@vincenzoimp/pitcrew

A crew for the pi coding agent: roles instead of model names, a per-session mode, child agents you can watch and steer, a guard, a verify gate, skills, and an unattended runner.

Packages

Package details

extensionskill

Install @vincenzoimp/pitcrew from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@vincenzoimp/pitcrew
Package
@vincenzoimp/pitcrew
Version
0.1.2
Published
Nov 13, 2026
Downloads
699/mo · 699/wk
Author
vincenzoimp
License
MIT
Types
extension, skill
Size
1,005.6 KB
Dependencies
6 dependencies · 1 peer
Pi manifest JSON
{
  "image": "https://raw.githubusercontent.com/VincenzoImp/pitcrew/main/docs/assets/screen-agents.png",
  "skills": [
    "./skills"
  ],
  "prompts": [],
  "extensions": [
    "./extensions/roles/index.ts",
    "./extensions/rules/index.ts",
    "./extensions/mode/index.ts",
    "./extensions/guard/index.ts",
    "./extensions/verify-gate/index.ts",
    "./extensions/status/index.ts",
    "./extensions/context/index.ts",
    "./extensions/handoff/index.ts",
    "./extensions/web/index.ts",
    "./extensions/agent/index.ts"
  ]
}

Security note

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

README

What you get

  • Roles, not model names. You talk to lead; worker, scout, reviewer and second do the rest. A profile says which model and thinking level answers for each role, per session.
  • Child agents you can see. The lead starts children in the background; /agents shows each one live, and you can steer it, answer it or stop it.
  • A mode per piece of work. stop says how far the agent goes alone (a plan, local commits, a PR, a merge); rigor says which checks run.
  • Guard rails. The guard blocks catastrophic commands and asks before risky ones; the verify gate means "done" only when the repo's own check passed on the current tree.
  • Unattended delivery. pitcrew run takes the issues of a milestone one by one: worktree, fresh worker, checks, PR, review, merge or stop, cleanup. Run it again to continue.
  • Web tools, 14 skills, and a footer with context use, cache, cost and subscription windows.

It is one pi package. It works with any provider pi supports: a subscription, an API key, or a local server.

Quick start

  1. Install pi and log in to a provider (/login in pi).
  2. Set up pitcrew in a terminal (it asks one or two questions):
    npx @vincenzoimp/pitcrew init
  3. In a repo of yours, run pi and ask for something small: "the test for X fails, find out why and fix it".

pitcrew init checks Node and pi, registers the package, adds the baseline settings without changing yours, proposes a profile that works with your logins, and lists the optional parts (GitHub CLI, Exa search key, Claude bridge, Docker). Model calls are billed by your provider (your plan, its credits, or API prices); the footer shows the cost and the plan windows. More: getting started.

Concepts

Concept In one line
Role a model slot: lead (you talk to it), worker, scout, reviewer, second
Profile which model and thinking level answer for each role; examples for claude-bridge, codex, anthropic, openai, google, local; with none set, the first in this order whose models you can use
Session mode stop (how far the agent goes alone) and rigor (which checks run), per session
Tracking a repo setting: where plans and work items live (none, local, github)
Agent, child an agent is a definition (worker, scout, reviewer, researcher); a child is one running instance
Skill method for one kind of work; the model picks some, you start the others with /skill:<name>
Guard blocks catastrophic commands, asks before risky ones, holds the delivery rules
Verify gate "done" means the repo's own check passed on the current tree
Runner pitcrew run: delivers the issues of a target unattended, resumable
Sandbox the container pitcrew run uses for repos that are not yours

Every word: glossary. Three diagrams: how it works.

Commands

In a session
/profile, /role show or change the role map of this session
/mode show or change stop, rigor and the switches
/agents (alt+a) the panel of child agents
/handoff [goal] go on in a fresh session with the same mode and roles
/guard, /stats, /skill:<name> guard dialogs on or off; usage and cost; run a skill
In the shell
pitcrew init set up, or check a setup
pitcrew run <milestone | issue | label:x> deliver GitHub issues unattended; pitcrew status for the digest
pitcrew tidy remove the worktree and branches of merged work, with proof
pitcrew check lean-code checks on a branch
pitcrew box the sandbox container

All of them, with flags and keys: commands.

What it costs

Measured once each with pi 1.5.2, so read them as orders of size. Details and models: live tests.

Thing Cost
Standing context: global rules and model-picked skill descriptions about 1,300 tokens in every request
First request of a session about 4,700 tokens natively; about 7,700 through the Claude bridge
First request of a scout child about 2,400 tokens natively; about 4,600 through the bridge
web_search Exa: about $0.012 per search (free monthly credit), no model tokens; Codex: 8,700 to 13,200 tokens of the plan
pitcrew run, one small issue, stop pr about 75,000 tokens, 47 s (local model)
code-review of a small risky change about 180,000 tokens sent by the lead, plus 77,000 by its children
Sandbox start 1.9 s

Limits

  • Sessions are not sandboxed. Commands run as you, on this machine. The guard catches mistakes with pattern rules; it is not a security boundary. Only pitcrew run uses the container.
  • No fallback model. If a role's model is missing or refuses for quota, the request fails with the reason; after the reset you resume (pi --continue, /agents resume, the same pitcrew run).
  • Small local models overflow on large reads and research fan-out, and often do not load a skill by themselves: name it.
  • Claude through pi-claude-bridge (optional) adds Claude Code's own prompt and drops some prompt channels; pitcrew uses only the channels that pass (claude-access).
  • pitcrew run works one issue at a time, on GitHub only.

When something goes wrong: troubleshooting.

Docs

Getting started · Install · How it works · Commands · All docs

Contributing

Issues and pull requests are welcome: read CONTRIBUTING first. Security problems: SECURITY. Changes: CHANGELOG.

Licence

MIT. Text and code from other projects keep their licences: THIRD_PARTY_NOTICES. Versions tested against: UPSTREAM.