Claude Agent System - Parallel agent swarms for planning, implementation, and code review
npx claudepluginhub kasempiternal/claude-agent-systemClaude Agent System - Parallel agent swarms for planning, implementation, research, and code review
Turn Claude into your personal development team. Plugin skills that handle everything — from deep planning through implementation to code review, with parallel agent swarms and automatic quality gates.
v7.20.0 — State directory moved out of
.claude/All CAS state (plans, mailboxes, wave files) moved from
.claude/plans/to.cas/plans/. Claude Code treats.claude/as a sensitive config directory, causing every plan/mailbox write to trigger a permission prompt — even indontAskmode. This blocked all swarm skills (Hydra, Legion, Siege, Spectre) in practice. 112 path references updated across 27 files. Also includes v7.19.0: all swarm teammates can now invoke any installed plugin skill via the Skill tool (24 templates updated).Report issues at GitHub Issues.
The Claude Agent System is distributed exclusively as a Claude Code plugin. If you previously installed via the legacy setup script, uninstall the old files first:
rm -rf .claude/commands .claude/workflows .claude/middleware .claude/agents /plugin marketplace add Kasempiternal/Claude-Agent-System /plugin install cas
/plugin marketplace add Kasempiternal/Claude-Agent-System
/plugin install cas
Done! You now have 14 skills: /zk, /spectre, /gonk-test, /siege, /legion, /pcc, /pcc-opus, /hydra, /review, /cyberconan, /systemcc, /l30, /setup-swarm, and /setup-hooks. The Gonk MCP server (headless browser E2E testing) is bundled and auto-registers on install.
"Vibe coding" = typing a prompt, accepting whatever the AI outputs, hoping it works.
This system is the opposite. Every skill enforces structure:
| What Vibe Coding Does | What This System Does |
|---|---|
| Blindly accepts AI output | Parallel code review agents |
| No validation | Build config detection + linting enforcement |
| No learning | Session memory - learns your patterns and mistakes |
| No quality gates | Decision engines with complexity/risk/scope analysis |
| Hope it works | Post-execution validation + auto-fix critical issues |
/gonk-test - E2E Frontend Testing with Gonk MCP NEWHeadless browser E2E testing, powered by direct Chrome DevTools Protocol. Describe what you want to test in plain English — Gonk launches a headless browser, navigates, interacts, asserts, and reports results. No visible browser needed.
/gonk-test https://localhost:3000
/gonk-test "test the login flow — fill email and password, click submit, verify dashboard loads"
/gonk-test "check that the checkout page has no JS errors and all API calls return 200"
/gonk-test "verify the landing page renders correctly on iPhone 15"
The Chrome extension MCP is slow and requires a visible browser. Maestro proved that headless, accessibility-tree-based testing is dramatically faster — but only works for mobile. Gonk brings this to the web.
| Chrome Extension MCP | Playwright MCP | Gonk | |
|---|---|---|---|
| Speed | Slow (extension relay) | Medium (Playwright layer) | ~5x faster (direct CDP) |
| Browser | Must be visible | Headless available | Headless, zero UI |
| Output | Screenshots | Accessibility snapshots | Hybrid AX+DOM with smart diffing |
| Framework awareness | None | None | React/Vue/Svelte/Angular component trees |
| Network | Basic logging | Basic logging | Full interception + mocking |
| Assertions | None | None | AI-native with self-correction suggestions |
| Flows | None | None | Maestro-style YAML record/replay |
| Token efficiency | Full snapshots | Full snapshots | Diff mode: ~50 tokens vs ~2000 |