invokerai-skills

InvokerAI agent routing brain delivered as Claude Code Skills.

What It Does

Three sub-skills orchestrate multi-agent execution:

• invokerai:setup — Scan your installed agents and build a domain→agent mapping (~/.invoker/agent-map.json). Run once or when agents change.
• invokerai:decompose — Break a task into a directed acyclic graph (DAG) of ordered steps with dependencies and parallelism flags.
• invokerai:spawn — Execute the DAG: select agents from the mapping, spawn them in dependency order, track via BEADS (if available), confirm completion.

Install

npx invokerai-skills

This copies the skill definitions to ~/.claude/skills/invokerai*. Requires ~/.claude to exist.

Options

• --user — Symlink full skill directories into ~/.claude/skills and ~/.cursor/skills
• --project — Symlink full skill directories into ./.claude/skills and ./.cursor/skills (current working directory)
• --claude-only — With --user/--project, install to Claude Code only
• --cursor-only — With --user/--project, install to Cursor only
• -h, --help — Print help

Examples

# Install to ~/.claude/skills (default)
npx invokerai-skills

# Symlink into ~/.claude/skills and ~/.cursor/skills
npx invokerai-skills --user

# Symlink into project-local ./.claude/skills and ./.cursor/skills
npx invokerai-skills --project

# Symlink to ~/.claude only
npx invokerai-skills --user --claude-only

Quick Start

1. Install

npx invokerai-skills

2. Set up agent mapping

/invokerai:setup

Run once to scan your installed agents and build ~/.invoker/agent-map.json.

3. Decompose a task

/invokerai:decompose
Task: <your task description>
Domains: <space-separated list: backend, frontend, testing, etc.>

Get back a DAG with ordered steps and dependencies.

4. Execute the DAG

/invokerai:spawn

Spawn agents in order, track via BEADS (optional), confirm completion.

Domain Decision Guide

Pass only domains where real work exists:

| Domain | Add when | Skip when | |--------|----------|-----------| | architecture | New subsystem design, cross-cutting redesign | Bug fix, additive feature, impl path obvious | | backend | Server routes, APIs, business logic, auth middleware | Pure UI work, DB-only schema changes | | frontend | UI components, styling, client-side state, rendering | Server-only work, no user-facing changes | | database | Schema changes, migrations, query optimization | No DB reads/writes in the task | | devops | CI/CD pipelines, Dockerfiles, infra config, deploy | App code changes only | | security | Auth flows, permissions, secrets, input validation, CVE fixes | Feature work, no trust boundary changes | | ml | Model training, inference, embeddings, vector search | Standard CRUD, no ML components | | testing | Writing/fixing tests, test infra, coverage gaps | Impl work where tests are a side effect | | documentation | API docs, READMEs, changelogs, docstrings, user guides | Code-only changes, no public surface | | mobile | iOS/Android native code, React Native, Flutter | Web-only work | | data | ETL pipelines, data transforms, analytics, reporting | App features with no data pipeline | | code-review | Reviewing a diff/PR, auditing quality/security | Active implementation (review ≠ build) |

Domain precision is critical: wrong domains → phantom steps → wasted agents.

• Pass only domains where real work exists.
• When unsure, under-specify — gaps will surface in decompose.
• Never add a domain speculatively.

Files

• install.js — Installation script (Node.js, no dependencies)
• skill/ — Skill definitions (4 skills: invoicerai, invoicerai-setup, invoicerai-decompose, invoicerai-spawn)
• package.json — NPM package metadata
• LICENSE — MIT

License

MIT — see LICENSE