cockpit

My personal marketplace of Claude Code (and Codex) plugins — skills, hooks and agents I reuse across the projects I work on.

I based this repo on the:

• loopkit — MIT, © 2026 Archive228
• ai-tools — MIT, © 2025 Martin Bens

So, yes, it's not an original repo in its core as everything else nowadays. And it's still the skillset I use as my day-to-day performance boost.

cockpit is MIT too — see LICENSE, which carries both upstream copyright notices.

Install

Non-interactive (both default to user scope, so it loads in every project):

claude plugin marketplace add elitongadotti/cockpit   # or a full git URL / local path
claude plugin install cockpit@cockpit

Pass --scope project (shared via the repo's .claude/settings.json) or --scope local (just you, just this repo) to scope it to one project — run it from inside that project. Then /reload-plugins (or start a fresh session) to activate; verify with claude plugin list.

The equivalents inside an interactive claude session are the /plugin marketplace add and /plugin install slash commands, which prompt for scope.

Test locally without installing:

claude --plugin-dir ./plugins/cockpit

--plugin-dir is dev-only: it loads the plugin for that one session, in that one directory, and nothing persists. Use the install steps above for day-to-day use across projects.

Build a plugin

A plugin is a folder with a manifest. Minimum viable plugin:

plugins/my-plugin/
└── .claude-plugin/
    └── plugin.json

1. Manifest — .claude-plugin/plugin.json:

{
  "name": "my-plugin",
  "description": "What it does, in one line.",
  "version": "0.1.0",
  "author": { "name": "Your Name" },
  "license": "MIT"
}

2. Add a skill — skills/<name>/SKILL.md (auto-loads when the description matches the task; you don't invoke it):

---
name: my-skill
description: One line — this is the trigger the model matches on.
---
# My Skill
Short, imperative guidance the model follows.

3. Add an agent — agents/<name>.md (a subagent you can delegate to):

---
name: my-agent
description: When to use this agent.
model: haiku
tools: [Read, Grep, Bash]
---
System prompt for the subagent.

4. Add a hook — hooks/hooks.json (the harness runs these on lifecycle events; format is identical to settings.json):

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          { "type": "command", "command": "\"${CLAUDE_PLUGIN_ROOT}/hooks/fmt\"" }
        ]
      }
    ]
  }
}

Referenced scripts sit next to hooks.json and must be executable (chmod +x). Use ${CLAUDE_PLUGIN_ROOT} for paths — never hardcode.

5. Register it in .claude-plugin/marketplace.json:

{
  "plugins": [
    { "name": "my-plugin", "source": "./plugins/my-plugin", "version": "0.1.0" }
  ]
}

Skill vs. hook (rule of thumb)

• Skill = how to do task X when it comes up — Markdown; the model decides to invoke it. Probabilistic.
• Hook = always do Y on event Z — JSON + script; the harness runs it every time. Deterministic.

Codex compatibility

Each plugin also ships a .codex-plugin/plugin.json and is listed in .agents/plugins/marketplace.json. See AGENTS.md for the full compatibility contract.

What's inside cockpit

Kind	Name	Purpose
skill	spec-first	Write PROMPT.md before a multi-step task
skill	read-the-trace	Root-cause from a stack trace
skill	sql-review	Review SQL/ORM for safety + perf
skill	pycegerb-yoda	PySpark + Iceberg: writes, schema, maintenance
skill	acceptance-audit	Audit a ticket's acceptance criteria against the branch
skill	leak-check	Catch secrets, PII and internal context before commit
skill	writing-commit-messages	Commit message about the focus, not the scaffolding
skill	pr-from-diff	Write an approvable PR description
skill	context-budget	Keep context lean on long sessions
skill	shift-notes	Between-session handoff file
skill	broken-window-check	Smoke-test last "done" feature
skill	using-cockpit	Routing rules (injected at session start)
agent	verifier	Reviews a diff assuming it's broken
hook	session-start	Injects using-cockpit as context
hook	pre-compact	Captures decisions before compaction