How to Use /goal in Claude Code and Codex CLI

/goal is here and it's a game changer. You type /goal, describe a stopping condition, and the agent just keeps going. Multiple turns, no prompting, no babysitting. It works until the job is done. Codex CLI shipped it first in late April, and Claude Code followed on May 12.

I tried it today and it wasn't there. Ran claude update, restarted, and boom. There it was. Fast, easy, and immediately useful.

---

What /goal Does

Normally when you give Claude Code (or Codex) a task, the agent works through it and stops when it thinks it's done. You review, maybe prompt again, repeat. /goal changes that loop. You set a completion condition, and the agent keeps working across multiple turns until that condition is met.

After each turn, a separate evaluator model (Haiku, in Claude Code's case) checks whether the goal has been satisfied. If not, the agent starts another turn automatically. You don't have to keep prompting it.

That's the key design choice: the model doing the work is not the same model deciding if it's done. Separation of concerns at the LLM level.

---

Both Companies Shipped It

OpenAI added /goal to Codex CLI first, landing it across five PRs in late April 2026. Anthropic followed in mid-May with Claude Code 2.1.139. The implementations differ under the hood but the UX is nearly identical: type /goal, describe your stopping condition, walk away.

Codex's version uses a SQLite-backed persistence layer with thread-level state tracking. Claude Code's version is lighter, using a fast evaluator model to gate turn boundaries. Both support pausing and resuming goals across sessions.

---

How to Use It

Set a goal:

/goal all tests in test/auth pass and the lint step is clean

Check status:

/goal

Clear it:

/goal clear

That's it. One goal per session. If you end the session while a goal is active, it picks back up when you resume with --resume or --continue.

Good goals have a measurable end state. "Make the auth flow better" is vague. "All tests in test/auth pass and coverage is above 80%" gives the evaluator something concrete to check. Include a turn limit if you don't want it running indefinitely: "or stop after 20 turns."

---

The Gotcha: You Might Need to Update First

This is what caught me today. I typed /goal and nothing happened because my Claude Code version was too old. The feature shipped in version 2.1.139 (May 12, 2026), and if you installed Claude Code more than a few days before that, you don't have it yet.

The fix:

claude update

That's the whole fix. Run it, wait a few seconds, restart your session. /goal should be there now. If you installed via npm instead, npm update -g @anthropic-ai/claude-code works too. Homebrew users: brew upgrade claude-code.

Same applies to Codex. If /goal isn't showing up, make sure you're on at least version 0.128.0.

---

When to Use It

This isn't for every task. If you're making a quick edit or asking a question, /goal is overkill. It's useful when you have a well-defined end state and want the agent to grind through it without you babysitting each step.

Good fits: getting a test suite green, fixing lint errors across a repo, migrating a batch of files to a new format. Bad fits: anything where you want to review each change before it happens, or where the stopping condition is subjective.

Worth noting that a vague goal can burn through your token budget fast. The agent will keep trying if it can't tell whether the condition is met. Be specific, set turn limits, and check in on long-running goals.

---

In Other News: Anthropic Made Their Call on Third-Party Agents

Anthropic finally decided what to do about OpenClaw and other third-party tools hitting Claude through subscriptions. Back in April they banned it outright. Now they're bringing it back, but on a meter.

Starting June 15, every paid Claude subscriber gets a separate monthly credit pool for "programmatic" usage (SDKs, CLI, third-party agents like OpenClaw). The credit matches your plan price: $20 for Pro, $100 for Max 5x and Max 20x. Once it's gone, you either stop or pay API rates out of pocket with extra usage enabled.

This means I can switch back to Claude as my OpenClaw backend without worrying about getting banned. They won't shut me down for using it. They'll just charge me. And honestly, that's fine. It's the right call. I'd rather pay overages than get cut off.

I also still have the $200 credit from Anthropic's April promotion, when they gave every subscriber a one-time extra usage credit (scaled by plan tier, claimed through the Usage settings page, expires 90 days after claiming). That cushion helps.

But it's not cheap. Between Claude Code overages and OpenClaw usage, I've spent about $45 in extra usage just today, and today was a light day. My ChatGPT Plus subscriptions hit their weekly token caps, so everything rolled over to Claude. That adds up fast.

The industry is clearly moving toward metered economics for agentic workloads. OpenAI has always billed API usage this way. GitHub is shifting Copilot to tokens and credits. Anthropic held out with flat-rate subscriptions longer than most, but the math doesn't work when subscribers are consuming thousands of dollars in compute on a $200/month plan. This was inevitable.

---