence without LLM cost
• Evidence collection — every tool call, command, and file change tracked automatically
• Steering — the loop detects stalls, over-research, repeated failures and nudges the agent
• Context cleaning — sanitize tool-call noise between nodes via code handlers
• Session persistence — loop state survives Pi session restarts
• Per‑node model routing — assign different models and thinking levels to different nodes
• loop_report tool — structured agent-to-loop handoff

Installation

# Install for all projects
pi install npm:@asm3r96/pi-loops

# Or install from git
pi install git:github.com/Asm3r96/pi-loops

# Or load directly
pi -e ./src/index.ts

After install, the extension auto-loads on Pi startup. The agent will also discover the bundled pi-loops skill from skills/pi-loops/, so it can guide users through designing and creating their own loops.

Quick start — build your first loop

After installing, ask the agent:

Help me build a loop that does X.

The agent will read the bundled skill and walk you through creating:

1. A loop definition in src/loops/my-loop/definition.ts
2. Optional code handlers in src/loops/my-loop/handlers.ts
3. Registration in src/loops/index.ts
4. Reload and run

The skill also includes a plain-text design reference at skills/pi-loops/references/github-issue.md. This is not runnable loop code; it is a walkthrough the agent can read to understand how a real multi-node loop can be structured.

Available commands

Command	Description
/loop start <definition> [input...]	Start a new loop
/loop status	Show current run state
/loop list	List available loop definitions
/loop cancel	Cancel active run
/loop pause	Pause active run
/loop resume	Resume paused/idle loop

Architecture

src/
├── index.ts              Extension entry: commands, events, tool registration
├── types.ts              Loop/workflow type definitions
├── runtime.ts            Generic loop runtime engine
├── nodes.ts              Code handler registry + gate/LLM execution
├── state.ts              Session persistence (save/restore via Pi session entries)
├── evidence.ts           Tool observation + steering triggers
├── tools.ts              loop_report custom tool
└── loops/
    └── index.ts          Add your loops here

Key node types

Type	Description
code	Deterministic logic (fetch, transform, validate, side-effect)
llm	Agent works inside bounded scope, reports via loop_report
gate	Validates artifacts and evidence deterministically
approval	Pauses for user decision
finalizer	Writes final result

The loop_report tool

LLM nodes signal completion by calling loop_report. The report includes a summary, artifacts, changed files, and confidence. It is a proposal — the next gate decides whether to pass, request a retry, or fail.

Adding a new loop

1. src/loops/my-loop/definition.ts — export a LoopDefinition
2. src/loops/my-loop/handlers.ts — write code handlers and register them
3. src/loops/index.ts — import and register

// loops/index.ts
import { createMyLoopDefinition } from "./my-loop/definition.ts";
import { registerMyLoopHandlers } from "./my-loop/handlers.ts";

export function registerLoopPackages(registerDefinition) {
  registerDefinition(createMyLoopDefinition());
  registerMyLoopHandlers();
}

Reload Pi and run:

/reload
/loop start my-loop input=value

License

MIT

Built with Pi.