By OpenAI shipped Codex CLI v0.133 on May 21 and quietly flipped a switch that most developers missed: /goal is now enabled by default. Every Codex CLI install now has persistent, resumable agent sessions out of the box. Set a goal, close your laptop, open it the next morning — Codex picks up where it left off. This is the session-persistence story that AI coding tools have been promising for months, and it subtly reshapes the Codex vs Claude Code comparison.
What Changed in v0.133
/goal existed before v0.133, but it was opt-in and experimental. Starting May 21, goals are enabled by default and backed by dedicated app-server storage. That storage distinction matters: the goal state isn’t just in your terminal history. It’s stored server-side, which means it survives network drops, token budget exhaustion, and laptop shutdowns. You’re not babysitting a terminal session anymore.
The TUI gives you four controls: create, pause, resume, and clear. The syntax is minimal:
# Set a goal
/goal Migrate all v1 API calls to v2. Don't touch DB schema. Validate: all tests pass.
# Check current goal and progress
/goal
# Pause without losing state
/goal pause
# Resume where you left off
/goal resume
# Install the latest version
npm install -g @openai/codex@0.134.0How to Write Goals That Actually Work
OpenAI is clear about when /goal makes sense and when it doesn’t. A good goal is bigger than a single prompt but smaller than a vague open-ended backlog. It needs a verifiable stopping condition — otherwise the agent keeps running indefinitely. The structure that works: state what to achieve, explicitly list what not to change, describe how the agent should validate progress, and define when to stop. Skip /goal entirely if you’re still exploring or making judgment calls mid-task. The feature shines on migrations, large refactors, deployment retry loops, and experiments with a clear “done” definition. See OpenAI’s official /goal use cases for more patterns.
What Else Shipped: v0.133 and v0.134
The /goal change arrived alongside meaningful updates to permission profiles and remote control. In v0.133, permission profiles gained list APIs, inheritance, managed requirements.toml support, and stronger Windows sandbox integration. The codex remote-control command now runs as a foreground process with daemon-style start/stop — useful for CI and long-running pipelines. A week later, v0.134 (May 26) added MCP improvements: per-server environment targeting, OAuth for streamable HTTP servers, and concurrent execution for read-only MCP tools. Local conversation history search and --profile as the primary profile selector round out the update. The full changelog is on the OpenAI Codex changelog. The AppShots feature — press both Command keys to send the frontmost app window to Codex as context — is the kind of small thing that becomes a daily habit fast.
How This Compares to Claude Code
Neither Codex nor Claude Code has persistent memory across sessions by default — context files (CLAUDE.md, AGENTS.md) are the standard workaround for both. What /goal adds is different from memory: it’s durable task state stored server-side. That’s a meaningful distinction. The community consensus from several May 2026 comparisons is blunt: “If you want an agent that survives across days, Codex just leapfrogged Claude Code.” That’s directionally accurate for the persistence story specifically.
The more complete picture:
| Feature | Codex /goal | Claude Code |
|---|---|---|
| Session persistence | Server-side durable state | Context files only |
| Goal TUI controls | Create / pause / resume / clear | Session-scoped, no native pause |
| Native cloud | Yes, via ChatGPT dashboard | Requires SSH or external platform |
| Model | GPT-5.3 Codex | Claude Sonnet / Opus 4.x |
| Avg monthly cost | $100–200/dev | $150–250/dev |
Claude Code’s approach isn’t wrong — it bets on model quality per session and verifiable stopping conditions rather than raw persistence. That philosophy works well for tasks that fit inside a session. Codex /goal targets work that doesn’t. Choose accordingly.
The Practical Takeaway
If your codebase has a migration, a refactor, or a class of error you’ve been putting off because it will take multiple hours of active attention, /goal is worth trying. The key is writing the goal well — a vague objective produces vague results, same as any other prompt. Tight stopping conditions, explicit exclusions, and a validation strategy are what separate useful /goal sessions from runaway agents. And since longer runs mean more tokens, set a budget before you start: Codex GPT-5.3 runs $14/M output tokens, and a full-afternoon refactor will produce a lot of output. Codex Mini at $3/M output is a reasonable option for lower-stakes goals where cost matters more than speed.
