Codex supports agent-written `/goal` prompts for spawned threads

• OpenAI's own docs position /goal as a persistent objective that can keep Codex working for multiple hours, while skirano's demo shows a newer pattern where Codex writes that goal for itself and for each spawned agent.
• thsottiaux's explanation describes agent-visible /goal as a generalization of meta prompting, and the official subagents docs confirm that Codex can spawn parallel agents and route follow-up instructions across their threads.
• The mobile side of this matters because reach_vb's guide treats the phone as a control plane for the host machine, and OpenAI's mobile announcement plus release notes say iOS now supports worktrees, environment setup, inline review comments, and /goal management.
• Users are also converging on the same caveat: according to omarsar0's reply, a weak self-written goal can quietly waste tokens, while onusoz's GEPA note says long solver loops still need explicit iteration bounds.

You can read OpenAI's goal-mode guide, check the official slash-command docs for /side, and browse the mobile rollout post. The weirdly useful bit is that reach_vb's workflow list and skirano's subagent demo are describing the same shift from two ends: the phone is becoming the steering surface, and /goal is becoming something the agent can author for itself.

OpenAI's mobile announcement frames Codex on phone as a way to answer questions, review findings, change direction, and approve what comes next while the real work runs on a laptop, devbox, or remote environment. reach_vb's guide makes that more concrete.

The workflow list in reach_vb's post lines up closely with OpenAI's Jun. 9 release notes, which added branch selection, worktree creation, environment setup scripts, inline review comments, and /goal support in ChatGPT for iOS.

The useful commands break down like this:

• /side: an ephemeral fork for a focused follow-up, per the official slash-command docs
• /plan: shape a multi-step implementation path before committing to execution, per the Codex manual
• /goal: set or manage a durable objective from mobile, per the goal-mode guide

The striking change in skirano's demo is not just that Codex can run a /goal, it is that the user no longer writes the goal text by hand. In a follow-up reply, skirano reduces the setup to a single request.

According to thsottiaux's explanation, Codex can see and set its own /goal, because OpenAI exposes new capabilities to the agent as tools for the agent too. That is a clean extension of what the official goal-mode docs already say: the goal text acts as both starting prompt and completion criteria.

The subagent piece is also real, not just a demo flourish. OpenAI's subagents docs say Codex can spawn specialized agents in parallel, route follow-up instructions, wait for results, and close agent threads, which is exactly the shape skirano's video is leaning on when it generates a main goal and per-agent goals.

Once the agent can draft its own objective, the new bottleneck is goal quality. omarsar0's longer reply argues that the agent often has the best context to write a strong goal, but his later reply says he now reviews that goal and runs it through a goal grader so a bad objective does not just burn tokens for hours.

That caution matches both the official Goals cookbook, which describes a goal as a user-controlled completion contract, and the rougher field report in onusoz's GEPA thread, where a long /goal loop stopped after one iteration and pushed the user toward explicit min_iter and max_iter bounds.

One thing the official docs say plainly, and the social posts mostly imply, is that long runs still need a validation loop. The goal-mode guide recommends goals with a clear target and a checkable definition of done, which helps explain why users are adding graders, verifiers, and iteration caps around the self-authored-goal pattern.

A smaller but revealing detail came from sound4movement's post: after a Mac mini powered off overnight, the Codex app relaunched at login and resumed the long-running /goal. That makes the workflow feel less like a chat session and more like a background job with a UI attached.

That report also fits the mobile-control-plane framing in OpenAI's remote-connections docs, which describe using the ChatGPT mobile app to control a connected Mac or Windows host. sound4movement's example adds the missing operational detail that at least one goal run survived an interrupted machine and continued once the host came back.