vladolaru/claude-code-plugins

My personal Claude Code and Codex plugin marketplace. These open source tools cover development workflows, code review, knowledge capture, prompt optimization, and agent safety.

Each plugin started as something I needed for my own work. They're opinionated, actively maintained, and battle-tested on real projects.

Quick Start

Claude Code

# Add the marketplace
/plugin marketplace add vladolaru/claude-code-plugins

# Install what you need
/plugin install pirategoat-tools@vladolaru-claude-code-plugins
/plugin install dex@vladolaru-claude-code-plugins
/plugin install prompt-engineer@vladolaru-claude-code-plugins
/plugin install image-optimizer@vladolaru-claude-code-plugins
/plugin install yoloing-safe@vladolaru-claude-code-plugins
/plugin install caffeinate-claude@vladolaru-claude-code-plugins

# Restart Claude Code to activate

Claude Code commands use slash syntax, for example /pr-review 42.

Codex

# Add the same marketplace
codex plugin marketplace add vladolaru/claude-code-plugins

# Install what you need
codex plugin add pirategoat-tools@vladolaru-claude-code-plugins
codex plugin add dex@vladolaru-claude-code-plugins
codex plugin add prompt-engineer@vladolaru-claude-code-plugins
codex plugin add image-optimizer@vladolaru-claude-code-plugins
codex plugin add yoloing-safe@vladolaru-claude-code-plugins
codex plugin add caffeinate-claude@vladolaru-claude-code-plugins

Codex exposes command workflows as explicit plugin skills, for example $pirategoat-tools:pr-review 42. Plugins with hooks require the normal Codex trust review before those hooks run.

Plugins

Plugin	What it does
pirategoat-tools	Development tools - domain reviewer agents, pipeline and cross-validation agents, shared skills, rich feedback loops
dex	Knowledge capture — frictionless capture of learnings, patterns, and decisions from conversations into agent-first docs
prompt-engineer	Prompt optimization — evidence-grounded pattern attribution with human-in-the-loop approval gates
image-optimizer	Image optimization — measured per-file picks under quality profiles, banding checks and a 1:1 review, for PNG, JPEG, GIF, SVG, WebP, AVIF
yoloing-safe	YOLO mode safety net - PreToolUse guardrails that block destructive commands and gate risky ones with host-aware behavior
caffeinate-claude	Mac sleep prevention - keeps your Mac awake during Claude Code and Codex sessions, multi-tab aware

pirategoat-tools

The main plugin. Domain reviewers running in parallel, two pipeline agents (reconciliation + decision critic), two cross-validators (Gemini + Codex), and utility agents (technical writer + repo-reviewer-adapter). Plus shared skills covering testing patterns (190KB reference library), software architecture (87KB pattern library), WordPress/WooCommerce development, Figma-to-code workflows, and browser automation.

Full documentation → | Changelog

dex

Frictionless knowledge capture based on compound engineering — each unit of work should make the next one easier, not harder. Fire /dex after any engineering work, confirm with a single selection, done. Knowledge lives as agent-first documents in .claude/docs/ with optional promotion to CLAUDE.md.

Full documentation → | Changelog

prompt-engineer

Systematic prompt optimization through a 5-phase workflow with human approval gates between phases. Every technique recommendation is evidence-grounded — quoted trigger conditions from a comprehensive research-backed reference library (50+ techniques).

Full documentation → | Changelog

image-optimizer

Measured image optimization: tries encoder settings per file, keeps the smallest that passes a quality profile (lossless by default), and shows a side-by-side review before anything changes. One command: /optimize-images path/to/images.

Full documentation → | Changelog

yoloing-safe

YOLO mode safety net. A PreToolUse hook that evaluates every tool call against safety rules - blocks the irreversible stuff (rm -rf, npm publish, credential access), gates risky-but-maybe-intentional actions (force push, brew install, terraform destroy), and lets everything else fly. Claude Code presents confirmation prompts for the ask tier; Codex fails closed with configuration guidance because its hook protocol cannot surface those prompts. Configurable credential patterns and zero-access paths. Zero dependencies, zero config needed.

Full documentation → | Changelog

caffeinate-claude

Keeps your Mac awake during Claude Code and Codex sessions using macOS caffeinate. Supports multiple tabs, handles crashes gracefully with a 1-hour safety timeout, and cleans up automatically when all sessions end. Zero config needed.

Full documentation → | Changelog

Repository Structure

vladolaru-claude-code-plugins/
├── .agents/
│   └── plugins/
│       └── marketplace.json      # Generated Codex marketplace
├── .claude-plugin/
│   └── marketplace.json          # Canonical marketplace registry
├── plugins/
│   ├── pirategoat-tools/         # reviewer agents, shared skills, review commands
│   ├── dex/                      # knowledge-capture commands, shared skill, tests
│   ├── prompt-engineer/          # optimize-prompt command, shared skill, reference library
│   ├── image-optimizer/          # optimize-images command, image-optimization skill, imgopt scripts, tests
│   ├── yoloing-safe/            # PreToolUse safety hook, tests
│   └── caffeinate-claude/       # macOS sleep prevention hooks
├── scripts/
│   └── generate_codex_compat.py  # Claude source to Codex adapters
├── CLAUDE.md
├── LICENSE
└── README.md

Every plugin directory contains a generated .codex-plugin/plugin.json. Plugins with commands also contain generated adapters under codex-skills/, isolated from Claude Code's top-level skill discovery.

Dual-Host Maintenance

Claude Code definitions are the canonical authoring format. Codex manifests and command-to-skill adapters are deterministic generated output. This avoids maintaining two independent prompt trees while leaving shared skills, scripts, hooks, and reviewer definitions directly reusable by both hosts.

When a plugin, command, or marketplace entry changes:

python3 scripts/generate_codex_compat.py
python3 scripts/generate_codex_compat.py --check

Do not hand-edit files marked GENERATED FILE - DO NOT EDIT. The repository test suite and the dual-host compatibility workflow run the drift check on every pull request.

The generator intentionally does not copy Claude model labels into Codex. The review pipeline instead dispatches native Codex subagents and gives each one the canonical reviewer definition at runtime. This keeps reviewer instructions synchronized without pretending the two hosts have identical model routing.

claude-code-codex-bridge was a useful reference for host mappings and remains useful for converting user or project configuration. Installing this marketplace does not require the bridge.

Design Patterns

Reusable patterns for building new skills and commands are documented in docs/patterns/.

• Step-by-step prompt injection — Enforce analytical discipline across multi-phase workflows by driving each step from a Python CLI script. Claude calls the script once per step; the script injects that step's instructions and nothing else. Includes script template, skill file structure, testing checklist, and reference implementations from decision-critic.
• Curated-context pipeline — Multi-mode LLM pipelines where a script acts as context curator — reads all state, presents pre-digested briefings, controls flow via condition-driven step routing, and manages split file-based state. Covers pipeline identity anchoring, artifact discipline, and conversational output.

External Tool References

Integration guides for spawning external CLIs as subprocesses, documented in docs/.

• Codex CLI reference — Integrating with the OpenAI Codex CLI: prompting contracts, structured output, headless review invocation, sandbox behavior, and Structured Outputs schema constraints.
• Claude Code CLI reference — Spawning Claude Code CLI as a subprocess: nesting guard status, isolation flags, structured output via --json-schema, settings hierarchy, and Python integration pattern.

Development

git clone https://github.com/vladolaru/claude-code-plugins.git
cd claude-code-plugins

# Add as a local Claude Code marketplace
/plugin marketplace add /path/to/claude-code-plugins

# Install in Claude Code
/plugin install pirategoat-tools@vladolaru-claude-code-plugins

# Add as a local Codex marketplace
codex plugin marketplace add /path/to/claude-code-plugins

# Install in Codex
codex plugin add pirategoat-tools@vladolaru-claude-code-plugins

# Verify generated host adapters
python3 scripts/generate_codex_compat.py --check

Credits

• solatis/claude-config — prompt-engineer skill and reference documents
• zl190/md-slides — creating-md-slides and marp-slide-quality skills
• Jim Humelsine's blog — testing-patterns (190KB) and software-architecture (87KB) reference libraries, synthesized from his excellent architecture and testing series
• Every, Inc. — compound engineering methodology that inspired dex

Resources

• Claude Code documentation
• Plugins guide
• Skills guide
• Codex plugins guide

License

MIT — see LICENSE.