pi-caw

pi-CAW: reusable Roles, Workflows and journaled Runs for Pi hosts

Packages

Package details

extensionskill

Install pi-caw from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:pi-caw
Package
pi-caw
Version
0.6.0
Published
Oct 15, 2026
Downloads
791/mo · 38/wk
Author
tohman233
License
MIT
Types
extension, skill
Size
2.9 MB
Dependencies
0 dependencies · 2 peers
Pi manifest JSON
{
  "skills": [
    "./skills"
  ],
  "extensions": [
    "./extensions/pi-caw.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

Reusable, traceable workflows for Pi

English | 简体中文

MIT Licensed

pi-CAW 0.6.0 turns repeatable multi-step work into reusable Workflows and journaled Runs for Pi. Its Workbench lets Pi users create, import, configure and execute workflows with declared inputs, resources and outputs, keeping reusable steps focused on the material they need.

Requirements

  • Node.js 22.19 or newer.
  • A Pi Host with compatible @earendil-works/pi-coding-agent (1.0.0 or newer) and @earendil-works/pi-ai interfaces.

Installation

Install the public GitHub package in Pi:

pi install git:github.com/TohmaN233/pi-CAW

From a local checkout, load the extension for one Pi launch with pi -e ./extensions/pi-caw.ts, or install the local package and its Skill into Pi settings with pi install ..

Configuration

In a Workflow's Models tab, choose a Pi model and supported thinking level for each child Provider slot it uses. Imported Workflows keep their own local bindings, which start empty even when the same slot ID is configured globally. Pi manages provider authentication and model transport.

Main has no separate model binding: it inherits the initiating chat's model and thinking level. Its default worker mode uses a fresh context with declared inputs and resources. orchestration continues the current chat and uses native Pi tools; it requires Cooperative execution.

Quick start

In Pi, open the Workbench:

/caw
  1. Choose system.build-workflow for a task brief or system.skill2workflow for a Skill source, then provide its required inputs.
  2. Bind any child Provider slots the selected authoring Workflow uses, then start it.
  3. Review the resulting Draft and publish it through the human approval step.
  4. Select the Ready Workflow, provide its workspace, declared inputs and read/write authorization, then start a Run.

What it does

  • Provides reusable Roles for focused helpers and a visual Workbench for Workflow graphs, model choices and Run inspection. Execution messages load when you expand their node details.
  • Ships two authoring Workflows: system.build-workflow turns a brief into a reviewed Workflow Draft; system.skill2workflow starts from a Skill snapshot. Host compilation and validation, independent review and human publication are part of authoring.
  • Imports codex.workflow.package and pi-caw.workflow.package while preserving the source graph, revision and original package envelope. Source integrity and local launch readiness are checked separately; unsupported dependencies, capabilities or missing local model bindings can prevent a Run.
  • Supports semantic repair through declared Workflow loops and fan-out correction. Fresh-review loops use a finite max_rounds; an explicitly qualified, read-only current-chat Main review can use max_rounds: null and stop on acceptance. Fan-out workers correct unresolved items while accepted siblings are preserved, within the pinned worker-attempt budget.
  • Gives authoring its own default of up to three automatic semantic repairs before it waits for user-guided continuation. This authoring budget is separate from Workflow execution repair.
  • Corrects invalid current-chat Main submissions in the same chat within the pinned turn budget, retaining staged results and completed effects. Mechanical failures remain visible for diagnosis.

Optional external collection: workflow-package. Each imported Workflow still needs to pass pi-CAW's compatibility, binding and local-readiness checks; this link does not imply that every package runs unchanged in Pi.

Documentation

Development

With the project dependencies installed, the repository defines these checks:

npm test
npm run check
npm run check:web
npm run audit:publication

Attribution and license

pi-CAW is an independent Pi adaptation of Codex Agents Workflow. See NOTICE for attribution and LICENSE for the MIT license.