- NEVER use
pkillor broad process-killing commands (likepkill -f "claudeup"orpkill -f "Codex"). This kills all Codex CLI sessions running on the machine. Instead, ask the user to restart applications manually or close specific windows. - Do not use hardcoded paths in code, docs, comments, or any other files.
Repository: Magus Purpose: Professional plugin marketplace for Codex Owner: Jack Rudenko (i@madappgang.com) @ MadAppGang License: MIT
| Plugin | Version | Purpose |
|---|---|---|
| Code Analysis | v4.0.2 | Codebase investigation with claudemem MCP, 13 skills |
| Multimodel | v2.6.2 | Multi-model collaboration and orchestration |
| Agent Development | v1.5.5 | Create Codex agents and plugins |
| SEO | v1.6.5 | SEO analysis and optimization with AUTO GATEs |
| Video Editing | v1.1.1 | FFmpeg, Whisper, Final Cut Pro integration |
| Nanobanana | v2.3.1 | AI image generation with Gemini 3 Pro Image |
| Conductor | v2.1.1 | Context-Driven Development with TDD and Git Notes |
| Dev | v1.39.0 | Universal dev assistant with workflow enforcement, 47 skills |
| Designer | v0.2.0 | UI design validation with pixel-diff comparison, 6 skills |
| Browser Use | v1.0.0 | Full-platform browser automation, 18 MCP tools, 5 skills |
| Statusline | v1.4.1 | Colorful statusline with worktree awareness |
| Terminal | v2.1.0 | Intent-level terminal: run, watch, observe, repl, tui + ht-mcp/tmux-mcp |
Claudish CLI: npm install -g claudish - Run Codex with OpenRouter models (separate repo)
Codex/
├── AGENTS.md # This file
├── README.md # Main documentation
├── RELEASE_PROCESS.md # Plugin release process guide
├── .env.example # Environment template
├── .Codex-plugin/
│ └── marketplace.json # Marketplace plugin listing
├── plugins/ # All plugins (12 published, 3 unlisted)
│ ├── code-analysis/ # v4.0.2 — 13 skills, 1 agent, claudemem MCP
│ ├── multimodel/ # v2.6.2 — 15 skills
│ ├── agentdev/ # v1.5.5 — 5 skills
│ ├── seo/ # v1.6.5 — 12 skills
│ ├── video-editing/ # v1.1.1 — 3 skills
│ ├── nanobanana/ # v2.3.1 — 2 skills
│ ├── conductor/ # v2.1.1 — 6 skills
│ ├── dev/ # v1.39.0 — 47 skills, workflow enforcement
│ ├── designer/ # v0.2.0 — 6 skills, pixel-diff design validation
│ ├── browser-use/ # v1.0.0 — 5 skills, 18 MCP tools
│ ├── statusline/ # v1.4.1 — 1 skill
│ ├── terminal/ # v2.1.0 — 2 skills, ht-mcp + tmux-mcp
│ └── (go, instantly, autopilot — unlisted)
├── autotest/ # E2E test framework
│ ├── framework/ # Shared runner, parsers (Bun/TS)
│ ├── coaching/ # Coaching hook tests
│ ├── designer/ # Designer plugin tests (12 cases)
│ ├── subagents/ # Agent delegation tests
│ ├── team/ # Multi-model /team tests
│ ├── skills/ # Skill routing tests
│ ├── terminal/ # Terminal plugin tests (9 cases)
│ └── worktree/ # Worktree tests
├── tools/ # Standalone tools
│ ├── claudeup/ # TUI installer (npm package, v3.5.0)
│ ├── claudeup-core/ # Core library
│ └── claudeup-gui/ # GUI version
├── skills/ # Project-level skills
│ └── release/SKILL.md
├── ai-docs/ # Technical documentation
└── docs/ # User documentation
.Codex-plugin/marketplace.json— Marketplace listing (update when releasing!)plugins/{name}/plugin.json— Plugin manifest (version, components, MCP servers)plugins/{name}/.mcp.json— MCP server config (if plugin has MCP servers)RELEASE_PROCESS.md/skills/release/SKILL.md— Release process docsautotest/framework/runner-base.sh— E2E test runner entry point
# Run a test suite (all use autotest/framework/ shared runner)
./autotest/terminal/run.sh --model Codex-sonnet-4-6 --parallel 3
./autotest/coaching/run.sh --model Codex-sonnet-4-6
./autotest/designer/run.sh --model Codex-sonnet-4-6
./autotest/subagents/run.sh --model grok
# Run specific test cases
./autotest/terminal/run.sh --model Codex-sonnet-4-6 --cases environment-inspection-08
# Analyze existing results
bun autotest/terminal/analyze-results.ts autotest/terminal/results/<run-dir>Required:
APIDOG_API_TOKEN=your-personal-token
FIGMA_ACCESS_TOKEN=your-personal-tokenOptional:
GITHUB_PERSONAL_ACCESS_TOKEN=your-token
CHROME_EXECUTABLE_PATH=/path/to/chrome
CODEX_API_KEY=your-codex-keyPlugin System Format:
- Plugin manifest:
.Codex-plugin/plugin.json(must be in this location) - Settings format:
enabledPluginsmust be object with boolean values - Component directories:
agents/,commands/,skills/at plugin root - MCP servers:
.mcp.jsonat plugin root (referenced as"mcpServers": "./.mcp.json"in plugin.json) - Environment variables: Use
${CLAUDE_PLUGIN_ROOT}for plugin-relative paths
Quick Reference:
# Install marketplace
/plugin marketplace add MadAppGang/magus
# Local development
/plugin marketplace add /path/to/CodexEnable in .Codex/settings.json:
{
"enabledPlugins": {
"code-analysis@magus": true,
"dev@magus": true,
"terminal@magus": true
}
}IMPORTANT: For complex tasks, prefer delegating to specialized agents via the Task tool rather than handling inline. Delegated agents run in dedicated context windows with sustained focus, producing higher quality results.
| Task Pattern | Delegate To | Trigger |
|---|---|---|
| Research: web search, tech comparison, multi-source reports | dev:researcher |
3+ sources or comparison needed |
| Implementation: creating code, new modules, features, building with tests | dev:developer |
Writing new code, adding features, creating modules - even if they relate to existing codebase |
| Investigation: READ-ONLY codebase analysis, tracing, understanding | code-analysis:detective |
Only when task is to UNDERSTAND code, not to WRITE new code |
| Debugging: error analysis, root cause investigation | dev:debugger |
Non-obvious bugs or multi-file root cause |
| Architecture: system design, trade-off analysis | dev:architect |
New systems or major refactors |
| Agent/plugin quality review | agentdev:reviewer |
Agent description or plugin assessment |
Key distinction: If the task asks to IMPLEMENT/CREATE/BUILD -> dev:developer. If the task asks to UNDERSTAND/ANALYZE/TRACE -> code-analysis:detective.
NOTE: Skills use the Skill tool, NOT the Task tool. The namespace:name format is shared by both agents and skills -- check which tool to use before invoking.
| Need | Invoke Skill | When |
|---|---|---|
| Semantic code search, claudemem CLI usage, AST analysis | code-analysis:claudemem-search |
Before using claudemem commands |
| Multi-agent claudemem orchestration | code-analysis:claudemem-orchestration |
Parallel claudemem across agents |
| Architecture investigation with PageRank | code-analysis:architect-detective |
Architecture-focused claudemem usage |
| Deep multi-perspective analysis | code-analysis:deep-analysis |
Comprehensive codebase investigation |
| Database branching with git worktrees (Neon, Turso, Supabase) | dev:db-branching |
Worktree creation with schema changes needing DB isolation |
| Interactive terminal: run commands, dev servers, test watchers, REPLs | terminal:terminal-interaction |
Task needs TTY, interactive output, long-running process, or database shell |
| TUI navigation: vim, nano, htop, lazygit, k9s, less | terminal:tui-navigation-patterns |
Navigating TUI apps, sending key sequences, reading screen state |
| Claudish CLI usage, model routing, provider backends | code-analysis:claudish-usage |
Before ANY claudish command — bare model names, no prefixes |
Version History: See CHANGELOG.md | Detailed Notes: See RELEASES.md
Git tag format: plugins/{plugin-name}/vX.Y.Z
Plugin Release Checklist (ALL 3 REQUIRED):
- Plugin version -
plugins/{name}/plugin.json->"version": "X.Y.Z" - Marketplace version -
.Codex-plugin/marketplace.json-> plugin entry"version": "X.Y.Z" - Git tag -
git tag -a plugins/{name}/vX.Y.Z -m "Release message"-> push with--tags
Missing any of these will cause claudeup to not see the update!
Claudeup Release Process:
- Update
tools/claudeup/package.json->"version": "X.Y.Z" - Commit:
git commit -m "feat(claudeup): vX.Y.Z - Description" - Tag:
git tag -a tools/claudeup/vX.Y.Z -m "Release message" - Push:
git push origin main --tags
The workflow .github/workflows/claudeup-release.yml triggers on tools/claudeup/v* tags (builds with pnpm, publishes to npm via OIDC).
Maintained by: Jack Rudenko @ MadAppGang Last Updated: March 3, 2026