8/10
Strong
Codex is the safest first install on the facts given. The stated strength is sandbox defaults, plus ChatGPT-plan access, and the cost is a ChatGPT plan or API. Install is the OpenAI Codex docs, official and not a single command here. Where it runs is not stated, so this score does not assume a machine. Scored 8 for day-one safety on a real repo, not for project memory, which this brief does not claim.
- Cost
- ChatGPT plan or API
- Where it runs
- Not stated
- Standout trait
- Sandbox defaults and ChatGPT-plan access
- Project file
AGENTS.md
- Install
curl -fsSL https://chatgpt.com/codex/install.sh | sh - Open a repocd /path/to/your/repo
- Give one taskRun codex. Ask for one reversible edit in a single file. The sandbox default is the reason this is the first write.
- Review the diffgit status then git diff.
KeepName sandbox defaults before the first write. That is the stated first-run advantage.
Best practices
- Put test commands, writable files, the definition of done, and commit conventions in AGENTS.md, then call them through a reusable /review. @OpenAICodexCli
- Keep the tool list small and the prompt tight. Raise reasoning effort only when the task needs more depth. @OpenAICodexCli
- Leave model_reasoning_effort at the default, medium, so it can vary effort on its own. @embirico
- Use /agents to see running tasks and jump between them. Use /fork to try another direction in a worktree. @OpenAIDevs
Further reading
- Codex CLI Guide (2026): How It Works, Costs, Models AI Builder Club · Third-party