OpenAI Codex is easy to take for a smart autocomplete in the terminal and stop there. But it is not a suggester - it is an agentic runtime: it receives context, chooses a tool, checks the approval and the sandbox, performs an action and continues the loop until the task is solved or stopped. The difference is fundamental. Once you understand that it is a controllable environment with its own trust boundaries, configuration and precedence, it turns from a tool for quick edits into a reliable part of the engineering process - from the first run to CI and corporate policy.

This book takes Codex from installation to fine-tuning. First - a checkable foundation: a map of surfaces (CLI, IDE, desktop, cloud, API), installation and updates, authorization and the source of billing, the first run and the agentic loop itself. Then - daily work: the TUI and composer, the task contract, plan mode, implementation with verification, review, sessions, git and worktrees, models and context control. Next - trust: permission profiles, approvals, the sandbox, execpolicy, prompt injection and secrets. After that - configuration and extension: config layers and AGENTS.md, memories, skills, custom agents, plugins, hooks, MCP. Then - interfaces and automation: desktop and IDE, cloud, GitHub Code Review, Slack and Linear, codex exec, the SDK, the App Server. And finally - production: the enterprise profile, diagnostics and five references to the official interface.

The material rests on one distinction. Where a fact, a command name, a key, a path or a precedence order is directly described by OpenAI, the book marks it as official and cites the primary source. Where a technique follows from documented behavior but is not a product requirement, it is an engineering conclusion - useful but not passed off as a rule. This split matters more than convenient generalizations: a setting becomes useful only when you understand its source, its allowed value and the cost of a mistake.