Mastering Codex Commands: A Complete Beginner‑to‑Pro Guide

Codex is an AI‑driven coding collaborator that operates directly on local project files, unlike pure chat models such as ChatGPT. It requires precise / ‑prefixed commands to perform actions such as reading code, modifying files, executing terminal commands, and conducting code reviews.

Input Types

Two input categories exist: ordinary prompts that describe a specific programming need (e.g., “fix the parameter validation error in the login API”), and slash commands (starting with /) that act as operation switches, such as changing the model or setting permissions.

Key Reminder

Slash commands must appear at the very beginning of the input line; otherwise Codex treats them as normal text and ignores the command. When a task is running, pressing Tab after a slash command queues it for execution after the current task finishes.

Onboarding Workflow

Navigate to the project directory and verify the path with /status.

Run /init to generate an AGENTS.md skeleton and add project‑specific rules (package manager, prohibited directories, coding standards, etc.).

Use /status again to confirm the model, permissions, and directory; start with “Read Only” mode for safety.

Let Codex familiarize itself with the code base, e.g., /mention src/ analyze core logic, without modifying files.

For large tasks, issue /plan to obtain a step‑by‑step execution plan, risk points, and verification methods.

After confirming the plan, enable write permissions with /permissions and let Codex apply changes.

Run /diff to review the exact Git diff and ensure no unintended modifications.

Before committing, invoke /review for a self‑audit of bugs, regressions, missing tests, edge cases, and security risks.

If the conversation becomes sluggish, use /compact to compress the history and free tokens.

Finish the session with /new or /clear for a fresh task, or /quit to exit (remember to save your work).

Core 10 Commands (cover ~90% of daily development)

/init : Generates AGENTS.md so Codex can quickly learn project structure, dependencies, and conventions.

/status : Shows the current model, permission mode, working directory, token usage, and other session details.

/model : Switches the active model and inference intensity; GPT‑5 is the recommended default for complex tasks.

/permissions : Sets the approval mode – “Read Only”, “Auto”, or “Full Access” – controlling whether Codex may write files or run commands.

/plan : Enters planning mode, outputting a detailed execution plan before any code changes.

/mention : Limits Codex’s focus to specific files or directories, improving relevance and speed.

/diff : Displays the Git diff (including untracked files) so developers can verify exact changes.

/review : Performs a comprehensive self‑review for bugs, regressions, missing tests, edge cases, and security issues.

/compact : Summarizes long conversations to release token capacity while preserving key objectives.

/resume : Restores a previously saved session, enabling multi‑day work without losing context.

Advanced Commands

/fast : Enables a faster, lower‑precision mode for lightweight tasks.

/personality : Changes Codex’s response style (friendly, pragmatic, or none).

/copy : Copies the latest output to the clipboard (or via Ctrl+O / Cmd+O).

/agent : Switches to a sub‑agent thread for parallel sub‑tasks within the same overall job.

/fork : Creates an independent session branch, analogous to a Git branch, for experimental solutions.

/apps , /plugins , /mcp : Browse and invoke external applications, plugins, or Model Context Protocol tools.

/ps / /stop : Monitor and abort long‑running background terminal tasks.

/experimental : Turn on experimental features (use with caution).

/debug-config : Prints detailed configuration and permission diagnostics when Codex ignores commands.

/statusline , /title , /keymap : Customize the TUI status bar, window title, and keyboard shortcuts.

/sandbox-add-read-dir : Windows‑only command to grant the sandbox read access to an extra directory.

/logout , /feedback , /quit , /exit , /new , /clear : Basic session management commands.

Common Questions

Command lists vary by Codex version, OS, installed plugins, and experimental flags; use / to view the local list.

Codex may “mis‑edit” code if permissions are too permissive, prompts are vague, or /diff is not checked; start with “Read Only” mode. /compact compresses history but retains the task goal, key steps, and completed results; add a prompt to preserve specific files if needed. /new opens a fresh dialogue while keeping history; /clear resets the current context and clears the terminal.

After modifications, always run /diff, then local tests, and finally /review before committing.

Codex differs from ChatGPT by directly accessing the local filesystem, executing commands, and reading Git history. /fork creates an independent branch; /agent creates a sub‑task that remains linked to the original session.

Use /apps to insert $app‑slug and call external services such as databases.

If an experimental feature bugs out, disable it with /experimental and restart Codex; report serious issues via /feedback.

Update Codex via npm install -g @openai/codex@latest (npm) or brew upgrade codex (Homebrew).

Officially supported OSes are macOS and Linux; Windows support is experimental and recommended through WSL.

References

https://developers.openai.com/codex/cli https://developers.openai.com/codex/cli/slash-commands

https://x.com/dengdry/status/2052026438442725525