Project

Somnio AI Tools

Open-source CLI that installs and runs AI agent skills, health checks and security audits across AI coding CLIs. I'm the

  • #ai
  • #open-source
  • #cli
  • #developer-tools

Somnio AI Tools is an open-source (MIT) repository published by Somnio Software at github.com/somnio-software/somnio-ai-tools. It packages AI agent skills, project health checks and security audits for Flutter and NestJS projects and installs them into AI coding CLIs such as Claude Code, Cursor and Gemini. I'm a core contributor: the #2 contributor by commits, with 15 merged pull requests. This note describes only my public, merged work.

The problem

Audit plans and rules written once need to run in many different agents, each with its own file layout and conventions, and a long audit run should not burn a model's whole context window or hide what it cost.

What I built

  • The CLI itself. My first PR introduced the Dart CLI (somnio) that transforms the repository's source-of-truth plans and rules into agent-specific formats and installs them globally, with init, status, update, uninstall and an add wizard to register new technology bundles.
  • somnio run. Executes a health audit step by step from the target project, each rule in a fresh AI context, with artifacts saved per step and a final report. A deterministic pre-flight handles tool installation, version checks and test coverage without AI. Later PRs added interactive agent and model selection, validation of the --model flag against the chosen agent, and per-step token usage (input/output tokens, cost, duration) with an aggregated summary.
  • More agents, then a registry. I added Cursor CLI support and a guided somnio setup command that detects and installs missing AI CLIs. I then replaced the hard-coded three-agent design with a data-driven registry of 17 agents (12 CLI, 5 IDE-only), where adding an agent is a single config entry and token parsing is pluggable per agent.
  • Workflow engine. somnio workflow defines repeatable multi-step, multi-model workflows; steps without dependencies run concurrently via wave planning. The PR shipped with 55 tests.
  • Security audit plan. A framework-agnostic security audit (sensitive files, secret patterns, dependency vulnerabilities) centralized into one module, plus a cheapest-model fallback in the runner.
  • Agent Skills plugin architecture. I restructured the repository to the Agent Skills standard so it installs as a Claude Code plugin and via skills.sh, converting 50 YAML rule files to Markdown and adding parallel subagents for health and security audits.

Related