Codex CLI Cheat Sheet — EdgeSet.dev

⭐ Most Used Commands

Command	Explanation	Typical Usage
codex	Start an interactive session in the current directory	Beginning a coding session — Codex reads your files and waits for instructions
codex "prompt"	Run a one-shot prompt and exit — no interactive session	Quick tasks from the shell: `codex "add error handling to server.py"`
codex -a auto-edit	Auto-apply file edits without asking; still prompts before running shell commands	Faster editing sessions where you trust the changes but still want shell control
codex -a full-auto	Fully autonomous — applies edits and runs commands without prompting	CI pipelines, sandboxed environments, or tasks you've already reviewed the approach for

⭐ Most Used Slash Commands

Command	Explanation	Typical Usage
/resume	Reload a previously saved session — picks up context, files, and conversation history	Continuing work from a previous day or after closing the CLI mid-task
/compact	Summarises conversation history to free up context while keeping the working thread	Session is getting long but the task isn't done — preserve context without starting over
/plan	Enter planning mode — Codex proposes an execution strategy and waits for approval before acting	Before a risky or multi-step task; when you want to review the approach before anything is touched
/status	Display current session config: model, approval mode, and token usage	Quick check on what model you're using or how much context has been consumed
/new	Start a fresh task in the same CLI instance — clears conversation without restarting the process	Moving to a completely unrelated task without closing and reopening the terminal

💬 Other Slash Commands

/diff Review Git changes in the current repo before committing

/fast Toggle performance optimisation — faster responses on supported models

/fork Branch the current conversation — explore an alternative approach without losing the original thread

/model Switch AI model mid-session without restarting

/permissions Adjust approval requirements on the fly — e.g. switch between auto and read-only mid-session

/personality Change communication style: friendly, pragmatic, or neutral

🛡️ Approval Modes

suggest

Default. Codex proposes changes and explains what it would do — you review and apply manually. Nothing is written or run without your action. Safest mode for getting started.

auto-edit

Middle ground. File edits are applied automatically; shell commands still require approval. Good for iterative coding sessions where you want speed but keep control over what runs.

full-auto

Fully autonomous. All edits and commands are applied without asking. Use inside a sandbox or CI environment. On supported platforms Codex runs network-isolated by default.

🚩 CLI Flags

-a / --approval-mode Set approval mode: suggest, auto-edit, full-auto

-m / --model Model to use — e.g. codex-1, o4-mini, o3

-q / --quiet Suppress interactive UI — pipe-friendly for scripting

--project-doc <path> Load a specific context file instead of AGENTS.md

--no-project-doc Skip loading AGENTS.md entirely

--dangerously-auto-approve-everything Skip all prompts — for trusted CI environments only

--version Print installed version and exit

--help / -h Show available flags and usage

📄 AGENTS.md

Place an AGENTS.md at your project root to give Codex persistent context — stack, conventions, commands to know, things to avoid.

Codex reads AGENTS.md at the start of every session. It's the Codex equivalent of Claude Code's CLAUDE.md.

What to put in it: tech stack, test commands, code style rules, file layout, anything you'd tell a new engineer on day one.

Use --project-doc <path> to load a different file, or --no-project-doc to skip it entirely.

⚠️ Keep AGENTS.md concise. Focus on non-obvious rules and constraints — Codex can already read your code.

⌨️ Keyboard Shortcuts

Enter Submit message

Ctrl+C Interrupt / cancel the current operation

Ctrl+D Exit Codex

↑ / ↓ Navigate prompt history

h Show inline help during a session

q Quit the session

🧠 Models at a Glance

codex-1 Default. OpenAI's code-optimised agentic model — best balance of speed, cost, and coding ability

o4-mini Fast, cost-effective reasoning — good for straightforward tasks and scripting

o3 Most powerful reasoning — use for complex architecture or hard debugging; slower and pricier

gpt-4.1 Latest GPT-4 generation — strong general-purpose coding and instruction following

Switch with codex -m o3 "prompt" or set a default in ~/.codex/config.yaml.

⚙️ Setup & Config

npm install -g @openai/codex Install Codex CLI globally

npm update -g @openai/codex Update to the latest version

API key: set OPENAI_API_KEY as an environment variable, or add it to ~/.codex/config.yaml.

Config file: ~/.codex/config.yaml — set your default model, approval mode, and other preferences so you don't have to pass flags every time.

Example config.yaml:
model: codex-1
approvalMode: auto-edit

🤖 Scripting & Headless Use

Run one-shot tasks from the shell — useful for CI, makefiles, and automation pipelines.

codex -q "prompt" Quiet mode — suppresses UI, outputs result only. Pipe-friendly.

Pipe input: cat file.py | codex -q "add docstrings to every function"

Redirect output: codex -q "generate tests for auth.py" > tests/test_auth.py

In CI, combine flags: codex -q -a full-auto --dangerously-auto-approve-everything "fix lint errors"

💡 Tips & Patterns

Start with suggest mode while learning the tool. Switch to auto-edit once you're comfortable, and reserve full-auto for sandboxed CI or pre-reviewed tasks.

AGENTS.md is your first lever. Put project conventions, preferred libraries, and "never do X" rules in it — Codex reads it every session and it shapes every response.

Use --project-doc to switch contexts — e.g. one AGENTS file for backend work, another for frontend. Pass it per-invocation without touching the default file.

Sandbox full-auto. On macOS, full-auto mode runs in a network-isolated sandbox by default. On Linux, use Docker or a VM to get equivalent isolation.

Provide a clear task scope. "Refactor the auth module to use JWTs" works better than "improve the code". The more specific your prompt, the fewer back-and-forth rounds.

Interrupt safely. Ctrl+C cancels the current operation mid-stream. Codex will not apply partial changes — you can redirect or refine your prompt and try again.

Check your version. Codex is actively developed — run codex --version and npm update -g @openai/codex regularly to pick up new models and fixes.