@nidvue/teamwork

Nidvue Teamwork is a skill for your coding agent - not a standalone application. Install it like any other skill (one clone, no package manager), then ask your agent for the team:

> use the teamwork skill: Build a URL shortener with a CLI, tests and docs

That prompt becomes a team of AI agents that plans, builds, tests and independently reviews its own work until every acceptance criterion has executed evidence. The agent host supplies the LLM API, and every agent's model is configured independently (teamwork model set <agent>): a strong model for the Reviewer, a cheap one for the mechanical lanes. The engine (teamwork.py, stdlib only) ships inside the skill; installing it as a package is optional.

• Developed by AI agents, for AI agents. This repository is built by its own team: reviewed tasks in Archive, the learning loop in Evolution, and all but two commits (a merge and a dependabot bump) by a role mailbox.
• One prompt in, a whole company out. The 30-second quickstart is the whole install.
• It improves from its own failures. A failure becomes a lesson, a lesson becomes a skill, and a skill reaches the core only after two cuts; its own verdict is confounded, because the first adoption changed the metrics and the classifier together (Evolution 9.2).

What makes it different

Five more ship inside the skill - engine commands, no plugin, no daemon:

Built by itself

The team is the skill's first user: it decomposes, dispatches to sub-agents that fan out further, reviews every deliverable and merges it - this repository is that output. The learning loop is recursive too: a failure becomes a lesson, a lesson a skill, a skill reaches the core after two cuts (Evolution 9.2).

The sourced comparison is in Alternatives.

Install the skill

Clone the skill into the directory your agent host scans - no daemon, no hosted service, nothing to build:

# project-local: plain files in your repository, not a nested git repo
mkdir -p .agents/skills
git clone --depth 1 https://github.com/Nidvue/Teamwork.git .agents/skills/teamwork
rm -rf .agents/skills/teamwork/.git

Prefer a user-level install (~/.agents/skills/teamwork), a package manager (npm install @nidvue/teamwork) or the optional engine wheel? Installation compares all four channels, their commands and their requirements. Dropping the nested .git is deliberate: the skill must be committed as plain files, not a gitlink.

30-second quickstart

# 1. create the project and install the skill into it (one clone, nothing to
#    build; drop the nested .git so the skill is committed as plain files, not a
#    submodule)
$ mkdir my-project && cd my-project
$ mkdir -p .agents/skills && git clone --depth 1 https://github.com/Nidvue/Teamwork.git .agents/skills/teamwork && rm -rf .agents/skills/teamwork/.git

# 2. ask your agent for the team, from the project root
> use the teamwork skill: Build a URL shortener with a CLI, tests and docs

The CEO writes AGENTS.md, staffs the roster, freezes the interfaces, records the acceptance criteria and opens the lanes. The hidden, gitignored workspace appears at .pi/Teamwork/Teams/<Team>/; project output is committed. Read the full walkthrough in Quickstart.

What you get

• 16 role charters - the 15-role org plus a Team Leader per feature, each with its own memory and queue.
• A two-stage quality gate - author self-check, then independent review; only the Reviewer merges.
• Parallel lanes, no plugin - swarm run sizes each wave from the work.
• Built-in memory and a mailbox - role-shared memory, a ctx:// tree and a prioritised agent-to-agent queue.
• Live state - status and its JSON feed, plus the bundled status widget.
• Layered tests - one owner per layer, gaps shown by tests plan.
• Token-true budget - harness budget measures the startup read; trim it, never raise the cap.
• Automatic model routing - teamwork model auto puts the strong model on the decision lanes and the economy one on the mechanical lanes.
• Portable - CPython 3.11-3.13 on Linux, macOS and Windows; stdlib only.

Documentation

Contributing, security and license

• Contributing - work on your own branch, run pytest -q and ruff check, open a PR; a Reviewer - never the author - merges it. Read CONTRIBUTING.md and the code of conduct.
• Security - report privately via GitHub security advisories; see SECURITY.md.
• License - AGPL-3.0-or-later, copyright (C) 2026 Nidvue (NOTICE). The strong copyleft is deliberate: the audit guarantees should survive modification.
• No personal data, no local leakage - scripts/privacy_scan.py fails the build on names, emails, home paths, credentials or agent-host state.