Codex CLI 0.128.0 Ships /goal: A Persistent Loop That Won't Stop Until It's Done

Codex CLI 0.128.0 shipped on April 30 with a new /goal command that adds persistent, loop-based execution to the terminal agent. You set an objective, and Codex keeps iterating toward it across turns until it either achieves it or hits your token budget.

The feature is OpenAI’s take on what the community often calls the “Ralph loop”: an agentic pattern where the model re-evaluates progress after each turn and continues until a terminal condition is reached. Before this, Codex would complete a task and stop. With /goal, it keeps going.

How It Works

The /goal command accepts an objective string and tracks it through a small state machine. Possible states are pursuing, paused, achieved, unmet, and budget-limited. Two injected prompts handle the mechanics: goals/continuation.md drives the looping logic, and goals/budget_limit.md enforces token guardrails.

The basic commands:

• /goal <objective> — start a new goal
• /goal pause — suspend the loop without losing context
• /goal resume — continue from where it left off
• /goal clear — end the goal entirely

Simon Willison noted that the feature works as described. The agent loops, checks whether the goal is met, and continues until it isn’t. The token budget cap prevents runaway execution.

What Else Shipped in 0.128.0

The /goal feature isn’t the only thing in this release. The changelog includes several other changes that improve day-to-day use.

The new codex update command lets you upgrade the CLI without going back to npm or your package manager. Configurable TUI keymaps let you remap the terminal UI shortcuts. Plan-mode nudges now prompt toward structured planning before jumping into execution.

On the infrastructure side, permission profiles got expanded defaults, sandbox CLI profile selection, and working directory controls. The multi-agent system (MultiAgentV2) has more explicit configuration for thread caps and wait times. Plugins got marketplace installation, remote bundle caching, and external agent config import.

One notable removal: --full-auto is deprecated. OpenAI is steering users toward the explicit permission profile system instead of a blanket full-auto flag.

Why /goal Matters

Most agent interactions today are still turn-by-turn. You ask, it answers, you review, you ask again. /goal shifts the model to operate more like a worker than an assistant: you hand it an objective and it handles the iterations without requiring you to re-prompt each time.

The token budget cap is what makes this practical. Without it, a goal that can’t be achieved would just run forever. With it, you set an upper limit on how far Codex will chase something before surfacing an unmet state so you can intervene.

The feature is in the 0.128.0 release, available now via codex update or the standard install path. The slash-command documentation doesn’t yet reflect /goal, which OpenAI has an open issue to fix.

---

Sources: Codex CLI 0.128.0 release, Simon Willison’s notes on /goal, Codex CLI docs