Opinionated Codex configuration with tooling and workflow defaults.
Caution
This project is under active development. Repository updates may change agent behavior, command policies, tool access, and workflow defaults at any time.
Review the configuration before using it as ~/.codex, especially rules, skills, and hooks that may affect security-sensitive workflows.
Use at your own risk. The author is not responsible for how you use this configuration, its rules, or agent outputs.
├── agents/ # Custom subagents
├── docs/ # Project documentation
│ ├── rationale.md # Design rationale
│ └── readme.md # Project overview
├── rules/ # Command policies (Starlark)
├── skills/ # Agent skills
│ ├── github_delivery/ # GitHub delivery overlay
│ ├── search_guide/ # Search and source-selection overlay
│ ├── software_engineering/ # Application engineering ground skill
│ └── system_engineering/ # System-boundary engineering ground skill
├── AGENTS.md # Global instructions
├── config.toml # Configuration
├── hooks.json # Lifecycle hooks
└── license # Project license
Each component is independent. Removing one does not break the rest.
This repository is a portable Codex configuration — clone it into ~/.codex and the agent can pick up the whole baseline: model settings, sandbox policies, custom subagents, command rules, skills, hooks, and MCP servers.
The goal is a reproducible, opinionated baseline for how Codex should behave. Not a framework. Not a plugin system. A configuration.
Explicit over implicit. Every important behavior is declared in a file. No hidden project defaults, no magic environment variables, no runtime drift.
Composable layers. Components are independent. config.toml works without agents/. skills/ work without rules/. Add what you need, ignore what you don't.
Readable source of truth. If you can read TOML, Markdown, and Starlark, you can understand the entire configuration. No abstractions, no generated files.
Declarative over imperative. Configuration and skills describe what the agent should be. Scripts may exist as explicit skill resources, but setup does not depend on state-mutating scripts, conditional logic, or platform-specific branching.
No unexplained black boxes. Every project behavior traces back to a file in this repository. If something looks like magic, it is not — follow the import chain.
See rationale for the reasoning behind selected defaults, skills, rules, and agent behavior.