Codex CLI gets long-horizon /goal mode

The interesting part is not the headline task length, it’s the shift from turn-by-turn prompting to a durable objective that can survive compaction, retries, and long stretches of unattended work.

• –`/goal` changes the unit of work from “answer this prompt” to “keep advancing this objective,” which is what long-running engineering tasks actually need
• –The value here is mostly workflow reliability: fewer interruptions, less context drift, and better recovery when the agent hits a bad path
• –That matters most for backend refactors, migrations, and bug hunts where progress comes in small checkpoints over many hours
• –OpenAI’s open-source CLI makes this easier to inspect and trust than a black-box agent, which is a big part of why Codex keeps mattering to AI developers
• –The competitive pressure is obvious: this is Codex trying to close the ergonomics gap with other terminal-first coding agents while leaning harder into autonomous execution