Codex users report better compaction and Colab control after v0.133.0

• OpenAI's Codex update cycle centered on /goal, which OpenAIDevs' launch thread said now runs across the app, IDE extension, and CLI for hours or days, while LLMpsycho's v0.133.0 summary tied that rollout to the CLI release that made goals default.
• The same wave shipped remote Mac control while locked, which OpenAIDevs' locked-use post described as phone-driven computer use with the screen off, and users like reach_vb's Colab run report immediately pushed it into browser automation workflows.
• Hands-on reports from bclavie's compaction comparison, dejavucoder's reply, and steipete's repost of a compaction claim converged on the same theme: Codex compaction feels materially stronger than Claude Code in long sessions.
• The product surface widened at the same time: OpenAIDevs' plugin-sharing post added team plugin distribution, jxnlco's skills-repo tip pointed users to openai/skills, and mattlam_'s release read flagged new lifecycle hooks that look like groundwork for more durable goal orchestration.
• The rollout was not perfectly clean, because dkundel's setup fix said some app installs still needed manual feature enablement, while swyx's regression report said the new build introduced approval pauses even in Full access mode.

You can scan the v0.133.0 release page, the goal mode docs, the locked computer-use docs, and the remote connections docs. There is also an official openai/skills repo, a new skills installer tip from jxnlco, and a practitioner video from Matt Pocock's /handoff walkthrough that frames how users are working around context limits.

OpenAI moved /goal out of experiment status and into the center of Codex. According to OpenAIDevs' launch thread, goal mode now works in the app, IDE extension, and CLI, and lets Codex keep pushing toward a milestone across long-running sessions.

The CLI release framed the same change more mechanically. mattlam_'s release read said /goal now has a dedicated SQLite database tracking active goal state, status, and token usage, and that extensions can observe tool, subagent, and turn lifecycle events.

That pairing matters because it makes /goal look less like a prompt shortcut and more like a durable execution layer:

• goal state is now stored separately, per mattlam_'s release read
• continuations were already patched in v0.132.0 to stop on usage limits or repeated blockers, per mattlam_'s v0.132.0 note
• users can pause and steer a running goal mid-conversation, per swyx's pause-and-adjust post and dkundel's mid-conversation tip

The rollout also had a small papercut. dkundel's setup fix said some app users still had to run codex features enable goals and restart before the default actually showed up.

The flashiest product change was remote computer use on a locked Mac. OpenAIDevs' locked-use post said Codex can keep using apps from a phone even while the Mac screen is off, and dkundel's clarification added that the feature specifically targets computer use and still displays on-screen when Codex is controlling the machine.

Practitioners immediately treated that as more than a demo. In reach_vb's Colab run report, Codex used Chrome against a signed-in Colab session, handled runtime and editing issues, trained a 10.65 million parameter transformer in about 19 minutes, then used subagents to audit the result. Later, jxnlco's Cloudflare post and jxnlco's follow-up described Codex setting up D1 and editing Workers directly from the dashboard.

Those reports sit on top of the earlier mobile-to-desktop handoff OpenAI described in its remote connections docs and OpenAIDevs' mobile thread. The new locked-use mode removes the last obvious requirement that the Mac stay awake and foregrounded.

OpenAI did not headline compaction in the launch posts, but users did. bclavie's comparison said switching between Codex and Claude Code creates "massive whiplash" because Codex survives compaction without seeming to forget where it was, and dejavucoder's reply made the same point more bluntly.

That shows up in workflow choices. PerceptualPeak's question said GPT 5.5 in Codex appears capped at a 258k window while GPT 5.4 gets 1M, but the thread treated stronger compaction as a reason the smaller window might still be viable. mattpocockuk's /handoff skill post points to the same pressure from another angle, replacing /compact with a handoff pattern that moves work between agents instead of relying on one swollen thread.

The strongest anecdote came from AlemTuzlak's migration post, who said a task that took three to five days with Opus took one day after moving the same setup to Codex. That is still a single-user report, but it matches the broader pattern in the compaction chatter.

The other story inside v0.133.0 is ecosystem shape. OpenAIDevs' plugin-sharing post said Business workspaces can now distribute custom plugins across teams, while OpenAIDevs' analytics post added org-level tracking for active users, credits, tokens, runs, plugin usage, and lines of code generated.

Users are already treating those surfaces as installable operating layers rather than one-off prompts:

• jxnlco's skills-repo tip pointed users to the official openai/skills repository and said Codex can scan it and install relevant skills
• NickADobos' reply suggests some users assumed open source skills would already appear inside the app's skills page
• AlemTuzlak's thread described a reusable /doc-writing skill for maintaining TanStack AI docs
• mattlam_'s release read highlighted extension lifecycle hooks, which makes plugins and skills easier to wire into long-running agent flows

Appshots fits the same pattern. OpenAIDevs' Appshots launch said pressing Command-Command on Mac attaches both a screenshot and extracted text from the active app window, including content beyond what is visible onscreen, and gdb's post shows the feature already surfacing inside the Codex app.

The clean launch narrative broke in a few places. swyx's regression report said updating the app for /goal led to repeated approval pauses despite Full access mode being enabled, which he described as a step back in autonomy.

OpenAI kept shipping small UI fixes around the same window. OpenAIDevs' diff markers post added optional classic + and - markers for diff review instead of color-only bars, and reach_vb's app-only workflow post said the Codex diff and file viewer had already become good enough that he had not opened an IDE in more than a month.

That combination is a useful tell on where Codex is heading. The big features are long-running goals, computer use, and shared plugins. The small ones are approval behavior, diff visibility, and app ergonomics, which is exactly the layer that determines whether users stay inside Codex all day or bounce back to an IDE.