Codex Task Links on iOS: Previews and Approval Updates

Codex task links on iOS now open directly in ChatGPT for iOS 1.2026.272, according to OpenAI’s October 7, 2026 changelog. For people checking a coding task away from their computer, the useful change is a more direct route into the particular task they want to inspect. Our recommended review order is simple: confirm the task, inspect its preview and changes, then decide whether the requested next action fits the work you authorized.

What changed for Codex task links on iOS

The October 7 entry names two new features: inline page previews in responses and support for opening Codex task links directly on iOS. It also reports improvements to the approval picker, streaming in long conversations, new tasks keeping the selected computer while it reconnects or is offline, and worktree files matching the desktop checkout.

OpenAI separately lists fixes for empty task lists after connecting and status updates appearing as questions. The release follows the October 2 version covered in our ChatGPT iOS widgets and usage update. That earlier article covers widgets; this one focuses on entering and reviewing a specific task.

Check the task before sending a follow-up

OpenAI’s mobile guide describes following tasks, sending instructions, approving actions, and inspecting responses, changed files, diffs, and test results. Tasks can run in Cloud or on a connected computer; availability depends on rollout and workspace settings. Those execution choices matter when a link takes you straight into an existing conversation.

We recommend using a short identity check before adding instructions. Compare the opened task’s title and recent request with the task you intended to reach. If you have several similarly named conversations, use a distinctive detail from your original request, such as the filename or the expected change, rather than relying only on the title. Then check any visible computer or project context. If the destination is unclear, return to your known task list instead of starting a second task to compensate.

For example, “Change the checkout button label in the staging branch” and “Review the production checkout” should lead to different decisions, even if both conversations contain the word checkout. A mobile follow-up should preserve that distinction. Send the specific correction or question you need, with the affected path, rather than assuming the linked conversation has the context of another task.

Use inline previews as a review aid

The changelog does not specify which page types are supported or say that every preview is editable. For a website task, we recommend treating a preview as one piece of evidence alongside the changed files and test results described in the mobile guide.

Begin with the requested visual result: the heading, label, spacing, or interaction you asked to change. Compare what you can see with that acceptance condition. Next, inspect the associated diff where available. A convincing preview does not answer whether an unrelated file changed, whether the intended checkout contains the edit, or whether a relevant test passed.

If the preview and diff disagree, ask the task to identify the artifact and changed path it is showing before authorizing further work. For a fuller review checklist, our frontend release-check skill separates visual checks from release decisions. Opening a preview is a useful inspection step; deployment remains a separate decision.

Review permission choices in the current task

OpenAI’s mobile guide says users can review requested commands and actions before work continues on a connected computer. The release note does not document the new picker’s exact labels or defaults.

Our recommendation is to read the requested action in context, not choose by the visual prominence of a button. Ask what resource it affects and whether that operation is part of the task. A test command, a file edit, and a production deployment have different consequences. If a request goes beyond the agreed change, ask for a narrower action or explanation.

Also distinguish progress information from a request for input. If the task is merely reporting a stage, avoid sending an instruction that unintentionally changes its scope. If it actually needs a decision, answer that decision explicitly. The related Codex CLI reconnect and permissions update concerns the terminal client, not the iOS picker’s undocumented controls.

Confirm the host and checkout after reconnecting

OpenAI’s remote-connections guide explains that host-based work uses the connected computer’s resources and permissions. Remote access stops when the host sleeps, loses its network connection, or closes the app. Keeping a selected computer during a connection interruption should therefore not be read as a promise that its local commands execute while it is unavailable.

The worktree documentation makes another distinction: worktrees are Git checkouts on the connected computer or remote development environment, not repositories running locally on the phone. After a reconnection, our suggested check is to compare the task’s expected changed path with the actual diff on the desktop checkout used by that task.

An original, untested acceptance exercise is to choose one already-authorized heading change in a disposable project. Record the task title, intended host, expected file, and acceptance condition. Open its existing task link on iOS, inspect the available preview and diff, and compare those observations with desktop. Record a mismatch precisely: wrong conversation, wrong host, missing file, or different rendered text. Do not deliberately disconnect an active production task to test recovery. This exercise is proposed review guidance, not a claim that we have measured the release’s behavior.