Codex adds /goal mode for long-running tasks with remote control preview

• OpenAI's new experimental /goal mode lets Codex keep pursuing a durable objective for hours, with explicit controls for pause, resume, and clear, according to reach_vb's /goal demo and the official Follow a goal docs.
• The shipped 0.129.0 release made goals easier to find and bundled broader persistence work, including a redesigned resume and fork picker, workspace-aware diffs, and shared plugin workflows, per LLMpsycho's release summary and the official 0.129.0 release notes.
• Early users are already running /goal for long stretches, with steipete's 17 hour 59 minute screenshot, mattlam_'s 7 hour run, and mattlam_'s rate-limit screenshot all pointing to much longer unattended sessions than normal turn-based CLI use.
• A remote-control surface is also taking shape: testingcatalog's settings leak showed a Connections panel for keeping incoming sessions alive, while mattlam_'s post about remote control framed the rollout as local session handoff across devices.
• The remote-work story is still half official, half preview: an OpenAI GitHub issue about SSH remote projects and an older user request for mobile remote control line up with the UI evidence, but OpenAI has not published a dedicated remote-control announcement yet.

You can read the official goal guide, skim the full 0.129.0 release notes, and inspect an SSH remote projects issue that walks through Settings > Connections and New remote project. The weird part is how fast the ecosystem moved around it: RepoPrompt added native /goal support, community screenshots are already showing 7 hour and 18 hour runs, and a leaked Connections pane suggests Codex is being stretched from terminal agent into something closer to a persistent cross-device workspace.

The official Follow a goal docs describe /goal as a way to keep Codex working toward one durable objective instead of stopping after one normal turn. OpenAI says it is for tasks with a clear target, a validation loop, and enough room for autonomous progress, and that runs can continue for multiple hours.

The basic contract is simple:

• Enable it in /experimental, or set goals = true under [features] in config.toml, per the docs.
• Set a goal with /goal ..., inspect status with /goal, and control the run with /goal pause, /goal resume, or /goal clear, per the same guide.
• Define "done" up front. In reach_vb's explanation, the pattern is mission, completion criteria, then let Codex keep going.

OpenAI's own examples center on migrations, big refactors, long-running experiments, prototypes, and prompt-optimization loops. That matches yacineMTB's training-run example and yacineMTB's bisect example, which both treat /goal as a wrapper for work that would otherwise need constant re-prompting.

The first useful reveal is not that /goal exists. It is that people immediately used it for very long sessions. steipete's screenshot shows "Goal achieved (17h 59m)," while mattlam_'s post says one run was still going after seven hours.

That changes the shape of Codex usage in a few concrete ways:

• sound4movement's walk-away post describes the loop as /goal, leave, come back, goal cleared.
• mattlam_'s screenshot shows OpenAI had already "10x'ed" that user's Codex rate, alongside a live "Pursuing goal" session.
• TheRealAdamG's retweet of andrewchen says the feature looks likely to multiply token use sharply.

One open question from the tweet pool was whether OpenAI expected this kind of extended unattended behavior, or whether users were just pushing an interactive feature to an extreme. The official docs resolve that pretty directly: they tell users to set checkpoints, define verification artifacts, and expect compact progress updates during multi-hour runs.

The second-order effect landed almost immediately in tooling around Codex. pvncher's release note says RepoPrompt 2.1.23 added native support for experimental /goal, and pvncher's orchestration screenshot shows the pattern clearly: a higher-level workflow dispatches a builder agent, waits on checkpoints, and keeps monitoring targeted tests until the goal run finishes.

That screenshot surfaces a structure worth pulling out:

• Decompose one tightly coupled implementation item.
• Dispatch a narrow-scope builder run.
• Wait in bounded intervals.
• Watch for validation and edge-case fixes.
• Finish with a spot check against the plan.

doodlestein's orchestration post describes the same general move from another angle, pairing goal mode with a custom orchestration tool and reusable "super skills." The result is that /goal is already being treated less like a standalone slash command and more like a long-horizon primitive other tools can wrap.

Remote control is more provisional, but the pieces line up. testingcatalog's screenshot showed a new Connections settings panel with toggles for "Allow other devices to connect" and "Keep connection alive," plus an authorize flow for remote control. Then testingcatalog's correction narrowed the claim: the keep-alive toggle is for incoming connections, where you control Codex from other devices.

That wording matches an OpenAI GitHub issue on SSH remote projects, where a user describes the current desktop flow as Settings > Connections, refresh, enable the discovered SSH host, then create a new remote project. It also matches an earlier GitHub discussion asking for remote control from the ChatGPT app, specifically for a headless Codex running on a laptop or desktop and steered from mobile.

The community read on rollout is blunt. In mattlam_'s post about remote control, the flow is start a session locally, remote control it, then continue on Codex iOS. There is still no dedicated official announcement page for that handoff flow, so for now the strongest evidence is the settings UI, the SSH remote-project issue, and user posts saying the feature has started to ship.

The broader 0.129.0 release matters because /goal did not land alone. The official release notes bundle three related changes that all make longer-lived sessions easier to manage:

• A redesigned resume and fork picker, plus raw scrollback mode and workspace-aware /diff.
• Plugin workspace sharing, access controls, source filtering, and remote bundle sync.
• Browsable /hooks, including hooks that can run before or after compaction.

mattlam_'s demo of /resume and /fork adds the missing product texture: searchable past sessions, preview summaries, and full transcript viewing. testingcatalog's Usage tab post adds one more fresh clue that did not appear earlier in the story, a new Usage settings view. Paired with long unattended runs and the rate-limit screenshots above, that suggests OpenAI is adding more visibility right as Codex starts consuming far more session time than a normal prompt-response loop.