Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog, and versions follow Semantic Versioning.

Unreleased

1.0.0 - 2026-11-11

Added

• A standalone, MIT-licensed repository for the swarm skill. Install it with npx skills (skills.sh), the Claude Code plugin marketplace, the Codex plugin marketplace, or pi install.
• Release assets: swarm-skill.zip, which passes the strict Agent Skills spec (claude.ai, Cowork, Claude Skills API, or unzipped into any agent's skills folder), and swarm-plugin.zip, which carries the Agent Plugins, Claude Code and Codex manifests (OpenAI plugin portal, Codex desktop, claude --plugin-url). Both zips are deterministic, and the release lists them in SHA256SUMS with build provenance.
• Codex: agents/openai.yaml, with allow_implicit_invocation: false. The plugin also sets requires_local_executor.
• swarm skills searches the Agent Skills locations (.agents/skills, ~/.agents/skills, ~/.config/agents/skills), pi's and Codex's folders, and project folders from cwd up to the repo root. It follows the symlinks npx skills creates, lists each skill once, and names plugin skills from Agent Plugins and Codex manifests too.
• swarm audit flags access to ~/.codex/.
• Unit and integration tests for the workbench, with a stub pi for spawn. A validator for the spec, every manifest and OpenAI's listing limits. One generator for all manifests.

Changed

• disable-model-invocation is now a top-level SKILL.md field, so Claude Code and pi honour it. As metadata.disable-model-invocation it had no effect.
• allowed-tools is space-separated, as the spec asks. schema.sql moved to assets/.
• SKILL.md no longer assumes one company's stack. Worktree setup, network hosts and lead-only work (specs, the live app, the final review) are now written for any repo. The concrete pnpm + submodules + in-memory MongoDB setup moved to references/worktree-setup.md as a worked example.